Morbidity and Mortality: How Do We Value the Risk of Illness and Death? Proceedings of Session I: Risk Assessment and Valuation of Health Effects from Air Pollution (including Introductory Remarks)


Morbidity and Mortality: How Do We Value the Risk of

Illness and Death?

PROCEEDINGS OF SESSION I: RISK ASSESSMENT AND VALUATION OF
HEALTH EFFECTS FROM AIR POLLUTION (INCLUDING INTRODUCTORY

REMARKS)

A WORKSHOP SPONSORED BY THE U.S. ENVIRONMENTAL PROTECTION
AGENCY'S NATIONAL CENTER FOR ENVIRONMENTAL ECONOMICS AND
NATIONAL CENTER FOR ENVIRONMENTAL RESEARCH

April 10-12, 2006

National Transportation Safety Board
Washington, DC 20594

Prepared by Alpha-Gamma Technologies, Inc.
4700 Falls of Neuse Road, Suite 350, Raleigh, NC 27609

ACKNOWLEDGEMENTS

This report has been prepared by Alpha-Gamma Technologies, Inc. with funding from
the National Center for Environmental Economics (NCEE). Alpha-Gamma wishes to
thank NCEE's Maggie Miller and the Project Officer, Cheryl R. Brown, for their
guidance and assistance throughout this project.

DISCLAIMER

These proceedings have been prepared by Alpha-Gamma Technologies, Inc. under
Contract No. 68-W-01-055 by United States Environmental Protection Agency Office of
Water. These proceedings have been funded by the United States Environmental
Protection Agency. The contents of this document may not necessarily reflect the views
of the Agency and no official endorsement should be inferred.

-------
Table of Contents

Introductory Remarks

William H. Farland, Chief Scientist, Office of the Science Advisor, and Acting Deputy
Assistant Administrator for Science, Office of Research and Development, U.S. EPA

Session I: Risk Assessment and Valuation of Health Effects From Air
Pollution

Session Moderator: Ron Shadbegian, U.S. EPA, National Center for Environmental
Economics

Willingness To Pay for Improved Health: A Comparison of Stated and
Revealed Preferences Models

Michael Hanemann, University of California-Berkeley, and Sylvia Brandt,
University of Massachusetts-Amherst

Individual Preferences and Household Choices: The Potential Role of
Dependency Relationships

Mary F. Evans, University of Tennessee-Knoxville; Christine Poulos, Research
Triangle Institute; and V. Kerry Smith, North Carolina State University

Preliminary Results From a Daily, Time-Series Study of Air Pollution and
Asthma in the San Francisco Bay Area

Charles Griffiths and Nathalie Simon, U.S. EPA, National Center for
Environmental Economics

Discussant: Bryan Hubbell, U.S. EPA, Office of Air Quality Planning and
Standards

Discussant: Glenn Blomquist, University of Kentucky
Questions and Discussion

-------
U.S. EPA NCER/NCEE Workshop
Morbidity and Mortality: How Do We Value the Risk of Illness and Death?

Washington, DC
April 10-12, 2006

Introductory Remarks
William H. Farland, Chief Scientist, Office of the Science Advisor, and Acting
Deputy Assistant Administrator for Science, Office of Research and Development

This is the 13th of the Economy and Environment Series workshops that have been
sponsored by the Office of Research and Development's National Center for
Environmental Research and the Office of Policy, Economics, and Innovation's National
Center for Environmental Economics. The opportunity here is really to bring together a
group of colleagues to think about some of the issues around approaches to valuation of
human health effects—both mortality and morbidity. These kinds of issues are
particularly important to the Agency.

The group that we expect to be interacting with over the next two-and-a-half days makes
up a broad section of the scientific community and the economic community. Clearly,
our colleagues from the National Center for Environmental Economics will be
represented here as well as our Science To Achieve Results grantees, who have the
opportunity to use Agency resources to explore many of these kinds of issues. We also
expect EPA economists and other scientists from a number of our program offices as well
as ORD and OPEI. In addition, we had a broad solicitation and have a number of
researchers from academic institutions and other federal agencies, and we're particularly
pleased to welcome our international guests for this program.

Again, our purpose here is really to learn more about research that improves our
understanding with regard to the valuation of health outcomes, and to assure that the
research that is done really feeds into opportunities for improving how we make our
decisions. Now, I want to spend just a few minutes on this idea of the importance of the
workshop, because as we look at research results, as we begin to think about the kinds of
data that come out of the work that you all do, whether they're out of the academic or
federal community, these results need to be credible and relevant and timely so they can
inform the kind of policy decisions that EPA and many other federal agencies are making
every day.

Clearly, the value of cost/benefit analysis is very high. If we look at things like the
Executive Orders that are coming out and new legislation like the Safe Drinking Water
amendments that really focus on having this type of information in order to judge the
quality and relevance of particular options for environmental decisions, we see that this
becomes extremely important. It really affects the way we do business now at the
Agency for many of the things that are important to us. Clearly, accurate cost/benefit
analysis leads to better decisions. It allows us to really understand, where appropriate,

W. Farland, Introductory Remarks

1

-------
the role of economics in terms of environmental decision making, and we're really proud
of the good relationship that has developed between the Office of Research and
Development and OPEI with regard to these particular issues and advancing the state of
the science.

So, as I said, the idea of human health valuation is a very important topic for the Agency.
Clearly, it drives much of what we do. Particularly, as you begin to look at the issue of
morbidity and mortality, we realize that our community has recognized that getting a
handle on approaches to assessing morbidity is really a top priority for the field, and it
strongly highlighted in the most recent environmental economics research strategy that
the Agency has put forward. The challenge of trying to deal with valuation of morbidity
is one that we'll talk a little bit more about, but it clearly is something that we need to put
our collective best minds toward as we begin to think about how we approach it.

As you know, mortality valuation is always high profile and plays a very, very large role
in many of the decisions that we make. It is the one area where we have made some
significant progress in valuation. For some of the most recent Clean Air Act decisions,
for instance the particulate matter decisions that have led to rules recently, more than 90
percent of the monetized benefits come out of these mortality valuations. So, again, we
recognize the importance of mortality in trying to get the approaches to mortality
valuation right, but we also realize the issue of how we approach morbidity and the very
strong role that non-monetized benefits currently play and the importance of trying to
value those.

So, over the next two-and-a-half days you'll have an opportunity to deal with a number of
very important issues. Some of these are old favorites—if we think about asthma,
pollutants in drinking water, lead paint and IQ loss, PCBs, and so on. Those are issues
that we've all been dealing with for quite a while, but the field is really opening up for us.
There are opportunities here for us to really emphasize things like children's health
valuation. As you begin to think about that, think about it from the standpoint of children
being differentially susceptible, in many cases, to some of the pollutants that we're
dealing with. So, clearly their life stage and their lifestyle—their behavior—is very
different from that of adults. At the same time, children are not full economic actors, if
you will, so trying to look at the valuation of children's health impacts and even dealing
with the issue of early life exposures leading to later life impacts becomes a real
challenge for us. Clearly there are opportunities here for methodologic advances in
valuing health risk reductions and modeling household decision making—again, a
challenge for us as we think about the fact that our households are changing and the types
of situations that we're looking at now are particularly important.

We're looking for an opportunity to deal with the issue of updates on cost-effectiveness
analysis, mortality valuation, and efforts to include economic questions in our large-scale
health surveys. Clearly, this is an integration of the field of monitoring and modeling
health effects with the economic valuation of those effects. We are expecting to have a
real opportunity to move forward.

W. Farland, Introductory Remarks

-------
One of the things that we're intending to do later on in the session is a panel discussion
on the pros and cons of web surveys. Many of you are building web surveys into your
protocols, and in the most recent OMB guidance there has been some concern raised
about the use of web surveys. This is an approach that is heavily in use, and it's
something that we need to look at very carefully so we understand the pros and cons.

Finally, at the end of the session, during the last half day, is something that we've not
done before. It is an opportunity to focus, in depth, on a particular set of research results
from a single grant that EPA has funded. This is a grant that has gone to UCLA and
Oregon State investigators, and we're going to be very interested in your feedback on this
particular approach for the workshop. We will focus in on a particular set of results and
really have a half-day, in-depth discussion on that.

So, clearly, the research results that we are going to be talking about are going to be very
important, and they are currently being used by the Agency. We're looking for
opportunities to do things even better than we have in the past. Clearly, the record shows
that previous results have been used in important analyses—our Section 812 report in the
Clean Air Act, which actually lays out costs and benefits of Clean Air Act decision
making, has illustrated the way that the Agency has been successful in laying out the
economic benefits of the Clean Air Act. It is actually among the leading regulatory
actions with regard to monetized environmental benefits of any of the actions that we
take.

The results of our research have been cited by OMB in their guidance on mortality
valuation. Some of you may remember the discussions that came forward on exactly
how we were going to value mortality for the elderly and the advice that we got from
OMB with regard to the so-called "senior death discount" and not discounting the issues
with mortality later in life.

Certainly we expect the future to hold expanded use of these techniques, and we're
looking forward not only to improving the approaches that we use for this, but also to
being able to demonstrate results—to deal with things like our program evaluation ratings
and other opportunities that we have to demonstrate how research is used to inform
decision making and how those decisions can lead to improved environmental results.

So, looking toward the future, what research will we be looking to fund and support?
Clearly, some of that will depend on the kinds of discussions that we have over the next
two-and-a-half days. A lot of these issues have been foreshadowed in the Environmental
Economics Research Strategy, and this provides a good opportunity for us to work within
a framework of important research needs. Some of the things that are clearly going to be
part of that research that we fund have to do with the fact that they will be results-
oriented types of work; they will be things that we can apply routinely; they will focus on
issues such as the question of benefit transfer, which is something that is particularly
important as you get into understanding particular situations and applying that to a
broader population or a broader situation. Another important issue will be the question of
marginal risk changes, so that we can really get at the question of how we monetize over

W. Farland, Introductory Remarks

-------
time with changes that occur with regulatory activities. It's very clear that the field is
moving toward an inter-disciplinary approach, much like many of the fields that we
interact with at EPA. We're looking forward to meetings like this one to really have an
opportunity to hear from various disciplines that have a role to play in the important
research that's going on.

With that I'll close and wish you well in terms of the research discussions that will occur
over the next two-and-a-half days. I'd like to thank Will and the other organizers from
both of the offices who co-sponsored this. I look forward to a very successful workshop.
Thanks.

W. Farland, Introductory Remarks

4

-------
DRAFT

Willingness to Pay for Improved Health: A Comparison of Stated and

Revealed Preferences Models

W. Michael Hanemann
Department of Agricultural & Resource Economics
University of California, Berkeley
hanemann@are. b erkel ey.edu

Sylvia Brandt
Department of Resource Economics
University of Massachusetts, Amherst
brandt@resecon.umass.edu

Prepared for presentation at the EPA Workshop Morbidity and Mortality: How Do We Value the
Risk of Illness and Death? Washington, DC, April 10-12, 2006.

This research was funded through the US EPA-STAR Valuation of Human Health Program (R-
82966501, Valuing Reduced Asthma Morbidity in Children). This article has not been formally
reviewed by the EPA. The views expressed in this document are solely those of the authors and
the EPA does not endorse any product or commercial services mentioned in the publication.

-------
Abstract:

In this paper we discuss two approaches to estimating the willingness to pay (WTP) for reduced
asthma morbidity, contingent valuation and health production function. The study population
includes 250 children ages 5-11 with clinically diagnosed asthma, residing in a section of Fresno
County, California. Asthma symptoms, including coughing, wheezing and/or shortness of breath,
ranged from mild and intermittent to severe and persistent in this group. Detailed health
measures (including atopy and pulmonary function), utilization of health services, levels of
antigens in the households and exposures to criteria air pollutants were collected as part of a five-
year epidemiological study. We administered two economic surveys to measure 1) households'
perceptions of risks to an asthmatic child, 2) averting and/or mitigating actions taken, and 3)
households' stated willingness-to-pay for a reduction in their children's asthma morbidity.

In the health production model the health outcome is a function of exposure to asthma triggers,
mitigating and averting behavior and household's perceived risks. We find that variation in WTP
is explained by attitudes towards asthma specific health investments including concerns of
associated risks and perceived effectiveness. The survey data indicate that households select
from a small number of discrete health investments and that most risk reducing behavior are
daily behavioral modifications with no relevant market prices.

We argue that the discrete nature of health investments and socio-cultural patterns of health care
utilization make the revealed preference approach inadequate for the case of asthma. As an
alternative we present a contingent valuation scenario that was specifically developed to
minimize systematic variation in preferences for characteristics related to the scenario rather than
the reduction in asthma morbidity. For this purpose, guided by extensive testing in focus groups,
we selected a scenario based on a hypothetical asthma monitor that provides to the wearer an
indicator of current asthma status.

-------
INTRODUCTION

The economic concept of value implies a tradeoff. The monetary value of any item is
defined in economics as the amount of money that a decision-maker - an individual, a
household, or a firm, depending on the context - would be willing to exchange for the item. That
monetary amount measures the worth of the item in monetary units in the sense that the
exchange of this monetary amount has the same impact on the decision maker's wellbeing
(utility) as the item itself.1 The challenge for economic measurement is to identify a trade-off
through which value can be measured. Revealed preference approaches work by observing actual
choices by decision-makers and inferring the trade-off underlying these choices. Depending upon
the nature of the choice (whether it is a discrete, continuous, or mixed discrete/continuous
choice) the choice behavior may reveal the trade-off either directly (a simple discrete choice) or
indirectly (the cases involving continuous choices) by permitting the identification of an
underlying set of preferences which had motivated the observed choice behavior. In the latter
case, the trade-off is inferred from the recovered preferences underlying the observed choice
rather than directly from the observed choice itself. Stated preference approaches work by
placing subjects in a survey or experimental setting and confronting them with choices that,
directly or indirectly, reveal their preferences.

In the context of valuing health outcomes, the standard revealed preference approach
assumes that health-related choice behavior reflects preferences for health outcomes that are
generated by a perceived health "technology". This separation between preferences and
production requires the researcher to differentiate between behavior that is an end in itself and
behavior that is a means to an end. Consider, for example, assessing the value of good water
quality at a beach from this perspective. In the case of amenity value, an individual's choice of
which beach to visit (trading off cleaner but more distant beaches versus dirtier but closer
beaches) bears directly on the trade-off of interest since going to a nice beach is presumably an
end in itself. In the case of health outcomes, an individual's choice of which precautions to take
(spending money to purchase goggles, taking an antibiotic before going surfing, etc) is a means
to an end - namely, good health - rather than an end in itself from which the individual derives
enjoyment per se. In the latter case, the valuation analyst has to disentangle the production
component from the pure preference component that underlies the sought-after trade-off. We
suggest that this complication may sometimes tilt the balance in favor of stated preference rather
than revealed preference as the preferred valuation approach2.

1	Generically, there are two ways to formulate the exchange: the maximum amount that the individual would be
willing to pay (WTP) to obtain the item, if it is favorable, or to avoid it, if unfavorable; and the minimum amount of
money that the individual would accept (WTA) to forego the item, if it is favorable, or to endure it, if unfavorable.
The relationship between WTP and WTA is a separate issue that will not be pursued here. For simplicity, the
discussion below focuses on the WTP measure of welfare.

2	An important consideration in modeling health outcomes for children is the question of the identity of the decision
maker. The decision maker is surely not the child but rather one or both of the parents; therefore, the framework is
the household rather than individual decision making. Making the household the unit of analysis raises several
important but difficult analytical issues that are addressed in other literature. In this paper we focus on the
relationship between health preference function and health production function, and we make the simplifying
assumption that household decisions regarding children's health reflect a unitary model of household preference and
production.

-------
We examine the application of the revealed preference and stated preference approaches to the
valuation of reduced asthma morbidity. Our economic study was done in collaboration with an
epidemiological study that was the most detailed socio-demographic, indoor air quality and
pollution monitoring data collection effort to date (California Air Resources Board). Findings
from multiple focus groups and two economic surveys suggest that the discrete nature of health
investments and socio-cultural patterns of health care utilization make the revealed preference
approach inadequate for the case of asthma. As an alternative we present a contingent valuation
scenario that was specifically developed to minimize systematic variation in preferences for
characteristics related to the scenario rather than the reduction in asthma morbidity.

This paper is organized as follows. In Section One we describe the epidemiological study and
economic surveys used to collect household level data. Second, we summarize the average
households expenditures related to asthma morbidity and conceptual limitations to using these
costs as a measure of value. In the third section we present the standard household health
production model. Conceptual limitations to the standard model are presented in the fourth
section. Fifth, we present empirical evidence of these complexities and their implications for the
household production model. Next we discuss how we used the findings from the first economic
survey to create a contingent valuation scenario. Concluding remarks are included last.

1. Empirical Study

A. Study Setting

This project is a collaboration with an extensive epidemiological study of the effects of air
pollution on asthmatic children [Fresno Asthmatic Children's Environment Study, FACES], The
study is located in Fresno, California, which has highest rate of asthma hospitalizations in
California at 28.8 per 10,000 (California Facts, 2003). Located in the Central Valley of
California, Fresno County has a population of 815,734 and this population has increased by
19.8% since 1990. Forty-four percent of the population is of Hispanic or Latin origin, followed
by forty percent of white origin, eight percent Asian and five percent African-American. The
Fresno population has lower median income, less education, poorer living conditions and a
greater percent of residents below the poverty line as compared to the rest of CA. For example,
median household income for 2001 was $34,725 as compared to $47,493 for California. The
proportion of residents with a high school degree was 67.5% as compared to 76.8% for the rest
of the state, and the proportion of residents below the poverty line was 22.9 % while that in CA
was 14.2% (US Census data, 2000).

The FACES cohort included children with clinically diagnosed asthma, residing in a section of
Fresno County, California3. Children were 6-10 years of age at intake and were followed for
approximately 4 years. The study population included children who had a physician's diagnosis
of asthma and at least one of the following: 1) reported utilization of or valid prescription for
asthma medication in the previous 12 months; or 2) symptoms consistent with asthma in the past
12 months; or 3) an emergent asthma visit or hospitalization in the past 12 months. The
requirements for asthma medication use, symptoms, or health care utilization are to minimize the
chance of enrolling subjects whose asthma is quiescent (remission). Children who meet these

3 FACES has been recruiting households for the survey since 2000.

-------
criteria may be enrolled regardless of the severity of asthma. Children with major comorbidities
that would confound the measurement of pulmonary function were excluded.

The FACES study screened 473 households, completed baseline interviews for 241 households,
and retained 205 participating households. The major reasons households who inquired about the
study were ineligible to participate include: other chronic disease, lived in house for less than
three months, child sleeps at home less than five nights/week, and family planned to move within
two years (Mann, 2003).

Demographics and characteristics of the FACES cohort are in tables 1. The percentage of blacks
enrolled in the FACES program (13.7%) is greater than the percentage of blacks for the Fresno
population (5.3%), while Asian Americans are underrepresented. The average age of children in
the FACES cohort is between eight and nine years. The majority of the interviewed households
were covered by health insurance (90.3%). Almost 70% households had at least one parent who
was affected by asthma. One observable characteristic of the FACES cohort that differs from the
Fresno population is the frequency of smoking in the home.

Table 1: Demographics of FACES Cohort

Demographics

Relative Frequency

Race

White

40.7

Hispanic

41.8

African American

13.7

Other

3.8

Male

At least one parent employed

At least one parent completed high school

Participant's health status

Ever prescribed oral steroids

One or more hospitalization(s)

Positive skin test to at least one antigen

FEV1

1.06 (range, 0.51-1.59)

Any smokers in home

23%

Asthma severity

step 1

step 2

step 3

step 4

Note: Based on baseline interviews completed as of June 30, 2002 (n=182). Severity scores based on the

NHLBI guidelines.

B. Economic Surveys

Two economic surveys were conducted in the FACES cohort. The first survey, a written mail
survey was conducted February to August 2004. This survey included detailed questions on
asthma related expenditures, asthma related symptoms and activity limitations, and health

-------
beliefs. A total of 202 households completed the first survey (representing 209 children with
asthma). The second survey contained a contingent valuation scenario and was conducted
October 2005-February 2006. The purpose of completing two surveys was to explore the
strengths and limitations of the two approaches to valuing children's health: stated preferences
and household health model. In this paper we present the findings from the first survey to
motivate our design for the contingent valuation scenario.

The median household size those completing for survey one was 4, (range: 2-9), and 41% of the
households had two children under the age of 18. The survey respondent was typically the
household member who interacted with the healthcare provider: 95% responded that they are the
ones to take children to medical appointments. Employment status of the respondent varied: 37%
were employed full-time, 27% were employed part-time, 11% were not employed but were
looking, and 24% were not employed outside the home and were not looking for employment.
The distribution of household income for participants that completed the first economic survey is
reported in Table 2.

Table 2: Household Income

Household income

Relative Frequency

Less than $10,000

11.5

$10,000 to less than $20,000

11.5

$20,000 to less than $30,000

8.2

$30,000 to less than $40,000

15.9

$40,000 to less than $50,000

11.5

$50,000 to less than $75,000

19.2

$75,000 to less than $100,000

12.6

$100,000 or more

9.3

Note: Total responses =182

Tables 3 and 4 describe the asthma status participants who completed the first economic survey.
Following the GINA recommendations, asthma severity was based on frequency of daytime and
nighttime symptoms. Less than 20% of the children had severe day or night symptoms. More
children had moderate symptoms in the day than at night (43% versus 35%), and consequently
more children had mild symptoms at night than during the day (46% versus 38%). There was a
high degree of correlation between day and nighttime severity.

Table 3: Count of Asthma Severity by Day and Night Symptoms

Nighttime Symptoms

Daytime symptoms

Mild

Moderate

Severe

Total

Mild

Moderate

Severe

Total

202

Note: Mild is defined as symptoms less than two times a week. Moderate is defined as symptoms 3-5
times a week. Severe is defined as symptoms every day.

-------
The majority of respondents had prescriptions for a rescue and controller medication.

Table 4: Frequency of Medication Usage

Number of Medications

Percentage who have prescriptions for:

Control Medication

Rescue Medication

None

3 or more

Of the survey respondents 79% reported that no one smokes in the home: of respondents that
reported smoking in the home 58% reported that the father smoked, 28% reported maternal
smoking and 20% reported that another adult smokes in the home. These smoking rates are
below that for the Fresno population, but are inline with those for the FACES cohort.

2. Costs related to Asthma

A. Health expenditures

Direct expenditures on asthma were broken into four categories: fixed costs4, household supplies,
pharmaceuticals (prescription and over-the-counter) and alternative therapies. Variable costs are
the sum of supplies, pharmaceuticals and alternative therapies.

Table 5: Asthma Related Expenditures

Median

Mean

Standard
deviation

Fixed costs

240

357

716

202

Variable costs

110

139

114

199

Household supplies

202

Pharmaceuticals

202

Alternative therapies

199

Note: Fixed costs included service costs or purchases of: air filters, allergy mattress covers and/or pillow
covers, humidifiers, dehumidifiers, air conditioners, HEPA vacuum cleaner, landscaping of yard, carpet
removal, pest extermination, mold/mildew removal, removing pets, humidity gauge, nebulizers, peak flow
meters, spacers for inhalers, replacement for window coverings, and fans. Household supplies include:
replacement filters for air filters, filters for air conditioners, HEPA vacuum filters, heater filters, cleansers
for mold and mildew, hypoallergenic or non-aromatic cleaners, allergen control sprays, sprays for pest
removal. Pharmaceutical costs included: prescription asthma or allergy medication, over the counter
allergy or asthma medications, herbal remedies, and home remedies. Costs related to alternative therapies
include visits to: chiropractor, acupuncturist, doctor of osteopathy, homeopathic/herbalist, nutritionist,
spiritual healer, or other alternative health care provider.

Indirect costs related to asthma morbidity include employment impacts and time used for
planned or unplanned medical visits (including in-clinic and emergency room). Of the 202

4 For fixed investments we delineated 1) purchased specifically to help asthma 2) purchased prior to asthma
diagnosis 3) never purchased.

-------
households, 43% reported that they usually needed to take time off from work to take their child
for medical appointments, and the median time taken off from work for each medical
appointment was 75 minutes (mean 83 and standard deviation 37 minutes). Twenty- four percent
of the families reported that they had gone to the emergency room for the child's asthma in the
previous 12 months, requiring a median of 215 minutes per visit (mean 211 and standard
deviation of 128)5. Although lost work time due to asthma is one indirect cost of asthma
morbidity, 18% of the households reported changes to their employment due to the frequency or
severity of their child's asthma. Of those that had changed their employment status: 28% had
been subjected to employers reducing their work hours due to previously missing work for
child's illness; 22% had been fired or laid-off due to missing work for child's asthma; 38% had
chosen to not seek employment outside the home due to child's asthma; 32% chose to work only
part-time due to asthma; 27% worked fewer hours during asthma seasons. These statistics
suggest that limiting indirect costs to workdays lost underestimates the impact of asthma
morbidity on household income.

B. Limitations to Costs as a Measure of Value
There are three reasons why using expenditures and indirect costs as measures of value of
reducing asthma morbidity has shortcomings. The first is that there is an important difference
between the concepts of cost and value, the latter being the concept of interest to economists.
The second issue is the critical distinction between marginal and non-marginal value. The last
limitation is that these cost miss larger impacts of asthma morbidity on households. We discuss
each of these in this section.

Economists have been aware of the fundamental distinction between what things cost and what
they are worth ever since Adam Smith exposited the diamonds-water paradox (water is essential
for life but inexpensive, while diamonds are entirely inessential but extremely expensive). Both
may be important to know, but they are different things. What something costs is a question
about supply; what it is worth is a question about demand. What something costs is objective and
a matter of production and engineering; what it is worth is subjective and a matter of preference
and taste.

The distinction between marginal and total value also underlies the diamond-water paradox is. At
the margin, an additional kilo of water may have a lower value to people than an additional kilo
of diamonds because water is abundant while diamonds are rare, but the total value of all water
to mankind is likely to be greater than the total value of all diamonds: if we lost all access to
water, people would surely judge this a greater harm than losing all access to diamonds.

Similarly, Dupuit emphasized the distinction between marginal and infra-marginal value. Dupuit
observed that, for any consumer and any commodity, while the last unit of an item to be
consumed would be just equal in value to its price (which is why it is the last unit to be
consumed), the infra-marginal units would be worth more than the price because of the
phenomenon of diminishing marginal utility. For the infra-marginal units, the consumer would
be willing to pay more than the price. Hence, the consumer receives a sort of "profit" or
"surplus" on these units which he would lose if the item were not available at that price. This
observation provides the foundation for Dupuit's concept of consumer's surplus - the excess of

5 Total time spent on medical visits and emergency room visits will be reported in future work.

-------
what a consumer would be willing to pay for an item over and above what he actually does pay.
If the consumer had valued all units of the item exactly at their price his consumer's surplus
would be identically zero, but this does not generally happen. Dupuit's larger point was that
value is measured by reference to the consumer's demand curve - by what we now call the
Marshallian consumer's surplus6.

The third limitation of the expenditure approach is that beyond the indirect and direct costs there
are psychosocial impacts on the households from asthma morbidity. These impacts include
changing family activities, interactions with peers and the burden of uncertainty surrounding the
status of a child's asthma. For example, in focus groups parents frequently discussed difficulties
in communicating their child's needs to school officials and to physical education teachers in
particular. In our survey 21% (out of 170 responding) disagreed with the statement "My child's
classroom teachers are helpful with my child's asthma needs." and 24% (out of 140 reporting)
disagreed with the statement "The physical education teacher works with us to include my child."
Other impacts were restrictions on normal childhood play: 38% of respondents (n=200) reported
that they restricted the amount of child's activity more often specifically due to asthma; 46% of
respondents (n=201) reported that they restricted the amount of time outside more often
specifically due to asthma; 44% of respondents (n=201) reported that they restricted the where
the child could play or visit more often specifically due to asthma. A more dramatic, though less
frequently reported change (11 households out of 199 reporting), was moving to a new home to
avoid asthma triggers and to improve the child's asthma. The frequency of these impacts and
extent to which they affect quality of life suggests that using expenditures to measure value of
reduced morbidity misses the complexities of how asthma affects household behavior.

In short, economic valuation is genetically about what things are worth, not what they cost.

While costs may provide some information about how households value health, costs alone are
not adequate measures. In the next section we present a standard household health production
model and discuss some conceptual challenges in applying it to valuing children's health.

3. Revealed Preference Approach to Valuing Reduced Asthma Morbidity

A. Health Production Function

We begin with a standard model for household health and to then proceed to show how it is a
special case of the Lancaster-Maler utility model. The critical characteristic of the indirect utility
function of the household health production model is that it has the same structure of the indirect
utility function produced by the Lancaster-Maler utility model, and hence the implications of the
Lancaster-Maler for welfare measurement are applicable to the case of the health model. We
begin by developing the indirect utility function for the health model.

In the standard model, marketed commodities are divided into two groups, those which have
some relation to health (z) - either in preventing ill health or in curing illness once it occurs -
and those which have no relation to health (x). The corresponding price vectors are denoted px

6 Hicks (1941, 1943) formalized Dupuit's and Marshall's concepts of the difference between the value of infra-
marginal units price and their concept of consumer' surplus. Hicks formalized what Dupuit and Marshall asserted in
terms of what he called the compensating and equivalent variation.

-------
and pz with individual elements denoted p, and pj, respectively. One could further subdivide the
health related market consumption activities into those which promote good health and prevent
illness (e.g., taking asthma control medication regularly), zA, sometimes called averting
behaviors, and those which reduce the adverse effects of falling ill (e.g. taking an asthma rescue
medication), zM, sometimes called mitigating behavior, so that z = (zA, zM). For our present
purpose, we can just work with the vector z. Health status could be a scalar or vector of health
states or outcomes but, for simplicity, we will treat H as a scalar here.7 Finally, q is some
measure of environmental pollution that affects health. Thus, for the household there is a health
production function given by:

H = H(z,q)

where

z is a vector composed of averting behaviors (za) and mitigating behaviors (zm).

There are several alternative formulations of the household's preferences, depending on what
enters the household's utility function. Obviously, household health (H) and the consumption of
non-health-related market commodities (x) enter the utility function. The question is whether any
of the elements of z and/or q enter the utility function as well. The point is that, while z and q
affect household utility indirectly through their influence on health/illness, H, they could also
affect household utility directly if it cares about q or z for motives unconnected with their effect
on H. The empirical evidence from our surveys suggests that both z and x are important elements
of the household's utility. Thus we will use the most general case is where all of the variables
affect household utility directly, and the household maximizes utility subject to the health
production function and a budget constraint (Y= income):

Maxxz U = U(x, z, q, H)

subject to H = H(z,q) and pixi + pjzj = Y

The result is a set of ordinary demand functions for all market goods, both non-health-related and
health-related, x; = xi(px,pz, q, Y) and zi = Zj(px,pz, q, Y) and a corresponding indirect utility

function v(px, pz, q, y).

If we compare the indirect utility function produced in the health model v(px,pz, q, y) to that in
the generalized Lancaster-Maler utility model, v(p,q,y), we can see that the former is a special
case of the latter in which prices have been partition into non-health related good and health-
related goods. The difference between the Lancaster-Maler model and the standard household
health production model is simply that the household health production model makes the health
production function explicit and implies that the production function can be estimated separately
from the pure preferences represented by U(x,z,q,H).

7 In this highly simplified version of a unitary model we are not bothering to distinguish between the health or
illness of the different members of the household.

-------
B. Welfare Measurement with the Household Production Model

Recall that within the Lancaster-Maier framework, a consumer's utility depends not only on his
consumption of market commodities, denoted by the vector x, but also on some other items, q;
the utility function is thus u(x,q). While the consumer controls the level of x, subject to his
budget constraint, q represents some things that affect the person's welfare but which he does not
control. The Generalized Lancaster-Maler model provides both a theory of how q affects the
consumer's choice of market commodities (x) and a framework for welfare evaluation of
changes in q. The specific implication for purposes of valuing morbidity in children is that the
Hicksian compensating and equivalent variation are expressed in terms of the indirect utility
function. In the most general case, all elements (p',q',y') can change to a new level (p",q",y") and
indirect utility can change from v(p',q',y') to v(p",q",y"). Then the compensating variation for this
change is the quantity C such that

v(p",q",y"- C) = y(p',q',y'),
while the equivalent variation is the quantity E such that

v(p",q",y") = v(p',q',y'+ E).

If the change is an improvement in the sense that u"> u', the quantity C measure the consumer's
willingness to pay (WTP) for the securing the change, while E measures her willingness to
accept to forego it, and vice versa if the change entails a reduction in utility. We can use the
concepts of compensating and equivalent variation as measures of the economic value of a
change in environmental health risks for children.

Consider two important polar cases with regard to the impacts of the change in environmental
health risks: A) The change in environmental health risks could simply and automatically trigger
a reduction in the family's disposable income, but with no other concurrent effect, so that the
change is from (p,q,y') to (p,q,y"). (B) The change in environmental health risks could simply
trigger a change in q, with no other concurrent effect on p or y, so that the change is from (p,q',y)
to (p,q",y). In the first case, the direct effect of the environmental change is that the household
has less disposable income but everything else remains the same: the impact is equivalent to a
lump-sum reduction in income. The only impact is a purely monetary loss and the economic
value of this is the monetary loss itself. In the second case, by contrast, the direct effect of the
change is a loss of utility - wellbeing - for the household, C and E are different, and they
represent alternative ways of expressing this loss monetarily in terms of a loss of income that is
equivalent in the magnitude of its impact on the household's wellbeing.

The practical implication of the distinction between (A) and (B) is that, in the first case, one can
get along with information on the magnitude of the monetary loss without necessarily knowing
anything about the structure of household preferences, u(x,q), while, in the second case, one
cannot avoid the need to know about household preferences. In that case, the comparison
between revealed- and stated-preference approaches to welfare measurement will hinge on the
relative ease and reliability of the two approaches in providing an insight into the structure of
household preferences. We argue below that characterizing household preferences is essential to
defendable welfare measurement and present the limitations of using the household health

-------
production model, as typically applied, to measure welfare changes. In the next section we
present two areas in which these limitations arise:

•	Use of the health cost function to estimate the value of a change in health due to a
change in pollution.

•	Validity of a production function for health.

4. Conceptual Limitations of the Household Production Model for Welfare Measurement

A. Health Cost Function
First we describe a common use of the health cost function to estimate the value of a
change in health due to a change in pollution. For purposes of illustration, we assume that
there are no changes in the price of any market goods (px, pz) or income (Y), and
environmental quality changes from q0 to qi. This scenario is equivalent to Case B
described above, and here we describe the limitations to the standard approach to
estimating a welfare measure in this case. Suppose the change is for the worse, so that
U0 = v(px,pz,qo,Y) >UX= v(px,pz,ql,Y) . In this case the equivalent variation measure

(denoted E above) is the household's willingness to pay to avoid the change. The
marginal WTP to avoid is given by:

Moreover, by suitable manipulation of the first first-order conditions for the solution to
the household's maximization problem, one obtains

Recall that we presented the most general utility specification in which environmental
quality, q, entered directly into the household preference function. If instead we restricted
environmental quality to entering the health production function only, in which case the
preference function takes the form U(x,z,H), then the first term above would drop out.
Then under this special case the expression becomes:

dE

i E= 0

dq

vq(p*,pz,gJ)
vy(px,pz,qJ)

vq(Px,Pz,
-------
health cost function, c(pz, q, H). (3) Given the health cost function, calculate the marginal cost of
pollution, Cq(pz, q, H) and assess the value of the given change in pollution, dq, as the product

Value of health damage = cq(pz, q, H) dq .

An attractive feature of the expression above for researchers who use it is that it is only
requires information derived from the health production function and it avoids the need to
use information about the household's utility function. We believe there are two major
flaws in this approach. First, as discussed in the second section, the appropriate measure
for valuing welfare change reflects the difference between the margin and infra-marginal
unit. The expression above considers only the market-clearing price not the Marshallian
consumer's surplus. Our second concern with this approach is the assumption that neither
q nor z enters the utility function directly: The construction of the utility function as
U(x,H), which omits both the health inputs and environmental quality seems contrary to
observations about household preferences8.

In Section Five we present empirical evidence of three ways in which health
averting/mitigating behaviors are central to the concept of the preference function.

B. Validity of a Production Function for Health

While the notion of a health production function is illustrative in the discussion of household
choice, the extent to which it captures the complexity of trade-offs in the household is
questionable. The conceptual concerns regard the deviation between objective and subjective risk
assessments and the degree to which health is determined by individual choice.

In the literature on revealed preference valuation of market commodities based on their
attributes, researchers have often found that here is a divergence between the objective measures
of attributes and people's perceptions of them. Whether people see a beach as clean, an
automobile as safe or comfortable, a computer as high-tech looking, say, is a matter of

g

The household production literature often makes reference to an approach to welfare measurement derived from
work by Bockstael and McConnell (1983) based on the demand function for z's or I. Bockstael and McConnell do
permit q to enter the utility function directly, but not the z's. they show that that the Hicksian measure of WTP for a
change in q can be measured exactly from information about the demand function for health, H, or the demand
function for one or more of the z's that are input to the production of health. There are two qualifications that are
critical to the application of their result to the health production context. The first qualification is that their result is
about the area under the compensated demand function for H or for the z's, not the ordinary demand function. If
there are income effects in the demand for H or for the z's, the two demand functions are different and it is not valid
to use the area under the ordinary demand function as an approximation to the area under the compensated demand
function. These areas involve a price change from the current "price" (marginal cost) of H to the cut-off price at
which the demand for H would become zero, which is by no means a marginal change. Hanemann (1980) showed in
an analogous situation that the difference in areas can be quite substantial. The second qualification is that
preferences satisfy Maler's (1971, 1974) property of weak complementarity with respect to either H or the z's. In
this context, weak complementarity implies , that, if a person is in poor health (I = 0), she is indifferent to a change
in air quality (q). That seems unlikely to be true. Indeed, one can imagine circumstances under which, as long as the
person is still alive, a worsening in air quality becomes more serious to her when she is ill (I = 0) than when she is
healthy (I > 0). Bockstael and McConnell were thinking of a household production function for recreation, not
health, when they wrote their paper about weak complementarity and the fact is that their analysis seems ill-suited to
health applications.

-------
perception. How people see these attributes can be quite different from how an expert would
assess them. But, people's choices are likely to be based on their own perception and
understanding of the attributes, not on those of the experts. Therefore, researchers often find that,
to model choice behavior successfully, they need to elicit the decision makers' subjective
perceptions of attributes involved in the choices. The same can be true of household decision
making on health production. What may matter is what the household - the parents - see as
efficacious courses of action, not what the medical experts or the econometricians determine to
be efficacious. Another way of making the same point is to suggest that, while households'
decisions are based on ex ante expectations of the effectiveness of health producing actions, what
the econometrician measures when fitting a health production function is the ex post outcome. If
there were perfect knowledge or rational expectations, the ex ante expectation and the ex post
outcome would coincide. To the extent that these conditions are not met, the ex post household
production estimated by the econometrician might be misleading as a guide to understanding
household choice behavior. If this is so, it has the potential to bias not only the estimation of the
production function but also the estimation of household preferences.

The concept of a household production function implies that the household exercises a degree of
control over its member's health that is exaggerated and unrealistic from at least two
perspectives. First, postulating a household production function H= H(z,q) implies that, for given
q, the household can in principle attain any desired level of health, H, providing it has sufficient
financial resources to cover the cost of the required z's. If it is rich enough to purchase sufficient
z's it can make itself as healthy as it wants, regardless of what might befall it in terms of q. From
introspection, this notion is implausible. Second, the notion of an interior solution to the
household's health production decision is unrealistic. It is conventionally assumed that, in the
context of the household's production function, the z's are finely divisible, so that the household
arrives at an exact, interior solution to its optimization decision. Households often face a limited
and constrained set of options. These constraints may be imposed by the structure of the
healthcare sector and the nature of averting/mitigating behaviors.

Thus, while the notion of the household's production of its own health certainly has some basis
in reality, it can be pushed too far. People can look after their own health, but this does not mean
that they can achieve any desired health outcome; therefore some levels of H are not attainable,
regardless of the input of z's. The production function H(z,q) is likely to be bounded and it may
have some flat segments. Similarly, people do not have an unlimited array of options and
therefore they are more likely to be at corner solutions than interior solutions in their household
production decisions. As Bartik (1988) has noted: "Defensive options often may be limited; for
example, a household seeking to reduce the effects of toxic waste on its water supply might be
able to defend itself only by a water filter, bottled water, or moving away."

If these doubts are justified, this can have important implications for health valuation. The usual
first-order conditions do not hold and the simple approximations are apt to be unreliable. The
household's marginal WTP for improved health might be considerably larger than the marginal
cost, cq, but it may have no viable option for further action. Also, in this case non-marginal
valuation can become more complicated because of the need to specify a realistic, non-
monotonically increasing health production function or a limited choice set with a few discrete
alternatives.

-------
We suggest that these concerns lead to five practical limitations of the approach: 1) defining the
health outcome, especially for chronic, episodic conditions such as asthma, is nontrivial 2)
households' have varying degree of control for relevant health inputs 3) there is a probable
divergence between an objective physiological health production model and the household
subjective perception 5) there is likely to be endogeneity between choice of z and H. We present
evidence for each of these in Section Five.

5. Empirical Evidence of the Limitations of the Household Production Model for Welfare
Measurement

A. Health Cost Function
In Section 4. A, we suggest that simplifying the household model such that environmental
quality, q, and health inputs, z, appear only in the health production function and not the utility
function, does not capture the complexities of household tradeoffs. In this section we present
three examples of ways in which the averting and mitigating behaviors play important roles in
the household preference function.

1. Preferences for Health Inputs
Although the majority of the households reported that their children took medications to treat
asthma, 30% reported that they had concerns about those medications. A surprisingly common
concern was that taking asthma medication as prescribed could lead a child to become addicted
or dependent to asthma medication (49% agree or strongly agree, n=187). Other households
reported that they believed that having to take medications regularly was embarrassing to
children (23% agree or strongly agree, n=189). These concerns over medications affect how a
household perceives their benefit and omitting their consideration distorts the model of
household choice. In addition to these general concerns, households reported both that their child
experienced negative side-effects from asthma medications as well as a belief that these drugs
presented a risk to children's health. Table 6 reports the frequency of that household reported
their child experiencing side-effects for specific drugs as well as the frequency that household
reported that they believe children in general experience side-effects. Note that 29% of
households reported their child having side-effects from oral steroids, 27% reported side-effects
from rescue medications (albuterol) and 16% reported side-effects from control medications
(inhaled steroids). Medication is the largest category of direct costs for a typical asthmatic;
however, the market prices of these medications do not reflect the perceived costs to households.
Experiencing negative side-effects is likely to have substantial influence on household behavior
and this should be incorporated into the preference function to adequately measure and welfare
changes.

-------
Table 6: Frequency of Reported Side-Effects and Perceived Side-Effects

Medication

Personal

Others

Oral steroids

29

37

Albuterol

27

20

OTC allergy medications

18

13

OTC cold/flu medications

18

15

Inhaled steroids

16

16

Antiobiotics

16

9

OTC asthma medications

13

10

Intal

10

5

Tylenol

8

3

Vitamins

5

5

Note: Personal is the percentage of households reporting their child having side-effects as a result of that
medication. Others is the percentage of households reporting they believe children generally have side-
effects as a result of that medication

2. Adherence to Prescription Medication
Even if prescription medication did not enter the household's preference function directly,
there would still be difficulties in using the observed household choice. This difficulty stems
from the imperfect information and uncertainty surrounding actual prescription medication
usage. Much attention has been given in the public health literature on the discrepancies
between national guidelines on prescribing asthma medication and actual prescribing patterns.
Furthermore, even if the prescriptions do meet the asthma management guidelines, there is
ample evidence of non-compliance on the household level, and moving households towards
appropriate usage of medication is a major goal of many asthma interventions. The first hurdle
to adherence to prescription medication is ensuring that the prescription is filled: in our sample
25% of the households (n=210) report that they had at some time been given a prescription
that they were not able to buy because it was too expensive. It is unclear whether not filling
the prescription reflects a fully informed trade-off by the household or the result of a
subjective assessment that underestimates the benefits of medication usage. For example, use
of asthma medication to control the chronic inflammatory component of asthma is a case of
investment under uncertainty. In order to decrease inflammation, control medications need to
be taken consistently for 4-6 weeks which leads to a delay between taking the medication and
experiencing the benefits. After this fixed investment, the benefit is the reduction of the
probability of an asthma exacerbation. Families may be unwilling to make the investment in a
prescription medication it the benefits are uncertain and occur in some future period. A more
complex issue is the difficulties of communicating to households that the benefits will not be
realized until after the initial investment, and there is substantial evidence in the public health
literature that households confuse the delayed benefits of control medications with the more
immediate benefits of rescue medications.

Within our sample, 17% of those prescribed a control medication were not taking the
medication in the manner in which it is intended and of those prescribed a rescue medication,
43%) were not taking the medication as intended. These patterns suggest an over-reliance on
rescue medications, which has been reported in other populations (Boschert, Sadof, Brandt,

-------
2006). The frequency with which households incorrectly use the rescue medication not only
reflect what can be thought of as an inefficiency in health production in the current period, but
it also perpetuates the inefficiency into future periods. Approximately 15% of the households
reported that the prescription medication used by their child had either worsened their child's
asthma or left it unchanged. These assessments are likely to drive the choices over
medications in the next period, and if they are a result of non-adherence, then "non-optimal"
choices could be perpetuated.

While the causes for non-compliance are complex and not well understood, it does suggest
that using observed expenditures on medications may be confounded by factors other than
preferences over health states.

3. Non-market Behavioral Choices
In addition to using medication to treat both components of asthma, chronic inflammation and
acute bronchial constriction, standard asthma management guidelines include recommended
behavioral changes. Most of these behavioral changes are focused on reducing exposure to
possible asthma triggers (Boschert, Sadof, Brandt, 2006). Table 7 lists the changes undertaken on
a regular basis as well as large, one-time changes that were made to prevent asthma
exacerbations.

Table 7: Non-market Household Mitigating and Averting Behaviors

Routine Behaviors

Frequency

N

Check for smog alert 2 or more times a week in summer

83%

190

Change activity on high smog days

78%

189

Dusting frequently

59%

200

Vacuuming frequently

59%

200

Mold removal

51%

197

Close windows

46%

199

Restrict amount of child's time outside

46%

201

Restrict where child can play

44%

201

Restrict amount of child's activity

38%

200

Limit where pet can spend time

31%

154

Restrict child's diet

15%

196

One-Time Changes





Move household to avoid triggers

6%

199

Stopped smoking

15%

189

In focus groups parents reported that these averting behaviors required both substantial time
investments and a level of persistence, which in itself was a burden on household relations. A
second observation from the focus groups was that when households were asked what they do to
prevent asthma exacerbations, the first reported changes were of the type listed above, not
purchases nor taking medication. Our interpretation is that these behaviors are pertinent to how
household's perceive the impact of asthma, because these are behaviors that they must maintain
over time. The implication for data analysis is that because there are no markets for these regular
choices the market data that are used in estimating a cost function are incomplete. Second,

-------
households described the fatigue from constantly monitoring their child's health and modifying
the home to reduce triggers and this psycho-social burden is not reflected in any market data.

B. Validity of a production function for health.

In section 4.b we discussed conceptual limitations of the health production function. Here we
discuss findings from the survey that suggest how these conceptual limitations apply to the case
of valuing asthma morbidity.

1. Defining health status
Characterizing asthma severity is the subject of substantial epidemiological research, because of
the difficulty of capturing the natural variation in frequency and degree of symptoms. In our
focus groups we asked households to describe what they consider "typical", "good" and "bad"
asthma days. This process generated a set of impacts commonly used to describe asthma
morbidity and included many impacts in addition to the standard asthma symptoms or healthcare
utilization. The impacts considered important to households included symptoms (wheezing or
coughing, shortness of breath, black under eyes, increased mucous/phlegm or sputum, ribs
showing, easy of breathing), activity limitations (interrupted playtime, ability to walk stair, ride
bike or jump rope, ability to talk and sing) and social impacts (avoiding places with triggers,
restricting time outdoors). Social impacts were used 79% of the time to describe asthma
morbidity, activity limitations were used 96% of the time, and physical symptoms were used
98%) of the time. This finding is consistent with the literature that suggests that households tend
to describe the severity of their medical condition in terms of activity limitations or impact on
quality of life, where as medical professional tend to categorize severity based on frequency of
physical symptoms. For the purposes of estimating welfare effects of morbidity, it is the
household perspective that matters and drives behavior. Furthermore these impacts affect
households much more regularly than do the extreme events of emergency room visits or
hospitalizations. Table 8: Asthma Related Morbidity, presents the percentage of household who
report that their child experiences limited activity levels, a social impact or a physical symptom
on each type of asthma day.

Table 8: Asthma

delated Morbidity



Percent experiencing



Limited activity level

Social impact

Physical Symptom

Good day

13

29

28

Normal day

47

43

63

Bad day

93

73

97

These data suggest three modifications of the standard household health models. First, welfare
estimates that are based on avoiding unplanned medical visits, emergency room visits or
hospitalizations miss the larger more routine impact of asthma morbidity on quality of life. The
vector that describes health status should include quality of life impacts that are disease and age
specific and should not be limited to physical symptoms. Last, because the outcome of interest is
a vector of related impacts, data analysis should utilize a multivariate approach with both
symptoms and psycho-social outcomes.

-------
2. Constrained Choice Sets
Reducing exposure to asthma triggers is an important part of asthma management and is often
described as the most important behavioral change a household can take. In our sample,
households were able to identify triggers for their child. The ten most commonly reported
triggers for asthma exacerbation were: an existing cold/flu (86%), air pollution (74%), pollen
(66%), exercising outdoors (63%), tobacco smoke (53%), dust (53%), outdoor smoke (52%),
strong winds (50%), mold (40%), animal dander (33%). Given the parents' perception of asthma
triggers, the household model would then posit that households would make the relevant choices
to reduce (avert) these risks; however, reducing exposure is not feasible for an important class of
asthma triggers. The total height of the bars in Figure 1 shows the frequency households
reported the exposure to be an asthma trigger for their child. The shaded are of the bars indicate
the degree to which households that reported the item as a trigger felt they had control over their
child's exposure. While households reported the ability to limit or reduce exposure to many of
the commonly cited asthma triggers (exercising outdoors, tobacco smoke, dust, mold, animal
dander, cleaning solutions, food allergies, and roaches), this was not uniformly the case. Of the
top ten potential triggers, there were five triggers that more than 20% of the household reported
that they had no control over their child's exposure (pollution 41%, strong winds 31%, cold/flu
28%, pollen 25%, and outdoor smoke 20%).

Figure 1: Perception of control of asthma triggers

Degree of control

~	total

~	some
¦ none

Trigger

One assumption of the health production function is that the household is able to purchase any
number of units of inputs if they are willing to pay the price. The reality was different for 11% of
our sample (n=202) who reported difficulties in making appointments with their medical doctor

-------
when needed, and the 7% (n=202) who report that their medical doctor is not helpful when their
child's asthma worsens.

These findings suggest that one of the cornerstones of asthma management, averting risk through
reducing exposure to triggers, does not readily translate into a production function framework,
because the ability to control these exposures is limited. In addition to limited control of asthma
triggers, households also reported limitations in the quantity and quality of available medical
inputs. The results of our survey corroborate our concern that there may be no interior solutions
to the household's maximization problem as commonly formulated.

3. Divergence Between Objective and Subjective Health Production Function
A fundamental assumption of the household production model is that households perceive a
health production function and make health choices accordingly. We found substantial
differences in how households conceptualize the process that determines asthma status. Of our
sample, 16% (n=191) disagreed that asthma can be managed so that a child does not have
symptoms; 13% (n=189) agreed with the statement that asthma episodes can cause problems but
are not really harmful or dangerous; 58% (n=189) agreed with the statement that asthma episode
usually occur without warning; 16% (n=202) report being uncertain about what to do when a
child begins to have asthma symptoms. These statistics suggest that households often have
incomplete and imperfect information with which to make choices, in other words there is
substantial divergence between an objective physiological model of health production and the
household's perceptions. They also suggest that rather than households conceiving of a
production function, they consider their child's asthma status as exogenous and try to optimize
welfare given the asthma status.

We present two avenues in which these divergences may arise. First, prior to making a health
related expenditure, households do not have an assessment of how helpful an input will be. To
explore this, we asked households both the amount spent on fixed health inputs and their ex post
evaluation of the effectiveness of each investment. Table 9 lists the households' assessments of
purchases, and three patterns should be noted. First, although each of these investments are
commonly suggested averting/mitigating behaviors none were unanimously helpful (ranging
from 44% to 98% reporting that the investment was helpful). Second, the investment that was
most often reported to be helpful was the nebulizer, which provides relief during a current
asthma exacerbation followed by a spacer which helps in delivery of medication, while those
investments that were less likely to be reported as helpful were those that reduce triggers thus
reducing the probability of an exacerbation in the future (e.g. air filters, removing pests, and
HEPA vacuums). Third, households perceived a value to the peak flow meter, which provides
information useful in asthma management but which in itself does not reduce triggers or alleviate
symptoms. These patterns suggest that households' ex ante expectation and ex post outcome do
not coincide for all investments, and the time frame for delivery of benefits may play an
important role in household's subjective assessment.

-------
Table 9: Household Investments and Assessments

Investment

Air filter
Mattress cover
Pillow cover
Humidifier
HEP A Vacuum
Removed carpet
Removed pests
Remove mold
Nebulizer
Peak flow meter
Spacer	

% purchased

% reported helpful

N

200

200

201

202
202

201

202
201
199

199

200

30
28
27

31

38
25
25

39
31
35
54

69

73
82

74

70
78
44

72
98

73
93

These empirical observations substantiate our concern that a production function for health may
impose a relationship between health and choice that is artificially strict and complete.

4. Endogeneity Between Health Inputs and Health Status
One source of endogeneity between the choice of health inputs (z) and asthma morbidity (H)
arises from families "benchmarking" their concept of what is normal or attainable respiratory
health. For example, one observation in asthma case management programs is that families either
do not perceive their respiratory difficulties as asthma symptoms or they come to accept the
asthma symptoms as unavoidable. Even within the FACES population, we found that the concept
of asthma control deviated from the standard medical concept of asthma control (normal or near
normal respiratory function). In our survey, 26% of the households who described their child's
asthma as well to completely controlled, actually would be classified as moderate to severe based
on the frequency of their daytime symptoms during the winter 2003-2004. As was shown in
Table 8: Asthma Related Morbidity, even on a "normal day" from the perspective of the
household, children commonly experienced limitations in their level of physical activity (47%),
constraints on social interactions (43%) and symptoms (63%). As households that have children
who routinely experience asthma morbidity come to expect these impacts as normal or "best
possible" level of asthma control, their health investments will reflect this perceived limitation.
This benchmarking could be thought of as a creating categories of households with differing
perceptions of the frontier for H(z,q), and the perceived frontier would be correlated with the
unobservable characteristics that affect the baseline asthma severity.

In the case of household health, the typical instrumental variables approach to the problem of
endogeneity is confounded by the complexity of fully specifying the production model. As
shown by Griliches and Mairesse (1999), instrumental variables estimates of the production
relationship will not produce valid estimates of the coefficients on the health inputs if there are
omitted variables from the health production function that are correlated with the elements in the
vector of health inputs, z. As shown in the previous sections, households vary in their perceived
risks and benefits of health inputs, and unless a model can adequately capture the factors that
determine this variation instrumental variables will be an incomplete solution.

-------
7. Stated Preference Approach

A major advantage of the stated preference approach relative to the revealed preference approach
is that it allows the researcher to create a trade-off with which to confront survey respondents,
thus it makes it possible for the researcher to control the specification of the household
production function. Instead of having to estimate an unknown production function, commingled
with unknown household preferences, and complicated by the household's unknown subjective
perceptions of what it can do to protect or improve its health, the researcher may be able to
create his own specification of the trade-off, thereby limiting the unknowns to be estimated from
the data to the respondent's preferences.

A second survey was conducted in-person at the FACES office over October 2005-February
2006. This second survey included questions on frequency of asthma symptoms that correspond
to the updated Gina asthma severity classification (Luppi, 2004), severity of asthma symptoms,
asthma triggers, asthma specific health beliefs, rating and ranking of the impact of asthma on
quality of life, causes of household stress and a contingent valuation scenario. A total of 130
FACES households completed the second survey9.

After the interviewer completed the health status questions, (s)he presented the contingent
valuation scenario. The participant was given a brochure that described a hypothetical asthma
monitor that could be worn like a watch. The description of how the watch worked included a
diagram of a normal airway and an asthmatic airway with both constriction and inflammation.

Figure 2: Brochure Diagrams

Normal Airway	Asthmatic Airway

Asthma Monitor

9 We began conducting the contingent valuation survey in Oakland, California in March 2006. Our target is an
additional 200 surveys by the end of 2006.

-------
The brochure explained that the watch monitors the level of oxygen in the child's blood and
provided an indication when it varied. A green face on the watch indicated that oxygen was
optimal whereas a yellow indicated caution and a red face indicated an emergency. By
monitoring the child's asthma, it was suggested, action could be taken to stop the asthma from
progressing to the point that physical symptoms developed. The hypothetical monitor, the
BreatheRight watch, was said to have been shown to cut the number of days with asthma
symptoms by one-half. We used a one and a half bounded dichotomous choice format to elicit
bids for the hypothetical scenario. Initial bids were based on the distribution of responses from a
pilot of twenty-two non-FACES households in the Fresno area conducted in August of 200510.
Stating bids and subsequent bids were updated following Cooper, Hanemann and Signorello
(2002).

We crafted the hypothetical scenario to have six characteristics relevant to the findings of the
first survey that were discussed in Section 5:

1.	The scenario reduced morbidity without relying on medication, and thus would not be
confounded by preferences for medication.

2.	The device did not require behavioral changes to be effective, thereby reducing the
issue of non-adherence.

3.	The tool reduced both the physical symptoms and the stress of monitoring the child's
asthma, which addresses the larger issue of how asthma morbidity affects quality of life.

4.	The device helped families communicate quantitative information about their child's
asthma, improving access to health care when needed.

5.	The instrument provided objective information on the child's health status and assisted
families in assessing health risks and effectiveness of averting and mitigating behavior.

Results from the contingent valuation survey will be reported in future research.

Conclusions

In this paper we discuss two approaches to estimating the willingness to pay (WTP) for reduced
asthma morbidity, contingent valuation and health production function. In the health production
model the health outcome is a function of exposure to asthma triggers, mitigating and averting
behavior and household's perceived risks. We find that variation in expenditures is explained by
attitudes towards asthma specific health investments including concerns of associated risks and
perceived effectiveness. The survey data indicate that households select from a small number of
discrete health investments and that most risk reducing behavior are daily behavioral
modifications with no relevant market prices.

We argue that the discrete nature of health investments and socio-cultural patterns of health care
utilization make the revealed preference approach inadequate for the case of asthma. As an
alternative we present a contingent valuation scenario that was specifically developed to
minimize systematic variation in preferences for characteristics related to the scenario rather than
the reduction in asthma morbidity. For this purpose, guided by extensive testing in focus groups,
we selected a scenario based on a hypothetical asthma monitor that provides to the wearer an
indicator of current asthma status.

10 Participants for the pilot were recruited through a newspaper ad in the local paper (Fresno Bee) and a recruitment
table at the American Lung Association's annual walkathon in Fresno.

-------
References

To be added

-------
Individual Preferences and
Household Choices: The Potential

Role of Dependency Relationships

Mary F. Evans, Christine Poulos, and V. Kerry Smith

-------
Background

~	STAR grant: applying weak substitution to value air
quality improvements that improve health

~	3 phases of research for two study populations:



Children

Older adults

1. Develop theoretical model



2. Verify environmental health
impacts using secondary data

ongoing



3. Survey parent-child and older adult-
caregiver pairs to value air quality
improvements



Focus groups
in 2005

-------
Motivation

~	Phase 3 requires understanding of dependency
relationships (young child/adult parent, older
adult/caregiver).

~	Central to characterizing and interpreting behavior,
and to designing stated preference surveys:

¦ Intra-household allocation process

-------
Purpose

~	Describe a conceptual framework and testable
theoretical model of dependency relationships that
are expected to influence the value of environmental
quality improvements that affect children and older
adults

~	Describe internet survey activities that:

¦	Informs conceptual structure

¦	Complement/substitute for focus group activities and/or
cognitive interviews

-------
Theoretical background

~	Collective model (Chiappori and co-authors) permits
the recovery of individual preferences from
household behavior. Offers strategy for analyzing
household structure and dependency relationships.

~	What information is used for identification?

Chiappori: Focus on the intensive margin.

Our proposal: Focus on the extensive margin.

~	Current tests of the collective model rely on
observing the full system of demands.

~	Is there an alternative test that does not rely on this
information?

-------
Review of Chiappori collective household
framework

~	Background

Household is best viewed as a collection of individuals
with different preferences; analysts usually observe
household not individual demands for goods and services.

Household "behavior" is unlikely to be described
adequately by the unitary model - i. e. treating choice as if
it was motivated by a single agent's decisions.

~	Assumptions

Each household member has own preferences that are
known by other household members

Collective decisions of household are Pareto efficient

-------
Focus of Chiappori and co-authors' research

Browning and Chiappori [1998], Chiappori and Ekeland [2006;
forthcoming]	

~ Key Issue: What does the efficiency assumption
imply for household demands and specifically for
the matrix of Hicksian price effects?

¦ Question directs attention to the	margin of

choice

-------
Basic structure of argument

Context and definitions

~ Assume two members (I and II) of household

¦	P = prices of market goods (T x 1 vector)

¦	X = quantity of market goods consumed by household,

I i II v

x +x — X

¦	Z = private good that is public consumption to members
of household

¦	U'(x',xJ,z) = /th individual's utility function (/' ^ /)

¦	y = household income,	y

-------
Key efficiency assumption

~ There exists a differentiable, homogeneous of degree
zero, function \x(p, y) such that for any (p, y) the vectors
(x1, x", Z) are solutions to the following optimization
problem:

Max ju{p, yp1 (x7, x11 , Zj+ (l - (x7, ,

~ Yields household (Marshallian analog) demand

~ Expenditure minimization problem yields (Hicksian
analog) demand functions

functions: fs(p,y,M)

-------
Key generalization

~ Duality implies, holding p. constant

~

dhr

v. + or,

dp.	dp

f,

Allowing p. to vary with prices and income,

f	f)

J t	/-<. j t J

opt opt dy djU opt dy

Element of matrix of
Hicksian price effects
(pseudo-Slutsky matrix)

Element of
conventional
Slutsky matrix
(symmetric)

Leads to matrix that is at
most rank one (in two-
person household) =
basis of test

-------
Illustrative example

~	Provides intuition for alternative test of the
collective model

~	Informs development of choice questions

~	Assumptions

¦	Up to a two-person household

¦	Linear indirect utility function

~ Form varies according to structure of household

¦	Consider change in indirect utility from improvement in
air quality, q,that reduces the amount of care giving time
required for self or another individual (small child,
teenager, older adult)

-------
Individual only (no altruism, no income sharing)

~	Indirect utility function with initial air quality

Vq = (Xq + CC^W + C^2.y ^3^0 —

¦	with wthe wage rate, ynon-wage income, (fixed) care
giving time to self, <7o initial level of air quality.

~	Indirect utility function with improved air quality

= (Xq + a^w + c^2 iy ~ ^3^1

¦	with Tthe cost of the program to improve air quality (and
reduce care giving time).

~	Change in indirect utility

AV = AL+a3(ql-q0)-a2T

-------
Altruism (no income sharing)

~	Indirect utility function with initial air quality

V0 =a0+alw + a2y + a4h{q0)-AL

¦ with hdescribing the health of the dependent as a function
of air quality, Lcare giving time to dependent.

~	Indirect utility function with improved air quality

V{ = a0 + «1w + «2(y-7T)+a4/z(g1)

~	Change in indirect utility

AV = XL +aA\h{g^)-h{g^-a2T

-------
Income sharing (no altruism)

~	Indirect utility function with initial air quality

F0 =	a0 +bxy + w + b3q0 + b4L]-

¦ where the term in brackets represents the individual's
share of household income.

~	Indirect utility function with improved air quality

Vl = [b,+bl{y-T)+b1w + b,q{]

~	Change in indirect utility

AV = (A - a2bA )L + a2b3 (qt - q(j)- a2b{T

-------
Altruism and income sharing

~	Indirect utility function with initial air quality

V0 = a0 + bxy + w + b3q0	+ b4L\+a4h(q

~	Indirect utility function with improved air quality

Fj =	a0 [b0 +b1(y~T)+b2w + b3q1]+a4h(ql)

~	Change in indirect utility

AV = (b - a2bA)L + a2b3 -q\h(c]\) - h(q())] -

-------
Matrix of household types

Label

Coefficient on

A q

Coefficient
on L

Coefficient
on T

Number in
household

Individual
only

a3

A

-a2

1

Altruism only

?
•

A

-a2

1

Income
sharing only

a2b3

A — cc

— a2bx

2

Altruism and

income

sharing

?
•

A — c

— oi2b

2

-------
Choice questions

~	Objective: understand how people make choices
(not to obtain WTP estimates for policy use)

~	Strategy:

¦	Two double-bounded questions per respondent

~	One focusing on respondent

~	Another focusing on another individual: young child (2-5),
teenager (13-17), or older adult (>62)

¦	Design attributes constant within questions for
respondent

Interval censored model to estimate valuation function

-------
Choice questions

~	Respondent asked to suppose that subject (self or other
individual) has asthma

~	General choice question: "Would you pay $T for a
program that will improve air quality and reduce the
amount of time you allocate to care giving for [yourself /
young child / teenager / older adult] from L to zero?

Double-bounded question with respect to T

L (2 levels), T (4 levels) randomly assigned

Sequence of responses to each double-bounded question
sort individuals into four bins (yes/yes, yes/no, no/yes,
no/no)

-------
Survey details

~	Add survey questions to weekly internet panel

¦	1000-2000 respondents (18+ years); $600-$ 1000/question

¦	Socioeconomic data provided: education, income,
household size and composition

¦	Data available in days

~	Questions

¦	Choice questions

¦	Additional characteristics (presence and age of children in
HH, presence and age of older adults in HH, asthma in
HH members)

-------
Conclusion

~	Dependency relationships and household structure
are potentially important in estimating the benefits to
improved environmental quality.

~	Proposed tests have focused on the	margin.
We propose an alternative that focuses on the

extensive margin that does not require observation of
full system of demands.

-------
Air Pollution and Asthma:

Preliminary Results from a Daily Time-Series Study of San Francisco

By

Charles W. Griffiths and Nathalie B. Simon
U.S. Environmental Protection Agency

Paper prepared for NCEE and NCER Sponsored Workshop on
"Morbidity and Mortality: How Do We Value the Risk of Illness and Death?"

April 10-12, 2006

-------
Air Pollution and Asthma:

Preliminary Results from a Daily Time-Series Study of San Francisco

By Charles W. Griffiths and Nathalie Simon1

Asthma is a chronic lung disease characterized by intermittent, recurring episodes of
wheezing, breathlessness, tightness of the chest, and coughing. These episodes are caused by
inflammation of the airways that carry air into and out of the lungs. Asthma is considered to be a
growing problem in the United States, especially among children. The prevalence of asthma
increased 46 percent between 1982 and 1993 in the United States. While increases in prevalence
have been documented in all age, race, and gender groups, the increase has been most significant
among children — individuals under the age of 18 — with a staggering 80 percent since 1982.

While the exact causes of the illness remain unknown, asthma attacks can be triggered by
a number of factors including exposure to allergens (e.g., dust mites, pollen, mold, pet dander,
and cockroach waste), strong fumes, respiratory infections, exercise, dry or cold air, as well as
air pollution (including ozone and particulate matter). Despite recent efforts to reduce ambient
levels of air pollution, approximately 46 million people lived in counties that did not meet the air
quality standards for at least one of the six criteria pollutants in 1996. The combination of poor
air quality with other triggers is often most extreme in urban centers where a disproportionate
number of minority and low income households reside.

A relatively large number of studies exist that focus on the temporal relationship between

1 Disclaimer: The views expressed in this paper are those of the author(s) and do not necessarily represent
those of the U.S. Environmental Protection Agency. In addition, although the research described
in this paper may have been funded entirely or in part by the U.S. Environmental Protection Agency, it
has not been subjected to the Agency's required peer and policy review. No official Agency endorsement
should be inferred.

Acknowledgements: The authors wish to thank Jessica Sloan, research assistant to this project, for all of
her efforts.

2

-------
air pollution and asthma attacks resulting in Emergency Room visits or Hospital Admissions.
However, these studies by design focus on severe outcomes and miss milder ones - asthma
attacks that are alleviated through medication use and do not require immediate medical
attention. Those studies that do examine milder forms of asthma symptoms (e.g., respiratory
symptoms or increased use of asthma medication) generally take the form of cohort or panel
studies and have tended to focus on children. This paper presents the preliminary results of a
time-series analysis of the effects of acute exposure to ambient air pollution on the incidence of
asthma attacks, as measured by prescription counts for short term "quick relief medications.
Using prescription data from San Francisco, California, we estimate the relationship between
exposure to ozone and PM10 and asthma symptoms.

Background

The relationship between short-term increases in ambient levels of air pollution and
asthma outcomes has been documented in a number of venues using two types of studies: daily
time series studies and cohort or panel studies. Daily time-series studies have been used to
model the relationship between air pollution and a number of health outcomes including daily
mortality and other relatively severe respiratory outcomes such as hospital admissions,
emergency room visits and doctor visits. A relatively large segment of these studies have
focused on asthma. In a study by Walters et al. (1993), for instance, daily levels of S02 and
black smoke were found to have a positive association with hospital admissions for asthma in
Birmingham, UK. A similar result was found in Birmingham, Alabama in a study focused on
hospital admissions due to pneumonia and Chronic Obstructive Pulmonary Disease (of which
asthma is a component) among elderly inhabitants (Schwartz 1994). Also found was a positive

3

-------
association between air pollution levels and doctor visits for asthma in London (Hajat et al.
1999). In Barcelona, Spain, a positive association between emergency room visits for Chronic
Obstructive Pulmonary Disease and air pollution levels was found (Sunyer et al. 1993). While
these studies are indicative of the detrimental effects of short-term increases in air pollution on
rather severe asthma outcomes, they give no indication of the effects of air pollution exposure on
asthma outcomes that are milder in nature.

Panel or diary studies can provide some indication of the effects of air pollution on less
severe health outcomes. They model symptoms experienced by panel members as a function of
air pollution levels. While most cohort studies are focused on children, some studies have found
positive and significant effects of air pollution exposure on exacerbation of asthma symptoms at
other ages. Neukirch et al. (1998) found measurable short-term effects of low-level air pollution
in Paris France on nonsmoking asthmatic adults diagnosed with mild or moderate asthma.
Similarly, Newhouse et al. (2004) found that ozone concentrations on the previous day were
associated with a number of symptoms including wheezing, headache, and fatigue in their panel
of 24 individuals aged 9-64 with physician diagnosed asthma. Ostro et al. (1991) also found a
strong association between daily air pollution levels (specifically airborne acid aerosols,
particulates, and sulfates) and increased asthma symptoms among a panel of asthmatics in
Denver, Colorado. Similar results have been reported in the Utah Valley (Pope et al. 1991),
Glendora California (Krupnick et al. 1990), and the Netherlands (Hiltermann et al. 1998) among
other places.

While diary studies are useful in isolating the effects of short-term increases in pollution
on milder outcomes, these studies face several difficulties. Among these difficulties, as noted by
Schwartz et al. (1991), is the fact that daily symptom rates are often highly correlated from one

4

-------
day to the next and the heterogeneity among subjects causes dependencies in the data. Some
study results are also limited by the availability of particulate pollution measures while others are
limited by panel size or length of study period.

In contrast to the studies described above, our study examines the effect of short term or
acute exposures to air pollution on asthma symptoms as measured by the purchase of quick relief
asthma medications in San Francisco, California. Zeghnoun et al. (1999) explore a similar
relationship in Le Havre, France and find statistically significant effects of black smoke, N02
and S02 on respiratory drug sales for mucolytic and anti-cough medications for children and
adults. Our study in contrast is focused on quick relief asthma medications. We hypothesize
that acute exposure to air pollution may make an individual more susceptible to asthma attacks,
causing an increase in the use of quick relief medications.

Methodology

This study looks at the effects of differences in short term air pollution exposures on the
occurrence of asthma attacks, where asthma attacks are proxied by the number of prescriptions
for quick relief asthma medication filled. The total count of prescriptions for quick relief asthma
medication is explained using measures of asthma triggers and other cofactors. The study
utilizes a dataset of asthma drug prescriptions for a large percentage of the pharmacies in the
state of California and GIS layers of spatial factors.

In this study, our "health" outcome (filling asthma prescriptions) is not a "direct" effect of
air pollution exposure, but rather a secondary effect. That is, the true sequence of events goes as
follows: short-term exposure to air pollution makes an individual more susceptible to asthma
triggers leading to an exacerbation of asthma symptoms which in turn causes an increase in

5

-------
asthma medication use. The increase in asthma medication use eventually (perhaps with a lag)
leads to the filling of a prescription. The urgency with which a prescription needs to be filled
will vary across individuals and their initial stock of asthma medication, making short term
effects more difficult to observe.

An individual suffering from asthma will use his inhaler with some probability based
upon the amount of pollution present, current weather conditions, and seasonal factors.

Pr(Inhaleruse)t = f(Pollutiont, Weathert, SeasonalFactorst) (1)

We assume that each day the individual makes an independent decision, where the choice is
whether or not to use the inhaler based on the contemporaneous pollution, weather, and seasonal
factors. If we were able to witness the individual's use of his inhaler, then we could model this
behavior using daily cofactors. Minor modification would be possible if asthma attacks were
serially correlated, or if the use of an inhaler one day was related to the conditions on the
previous day as well as the current day.

In our case, we do not witness the individual's use of the inhaler, only the purchase of a
new inhaler when the old one is empty (or close to empty). Therefore, the observable event, the
number of prescriptions, is a function not only of the contemporaneous factors, but the total
amount of pollution and weather conditions over the recent past, as well as the seasonal factors.

(Numberof Prescriptions^ = f(Polluticnt, Pollution^,..Pollutiont m, Weathert,

Weathert 4Weathei;_m, Seasonal Factorst)	(2)

The number of days, m, which defines the recent past needs to be long enough to capture the
signal of inhaler use, but should not be so long as to add additional noise to the model. Seasonal
variation does not need to be modeled as an aggregation over time since it can be captured using
other methods - in this paper, through the use of continuous trigonometric cycles of varying

6

-------
length. If the number of prescriptions follows a Poisson distribution, then we can model the
mean incidence rate, that is, the number of prescriptions on any given day, as

rt = exp(a + /3[Pollutiont + Pollutiont_x +... + Pollutiont_m ]

+ v[Wealhei- + yWeathei] , +... + ymWeathert_m\	(3)

+ 8 Seasonal Factorst

Since each day should be given the same weight, P and y can be estimated for the sum of
pollution and weather cofactors over the recent past.

Since we are modeling a single CSMA over time, we assume that the exposure rate does
not change from one day to the next and, therefore, do not include it explicitly in the estimation.

Modeling the number of prescriptions in this fashion means that P is then a semi-
elasticity of the impact of a one-unit change in pollution. In other words,

fpl-P	<4)

dP /',

where a one unit change in pollution produces a P percent increase in inhaler prescriptions.

Data

The number of prescriptions for quick acting asthma medication was obtained from
NDChealth (hereafter, NDC), a Phoenix-based company that maintains prescription-related data
for marketing research. NDC maintains two datasets of use for this study, a "retail pharmacy"
database and a "patient" database. The pharmacy database contains dispensing records from
approximately 36,000 pharmacies nationwide, and captures approximately 70% of the volume of
traditional pharmacy-dispensed prescriptions. Hospital, military and mail order pharmacies and
prescriptions dispensed to institutionalized patients are not included in this database, which may

7

-------
pose a problem in the future as mail order prescriptions grow, but is probably not important here.
The patient database is a subset of approximately 14,000 of the pharmacies in the pharmacy
database. The patient database is a more complete database, in many cases including the patients
age and gender, along with a unique patient identifier so that the history of a patient may be
followed. Not included in the database, and unknown to NDC, is any information that could
personally identify a patient (such as a name, address or phone number) and NDC has been very
careful not to release any individual patient data, even with the anonymous identifier.

Prescription data were provided for San Francisco by NDC, segregated by the level of
asthma severity of the patient. Asthma severity is classified as mild intermittent, mild persistent,
moderate persistent, and severe, based upon the number and combination of prescriptions that
the patient fills for both quick-relief and maintenance asthma medicine over the 12 month
calendar year (NIH, 1997). Generally, asthma medications fall into one of two categories: (1)
short-term treatments intended to provide quick relief in the event of an asthma attack and (2)
long-term maintenance therapies intended to prevent asthma attacks. Mild asthmatics are those
patients prescribed a quick-relief medication only. Patients with mild persistent asthma not only
are prescribed a quick-relief medication but are also prescribed a single controller or
maintenance therapy. Moderate asthmatics are prescribed two controllers operating by different
modes of action in addition to the quick relief medications, while severe asthmatics are
prescribed three controllers with different modes of action. Should an individual's asthma
severity level shift over the 12 month period, the individual is assigned to the most severe of the
categories for which he/she qualifies. A list of the quick acting and controlling asthma
medication is listed in Table 1.

8

-------
Table 1: Asthma Medication

Symptomatic Therapy (Quick Relief)

Albuterol

Bitolterol

Isoetharine

Metaproteronol

Pirbuterol

Terbutaline

Controller Therapy (Long-term preventative)

Inhaled Corticosteroids
Beclomethasone
Budesonide
Flunisolide
Fluticasone
Triamcinolone
Leukotriene Antagonists
Motelukast
Zafirlukast
Zileutin
Long Acting Beta Agonists

Salmeterol
Xanthine Derivatives
Aminophylline
Dyphylline
Oxtriphylline
Theophylline
Mast Cell Stabilizers
Cromolym
Nedocromil

medication in a five digit zip code for each quarter from 1998 to 2001 were used. Data are
given by dispense quarter and the zip code of the dispensing pharmacy. These data are further
disaggregated by asthma severity.

The prescription data used in this analysis are limited in the following way. They only
include counts of prescriptions for quick relief asthma medication from those pharmacies that
"consistently" report this information. "Consistent" reporting is defined by NDC as pharmacies

9

-------
for which fewer than 11 days of data are missing in any 30 day period.

The air pollution data are publicly available from the California Air Resource Board.
Daily observations on the levels of PM10, S02, NOx, and ozone are available for 71 monitors in
San Francisco.

The weather data come from the National Climatic Data Center. Daily observations for
the average, minimum, and maximum temperature, as well as relative humidity, and the
minimum and maximum relative humidity were obtained for 97 active weather stations in San
Francisco.

The summary statistics for the data used in this analysis are listed in Table 2.

Table 2: Summary Statistics

Variable

Number of







(units)

Observations

Mean

Min

Max

Total Daily Prescriptions (number)

1428

852.0

28

1714

Mild Intermittent

1428

389.5

15

950

Mild Persistent

1428

276.4

9

579

Moderate

1428

128.8

1

260

Severe

1428

57.3

3

122

Daily Minimum Temperature (°F)

1672

39.69

15

67

Daily Average Relative Humidity (%)

1672

68.09

37

98

Daily Average PM10 (|ig/m3)

1499

38.01

2.60

227.72

Max of Ozone lHr (ppm)

1672

0.08

0.04

0.18

Results

We estimate equation 3 above using a standard negative binomial regression - the more
generalized form of the Poisson model. Total counts of prescriptions filled over the course of a

10

-------
week are regressed against minimum temperature, relative humidity, pollution measures, time
trends and trigonometric terms designed to capture cyclical trends ranging from 1 year to 2.4
months in length. Weekly counts are used to minimize day of week effects and effects of
pharmacy closures due to holidays. Temperature and pollution measures are summed over 180
days preceding the weekly counts. As described above, summing these factors in this way
allows us to more accurately capture the effects of pollution and weather on the dispensing of
quick relief asthma medications. A 180-day window was selected upon inspection of the plot of
the residuals of prescription counts (with seasonal variation and time trends removed) against
pollution measures.2

The model was constructed in a step-wise manner in which we first added our season and
time controls, using the Akaike Information Criterion to inform the choice of model before
incorporating other factors. Once the seasonal factors were selected, we incorporated
meteorological metrics including temperature and relative humidity. Minimum, maximum and
daily average measures were tested against one another. Minimum daily temperature and average
daily relative humidity provided the best fit according to AIC.

With weather factors controlled for, we added ozone and PM10 measures to the model.
We experimented with 8-hour and 1-hour ozone measures, and found that daily maximum
observations of 1-hour ozone readings provided the best fit. Daily average PM10 was similarly
selected for inclusion.

Once the model construction was completed for counts of total prescriptions, we applied
the same model construct to counts of prescriptions by severity level. Results for all five
regressions are reported in Table 3.

2 Troughs and peaks in the residuals were matched with those observed in the pollution data.

11

-------
Table 3: Regression Results

Variable Name

Total Prescriptions

Mild

Mild Persistent

Moderate

Severe

Coeff.

Std.
Error

Coeff.

Std.
Error

Coeff.

Std.
Error

Coeff.

Std.
Error

Coeff.

Std.
Error

trend

-0.0001

0.0001

-0.0001

0.0001

-0.0001*

0.0001

0.0000

0.0001

-0.0002*

0.0001

Year 1999

0.1238*"

0.0315

0.1206**

0.0386

0.0978***

0.0302

0.1299***

0.0269

0.2836***

0.0290

Year 2000

0.2051***

0.0589

0.2285***

0.0714

0.1473**

0.0568

0.1559**

0.0506

0.4056***

0.0545

Year 2001

0.0062

0.0884

0.0097

0.1075

-0.0867

0.0851

0.0354

0.0758

0.3489***

0.0816

sinlyr

-0.4187***

0.0302

-0.5011***

0.0372

-0.3914***

0.0287

-0.3140***

0.0255

-0.2129***

0.0274

coslyr

0.2556***

0.0098

0.3547***

0.0120

0.1974***

0.0093

0.1440***

0.0082

0.1044***

0.0087

sin6mo

-0.0707***

0.0058

-0.0924***

0.0070

-0.0774***

0.0055

-0.0221***

0.0049

-0.0101*

0.0053

cos6mo

0.0102**

0.0034

0.0179***

0.0042

-0.0034

0.0032

0.0075**

0.0029

0.0111***

0.0031

sin4mo

0.0317***

0.0046

0.0477***

0.0056

0.0207***

0.0044

0.0167***

0.0039

0.0177***

0.0042

cos4mo

0.0286***

0.0034

0.0447***

0.0042

0.0200***

0.0033

0.0042

0.0029

0.0091**

0.0031

sin3mo

-0.0162***

0.0040

-0.0194***

0.0049

-0.0177***

0.0039

-0.0080*

0.0034

-0.0112**

0.0037

cos3mo

-0.0102**

0.0034

-0.0075*

0.0042

-0.0158***

0.0033

-0.0095**

0.0029

-0.0061**

0.0031

sin2.4mo

-0.0191***

0.0037

-0.0220***

0.0046

-0.0213***

0.0035

-0.0082***

0.0032

-0.0105**

0.0034

cos2.4mo

0.0200***

0.0035

0.0331***

0.0042

0.0105***

0.0033

0.0064*

0.0029

0.0002

0.0031

Minimum
Temperature

-0.0003***

2e-05

-0.0004***

2e-05

-0.0003***

2e-05

-0.0002***

2e-05

-0.0001***

2e-05

Average Relative
Humidity

0.0002***

2e-05

0.0002***

2e-05

0.0002***

2e-05

0.0001***

2e-05

0.0001***

2e-05

Average PM10

2e-05***

5.13e-06

2e-05***

6.29e-06

2e-05***

4.90e-06

le-05**

4.36e-06

5.89e-06

4.65e-06

Max of Ozone lHr

-0.0536***

0.0085

-0.0705***

0.0105

-0.0397***

0.0081

-0.0462***

0.0072

-0.0279***

0.0077

constant

9.5639***

0.2153

9.0780***

0.2633

8.1843***

0.2059

7.4835***

0.1833

6.0107***

0.1956

Note: ***= statistically significant at 99% confidence level; **=statistically significant at 95% confidence level;

*=statistically significant at 90% confidence level

12

-------
Looking across the five models presented in Table 3, we find mixed results. Minimum
temperature enters all five equations with the expected sign. As temperature decreases, we
expect to see an increase in asthma symptoms as exhibited in the results. The decrease in the
magnitude of the effect across severity levels is not altogether surprising given the increased use
of maintenance therapies as asthma severity increases. This is in keeping with a recent study by
Delfino et al. (2002) that found stronger associations between air pollution and exacerbation of
asthma symptoms among asthmatic children who were not taking anti-inflammatory
medications.

The effects of relative humidity are somewhat puzzling, however, as we expected dryer
air to exacerbate asthma symptoms. In our results, it seems we have the opposite effect — as
relative humidity increases, so does the number of asthma prescriptions filled.

The effect of pollution exposure on asthma symptoms is also mixed. Daily levels of
PM10 have a positive and significant effect on asthma prescriptions with a 1 microgram per
cubic meter increase in PM10 resulting in a 0.00002 percent increase in asthma prescriptions
each week (or approximately 0.12 additional prescriptions). Assuming each inhaler contains
approximately 200 metered doses of the quick relief medication with the recommended usage to
relieve symptoms being 2 doses, this translates to approximately 12 "attacks" per week. The
effect of ozone exposure however is contrary to what we expected, with our model showing
negative but statistically significant effects across all severity levels.3

Next Steps

Our regression results are unsatisfying in many respects. We fully expected to see a

3 We intended to incorporate the effects of NOx and S02 as well, but the data coverage was too spotty for these
pollution measures.

13

-------
negative (and statistically significant) effect of ozone exposure on asthma prescriptions. Still, our
analysis is preliminary and we hope that a number of improvements to our model will likewise
improve our results.

First, we recognize that our negative binomial model is rather crude. The model does not
take into account the autoregressive nature of the data and as such could be producing biased
results. Now that we have selected the appropriate terms to incorporate into the model, we
intend to explore the use of generalized additive models.

We also recognize that we applied the model we constructed for total prescriptions to
other endpoints (counts by severity level) with little regard for whether it produced the best fit in
these other contexts. Ideally, we would apply the same methodology for constructing the model
to these other endpoints - controlling for the factors we believe to be important (e.g., weather,
pollution, and seasonality) but using the AIC to assess which measures best control for these
factors and eventually applying generalized additive models here as well.

The selection of the window over which to sum our observations was admittedly rather
ad hoc. We intend to explore other more "rigorous" means of selecting the proper window. One
option is to conduct a rather crude "expert elicitation" and identify (fewer than 10) physicians
whom we could survey on this point. Essentially, we would want to learn from them
approximately how often their patients refill their prescriptions.

In addition, we have access to the counts of prescriptions by age group as well as by
severity. While time did not permit us to present the results of regressions by age here, we
intend to explore the effects of pollution exposure on asthma prescriptions by age group in future
analyses.

14

-------
References:

Delfino, Ralph J., Robert S. Zeiger, James M. Seltzer, Donald H. Street, and Christine E.

McLaren (2002). "Association of Asthma Symptoms with Peak Particulate Air Pollution
and Effect Modification by Anti-inflammatory Medication Use. " Environmental Health
Perspectives. 110(10): A607-617.

Hajat, S., A. Haines, S. A. Goubet, R.W. Atkinson and H.R. Anderson (1998). "Association of
Air Pollution with Daily GP Consultations for Asthma and Other Lower Respiratory
Conditions in London." Thorax. 54: 597-605.

Hiltermann, T.J.N., J. Stolk, S.C. van der Zee, B. Brunekreef, C.R. de Bruijne, P.H. Fischer, C.B.
Ameling, P.J. Sterk, P.S. Hiemstra, and L. van Bree (1998). "Asthma Severity and
Susceptibility to Air Pollution." European Respiratory Journal. 11: 686-693.

Krupnick, Alan J., Winston Harrington, Bart Ostro (1990). "Ambient Ozone and Acute Health
Effects: Evidence from Daily Data." Journal of Environmental Economics and
Management. 18:1-18.

Neukirch, Francoise, Claire Segala, Yvon Le Moullec, Myriam Korobaeff, and Michel Aubier
(1998). "Short-term effects of low-level winter pollution on respiratory health of
asthmatic adults " Archives of Environmental Health. 53(5): 320-9.

Newhouse, C.P., B.S. Levitin, and E. Levitin (2004). "Correlation of Environmental Factors with
Asthma and Rhinitis Symptoms in Tulsa, OK. Ann. Allergy Asthma Immunol. 92: 356-
366.

Ostro, Bart D., Michael Lipsett, Matthew B. Wiener, and John Seiner (1991). "Asthmatic

Responses to Airborne Acid Aerosols." American Journal of Public Health. 81: 694-702.

Peters, Annette, Inge F. Goldstein, Ulrich Beyer, Kathe Franke, Joachim Heinrich, Douglas
Dockery, John D. Spengler, and H. Erich Wichmann (1996). "Acute Health Effects of
Exposure to High Levels of Air Pollution in Eastern Europe." American Journal of
Epidemiology. 144(6): 570-81.

Pope, C. Arden, Douglas W. Dockery, John D. Spengler, and Mark E. Raizenne (1991).

"Respiratory Health and PM10 Pollution." American Review of Respiratory Disease. 144:
668-674.

Schwartz, Joel (1994). "Air Pollution and Hospital Admissions for the Elderly in Birmingham,
Alabama." American Journal of Epidemiology. 139(6): 589-98.

Schwartz, Joel, David Wypij, Douglas Dockery, James Ware, Scott Zeger, John Spengler, and
Benjamin Ferris, Jr. (1991). "Daily Diaries of Respiratory Symptoms and Air Pollution:
Methodological Issues and Results " Environmental Health Perspectives, 90: 181-187.

15

-------
Sunyer, Jordi, Marc Saez, Carles Murillo, Jordi Castellsague, Francesc Martinez, and Josep M.
Anto (1993). "Air Pollution and Emergency Room Admissions for Chronic Obstructive
Pulmonary Disease: A 5-year Study." American Journal of Epidemiology. 137(7): 701-5.

Walters, Sarah ,R.K. Griffiths, J.G. Ayers (1994). "Temporal Association Between Hospital
Admissions for Asthma in Birmingham and Ambient Levels of Sulphur Dioxide and
Smoke." Thorax, 49: 133-140.

Zenghnoun, Abdelkrim, Pascal Beaudeau, Fabrice Carrat, Veronique Delmas, Onealy

Boudhabhay, Francois Gayon, Dominque Guincetre, and Pierre Czernichow (1999). "Air
Pollution and Respiratory Drug Sales in the City of Le Havre, France, 1993-1996."
Environmental Research, Section A, 81: 224-230.

16

-------
No documents are available regarding Bryan Hubbell's
discussion comments.

-------
Comments by Glenn Blomquist on
"Risk Assessment and Valuation of Health Effects from Air Pollution"
EPA Workshop on Morbidity & Mortality: How Do We Value the Risk of Illness & Death?

April 10, 2006, Washington, DC

Comments on "Willingness to Pay for Improved Health: A Comparison of Stated and Revealed
Preference Models" by W. Michael Hanemann and Sylvia Brandt

This paper discusses two approaches to valuing changes in asthma morbidity among
children: (1) contingent valuation and (2) household production of health. The study is designed
to facilitate comparison of estimates based on the two approaches by estimating willingness to
pay (WTP) for the same households and children. While the comparison is incomplete at this
stage of the project, the potential for comparison between the stated WTP in contingent valuation
and WTP inferred from household averting and mitigating behavior is great. The focus of this
paper is on the results of Hanemann and Brandt's survey of households with asthmatic children,
parents' perceptions of risks of asthma attacks, and their behavior with regard to those risks.

The study group was recruited by Fresno Asthmatic Children's Environment Study
(FACES) in Fresno, California. All children had clinically diagnosed asthma and nearly 70% of
the households had at least one parent who was affected by asthma. These households are
familiar and experienced with asthma. The written, mail survey of 202 households was
conducted in 2004. Information about asthma severity, medication use, asthma-related
expenditures, items and services purchased, household income, and time spent dealing with
children's asthma was collected. The comprehensiveness and detail are incredibly good. Any
doubt that asthma health risks are partly endogenous surely disappears when confronted with
these data.

Table 5 is particularly informative. Seventeen types of fixed costs (expenditures) are
listed including purchases of air conditioning, air filters, pest extermination, carpet removal, and
pet removal. Eight types of variable costs for household supplies are listed including heater
filters, cleansers for mold, and hypoallergenic cleaners. Four types of pharmaceutical costs are
listed including prescription asthma medication, over-the-counter drugs, and herbal remedies.
Seven types of alternative therapies are listed including nutritionists. The time periods for fixed
and variable costs are not specified in Table 5, but if they are (recklessly) lumped together
average (mean) expenditures appear to be roughly 1% of median household income for group.

As Grossman (1972) showed more than 30 years ago, time inputs in the production of
health matter. My educated guess is that ways in which households change their time allocations
due to their children's asthma will be at least as important as changes in money expenditures.
About this behavioral response, Hanemann and Brandt have information reminiscent of diary
data collected for EPA for a small sample of adult asthmatics in the 1980s. These data for adults
are extraordinary and, as evidenced by the recent study by Yen, Shaw, and Eiswerth (2004), are
still being gleaned. The new data for households with asthmatic children should be at least as
useful. They deserve their own table comparable to Table 5. Hanemann and Brandt report that
time costs are sizable. They should be able to estimate them well. They report that 43% of the

1

-------
households took time off from work to take a child to a medical appointment and that the
average (mean) time taken off was 83 minutes. They report 24% of the households took time off
from work in the previous year to take a child to an emergency room with an average (mean) of
211 minutes spent. If wage data are available for each household, dollar values of these time
costs can be estimated and added to the dollar expenditures on marketed goods and services. If
wage data are not available, then given the available information on household income and
characteristics of the household, Hanemann and Brandt should be able to use a data set such as
the Current Population Survey and estimate wages and/or shadow wages for each of the
members of the household. While these time costs may be less than 1% of annual household
income, there are indications of substantially higher time costs. These higher costs are due to
changes in employment status due to having hours reduced (28%), being fired or laid off (22%),
or choosing to work fewer hours during asthma season (21%). More than two-thirds (10%) of
the households reported that a parent had chosen to work part-time or be a stay-at-home parent
due to a child's asthma. While the value of time not spent at work for those who have decided
not to work in the market is not zero, the value of these time costs could be much greater than the
money expenditures for some households. We should look forward to estimates that exploit
these data using the best techniques that labor economics has to offer. The fact that some of the
household time is unpriced is a barrier that can be overcome.

Hanemann and Brandt note several conceptual and empirical limitations of the household
production approach that they believe are threats to the "validity" of the production function
approach to estimating WTP for reducing children's asthma morbidity. My view is that their
concern is legitimate, but that none of the limitations is a fatal flaw that should prevent them
from making a meaningful comparison with the WTP estimates from contingent valuation. My
assessment is that a number of high quality studies have been done valuing changes in morbidity
using a household production approach despite limitations. A recent example is Dickie's (2005)
article on valuing children's health, work for which he received the Georgescu-Roegen Prize.
Hanemann and Brandt have excellent information on what goes on in, what to some is, the black
box of household production of health. Perhaps it is the richness of the information that makes
them hint that a meaningful comparison cannot be made. For example, they report that more
than 20%) of the households said they had no control of the top ten potential asthma triggers.
Viewed differently, the fact that nearly 80%> said they had at least some control would be
reassuring to many who would apply the household production approach.

In Table 7, nonmarket, household averting and mitigating behaviors is listed based on
Hanemann and Brandt's first survey. They include activities such as checking for smog alerts,
closing windows, and restricting where the child can play as well as parents giving up smoking
cigarettes. These activities can be difficult for researchers to value in dollars for a household
production estimate, as they note. This information guided them in designing their contingent
market for a hypothetical BreatheRight watch for monitoring. In addition to trying to put dollars
values on the activities directly, they might consider adding some contingent time tradeoffs
and/or contingent behavior questions to value the nonmarket averting and mitigating behaviors.

All in all, I think that Hanemann and Brandt believe that they have laid the groundwork
for a first-class contingent valuation study to estimate the values that parents place on improved
control of their children's asthma. I share that belief and think further that they have laid the

2

-------
groundwork for a first-class household production study, and comparison between the two. I
look forward to seeing the successful completion of all three.

Comments on "Individual Preferences and Household Choices: The Potential Role of
Dependency Relationships" by Mary F. Evans, Christine Poulos, and V. Kerry Smith

Since this project is a work in progress and no paper was available, my comments, unlike
good Kentucky bourbon, have aged only one hour. Here are a few quick reactions. One is that
the motivation for valuing changes in the environment is not entirely clear. Modeling
households as groups in which individuals' roles are treated as economic decisions is offered as a
way to gain insights into household relationships. Evans, Poulos, and Smith believe that insights
gained could be used as an alternative to focus groups in the development of surveys and survey
instruments. Presumably the values that are elicited for changes in environmental quality could
depend on which member of the household is asked. My comment is that several good reasons
exist for using focus groups and that even if we learn something about household relationships,
there are still potentially great benefits to focus groups. If focus groups are held, the marginal
cost of exploring household relationships is probably low.

A second comment is that it would be interesting to work through the system of equations
for an exogenous change in a quantity, as we often do in environmental economics, instead of a
change in price. I am not sure that makes sense based on the brief presentation, but it might be
worth considering.

My third and last comment is that the household structures presented did not include the
one that I consider the most important, namely a household with at least three individuals.
Because I have been father in a household of four in which two parents were involved in raising
a child with chronic asthma, I think that an important household configuration has been omitted
in the modeling so far. In our household, I think the values you would have elicited in surveys
would have been fairly close regardless of which of us you asked. It would be good to have a
model that allowed for that situation. What comes from this research could be fascinating.

Comments on "Air Pollution and Asthma: Preliminary Results from a Daily Time Series Study
of San Francisco" by Charles W. Griffiths and Nathalie B. Simon

This paper is similar to Hanemann and Brandt's in that deals with morbidity related to
asthma and air pollution and studies residents of a city in California. The idea is that since air
pollution can trigger asthma attacks, the pattern of filling prescriptions for medications that
relieve asthma symptoms, should be influenced by the pattern of air pollution. Air pollution
episodes lead to prescription episodes with some lag. My main comment is that the probability of
inhaler use, as shown in equation 1, should be broadened to incorporate human behavior. In the
context of household production of health, the use of a market input such as an inhaler will
depend on exogenous factor such as pollution, weather, and seasonal factors, as Griffiths and

3

-------
Simon indicate. In addition, however, inhaler use will depend on averting and mitigating
behavior such as limiting outdoor exercise, and the myriad of things documented by Hanemann
and Brandt. Inhaler use will depend on how well the individual manages controller therapy. Use
of these long-term maintenance drugs is an investment that yields a return weeks later when
asthma attacks are prevented or reduced in severity.

Weekly count data of prescriptions for quick relief asthma medication for San Francisco
for the years 1998-2001 were analyzed using a standard negative binomial regression. Increases
in average PM10 increased counts and increases in average minimum temperature decreased
counts as expected. Increases in average relative humidity are found to increase prescription
counts, a result which Griffiths and Simon did not expect. As one who lives in the Bluegrass
Region of Kentucky and associates humidity with lush growth, abundant pollen, thriving mold,
and other asthma triggers, I am not surprised by the positive sign on humidity. The result that
ozone is associated with a decrease in prescription counts is unexpected. My suggestion, based
on my main comment, is to think about inhaler use as determined by individual factors to see if
that suggests other variables. If a pollution episode is correctly anticipated, asthmatics are
carefully engaged in averting behavior, and long-term maintenance drugs are effective, the
increase in use of rescue drugs and increase in prescription counts will be small. This
explanation does not distinguish between PM10 and ozone, but perhaps it will lead to a better
specification that produces results that are more in line with expectations.

My last comments are that I agree with Griffiths and Simon that they should do more time
series diagnostics and more sensitivity analysis of the 180 day window that they use for pollution
measures. The windows for PM10 and ozone may be different.

References

Dickie, Mark. "Parental Behavior and the Value of Children's Health: A Health Production
Approach" Southern Economic Journal 71 (April 2005): 855-872.

Grossman, Michael. "On the Concept of Health Capital and the Demand for Health" Journal of
Political Economy 80 (March 1972): 223-255.

Yen, Steven T., W. Douglas Shaw, and Mark E. Eiswerth. "Asthma Patients' Activities and Air
Pollution: A Semiparametric Censored Regression Analysis" Review of Economics of the
Household2 (March 2004): 73-88.

4

-------
Summary of the Q&A Discussion Following Session I

Reed Johnson, (RTI)

Directing his comment to Mary Evans and Christine Poulos, Dr. Johnson stated, "I
thought I followed your model pretty well, Mary, and then you offered the format of the
question you were going to ask, which was as I recall: For a given improvement in air
quality that would reduce the amount of care giving time, how much would you be
willing to pay? I don't quite understand, in the context of the other presentations we've
heard today, what you're assuming about the household production function. That is, are
you assuming some fixed proportion of time for a decrease in . . . ?"

Mary Evans, (University of Tennessee)

"Yes, at least in the pilot survey the care giving time is going to be exogenous, so that
will be determined in the stated preference survey. So we exogenously specify both the
initial care giving time and the reduction in the final care giving time."

Reed Johnson

Dr. Johnson continued, "And what are they going to get in terms of improved health? Is
it possible to value the time independently of the change in the health outcome
experienced in the household?"

Mary Evans

Dr. Evans clarified, "In some household structures, they'll value both the health impact as
well as the reduction in time allocation."

Reed Johnson, (RTI)

Dr. Johnson responded, "So if it's not altruistic, they wouldn't value the health outcome."
Mary Evans

"Right. If there's only that income-sharing model, the only way that air quality would
impact that particular model is through the mu function essentially—through the
individual's share of the household income."

Reed Johnson

Dr. Johnson added, "I guess it wasn't clear to me how the model handles substitution
among various household production inputs and how that plays out in the willingness to
pay."

Lauraine Chestnut, (Stratus Consulting, Inc.)

Ms. Chestnut addressed Charles Griffiths: "Regarding the negative ozone coefficient—
that colder temperatures give higher medicine use—I was just wondering how much you
looked into the correlation between that, since the ozone tends to be worse in warmer
weather."

Session I Q&A

1

-------
Charles Griffiths, (U.S. EPA, NCEE)

Dr. Griffiths responded, "That's actually an excellent point that was raised to me just
recently, which is why I kept saying that the way we've modeled ozone is counter-
intuitive." Acknowledging that the factor "currently enters in straight," he offered that he
"should account for the fact that the ozone effect may be seen only during a certain
season." He said, "It may be washed out by the fact that I'm not accounting for the
seasonality of the ozone effect."

END OF SESSION I Q&A

Session I Q&A

2

-------
Morbidity and Mortality: How Do We Value the Risk of

Illness and Death?

PROCEEDINGS OF SESSION II: ISSUES WITH MORBIDITY VALUATION

AND

KEYNOTE ADDRESS BY BRIAN MANNIX

A WORKSHOP SPONSORED BY THE U.S. ENVIRONMENTAL PROTECTION
AGENCY'S NATIONAL CENTER FOR ENVIRONMENTAL ECONOMICS AND
NATIONAL CENTER FOR ENVIRONMENTAL RESEARCH

April 10-12, 2006

National Transportation Safety Board
Washington, DC 20594

Prepared by Alpha-Gamma Technologies, Inc.
4700 Falls of Neuse Road, Suite 350, Raleigh, NC 27609

ACKNOWLEDGEMENTS

This report has been prepared by Alpha-Gamma Technologies, Inc. with funding from
the National Center for Environmental Economics (NCEE). Alpha-Gamma wishes to
thank NCEE's Maggie Miller and the Project Officer, Cheryl R. Brown, for their
guidance and assistance throughout this project.

DISCLAIMER

These proceedings have been prepared by Alpha-Gamma Technologies, Inc. under
Contract No. 68-W-01-055 by United States Environmental Protection Agency Office of
Water. These proceedings have been funded by the United States Environmental
Protection Agency. The contents of this document may not necessarily reflect the views
of the Agency and no official endorsement should be inferred.

-------
Table of Contents

Keynote Address

Brian Mannix, Associate Administrator, U.S. EPA, Office of Policy, Economics, and
Innovation

Introduction by: Al McGartland, Director, U.S. EPA, National Center for Environmental
Economics

Session II: Issues with Morbidity Valuation

Session Moderator: William Wheeler, U.S. EPA, National Center for Environmental
Research

IOM and Cost Effectiveness

Nathalie Simon, U.S. EPA, National Center for Environmental Economics

Altruism and Environmental Risks to Health of Parents and Their Children

Mark Dickie and Shelby Gerking, University of Central Florida

Is an Ounce of Prevention Worth a Pound of Cure?

Ryan Bosworth and Trudy Cameron, University of Oregon; and J.R. DeShazo,
University of California-Los Angeles. Presented by: J.R. DeShazo.

Discussant: Kelly Maguire, U.S. EPA, National Center for Environmental
Economics

Discussant: Kevin Boyle, Virginia Polytechnic Institute
Questions and Discussion

-------
U.S. EPA NCER/NCEE Workshop
Morbidity and Mortality: How Do We Value the Risk of Illness and Death?

Washington, DC

April 10-12, 2006

Keynote Address
Brian Mannix, Associate Administrator,

U.S. EPA, Office of Policy, Economics, and Innovation

He [A1 McGartland] told you I'm not a lawyer, but he didn't say what my background
is—I'm actually a chemist, although you couldn't tell by my career. . . . What I want to
do is step back and take a look at the metrics that we use to describe the benefits of
mortality reductions that we attribute to environmental regulations. In particular, I want
to raise questions about the statistical robustness of the "lives saved" metric that is now
commonplace. I should say that years ago I was an advocate for VSL analysis at the
beginning of my career, and I encouraged EPA to focus on lives saved. Now that I'm
back at EPA (and I did start here at EPA in 1977), I'm surprised at how much progress
has been made in incorporating VSL into Agency analyses and decisions. I'm surprised,
too, that I'm not very comfortable with where that progress has left us, and I'm most
surprised to find that the most serious difficulty in my mind turns out not to be with the
"V" but with the "SL." That is, economic valuation of mortality benefits is a tractable
problem analytically and politically, but figuring out the right metric for mortality
benefits is much more problematic. I'll illustrate this with a contrived example:

Suppose on Monday a hospital in a small town publishes a press release announcing that
over the busy weekend it had managed to save a dozen lives. The local TV station sends
down a camera crew and asks if it can interview a few of the lucky survivors. The ER
nurse tells them, "I'm sorry, that won't be possible—he died." "What do you mean—
who died?" the reporter asks. "The man who was having the heart attacks," the nurse
replies. "We managed to save him 12 times in 13 attempts."

The point of this story is that while we can easily count "lives" or "deaths," we cannot
easily count "lives saved." It is not well defined, and it is inherently unbounded. The
airbag may save your life in the event that your brakes fail, but how many times has your
life been saved when the brakes didn't fail? The number of lives saved during my
commute this morning is already beyond my ability to reckon. In some narrow context,
we might be able to come up with a workable definition of a "life saved." As a lifeguard,
Ronald Reagan would put a notch in a log every time he saved a life, and I don't doubt
that it was accurate and meaningful. If he had kept the notched log during his presidency,
however, I can't imagine how we would come up with an accurate count—or interpret it
if we had one.

I don't believe it is possible to come up with a definition of "lives saved" that is robust,
that can be applied to a wide variety of situations, and that can be aggregated in a

B. Mannix, Keynote Address

1

-------
statistically meaningful way. The underlying difficulty is that "lives saved" lacks a time
dimension. We know that all lives are temporary, and while the valuation problem is
quite complex, we are generally in agreement that a longer life is better than a shorter
one. If we don't capture the time dimension, we are unlikely to come up with a metric
for mortality that is versatile and that behaves well in statistical usage.

There is a standard statistic for measuring longevity that everyone is familiar with—the
expected value of the length of life, or life expectancy. It has several advantages in
communicating with the public. Everyone has a pretty good idea of what it measures.
People also have a good sense of what the units mean. They may have a great deal of
difficulty picturing what a "ten to the minus six" risk of death is, but they know how long
a minute is and how long 10 years is, and that covers more than six orders of magnitude.
This also solves the problem of divisibility—some find it difficult to think about a
fraction of a life saved or about the same life being saved multiple times, but they have
no trouble dividing time into units of arbitrary size. The public will also have less
difficulty attaching a monetary value to changes in life expectancy, even those who
cannot imagine attaching a finite value to a life saved.

I should mention here that just yesterday I saw a new Ford commercial in which Bill Ford
says, "Every life saved is worth it." I couldn't agree more. The real advantage of using
life expectancy, though, is that it is a well-defined and well-behaved summary statistic
that reflects mortality risks across an entire population, including risks of all kinds and at
all ages without discriminating against any particular subgroup. Let's suppose we're
evaluating a range of policy options, all of which have a small marginal effect on
mortality risks. If we take as our mandate to maximize life expectancy using limited
resources, we can easily solve the problem. We know that the solution will give us a
cost-effectiveness criterion, a fixed dollar amount per incremental year of life
expectancy. The decision rule would be to adopt those measures that met the cost-
effectiveness criterion and to avoid committing resources to those that didn't.

Note that if we use another decision criterion in place of this one, we will get a shorter
life expectancy for the same expenditure of resources. If we use a VSL rule, for example,
we might save more lives, whatever that might mean, but on average, people will live
shorter lives. In most cases I think the two criteria would likely lead to similar outcomes.
When they don't, however, we have to ask whether policies that cause a shorter life
expectancy can really be said to be improving public health. Similarly, if we adopt
maximum life expectancy as the goal, but make adjustments to the metric for age,
quality, or willingness to pay, the result will be that people live shorter lives—better,
maybe, in some sense, but shorter. I believe this creates a strong presumption for using
life expectancy as a standard metric in evaluating regulatory decisions, using a flat VSLY
(value of a saved life year) as the cost-effectiveness criterion. As the first-order
approximation of mortality benefits, I think this is vastly superior to the VSL approach,
and I think anyone advancing some other decision rule needs to explain how we can
justify adopting policies that will lead to a shorter life expectancy. I don't rule out that
such justifications may exist, but I think we should be cautious in entertaining them.

B. Mannix, Keynote Address

2

-------
A1 McGartland has pointed out to me that there's a contradiction here. I embrace the use
of willingness-to-pay data in figuring out what our cost-effectiveness criterion should be,
but I shrink from looking any deeper into the data to find out how it might vary from
group to group or person to person. I think this is a contradiction I can live with. An
individual, perhaps because he is wealthy, who is willing to pay and does pay much more
than average to reduce his own mortality risks, should certainly be able to do so.
However, I am not ready to concede that that same individual is entitled to tilt public
health measures in his favor simply because he is willing to pay—but does not pay—for
them. When writing rules or spending public funds, there is an egalitarian consideration
that does not apply when individuals are spending their own money. As analysts we may
feel that we can improve the analysis by making adjustments for age or quality or to
incorporate the latest willingness-to-pay data, but as a government official I'm reluctant
to go very far down that road. In part, that's because I question whether government has
any legitimate business making such adjustments, and in part it's because if the
government did get into that business, the adjustments would likely be made according to
the rules of politics, not necessarily those of economic analysis. So, perhaps a flat VSLY
is desirable for the same reason that a flat tax is appealing to some—it minimizes the
opportunities for making mischief.

I'll stop there and look for reactions.

(A question and answer session followed.)

B. Mannix, Keynote Address

3

-------
loM Committee
Recommendations and Cost
Effectiveness Analysis at EPA

Nathalie B. Simon, USEPA

Presentation Prepared for "Morbidity and Mortality:
How do We Value the Risk of Illness and Death
April 10-12, 2006

Disclaimer

The views expressed in this presentation are entirely those
of the presenter and do not necessarily represent those of
the USEPA. No endorsement by the Agency should be
inferred.

-------
A Very Brief History

•	In 2003, OMB released Circular A-4

-	Requiring Agencies to perform health-based cost effectiveness
analyses of economically significant rules where health is the primary
benefit

-	Prior to this, EPA seldom performed CEAs using health-related quality
of life measures.

•	Later that year, the Institute of Medicine (loM) convened a
committee, at the request of John Graham and with funding from
several Agencies, to address a number of questions on how best to
perform CEAs in a regulatory context.

•	Specifically, the Committee was asked to:

-	Describe current agency practices in estimating benefits and costs of
regulatory actions

-	Review measures currently used in CEAs to aggregate health
improvements

-	Develop criteria for choosing among available measures

-	Assess the various measures for data requirements, feasibility,
theoretical validity and ethical implications

-	Recommend measures appropriate for federal agency use

•	loM released its long anticipated report in January 2006

Recommendation 1: Regulatory CEAs that integrate
morbidity and mortality impacts in a single effectiveness
measure should use the quality-adjusted life year to
represent net health effects.

Recommendation 2: Regulatory analyses should report
four measures of cost- effectiveness:

-	Compliance cost per death averted

-	Compliance cost per life year gained

-	A health-benefits-only ratio using the net change in
QALYs as the outcome measure.

-	A comprehensive ratio using QALYs as the outcome
measure and incorporating the value of other benefits
as offsets to compliance costs.

-------
Recommendation 3: The life year and QALY estimates
used in regulatory analyses should reflect actual population
health as closely as possible

Recommendation 4: Incremental cost-effectiveness ratios
are generally the most useful summary measure for
comparing different regulatory interventions.

Recommendation 5: In addition to reporting effects in the
aggregate, regulatory analyses should report QALY
impacts separately for each health endpoint.

Recommendation 6: The reporting of all CEA results
should be accompanied by information on related
uncertainties and non-quantified effects.

Recommendation 7: Regulatory analyses should not
assign monetary values to estimates of health-adjusted life
years as a method for valuing health states.

Recommendation 8: The regulatory decision-making
process should explicitly address and incorporate the
distributional, ethical, and other implications of a proposed
intervention along with the quantified results of BCA and
CEA.

Recommendation 9: Policy makers and program
administrators should work to ensure the substantive
involvement of a broad range of individuals and groups at
all stages of policy development for regulating risks.

6

-------
Recommendation 10: A high research priority should be
improving the data used to assess the health risks (effects
on incidence of particular types of illness, injuries, and
deaths, and the duration and latency of effects) addressed
by regulatory actions.

Recommendation 11: The Department of Health and
Human Services (DHHS) and other federal agencies
should collect HRQL information through routinely
administered population health surveys and other major
studies and data collection efforts related to risk
assessment and monitoring.

Recommendation 12: DHHS should coordinate, with the involvement
of federal regulatory offices and agencies, the development of an
integrated research agenda to improve the quality, applicability, and
breadth of HRQL measures for use in regulatory CEA. The Committee
identifies the following areas as priorities for research:

-	Current elicitation methods such as the standard gamble and time trade-
off, while theoretically well founded, may be difficult for respondents to
understand and prone to generate inconsistent responses. Research to
facilitate improved methods is needed. In addition, methods for eliciting
societal values for investments in health (in contrast to individual
preferences for health states), such as person trade-off techniques,
should also be investigated.

-	Methods for measuring children's health-related quality of life, including
characterization of the impact of illness and injury and the valuation of
these impacts, need continued development and refinement.

-	Methods to correlate QALY estimates based on different generic HRQL
indexes should be developed so that estimates from different underlying
valuation studies are consistent and can be used in the same analysis.

-------
Some Key Issues for EPA

What do we do if BCA and CEA produce different rankings of
policies/programs... or if the four CEA ratios produce different
rankings?

We sometimes have policies that reduce morbidity in one
population and prevent deaths in another. How do we combine
these effects using a single HRQL index?

Are QALYs biased since they undervalue health gains to the
disabled, elderly and chronically ill?

How should we report uncertainty? Do we somehow develop
ranges of QALYs as we've been asked to do with benefits
estimates?

What do we do when we are unable to quantify the impacts but
we know what they are?

How will the CEA results be used? Will they be used to construct
league tables? These may lead to systematic bias away from
environmental policies to direct health policies simply because we
are unable to quantify some of the environmental effects.

Next Steps for EPA

•	The Agency is in the process of updating its Guidelines
for Preparing Economic Analysis

•	A cross-office workgroup was recently convened to write
a chapter on CEA for the Guidelines

-	Working through loM recommendations

-	A public review draft is anticipated by end of calendar year.

For those of you interested in learning more about the loM
report:

see "recent reports" at www.iom.edu

10

-------
Altruism and Environmental Risks to Health of
Parents and their Children*

Mark Dickie
and

Shelby Gerking**

Department of Economic
University of Central Florida
Orlando, FL 32826

June 2, 2006

Running Title: Altruism and Environmental Risks
JEL Codes: Q51, Q58, D13, D64.

*The US Environmental Protection Agency (USEPA) partially funded the research described here under
R-82871701-0. The research has not been subjected to USEPA review and therefore does not necessarily
reflect the views of the Agency, and no official endorsement should be inferred. Gerking acknowledges
the hospitality of CentER, Tilburg University, where this research was begun, as well as support from
Visiting Grant B46-386 from the Netherlands Organization for Scientific Research (NWO). For
numerous constructive comments on earlier drafts, we thank Anna Alberini, Kevin Boyle, Erwin Bulte,
Glenn Harrison, Kyung-So Im, Bill Schulze, Kerry Smith, Aart de Zeeuw, participants in the workshop
on Valuation of Children's Health organized by the National Policies Division, OECD Environment
Directorate, participants at the Conference on Valuing Environmental Health and Risk Reduction to
Children, sponsored by USEPA's National Center for Environmental Economics and National Center for
Environmental Research and University of Central Florida and seminar participants at Tilburg University.

** Corresponding author. E-mail: Shelbv.Gerking@bus.ucf.edu. Mail: Department of Economics,
University of Central Florida, P.O. Box 161400, Orlando, FL 32816-1400. Phone: (407) 823-4729. Fax:
(407) 823-3269.

-------
Altruism and Environmental Risks to Health of
Parents and their Children

ABSTRACT

This paper tests an equilibrium condition from a model that incorporates: (1) altruism of
parents toward their young children and (2) household production of latent health risks. The
model demonstrates that an altruistic parent's marginal rate of substitution between an
environmental health risk to herself and to her child is equal to the ratio of marginal risk
reduction costs. Econometric estimates support this prediction based on data from a stated
preference study involving 488 parents of children aged 3-12 years. This outcome implies that
parents reallocate family resources to at least partly offset the effectiveness of public programs
that aim to reduce their children's environmental risks.

Key words: Altruism, household production, environmental risk, child health.

1

-------
Altruism and Environmental Risks to Health of
Parents and their Children

1. Introduction

Special protection of young children from environmental hazards has become a
worldwide priority in government policies to improve human health.1 Effectiveness of these
measures depends on what steps parents voluntarily take to keep children out of harm's way. If
parents are naive about hazards, do not care about their children, or lack the resources to protect
their health, implementation of well-designed public policies to increase protection of children
may have the intended effect. On the other hand, if parents are informed, altruistic, and
sufficiently well off financially, measures aimed at increasing protection of their children from
particular hazards will be offset to some extent as parents redistribute family resources. In any
case, the fundamental tension between altruism and self-interest in family exchange looms as a
crucial behavioral factor determining the effectiveness of government policies to protect
children's health.

What is known about altruism in families? Several prominent empirical studies (e.g.,
Cox and Rank 1992, Altonji, Hayashi, and Kotlikoff 1992, 1997, Laitner and Juster 1996) do not
support the implication of altruism for transfer-income derivatives in examining inter-household
financial transfers between parents and adult children. Other papers (e.g., Liu et al. 2000,
Jenkins, Owens, and Wiggins 2001, Nastis and Crocker 2003, Agee and Crocker 2004, Dickie
and Messman 2004) look at how parents protect themselves and their pre-teenage children from

1 For example, Executive Order 13045 (Federal Register, 1997) directs U.S. federal executive branch agencies to
assign a high priority to addressing health and safety risks to children, coordinate research priorities on children's
health, and ensure that their standards take into account special risks to children. The U.S. Environmental Protection
Agency has formulated a seven-step strategy to protect children's health (U.S. EPA 1996). Some of the more visible
federal decisions in which protection of children's health figured prominently include tightening of air quality
standards for ozone and particulate matter and implementation of the 1996 Safe Drinking Water Act Amendments
and the 1996 Food Quality Protection Act. Scapecchi (2006) summarizes similar efforts undertaken in other
countries.

2

-------
environmental and other hazards. In this branch of the literature, altruism is sometimes
mentioned as a possible parental motivation, but equilibrium conditions implied by altruism are
not tested.

This paper tests a model of altruistic family behavior (Becker 1974, 1981 and Barro
1974) that incorporates household production of latent health risks. The model demonstrates that
the parent's marginal rate of substitution between risks faced by herself and her child is equal to
the ratio of marginal risk reduction costs. This prediction is tested using survey data on skin
cancer risks faced by 488 parents in Hattiesburg, MS and their biological children between the
ages of 3 and 12 years. Marginal rates of substitution are obtained from stated preference values
for a hypothetical sun lotion. While stated preference valuation remains a controversial method
of obtaining willingness to pay for reduced environmental risk, its application here supports
consistent estimation of the desired marginal rates of substitution because of the way the survey
(described more fully later on) is designed. Test outcomes support the model and imply that
parents are altruistic toward their young children.

2. Conceptual Framework
2.1 Model

This subsection presents an extension of Becker's (1981) model of altruism that
incorporates household production of latent health risks. The model envisions a "family"
composed of one altruistic parent and one child. Because only one child is included in the
model, the analysis focuses on how parents allocate resources between themselves and their
children, rather than on how parents make tradeoffs among different children. By including only
one parent in the model, a unitary perspective is adopted in which possible divergent interests
between parents in a family are not considered. Although the unitary model has been rejected in

3

-------
several empirical tests (e.g., Lundberg et al. 1997), tests presented in Section 4 reveal no
significant differences in valuation of latent health risks between fathers and mothers.

To facilitate treatment of latent health risks, assume that the parent has two periods of life
remaining while the child has three. During the present period (t = 0), the parent receives all
family income, purchases market goods for her family, and behaves as a paternalistic altruist in
that she derives utility from her own consumption as well as from the combination of goods that
she provides to her child.2 Thus, the parent allocates goods to the child according to her own
views as to what is best and disregards the child's preferences (if any) except in situations in
which they are congruent with her own. In period t = 7, the child will be an adult with his own
income, which the parent may supplement with transfers, and will make his own consumption
decisions. In this period, the parent will derive utility from her own consumption and may also
derive satisfaction from the level of utility achieved by the child. The model therefore envisions
that the parent's altruism may switch from paternalistic altruism to the more all-encompassing
concern for the child's well-being considered by Altonji, Hayashi, and Kotlikoff (1997) after the
child is mature enough to exhibit well-defined preferences and the parent can no longer dictate
the combination of goods that the child will consume.3 In the third and final period (t = 2), the
child continues to receive income and purchase market goods while the parent is deceased.

The survey, described more fully in Section 3, elicits willingness to pay to reduce two
latent environmental health risks facing both the parent and the child. In the model, these two
risks are denoted a and b. To consider a latency period that is longer for the child than for the
parent, assume that the events at risk may occur in the last period of either individual's life.

2	Paternalistic altruism is more fully discussed by Jones-Lee (1991, 1992)

3	Both types of altruism are incorporated into the model to assist in clarifying the interpretation of statistical tests
presented in Section 4. All-encompassing concern for another's well-being has also been termed "benevolence"
(Bergstrom 2006) or "pure" altruism (Jones-Lee 1991, 1992).

4

-------
Constraining the lifetime risk to lie in a single period simplifies the task of communicating
changes in risk to survey respondents (see Section 3). Perceptions of the /th latent risk to the z'th
person are denoted R. , where superscript j distinguishes between the two risks (a and b) while

subscript i distinguishes the parent (p) from the child (k). Perceived lifetime risks are influenced
by the use of market goods that otherwise have no utility:

where Gjt denotes individual Vs use in period t of a market good affecting the /th risk.

Simplifying assumptions here are that: (1) the risk production functions do not shift over time,
(2) the child when grown is assumed to share his parent's assessment of both risks, and (3)
marginal products of the (/' are strictly negative in both production functions.

When the child begins to make his own consumption decisions as an adult in period 1=1,
he will maximize his lifetime utility given by Uk(Ck0,Cki,Ck2,Rk ,Rk) subject to his perceived

risk production functions given in equation (1), the choice of (Ck0,Gk0,Gbk0) that already will
have been made by the parent, and his lifetime budget constraint,

T + ykl+(\ + ryyk2 = Ckl +PaGakl +PbGbkl +(1 + r)~1[Ck2 + PaGak2 +PbGbk2]. Here and in

equations (2) and (3), variables yit and Cit respectively denote individual i 's income and
consumption of an aggregate market good in period t, T denotes the income transfer from parent
to child in period t= 1 ( '/ > 0), r denotes the market interest rate and P1 denotes the market price
of the protective good affecting the /th risk.

In period t= 0 the parent maximizes the utility function

R'P = Rp(Gp0,G}pl),

r: r:(g:.g:.g:2). j = a,b.

(i)

Up (Cp0, Cpl, Ck0, Rap, Rbp, ,Rl) + ijU'l (Ck0, G;o, Gbk0 J,ykl,yk2,r,Pa,Pb)

(2)

5

-------
subject to the four perceived risk production functions in equation (1), the restriction T > 0 and
her lifetime budget constraint

where rj >0 is the weight the parent places on the child's lifetime utility and U*k(») denotes the

indirect utility function from the child's maximization problem. When t = 0, the parent chooses
quantities of all market goods that she and her young child use and when t = 7, the parent makes
these choices only for herself while deciding how much income to transfer to her child.

The parent's paternalistic altruism in period t = Ois reflected in her concern for her
child's present consumption and his risk . If ij = 0, the parent has no further concern for the
child in future periods and will not care how his future choices may affect the lifetime risk he
ultimately faces. If r] >0, the parent continues to care about the child in the future, but she
exhibits benevolence or all-encompassing altruism in that she respects the child's adult
preferences and cares about his overall level of well-being rather than the specific bundle of
goods he consumes.

First order conditions4 for period t = 0 quantities imply that for j = a, b

4 These equations make use of the relationships dll"k / dGk0 = (dUk / dRk)(cR'k / dGk0). Equations (4) and (5) also

make use of the assumption that the parent exhibits paternalistic altruism only in period t=0. Thus her paternalistic
altruissm encompasses concern for how her present choices affect her child's risk but does not extend to concern for
how his future choices may alter his risk. Any concern for the child in future periods is reflected by 77 > 0, not by

dU / dRk. This assumption means that the parent does not have to consider the dependence of her child's future
choices on her decisions today. A more formal analysis of this point is available on request.

» + 0 + O"1^i =C,o +ck 0 +Pa(G;0 + Gak0) + Pb(Gbp0 +Gbk0)
+ (1 + ryl [Cpl +T + PaGapl + PbGbpl I

(3)

dUpISCp(s=dUpl5Ck(s + 71dUkldCk(s

(.dUp !dRJp)(dRJp Wp0) = (dUp/dRIXdRl /dGi0) + V(dUk/dRl)(dRl / «&).

(4)

6

-------
Thus, in period t = 0, the model predicts the familiar result that if both individuals consume C
and G in positive quantities, the parent's marginal rate of substitution between the child's
consumption of C (G) and her own consumption of C (G) is equal to unity.5 This outcome holds
independently of the magnitude of rj, the weight that the child's utility receives in the parent's
utility function, and also holds if the parent exhibits either type of altruism. If instead the parent
exhibits neither type of altruism (i.e., is not an altruist toward the child), then these marginal
rates of substitution equal zero. If the parent exhibits either or both types of altruism toward the
child but does not care about her own consumption or about the level of risks that she faces, then
these marginal rates of substitution are arbitrarily large.

In periods t = 1 and t = 2, first order conditions imply that
dUp/dCpl=Ap(l + ry

dUJdC^Ul + ry-'	t = 1,2

(dUp / dR'p )(dR'p / dG'pl) = ApPJ (1 + rr1	j = a, b	(5)

(.dUkldRi){dWkldGlt) = \P]{\ + rr	j = a,b t = 1,2

il\ = Ap(l + ryl	ifT> 0.

Equation (5) shows that if ij > 0 and ifr > 0, then in period t I the parent's marginal rate of
substitution between the child's consumption of C (G) and her own consumption of C (G) also is
equal to unity. In the case in which // >0, therefore, transfers from the parent to child ensure
that the parent's marginal rate of substitution between the child's consumption of market goods
and her own consumption of market goods is equal to unity in all periods in which both
individuals are alive. If rj > 0, but T = 0 (as may occur in period t = 1 if the child is rich and the
parent is poor) then the parent's marginal rates of substitution between her child's consumption

5 Throughout the paper, the convention adopted for calculating marginal rates of substitution is that the parent's
marginal utility of the child's consumption is in the numerator and the parent's marginal utility of her own
consumption is in the denominator.

7

-------
and her own consumption are positive, but in general are not equal to unity because

^ Ap(l + r). On the other hand, if the parent is a paternalistic altruist only and has no

concern for the child's well-being after period t = 0 has ended (rj = 0), then in period t = 1 the
parent's marginal rates of substitution between her child's consumption and her own
consumption are equal to zero. Finally, just as in period t = 0, if the parent cares about her
child's well-being but not about her own consumption of market goods, then her marginal rates
of substitution between the child's consumption and her own consumption become arbitrarily
large.6

The empirical analysis presented in Section 4 looks at risk reduction, not consumption

ofG;. So, in period t 0, the first order equation for GJ in (4) is rewritten as equation (6) to

show that when corner solutions are set aside, the parent's marginal rate of substitution between

risk to her child and risk to herself is equal to the ratio of marginal products of a risk-reducing

market good that both individuals consume.

(dUpldRl) + r,(dUkldRl) (dRJp / dGJp0) MC^	. ,
	-	:	=	-	-— =	— / = a,b.	(6)

(dUp/dVp)	(PRIISGU) MC^

The ratio of marginal products, in turn, equates to the ratio of present value marginal costs
because the price per unit of GJ is the same no matter who uses it.

Equation (5) also implies that each individual equates the present-value marginal costs of
risk reduction over time, provided that risk production functions are constant over time. Thus, in

6 Equation (5) also implies that when the parent and child consume positive quantities of all goods in all periods, the
inter-temporal marginal rate of substitution between consumption of C (G) in period I I and consumption of C (G)

in period t equals the discount factor (1 + /') 1 for both the parent and the child. The inter-temporal marginal rate of
technical substitution between risk-reducing goods in different periods likewise equals the discount factor for both
individuals. If 7] > 0 and if T > 0, then the parent's marginal rate of substitution between her child's consumption
of C (G) in period I I and her own consumption of C (G) in period t is equal to the discount factor as well.

8

-------
period t = 7, the present-value marginal cost of risk reduction for the parent will be the same as
in period t = 0, and the present-value marginal cost of risk reduction will be the same for the
child in periods t = 1 and t = 2. In addition, if rj > 0 and T > 0, then the marginal costs of risk
reduction for the child are the same in all three periods.7 Evidently, the parent's all-
encompassing concern for the child's well-being together with her monetary transfers enables
her to choose marginal cost of risk reduction values that the child will use for the rest of his life.
In consequence, if rj > 0 and T > 0

tjidUk / dR[) {SRj / dGJpl) MC{	, , .

			— =	-	— =	—	l = a-,b t = 1,2.	(7)

(dUp/dRJp) mUdGl) MCJp

On the other hand, this marginal rate of substitution equates to zero if rj = 0 and will not equate
to the marginal cost ratio if either T = 0 or if the parent does not care about risk to herself.

Together, equations (6) and (7) imply that if rj > 0 and T > 0, and both the child and
parent consume positive quantities of all market goods in all periods when they are alive, then
the parent's marginal rate of substitution between her child's and her own latent risk equals the
ratio of present-value marginal costs of reducing risk in any period. Three further implications
of equations (6) and (7) are that even if the parent is a paternalistic altruist in period t=0 and if rj
> 0 and T > 0: (1) the ratio of marginal risk reduction costs for the child and the parent is not
expected to equal unity because the technologies used to produce perceived risk reduction may
differ and, even if the technologies are the same, levels of perceived risk faced by the two people
may not be the same, (2) for either individual, the ratio of marginal costs for reducing the first
risk need not equal the ratio of marginal costs for reducing the second risk, and thus (3) for either

7 MCJp = (1 + ry'PJ /(dRJp /dGJpl), i = 0.1. and MCJk = (1 + r) /(dRi / dGJb), t = 0,1,2.

9

-------
individual, the marginal rate of substitution between the two types of risks equals the
corresponding ratio of marginal costs in reducing the two risks.

Empirical estimates described in Section 4 test the null hypothesis that the equilibrium
conditions stated in equations (6) and (7) hold. This test is facilitated by considering percentage
risk changes rather than changes in risk by absolute amounts. For instance, when the parent and
child experience the same percentage reduction in a risk, the ratio of marginal products in
equation (6) equals the ratio of initial risk levels, as illustrated below for period t = 0s

(.dRl/dG< ) R

P ~ ~	j = a,b

,>j	J '

'kOJ "'Sfc

(dR>/dG>0) R>

Thus, in this case, as shown in equation (8), the parent's marginal rate of substitution between
equal percentage risk changes for herself and for the child equates to unity.

[(dUp /8Rl) + r,(dUk /8Rl)]R< (dR]p /dGL)/Ri

(dU p /dR3p)RJp	(dRlldGl0)IRl

= 1	j = a,b (8)

If rj > 0 and T > 0, then the corresponding condition will hold for periods t = 1 and t = 2, as
shown in equation (9).

r^dUkldRl)Rl _(dR3/dG3pl)/R3

(dU /dRi)R{ (dR]k /dG]kt)lR]k

= 1 j = a,b t = 1,2 (9)

Evidence that equation (8) holds supports the notion that parents are altruistic toward
their children, but does not indicate whether parents are paternalistic altruists only, whether
parents only exhibit the broader type of altruism associated with rj > 0 and '/' 0, or whether

This outcome also yields a useful corollary for transferring adult morbidity estimates to children when equal
proportionate changes in risk to both groups are considered. If the parent and child experience the same percentage
reduction in risk, the ratio of marginal products in equation (4) equals the ratio of initial risk levels. This means that
the ratio of the parent's willingness to pay to reduce risk to the child to the parent's willingness to pay to protect
herself equates to this ratio of risks. The ratio of actual risks faced might be estimated in some cases using existing
health science and biomedical information. The ratio of perceived risks might be established by studies of parents'
perceived risks to children and to themselves.

10

-------
parents exhibit both types of altruism. Evidence that equation (9) holds, on the other hand, says
nothing about paternalistic altruism, but supports the notion that rj > 0 and T > 0. Evidence
supporting equations (8) and/or (9) does not indicate whether rj or the provisions the parent

makes for the child (Ck0,G^0,T) are large or small. As discussed more fully in Section 4, tests

applied do not distinguish between paternal and all-encompassing altruism and do not identify
the value of rj if rj > 0 9
2.2 Policy implications

The model developed in the previous subsection suggests that effectiveness of
government programs aimed at reducing risk through behavior modification will be
compromised to some extent because they motivate parents to reallocate family resources, as
illustrated by the following three examples.10 First, suppose that in a country composed of M
identical families11, the government initiates an administratively costless program in time period
t = 0 to provide special protection of children from risk, as envisioned by Executive Order
#13045 (Federal Register 1997) in the United States and by similar policies pursued by other
countries (Scapecchi 2006). Assume that: (1) the government has access only to the "family
technology" for risk reduction described by equation (1), (2) the program provides the parent

9	The model presented can be modified or extended in a variety of ways without altering the basic result that the
altruistic parent's marginal rate of substitution between her child's and her own risk equals the ratio of marginal
costs of risk reduction. For example, a discounted expected utility model in which individuals produce risk but
probabilities condition expectations rather than utility itself also implies equality between the parent's marginal rate
of substitution and the ratio of risk-reduction costs.

10	Although the model does not address issues related to government risk information provision or how parents
might respond to such information, it is at least plausible that such programs might be more effective than behavior
modification programs. Also, along these lines, note that if in addition to paternalistic altruism, rj > 0 and T 0.
parental learning about risks will be retained by the child through adulthood in the sense that his marginal costs of
avoiding a risk are equated through all periods of his life. In this situation, parental learning may be passed to future
generations as well, but a formal investigation of this matter would require reformulating the model to allow the
child to have children of his own as, for example, in Becker (1974).

11	Further examples based on heterogeneity of parent incomes, two-parent families, and families with multiple
children easily can be constructed based on those presented below. Similar examples also can be developed for
models where government policy operates by determining the level of an environmental hazard that affects child
and/or parent risk rather than by providing G, although in that case the rate of substitution between G and the
environmental hazard in the risk production functions must be considered.

11

-------
with an extra unit of G earmarked for the child's use, (3) the program is financed by levying a
tax on each parent in the amount of $P, the price per unit of G ,12 and (4) and parents exhibit
one or both types of altruism. As long as prices of market goods and the parent's income remain
unchanged, parents and children in each family end up consuming the same quantities of all
goods as before. In consequence, the program does not alter behavior and has no effect on the
level of risk faced by either person.

Second, suppose instead that the government program sets out to protect everyone (i.e.,
both adults and children) from risk by giving each family one unit of G for either person to use,
rather than earmarking it for the child's use. In this situation, each family simply "purchases"
one unit of G for $ P from the government rather than from the private market. Again, if
incomes and market prices remain unchanged and parents behave altruistically, each family
member consumes the same quantities of C and G as before so that the program has no effect on
behavior or on risk levels faced by either parents or children.

Third, suppose that the government is more efficient than families in lowering risk,
perhaps because of economies of scale in providing risk reduction. In this case, each family
might receive more than one unit of G in return for the tax payment of $ P, thereby
experiencing the equivalent of an increase in income. Pure paternalistic altruists would then
divide the income increase between their own consumption of C and G in periods t = 0 and t = 1
and their child's consumption of these goods in period t = 0, with the increment in G allocated
between the parent and the child so that the parent's marginal rate of substitution between risk to
the child and risk to herself remained equal to the ratio of marginal costs of risk reduction. If in
addition to or instead of paternalistic altruism, parents also exhibit all-encompassing altruism

12 Becker (1981, Chapter 8) presents a closely related example with extended discussion in the context of an income
transfer between an altruistic person and his/her spouse.

12

-------
with rj> 0 and T > 0, more substitution possibilities arise because a portion of the income
increase could be transferred to the child for use later in his life. Thus, while the program could
succeed in lowering risk, the efficiency gain is diffused because both family members now
consume more of all goods in the present period and possibly in future periods.

3. Data and Experimental Design

3.1 Background

Field data were collected from parents of pre-teenage children during summer of 2002
using a self-paced, interactive, computerized instrument.13 An early version of this instrument
was used in a pilot study of parents' willingness to pay to reduce perceived skin cancer risks
(Dickie and Gerking 2003). Two subsequent versions of the instrument were pre-tested and de-
briefing sessions with pre-test participants guided development of the final version. Parents who
participated in this study were residents of the Hattiesburg, MS metropolitan statistical area and
were initially identified by random digit dialing. When calls reached adults, interviewers asked
whether they had at least one biological child between the ages of 3-12 living at home, and
whether they were willing come to the University of Southern Mississippi to participate in a
federally funded study of health risks to parents and their children. Biological children were
singled out for inclusion in the study because skin cancer risk is partly determined by genetic
characteristics inherited from parents (e.g., fairness of skin and sensitivity of skin to sunlight).
Parents were offered a $25 payment for participating in the study.14

13	A more complete description of these data is provided in Dickie and Gerking (2006).

14	Approximately 30% of calls to presumed working residential numbers yielded no contact with an adult after three
attempts at different times of day and days of the week. In 64% of cases in which a call reached an adult, the adult
declared that the household did not meet eligibility requirements (had no biological children aged 3-12 living at
home). Parents agreeing to participate in the study constituted 3.5% of working residential numbers, 5% of contacts
with adults, and 14.3% of contacts with adults who did not declare the household ineligible. Finally, 68% of persons
agreeing to participate completed the instrument.

13

-------
The sample consisted of 610 parents; children did not participate.15 Of the parents, 75%
were white, 20% were African-American, and 5% were members of other races. Data from the
122 African-American parents are not considered further in this paper (but are analyzed in
Dickie and Gerking 2006) because blacks face low levels of risk and therefore have fewer
incentives than whites to think about precautions against solar radiation exposure and how their
own risk might differ from that of their children. Of the 488 non-black parents, 25% were male,
75%) were under the age of 40, mean household income was $60,000 per year, 83%> were
married, and 60%> worked full time. Parents generally were aware of skin cancer: 83%> knew
someone personally who had been diagnosed with this disease, 18%> knew of someone (public
figures, friends, or relatives) who had died from skin cancer, and 82% had considered the
possibility that one of their children might get skin cancer. At an early stage in the interview,
one biological child aged 3-12 of each parent was randomly selected (if there was more than one
in this age range) and designated as the sample child. Questions asked mainly focused on the
parent and the sample child. Half (50.4%) of the sample children were male and the average age
of sample children was 7 years.

3.2. Elicitation of Risk Beliefs

Two types of risk to both parents and children were elicited: (1) the unconditional risk of
getting skin cancer during one's lifetime and (2) the conditional risk of dying from this disease
given that it occurs.16 Parents Slovic, Paul, Baruch Fischhoff, and Sarah Lichtenstein made

15	Responses from 25 parents were disregarded either because they did not answer all questions (21 parents) or
because they did not follow instructions given by the experiment administrator (4 parents).

16	The ability of respondents to understand the risk concepts presented and to clearly distinguish between these two
types of risk was a concern from the beginning of the study because of difficulties people have thinking about
probabilities (Slovic, Fischhoff and Lichtenstein 1985). This concern was amplified for the present study because
few previous surveys have dealt with compound risks. In de-briefing sessions conducted after the pre-tests, the
meaning of the morbidity risk and conditional death risk questions were extensively discussed with participants.
Participants suggested a number of wording changes in the questions, but through this discussion and through their
direct statements, they demonstrated facility with the risk concepts involved.

14

-------
preliminary assessments of lifetime skin cancer risk using an interactive scale similar to that used
by Krupnick el a/. (2002) and Corso, Hammitt, and Graham (2001). The scale, which underwent
a number of design changes based on the pre-tests, depicted 400 squares in 20 rows and 20
columns and all 400 squares were initially colored green. Parents changed green squares to red
ones to represent amounts of risk. Before using the scale to estimate skin cancer risk, parents
practiced using the risk scale for an unrelated event (a possible auto accident) and were told
about the meaning of "chances in 400". Also, they were told to consider only the chances of
getting skin cancer (or of getting it again if they had already had it), rather than how serious the
case might be. Parents then used the risk scale to estimate lifetime chances of getting skin
cancer, for themselves and then for their sample child. Frequency distributions of these
responses presented in Table 1 indicate considerable variation in risk estimates with some
parents believing that skin cancer is highly unlikely and a smaller number of parents believing
that skin cancer is inevitable. Risk estimates tended to pile up at the 5, 10, 15, etc. percent
marks.

As shown in Table 2, parents estimated that their own lifetime risk of getting skin cancer
exceeded that of their sample child (26.9% vs. 22.5%). The null hypothesis that mean perceived
skin cancer risks are equal for parents and children is rejected at the 1% level in a matched-
samples test. This outcome may reflect a number of factors possibly including parents' beliefs
that they take greater precautions to protect their children from skin cancer risk than their parents
did in an earlier period when less was known about the hazards of solar radiation exposure.
Parents also appear to have overestimated skin cancer risk. Ries el al. (1999) found that whites
have a lifetime chance of 21% of getting either melanoma or non-melanoma skin cancer. The
fact that the survey introduced the possibility of getting skin cancer again if the parent had

15

-------
already had it does not appear to be an important complicating factor in this regard. Sample
parents are relatively young and 4.3% reported having been previously diagnosed with this
disease.

Parents were given an opportunity to revise their beliefs about the chances of getting skin
cancer after receiving information about this disease. They were told that: (1) according to the
National Cancer Institute, the average person in the United States has a lifetime risk of getting
skin cancer of 18% and (2) a person's risk may differ from this average because of skin color and
sensitivity to sunlight, family history of skin cancer, amount of time spent in direct sunlight,
experience with sunburns, and use of sun protection products. Parents were questioned about
observable skin characteristics, sun exposure history, and use of sun protection products both for
themselves and their sample children. Over 90% of parents and 97% of children use sun
protection products such as sun lotion. Children use sun protection products a greater fraction of
the time that they are outside and use products with a higher sun protection factor than do their
parents (Table 3). About 40% of parents revised their own lifetime risk estimates, but upward
and downward revisions balanced to yield zero mean revision. Revised risk estimates for
children were on average 2 percentage points lower than initial risk estimates.

To obtain a rough indication of beliefs about latency of skin cancer risks, parents were
asked, "Suppose you do get skin cancer sometime in the future. At what age do you think you
would get it for the first time (or for the next time if you have already had it)?" Responses to this
and a parallel question about the children are summarized in Table 4. About 65% of parents saw
skin cancer as a disease that would strike them or their children at age 50 or later. Based on the
midpoints of the age intervals listed in Table 4, parents on average expected that skin cancer, if it
occurs, would strike them at age 53 or their children at age 55. Comparing expected age at onset

16

-------
to current age, the average implied latency period is 18 years for parents and 48 years for
children, a difference that is significant at the 1% level. These rough measures of perceived
latency suggest that parents see skin cancer as a disease that occurs later in life and see their
children's risk as lying farther in the future than their own.

Parents also provided estimates of mortality risk from skin cancer both for themselves
and for their sample children assuming a doctor had diagnosed this disease. Parents were
unaware that they would be asked about the likelihood of dying from skin cancer when they
answered the previously described questions about getting this disease.17 Parents provided their
perceptions of conditional mortality risk of skin cancer given a diagnosis of this disease using the
previously described risk scale. Table 1 presents the frequency distribution of responses. About
two-thirds of parents believed that their conditional risk of death given a diagnosis of skin cancer
is 10% or less and about three-fourths of parents believed that if similarly diagnosed, their
sample child's conditional risk of death is 10% or less. Many parents felt that the conditional
risk of death is less than 5% both for themselves and for their children. This outcome suggests
that parents were aware that skin cancer is seldom fatal. Parents reported higher mean
conditional death risk estimates for themselves (12.1%) than for their sample children (9.4%), a
significant difference at the 1% level.

3.3 Experimental Design and the Choice Experiment

Parents valued risk reductions by expressing willingness to pay for a hypothetical sun
lotion.18 The product was described using labels (see Figure 1 for an example) designed to look
like those on bottles of over-the-counter sun lotions. Except for differences in the type and

17	Respondents were instructed not to look ahead or to go back to previous questions but rather to see the experiment
administrator if they needed to correct a mistaken answer. Data from 4 respondents who did not comply with this
instruction were among the previously mentioned observations that were deleted.

18	This approach also was used in a recent cross-country study of skin cancer risks (see Brouwer and Bateman
2005).

17

-------
amount of skin cancer protection offered, the labels were identical in all respects to control for
other possible motivations for purchasing sun lotion, such as to prevent sunburn or to get a
suntan and to guard against aging or wrinkling of skin (see Dickie and Gerking 1996). Eight
labels were used in the study: Four labels varied reductions in risk of getting skin cancer
(10%/50% for parent/child) and four labels varied reductions in conditional death risk (10%/50%
for parent/child).19 As demonstrated in Section 2, use of percentage changes simplifies the
econometric tests. Use of percentage changes in risk also has an advantage over presenting
absolute risk reductions in that the post-treatment risk levels always are non-negative.20

Each parent was randomly assigned two of the eight labels and asked for willingness to
pay for each.21 One of the assigned labels offered reduced risk of getting skin cancer and the
other offered reduced conditional death risk from skin cancer. Labels were presented one at a
time in randomized order. After parents were given time to read a label as if considering buying
the product for the first time, they were shown their previously marked risk scales both for
themselves and their children showing the level of perceived risk the parent originally indicated,

19The survey presents exogenous changes in risk to avoid issues that arose in a previous study (Dickie and Gerking
1996) in which risk changes were treated as endogenous. In the earlier work, labels were presented without the
stated risk changes and respondents indicated the amount by which risk would be reduced if the product were used
as directed. Survey participants, however, expressed little confidence in their response to this question and
responses obtained were unavoidably correlated with unobserved participant characteristics. In the present context,
telling parents what to believe about the magnitude of risk change is at least arguably better than asking a difficult
question. Also, random assignment of labels means that risk changes are orthogonal to respondent characteristics.
Nonetheless, because changes in risk actually are endogenous, interpretation of the econometric estimates presented
in the next section must necessarily be guarded.

20	Data on actual purchases of currently marketed sunscreen lotions would not support valuation of the two risks
separately from other motivations for using sunscreen (Dickie and Gerking 1991, 1996) and would not reflect
random assignment of exogenous risk changes. These two features of the field study are critical for estimating the
marginal rate of substitution.

21	Means of the four perceived risks, family income, number of children in the family, and age and gender of parent
and children were compared across labels, separately for the four morbidity labels and four conditional mortality
labels. Statistical tests fail to reject the null of a constant mean across labels at 10% for all characteristics except
gender of parent across the four morbidity labels. With that one exception, the randomly assigned labels are
orthogonal to important parent and child characteristics.

18

-------
and the risk reduction the sun lotion would offer. In this way the magnitude of the risk change
for the parent and the child was described in absolute as well as in percentage terms.

For the first of the two labels, parents were asked, "Now please think about whether you
would buy the new sun protection lotion for yourself or your child. Please do not consider
buying it for anyone else. Suppose that buying enough of the lotion to last you and your child
for one year would cost $X. Of course, if you did buy it, you would have less money for all of
the other things that your family needs. Would you be willing to pay $X for enough of the
sunscreen to last you and your child for one year?" The value of X was randomly selected from
among nine values ranging between $20 and $125. The narrative also reminded parents that
lifetime use of the sun lotion is necessary to obtain the stated skin cancer protection benefits. For
the second label, parents were told, "Suppose that instead of the previous label, we showed you
the following label." Willingness to pay then was elicited as before.

4. Empirical Estimates
4.1. Methods and Interpretation

Following Cameron (1988), the null hypothesis that parents' stated purchase intentions
for the hypothetical sun lotion are consistent with equations (8) and (9) is tested based on a
specification of the willingness-to-pay function rather than on an explicit specification of a
difference in random utility functions. The approach taken uses the model developed in Section
2 to derive present period (t = 0) willingness to pay ( WTP1) for the hypothetical sun lotions to
reduce the unconditional risk of getting skin cancer (j = a) and the conditional risk of dying
from this disease if it is contracted (j = b).

Each new sun lotion is treated as a newly available private good that if purchased would
provide an increment, Sjt, in the planned amount of protective goods that was optimal in the

19

-------
absence of the new sun lotion. If individual i uses sunscreen j during period t then dG]it = Sft = 1;

otherwise dG]t = S]it = 0. The resulting changes in lifetime risk are dlij

Parents participating in the field study were told the lifetime risk reductions that would
result from use of the new sun lotion and that achieving these risk reductions would require
lifetime use of the product. Therefore assume that the parent would prefer not to purchase the
sun lotion for herself now, unless she envisioned continuing to use it in the future. Likewise, she
would prefer not to purchase the sun lotion for her child now, unless she believed that he would
find it in his interest to use it in the future. Also, the first period's supply of the sun lotion is
offered as a single purchase decision for the parent and child together, rather than as a separate
purchase decision for each. In consequence, the parent decides that neither she nor her child will
use the sun lotion at all (S.t = S = 0), or that both will use it now and in the future (S.t = S = 1).

The possibility that only one of the two individuals would use the sun lotion is addressed below.

Suppose that the required expenditure for the lotion for the parent and child together
during t = 0 is denoted X1, and that in subsequent periods, when the child makes his own
allocation decisions, each individual may purchase the sun lotion in an amount for one person at
half of this expenditure, X112. Then the parent's maximal lifetime utility assuming continuing
use of the sun lotion is U*p(yp0 -XJ ,ypl -X112 ,ykl -X112,ykl - XJ / 2,r,Pa ,Pb ;S = 1), where

22 This specification assumes that users of the new sun lotions would not neutralize the risk reductions by making
other substitutions, for example by spending more time outdoors in sunlight. In two previous skin cancer surveys,
attempts were made to account for possible substitutions that might influence endogenously perceived risk changes
associated with hypothetical sun lotions. In Dickie and Gerking (1996), an indicator for whether respondents used
current sunscreen in order to stay outdoors longer was not significantly related to the perceived risk reduction
associated with a hypothetical sun lotion. In Dickie and Gerking (2003), respondents were asked whether using a
hypothetical sun lotion would lead them or their children to spend more time outdoors in sunlight. Fewer than 10%
of parents responded affirmatively, and indicators for this type of substitution were not significantly related to
perceived risk changes associated with the hypothetical sun lotion, or with willingness to pay for it. These results
suggested that the possibility of offsetting substitutions would not be a major factor considered by parents when they
initially evaluated the new sun lotions and consequently no questions concerning this type of behavior were included
in the present study.

20

-------
£/*(•) denotes the indirect utility function and where dU*p / cyk! = 0 if // = 0. Derivatives of this
function include

Of/; / as) = (at/, / SR'r )dR'r + (dur / a/? «+r,(dut/ sr- )dRt

(a{/;/ax') = -[v(i/2)(i+r)-'(V^)+(i/2)(i+rrJH] <10)
=-*M"/2y+rr

t=o V J

where the dRj denote the lifetime risk changes resulting from use of the sun lotion in all periods
and nt denotes the number of users of the sun lotion in period t whom the parent cares about (if
rj > 0, n0 = 2 = , n2 = 1 because the parent cares about the child in all periods, while if
r] = 0, n0 = 2, nx = 1, n2 = 0 because the parent cares about the child only in t = 0). As shown in
equation (10), the child's decision to purchase the sun lotion in periods t = 1 and t = 2 affects the
parent's welfare if rj> 0.

The parent's willingness to pay for the sun lotion per period, WTP1, is the value of
X1 that equates I/ (•) = U, where U denotes the parent's maximal lifetime utility if neither she
nor her child uses the sun lotion. Applying the implicit function theorem to this identity and
using equation (10) implies that marginal willingness to pay for the first period of sun lotion use

is

d(WTPj) = (1/ A )

(dUp/dR;)(dRp

+ [(dUp / 8RI) + r,(dUk / 8RI)]«)

Z / R{))	j = a, b,

In this equation8Jp = ~(dUp /dRp)Rp / Ap and53k - ~[(dUp /cRj.) + i](8f /,. /cRj. )\R£ / Ap denote
the parent's marginal willingness to pay for proportionate reductions in her own and her child's

(11)

21

-------
lifetime risk, and /? =

S(V2)(i+/T

. t=0

-1

denotes the fraction of the present value of total

planned expenditures on the sun lotion that occur in the first period. Because /? < 1, coefficients
of lifetime risk reductions understate the parent's marginal willingness to pay for risk reduction;
i.e., first-period expenditures on sun lotion do not reveal the full willingness to pay for lifetime
risk reduction. Nonetheless, the ratio of coefficients of lifetime risk changes
/?£>/ / pSJp = \(dUp / cRj.) + i](8f I k / cRj: )R{ ]/((dUp / cRJp )Rp) equals the parent's marginal rate of

substitution between equal percentage risk changes for herself and for the child. If the parent is
altruistic, this marginal rate of substitution equals unity.23

For econometric estimation, equation (11) is specified for parent h as

WTPj = ro+K [a;/Rp\h + ri [_K IR]k\+controls,+s]h.	(12)

In equation (12), AJp and A[ are interpreted as the discrete reduction in the /th risk for the parent
and the child that would occur if the sun lotion was used, the Rj denote the last estimate the /th
perceived risk elicited for individual i in the field study, and y] = P^!, i=P,k- Thus the variables
in square brackets denote the percentage risk reductions (divided by 100) shown on the sun
lotion labels for the /th type of risk and take the value 0.1 or 0.5. Treating the y' as constants

implies that willingness-to-pay per unit of risk reduction dWTP1 / dAj = y- / R- decreases with

23 Nonmonetary costs of using the sun lotion such as time costs of ensuring proper application and disutility from
odor or other product attributes are assumed equal for parent and child. The description of the sun lotion attempted
to minimize time requirements by indicating that one application would last all day and to control for potential
sources of disutility such as odor, allergic reactions and blocking of pores. The description was constant across all
labels. To the extent that nonmonetary costs differ between parent and child, however, the costs would be

confounded in the 8- coefficients.

22

-------
the magnitude of perceived risk initially faced.24 Also: (1) controls refers effects on willingness
to pay of measured parental characteristics such as income and family size, and (2) sJh denotes a

random disturbance term with standard properties included to capture unobserved characteristics
of parent h. These characteristics might include willingness to try new products, the ability to
process the information presented on the sun lotion label, evaluation of joint outputs such as
sunburn protection and skin aging, as well as other factors that influence whether the product
would be purchased.

Five aspects of equation (12) warrant further discussion before turning to the results of
estimation. First, altruism implies thatyJk / yJp = 1. But a test of this hypothesis does not
distinguish between types of altruism that may motivate parents' stated intentions to purchase the
sun lotion, because dU / dRJk and // are not separately identified; both are components of yj..

Distinguishing between the types of altruistic motivations considered in Section 2 must await
further research that contrasts parental behavior toward both young and adult children. In any
case, the test does not rest on directly estimating WTP for risk reduction, but instead on
estimating the ratio of estimated contributions of risk reduction to willingness to pay. This
means that y]k and yJp must be consistently estimated, but it is not necessary to obtain a consistent

estimate of y:'.

Second, the percentage risk reduction variables are randomly assigned experimental
design points. Thus, they are orthogonal to other experimental design points as well as to parent

24 In other words, the marginal value of risk reduction dUi / cR diminishes as R.' rises so that y1. remains constant.

To test the adequacy of this specification, which treats willingness to pay as a linear function of percentage risk
changes, separate regressions were run for low-risk and high-risk groups. The null hypothesis that slope coefficients
in both the morbidity and conditional mortality equations are equal in the high and low risk groups was not rejected
at conventional levels. This result occurred whether morbidity risk or conditional mortality risk of the parent or the
child was used to distinguish between low and high risk groups. The test was based on the first specification
reported in Table 5 below.

23

-------
characteristics included in controls and to parent characteristics captured by sJh. This means that
if the functional form of equation (12) is correct: (1) endogeneity problems in estimating
the y. are avoided and (2) estimates of the y. are unaffected by the choice of variables to include
in controls.

Third, willingness to pay for the sun lotion is treated in an errors-in-variables framework
in which stated willingness to pay (W,' ) by parent h to reduce the /th risk differs from true

willingness to pay (W'/'P,') by both systematic and random factors according to

W> = WTPfl +a]h = WTPfl +aJ +vJh, j = a,b.	(13)

In equation (13), a1 is the nonzero mean of aJh and vJh is a random disturbance, a' is assumed to

represent systematic misstatement of true willingness to pay. For example, parents may misstate
willingness to pay because the choice of whether to buy the sun lotion was presented as a
hypothetical question and/or may not have been adequately considered in light of preferences,
and financial constraints.25 Also, vJh captures unobserved parent-specific heterogeneity as well
as purely random factors that may affect a parent's stated willingness to pay for the label
presented. The vJh are assumed to be normally distributed with mean zero and constant variance

and the possibility that E(yahv b„)* 0 motivates joint estimation of willingness-to-pay equations for
the two types of risk.

The marginal rate of substitution (yJk /yJp) is estimated by substituting equation (13) into
equation (12) to obtain

25 As discussed by Carson, Groves, and Machina (2000) the overstatement of purchase intentions arising from
incentive incompatibility of hypothetical, binary discrete-choice questions for private goods is unrelated to the scope
of the good and its costs. Also, joint benefits of the sun lotion are held constant across labels but the parent's
evaluation of any perceived difference between joint outputs of the lotion and existing products would be reflected
in the constant term.

24

-------
wj =(ri +a1) + r1P[_k1pIRiP\h + Y1k[_KIRl\ + controlsh+slh+vl, j = a,b. (14)

Notice that estimators of the constant term (y'A) will be inconsistent if, as expected, a ' ^ 0.

Also, estimators of coefficients of parent characteristics included in controls will be inconsistent
if the controls are correlated with the composite error (a>]h = sJh +v]h). Nevertheless, consistent

estimators of y]k and yJp still can be obtained as long as equation (14) is correctly specified,

because the two risk reduction variables are experimental design points that were assigned
independently of parent characteristics.

Fourth, the dependent variable (stated willingness to pay for a one year's supply of
sun lotion) is latent: Parents only were asked to state whether they would be willing to make a
randomly assigned expenditure. Parents are assumed to answer in the affirmative if W,' > PI,

where PJ denotes the expenditure for a one year supply of sun lotion j that was randomly
assigned to parent h. Thus a parents states that she will purchase the sun lotion if

ioJh / aJ < (f0; + a1) / aJ + (yJp / aJ) [AJp / RJp ] + (y]k / a1) [AJk / R]k ] - (1 / a1 )P>,

where the controls are suppressed for notational simplicity, E(a>Jh) = 0 and var(®^ ) = (a1 )2, and
0)]h is symmetrically distributed. These features together with an assumption of normally
distributed composite errors that have an expected non-zero covariance across equations
E(a^abh) = oah ^ 0 motivates estimation by bivariate probit, where p-oab! oaoh.26 Following
Cameron and James (1987), the coefficient of the randomly assigned sun lotion price is
interpreted as an estimate of -Ha1 that can be used to recover unnormalized coefficients of risk
reductions (ytJ) from the normalized estimates of y/ /oJ.

26 Of course, the assumption of normally distributed errors will not be exactly satisfied when non-normally
distributed parent characteristics (e.g., income) are not included as covariates.

25

-------
Fifth, a concern is that use of stated preference data to estimate the willingness to pay
function will result in a comparatively large variance of the composite error (a>]h = sJh +v]h ).

Stated preference data are often "noisy" and this feature could lead to wide confidence intervals
around the estimated values of marginal rates of substitution, thus making it more likely that the
null hypothesis being tested will not be rejected.

4.2 Results

Full information maximum likelihood bivariate probit estimates are shown in Table 5.27
Sample means of covariates are presented along with the regression estimates. Two pairs of
estimates are reported. The first uses only design points as covariates and the second shows the
outcome when two controls for parent characteristics (family income and number of children in
the family) are added. Two design points measure skin cancer risk changes for the parent and
the child (see equation (14)) and a third measures the randomly assigned sun lotion price. A
fourth design point variable is added to control for the order in which the morbidity and
conditional mortality labels were shown.

Consider first the pair of estimated regressions that use only design points as covariates.
The estimated value of p (=0.778) is positive, as expected, and significantly different from zero,
indicating an efficiency gain from joint estimation of the two equations. The coefficients of the
required annual expenditure are negative and differ significantly from zero at 1%, suggesting that
parents were more reluctant to purchase the sun lotion at higher costs than at lower costs.
Additionally, coefficients of variables measuring percentage reductions in the two types of risk
to both parent and child are positive and significantly different from zero at the 1% level in each

27 Ordinary least squares estimates were used as initial values in computing the binomial probit estimates used as
starting values for the bivariate probit routine. Coefficient estimates and estimates of the marginal rate of
substitution between child and parent risks from the binomial probit estimates are broadly consistent with those
reported in Tables 4 and 5, but are less precisely estimated.

26

-------
of the two equations. This outcome suggests that parents are willing to pay more for larger than
for smaller reductions in the two types of risk and is consistent with the conceptual model
presented in Section 2. Comparing these coefficients to the estimated intercept, however,
appears to suggest that increases in risk reduction do not bring about proportionate increases in
willingness to pay. Many previous studies have found that stated willingness to pay does not
increase proportionately with increases in risk reductions (see Hammitt and Graham 1999 for
further discussion of this issue). Nevertheless, this conclusion may not apply because the
(unnormalized) intercepts actually are estimates of (y;\ +aJ) rather than v:', and a1 > 0 if
parents tend to overstate purchase intentions. Also, as mentioned previously, coefficients
understate willingness to pay for reduced risk because /? < 1. Estimates show that the order in
which the morbidity and conditional mortality labels were presented is unimportant.

When controls for income and family size are introduced, estimates again indicate
positive correlation between the errors in the two equations (0.788). Coefficients of family
income are positive while coefficients of the total number of children in the family are negative
as expected. These coefficients, however, are not consistently estimated if income and family
size are correlated with unobserved family characteristics influencing the sun lotion purchase
decision. Income coefficients are significantly different from zero only at the 10% level under a
two-tail test, suggesting a weak tendency for parents' willingness to pay to increase with income.
The small effect of income may simply reflect the relatively low costs of the sun lotion, with the
highest cost reaching only about $10/month. Coefficients of the number of children are
significant at the 1% level, providing evidence that parents reduce protective expenditures per
family member when more children are present. Because the risk change variables are
orthogonal to these parent characteristics, coefficients and standard errors of risk changes are

27

-------
little altered from their corresponding values discussed previously. Supplementary regressions
(Appendix) specified like those in the last pair of columns but also including covariates for
marital status, education, age and gender of parent, age and gender of child, and whether a close
relative had been diagnosed with skin cancer also demonstrated this same result. Only two of the
additional 14 coefficients differed significantly from zero at 10%.28 Also, in this expanded
regression, coefficients of the risk change variables were almost unchanged as compared with
those presented in Table 5.

Table 6 reports tests of whether the equilibrium condition implied by altruism holds
(jI / ^ —1 = 0, j = a,b). Column (2), Table 6, labeled "full sample," reports results based on

Table 5 estimates that control only for design points. Standard errors are computed using the
delta method. As shown, the null hypothesis that this equilibrium condition holds is not rejected
at conventional significance levels in either the unconditional morbidity or conditional mortality
equations. This null hypothesis also is not rejected using a Wald test of the restriction
Yl I Yp~ 1 = 0 in both equations jointly.

Remaining columns of Table 6 summarize outcomes of parallel tests in six subsamples
defined according to the gender of parent, gender of child, and age of child. Results for
subsamples were obtained by re-estimating the willingness-to-pay equations separately by
subsample using only the four experimental design points as covariates. Parent gender is
considered because the unitary model assumes that families act as if maximizing a single utility
function, so that decisions made by mothers should be consistent with those made by fathers.
Gender and age of child are considered because parental marginal rates of substitution should not

28 The two variables with significant coefficients were parent gender in the morbidity equation and child age in the
conditional mortality equation. Also, in regressions including only experimental design points and the constructed
measures of perceived latency for parents and children, three of the four latency coefficients were negative as
expected, but none was significant.

28

-------
differ between children as long as marginal costs of risk reduction are the same, as in this field
study. As shown in Table 6, results are consistent with the hypothesis yj. / yJp = 1 in all six

subsamples. Furthermore, likelihood ratio tests detect no significant differences in willingness to
pay functions by gender of parent, or by age or gender of child.29

Although not reported in Table 6, a comparable analysis was undertaken based on
subsamples defined by family income, by age and education of parent, and by presence of one
versus more than one child in the family. This analysis is motivated by the assumed constancy
of coefficients of the willingness to pay functions, relative to the possibility that the marginal
utility of income, the /? term, or other parameters may vary with characteristics of the parent.30
Also, the model in Section 2 includes only one child in the family and the survey asked parents
to consider using the sun lotion for only one of their children, even though most parents in the
sample reported having more than one child. However, the null hypothesis that parameters of
willingness to pay functions are equal between families with high or low income, or between
parents with and without college educations, or between older and younger parents, or between
single or multi-child families, is not rejected. Also, the hypothesis y{ I y]p = 1 is not rejected in
any of these additional subsamples.

29	The null hypotheses that slope coefficients of the equations do not differ by gender of parent, or by gender or age
of child, after allowing for different intercepts, were each separately tested using likelihood ratio tests. Results
indicated that the null hypothesis would not be rejected at conventional significance levels in any comparison.
Further analysis of the role of parent gender was conducted by re-estimating the model in the last two columns of
Table 5 while including a dummy variable for parent gender and interactions of this variable and all covariates. The
only statistically significant difference between male and female parents was found in the coefficient of the number
of children in the morbidity equation, where female willingness to pay for the sun lotion declined less than male
willingness to pay with increases in the number of children. Coefficients of risk changes, annual cost and income
appear to be the same for mothers and fathers. Also, outcomes of all of these tests by parent gender are the same if
the comparison is restricted to married parents.

30	A related issue involves whether parents differed in their perceptions of available substitutes for the hypothetical
sun lotion. The survey would have been improved had parents been asked how skin cancer risks could have been
reduced by the amounts shown on the labels if the product were not available or if they chose not to buy it. In the
absence of this information, we assume that either substitution opportunities are negligible or are the same for all
parents.

29

-------
The analysis presented assumes that the parent would use the sun lotion for herself and
her sample child but not for anyone else. The apparent decline in willingness to pay for the sun
lotion with increases in the number of children in the family (Table 5) along with the lack of
significant differences in slope coefficients of willingness to pay functions between single- and
multiple-child families suggests that parents did not envision using the sun lotion to protect
additional children when stating their purchase intentions. Also, parents who indicated that they
would buy the sun lotion were asked about the intended users. The majority of parents indicated
that the lotion would be used for the parent and the sample child (85% for the morbidity labels
and 90% for the conditional mortality labels), with almost all of the remaining purchasers
intending to use the lotion for the child only.31 Excluding parents who envisioned purchasing the
sunscreen but using it for only one individual does not change the outcome of any of these
statistical tests. Additionally, because parents were told that achieving the stated risk reductions
required use of the lotion as directed, the above tests were performed again after adjusting the
risk change measures of Table 5 so that the risk change would be zero for the parent or child if
the parent did not envision that person using the sun lotion. The null hypothesis is not rejected
using these adjusted measures of risk changes.

Finally, empirical results obtained can be used to test another aspect of the model
presented in Section 2. Wald tests are carried out to determine whether the marginal rate of
substitution between the two types of risk for both parent and child are equal to the
corresponding ratio of marginal costs in reducing these risks. This amounts to testing whether
the cross-equation coefficient restrictions /¦' / y!- = 1, i = p,k, are valid. To control for different

values of a1 in the two equations, the tests were conducted using the unnormalized coefficient

31 Four parents who indicated that they would purchase one of the sun lotions envisioned using it for themselves
only (three for the morbidity labels and one for the conditional mortality labels).

30

-------
estimates. Standard errors of ratios of these coefficients were computed using the delta method.
In separate tests involving the coefficients of risk reduction for parents and children, the null
hypothesis is not rejected at conventional significance levels. Additionally, a joint test of the
null hypothesis for parents and children together yields the same result. These results are
consistent with altruism and suggest that parents responded to the assigned changes in the two
types of risk consistently with the theoretical model of Section 2.32
5. Summary and Conclusions

Special protection of young children from environmental hazards has become a
worldwide priority of government policies to improve human health. The fundamental tension
between altruism and self-interest in families looms as the crucial behavioral factor determining
the effectiveness of these policies. This paper estimates parents' marginal rates of substitution
between skin cancer risks faced by 488 parents and their children between the ages of 3 and 12
years. A model of altruistic family behavior that incorporates household production of latent
health risk guides the estimates. The model demonstrates that the marginal rate of substitution
between risks faced by the parent and child is equal to the ratio of marginal risk reduction costs.
Resulting empirical estimates then focus on whether this equality holds.

Tests rest on an examination of stated preference values for a hypothetical sun lotion.
Although stated preference valuation is a controversial method of obtaining willingness to pay to
reduce environmental risks, it supports consistent estimation of parents' marginal rates of
substitution between health risks to themselves and corresponding health risks to their children in
the field study described here. Consistent estimation of marginal rates of substitution is made
possible by: (1) allowing for both systematic and random errors in parents' stated willingness to

32 The outcome of this test reinforces the conclusion that respondents sensibly considered the compound
probabilities involved in the study.

31

-------
pay for the sun lotion and (2) randomly assigning skin cancer risk reductions offered by sun
lotion to the sample of parents. Together, these innovations imply that the skin cancer risk
reductions assigned are orthogonal both to parent characteristics and to errors parents may make
in assessing their willingness to pay for the sun lotion.

In the theoretical model, an altruistic parent's marginal rate of substitution between risk
to her child and risk to herself equates with the corresponding ratios of marginal skin cancer risk
reduction costs. This prediction is the basis of the null hypothesis for econometric tests using
data from the field study. The null hypothesis is not rejected, so test results support the notion
that parents are altruistic toward their young children. This outcome stands in contrast to
findings in related studies that present evidence against altruism of parents toward their children.
This study, however, looks at behavior of parents toward pre-teenage children living at home,
rather than behavior of parents toward their adult children who have formed households of their
own. An important implication is of findings from this study is that effectiveness of public
intervention programs to reduce environmental risks faced by children may be compromised to
some extent because parents will respond by redistributing family resources.

32

-------
References

Agee, Mark D. and Thomas D. Crocker. 2004. "Transferring Measures of Adult Health Benefits
to Children: A Review of Issues and Results," Contemporary Economic Policy 22: 468-482.

Altonji, Joseph, Fumio Hayashi, and Laurence Kotlikoff 1992. "Is the Extended Family
Altruistically Linked? Direct Tests Using Micro Data," American Economic Review 82,
1177-98.

Altonji, Joseph, Fumio Hayashi, and Laurence Kotlikoff. 1997. "Paternalistic Altruism and
Inter- vivos Transfers: Theory and Evidence," Journal of Political Economy 105: 1121-1166.

Barro, Robert J. 1974. "Are Government Bonds Net Wealth?" Journal of Political Economy,
82: 1095-1117.

Becker, Gary S. 1974. "A Theory of Social Interactions," Journal of Political Economy, 82:
1063-1094.

Becker, Gary S. 1981. A Treatise on the Family. Cambridge, MA: Harvard University Press.

Bergstrom, Theodore C. 2003. "Benefit-Cost Analysis and the Entanglements of Love,"
University of California, Santa Barbara manuscript.

Bergstrom, Theodore C. 2006. "Benefit-Cost in a Benevolent Society," American Economic
Review 96: 339-351.

Brouwer, Roy and Ian J. Bateman. 2005. "Benefits Transfer of Willingness to Pay Estimates
and Functions for Health Risk Reductions: A Cross-Country Study," Journal of Health
Economics 24: 591-611.

Cameron, Trudy A. and Michelle D. James. 1987. "Efficicient Estimation Methods for 'Closed-
Ended' Contingent Valuation Surveys," Review of Economics and Statistics 69: 269-276.

Cameron, Trudy A. 1988. "A New Paradigm for Valuing Non-market Goods Using Referendum

33

-------
Data: Maximum Likelihood Estimation by Censored Logistic Regression," Journal of
Environmental Economics and Management 15: 355-379.

Carson, Richard T, Theodore Groves and Mark J. Machina. 2000. "Incentive and Informational
Properties of Preference Questions." University of California, San Diego, Department of
Economics working paper.

Corso, Phaedra S., James K. Hammit and John D. Graham. 2001. "Valuing Mortality-Risk
Reduction: Using Visual Aids To Improve The Validity Of Contingent Valuation," Journal
of Risk and Uncertainty 23: 165-84.

Cox, Donald and M. Rank. 1992. "Inter-vivos Transfers and Intergenerational Exchange,"
Review of Economics and Statistics 74: 305-314.

Dickie, Mark and Shelby Gerking. 1991. "Valuing Reduced Morbidity: A Household
Production Approach," Southern Economic Journal 57: 690-702.

Dickie, Mark and Shelby Gerking. 1996. "Formation of Risk Beliefs, Joint Production and
Willingness to Pay to Avoid Skin Cancer. The Review of Economics and Statistics 78: 451 -
464.

Dickie, Mark and Shelby Gerking. (2003). "Parents' Valuation of Latent Health Risks to their
Children," in Risk and Uncertainty in Natural Resource and Environmental Economics,
(eds.) Justus Wesseler, Hans-Peter Weikard, and Robert D. Weaver, Edward Elgar:
Cheltenham, UK.

Dickie, Mark and Shelby Gerking. (2006). "Valuing Children's Health: Paternalistic

Perspectives," in Economic Valuation of Environmental Health Risks to Children, OECD:
Paris.

34

-------
Dickie, Mark and Victoria L. Messman. 2004. "Paternalistic Altruism and the Value of
Avoiding Acute Illness: Are Kids Worth More than Parents," Journal of Environmental
Economics and Management 48: 1146-1174.

Federal Register. 1997. Executive Order 13045 of April 21, 1997—Protection of Children from
Environmental Health Risks and Safety Risks, 62, FR Doc. 97-10695 (April 23).

Hamilton, William D. 1964. "The Genetical Evolution of Social Behavior, i and ii," Journal of
Theoretical Biology 7: 1-52

Hammitt, James K. and John D. Graham. 1999. "Willingness to Pay for Health Protection:
Inadequate Sensitivity to Probability?" Journal of Risk and Uncertainty 18: 33-62.

Jenkins, Robin R., Nicole Owens, and Lanelle B. Wiggins. 2001. "Valuing Reduced Risks To
Children: The Case Of Bicycle Safety Helmets," Contemporary Economic Policy 19: 397-
408.

Jones-Lee, Michael. 1991. "Altruism and the Value of Other People's Safety," Journal of Risk
and Uncertainty 4: 213-219.

Jones-Lee, Michael. 1992. "Paternalisticistic Altruism and the Value of Statistical Life," The
Economic Journal 102: 80-90.

Krupnick, Alan, Anna Alberini, Maureen Cropper, Nathalie Simon, Bernie O'Brien, Ron
Goeree, and Martin Heintzelman. 2002. Age, Health, and the Willingness to Pay for
Mortality Risk Reductions: A Contingent Valuation Study of Ontario Residents. Journal of
Risk and Uncertainty 24: 161-86.

Laitner, John and F. Thomas Juster. 1996. "New Evidence on Altruism: A Study of TIAA-
CREF Retirees " American Economic Review 86: 893-908.

35

-------
Liu, Jin-Tan, James K. Hammitt, Jung-Der Wang and Jin-Long Liu. 2000. "Mother's
Willingness to Pay for Own and Her Child's Health: A Contingent Valuation Study in
Taiwan," Health Economics 9: 319-26.

Lundberg, Shelly J., Robert A. Pollak, and Terence J. Wales. 1997. "Do Households and Wives
Pool Their Resources? Evidence from the United Kingdom Child Benefit", Journal of
Human Resources, 32, 463-480.

Nastis, Stefanos A. and Thomas D. Crocker. 2003. "Pregnant Mother's Valuation of Own and of
Child Health," presented at USEPA/University of Central Florida conference, "Valuing
Environmental Health Risk Reductions to Children," Oct 20-21,

http://vosemite.epa.gov/EE/epa/eerm.nsf/vwRepNumLookup/EE-Q47570penDocument.

Ries, L A G., M P. Eisner, C.L. Kosary, B.F. Hankey, B.A. Miller, L. Clegg, A. Mariotto, M P.
Fay, E.J. Feuer, B.K. Edwards (eds). 2003. SEER Cancer Statistics Review, 1975-2000,
National Cancer Institute. Bethesda, MD, http://seer.cancer.gov/csr/1975 2000.

Scapecchi, Pascale. 2006. "Valuation Differences between Adults and Children," in Economic

Valuation of Environmental Health Risks to Children, OECD: Paris.

Slovic, Paul, Baruch Fischhoff, and Sarah Lichtenstein, "Regulation of risk: A psychological
perspective," in Roger Noll (ed.), Regulatory Policy and the Social Sciences (Berkeley:
University of California Press, 1985).

U. S. Environmental Protection Agency. 1996. A National Agenda to Protect Children's Health
from Environmental Threats, Washington, D.C.: U.S. Government Printing Office.

36

-------
Table 1. Frequency Distribution of Parents' Perceived Risks.

N=488.



Risk of Getting
Skin Cancer3

Conditional Risk of
Dying from Skin Cancer

Risk Range (%)

Parents

Children

Parents

Children

0-4.75

53

46

78

111

5-9.75

24

48

140

169

10 - 14.75

53

78

112

97

15 - 19.75

55

62

59

40

20 - 24.75

55

59

33

28

25 -29.75

61

63

22

17

30 -34.75

39

32

9

5

35 - 39.75

22

16

7

5

40 - 44.75

33

23

4

5

45 -49.75

6

4

5

1

50 - 54.75

49

29

16

9

55 - 59.75

4

2

1

1

60 - 64.75

5

5

0

0

65 - 69.75

0

1

0

0

70 - 74.75

4

2

2

0

75 - 79.75

6

5

0

0

80 - 84.75

2

3

0

0

85 - 89.75

2

2

0

0

90 - 94.75

9

5

0

0

95-100

6

3

0

0

initial risk assessment.

37

-------
Table 2. Parents' Mean Risk Perceptions (%).

Sample

Risk of Getting
Skin Cancer3

Conditional Risk of
Dying from Skin Cancer

Sample
Size

All Parents

26.93

12.05

488

All Children

22.46

9.36

488

Mothers

29.17

12.46

368

Fathers

20.08

10.82

120

Daughters

22.31

9.38

242

Sons

22.61

9.33

246

Children aged 3 to 7 years

23.84

10.10

275

Children aged 8 to 12 years

20.68

8.39

213

initial risk assessment.

38

-------
Table 3. Use of Sun Protection Products.

Fraction of Time
Outdoors that Sun
Protection Products Used

Parents

Children

Never

44

15

Less than half

115

80

About half

109

106

More than half

91

106

Always/almost always

129

181

Sun Protection Factor
Normally Used

Parents

Children

Less than 15

67

15

15 to less than 30

185

103

30 or higher

192

355

39

-------
Table 4. Frequency Distribution of Expected Age at Onset.

N=488

Age Range (years)

Parents

Children

Before age 40

45

68

40-44

63

42

45-49

64

52

50-54

111

84

55 - 59

61

66

60-64

84

55

65-69

41

46

70-74

13

49

75-79

1

12

Age 80 or later

5

14

Mean age at onset (years)

53

55

Mean age (years)

35

7

Implied mean expected





latency period (years)

18

48

40

-------
Table 5. Willingness to Pay to Reduce Skin Cancer Risks: Bivariate Probit Estimates (N=488).

Sample Mean (Std.

Dev.) or Proportion	Coefficients (Standard Errors)





Conditional



Conditional



Conditional

Covariate

Morbidity Mortality

Morbidity Mortality

Morbidity Mortality

(Parameter Notation)

Risk

Risk

Risk

Risk

Risk

Risk

Parent's Percentage Risk Reduction

0.289

0.302

0.912

0.717

0.901

0.739

(^/cr;)

(0.200)

(0.200)

(0.272)

(0.267)

(0.274)

(0.267)

Child's Percentage Risk Reduction

0.300

0.299

0.854

1.426

0.843

1.487

(rile1)

(0.200)

(0.200)

(0.270)

(0.267)

(0.275)

(0.272)

Cost of Sun Lotion ($/year)

64.518

64.150

-0.011

-0.011

-0.011

-0.011

(-1 la1)

(34.520;

) (34.897)

(0.002)

(0.002)

(0.002)

(0.002)

Order (=1 if risk change in column

0.488

0.512

-0.149

-0.087

-0.151

-0.105

presented last, 0 if first)





(0.122)

(0.122)

(0.126)

(0.125)

Family Income ($10,000/year)



5.957





0.028

0.029





(3.569)





(0.018)

(0.017)

Number of Children in Family



2.078





-0.190

-0.004





(0.952)





(0.069)

(0.068)

Constant





0.733

0.520

0.981

0.347

(Oo +GC])/CJ])





(0.171)

(0.170)

(0.251)

(0.229)

Error Correlation







0.778



0.788

(P)







(0.044)



(0.044)

Log-Likelihood





-

512.553



505.391

41

-------
Table 6. Estimates of y]k / y3 and Altruism Tests.

Estimates of yj. / y3 (Standard Errors) and Tests of Altruism



Full
Sample

Mothers

Fathers

Daughters

Sons

Child Age
3-7

Child Age
8-12

Morbidity ratio

0.936

0.927

0.88

0.902

0.96

1.438

0.441

(raJraP)

(0.415)

(0.456)

(0.678)

(0.777)

(0.503)

(0.766)

(0.462)

z-test ratio= 1 (p)

0.878

0.873

0.860

0.900

0.937

0.568

0.226

onditional Mortality

2.005

1.816

3.746

1.512

3.003

5.018

0.661

ratio (yhk /ybp)

(0.853)

(0.837)

(6.133)

(0.702)

(2.688)

(4.962)

(0.417)

z-test ratio= 1 (p)

0.240

0.329

0.654

0.465

0.456

0.418

0.416

Wald test, both















ratios=l (p)

0.493

0.608

0.883

0.761

0.750

0.601

0.398

Sample Size	488	368	120	242	246	275	213

LR test, equal
parameters between

groups (p)	0.975	0.958	0.214

42

-------
Developed with dermatologists to protect
skin from haimful effects of sun exposure.

Skin Cancer	Ultra

Protection	Waterproof

SPF		Pa iso I® 1789

"Making the outdoors safer
you and youi family."

43

-------
(Back of bottle)

New SkinSaver® sun protection lotion.

VvWvVA-WWv-vVVvVV vVWvV%'vVV-.-\WvVVVX-Vv\W.-V-.-,vWv\W^V\WvWvVWvV%'vW-/v

Six

Skin Cancer Protection

v Used as directed in clinical trials. SkinSaver
reduced risk of skin cancer by:

10% for Adults	10% for Children

i**



v Used as directed in clinical trials, SkinSaver had
no effect on the risk of dying if skin cancer
occurred.

More Skin Protection

WvWVWWW

Parsol®1789	3PF

Protects against
premature skin aging

Protects against
sunburn

More Added Featuies
Ultia long-lasting waterproof formula - One application lasts all day
c Non-com edoaenic—Won't block pores * Oil-free-Won't feel greasy*

VvWvWV\WAArWvVvW	+W.VvW/	•*

	* Hvnoallei tienic * PABA-fiee * Unscented *	

DIRECTIONS: Apply geneiously and evenly to all exposed areas of
shin at least 15 minutes before sun or water exposure.

ACTIVE INGREDIENTS:	2-ethylhexyl	tlPJPMlSfe	1

44

-------
APPENDIX: Supplemental Data and Empirical Results

Table A-l

	Hypothetical Sun Protection Product Labels

Percent Change in	Percent Change in

Morbidity Risk	Mortality Risk

Label

Parent

Child

Parent

Child

A

10

10

0

0

B

10

50

0

0

C

50

10

0

0

D

50

50

0

0

E

0

0

10

10

F

0

0

10

50

G

0

0

50

10

H

0

0

50

50

1

-------
Table A-2. Sample Means by Experimental Design Point.

Morbidity Risk	 	Conditional Mortality Risk

Label

A

B

C

D

E

F

G

H

Percentage risk change for parent

10

10

50

50

10

10

50

50

Percentage risk change for child

10

50

10

50

10

50

10

50

Perceived risk of getting skin cancer for parent

30.26

25.58

26.19

25.44

27.40

25.08

27.59

27.63

Perceived risk of getting skin cancer for child

23.37

22.88

22.18

21.27

23.13

18.90

23.47

24.27

Perceived conditional risk of dying from skin cancer

















for parent

11.89

11.89

12.05

12.41

11.83

10.66

13.21

12.47

Perceived conditional risk of dying from skin cancer

















for child

9.30

8.76

9.85

9.59

8.58

8.73

10.43

9.66

Family Income ($10,000/year)

5.67

6.49

5.99

5.66

6.03

6.00

6.14

5.67

Number of Children in Family

2.10

2.10

2.04

2.07

2.23

1.97

2.05

2.07

Parent is female

0.85

0.78

0.68

0.70

0.78

0.78

0.73

0.74

Child is female

0.45

0.53

0.46

0.55

0.46

0.56

0.52

0.45

Child age

7.18

7.12

7.25

6.72

6.95

7.40

6.86

7.07

Sample Size

130

127

114

117

121

120

124

123

11

-------
Table A-3. Willingness to Pay to Reduce Skin Cancer Risks: Bivariate Probit Estimates (N=488).

Parent's Percentage Risk Reduction
Child's Percentage Risk Reduction
Cost of Sun Lotion ($/year)

Order (=1 if risk change in column
presented last, 0 if first)

Parent Perceived Latency Period

Child Perceived Latency Period

Family Income ($10,000/year)

Number of Children in Family

Parent is Married

Parent is College Graduate

Parent Age

Parent is Female

Child Age

Child is Female

Close Relative of Parent Diagnosed
with Skin Cancer

Constant
Error Correlation

Log-Likelihood	

Mean (s.d.) or
Proportion	Coefficients (Standard Errors)

Cond.



Cond.



Cond.

Morb. Mort.

Morb.

Mort.

Morb.

Mort.

Risk Risk

Risk

Risk

Risk

Risk

0.289 0.302

0.990

0.711

0.918

0.749

(0.200) (0.200)

(0.300)

(0.277)

(0.278)

(0.271)

0.300 0.299

0.849

1.412

0.850

1.384

(0.200) (0.200)

(0.279)

(0.288)

(0.271)

(0.272)

64.518 64.150

-0.011

-0.011

-0.011

-0.011

(34.520) (34.897)

(0.002)

(0.002)

(0.002)

(0.002)

0.488 0.512

0.026

0.024

-0.146

-0.104



(0.022)

(0.021)

(0.123)

(0.123)

18.092





-0.045

-0.077

(9.811)





(0.072)

(0.073)

48.148





0.012

-0.028

(12.239)





(0.060)

(0.059)

5.957

0.026

0.024





(3.569)

(0.022)

(0.021)





2.078

-0.194

-0.022





(0.952)

(0.073)

(0.073)





0.830

0.104

-0.019







(0.182)

(0.188)





0.576

0.081

0.072







(0.138)

(0.137)





35.117

-0.004

-0.004





(6.63)

(0.012)

(0.011)





0.754

0.271

0.161







(0.154)

(0.149)





7.070

0.011

0.051





(2.937)

(0.025)

(0.025)





0.496

0.156

0.061







(0.128)

(0.127)





0.252

0.036

-0.177







(0.150)

(0.158)







0.063

0.520

0.751

0.815



(0.144)

(0.170)

(0.303)

(0.296)

0.791	0.777

(0.0443)	(0.0445)

-500.121	-511.018

111

-------
Is An Ounce of Prevention Worth a Pound of Cure?

Ryan Bosworth and Trudy Ann Cameron
Department of Economics
University of Oregon

and

J R. DeShazo
Department of Public Policy, UCLA*

Corresponding author:

Trudy Ann Cameron

R.F. Mikesell Professor of Environmental and Resource Economics

Department of Economics, 435 PLC

1285 University of Oregon

Eugene, OR 97403-1285

Email: cameron@uoregon.edu

Phone: (541) 346-1242; Fax: (541) 346-1243

* For their helpful comments, we are very grateful to Maureen Cropper and V. Kerry
Smith, as well as to several participants at Camp Resources XII, Wilmington, NC,
August 2004. This research has been supported by the US Environmental Protection
Agency (R829485) and Health Canada (Contract H5431-010041/001/SS). This work has
not yet been formally reviewed by either agency. Any remaining errors are our own.

1

-------
Is An Ounce of Prevention Worth a Pound of Cure?

Abstract

We examine how preferences for prevention and treatment policies vary with
individual characteristics and policy attributes, which include costs to the individual, the
prevalence of the public health problem (numbers of illnesses and deaths), the extent to
which each policy reduces illnesses and deaths, the type of health risk (disease) and, for
prevention policies, the underlying cause and the time horizon for the policy. Individuals
do prefer prevention policies to treatment policies, although at a rate considerably less
than the 16 to 1 ratio implied by the "ounce of prevention..adage. Preferences also
differ substantially by the characteristics of the respondent or policy.

JEL Classifications: 112, J17, J28, Q51

Keywords: Prevention, Treatment, Morbidity, Mortality, Public Health

-------
1 Introduction

Is an ounce of prevention really worth a pound of cure? Some polices can prevent
illnesses by providing a cleaner environment or safer roads. Other policies can allocate resources
to help treat those who are already sick or injured. Should we allocate additional resources to
help those who are already sick, or should we spend more on measures that will help people
avoid illnesses in the first place? Policy makers, at one level or another, must often make these
types of tradeoffs when allocating resources for community health improvements.

Previous economic research that directly examines differences in preferences for
treatment and prevention policies in a utility-theoretic or willingness-to-pay (WTP) framework is
sparse. The existing literature most closely related to our own includes Corso et al. (2002),
Hammit and Liu (2004), and Subramanian and Cropper (2000).

Corso et al. (2002) use survey data to assess preferences for treatment policies and
prevention policies that provide equivalent mortality risk reductions and find that WTP for
treatment policies is much higher. The research reported herein uses a more detailed survey
instrument and can exploit a richer set of data on responded characteristics to better understand
the systematic differences that tend to make prevention policies preferable to treatment policies.

Hammit and Liu (2004) do not address the difference in prevention and treatment policies.
However rather, their research is related to the present study because they investigate in the
impact of latency and disease type on WTP. By considering two different latencies and two
disease types, Hammit and Liu find that WTP for risk prevention declines with latency and that
WTP to avoid a specific cancer risks is moderately larger than WTP to avoid a non-cancer
chronic disease with similarly severe symptoms. We also find evidence that WTP declines with

-------
latency and that WTP is relatively higher for policies that prevent cancer risks. However, we find
that WTP is relatively lower for policies that provide treatment to cancer victims.

Subramanian and Cropper (2000) investigate the relationship between WTP for public
risk reduction policies and the qualitative factors (such as funding fairness or blame for the
health risks) associated with those policies. Relative to their study, we provide a more extensive
analysis of how WTP for both treatment and prevention policies is related to a broader variety of
both quantitative and qualitative factors.

The question of how to value various prevention or treatment policies has been discussed
extensively in the Quality Adjusted Life Years (QALY) literature. For example, Gyrd-Hansen
(2004) finds that how individuals value health increments depends on whether the question is
framed as an individual or social choice. Richardson and Nord (1997) find evidence that
individuals feel that the distributional consequences of health programs are important and should
be included in the evaluation of any health policy. The difficulty of using private choices to
make public policy is clearly illustrated in Ubel et al. 1996). These authors find that subjects in
an experimental setting strongly rejected the health rationing choices derived from their own
utility responses. Finally, Nord (1994) argues that for the QALY to be a generally empirically
meaningful concept, it needs to be interpreted as a measure of social value, rather than of private
value.1

The findings of this paper, related papers, and the QALY literature have important policy
implications. When policy makers allocate resources devoted to public health policies, several
strategies are possible. One strategy is to observe individual preferences about tradeoffs between

1 QALYs are most useful for cost-effectiveness analysis of alterative medical therapies. They focus on physical
measures of health status and involve the standardization of health decrements relative to a year of perfect health
(where death is 0 and perfect health is normalized as 1.) For a brief overview, see Appendix A, available from the
authors.

4

-------
private risk and private income and then use these preferences to make public policy. A policy
maker who relies on estimates of the value of a statistical life (VSL) to make public policy would
be following this strategy. Another strategy would be to directly estimate the demand (or WTP)
for the public health program in question and compare aggregate WTP to the cost of the policy.
Similarly, the policy maker may simply hold a referendum on a proposed project. Researchers
have noted that using private preferences to make public policy may be problematic. For
example, Ubel et al. (1996) provide experimental evidence that in individuals may soundly reject
the public policy implications of their own preferences.

The research reported in this paper focuses on directly estimating the demand for public
health policies. We feel that this strategy has several important advantages. First, individual
preferences for public policies that reduce risk may be very different from the preferences that
individuals have for private risk reductions. For example, individuals may be willing to pay for a
policy that reduces drinking water contaminants because in reduces risk, provides ecosystem
benefits, or because individuals feel that clean drinking water is a "right" that should be enforced
by government. Individuals may also have altruistic motives or different notions of fairness that
influence WTP for public polices. This issue is especially important when considering polices
that affect children, the elderly, or economically disadvantaged groups. Assessing the demand
for public health policies directly also allows us to see how heterogeneity in preferences for
private risk-reducing polices compares with the heterogeneity in preferences for public polices.

Another advantage of our research strategy is that it allows us to assess what attributes of
public health policies individuals view as desirable, as well as how different sociodemographic
groups value different kinds of public health policies.

5

-------
In this study, we use data collected from two analogous surveys of demand for health-
related public policies. These surveys were designed to allow us to compare preferences for
treatment policies with preferences for prevention policies. The surveys were designed by Trudy
Ann Cameron at the University of Oregon Economics Department and J.R. DeShazo at UCLA.
The data were collected, either via computer or Web-TV interface, by Knowledge Networks, Inc.
The hypothetical treatment and prevention policies presented to respondents follow a
randomized design that allows for the investigation of heterogeneity in preferences along several
dimensions. The analysis is also enriched by the availability of a wide variety of individual-level
sociodemographic variables.

Our analysis uses data from two conjoint stated preference surveys of demand for risk
reducing policies that were administered to a nationally representative sample of over 1,500
individuals each. In addition to respondent's answers to the policy questions, we elicited
individual-specific measures of the incidence of the perceived private benefits of each policy as
well as a measure of attitudes toward government intervention.

The basic empirical framework used for analyzing respondents' stated survey choices is
developed in the context of the prevention policy survey in Bosworth, Cameron, and DeShazo
(2005) (hereafter BCD). In the present paper, we develop complementary analyses for the
analogous treatment policy survey and compare the two types of demands.

While both prevention policies and treatment policies can lead to improved community
health outcomes, the presence of systematic differences in consumer preferences for treatment
and prevention policies would indicate that resources could perhaps be allocated more efficiently.
We seek to establish key differences and similarities across policy types and to provide policy

6

-------
makers with improved information about the potential welfare effects of different types of policy
options.

Both the prevention and treatment scenarios presented to respondents vary in elicitation
format, as well as in the specific type of illness threat that is addressed, the number of individuals
who would benefit from the policies, and the size of the affected community. Basic respondent-
level variables include the age, gender, income, education level, and ethnicity of the respondent.
We describe the survey design in detail below.

2 Survey Design

In 2003, we conducted two distinct national stated preference surveys. For each, the
sample size is approximately 1,500.2 The two surveys that provide the data used for this paper
were designed to elicit demands for policies which are publicly financed and which benefit many
individuals (i.e., public goods), rather than privately paid programs with just individual benefits:

(a)	The public "prevention" survey concerns policies that reduce contaminants that cause
illness (i.e., air pollution, water pollution, food safety problems; see Cameron and DeShazo,
2005a). In terms of broader impacts, we intend these prevention policies to be analogous to
the real public policies that lie within the purview of the Environmental Protection Agency
and the Department of Agriculture.

(b)	The public "treatment" survey concerns public provision of remedial medical interventions
to individuals who are ill or injured and which increase their chances of recovery (i.e.
devices, therapies, and procedures; see Cameron and DeShazo, 2005b). In terms of broader
impacts, these treatment policies are intended to mirror the kinds of real public policies

2 A third survey was also conducted, concerning demands for priced programs that benefit only the purchaser (i.e.,
private goods). A large pre-test for this third survey, involving over 1000 respondents, was conducted for Canadian
consumers.

7

-------
achieved, for example, through the regulatory decisions of the Food and Drug
Administration or the medical research funding decisions of the National Institutes of
Health.3

For conformability, the survey instruments for the prevention and treatment studies have
initial and concluding modules that are very similar. Where they differ is in the key policy
choice scenarios. Each policy is described as preventing or treating a named illness or injury.
The illnesses in the prevention survey are attributed to a particular exposure pathway (i.e., air,
water, food). The effectiveness of these policies is described in terms of the numbers of illnesses
prevented (or successfully treated) and the number of deaths prevented. For the individual's
community to enjoy the policy, he or she must pay costs in the form of higher taxes (expressed
both per month and per year). Each of the five choice sets consists of two explicit policies plus
the option to choose neither. We planned these two surveys so that it would be straightforward to
pool their data, to test whether the subsets of corresponding utility parameters are identical, and
to impose common preference parameters as warranted.

Details about the two surveys:

Module 1 (Introduction) - In the first modules of both the prevention and the treatment surveys,
respondents are asked:

(a)	whether they have themselves suffered from each of a range of health threats, or whether
family members or friends have experienced these problems.

(b)	to think about their family health histories, and to assess their own degree of risk from
each of seven major classes of health threats.

3 Of course, much research and development for medicines is also undertaken privately by pharmaceuticals firms,
but these products must still be approved by the FDA.

8

-------
(c)	whether there is room to reduce their health risks by improving their lifestyle or habits,

(d)	whether such improvements would reduce their risks of each of a range of health threats

The treatment survey asks the respondent to rate the likelihood that they would recover from
each of a list of illnesses if they experienced it, given the quality of their current health care plan.

Both surveys then ask respondents to put aside their personal health concerns and to rate
the prevalence of each class of illness and injury in their community (with "community" defined
for them explicitly as a randomly assigned number of people living around them).

Module 2 (Tutorial) - The second module of each survey introduces the ideas of public
prevention policies and public treatment policies, according to the topic of the survey, and begins
to train the respondent how to interpret the summaries of policy attributes that will eventually be
incorporated into compact choice tables. Eight pages of the prevention survey (and eleven pages
of the treatment survey) are devoted to the tutorial process, where the information to be
summarized in each row of the upcoming choice sets is unfolded one row at a time, with careful
and clear explanations. These tutorial pages also include comprehension-testing questions to
confirm that the respondent understands key attributes of the choices.

Module 3 (Choice Scenarios) - The third module of both surveys contains the five-different
stated choice exercises. Each choice, with its preamble and debriefing questions, occupies a set
of four survey pages.

9

-------
In the prevention survey: The complex choice table is preceded by a page that first describes
each policy in words, such as "Policy B reduces types of pesticides in foods that cause adult
leukemia. New growing techniques and standards would reduce food contaminants that cause
leukemia in adults." On the next page, the respondent studies the complete set of attributes of the
two alternatives and makes a choice (which can include selection of "Neither policy"). See
Figure 2 for an example of a prevention policy choice set.

The key attributes of the hypothetical prevention policies presented to respondents in our
prevention survey include the number of cases (illnesses) prevented, the number of deaths
avoided, the duration of the policy and the cost of the policy to the respondent. These prevention
policies also vary in terms of the underlying cause to which the health effects are attributed: (e.g.
an environmental cause (water contaminants, air pollution, pesticides in food); or a non-
environmental cause (traffic accidents)). Prevention policies also vary by the specific type of
illness or injury that is addressed. These include: cancer (general), colon/bladder cancer,
leukemia, asthma, heart disease, heart attack, lung cancer, stroke, respiratory disease, and traffic
accident injuries.

In the treatment survey: Each choice exercise also involves a set of four survey pages. The
first page again describes the two policies in more detail. For example, "Policy A treats children,
adults, and seniors who have leukemia. Those helped will be 25% children, 25% adults, and
50% seniors (i.e. 25/25/50 mix). Then the choice table is presented, containing its summaries of
each program. See Figure 3 for an example of this type of choice set.

The duration of the policy and the policy cost are also key attributes of the policies
described in the treatment survey. However, rather than the number of avoided illnesses or

10

-------
deaths, the treatment policies include the number of increased recoveries as well as the number
of avoided deaths. Treatment policies vary in terms of the demographic group that would most
benefit from the policy (men, women, children, adults, seniors, or some combination of these
groups) as well as the specific health threat addressed. For the treatment policies the list includes:
prostate cancer, breast cancer, colon/bladder cancer, leukemia, lung cancer, asthma, heart attack,
heart disease, stroke, respiratory disease, traffic injuries, and skin cancer.

In both surveys, to follow up on the choice exercise, the respondent is then asked how
difficult it was for them to make up their mind on the previous screen with the choice table, and
is then asked to reply directly to the question "To what extent would each policy directly benefit
you or your family?" This question was asked about each of the two policies in the choice set
just considered. Finally, any respondent who selected the "Neither policy" alternative was given
an opportunity to check which reasons explain their choice. Some of the available answers
constitute reasons that reveal choice-scenario rejection on the part of the respondent (e.g.
disbelief that the policy would achieve what was advertised).4

Module 4 (Follow-up) - This module of each survey asks a number of auxiliary
questions. Among these, the most relevant one for this paper is the question that invites
respondents to both the prevention and treatment surveys to rate how involved they feel their
government should be in regulation environmental, health, and safety hazards.5

The final three pages of each survey instrument are devoted to a hypothetical choice
about how to take some lottery winnings, either as a lump sum now, or as a series of payments

4	These answers can be used to limit the estimating sample, if no other economically admissible reason for choosing
"Neither Policy" is selected.

5	See Cameron and DeShazo (2005a) and (2005b) for more about Module 4.

11

-------
spread out over several years. These choices are used to estimate individual-specific discount
rates. An understanding of these individual discount rates is important to our analysis of the
policy choices in both the prevention and the treatment studies, since the prevention and
treatment policies under consideration in the choice sets in each survey are described as having
different durations.

3 Theoretical Framework

This section explains a simple framework that will allow us to analyze, simultaneously,
preferences for prevention policies and treatment policies. As summarized in Figure 1,
individuals can be in one of three "health" states: healthy, sick, or deceased. Prevention and
treatment policies can both help to decrease the rate of flow from the "sick" to the "deceased"
state. Prevention policies work by decreasing the flow from the "healthy" state to the "sick" state,
while treatment policies work by increasing the flow from the "sick" state to the "healthy" state.
As part of our analysis, we test for systematic differences, across both respondents and health
threats, in individuals' implied preferences over how these flow changes are achieved. We also
provide estimates of the relative values of changes in various flows.

Figure 1

For our most basic specifications, we let the utility of a policy (treatment or prevention)
depend on the number of avoided illnesses (or increased recoveries), the number of avoided
deaths, and the duration of the policy (the length of time the policy is in effect). The individual's

12

-------
income can also be expected to influence utility. Thus, individual Vs indirect utility from policy j
can be represented as:

Vjt = fi(Y, ) + S]f (Avoided Illnessesn)

+ S2g(Avoided Deathsn )	(1)

+ dhiDuration^ )

Where Y, is income, /? i s the marginal utility of income, Si is the marginal utility of an increase in
f 82 is the marginal utility of an increase in g, and S3 is the marginal (dis)utility of an increase in
h6 S3 captures preferences over the length of time the policy lasts, where S3 can be interpreted
as the marginal disutility experienced when the benefits of the policy are spread more thinly
across time.7

We also employ the dummy variable POLj —equal to 0 if the alternative is "neither
policy" (i.e., the status quo) and equal to 1 if it is one of the policy options. The coefficient on
this dummy variable, 9, serves a function similar to that of an intercept shifter in a regression
model. It captures the average effect on person V s utility of all other unobserved factors,
associated with any affirmative policy in the choice set, for which we do not explicitly control in
the random-utility model. 9 merely shifts the entire utility level and is interpreted as the average
effect of unspecified factors on utility of any policy j relative to the status quo (Train 1986, pp.
21-27). Allowing 9 to vary systematically with individual- or policy-specific attributes, as we
will do, increases flexibility in estimation without sacrificing the utility-theoretic foundations of
the model.

6	An obvious objection to this simple linear-in-attributes specification is the implicit assumption that utility is
additively separable in these generic functions of avoided illnesses and deaths. In empirical work documented in
BCD (2005), we relax this assumption by including an interaction term between our functions of avoided illnesses
and avoided deaths. To minimize the complexity of our combined prevention and treatment specification, we omit
the interaction term in this paper. Of course, more elaborate specifications can be entertained.

7	BCD (2005) develops a structural utility-theoretic model with constant exponential discounting employed
explicitly. Since the ad-hoc model presented here provides a better fit, and is more comparable to previous research,
we use it in the empirical portions of this paper.

13

-------
Each choice set consists of two possible policies and a status quo alternative. We employ
a random-utility model that permits analysis with a multiple-conditional logit specification for
econometric estimation. To allow for a range of flexible estimation options, we assume at this
point only that /(0) =0, g(0)=0, h{0)=0 and that f g, and h are increasing in their arguments. The
utility level provided by policy j to individual i is thus:

Vp = P(Ji ~ c;j ) + 8]f(Avoided Illnesses/;)

+ S2g( A voided Deaths^)	(2)

+ 83h(Duration}i) + dPOLj + //,,

Where c.. is the annual cost of the policy and rjJj is the unobserved random component of total

utility. Total indirect utility over the time period of the policy, if the status quo option (neither
policy) is chosen, is given by:

Vm=W) + Vm	(3)

Since we assume thatf(0)=0, g(0)=0, and h(0)=0, it is convenient to normalize on the
level of indirect utility derived under the status quo. The perceived indirect utility difference that
we assume drives the stated choices of our respondents is:

AVp = P(-cij) +8xf (Avoided Illnessesn)

+ S2g(Avoided Deathsjt)	(4)

+ S3h(Durationp ) + dPOLj + s*jt

where the //,, are distributed extreme value and s*t = //,, - //„,.

For this homogeneous-preferences case, it should be noted that the parameters
[i, 8X, 8:, and 8, represent the marginal indirect utilities that individuals associate with the
attributes of the policy while the parameter 0 represents overall increment to utility provided by
any policy, regardless of the other attributes. We can allow each of these marginal utility

14

-------
parameters to vary depending on whether the policy is a treatment or prevention policy and can
identify whether differences in preferences are attributable to different characteristics of the
policies or to a general difference in the value placed on any type of policy, independent of its
particular attributes.

There is no a priori expectation that the error dispersion in the choice model for
prevention policies will be identical to the error dispersion in the choice model for treatment
policies, although this is a testable hypothesis. If we wish to test the equivalence of the marginal
utility parameters across the two samples, it will be necessary to allow for distinct error variances.
(See Cameron et al. (2002)). We scale the level of indirect utility for the prevention policies and
treatment policies by Kp and Kt, respectively. Let 1 {Treatment) be a dummy variable equal to 1

for treatment policy choices and equal to 0 for prevention policy choices. Thus, the indirect
utility differences for the prevention policies and treatment policies are:

I	Pv	d\v

A Vtj	= — (—c ) H—- /(Avoided Illnesses,)

JI \\(Treatment)=0 JC	K

5'

H		 g(A voided Deaths )	(5 )

kp

5,	0	s

H—- h(Duration. ) +—POL + —

AT	AT



= — (-c ) +—/(Avoided Illnesses )

\(Treatment)=\ K. K.

+ ^Lg(Avoided Deaths )	(6)

Kt

+—h(Durationi) +—POL + —
K,	K,	K,

15

-------
Given that the scale of utility is arbitrary, we normalize by assuming Kp = 1 for the
prevention data set. The parameter Kt is freely estimated and is interpreted as the ratio of

dispersion of the unobserved portion of utility in the treatment sample to the dispersion of the
unobserved portion of utility in the prevention sample.

We wish to investigate whether the parameters [ip, S1 , S2 , S3p, and 9p are

systematically different from /?,, Su, S2l, S3t, and 0t , based on inferences from respondents'

choices among prevention policies and among treatment policies (from separate samples). By
introducing the dummy variable l(Treatment) as a shifter we can permit the marginal utilities of
avoided illness, avoided deaths, and policy duration vary systematically. With the pooled sample,
we can test for statistically significant differences in preferences. Of course, we can (and do)
allow all parameters (including k ) to vary with other individual- or policy-specific
characteristics as well.

3.1 Willingness to Pay

In the deterministic case, formulas for total WTP and marginal WTP are straightforward.
Point estimates of total WTP can be calculated by solving for the annual payment that would
make the individual just indifferent between (a) paying for the policy and receiving the benefits,
and (b) not paying for the policy and not receiving the benefits. Suppose we ignore the
symmetric and mean zero error term and the variance-covariance matrix for the maximum
likelihood estimates of the unknown preference parameters. We can set the utility difference in
equation (4) equal to zero and solve for c*. in terms of the parameter point estimates and the data.
Total WTP for policy j is thus:

16

-------
8xf (Avoided Illnessesn) + S2g(Avoided Deaths;. ) + 83h(Duration;. ) + 6P0Lj

T

c* _	v	J' ' 2o V	ji ! 3 V	JU	J_

Marginal WTP (MWTP)--WTP for incremental changes in one of the attributes of the
policy—is calculated by taking the derivative of total WTP with respect to that attribute. For
example, MWTP for one-unit increase in g(Avoided Deaths ) is simply:

dc,.

p

	= —	(8)

dg(Avoided Deathsp) (5

In our empirical work, we use a shifted log specification for the functions/ g, and h. The
MWTP in equation (8) above is therefore roughly interpreted as marginal willingness to pay for
a 1% increase in avoided deaths.8

4 Results

There are five different threads to our empirical results. Section 3.1 discusses our most basic
specifications; Section 3.2 describes effects related to the size of the affected population. Section
3.3 details results relating to the socio-economic status of the respondent. Section 3.4 discusses
differences in preferences for cancer and non-cancer policies. Finally, section 3.5 reports the
results of models designed to investigate preference heterogeneity according to policy attributes.9

4.1 Basic Specifications

For our most basic specifications, reported in Table 2, we follow the standard practice in the
choice literature and specify a utility function that is linear and additively separable in some

8	Crude confidence bounds of fitted WTP and MWTP, reflecting estimation precision, can always be calculated by
sampling from the joint (asymptotically normal) distribution of the maximum likelihood parameters and building up
a sampling distribution for each calculated quantity. Of course, since total WTP is a function of policy attributes (see
equation (7)), this sampling distribution will differ across policies.

9	In a separate paper using only the "prevention" survey (BCD 2005), we submit our inferences to numerous
robustness and validity checks. We also assess scope effects, order effects, sample selection biases and, through our
survey design, attempt to mitigate hypothetical bias associated with incentive incompatibility.

17

-------
function of the fundamental attributes of the policy. These fundamental attributes include the
number of avoided illnesses or increased recoveries, the number of avoided deaths and the
duration of the policy. In models where we pool the data from the two samples, we test whether
or not the marginal utilities associated with illnesses, deaths, and the duration of the policy can
be constrained to be the same across treatment and prevention policies by allowing the
coefficients on these fundamental attributes (as well as the generic policy dummy) to vary
systematically with a dummy variable that is equal to one if the policy is a treatment policy. We
also allow the error variance to differ across policy type.10,1 U2

In Table 2, the negative and statistically significant coefficient on the interaction term
between the Log {Death Reductions) variable and the treatment dummy, in models 3 and 4,
suggests that respondents place a higher value on deaths avoided via prevention policies than
treatment policies. However, the other utility parameters (including the marginal utility of
avoided illnesses) are not statistically different for treatment and prevention policies.

The estimates for our parsimonious model 4 in Table 2 imply that (yearly) marginal WTP for
a 1% increase in avoided deaths via a prevention policy is about $238, compared to about $142
for the same 1% increase in avoided deaths via a treatment policy. These estimates are,
unsurprisingly, almost identical to the estimates obtained for models 1 and 2 for the two separate
samples. The estimates in model 1 (the prevention sample) indicate MWTP of $245 for a 1%
increase in avoided deaths, while model 2 (the treatment sample) indicates MWTP of $138.

10	For estimation purposes, we constrain K to be positive by estimating the logarithm of this parameter. The
estimates of K in the tables below need to be exponentiated to conform to the model presented in section 2.

11	Models that pool the prevention and treatment samples (or allow for heteroscedasticity in other contexts) are
estimated via a heteroscedastic conditional logit optimization routine programmed by the authors using the software
package Matlab. Models that do not involve heteroscedasticity are estimated using the packaged conditional logit
routine in Stata software. The Matlab code is validated for special cases where either type of software can be used.

12	We use a shifted log specification to maintain the assumption that f(0)=0 and g(0)=0. For example, we use
log(Avoided Illnesses+1) rather than log(Avoided Illnesses)

18

-------
MWTP estimates for a 1% increase in avoided illnesses, in all models in Table 2, are about $70.
We also note that the variance of the errors associated with choices concerning treatment policies
appears to be larger by a factor of about e°'3639=1.43 9.

Loosely speaking, the question posed in the title of this paper may be answered as follows:
In terms of the estimated marginal utility avoided deaths, an "ounce" of prevention appears to be
worth only about two "ounces" of cure in the sense that individuals appear to be willing to pay
about twice as much for an incremental (1%) improvement in the number of deaths avoided via a
prevention policy than they are for the same incremental improvement via a treatment policy.
However, it should be noted that this difference applies only to avoided deaths. This result stands
in contrast to that of Corso et al. (2002) who find that respondents to general allocation questions
are willing to pay much more for treatment programs than for prevention programs.

There appears to be no statistical difference between the marginal amount individuals are
willing to pay to avoid illnesses and the amount they are willing to pay to increase the number of
recoveries. In terms of the diagram in Figure 1, individuals are willing to pay about twice as
much to decrease the flow from the "Sick" state to the "Deceased" state via prevention policies
as they are willing to pay to achieve the same net result via treatment policies. However, they are
willing to pay about the same amount to increase the flow from the "Sick" state to the "Healthy"
state as they are to decrease the flow from the "Healthy" state to the "Sick" state.

We also note that in all models in Table 2, the estimated marginal utility associated the
duration of the policy is statistically significantly negative, indicating that individuals generally
prefer policy of shorter duration (holding the total number of avoided illnesses and deaths
constant). This result is consistent with positive discounting and is similar to that found by
Alberini et al. (2004) who find that WTP to reduce mortality risk declines with latency. Ariely

19

-------
and Lowenstien (2000) show that in most cases individuals underweight the importance policy
latency, but that models (such as the research reported herein) that carefully and explicitly
describe policy attributes can increase the likelihood that respondents discount future benefits.
Cropper, et al. (1994) also finds that individuals generally values lives saved in the future less
than lives saved in the present.

4.2 Population Size Effects

While Table 2 provides an initial answer to the question that headlines this paper, there are
a number of additional insights that can be gleaned from more-general models built upon the
same framework. For example, most previous studies that ask respondents to value policies that
save a given number of lives do not ask respondents to consider the population size of the
affected community. The policies presented to respondents in our study offer potential reductions
in the number of illnesses and/or deaths in the community where the respondent lives. All
policies considered by a given respondent are described as affecting "their community", where
the community is described as a specified number of people living around the respondent. This
asserted community size is varied randomly across respondents. The results related to population
size heterogeneity reported in Table 3 do not exhibit the strong statistical significance
characteristic of the other results reported herein. This is unsurprising, however, given the fact
that population size is not a line-item attribute in the survey and that variation in asserted
population size occurs only across respondents, rather than policies or choice sets.

We endeavor in this study to evaluate willingness-to-pay for health improvements from a
public-goods perspective, so we must consider the size of the affected population. Note that this
issue does not arise in studies that attempt to estimate only private trade-offs between health and

20

-------
income. To make this issue clear, consider a simple example: Suppose that the leaders of a
community of 100,000 people are considering two policies (policy A and policy B) that are each
expected to reduce the number of deaths in the community. Policy A is expected to save the lives
of 100 individuals in the community by reducing the risk level of each individual in the
community. Policy B is also expected to save 100 lives, but the risk reductions from Policy B
will accrue only to the inhabitants of the western side of the community, (i.e. only 50,000 people
will see their risk level reduced.)

Even if the two policies cost the same, there is no a priori reason why a person (or a
community) should be indifferent between the two programs. The decision maker would be
choosing between providing a relatively large risk reduction to a smaller number of people, or
providing a relatively small risk reduction to a larger group of people. In our example, we would
expect that the 50,000 people on the west side of the community would prefer policy B to policy
A (if they are selfish), while the rest of a selfish population would presumably prefer policy A to
policy B. However, individual notions of fairness or altruistic preferences may lead to valuations
that differ from the purely selfish outcome. If individuals are concerned only about their own
health and income, we might expect that they would be willing to pay more for a program that
affects a smaller population, ceteris paribus.

Table 3 thus presents the results of more-general models that allow the estimated utility
parameters in our model to vary with the size of the population that will be affected by the policy.
Models 1 and 2 again show results for analogous models estimated on our separate samples. In
the less-restrictive pooled specification of model 3, we constrain the basic utility parameters (the
ones that Table 2 suggests can be constrained) to be the same across treatment and prevention
policies, but allow them to vary systematically with the size of the affected population. In the

21

-------
prevention sample, the marginal utility of avoiding illnesses appears to be lower when the
population size is larger. We also note that the coefficient on the interaction with population size
and the policy dummy appears to be statistically significantly negative in models 1 and 4,
indicating that individuals are less likely to choose either offered policy over the status quo
option when the affected population size is larger. Both of these effects suggest selfish behavior.
This result is consistent with the idea that social discount rates may be smaller than private
discount rates: a larger population size probably causes the specified health improvements to be
viewed as less of a private good and more of a public good. In other words, if it is not your life
that is saved, it doesn't matter as much when that life is saved. However, this tendency is not
apparent in the treatment sample.

4.3 Sociodemographic Effects

We now investigate how preferences for prevention policies appear to differ from
preferences for treatment policies according to the sociodemographic characteristics of the
respondent. Tables 4, 5, and 6 present these results. To keep the dimensionality of the parameter
space manageable, we allow only the coefficient on the policy dummy to vary with the
sociodemographic variables. Recall that the coefficient on the policy dummy captures how
individuals feel about any policy, relative to the no-policy status quo. Coefficients on the
interaction terms in Table 4 can be roughly interpreted as capturing the effect of change in a
given variable on the latent propensity to choose either of the two offered policy options over the
status quo.

Previous work that investigates how WTP for health benefits varies with the
sociodemographic characteristics of the individual includes Alberini et al. (2002), and DeShazo

22

-------
and Cameron (2005), Alberini et al. (2002) find that WTP declines with age, but only after age
70. DeShazo and Cameron (2005), however, find that WTP follows an inverted U-shaped profile.
Kartman et al. (1996) find that income is positively related to WTP to reduce the risk of angina
pectoris attacks. It should be noted, however, that these authors investigate WTP for private risk
reductions rather than the public choices considered here.

The treatment sample indicates lower WTP for females. This may reflect lower incomes
and higher marginal utility of income for women (not estimated here) or may reflect different
propensities to avail themselves of private health care services and diagnostic procedures. In fact,
the magnitude of this effect is relatively large in terms of estimated WTP. The estimates in
model 2 in Table 4 indicate that females are willing to pay about $248 less per year for a typical
treatment policy than males. However, there is no indication that females are less likely than
males to choose either prevention policy over the status quo in the prevention sample.

In contrast to the inverted U-shaped age profile found by DeShazo and Cameron (2005),
the separate prevention and treatment samples (as well as the pooled model) indicate that
willingness to pay has a U-shaped age profile. This profile reaches an estimated minimum at
about age 60 in the prevention sample and at about age 71 in the treatment sample. The curvature
of the age profile is also significantly less for the treatment sample.

The models in models 1 and 2 indicate that the income of the respondent, as a proxy for
general socioeconomic status, has opposite effects on the estimated WTP of the respondent in the
two samples. In particular, higher income individuals are less willing to pay for prevention
policies, while they are more willing to pay for treatment policies. These effects, while
statistically significant, are relatively small in magnitude: the estimates in model 1 indicate that a
$10,000 increase in annual income is associated with an estimated decrease of $0.34 in annual

23

-------
WTP for a typical prevention policy. Similarly, the estimates in model 2 indicate that a $10,000
increase in annual income is associated with an estimated increase of $0.44 in annual WTP for a
typical treatment policy. One plausible explanation for this difference may be the availability of
substitutes. Recall that the prevention policies work in one of four ways: cleaner air, cleaner
water, fewer pesticides in food, and safer roads. A high income individual can more easily move
to a cleaner location, drink bottled or filtered water, eat organic produce, and purchase safer
automobiles. However, there are relatively fewer substitutes for prevention policies that lower-
income people can exploit.

The results for years of education suggest that more highly educated people are more
likely to support prevention policies, but not treatment policies. The estimates in model 1 suggest
that, for prevention policies, one additional year of education is associated with an estimated
increase of $128 in annual WTP for a typical prevention policy.

Non-white individuals are generally more likely to support both prevention and treatment
public policies over the status quo, and more likely to support treatment than prevention policies.
The estimates from models 1 and 2 suggest that non-white individuals are willing to pay an
estimated $299 per-year more than non-white individuals for prevention policies and $660 more
for treatment policies.

Attitudes toward government intervention have a lot to do with individuals' receptivity to
publicly supported health policies. Tables 5 and 6 present separate results for the prevention and
treatment samples, and demonstrate the impact of the additional variable Government Preference.
After the choice scenarios, we presented individuals with the following question: "People have
different ideas about what their government should be doing. How involved do you feel the
government should be in regulating environmental, health and safety hazards?" Individuals were

24

-------
invited to indicate their preferred level of government involvement along a continuum ranging
from minimally involved (0) to heavily involved (7). While this variable is merely ordinal, we
limit the complexity of our estimating specification by treating it as an approximately continuous
variable.

The results in Tables 5 and 6 make the statistical importance of this (endogenous)
variable clear. The maximized value of the log-likelihood function is higher (in both the
treatment and prevention samples) when the variable Government Preference is included as a
single shifter on the 6 parameter (model 2) than when the entire suite of other sociodemographic
variables is included (model 1). Moreover, the third models of Tables 5 and 6 demonstrate that
there is almost no impact on the statistical significance of the other sociodemographic variables
when the Government Preference variable is included.

We conclude from this analysis that although the basic sociodemographic characteristics
of the respondent are important in determining choices, an individual's perception of the proper
role of government is relatively more predictive of their stated choices across proposed public
policies.

4.4 Cancer vs. Non-Cancer Policies

Previous research (e.g. Hammitt and Liu (2004)) has suggested that individuals may be
willing to pay more to reduce cancer risks than non-cancer risks, independent of the severity of
the symptoms of either type of disease.13 Cancers may simply instill greater fear than other
diseases. We address this interesting question by assessing whether or not individuals are
(broadly speaking) more likely to support policies that address cancer risks than other non-cancer

13 Other researchers, including Tsuge et al. (2005), and Magat et al. (1996) find that it may not be necessary to adjust
VSL estimates for cancer.

25

-------
risks. Our prevention policies survey asks about several different diseases, including cancers (in
general) lung cancer, colon/bladder cancer, and leukemia. The treatment policy sample is asked
about colon/bladder cancer, leukemia, lung cancer, prostate cancer, breast cancer and skin cancer.
We define an indicator variable, "Cancer," to be equal to 1 if the policy provides prevention or
treatment with respect to a major cancer.14 Table 7 presents results that utilize this variable to
differentiate between preferences for cancer versus non-cancer policies.

The results in Table 7 suggest that individuals are more likely to support a cancer
prevention policy than other types of policies, but less likely to support cancer treatment policies.
Since many types of cancers are still viewed as incurable, these findings seem plausible.

WTP calculations based on the estimates in Table 7 suggest that, in the prevention
sample, individuals are willing to pay about $310 more (annually) for a policy that avoids deaths
and illnesses from a major cancer than from other (non-cancer) illnesses or injuries. In the
treatment sample, however, estimates suggest that individuals are willing to pay about $130 less
per year for policies that address major cancer risks than for other types of policies. The pooled
sample provides similar estimates: increase WTP of about $280 for prevention of cancer, but
about $160 less for the treatment of cancer.

4.5 Heterogeneity by Policy Attributes

Tables 8 and 9 present parameter estimates for models that allow the coefficient on the
policy dummy to vary systematically with additional attributes of each policy. We find evidence
that individuals have statistically distinguishable preferences for some types of policies.

14 We exclude skin cancer from the list of "major" cancers because skin cancer is generally perceived as a less
serious health threat than the other cancers considered in our survey.

26

-------
Previous Research: Authors that investigate how WTP for public health benefits varies
with the source of the risk include Vassanadumrongdee and Matsuoka (2005), Carlsson et al.
(2004), and Chilton et al. (2002). Vassanadumrongdee and Matsuoka (2005) find that WTP for
reductions in the risk of disease from air pollution and the risk of traffic accident are comparable
while Chilton et al. (2002) find that the perception of risk influences WTP values for reducing
the risk of rail accidents. Carlsson et al. (2004) find that individuals in their sample are willing to
pay more to improve air travel safety than taxi travel safety. Subramanian and Cropper (2000)
find that the number of lives saved, as well as psychological risk characteristics are important
determinants of allocation decisions, and Krupnick and Cropper (1992) find that individuals who
have had friends or family with chronic lung disease are willing to pay more to reduce the risk of
chronic lung disease. Jacobsson et al. (2005) find that the altruistic component of WTP is greater
for more severe diseases. Wittenberg et al. (2003) find that their respondents "were 10 to 17
times more likely to allocate liver transplants or asthma treatment to patients they deemed not
responsible for their illnesses than to patients they deemed responsible for their conditions".

This study appears to represent the most comprehensive comparative analysis of
systematic variation in WTP for public health policies to date. The policies in our survey vary, as
reported above, in terms of basic attributes such as the number of lives saved of the cost of the
policy. The policies presented to respondents also vary in terms of the source of the risk, the
disease or health threat that is addressed, and the population sub-group that is affected, allowing
respondents to consider a wide range of substitute policies when making allocation decisions. In
Tables 8 and 9, we report the results of models that allow key utility parameters to vary with a
variety of additional policy attributes.

27

-------
Table 8 allows the coefficient on the policy dummy to vary by the type of disease. We
have chosen heart disease as the baseline disease (the omitted category) because it is one of the
most common causes of death and there are effective methods for both the prevention and
treatment of heart disease. The estimated coefficients on the policy-dummy interaction terms in
Table 6 are interpreted relative to the base case of heart disease.15

In the prevention sample, we see that individuals are more likely to support public
policies that prevent cancer (general), leukemia in children, and asthma in children than they are
to support heart disease policies. However, individuals are less likely to support prevention
policies that address leukemia in general, stroke, asthma in general, and traffic injuries.

In the treatment sample, there are no public policies that are statistically significantly
more likely to be supported than those for heart disease. However, public policies to reduce
colon/bladder cancer, leukemia, stroke, respiratory disease, asthma, asthma in children, lung
cancer, injuries, prostate cancer, and skin cancer are all less likely to be chosen relative to public
heart disease treatment policies.

Finally, Table 9 shows results for models that utilize additional sources of heterogeneity
in the type of health risk that are unique to either the prevention or the treatment surveys. As in
the choice set example in Figure 2, the scenarios concerning the public prevention policies vary
in terms of the underlying cause of the particular illness or injury. These causes include: air
pollution, drinking water contaminants, pesticides in foods, and traffic accidents. The
"prevention" model in Table 9 suggests that individuals prefer prevention policies that reduce air
pollution, drinking water contaminants and pesticides in foods, (relative to policies that reduce
the likelihood of injuries via traffic accidents).

15 We report in Tables 8 just the prevention policy and treatment policy results, separately. The pooled model has a
very large parameter space and offers few additional insights.

28

-------
The "treatment" model in Table 9 presents results for a specification which reflects the
fact that some treatment policies are targeted at specific socio-demographic groups. For example,
breast cancer treatment policies primarily benefit women and prostate cancer treatment policies
primarily benefit men. The results in the "treatment" model of Table 9 suggest, perhaps
unsurprisingly, that females are statistically significantly more likely than males to support breast
cancer treatment policies, while they are less likely than males to support prostate cancer policies.

Other types of policies may be targeted primarily at children, adults, or seniors. Notice in
the choice set example in 'Figure 3 that the choice scenarios presented to respondents make the
targeted beneficiary group explicit. When policies are designed to benefit more than one group,
the percentages of the benefits accruing to each group are included explicitly in the description
of the choice. For example, policy A in the Figure 3 (the treatment choice set example) treats
children, adults, and seniors who have leukemia. The percentage mix is given as 25/25/50,
indicating that 25% of the benefits would accrue to children, 25% to adults, and 50% to
seniors.16 We construct the continuous variables Percent Children and Percent Senior and allow
the coefficient on the policy dummy to vary systematically with these variables. We also utilize
the variables "Fewa/e", "Age65+", and "Kids" to distinguish how preferences differ for
treatment policies that affect particular groups. "Female" is equal to 1 if the respondent is female
(and 0 otherwise), Likewise, "Age65+" is equal to 1 if the respondent is age 65 or older, and
"Kids" is equal to 1 if the respondent lives in a household with any children under the age of 18.

The results in Table 9 suggest that respondents with children in the household are
statistically significantly more likely to support policies that benefit children while policies that

16 The tutorial portion of the treatment survey, which precedes these choice scenarios, explains the interpretation of
these proportions.

29

-------
benefit seniors are less likely to be supported. Interestingly, this apparent lack of support for
policies that benefit seniors is also shared by seniors themselves.

5 Conclusion

Policymakers face many tradeoffs when allocating funds for public risk reduction and
health improvement policies. Some policies can help prevent adverse health states while other
policies can allocate resources to help treat those who are already sick or injured. We find that
preferences for prevention and treatment policies differ in several important ways.

Individuals appear to have a preference for prevention policies over treatment policies.
This preference appears to be driven by a higher marginal value placed on lives saved via
prevention policies. We find that individuals are willing to pay about twice as much to avoid
deaths via prevention than they are to avoid deaths via treatment.

We also find that the size of the affected population affects preferences for both treatment
and prevention policies in ways that are generally consistent with selfish behavior. In particular,
individuals are less likely to support prevention policies when the affected population size is
larger. The size of the affected population has a much less pronounced effect on preferences for
treatment policies.

We find evidence of significant heterogeneity in WTP for prevention and treatment
policies according to differences in socio-demographic characteristics. We find that WTP has a
U-shaped age profile that reaches a minimum at about age 60 for both types of policies. High
income individuals are more likely to support treatment policies, while they are less likely to
support prevention policies. This seemingly strange result may be the result of a wider array of
preventative/risk mitigating options available to wealthier individuals.

30

-------
We also note that more highly educated people are more likely to support prevention
policies than less educated people, while there is no systematic heterogeneity in preferences for
treatment policies by education level. Females are less likely to support treatment policies, while
non-white (non-Caucasian) individuals are more likely to support both prevention and treatment
policies.

Respondents in our sample are more likely to support prevention policies that address
cancer risks than non-cancer risks, but are less likely to support major cancer treatment policies
than policies that treat other major illnesses or injuries. Respondents in both samples are less
likely to support policies that address stroke, leukemia, and asthma than policies that address
heart disease. We also find that females are more likely to support breast cancer treatment
policies and less likely to support prostate cancer treatment policies. Individuals with children
are more likely to support policies that benefit children, but seniors are not more likely to support
policies that benefit seniors.

We identify several areas of heterogeneity in preferences by individual and policy
attributes and find that respondents are more likely to choose policies that directly affect
themselves and/or their family members. We also find that individual perceptions of the proper
role of government are significant in explaining whether or not individuals support policy
changes over the status quo.

31

-------
References

Alberini, A., Cropper, M., Krupnick, A., Simon, N. (2002). Does the value of a statistical life

vary with age and health status? Evidence from the United States and Canada. Resources
for the future discussion paper 02-19. Washington, D.C.

Alberini, A., Cropper, M., Krupnick, A., Simon, N. (2004). Willingness to pay for mortality risk
reductions: Does latency matter? Resources for the future discussion paper 04-13.
Washington, D.C.

Andreoni, J. (1990). "Impure altruism and donations to public-goods - a theory of warm-glow
giving", Economic Journal 100, 464-477.

Arana, J.E., Leon, C.J. (2002). "Willingness to pay for health risk reduction in the context of
altruismHealth Economics 11, 623-635.

Ariely, D., Loewenstein, G. (2000). "When does duration matter in judgment and decision
making?" Journal of Experimental Psychology-General 129, 508-523.

Bosworth, R.C., Cameron, T.A., DeShazo, J.R. (2005). Advances in evaluating the demand for

risk prevention policies. University of Oregon, Department of Economics Working Paper.
Eugene, OR.

Cameron, Trudy Ann, and J.R. DeShazo (2005a) "Demand for Public Health-Risk Reduction Policies:
The Prevention Survey" http://darkwing.uoregon.edu/~cameron/vsl/public prevention framed.pdf

Cameron, Trudy Ann, and J.R. DeShazo (2005b) "Demand for Public Health-Risk Reduction Policies:
The Treatment Survey" http://darkwing.uoregon.edu/~cameron/vsl/public treatment framed.pdf

Cameron, T.A., Poe, G.L., Ethier, R.G., Schulze, W.D. (2002). "Alternative non-market value-
elicitation methods: Are the underlying preferences the same?" Journal of Environmental
Economics and Management 44, 391-425.

Carlsson, F., Johansson-Stenman, O., Martinsson, P. (2004). "Is transport safety more valuable in
the air?" Journal of Risk and Uncertainty 28, 147-163.

Chilton, S., Covey, J., Hopkins, L., Jones-Lee, M., Loomes, G., Pidgeon, N., Spencer, A. (2002).
"Public perceptions of risk and preference-based values of safety", Journal of Risk and
Uncertainty 25, 211-232.

Corso, P.S., Hammitt, J.K., Graham, J.D., Dicker, R.C., Goldie, S.J. (2002). "Assessing

preferences for prevention versus treatment using willingness to pay", Medical Decision
Making 22, S92-S101.

Cropper, M.L., Aydede, S.K., Portney, P.R. (1994). "Preferences for life saving programs - how
the public discounts time and age", Journal of Risk and Uncertainty 8, 243-265.

Cummings, R.G., Taylor, L.O. (1998). "Does realism matter in contingent valuation studies?"
Land Economics 74, 203-215.

DeShazo, J.R., Cameron, T.A. (2005). Two types of age effects in the demand for reductions in
mortality risks with differing latencies. University of Oregon, Department of Economics
Working Paper. Eugene, OR.

Fetherstonhaugh, D., P. Slovic, S. M. Johnson, and J. Friedrich. (1997). "Insensitivity to the
value of human life: A study of psychophysical numbing", Journal of Risk and
Uncertainty 14, 283-300.

Frederick, S., G. Loewenstein, and T. O'Donoghue. (2002). "Time discounting and time
preference: A critical review", Journal of Economic Literature XL, 351-401.

Hammitt, J.K., Liu, J.T. (2004). "Effects of disease type and latency on the value of mortality

32

-------
risk", Journal of Risk and Uncertainty 28, 73-95.

Horowitz, J.K., Carson, R.T. (1991). "A classification tree for predicting consumer preferences
for risk reduction", American Journal of Agricultural Economics 73, 1416-1421.

Horowitz, J.K., Carson, R.T. (1993). "Base-line risk and preference for reductions in risk-to-life",
Risk Analysis 13, 457-462.

Jacobsson, F., Carstensen, J., Borgquist, L. (2005). "Caring externalities in health economic
evaluation: How are they related to severity of illness?" Health Policy 73, 172-182.

Johannesson, M., Johansson, P.O., Oconnor, R.M. (1996). "The value of private safety versus the
value of public safety", Journal of Risk and Uncertainty 13, 263-275.

Kartman, B., Andersson, F., Johannesson, M. (1996). "Willingness to pay for reductions in
angina pectoris attacks." Medical Decision Making 16, 248-253.

Krupnick, A. J., Cropper, M.L. (1992). "The effect of information on health risk valuations",
Journal of Risk and Uncertainty 5, 29-48.

Magat, W.A., Viscusi, W.K., Huber, J. (1996). "A reference lottery metric for valuing health",
Management Science 42, 1118-1130.

Nord, E. (1994). "The QALY - a measure of social value rather than individual utility", Health
Economics 3, 89-93.

Richardson, J., Nord, E. (1997). "The importance of perspective in the measurement of quality-
adjusted life years", Medical Decision Making 17, 33-41.

Subramanian, U., Cropper, M. (2000). "Public choices between life saving programs: The

tradeoff between qualitative factors and lives saved", Journal of Risk and Uncertainty 21,
117-149.

Train, K. (1986). Qualitative choice analysis. Cambridge: The MIT Press.

Tsuge, T., Kishimoto, A., Takeuchi, K. (2005). "A choice experiment approach to the valuation
of mortality", Journal of Risk and Uncertainty 31, 73-95.

Ubel, P.A., Loewenstein, G., Scanlon, D., Kamlet, M. (1996). "Individual utilities are
inconsistent with rationing choices: A partial explanation of why Oregon's cost-
effectiveness list failed", Medical Decision Making 16, 108-116.

Ubel, P.A., Loewenstein, G., Scanlon, D., Kamlet, M. (1998). "Value measurement in cost-utility
analysis: Explaining the discrepancy between rating scale and person trade-off
elicitations", Health Policy 43, 33-44.

Vassanadumrongdee, S., Matsuoka, S. (2005). "Risk perceptions and value of a statistical life for
air pollution and traffic accidents: Evidence from Bangkok, Thailand", Journal of Risk
and Uncertainty 30, 261-287.

Viscusi, W.K., Aldy, J.E. (2003). "The value of a statistical life: A critical review of market
estimates throughout the world", Journal of Risk and Uncertainty 27, 5-76.

Wittenberg, E., Goldie, S.J., Fischhoff, B., Graham, J.D. (2003). "Rationing decisions and

individual responsibility for illness: Are all lives equal?" Medical Decision Making 23,
194-211.

33

-------
Figure 2 - Sample Prevention Choice Set

These two policies would be implemented for the 100,000 people
living around you. Would you be most willing to pay for Policy A,
Policy B, or neither of them?

Policy A	Policy B

Policy in
effect

Cases
prevented

Deaths
prevented

Cost to
you

reduces pesticides in foods that cause
colon and bladder cancer

reduces air pollutants that
cause heart attacks

over 5 years

over 10 years

100 fewer cases

200 fewer cases

10 fewer deaths over 5 years

5 fewer deaths over 10 years

$70 per month
(= $840 per year for 5 years)

$6 per month
(= $72 per year for 10 years)

Your	f

choice	Policy A

reduces pesticides in foods that cause
colon and bladder cancer

r

Policy B
reduces air pollutants that
cause heart attacks

c

Neither Policy

Next Question

34

-------
Figure 3 - Sample Treatment Choice Set

Recall that these two policies will be implemented for the 100,000
people living around you. Below we describe how many of these
people get sick and die, with and without these policies.

Would you be most willing to pay for Policy A, Policy B, or neither of
them?

Policy A

Policy B

How many
Policy will
affect, and
when

Increased
Recoveries

Deaths
prevented

Cost to
you

Your
choice

treats children, adults, and seniors
(25/25/50 mix) who have leukemia

treats seniors who have heart
disease

700 will get sick over 30 years

10,000 will get sick over 4 years

25 more full recoveries

50 more full recoveries

5 fewer deaths over 30 years

5,000 fewer deaths over 4 years

$6 per month
(= $72 per year for 30 years)

$35 per month
(= $420 per year for 4 years)

c

I Policy A
treats children, adults, and seniors
(25/25/50 mix) who have leukemia

c

Policy B
treats seniors who have heart
disease

c

Neither Policy

Next Question

35

-------
Table 1: Basic Policy-Specific Variables (orthogonal)3

Variable

Mean St. Dev. Min

Max

Description

Yearly Cost

Prevention	498

Treatment	498
Illness Reductions

Prevention	862

Treatment	841
Death Reductions

Prevention	101

Treatment	539
Duration

Prevention	13.8

Treatment	13.9
Population Size

Prevention	0.246

Treatment	0.758

351
346

1584
1550

464
1240

9.7
9.7

60
60

0
0

0
0

2
2

0.359 0.001
0.358 0.001

1200	Yearly cost of policy

1200	Yearly cost of policy

5000	Cases avoided

5000	Increased recoveries

5000	Deaths avoided

5000	Deaths avoided

30	Length of policy (years)

30	Length of policy (years)

1	Size of affected population (millions)

1	Size of affected population (millions)

aExcept for exclusions based on implausible combinations

-------
Table 2: Basic Specificationsb

Parameter

Variable

Separate Samples

Pooled Sample





(1)

(2)

(3)

(4)





Prevention

Treatment

Less

More





Sample

Sample

Restricted

Restricted

P

-(Yearly Cost/10,000)

5.680

4.9854

6.1162

6.1091





(9.72)***

(8.13)***

(10.90)***

(12.09)***

Sy

Log(Illness Reductions)

0.04046

0.0358

0.04156

0.04305





(5.77)***

(5.05)***

(5.92)***

(7.42)***



... • 1 (Treatment)

—

—

.001797

—









(0.48)



5i

Log(Death Reductions)

0.1394

0.0692

0.1455

0.15206





(11.16)***

(7.20)***

(11.72)***

(12.76)***



... • 1 (Treatment)

—

—

-.05804

-0.07738









(-3.26)***

(-6.15)***

&

Log(Duration)

-0.1688

-0.1198

-0.1586

-0.17048





(-7.39)***

(-5.14)***

(-6.92)***

(-8.72)***



... • 1 (Treatment)

—

—

-.03868

—









(-0.70)



e

Policy Dummy

-0.2371

-0.2657

-0.2649

-0.27051





(-3.30)***

(-3.48)***

(-3.67)***

(-4.51)***



... • 1 (Treatment)





-.04018

—









(-0.27)



Iti(k)

Heteroscedasticity Parameter

0

--

0

0



... • 1 (Treatment)

—

0

.3659

0.35857









(2.04)**

(4.01)***

Maximized log-likelihood

-8035.19

-7539.49

-15576.35

-15577.87

Maximized log-likelihood overall

-15574.68

-15576.35

-15577.87

Total sample

size (choices)

7556

7033

14589

14589

Total sample

size (respondents)

1531

1423

2954

2954

''All specifications use shifted log format for Log(X) variable. For example, Log(Death
Reductions) is actually Log(Death Reductions +1)

Test stat for restrictions in Model 3: 3.34 Critical value: 3.84: Fail to reject restriction
Test stat for restrictions in Model 4: 3.04 Critical value: 7.81: Fail to reject restrictions

-------
Table 3: Population Size Effects

Parameter

Variable

Separate Samples

Pooled Sample





(1)

(2)

(3)

(4)





Prevention

Treatment

Less

More





Sample

Sample

Restricted

Restricted

P

-(Yearly Cost/10,000)

6.2373

3.5863

5.7920

6.7188





(8.51)***

(1.57)

(8.14)***

(10.13)***



... • Population Sizea

0.0008

2.0745

3.5046

—





(0.00)

(0.27)

(0.70)



Sy

Log(Illness Reductions)

0.0545

-0.0277

0.0452

0.0486





(5.99)***

(-0.51)

(5.05)***

(6.92)***



... • Population Size

-0.0527

0.0885

0.0147

—





(-1.95)*

(0.74)

(0.50)



4p

Log(Death Reductions)

0.1324



0.1407

0.1414





(8.40)***



(8.85)***

(9.62)***



... • Population Size

0.1067



0.0668

0.0648





(0.11)



(0.55)

(1.27)

Six

Log(Death Reductions)

—

0.0919

0.0415

0.0352







(2.66)***

(1.47)

(1.37)



... • Population Size

—

-0.0271

0.1111

0.1055







(-0.41)

(1.13)

(2.32)**



Log(Duration)

-0.2107

-0.1083

-0.2011

-0.1901





(-7.37)***

(-1.37)

(-7.24)***

(-7.56)***



... • Population Size

0.1381

-0.0268

-0.0199

—





(1.39)

(-0.14)

(-0.17)



e

Policy Dummy

-0.0967

-0.4665

-0.1444

-0.1396





(-1.06)

(-1.56)

(-1.61)

(-1.65)*



... • Population Size

-0.5216

0.2623

-0.5030

-0.4708





(-2.45)**

(0.82)

(-1 25)

(-1.83)*

Iti(k)

Heteroscedasticity Parameter

—

—

—

—



... • Population Size

0.4114

0.0597

0.5356

0.4407





(0.67)

(0.05)

(0.98)

(2.07)**



... • 1 (Treatment)

—

—

0.1315

0.1320









(1.27)

(1.30)

Maximized log-likelihood

-8025.39

-7529.39

-15569.80

-15573.09

Maximized log-likelihood overall

15554.78

15569.80

-15573.09

Total sample

size (choices)

7556

7033

14589

14589

Total sample

size (respondents)

1531

1423

2954

2954

a Affected population size measured in millions.

38

-------
Table 4: Sociodemographic Effects

Parameter

Variable

(1)

Prevention

(2)

Treatment

(3)

Pooled

P

-(Yearly Cost/10,000)

5.7124

5.0104

5.8665





(9.75)***

(8.16)***

(11.03)***

Si

Log(Illness Reductions)

0.0419

0.0355

0.0424





(5.95)***

(4 99)***

(7.32)***

Szv

Log(Death Reductions)

0.1396

--

0.1406





(11.14)***



(11.59)***



Log(Death Reductions)

--

0.0702

0.0840







(7.29)***

(6.64)***

Si

Log(Duration)

-0.1688

-0.1204

-0.1601





(-7.38)***

(-5.16)***

(-8.11)***

e

Policy Dummy

-0.4793

-0.0264

-0.2737





(-1.62)

(-0.09)

(-1.19)



... • 1 (Female)

0.0203

-0.1244

0.0198





(0.43)

(-2.53)**

(0.42)



...• l(Female) l(Treatment)

--

--

-0.1705









(-2.18)**



...(Age/100)

-2.4105

-1.8901

-3.2836





(-2.32)**

(-1.89)*

(-3.72)***



...• (Age/100) l(Treatment)

--

--

2.0492









(2.20)**



...(Age2/10,000)

2.0078

1.6335

2.8272





(2.02)**

(1.71)*

(3.29)***



... (Age2/10,000)l(Treatment)

--

--

-1.8463









(-1.90)*



... • Income/10,000

-1.9524

2.2224

-1.8301





(-2.57)**

(2.90)***

(-2.42)**



... Income/10,OOOl(Treatment)

--

--

4.4788









(3.62)***



... Educ.Years/10

0.7071

0.1046

0.6941





(6.90)***

(0.98)

(7.06)***



... Educ.Years/10 1 (Treatment)

--

--

-0.5557









(-3.78)***



... 1 (Non-White)

0.17135

0.3770

0.1696





(2.94)***

(6.00)***

(2.92)***



... l(Non-White) l(Treatment)

--

--

0.2867









(2.66)***

ln(x)

Heteroscedasticity Parameter

0

--

--



... • 1 (Treatment)

—

0

0.1931









(1.62)



Age effect minimized at

60.7

70.8

41.1 (Prev)
128.3 (Treat)

Maximized log-
likelihood



-7998.81

-7509.59

-15509.46

-------
Table 5: Sociodemographic Effects (Prevention Sample)

Parameter

Variable

(1)

SES Only

(2)
Govt. Only

(3)

SES and Govt.

P

-(Yearly Cost/10,000)

5.7175

5.9084

5.9274





(9.75)***

(10.00)***

(10.01)***

Si

Log(Illness Reductions)

0.0419

0.0420

0.0432





(5.95)***

(5.92)***

(6.06)***

<$2p

Log(Death Reductions)

0.1396

0.1438

0.1438





(11.14)***

(11.40)***

(11.37)***

Si

Log(Duration)

-0.1688

-0.1730

-0.1728





(-7.38)***

(-7.51)***

(-7.50)***

e

Policy Dummy

-0.4793

-1.3082

-1.4867





(-1.62)

(-12.42)***

(-4 77)***



... • Government Preference

—

0.2073
(14.09)***

0.2022
(13.67)***



... • 1 (Female)

0.0204
(0.43)

—

0.0103
(0.21)



...(Age/100)

-2.4154
(-2.32)**

—

-1.9811
(-1.87)*



...(Age2/10,000)

2.0071
(2.02)**

—

1.5260
(1.51)



... • Income/10,000

-1.9524
(-2.57)**

~~

-1.8089
(2.34)**



... Educ.Years/10

0.7071
(6.90)***

—

0.0619
(5.95)***



... l(Non-White)

0.1714
(2.94)***



0.1329
(2.23)***

Maximized log-
likelihood



-7998.81

-7875.33

-7847.22

-------
Table 6: Sociodemographic Effects (Treatment Sample)

Parameter

Variable

(1)

SES Only

(2)
Govt. Only

(3)

SES and Govt.

P

-(Yearly Cost/10,000)

5.0104

4.9810

4.9977





(8.16)***

(8.07)***

(8.08)***

Si

Log(Illness Reductions)

0.0356

0.0381

0.0378





(4 99)***

(5.32)***

(5.26)***

<$2p

Log(Death Reductions)

0.0703

0.0695

0.0703





(7.29)***

(7 17)***

(7.24)***

Si

Log(Duration)

-0.1204

-0.1208

-0.1215





(-5.06)***

(-5.15)***

(-5.18)***

e

Policy Dummy

-0.0264

-1.0415

-0.6918





(-0.09)

(-9.76)***

(-2.31)***



... • Government Preference

—

0.1524
(10.36)***

0.1463
(9.86)***



... • 1 (Female)

-0.1244
(-2.53)**

—

-0.1235
(-2.49)***



...(Age/100)

-1.8962
(-1.89)*

—

-2.0106
(-1.98)**



...(Age2/10,000)

1.6330
(1.71)*

—

1.7432
(1.81)*



... • Income/10,000

2.222
(2.90)***

~~

2.2750
(2.94)***



... Educ.Years/10

0.1046
(0.98)

—

0.0710
(0.66)



... l(Non-White)

0.3770
(6.00)***



0.3312
(5.18)***

Maximized log-
likelihood



-7509.60

-7432.74

-7407.98

-------
Table 7: Cancer v. Non-Cancer Policies





(1)

(2)

(3)

Parameter

Variable

Prevention

Treatment

Pooled





Sample

Sample

Sample

P

-(Yearly Cost/10,000)

5.6529

4.9757

5.6253





(9.67)***

(8.12)***

(10.46)***

Sy

Log(Illness Reductions)

0.0404

0.0357

0.04067





(5.77)***

(5.03)***

(7.13)***



Log(Death Reductions)

0.1391

—

0.1363





(11.11)***



(11.27)***



Log(Death Reductions)

—

0.0691

0.0808







(7 19)***

(6.45)***

&

Log(Duration)

-0.1707

-0.1201

-0.1557





(-7.46)***

(-5.15)***

(-7.86)***

e

Policy Dummy

-0.3035

-0.2377

-0.3325





(-4.13)***

(-3.05)***

(-5.01)***



... • 1 (Treatment)

—

—

0.0984









(0.98)



... MajorCancer

0.1748

-0.0652

0.1805





(4.57)***

(-1.75)*

(4.73)***



... • MajorCancer (Treatment)

—

—

-0.2619









(.4 50)***

Iti(k)

Heteroscedasticity Parameter

0

—

0



... • 1 (Treatment)

—

0

0.1207









(0.97)

Maximized log-likelihood

-8024.78

-7537.96

-15563.38

Sample size

(choices)

7556

7033

14589

Maximized log-likelihood overall

-15562.74

-15563.38

Total sample size (choices)

7556

7033

14589

LR-test of restrictions in pooled model: (4) ~ 9,48, j2 = 1.28, fail to reject restricted

model.

42

-------
Table 8: Heterogeneity by Disease Type

Parameter

Variable

(1)

(2)

Prevention

Treatment

P

-(Yearly Cost/10,000)

5.6739

5.1971





(9.62)***

(8.42)***

Si

Log(Illness Reductions)

0.0398

0.0350





(5.65)***

(4.90)***

s2

Log(Death Reductions)

0.1439

0.0705





(11.41)***

(7.28)***

s3

Log(Duration)

-0.1781

-0.1180





(-7 71)***

(-5.02)***

e

Policy Dummy

-0.1508

0.0388





(-1.59)

(0.40)



... • Heart Disease

--

--



... • Heart Attack

-1.59

-0.0093





(-1.16)

(-0.11)



... • Cancer (General)

0.2582

--





(2.91)***





... • Colon/Bladder Cancer

0.0833

-0.2862





(0.92)

(-3.26)***



... • Leukemia

-0.4663

-0.4785





(-4.88)***

(-5.16)***



... • Leukemia in Children

0.2250

0.0197





(2.57)**

(0.11)



... • Stroke

-0.3380

-0.3764





(-3.59)***

(-4.28)***



... • Respiratory Disease

-0.0612

-0.1531





(-0.67)

(-1.70)*



... • Resp. Dis. in Children

--

-0.1774







(-1.03)



... Asthma

-0.5373

-0.3942





(-5.55)***

(-4.03)***



... Asthma in Children

0.1691

-0.3036





(1.93)*

(-2.19)**



... Lung Cancer

-0.0674

-0.5230





(-0.74)

(-5.78)***



... Traffic Injuries

-0.1877

--





(-2.32)**





... Injuries

--

-0.3486







(-3.73)***



... Injuries to Children

--

0.2580







(1.50)



... Prostate Cancer

--

-0.4677







(-5.20)***



... Breast Cancer

--

-0.0578







(-0.68)



... Skin Cancer

--

-0.8526







(-8.94)***

Maximized Log-likelihood

-7930.80

-7453.51

Sample Size (Choices)

7556

7033

-------
Table 9: Heterogeneity by Other Policy Attributes

Parameter Variable

(1)

Prevention

(2)
Treatment

P -(Yearly Cost/10,000)

5.6777

5.1482



(9.72)***

(8.35)***

<5,p Log(Illness Reductions)

0.0405

--



(5.78)***



<5[t Log(Illness Reductions)

--

0.0359





(5.04)***

S2 Log(Death Reductions)

0.1390

0.0714



(11.11)***

(7.38)***

<5$ Log(Duration)

-0.1693

-0.11896



(-7.40)***

(-5.07)***

0 Policy Dummy3

-0.3491

-0.2135



(-4.08)***

(-2.63)***

Cause of ailment





...¦ Air Pollution

0.0981

--



(1.75)*



... • Water Contaminants

0.1181

--



(1.80)*



... • Pesticides in Foods

0.2317

--



(3.55)***



Gender-specific illnesses; respondent gender





...¦ Breast Cancer

--

0.0865





(0.92)

... • Breast Cancerl(Female)

--

0.3655





(2.94)***

... • Prostate Cancer

--

0.1620





(1.72)*

... • Prostate Cancer l(Female)

--

-0.5533





(-4.03)***

"Affected group " choices





...¦ Percent Children

--

0.1220





(1.29)

... • Percent Childrenl(Age65+)

--

-0.2553





(-1.50)

... • Percent Childrenl(Kids)

--

0.6027





(4.24)***

... • Percent Seniors

--

-0.2295





(-4.50)***

... • Percent Seniorsl(Age65+)

~

0.0042





(0.05)

Maximized Log-likelihood
Sample Size (Choices)

-8028.5356
7556

-7475.99
7033

a Prevention: omitted category= traffic accidents;
Treatment: omitted category=all other illnesses or injuries

-------
Discussant Comments
Kelly Maguire
EPA Workshop

Morbidity and Mortality: How Do We Value the Risk of Illness and Death?

April 11,2006

Session II: Issues With Morbidity Valuation

Altruism and Environmental Risks to Health of Parents and Children by Mark Dickie and
Shelby Gerking

Is An Ounce of Prevention Worth a Pound of Cure? By Ryan Bosworth, Trudy Ann
Cameron, and J.R. DeShazo

Mark Dickie and Shelby Gerking's paper, Altruism and Environmental Risks, is very
interesting and well written. It is thorough and was a pleasure to read. The authors test a
model of altruistic family behavior using a sun screen that will protect against both the
risk of getting skin cancer, as well as dying from the cancer conditional on a positive
diagnosis. They employ a stated preference survey using adults in Mississippi. Adults
are asked their perceived risks of contracting and dying from skin cancer for both
themselves and their child, and then they are asked the willingness to pay (WTP) for a
sun screen that will reduce these risks by 10 percent or 50 percent, which are randomly
assigned. The results are used to test the existence of altruism in the family.

Altruism is an important concept to consider in economic analysis. Primarily our
concerns rest with the impact of altruism on valuation. In its simplest form, if parents, or
any individual for that matter, behave in an altruistic manner, then individual values for a
risk reduction will be compromised to the extent that they incorporate more than just the
individual's WTP. By summing individual values we would then risk double-counting or
over-estimating the total value for a risk reduction.

The paper could be more informative in this regard by including some discussion of the
different types of altruism. Paternalistic altruism exists when an individual has concern
for another's welfare, but is not necessarily concerned about the costs imposed on that
individual. In other words, the paternalistic altruist does not incorporate the other's
utility function into their own decision-making. Non-paternalistic altruism exists when
an individual cares about both the benefits and costs imposed on another. That is, the
non-paternalistic altruist fully accounts for the other's utility when making decisions. It
would be useful to have a discussion of the different types of altruism and how they relate
to this study, as well as valuation results.

Some additional questions that arose when reading the paper that could have implications
for the application of these results to policy include:

How do the results change when there are multiple children in the household? Do
parents adjust their WTP to account for the additional children?

1

-------
How do you account for two-parent versus single-parent households? If each
parent in a two-parent household is altruistic how does this affect the values for the child?

How do you account for other individuals who are altruistic towards children,
such as grandparents?

Overall, this is a well-written and interesting paper that sheds light on an important issue
for benefits analysis.

Ryan Bosworth, Trudy Cameron, and J.R. DeShazo also have an interesting and well-
written paper, "Is An Ounce of Prevention Worth a Pound of Cure?" They investigate
preference for treatment versus prevention policies over a wide variety of policy
attributes, types of illnesses and accidents, and respondent characteristics. There is a
substantial amount of information in a short paper. Their results show that individuals
are willing to pay almost double for prevention of death than for treatment of an illness
that can cause death. For example, WTP for prevention of a death is $245, whereas WTP
to treat an illness that causes death is $138. This is not surprising. There is disutility
associated with entering the diseased state and therefore individuals are willing to pay to
avoid entering that state. The authors also find that people are willing to pay equivalent
amounts to treat and prevent illnesses, at about $70 for both.

The largest contribution of this paper to policy is in terms of determining how a policy
maker may allocate resources. These results suggest that people would rather prevent
than treat outcomes. Again, this is not surprising and it would be useful for the paper to
explore more of why this might be the case. My sense is that it is related to either the
uncertainty associated with outcomes, or the stigma, or both.

In terms of uncertainty, people are WTP to avoid uncertain outcomes, particularly those
that result in death. Individuals would rather prevent cancer, than be in the state of
having cancer and facing the possibility of death and having to back out from that state.

These results are consistent with the approach we have found to be the case in the
manufacturing sector. Twenty years ago the Pollution Abatement Costs and
Expenditures (PACE) survey primarily addressed costs associated with treating pollution,
say installing scrubbers on a stack to treat emissions, or filters at water discharge areas to
treat water before release. Today, we are in a pollution prevention paradigm. The
treatment options have been addressed and we now focus on preventing emissions before
they are created. Much of the expenditures at manufacturing facilities that we see
through the PACE results support this notion.

It is also possible that stigma is driving these results. People would rather not enter a
disease or illness state that may have a stigma associated with it. Hence, they are willing
to pay more to avoid the stigma of being a survivor. It would be useful to explore these
ideas further in the paper.

Other questions that would be useful to address include:

What are the implied VSL or morbidity values that result from this study?

2

-------
What is the impact of the complex question design on results?

Overall, the paper is interesting and provides a useful discussion of how individuals value
treatment versus prevention programs.

3

-------
No documents are available regarding Kevin Boyle's discussion

comments.

-------
Summary of the Q&A Discussion Following Session II

Perry Beider, (Congressional Budget Office)

Commenting on the presentation of the Bosworth/Cameron/DeShazo paper and referring
specifically to the finding that "someone ideologically opposed to government
intervention would support a certain program once there was enough personal direct
benefit perceived from it," Mr. Beider asked if the researchers observed that it went the
other way also. In other words, was it observed that people who were generally in favor
of government intervention did not support policies if there was too little perceived
personal benefit?

Trudy Cameron, (University of Oregon)

Dr. Cameron responded that "it is sort of treated symmetrically—if it works in one
direction, then it works in the opposite direction also, just by the structure of the model."

JR. DeShazo, (UCLA)

Dr. DeShazo continued the response, adding: "But the effect isn't quite as large—the
ideology effect dominates. It's true that people strongly ideologically in favor of
government intervention are responsive to the size of the private benefits, but much less
so than at the other end of the continuum."

Bryan Hubbell, (U.S. EPA, OAQPS)

Also addressing the Bosworth/Cameron/DeShazo paper, Dr. Hubbell commented on the
finding presented toward the end of the paper that when people were asked whether they
prefer policies that help seniors or not, all of them said "no," including the seniors. He
commented that "we just throw out the term seniors as if that's a well-defined term," and
he added that he is curious to know whether the researchers worked to uncover an age
breakpoint for this phenomenon. He clarified by asking, "What age does a policy have to
affect before people will say that they'd rather not have that policy?—Is it 50? 55? 60?
65?—and is there any kind of declining support ratio at that point?"

Trudy Cameron

Dr. Cameron responded that the issue raised is an item of discussion on the Wednesday
agenda. She added, "For private preferences we have some very detailed and elaborate
analysis of age effects that are much richer than the simple quadratic thing that tends to
dominate most of the prior literature. In the public choices study, which was discussed
today, the distinction among beneficiaries is just defined in three groups—seniors, adults,
or children. It's left to the individuals to interpret whether they are a senior or not."

J.R. DeShazo

Picking up on the response, Dr. DeShazo added, "Actually, I think for the respondents we
did define the age intervals—60 or 65 was the cutoff."

Session II Q&A Summary

1

-------
Trudy Cameron

Dr. Cameron clarified, "But that's the beneficiaries—we have very detailed information
about the respondent's age, of course, so that can be much richer."

Douglass Shaw, (Texas A&M University)

Addressing the authors of both papers, Dr. Shaw asked, "What's the welfare measure?—
what is it reallyV He said that in Dr. Cameron's journal paper it is a "pretty careful
derivation of an option price, which is kind of what we think it should be."

Dr. Shaw also asked, "When you do the subjective risk estimates, are you going after just
the baseline risks, or are you also getting the subjectives on the risk changes?—and in
either one of the designs, did you look to see if things are adding up?" He expounded that
particularly in the Dickie/Gerking study there should be an obvious implication when
doing a conditional probability. Acknowledging that the authors said they can do
compound probabilities, which Dr. Shaw classified as "a very unusual result in the
literature," he asked whether they did anything simple also.

. /. R. DeShazo

Seeking clarification of the question, Dr. DeShazo asked, "Do you mean data analysis-
wise or with the respondents or . . . ?"

Shaw

Dr. Shaw stated, "On the latter one, you're sort of saying that the results support that you
can do compound probabilities, so obviously there's a law of probability between a
conditional and an unconditional probability, so did you kind of just ask them to do a
little experiment in the survey where you could verify that in fact they got that?"

DeShazo

Dr. DeShazo responded, "When I said you could do compound probabilities, I'm not sure
I meant that if you asked them the unconditional probability and then did the
multiplication on their conditional and their unconditional morbidity risk, would you
actually get exactly the same number. I think they were making tradeoffs between those
two risks that were consistent with the model, and I think they could distinguish between
the two risks and not be confused between them. But we didn't ask them "what do you
think the unconditional mortality risk is?" which you could use then to test whether they
were really doing the math right. So, I don't know that."

Dr. DeShazo continued, "The perceived risk is all baseline; we asked them "what do you
think the risk is?" The risk changes are exogenously assigned in the experimental
design—they just come packaged in the sunscreen." He added that for the welfare
measure they were looking at anti-willingness to pay for risk changes that would occur
later in life.

Session II Q&A Summary

2

-------
Unidentified Participant

The response to Dr. Shaw was clarified by explaining that the probabilities were
presented one at a time. For an example, "first the respondents were asked about the
probability of getting skin cancer, and after they wrestled their way through that question,
then they were asked about the chance they would die, given that they had it. So, we
don't really have any results that say people can juggle two probabilities at the same
time—and I'll bet they can't do it, just as you alluded."

New Questioner

"How do you go about measuring violations of rationality when people are answering
your survey questions?—is it a violation of transitivity assumption? If you do those
measures, what do you do with the results? Do you throw out people who are clearly
violating rationality?

. /. R. DeShazo

Commenting that it was a very good question, Dr. DeShazo replied, "You can and we
have looked at violations of rationality. We've also looked at how much attention people
spend absorbing the information that we've given them, and we have altered our sample
based on some minimum level of attentiveness that we felt they needed." He went on to
say that there is always the sticky issue of how much information is enough and how
much is too much. He added that he is "deeply concerned about the declining cognitive
efficiency of individuals when they're given too much information." Saying that "we are
all always given too much information—and we sort through it," he identified one of the
tasks for researchers presenting information to individuals is to ask, "Have we left out
something that is important?" Dr. DeShazo said he believes that if you give individuals
enough familiarity with the attributes that make up a program, they'll decide for
themselves which attributes are most and least important, and the proof (or disproof) of
that will show up in your statistical analysis.

Dr. DeShazo added that in addition to looking at time on task his research team asks
respondents, "How difficult was that choice?" In closing, he said that "in the context of
evaluating their risk judgments, we can actually look at whether or not they make
consistent decisions—we give them quizzes, basically, in the private version of the
survey."

Trudy Cameron

Continuing the response to the questioner, Dr. Cameron stated, "J.R. mentioned this
notion of how much attention people give to different aspects of a particular survey
design. J.R. with Herman Fermo has some pretty rigorous work that came out in 2002
looking at how the structure of the randomized design affects the amount of noise, the
choice inconsistencies that people make." She added that work that they're doing now, in
conjunction with another student at Oregon University, Dan Burkhart, "has to do with an

Session II Q&A Summary

3

-------
actual sort of optimal allocation of attention problem." Dr. Cameron asserted that this
deepens the model a bit by "acknowledging that what you think you're estimating as a
marginal utility in a choice model as a consequence of standard multiple-choice
specification is actually the product of some fractional attention, which may be very
small or very large, times the true underlying marginal utility that you would be
estimating if they were paying full attention—and that's producing some very interesting
results. That's just some fundamental broader research that will have some bearing on
these data as well."

Lauraine Chestnut, (Stratus Consulting, Inc.)

Saying that this might be getting back to Douglass Shaw's question about welfare
measure, Ms. Chestnut addressed this question to Drs. DeShazo and Cameron: "When
you're asking questions about public policies that affect the person and everybody else in
the community, how do you interpret that relative to the private valuation numbers that
we tend to want for benefit/cost analysis. So, the example of the responses for seniors—
and we've seen this in some other studies that ask these questions about public policy—
what does that mean for valuation purposes?"

Trudy Cameron

Dr. Cameron responded, "Going back to respond to Douglass's question, which I didn't
get a chance to: In the private choices survey, the model is highly structural and has to
do with discounted expected utility maximization getting to an option price. But, for
individual choices with respect to their own budgets and their own preferences for stuff
that happens to them, it's a little easier to do that. This may account for why we haven't
directly addressed much of the public choice stuff before—it's harder to come up with a
solid, theoretical model about how people should think about these public goods. So, by
its nature the public choices study with its two different surveys is very much more
exploratory. Perhaps the term descriptive would be better—we're sort of identifying the
stylized facts that need to be addressed in any further theorizing rather than starting with
a rigid model. Bosworth has a more structural specification with respect to discounting—
that's stuff he's working on now and finishing up for his third essay—but we figured
we'd start with just the high points of the actual description of people's choices."

./. R. DeShazo

Acknowledging that he is relatively young to this field, having been actively involved in
VSL literature only for 3 or 4 years, Dr. DeShazo said he finds it "a hard sell that we
should be using these private good estimates for public policies because preferences over
aspects of the public policy are so different. We could try to explain two different
things—their actual support, their behavior—and it seems to me that if you're interested
in their actual behavior with respect to these policies, you have to give them these
attributes of public policies that don't hold or don't exist for private programs. Also, that
behavior presumably reveals something about their perceived welfare from the public
policy. I think the challenge is really on those that want to use the private estimates,

Session II Q&A Summary

4

-------
because, to me, they seem like very different utility functions with very different
arguments in them."

Reed Johnson, (RTI)

Mr. Johnson said that he had actually "looked forward to a lively discussion on the IOM
Report, but I guess neither of the discussants were asked to comment on that. I'm afraid
that we have on our hands another NOAA Blue Ribbon Panel Report that's going to be
cited for the next 15 or 20 years, long after the evidence base that was used to make the
recommendations has become obsolete. I'm a little concerned that the panel was
constituted in a way that sort of biased it in favor of conventional ways of thinking about
health utility that don't really line up very well with the way most people in this room
think about utility—and I'd like to thank Alan Krupnick for his valiant efforts to try to
keep the process a little bit more honest in that respect."

He continued, "There are a couple of aspects of the recommendations that I find
troubling, in addition to Nathalie's points. For example, the quality recommendation is
that the quality should be elicited for a general population sample. I work a lot with
patient surveys and patient preferences and with some general population surveys. For
many of the particular outcomes of interest, it is difficult for people who have never
experienced that outcome to give meaningful values. . . . The general result is that
patients experience much less of a utility loss than the general population assumes that
they experience, partly because of adaptation and partly because people just imagine that
something is going to be a lot worse than the experience actually turns out to be."

Mr. Johnson added that the report includes a recommendation for more research, but he
said he thinks "the recommendation on gathering more data is stronger than the
recommendation for improving methods," and he said he would have liked to have seen a
much stronger advocacy for providing "measures of health utility that are both
theoretically correct and empirically robust."

In conclusion, Mr. Johnson stated that he feels the publication of the report is an
opportunity for groups like this to become engaged in trying to understand not only what
obligation EPA is going to have in terms of doing their analysis but also what we can do
to help encourage "more nuance of interpretation and more flexibility in use of methods."

Someone

"I thank Reed for that compliment. There were a lot of people on the committee that
worked hard to do what we did. What we were trying to do is to create separation
between measures of utility that we use in this literature, the economic valuation
literature, and the measures of quality-adjusted life years and so on that are used in this
other literature. . . . Hopefully that will serve the policy process and also serve our
profession."

END OF SESSION II Q&A

Session II Q&A Summary

5

-------
Morbidity and Mortality: How Do We Value the Risk of

Illness and Death?

PROCEEDINGS OF SESSION III: PANEL DISCUSSION ON THE USE OF THE

INTERNET IN VALUATION SURVEYS

A WORKSHOP SPONSORED BY THE U.S. ENVIRONMENTAL PROTECTION
AGENCY'S NATIONAL CENTER FOR ENVIRONMENTAL ECONOMICS AND
NATIONAL CENTER FOR ENVIRONMENTAL RESEARCH

April 10-12, 2006

National Transportation Safety Board
Washington, DC 20594

Prepared by Alpha-Gamma Technologies, Inc.
4700 Falls of Neuse Road, Suite 350, Raleigh, NC 27609

ACKNOWLEDGEMENTS

This report has been prepared by Alpha-Gamma Technologies, Inc. with funding from
the National Center for Environmental Economics (NCEE). Alpha-Gamma wishes to
thank NCEE's Maggie Miller and the Project Officer, Cheryl R. Brown, for their
guidance and assistance throughout this project.

DISCLAIMER

These proceedings have been prepared by Alpha-Gamma Technologies, Inc. under
Contract No. 68-W-01-055 by United States Environmental Protection Agency Office of
Water. These proceedings have been funded by the United States Environmental
Protection Agency. The contents of this document may not necessarily reflect the views
of the Agency and no official endorsement should be inferred.

-------
Table of Contents

Session III: Panel Discussion on the Use of the Internet in Valuation
Surveys

Session Moderator: William Wheeler, U.S. EPA, National Center for
Environmental Research

Panel Participants:

•	Nathalie Simon, U.S. EPA, National Center for Environmental Economics

•	J.R. DeShazo, University of California-Los Angeles

•	Shelby Gerking, University of Central Florida

•	Alan Krupnick, Resources for the Future

•	Jon Krosnick, Stanford University

•	Brian Harris-Kojetin, Office of Management and Budget

Questions and Discussion

-------
U.S. EPA NCER/NCEE Workshop
Morbidity and Mortality: How Do We Value the Risk of Illness and Death?

Washington, DC
April 10-12, 2006

Session III: Panel Discussion on the Use of the Internet in Valuation Surveys

Nathalie Simon, U.S. EPA, National Center for Environmental Economics

Will has asked me to sort of set things up for the panel discussion, so I'll talk through it a
little bit and then present the charge questions. The way I see it, there are three kinds of
internet surveys. There are those in which you recruit individuals into the sample using
standard random probability sampling—and then you ask people to actually complete the
survey over the internet, using a link that you provide. Then there are internet surveys
using standing panels, and there are two kinds of standing panels: those in which
individuals self-select into the panel and those in which the panel is created using
sampling techniques.

It seems to me that there are benefits to all those types of web-based surveys. Generally
speaking, once the survey is administered, you tend to have quicker turnaround on the
results. In addition, you often have lower costs with these types of surveys and lower
respondent burden. You can have greater accuracy as well—there's no interviewer bias
or data-entry mistakes to worry about. Generally, individuals are entering the data the
way they want to, and then they submit the results to you.

There is also greater flexibility in how the information is presented. You can have
complicated skip patterns programmed directly into the survey instrument—you can have
extensive use of graphics and color, which would be expensive or difficult to do using
other modes. You can also have more interactive questions and can basically tailor the
survey to individuals as they're going through it. Especially in the case of the standing
panels you have the availability of some unique information that has been collected prior
to the survey being administered.

You can also get information on time for question, and you can have extensive variable-
tracking information if you need it. In some cases, you also have the possibility of using
a voice-over, which can be very helpful in getting people to understand the questions that
are being asked and to take the time to listen to the questions as well as reading them.

Of course there are a number of problems associated with web surveys as well. With the
panel-based surveys you can often have low response rates. In fact, those of you who are
users of Knowledge Networks, if you start looking at the response rate from the time that
people are initially contacted to join the panel, the response rate is rather low. Given this
low response rate, non-response bias then becomes an issue. In other venues and at other
conferences, I've also heard a concern expressed that panels run the risk of creating

Session III Panel Discussion on Internet Use

1

-------
"expert survey takers"—I believe Reed [F. Reed Johnson] referred to them as "trained
seals" at one point. That is a concern, as well. If you're going to the same individual
repeatedly with surveys, do you create this expert survey taker?

There are other issues as well, especially with those surveys in which individuals are self-
selecting into the panel. Really, these result in little more than convenience samples. It's
often difficult to tell whether you're getting more than one individual from a household
and things like that. You may have problems with actually downloading the information
from the internet, and you may run into technology constraints as well.

Regardless of these problems, we are intrigued by the benefits associated with these web-
based surveys, especially given that telephone surveys are becoming more and more
difficult to do and mail surveys are also somewhat difficult—these modes pose problems
when we're dealing with complicated questions that do involve complicated skip patterns
or where we would like to use more complicated graphics.

As an agency though, to my knowledge, I think we've managed to use web-based surveys
only on a very limited basis, and generally these have been surveys that have been done
for research purposes. You'll hear more about one of these tomorrow, "Eliciting Risk
Tradeoffs for Valuing Fatal Cancer Risks." This was work done by Chris Dockins,
Melonie Sullivan, who is no longer with our office, and George Van Houtven, who will
be presenting the paper tomorrow. Two other surveys looked at willingness to pay for
water improvements, one designed by Kip Viscusi looking at eliciting willingness to pay
estimates for improvements to fresh water, and then a survey looking at coastal water
improvements.

But, again, these were surveys that were either couched in terms of pure research or
testing of survey instruments—so, they were pilot surveys. We have yet, at least to my
knowledge, to get approval for a web-based survey that would feed directly into a policy
analysis for one of our rules and regulations.

Faced with the problems associated with web surveys, but trying to balance those against
the benefits that we could exploit, I wonder whether there is a way to actually address
some of these issues. One of the most important ones, perhaps, is this issue of non-
response bias. It seems to me that non-response bias is perhaps more of an issue if you
are dealing with low response rates, but it seems that you have a potential for non-
response bias regardless of what the response rate is, unless of course you're dealing with
a survey where you have 100 percent compliance. So, it seems to me that one question is
"How do we address that?"—How do we go about trying to improve the
representativeness of the sample or "How do we test for sample representativeness?"

Thinking about all these things, we had asked our panelists here to think about several
questions as they were looking back over their own research and what they've done in
this field, [referring to a slide] We have the questions for the panelists up on the monitor
here. Basically, we've asked people to think about their experience with using the
internet as a survey mode and to think about the choice of the survey mode in their

Session III Panel Discussion on Internet Use

2

-------
research and to consider the tradeoffs between convenience, cost, and bias and to
comment on the key issues. Specifically, we've asked them to address these questions:

•	What special issues must be considered when using the internet as a survey
mode?

•	Are there special circumstances where it makes sense to use the internet for
stated-preference surveys?

•	What conditions or circumstances does the internet provide and under what
circumstances should the mode be avoided?

•	What specific follow-up analysis or testing should be conducted when using the
internet?

Brian Harris-Kojetin, Office of Management and Budget

I'm going to take a very "10,000-foot-level perspective" here and then focus in a little bit
and touch on some of the things that Nathalie mentioned, but I suspect that others in the
panel will focus even more deeply into the specific issues. For those of you who have
ever wondered, "Why does OMB review our surveys?" it is because we are required to
by law. The Paperwork Reduction Act requires that any information collection that is
sponsored or conducted by a federal agency go through this review, and the purpose of
this is to improve the quality and practical utility of information that is gathered by
federal agencies.

I want to make you aware of some new guidance that we recently issued in January of
this year. It's entitled "Questions and Answers When Designing Surveys for Information
Collection," and it covers a broad swath of things, but I'll just provide a brief overview of
that. Just so you know, the intended audience for this guidance is very broad. It's
intended to be used by people implementing the Paperwork Reduction Act—Chief
Information Officers, Program Managers, survey folks, who are out there in the front
lines doing this—and it covers a wide variety of different topics—everything from what
do you have to do in terms of some basic process issues in terms of submitting the
information collection requests or "OMB clearance packages," as they're more popularly
known to what kinds of different issues you need to address and explain and justify and
document here. I'm going to focus on a few of these that are related to some of the things
you're interested in here in terms of internet surveys.

Specifically, we have questions on when should agencies consider designing a survey.
Obviously, a survey is just one method in a social scientist's arsenal—it's appropriate
some kinds of questions and issues and not so much for others. We have a whole section
on sampling covering probability samples, coverage issues, sampling frames, . . .We
bring in the issue that Nathalie raised, too, in terms of non-probability samples—that
there are these internet panels out there that are essentially convenience samples. Even
though these panels often boast of their numbers, which can reach over a million people,
what you have is still an entirely self-selected convenience sample of "1.2 million"
people who had nothing better to do than stumble across a web site and say, "Sure."

Session III Panel Discussion on Internet Use

3

-------
I'll also touch on several points about the mode of data collection . . . also the
Government Paperwork Elimination Act (GPEA), which OMB is also in charge of
implementing. This required agencies to allow citizens electronic options for reporting to
the federal government. Although this law was not really written with surveys in mind, it
can be applied that way—it's more for people who are applying for benefits or things like
that or for businesses conducting transactions with the government. . . .

There's also a short section here in the guidance on stated preference methods. For those
of you familiar with OMB Circular A-4, there's nothing really new here.

Generally, agencies are being encouraged to do a lot of electronic reporting, but there are
some important stipulations in GPEA—agencies are encouraged to do this, "as practical."
So, if you're doing a very small-scale survey or if you're doing anything with fewer than
5,000 respondents you don't even really need to consider it—or if it's otherwise just not
practical or cost-effective.

Most federal agencies that use the internet for their surveys use it as one option in a
multi-mode survey. Looking across agencies, it's being used more and more for
establishment surveys or business surveys or surveys of organizations or institutions like
hospitals or schools, and sometimes it is being used as the sole mode for those or for
some specialized survey, such as a web site satisfaction survey. It's being used as the
sole mode pretty much exclusively in cases where your target population or some sub-
population of that has nearly universal web access. Not all businesses have that, but in
certain industry sectors you can really count on that. There are several post-secondary
school surveys that are now based exclusively on web collection.

One other thing that I want to point out is that web surveys are sometimes touted as
convenience, and with some of the things I've seen from agencies it's not clear if they're
thinking about the respondent or themselves. It can be very convenient for the
researcher sometimes to use a web survey, but not so much for the respondent. The
worst case scenario that I've seen in this regard is where the agency sends out a request to
a respondent saying, "Please do our survey on the web. What you can do is download
this, print out the PDF file, go fill it out, and then get on the web and put all the
information in." This is not more convenient for the respondents. Why not just mail
them the survey and let them mail it back? Why do they have to go through these extra
steps?

In terms of cost, web surveys are often portrayed as being less costly. This is true under
some circumstances, especially if it's a very simple survey that doesn't require much
complex programming or testing. We have a lot of government surveys that very quickly
become very complicated

In terms of bias and error reduction, we're looking for agencies to take these things into
account and talk about how they are dealing with them in terms of why they've chosen
the mode or modes that they're using.

Session III Panel Discussion on Internet Use

4

-------
In terms of choosing an internet survey as the preferred mode or one to be avoided, in
reviewing packages we're basically looking for a good understanding and justification of
how the agency is balancing some of the advantages and disadvantages—and Nathalie
mentioned a number of these. Email reminders are certainly cheap and convenient for
prompting respondents, especially if you can include that hyperlink in the email message
that will take them right there. That has advantages over sending them a postcard with a
long address that they have to type in. You do get faster data collection without delays in
receiving the data. For instance, respondents can't tell you, "Oh, I mailed that last
week,"—you know whether it's completed or not. Using visual aids and sometimes even
multimedia is another advantage, as is the ability to build in some of these edit-and-
consistency checks.

Disadvantages: Again, reflecting some of those issues mentioned earlier in terms of
coverage and non-response and measurement error. What is the sampling frame?—
Where did this sample come from?—Can you actually draw a random sample from your
target population?—How well are you covering your target population? There are issues
of response rates, in general—again, when they are used as a sole mode, web surveys
tend to have lower response rates than other modes. That said, they are more often used
as a mixed mode. Respondents have to be computer-literate and have access. There are
hardware and software differences that can affect your presentation. Finally, there are
some respondent concerns about confidentiality when giving information over a web site.

In terms of follow-up analysis or testing, I want to make two points. One is that pre-
testing is just as important [as follow-up]—have the questions been tested to determine
whether they are functioning as intended? When you're putting this on a web instrument,
you need to do the usability testing as well. As far as follow-up analyses to assess a
potential non-response bias—we all recognize that non-response rates don't indicate non-
response error—they're an indicator for the potential for non-response bias. We expect
that surveys collecting influential information should achieve high response rates, and
agencies need to consider how what they are doing is going to give them data of the
quality that they need. Our guidance, as many of you are probably aware, says that if an
agency is getting a response rate of less than 80 percent, they need to plan a non-response
bias analysis. There's a variety of ways of doing that—I think some of the people [fellow
panelists] are going to talk about some specific examples here. Bob Groves and Mike
Brick have taught a course now several times at a few federal agencies as well as to the
general public—this is in the joint program in survey methodology—on Practical Tools
for Non-Response Bias Analyses.

Shelby Gerking, University of Central Florida

I want to report on some joint work that Mark Dickie and I have done using web-based
surveys in valuation studies. We have some experience, at least, working with internet
panels. We've worked with the CentERpanel at Tilberg University in the Netherlands
and looked at willingness to pay for greater protection of seals there. That was back in
the early part of this decade. Also using CentERpanel, we've looked at willingness to

Session III Panel Discussion on Internet Use

5

-------
pay for reduced risk of pancreatic cancer. Using Knowledge Networks, we've looked at
blue-collar workers' willingness to pay for on-the-job-safety improvements. That was
another study done earlier in this decade. More recently we've looked at parents'
willingness to pay for reduced skin cancer risk to themselves and to young children ages
3 - 12. This last study is the one that I want to base my remarks on now, because it
serves as a side-by-side comparison to the computer-assisted study that Mark Dickie
reported on earlier.

The Knowledge Networks, or KN, Skin Cancer Survey in 2005 was transmitted to about
1200 panelists, and we, in one way or another, got down to 644 panelists with a child
between the ages of 3 and 12 years who actually did complete the survey that was
provided. The panelists completed the survey at home—there was about a 3-month
period for Knowledge Networks to design, pre-test, and field the survey and for
respondents to return a usable data set. It was a very smooth process, with very good,
helpful people to work with.

The comparison is with the Hattiesburg Skin Cancer Survey from 2002 that Mark Dickie
reported on. The survey was virtually identical, though not exactly identical, to the
Knowledge Networks survey. It consisted of a sample of 612 parents with at least one
child between the ages of 3 and 12 years. As Mark indicated, that survey was obtained
by random-digit dialing of Hattiesburg area residents, and it took about a hundred calls
from the poor students there to generate one completed survey. There were lots of hang-
ups and lots of reasons why people might say "No," but there was also an eligibility
problem, of course, because people had to have at least one biological child living at
home between the ages of 3 and 12 years—that accounts for a lot of the extra phone calls.
Respondents came to the University of Southern Mississippi campus and took this survey
in a computer lab there, so rather than just being able to off-load the survey to the good
folks at Knowledge Networks, we needed a lot of students and oversight to make sure
that we at least knew what was going on in this computer lab.

As to the cost, using Knowledge Networks cost us $82 per completed survey. This
excludes the investigator time needed to develop the survey—in other words, the clock
starts running when you hand the survey to Knowledge Networks. It includes all pre-test
costs and all of Knowledge Networks costs and all university indirect costs. With the
Hattiesburg survey, it cost about $123 per completed survey. Again, that excludes the
cost of investigator time used for survey development. Although it's not exactly the
same, I tried to make the comparison as much apples to apples as I could. Anyway, the
oversight that you need with one of these computer-assisted surveys is significant, and I
valued Mark's time and my time on that job at about 9 cents per hour. The Hattiesburg
cost includes the $25 participation fees provided to those who came to the computer lab
and took the survey, and it includes all pre-test expenses, programming costs, labor and
telephone charges, and university indirect costs. So, the Hattiesburg survey came at about
a 50 percent cost premium.

Session III Panel Discussion on Internet Use

6

-------
Data Quality:

As far as sample composition, the Hattiesburg sample was more highly educated than the
KN sample, and this is what you would expect, given that random-digit dialing was used
to recruit the survey. The sample was more highly educated than you would have
expected, given the census data for the Hattiesburg area. The Knowledge Networks
survey was more representative of the United States population, but I would call attention
to the fact that we're not really sure who completed all the surveys. When we were
debriefing pre-test participants, out of eight such persons that we spoke with (Knowledge
Networks had arranged the calls and was on the line also), we found out that one of them
was not the person who had completed the survey—it was that person's spouse, instead.
How widespread this problem is I have no idea—I'm not trying to condemn the
Knowledge Networks survey on the basis of one observation.

The average survey completion time for the Hattiesburg survey was 26 minutes, and we
had projected a completion time of 25 - 30 minutes, based on our own experience taking
the survey and the time it took pre-test respondents. Twenty-three percent completed the
survey in 20 minutes or less—you also want to know how many people just ripped right
through it and probably didn't pay too much attention to what they were doing. In the
KN survey, it took 1178 minutes for those respondents to complete the survey. One
interpretation is that these people obviously work much more carefully than they do in
Hattiesburg, but there are other interpretations as well that could be offered. One is that
if you're at home and you're doing this on the internet, you're free to look at the survey.
That's when the clock starts running, and then you say, "Yes, I see what this is—it looks
very interesting—I think I'll do it in three days." That's possible. Another possibility is
that you look at the survey and begin to do it but you decide to come back later to finish
it. Then when you return, you have to pick up where you left off and reconstruct your
train of thought. Seventeen percent of the surveys returned were "resumed interviews"—
this is how Knowledge Networks refers to a survey that exceeds 100 minutes. Actually, I
would classify a resumed interview as any that took from 30 minutes on, but this is how
Knowledge Networks furnishes the data. Thirty-nine percent of KN respondents
completed the survey in 20 minutes or less.

Another issue is the level of respondent engagement—the question distractions and
interruptions come in. Looking at the KN survey, you begin to look at that average
completion time, and you begin to think a lot about distractions and interruptions.

Imagine someone trying to complete the survey and the cat is climbing up the drapes, the
dog is barking, and the kids are playing with matches, someone's at the door, the
telephone's ringing—all these things could be happening at once, who knows? Or, none
of those things could be happening and someone just decided on their own that they
would rather complete the survey later. Again, who knows? Anyway, the possibility of
distractions and interruptions is certainly there.

With the Hattiesburg survey, where the respondents were completing the survey in the
university computer lab, about the only possible distraction would be someone teaching
calculus across the hall and a respondent might decide that they would rather go learn
about the quotient rule. I don't think this happened, though.

Session III Panel Discussion on Internet Use

7

-------
A number of people in the KN sample had taken a lot of surveys—presumably they were
experienced—that could be good, it could be bad—Reed referred to this sort of thing as
the trained seal effect. Who knows? With the Hattiesburg study, it was a fresh sample—
they hadn't participated in any previous surveys, at least none that we had done. There
was also more item non-response in the KN survey than in the Hattiesburg survey. In the
computer-assisted survey we had practically no item non-response, whereas in the KN
survey there was a lot.

Did changes in features of the hypothetical sun lotion that Mark described alter
willingness to purchase it in a predictable way? Well, with a change of price, yes. As the
price went up, willingness to buy the stuff went down. How about extent of risk
reduction? In the Hattiesburg survey, in a between-respondent comparison, we got
higher willingness to pay for larger risk reduction, so there's an external scope test there.
With the KN survey, again in a between-respondent comparison, we got significantly
lower willingness to pay for larger risk reductions. What are possible explanations for
the difference in outcome? I mentioned the greater education level of the parents in
Hattiesburg. Maybe better-educated people are just in a better position to do these
surveys than less-educated people. We did a variety of tests to try to detect whether
education level had any bearing on the outcome of the extent of risk questions that we
asked, and the answer was "no." It was just that the KN respondents, in general, were
poorer at this than the Hattiesburg sample.

As far as resumed interviews in the KN sample, if you just took out all the people who
took 100 minutes or more to complete the survey, would the basic results change? The
answer is "no." Was there a greater level of engagement on the part of the respondents in
the Hattiesburg survey? Maybe—I don't know—but it is a concern. One thing I wish we
could have generated was some within-respondent evidence as to how people respond to
changes in risk.

Alan Krupnick, Resources for the Future

Wow—those are quite problematic responses to that survey of Knowledge Networks, and
I don't want this panel to become a referendum or a judgment of Knowledge Networks,
but it's probably worth saying why we mention Knowledge Networks so much. There
may be people here who don't understand that. The reason that Knowledge Networks is
so attractive is because they made an attempt through random-digit dialing to convert
people to their panel who were not internet users. They gave them this special
technology, webTV. You don't need a computer to take these surveys when you have
this technology, so it deals with the problem of non-internet-users.

We (Maureen [Cropper], Nathalie [Simon], Anna [Alberini], and myself) did a national
U.S. mortality-based survey in the year 2000 or so for our mortality risk valuation work,
which has been reported in a couple of different journals. I wanted to talk a little bit
about our experiences, particularly in regard to some of the responses I have after

Session III Panel Discussion on Internet Use

8

-------
listening to Shelby's presentation. Then I want to give a little advertisement for what's
going to happen at Resources for the Future in October.

So, we've had experience with both Knowledge Networks and Ipsos Reed, which is a
Canadian firm that does probability-based internet sampling but doesn't have the webTV
technology. First, going through the work on mortality risk valuation, we basically had
exactly the same setup, although different locations, as Shelby. In our Canada sample, it
was a random-digit-dialed sample of people in Hamilton, Ontario that came to a central
location to take the survey on a computer. Then later we did a national sample using
Knowledge Networks on webTV or the computer. We got extremely close results on
both of those surveys. Many of you in the audience have seen our bar graphs—almost
equal responsiveness to the bids, which were basically PPP-corrected, so they were
equivalent bids across the two countries. We had significant external scope effects. We
had very little item non-response. Maybe this can be explained partly by the fact that we
were using the panel in its early days—by the time Shelby got to it, it was rather old.

The one benefit that we saw from Knowledge Networks that you can't get easily from
these in-person, self-administered surveys at centralized locations is that you can pick up
infirm or immobile people—if they're in your panel or however you get them. That's
important to many health surveys, so we thought that was a benefit from our work
although I can't prove it. We also looked at the timing issues—these people who take
100 minutes or more, and so on. As Shelby mentions, we didn't find any effects on
timing.

So, let me go to our Adirondack survey. This was done by Knowledge Networks in New
York, so our sample of people was panelists from New York state, where we estimated
the willingness to pay for improvements in the Adirondacks, and it was set up with an
external scope test framework. What we did here is we used two different modes—an
RDD mail survey and a panel internet survey. We had Knowledge Networks do both of
these for us. The survey for the two was as identical as we could make it, given the
difference in mode.

So, we did a few things. The first is that we looked at the demographics comparing the
two modes to each other and comparing them to the census. We did pretty well. There
were some observable differences across various samples, which we corrected using
weighted regression. Differences in observables really don't cause any major problems.
Then we used a Heckman selection analysis on the panel internet survey using KN's
panel data, so we know from the panel who was exposed to the survey and had an
opportunity to take the survey but chose not to. We did the analysis with that group and
with the group that did take them, and we did find some groups less likely to respond to
our survey—women, minorities, and the lower-educated were less likely to respond—but
we didn't' find any statistical effect of the unobservable component of response on
willingness to pay. Of course, the limitation of this kind of analysis is that we did not
look further back in the chain to compare our results to people who chose not to be on the
panel. So, that's going all the way back to the beginning of the RDD effort, and we
weren't able to do that.

Session III Panel Discussion on Internet Use

9

-------
Finally, we compared the frame mode, the RDD mail results for willingness to pay to the
panel internet results for willingness to pay, and we found that they were quite similar—
there was no statistical difference between those two. For what it's worth, that's what we
found.

Finally, I just want to mention what we're going to be doing in October. We'll be
hosting an OPEI-funded workshop on the general topic of sampling bias. It's called
"Sampling Representativeness: Implications for Administering and Testing Stated-
Preference Surveys." We're going to bring in experts—some of the people on the pane
here—survey researchers, statisticians, cognitive psychologists, and government officials,
including Brian [Harris-Kotejin] and others to help better define the problems and work
toward a solution. Our motivation here is this linkage that OMB makes between low
response rates and therefore unreliability of the surveys. Our view is that you could have
an 80 percent response rate that doesn't guarantee representativeness, or you could have a
10 percent response rate that does. What we need to do is decide what our performance
measures are going to be and then what protocols we need to follow—and I know OMB
is interested in defining those kinds of protocols—to permit us to take advantage of
internet technologies that are out there to get these surveys done at low cost, quickly, and
flexibly to give all the advantages that Nathalie mentioned and not give that up on what
may be a false goal of lowering non-response rates. What we want to lower is sampling
bias, and that's a different thing.

Jon Krosnick, Stanford University

I'm a professor of communication, political science, and psychology at Stanford
University, and I'm delighted to have the opportunity to speak with you this afternoon. I
make my life, among other things, focusing on survey methodology. Increasingly lately
I've found myself obsessed with mode—doing mode studies for a variety of reasons and
trying to answer the general question of: What impact does mode choice have on survey
outcomes?

As some of you no doubt know, there are lots of different sources of error in surveys.
One is coverage error. That is, if we're doing a telephone survey, we'll fail to reach
households that have no telephone access at the moment that we call. There is non-
response error. That is, people of particular types choose not to participate and therefore
bias the sample composition. Interviewers make errors in reading questions and in
hearing and recording answers. Respondents make errors in interpreting questions and in
doing inadequate memory searches for relevant information—integrating or reporting, as
well. When you put all of this together, if produces what we think of these days as "total
survey error"—that's sort of the sum of all of these errors. In order to provide the most
accurate measurements from a survey, we want to minimize all of these various sorts of
error. My focus during my few moments today is on how mode can impact the sum total.

Session III Panel Discussion on Internet Use

10

-------
There are various ways to think about how mode choice does have impact. As I've said
already, if you decide to do a telephone interview, you have coverage error—period.

That doesn't mean your results will be different from the results you would get if you had
overcome that coverage error, but it does mean that if you ask people a question like "Do
you have working telephone service in your house?" you will not get the right answer
because of the method you used to contact people.

But, there are some other cases in which mode differences are less predictable, less
expected, and less anticipatable. Let me say from the start here, my discussion is going
to focus on probability samples only. As you've heard already, there are internet survey
firms offering, at fabulous prices, internet surveys provided from non-probability
samples. We have done work on non-probability samples, and we find consistently that
those samples are less accurate in the data that they produce, sometimes dramatically
inaccurate. I personally don't take them seriously for the kinds of work that requires
generalization to populations, so I'm not going to spend any time talking about that
today. What I am going to talk about very briefly [referring to slide] are the four primary
"contender" modes these days and the considerations or variables associated with these
modes that can help differentiate between them. I'm not going to go into great detail, but
we could think through how face-to-face interviews, versus telephone interviews, versus
paper-and-pencil questionnaires could differ in the rapport and trust that the respondents
feel they have in researchers, in the confidentiality they feel their responses can be
assured, the modeling of commitment that a researcher or an interviewer might provide
and become contagious with respondents, and so on. There are lots of these different
factors and 10 minutes is not adequate time to go through this theoretical analysis.

What I do want to do, though, is very quickly skate you across a set of mode comparisons
leading to the ones we care about most on the internet. First of all, comparing face-to-
face with telephone interviews, you know that in the late 60's to early 70's when
telephone penetration in households became essentially universal, the appeal of the many
practicalities of the telephone attracted researchers to that mode, especially the reduced
cost. The question that arises is: Was there any price paid by saving that money and
moving to the telephone and not having to ship interviewers around the country, being
able to supervise them closely, being able to complete surveys much more quickly, and
so on. [Dr. Krosnick then showed a slide that listed "all the studies that had been done
comparing face-to-face to telephone interviewing before we did our work, and showing
all the design flaws that they suffer from that prevent you from being able to make any
inferences, unfortunately, from them about the question we care about."]

So, we did a study that used three different national experiments—a data set collected in
1976, another one in 1982, and another one in 2000—conducting the same survey side-
by-side, random-digit-dialed telephone nationally as well as face-to-face with area
probability samples. I want to just show you, without going into great detail, that for the
full samples [again, referring to slide] there was more reporting error in the telephone
data than in the face-to-face data across the board. The data show that the real cost of
moving to the phone is for the least-educated respondents—they get hit the hardest by the
added cognitive burdens of a telephone conversation. In addition, the telephone

Session III Panel Discussion on Internet Use

11

-------
respondents complained more often about how long the interview was lasting, they
expressed more dissatisfaction with the length of the interview, they said that the
interview was "too long" more often, and, amazingly, their interviews were shorter than
those of the face-to-face respondents. Is it surprising that people feel rushed on the
phone?—maybe not.

Interviewers also rated the respondents on the phone as "less interested" in the interview
process and "less cooperative" with the response process, and we found that the
telephone respondents were more likely to distort answers in socially desirable directions
than were the face-to-face respondents, who presumably developed a sense of rapport and
trust with their interviewers more effectively. In addition, the telephone respondents said
they were more uncomfortable discussing sensitive topics, and the interviewers rated the
phone respondents as being "more suspicious" than the face-to-face respondents.

Okay, that was very, very quick, but you get the bottom line, which is that in this contest
face-to-face wins.

What about a competition between telephone and paper-and-pencil, as we move closer to
the internet case? In this case, this is a study that we did for NASA, funded by the
FAA—a study of airline pilots who fly you and me around on commercial airplanes. This
was using a survey project called the National Aviation Operations Monitoring System.
A field experiment was involved—licensed pilots were interviewed and they were
randomly assigned either to be interviewed by telephone or self-administered
questionnaires, and they were asked factual questions. We built into the experiment a
measure of the accuracy of answers, and what we found was that the telephone provided
substantially more accurate responses than the paper-and-pencil questionnaires did. So,
in this case when you take the interviewer out and leave respondents on their own, the
quality goes down. In general, the respondents forgot events they should have reported
more on paper than they did when they were walked through the questionnaire by an
interviewer on the telephone.

The respondents answering the paper-and-pencil questionnaire actually realized that their
answers were less accurate. When we asked them to rate how accurate the answers were
as descriptions of their experiences, they reported significantly lower confidence in the
accuracy of their answers. The real story here is this one: Whereas it took 27 minutes on
average for the respondents to complete the interview by telephone, it took only 16
minutes for the paper-and-pencil respondents to complete that very same questionnaire.
They rushed through the questionnaire; they overlooked events and by failing to report
them, compromised the accuracy of the data they provided. As a result, the winner in this
little "race" is the telephone.

Now we move, finally, to your favorite topic: telephone versus internet. So, paper-and-
pencil and computer modes seem pretty similar—no interviewer involved, just answering
questions on your own—maybe we should be worried about this competition, maybe we
should be pessimistic. What do the data say? Well, we have two kinds of data [again,
referring to slide]. One is a lab experiment, where we brought a group of respondents

Session III Panel Discussion on Internet Use

12

-------
into our lab and randomly assigned them either to complete a questionnaire on a
computer by themselves in a cubicle or to complete the very same questionnaire over an
intercom system, being interviewed orally by an interviewer down the hall. What we
found is, depending on which measure of validity we looked at, . . . large majorities of
comparisons showed statistically significantly higher validity for the computer than for
the oral interview and no statistically significant differences suggesting the oral interview
was superior to the computer. So, interestingly, we find here that the computer yields
more-accurate reports than the oral administration. Furthermore, in the computer case,
manipulating the order in which response choices were presented to people had no
meaningful impact on those answers—54 percent versus 51 percent. However, on the
intercom we found a very pronounced order effect, where we manipulated the order of
choices and it produced a big difference in the answers people gave.

Lastly [again, referring to a slide], the pressures toward social desirability were more
powerful on the telephone than on the computer. On the computer, White respondents
were quite willing to say they were in favor of decreased government help for Black
Americans, whereas being interviewed on the intercom the plurality of respondents said
they supported increased help for Black Americans instead.

So, what are our conclusions? Well, face-to-face beats telephone. Computer beats
telephone. Telephone beats paper-and-pencil. So, one possibility is that face-to-face
produces better data quality than computer, which produces better data quality than
telephone, which produces better data quality than paper-and-pencil. If this were true, it
would sort of be the case that you get what you pay for—the more expensive the method,
the higher quality the data. . . . We shouldn't over-generalize here, but I guess what I
would say is I think there's a lot of promise in the data I've shown you for the potential
of the internet mode to produce valid data. The question is: Can it be accomplished
effectively?

JR. DeShazo, UCLA

Given all the discussion about the benefits, I don't think I'm going to cover the benefits.
Let me briefly tell you what my experience has been in the context of four surveys and
then talk about sample selection correction, because following up on Alan's point, I think
what we do want to reduce is sample selection bias. We, entirely through the efforts of
Trudy [Cameron], did go back to the random-digit-dial stage and evaluate sample
selection bias for both opinions that were expressed and the propensity of being our final
samples for the first three surveys that we did through Knowledge Networks.

[goes through a series of slide that describe the surveys they did]

We were very much concerned that our estimates of willingness to pay would not be
representative of the U.S. population, and so Trudy began thinking about how to go about
correcting for that. . . . One of the problems in random digit dialing is to figure out who
chooses not to be recruited by Knowledge Networks, but the problem doesn't stop there.

Session III Panel Discussion on Internet Use

13

-------
Here's a summary of the process so you can get an idea of the magnitude of the problem:
There's the initial random-digit-dialed contact, at which time individuals can select out of
the sample if they're not recruited. They could be recruited by Knowledge Networks and
not profile—that is to say, not enter their panel at time "t." Assuming they enter their
profile at time "t," they may at time "t +1" select out of the panel and not be active and
thus not be available to us when we draw our sample. Then, of course, the final selection
stage occurs if they are not drawn randomly or otherwise by Knowledge Networks as part
of our estimated sample. What we wanted to do is explore differences and describe the
systematic selection out of our estimated sample as a function of a set of individual
characteristics. . . .

One of our surveys gathered data on public opinion with respect to whether the
government ought to intervene in environmental health and safety programs. One
concern of ours was "did the panel have a liberal bias?" and we thought we could get at
this question by focusing on this question about the appropriateness of government
intervention. A more fundamental question, given that we are interested in estimating
demand and peoples' willingness to pay is: Does this selection process lead to a non-
representative sample that is going to express a biased willingness to pay? The second
approach goes about estimating marginal selection probabilities, conditional selection
probabilities, and then allowing the marginal utility associated with the attributes of the
programs we're interested in the peoples' willingness to pay for to depend on the
propensity to respond to the survey.

Approximately half a million individuals were contacted by Knowledge Networks or one
of their subcontractors. We placed a restriction on our sample—we wanted adults over
24 years of age . . .there were 1600 individuals that were recruited for the sample. The
nice thing about the random-digit-dial information that we were able to obtain is that we
could match it with census data. This was not easy and it took a huge amount of time.
Basically, the way we did it is we used individuals' addresses and their telephone
exchange and Trudy developed an algorithm to associate the probability that that
individual in either that address or telephone exchange would be associated with a
particular census tract. Then she very cleverly developed a set of 15 orthogonal factors
plus using data on voting behavior—basically, these propensities to participate or to
persist in the sample. This was extremely laborious, so much so that it justified a paper
by itself (Cameron and Crawford). Let me say that there are three papers that are
available on our attempts at sample selection correction.

These 15 orthogonal factors explain 88 percent of the variation [unintelligible words]
characteristics across tracks, so this is a very robust selection model.

Given the limited time, let me just get to the conclusions. For the first analysis on the
question of liberal bias, whether or not we were obtaining an average representation of
peoples' opinions as to whether or not the government should intervene via
environmental health and safety programs—we found that there was basically an
insignificant point estimate of bias in the distribution of attitudes toward regulation. So,

Session III Panel Discussion on Internet Use

14

-------
there was no appreciable effect that resulted from selection on the response item of
interest.

In the second analysis, we did find statistically significant but very, very, very tiny effects
on the key parameters across respondents' propensities to persist in the panel, so much so
that they were, in the context of our willingness to pay estimates, insignificant—and I'll
stop there.

END OF SESSION III

Session III Panel Discussion on Internet Use

15

-------
Summary of the Q&A Discussion Following Session III

Mary Evans (University of Tennessee)

"It's my understanding of these panels, such as Knowledge Networks and Harris
Interactive, that if you submit a fairly small number of questions they may sort of
piggyback your questions onto a larger survey. I'm wondering, first, if that may explain
some of the differences in Shelby Gerking's experience and Alan Krupnick's experience
with Knowledge Networks in particular. Secondly, I'm wondering if anyone is aware of
any studies that look at the effect this kind of piggybacking has on results, whether
there's a systematic bias."

1st responder (Gerking or Krupnick)

He responded, "The Knowledge Networks study that we did was not piggybacked on any
other" and added that, in fact, it was sufficiently long that Knowledge Networks
determined that a time constraint should be imposed on it—they wouldn't piggyback it
with another one.

2nd responder (the other one)

He added "and that's the same with ours. Ours was about 30-32 minutes on average, as
well, and there was no piggybacking, so that won't explain it."

3rd responder

This person clarified that there are two kinds of piggybacking. One is when your
questions go first before other people's questions, in which case there's no impact so who
cares? The other possibility is that your questions get added to the end of somebody
else's, and this creates two issues. He explained, "One is that your questions are now
appearing when respondents are more fatigued. Secondly, prior questions have been on
particular topics and have activated thinking in particular directions. There's plenty of
literature suggesting that fatigue and the content of prior questions can indeed influence
answers to later questions, so there's every reason to believe that that's problematic." He
continued on saying, "On the other hand, there's absolutely nothing unique to Knowledge
Networks or Harris Interactive in piggybacking, because if you take Alan's survey or any
survey that I've done, all of the questions at the end of the questionnaire are sort of
piggybacked on all the questions at the beginning of the questionnaire. So, anything that
comes late in the questionnaire could be influenced by what came earlier, just as in any
other case." He concluded by saying that although it's not unreasonable to ask if there's
impact of early questions on late questions, but it's not unique to those firms.

Here, the questioner made an unintelligible follow-on comment.

3rd responder

This person replied, "Absolutely," and he said he would repeat the comment so that
everyone could hear it. He summarized the comment, saying "that it would make a
difference on the results of willingness to pay for asthma if the prior questions were about
cell phones versus whether they were about asthma medications." He went on to say that
he doesn't think there's any doubt about that, "and it could very well be true that your

Session III Q&A Summary

1

-------
early questions in your questionnaire can influence the later ones regarding asthma, too.
In particular, there's one very well documented danger: If you ask early questions on
willingness to pay for cleaning up pollution in the ocean, people will feel as if they have
less disposable money available by the time they get to the asthma questions. We know
of that problem, and that will occur in any questionnaire as a result."

4th responder

"There is a related issue, as well, which is the expectations of respondents when they first
begin to take the survey. If they're seasoned panelists, they may be used to taking
surveys where the questions are similar to: Would you open an account with thus-and-so
bank if the account had these features and we threw in a free pizza? That's one kind of
question. Or, to go along with Jon Krosnick's presentation: Would you vote for
President Bush if he stood for election today?—Yes, No, Don't know. When you follow
such a question with one such as: Now, assume you're an asthmatic—would you pay for
this or that type of medication to control these or those kinds of symptoms?—then you're
just increasing the level of difficulty for these questions. If somebody was not expecting
to see something that difficult, maybe that would be a flag."

Trudy Cameron, (University of Oregon)

Dr. Cameron said she just wanted to acknowledge "the remarkable cooperation" that she
and Dr. DeShazo received from Knowledge Networks in doing the non-response study,
"going all the way back to the original RDD contacts." She specifically acknowledged
the hard work of Mike Dennis and Rick Lee as well as a consultant, Dale Culp. She
added, "If I had been them, I would have been very much more nervous about the
downside of this enterprise. All of us, collectively, heaved a sigh of relief when things
turned out pretty well, . . . but we put them way out on a limb, and we're very grateful
they did cooperate in providing that data." Adding that the exercise has been done as
much "at arm's length" as possible, she closed by saying that she is "comfortable that
what we're finding is the right stuff."

. /. R. DeShazo (University of California, Los Angeles)

Dr. DeShazo added, "These are firms—and they'll respond if we tell them what we need
and they have enough lead time and planning time. One of the challenges Knowledge
Networks had was that they hadn't thought to keep track of all of their random-digit-
dialed contacts. They had to go back and recover that and were uncertain as to whether
or not they could. So, whether we're expressing professional standards for data quality or
responding to OMB, I think that there's a market out there for data collection. If we
communicate our needs clearly, we're large enough demanders of the product that they
are going to be responsive."

Unidentified speaker

"We use Harris Interactive to get access to their chronic disease panel for surveying
patients. I've never tried to do a general population survey with them—I've done a
couple with Knowledge Networks. One of the marketing strategies that Harris uses, that
I believe they have implemented subsequent to Jon's study six years ago, is a fairly
sophisticated propensity weighting scheme, in which every other month they conduct a

Session III Q&A Summary

2

-------
random-digit-dial survey and an internet survey from their panel and then attempt to
devise a weighting scheme to match not only the demographics but the responses to
certain attitude questions, particularly attitude questions that screen well for people who
take internet surveys. Jon, are you aware of this sheme? Does it make sense to you? Do
you think it's fixing some problems? Knowledge Networks' argument is that we can
match the demographics but it doesn't really necessarily match people who are going to
join a panel and answer survey questions every week."

Jon Krosnick, (Stanford University)

Saying he was happy to comment on this, Dr. Krosnick responded, "The Harris
Interactive propensity weighting scheme is proprietary—they will not describe how they
do it to anybody—and they did have it in place at the time that we did our 2000 study,
which I showed you. We were provided with the proprietary propensity weights, and
when we analyzed the Harris data, both with the weights and without the weights, we
found that it did not change the substantive results at all—it didn't change the means or
the distributions of variables. What it did do was increase the standard errors of the
estimates. The reason for this is because when we looked at the weights, there were some
as large as 20 or 30 and some as small as 0.1 or 0.2. So, the weights are dramatic and
they didn't have any real impact on the results that we looked at. As it turns out, Harris
will not normally reveal the questions that they use in those parallel surveys to develop
the weights, but they actually accidentally sent us the questions. So, having seen the
questions, I can tell you that I'm not even slightly surprised that they don't do anything
helpful."

Dr. Krosnick continued, "The more recent study we've done, which I haven't mentioned
to you, is one in which we compared the same questionnaire administered by random-
digit-dial telephone, Knowledge Networks, and six other firms that use volunteer
samples, some of which do weighting by quotas on demographics and one of which
provided proprietary propensity weights. We found the same thing—the propensity
weights didn't change anything, and the volunteer samples were substantially less
accurate. So, my results that I showed you earlier and these new results are not focused
on demographics. The vast majority of our results comparing the reliability and validity
have to do with substantive measures of attitudes, beliefs, behaviors, and so on. In cases
where you can compare factual matters—like whether people have a driver's license or
not, whether they have a passport or not—and other figures where there are official
numbers to compare to, the probability samples from telephone and from Knowledge
Networks were equivalently accurate and the volunteer samples were notably less
accurate."

In clarifying how the panels process a shorter survey, Dr. Krosnick stated, "Their
panelists are answering questions every week, so they'll add your question to a survey
that's already going to go out anyway. How much does this cost them to add one more
question?—nothing—get your $500—fabulous.

Another responder

Session III Q&A Summary

3

-------
"I just wanted to mention with respect to the cost figures that were presented before—we
might have received a bulk discount. Our experience, just for the benefit of future
negotiations, was that the total cost for Knowledge Networks was less than $45 an
observation for a 30-minute survey."

Jon Krosnick

"Definitely a bulk discount."

Unidentified questioner

"Does that include university overhead?" When a responder replied, "No, it doesn't," the
questioner said, "Okay, that's part of the difference."

Jon Krosnick

Dr. Krosnick added, "There's also a very subtle but interesting issue on overhead for
those who care about this. The universities make a distinction between subcontracts and
service purchases. If it's a subcontract, you only pay indirects on the first—let's say
$20,000; if it's a service purchase, you're paying indirects on everything. You definitely
want to negotiate with your university to make it a subcontract so you don't pay more
indirects than necessary."

James Hammitt, (Harvard University)

Dr. Hammitt said he wanted to get some of the panelists' perspectives on a question
related to the cost issue. He continued, "When I first got involved in internet surveying,
it seemed to me that compared with phone surveys the fixed cost of setting it up might be
high but the marginal cost per respondent would be very much lower because you don't
need the live interviewer. With something like a Knowledge Networks panel, there's
obviously a cost to maintain the panel and an opportunity cost to use it up. Is it right that
the marginal cost per respondent will tend to be much lower with internet than with
phones, for example? It seems to me that that would have implications for how we
design surveys, because, as Jon has commented, there's a concern that if you ask people a
lot of questions they get tired out and the responses toward the end may not be very good.
However, if the marginal cost per respondent is low, we should just have very short
surveys of a very large sample, whereas with phone surveys there's so much cost
involved in getting somebody on the line who is willing to answer your questions that we
tend to go for a longer interview with them."

Jon Krosnick

Dr. Krosnick replied, "I think that's definitely misleading. Basically, when you think
about fixed costs of telephone interviewing—you have to hire a staff, you have to train
the staff, you have to have supervisors, you have to have facilities and machines and all
that—then once you get them in there, if they keep making more phone calls obviously
making one extra phone call doesn't require all that much more staff time. Similarly,
Knowledge Networks has to invest a bunch of money in recruiting a panel and then
equipping the panel and paying them incentives and keeping them all going every week.
My guess is that adding another respondent to the panel is actually considerably
expensive—you have to make recruiting phone calls and get them signed up and send

Session III Q&A Summary

4

-------
them the equipment and all that—and you have only so many people in your internet
panel. So, when you say that adding one extra respondent doesn't increase the cost very
much, that's sort of true, but the whole fixed cost scheme is pretty burdensome, I think.
You might say that you don't have to make a new phone call. Adding that marginal
respondent on the internet case isn't that expensive if you weren't going to use them
anyway that week, but it's not clear that Knowledge Networks doesn't want to use them
anyway."

Kelly Maguire, (U.S. EPA)

Addressing Brian Harris-Kojetin, Dr. Maguire stated that he had mentioned that "many
federal governments are moving toward using mixed modes," and she said she was
wondering whether any of the panelists have experience with using mixed modes. She
added that one of her concerns is that "when you start to use multiple modes within one
research study, you introduce other biases that become more problematic than say the
non-response bias that you're trying to correct in the first place."

Alan Krupnick (RFF)

Dr. Krupnick responded, "I mentioned in my remarks that we did use mixed mode—we
used a mail survey and the Knowledge Networks internet survey." He acknowledged that
the two surveys were "not exactly the same" due to the "issues you have to confront in
switching these modes"—but they were pretty close. He added, "Maybe we were
fortunate to have our willingness to pay estimates not be any different across these two
modes. If they had been different, then we would have faced the issue of trying to
explain why, but we didn't have to do that."

END OF SESSION III Q&A

Session III Q&A Summary

5

-------
Morbidity and Mortality: How Do We Value the Risk of

Illness and Death?

PROCEEDINGS OF SESSION IV: VALUING MORBIDITY AND MORTALITY:

PESTICIDES AND TOXICS

A WORKSHOP SPONSORED BY THE U.S. ENVIRONMENTAL PROTECTION
AGENCY'S NATIONAL CENTER FOR ENVIRONMENTAL ECONOMICS AND
NATIONAL CENTER FOR ENVIRONMENTAL RESEARCH

April 10-12, 2006

National Transportation Safety Board
Washington, DC 20594

Prepared by Alpha-Gamma Technologies, Inc.
4700 Falls of Neuse Road, Suite 350, Raleigh, NC 27609

ACKNOWLEDGEMENTS

This report has been prepared by Alpha-Gamma Technologies, Inc. with funding from
the National Center for Environmental Economics (NCEE). Alpha-Gamma wishes to
thank NCEE's Maggie Miller and the Project Officer, Cheryl R. Brown, for their
guidance and assistance throughout this project.

DISCLAIMER

These proceedings have been prepared by Alpha-Gamma Technologies, Inc. under
Contract No. 68-W-01-055 by United States Environmental Protection Agency Office of
Water. These proceedings have been funded by the United States Environmental
Protection Agency. The contents of this document may not necessarily reflect the views
of the Agency and no official endorsement should be inferred.

-------
Table of Contents

Session IV: Valuing Morbidity and Mortality: Pesticides and Toxics

Session Moderator: Jin Kim, U.S. EPA, Office of Pesticide Programs

ORD Activities With the National Children's Study and National Health and
Nutrition Examination Survey

Montira Pongsiri, U.S. EPA, National Center for Environmental Research

Use of Contingent Valuation to Elicit Willingness-to-Pay for the Benefits of
Developmental Health Risk Reductions

James K. Hammitt and Katherine Von Stackelberg, Harvard Center for Risk
Analysis, Harvard University

Parental Decision Making About Children's Health

Alan Krupnick and Sandra Hoffmann, Resources for the Future; Victor
Adamowicz, University of Alberta; and Ann Bostrom, Georgia Institute of
Technology

Value of Reducing Children's Mortality Risk: Effects of Latency and
Disease Type

James Hammitt and Kevin Haninger, Harvard Center for Risk Analysis, Harvard
University

Discussant: Lanelle Wiggins, U.S. EPA, National Center for Environmental
Economics

Discussant: F. Reed Johnson, Research Triangle Institute
Questions and Discussion

-------
t Integrating Economic and
Behavioral Questions into
National Health Surveys

US EPA NCER/NCEE Workshop
11 April 2006

1

National Human and

Nutrition Examination
_|_Survey (NHANES)

¦	Nationally representative, continuous,
longitudinal study

¦	Collects data on demographics, health
status, diet and chemical exposure

¦	2-year cycle, 5000 participants per year

¦	Oversampling of special populations

¦	Value: individual data can be linked to
objective, scientific health measures for
hypothesis testing

2

-------
Opportunity and Value-

I Added

¦	Improved estimates of the value of reduced
morbidity and mortality to justify
environmental health policies

¦	Better understanding of how people assess,
perceive and respond to risk

¦	Improved analyses of the accountability of
regulatory decision-making

¦	Design of targeted policies to minimize risk
or support risk-averting behavior

NCER Proposal -

Costs of asthma medication

¦ Costs of asthma medication:

- You have already said that you use one, or
more, of these medications to control your
asthma. How many months, out of the past
three months, did you need to take this
medication everyday or almost everyday?

4

-------
NCER Proposal —

Behavioral response to poor air quality

Some people change their activities on days when air pollution
is bad, while others go ahead with their activities as planned.
On days in the past year when you thought or were informed
air quality was bad, did you ever do anything differently,
provided you had the choice, such as:

¦	Restrict the amount of your time outside?

¦	Exercise indoors instead of outside?

¦	Choose less strenuous activities?

¦	Cancel activities?

¦	Avoid areas with heavy traffic?

¦	Take medication?

¦	Close windows of your house?

¦	Stay indoors?

¦	Did nothing differently

National Children's Study

¦	Longitudinal, 21yr follow up study of
100,000 children from birth to adulthood

¦	Data to be collected on physical, biological
and psychosocial environments, as well as
exposure to chemicals

¦	Hypotheses frame the study survey

¦	Adjunct studies can be proposed

¦	Funding for FY07 uncertain

-------
Next Steps

RFA on analysis of NHANES data

Reproductive/developmental
outcomes, air pollution and CVD,
drinking water contaminants and GI
llness

Other national health surveys

7

-------
Use of Contingent Valuation to Elicit Willingness-to-Pay for the Benefits of
Developmental Health Risk Reductions

Katherine von Stackelberg
Center for Risk Analysis, Harvard University
kvon@hsph. harvard. edu

and

James K. Hammitt
Center for Risk Analysis, Harvard University
IDEI and LERNA-INRA, Universite de Toulouse
jkh@harvard.edu

March 2006

-------
Abstract

We report several contingent valuation surveys to elicit willingness-to-pay for risk
reductions associated with decreases in exposure to a chemical, PCBs, in the
environment. We also develop Quality Adjusted Life Years (QALYs) from the survey
using either standard gamble or time-tradeoff elicitation methods to explore the
relationship between QALYs and willingness to pay (WTP), and to develop QALY
weights for subtle developmental effects. The results of the contingent valuation are
designed for incorporation into an integrated risk model to demonstrate the economic
impact of risk reductions. Respondents showed a positive and proportional relationship
between decreasing the risk of a 6-point reduction in IQ and WTP. Socioeconomic
variables were not statistically significant predictors of WTP, while behavioral variables
were strongly predictive and statistically significant. The range of mortality risks that
respondents would accept on behalf of their (hypothetical) 10-year-old child is 2 in
10,000 to 9 in 1,000 per IQ point, and WTP per IQ point is $466 (95% confidence
interval = $380, $520). QALY weights elicited via time tradeoff (reduction in life
expectancy) were statistically significantly different from QALY weights elicited via a
standard gamble. Respondents who answered questions about ecological endpoints first
were willing to pay a small additional amount when asked about human health effects,
but those respondents who answered questions about human health endpoints first were
not willing to pay any additional amount when subsequently asked about ecological
effects. WTP models demonstrate the importance of obtaining behavioral and cognitive
information from respondents when eliciting WTP and in tests of sensitivity to scope.

-------
1. Introduction

Potential health effects resulting from exposure to environmental chemicals can
range from severe terminal illnesses such as cancer to milder, systemic illnesses. One
category of effects that is receiving increased attention includes developmental and
reproductive effects, such as reduced fertility, low birth weight, genetic defects, and
cognitive deficits. The policy implications of these exposures have yet to be realized, in
part because the relationship between exposure and effects is not well quantified, and in
part because there is a dearth of data and information with which to quantify the benefits
of risk reductions associated with exposure to chemicals that exert these kinds of effects.
One such chemical, polychlorinated biphenyls or PCBs, contribute to the existence of fish
consumption advisories in virtually every state, indicating that this exposure has
important implications for public health. Other contaminants, such as mercury and lead,
also pose developmental risks.

Cannon et al. (1996) conducted a scoping study to evaluate the literature and data
available with which to quantify the value society places on avoiding potential effects
from in utero exposures to chemicals. Their primary finding was that there are very few
existing studies with which to quantify the monetary (or other valuation metric) of these
effects. Cost of illness techniques can be used to quantify the impacts of some birth
defects, but these would be restricted to fairly severe outcomes requiring ongoing
treatment and attention. For other, more subtle effects, such as mild cognitive deficits,
cost of illness and other related techniques are inadequate for capturing the range of costs
and for estimating welfare measures. In addition, the authors acknowledge that existing
cost of illness analyses related to the costs associated specifically with low birth weight (a
very nonspecific effect in terms of the relationship between exposure and outcome) do

-------
not reflect the total costs associated with the occurrence of these endpoints (Cannon et
al., 1996).

Stated preference methods have been used frequently for the evaluation of risk
reductions related to mortality (Hammitt and Liu, 2004; Hammitt and Graham, 1999) to
obtain estimates of the value of a statistical life (Alberini, 2005), and increasingly also to
value morbidity endpoints (Dickie and Gerking, 2002; Van Houtven et al., 2003, 2004;
Krupnick, 2004). Fewer studies have evaluated potential morbidity effects for risks and
exposures to children, which generally must be evaluated by parents (Dockins et al.,
2002). While imperfect, these methods provide policy makers with information on how
the general public might trade-off income against reductions in the risk of specific health
effects. The results of the surveys presented here contribute to the growing literature on
the relationship between WTP and reductions in risk of mild developmental delays.

2. Survey Design and Development

The surveys were designed over a one-year period and involved several informal
pilot surveys, focus groups, and a pretest. From the onset, the surveys were designed to
be administered over the Internet using a professional survey firm, Knowledge Networks.
The research goal was to evaluate whether a CV might provide a feasible method for
obtaining economic values for endpoints consistent with how they are expressed in a
typical risk assessment framework (drawing from the experience of the lead author at an
actual Superfund site) and explore how people respond to questions regarding potential
effects to children and wildlife as a result of exposure to a specific chemical in the
environment. To that end, there were numerous open-ended questions for which
respondents were invited to provide comments as they progressed through the surveys.

-------
These open-ended responses provide important insights into respondent motivations and
thinking short of actually sitting with the respondent.

The primary objective of the surveys was to elicit an approximation of the
monetized loss in utility consistent with economic theory experienced by respondents
resulting from potential effects associated with exposure to PCBs. Another objective of
the surveys was to measure WTP for risk reductions, consistent with the results that risk
assessments generate. The surveys were designed so that members of the general public
could follow and understand the issues, and the surveys asked various questions
throughout to gauge what respondents already knew (or thought they knew) concerning
chemicals in the environment and how they felt, in a general sense, about exposure to
chemicals (e.g., whether they thought it was a serious issue, or even feasible that the
kinds of effects described in the survey could really occur). The surveys are based on a
generic, non-specific site (although there are numerous actual PCB-contaminated
freshwater systems across the United States and it is likely that there is at least one
system in the general area in which the respondent lives); nonetheless, the surveys were
designed to be plausible and the payment vehicle realistic and believable.

Respondents to the survey are first told that government officials in their State are
responsible for allocating resources and are interested in individual opinions to inform
potential policies. The first question asks respondents to rate the importance of several
issues, including reducing crime, cleaning up the environment, improving education,
reducing taxes, protecting State waterways, improving library services, reducing air
pollution, and providing additional security at public events. The second question asks
respondents to consider whether current State budget allocations should be reduced or

-------
increased, keeping in mind that overall expenditures cannot be increased without an
increase in revenue. Respondents are reminded that State policy makers are responsible
for allocating resources, and that people may feel differently about these allocations
depending on their own beliefs and knowledge. Respondents are informed that State
policy makers are interested in learning how taxpayers feel about specific issues.

The survey then proceeds to set up the specific valuation question, which involves
the potential effects of a specific chemical (PCBs - we ask "have you ever heard of
PCBs?") in a large, unnamed freshwater system in the state in which the respondent
resides. This system is contaminated, and the company or companies ostensibly
responsible went out of business some years ago. Therefore, the State is contemplating
setting up a special "cleanup" fund to be funded through a one-time increase in the State
income tax.

We chose a payment vehicle that calls for a one-time increase in the State income
tax, to be kept in a fund earmarked for a cleanup remedy for the (unnamed) freshwater
system. The question states that the risk will decrease if the cleanup is conducted if the
income tax is raised by the bid amount for all, not just for the respondent (Johansson-
Stenman, 1998), which has been shown to generate values consistent with economic
theory. However, not all States have an income tax, and this was not explicitly
acknowledged. Another format might be to specify an increase in the property or local
tax for those States without an income tax; however, for the sake of consistency across all
respondents, we chose the income tax payment vehicle. The cleanup is described as
occurring over several years, and the survey also states that even after cleanup is
complete, it will still take several years for the wildlife receptors to recover. In addition,

-------
the risks will never go to zero. Respondents are presented with an initial bid randomized
from a bid vector ranging from $25 to $400. If the respondents agree to the initial bid,
they are presented with a bid that is double the first bid (if they agree to $400 initially,
then they are asked if they would be willing to pay at $800). If respondents do not agree
to the initial bid, then they are presented with a bid that is half as much ($10 if they did
not agree to $25 initially).

A particular issue that arises with double-bounded CV estimates from the
literature is a failure to achieve consistency (Hanemann, 1991; Hanemann and Kaninnen,
2001; McFadden and Leonard, 1993). We used a double-bounded dichotomous choice
(Hanemann, Loomis and Kanninen, 1991) which has been shown to substantially
increase the statistical power of the WTP estimate, at the expense of a downward bias in
the estimate because the second response is not incentive-compatible (Carson et al.,
2003). There is evidence that in some cases, responses to the second bid are inconsistent
with responses to the first bid. Some authors (e.g., Alberini, 1995) have shown that
pooling the responses to the first and second bids leads to some bias in the coefficient
estimates, but a gain in efficiency.

The bid vector for the second part of each survey (except combined) takes as its
starting point the next highest bid that was agreed to in the first part of the survey. One
could randomize the bid vector, but true randomization could lead to a bid being offered
for the combined valuation that would be less than what a respondent already agreed to
for an individual endpoint. One could randomize the bid amount offered for the combined
endpoints starting with the bid amount just above what had already been agreed to, but
that isn't true randomization. Therefore, we decided to offer the next highest bid

-------
following the one already agreed to (except in the case where a respondent said No-No to

the first bid: in that case, we randomized the combined bid as well). Table 1 shows the

relationship between the bid amounts for just the individual endpoints in the first part of

each survey and the bid amounts for the combined total across both endpoints.

There are a series of motivation and "confidence" questions, including:
D6. Thinking back on your responses for the tax you'd be willing to pay when thinking
about the potential effects ofPCBs on humans, how confident would you say you
were about whether you would be for or against this referendum on a scale of 1 to
5 where 1 is "Not confident at all" and 5 is "Very confident"?

The next set of questions asks about the confidence in responses for the endpoints

individually and jointly (Conf.Human; Conf.Total). Another question asks whether

respondents feel they can separate ecological and human endpoints in the valuation

question. Another set of questions asks about familiarity with PCBs, concern about

chemicals in the environment, and whether the respondent believes that PCBs really can

cause these effects in humans and animals (risk.baby; risk.wldlf; ChemConcern;

PCBConcern). Finally, respondents are asked to rate their trust on a one to five scale

concerning the information they receive from a number of sources, including different

web sites, print media, and television.

2.1. Endpoint Selection

Health effects resulting from environmental exposures can be acute (immediate)

or chronic (longer term). Acute effects can often be ameliorated if the source of the

exposure is removed (e.g., asthma attacks as a result of air pollution), while chronic

effects by definition tend to extend beyond the period of exposure (e.g., the asthma itself,

or the kinds of developmental effects explored here). In addition, with chronic effects,

there can also be a latency period (e.g., cancer, liver disease and other diseases that might

-------
not reveal themselves until long after exposure has ceased). The bulk of the WTP studies
found in the literature are for respiratory exposures (Van Houtven et al., 2003 provide a
meta-analysis of 136 studies) leading to episodes of asthma or angina attacks. This study
is designed to evaluate willingness to pay for a subtle effect (in humans) that occurs with
a fairly large probability (20% chance if exposed) relative to typical cancer risks at
Superfund sites.

The weight-of-evidence for a relationship between in utero polychlorinated
biphenyl (PCB) exposure and developmental outcomes has been well established and
continues to grow (Schantz et al., 2003). However, as with most epidemiological studies,
discrepancies exist among measures of exposure and the strength of the relationships
between the measures of exposure and developmental outcomes. Some of those
discrepancies are attributable to differences in analytical methods, particularly in older
studies (Longnecker et al., 2003) that had higher detection levels and less sophisticated
quantitation techniques. Both epidemiological as well as animal studies demonstrate
statistically significant increases in developmental delays and effects with increasing
maternal PCB exposure (Jacobson and Jacobson, 2002b; Jacobson et al., 2002; Levin et
al., 1988; Schantz et al., 1989, 1991; ATSDR, 2000). These effects can be seen in
newborns as measured by the Bayley Scales of Infant Development to older children,
measured either directly in terms of IQ or from other, related tests.

In terms of potential developmental effects, it is the in utero exposures that have
been most implicated in terms of effects (Jacobson et al., 1999; Jacobson and Jacobson,
2002b). Several studies have shown that although absolute doses of PCBs may be higher
during breastfeeding due to mobilization of PCBs stored in maternal lipid, the protective

-------
effects of breastfeeding itself together with other factors (e.g., nurturing home
environment) potentially ameliorate the detrimental effects of PCBs. The children who
showed the most statistically significant dramatic developmental delays were those
exposed in utero and who were not breastfed. Breastfeeding may therefore be protective
against developing these effects even if maternal body burdens are relatively high
(Jacobson et al., 1999; Jacobson and Jacobson, 2002a).

However, regardless of the exposure issues, there is a substantial body of
evidence that show declines in various cognitive responses across both human and animal
studies (summarized in EPA, IRIS, www.epa.gov/IRIS/; ATSDR, 2000), typically as a
result of in utero exposures. Much of our understanding of the implications of slight
declines in cognitive ability across a population is based on work done relative to lead
exposures (Schwartz et al., 1985; Schwartz, 1994). The research conducted in this area
shows that slight declines in IQ which are difficult to detect in individuals and which may
or may not lead to noticeable adverse effects on an individual basis are significant on a
population level in terms of a population shift in IQ. Other cognitive effects include other
kinds of developmental delays such as declines in reading comprehension to levels below
grade level, low scores on analytical tests and tests of simple math problems, and
behavioral responses.

The risk reductions used in the surveys are based on the results from Jacobson et
al. (2000) who present a linear relationship between lipid-normalized breast milk
concentration of PCBs and outcomes including a 6-point reduction in IQ and a 7-month
deficit in reading comprehension as evidenced by scores on the WISC-R at eleven years
for the Michigan cohort..

-------
2.3. Risk Reduction and Tests of Scope

Sensitivity to scope can take several forms. Typically, these are referred to as
regular embedding, (part-whole bias), and perfect embedding, or sensitivity of WTP to
the stated risk reduction. There are two "part-whole" aspects to these surveys: one is
within an endpoint, and the other is across endpoints. The human health endpoint doesn't
have quite the same part-whole property as the ecological version of the survey since the
potential human health effects of in utero exposures to PCBs include a panoply of
developmental effects, all or some of which may or may not occur. Indeed, as stated in
the survey:

"Studies involving children exposed while in the womb to PCBs have
shown that these children perform less well on a variety of developmental tests
throughout childhood. Government officials are interested in knowing whether
you would be willing to pay a tax to remove the source of the PCBs for the benefit
of protecting children exposed in the womb. Children that have been exposed to
PCBs have been shown to have slightly lower IQ than average children, read at
slightly below grade level, and are less able to perform simple math problems.
The chemical doesn't cause the exact same effects in every child, but it does cause
some effect in every child."

However, IQ does encompass general intelligence while reading comprehension
is but one component of intelligence, allowing us to explore differences and/or
similarities in the way respondents consider IQ versus reading comprehension as
endpoints. Reduction in IQ as an endpoint has been well-studied in the literature
particularly relative to exposures to lead and mercury. However, in terms of
developmental endpoints, there is enough interindividual variability in IQ that makes an
endpoint such as reading comprehension, which doesn't vary as much across repeated
tests of any one individual, potentially more interesting in terms of valuation.

-------
There has been increasing discussion in the CV literature concerning the effect of
the placement of a particular good or endpoint within a valuation sequence and the
influence that has on respondent valuation (Carson and Mitchell, 1995; Diamond, 1996;
Bateman and Willis, 2001). Different WTP estimates are obtained depending on the order
in which the benefits are presented, and additionally, the summation of the individual
WTP values is often not the same as the overall WTP obtained without specifying
individual endpoints. This is the issue of embedding, or part-whole bias, across
endpoints. We explore this by administering three different versions of the survey. Two
versions ask exactly the same set of questions except in opposite order (HHFirst,

Ecofirst), and one survey asks only about the combined set of potential effects and risk
reductions (human and ecological) to evaluate adding-up properties.

We evaluate perfect embedding by randomizing two different risk reductions for
each endpoint across respondents as shown in Table 2. That is, each respondent sees only
one risk reduction per developmental and ecological endpoint, but there are two risk
reductions for each endpoint randomized across each subsurvey. We focus a number of
the analyses on the risk reduction coefficient across surveys and endpoints.

2.4 Questions Related to Motivation

The survey contains a number of questions related to respondents' knowledge and
beliefs regarding chemicals in the environment, PCBs in the environment, potential
effects of PCBs, and trust in different sources of information (e.g., industry scientists,
media, and academia). The survey contains several follow-up questions designed to elicit
motivation for agreeing to a particular bid. One question asks respondents to rate on a
scale from not important to very important the specific reasons why they might be willing
to pay to reduce potential risks to unborn children. We asked this follow-up question if

-------
the respondent answered N-Y, Y-N, or Y-Y (e.g., they agreed to any offered bid). The
reasons include:

B5. People have lots of different reasons for voting for the program. Please rate
the importance of the following reasons why you might vote for the program:

I'm worried about the potential risk to my own unborn children
I'm worried about the potential risk to unborn babies generally
I support a cleanup no matter what the risk might be (I don 7 like the idea of
chemicals in the environment generally)

Some other reason: please specify

Likewise, for those respondents who answered N-N and were not willing to pay
any amount, we asked the following:

D4. The State is interested in knowing why you would vote against the program. There
are lots of different reasons why you might vote against the program, like it just
isn 7 worth that much money, or it would be difficult for your household to pay
that much even though you support the program, or you are opposed to dredging
as an alternative. Or there might be some other reason.

Isn't worth the money	1

Difficult for my household to pay	2

Don't believe the cleanup would work... 3

Some other reason, please specify: 	4

2.5 Quality Adjusted Life Years

All respondents see a set of questions designed to elicit utility weights for mild

cognitive effects using either a standard gamble or time-tradeoff question format. Utility

weights are typically elicited using a QALY index derived by questioning respondents

about specific health states. The QALY index is defined as the product:

q-T	(1)

where:

-------
q = a numerical gauge of the quality of the health index on a scale of zero to one
(typically zero is the health state equivalent to death and one is perfect health, although
values less than zero are possible for "worse than death" health states)

T = duration of health state

In one set of questions, respondents are asked to assume that they have a 10-year
old child with a mild cognitive deficit, and are then offered either a standard gamble (SG)
or time tradeoff (TTO) question concerning the mortality risk they would accept on
behalf of their child for a perfect cure. These two approaches, SG and TTO, are the two
primary methods used in the literature to elicit QALY weights (Gold, 1996).

The standard gamble offers the respondent a choice of a mild cognitive deficit in
the child (either the reduction in IQ or reading comprehension deficit) for the remainder
of the child's life (assumed to be 60 years) in comparison to a lottery of perfect health for
that duration versus death. Respondents are asked about the probability of death that
would be considered equivalent to a lifetime with a mild cognitive deficit. Table 3 shows
the specific probabilities which range from 2.5 in 10,000 to 40 in 10,000.

The other elicitation scheme uses time tradeoff. Under this approach, the survey
asks about years of longevity in perfect health a respondent would give up on behalf of
the (hypothetical) 10-year old child to avoid a mild cognitive deficit that lasts a lifetime
(60 years assuming a lifetime of 70 years). To correspond to the probabilities given
above, the question asks about weeks of longevity that respondents would be willing to
give up on behalf of an exposed child as shown in Table 3.

The question follows the same double-bounded dichotomous choice format as for
WTP. That is, respondents are shown a time-tradeoff or probability of death, and if they
respond "Yes", the followup questions asks about a larger number of weeks, or higher

-------
probability of death. If they respond "No," the number of weeks, or probability, is cut in
half. Respondents are shown a visual aid for the probability based on "dots" (Corso et al.,
2001). The QALY weight that is assigned is equal to 1 - mortality risk interval agreed to
by an individual respondent. The relationship between WTP and QALYs is given as:
WTP = 0O* {Aq* At/1 + s	(2)

where:

Aq = change in health related quality of life

At = specific time period applicable to the quality weight

In this survey, respondents are asked to assume they have a 10-year-old child with
the cognitive deficit, and what risk would they be willing to assume for this hypothetical
child for a perfect cure. In the analysis, we assume that the child would live to be 70
years, so the duration of this health state is 60 years. In theory, WTP should increase
proportionally relative to the gain in QALYs, which is testable under the hypothesis that
Pi = 1.

As with the WTP interval, the mortality risk that any given respondent agrees to is
observed as an interval rather than the single value. Therefore, it was necessary to
determine a single (conditional mean) mortality risk (or QALY weight, equal to 1-
mortality risk) for each respondent. This was done as follows. First, we assume that the
mortality risk interval for each respondent based on the two questions represents a single
risk distribution. For each individual respondent j, there exists an upper and lower bound
on the value, call these Uj and Lj, where Lj is the minimum risk agreed to (which could
be zero) and U/ is the maximum risk the respondent accepted. The likelihood for this
respondent is [F(U/') - F(L/')], where F is the cumulative distribution function (CDF) for
the assumed distribution, which depends on a small number of parameters (e.g., mean

-------
and variance for normal). The likelihood for the full sample is just the product over j of
the individual contributions to the likelihood, which depends on the parameters of the
distribution function. To maximize it, we calculated the first derivatives with respect to
the parameters and set them equal to zero.

2.6 Survey Administration

A professional survey firm, Knowledge Networks (KN), administered the survey
to a panel representative of the US general population via a web-based survey mechanism
during Spring 2005. The statistical foundation of the research panel stems from the
application of probability-based sample selection methodologies to recruit panel
members. The KN web-enabled panel is the only available method for conducting
Internet-based survey research with a nationally representative probability sample
(Couper, 2001; Krotki and Dennis, 2001).

The Knowledge Networks Panel, recruited randomly through Random Digit
Dialing, represents the broad diversity and key demographic dimensions of the U.S.
population. The web-enabled panel tracks closely the U.S. population on age, race,
ethnicity, geographical region, employment status, and other demographic elements. The
differences that do exist are small and are corrected statistically in survey data (i.e., by
non-response adjustments). The web-enabled panel is comprised of both Internet and
non-Internet households, all of which are provided the same equipment for participation
in Internet surveys. Internet-based surveys are increasingly showing favorable
comparisons to mail and telephone survey methods (Berrens et al., 2003).

There are four main factors responsible for the representativeness of the web-
enabled research panel. First, the panel sample is selected using list-assisted random digit
dialing telephone methodology, providing a probability-based starting sample of U.S.

-------
telephone households. Second, the panel sample weights are adjusted to U.S. Census
demographic benchmarks to reduce error due to non-coverage of non-telephone
households and to reduce bias due to nonresponse and other non-sampling errors. Third,
samples selected from the panel for individual studies are selected using probability
methods. Appropriate sample design weights for each study are calculated based on
specific design parameters. Fourth, nonresponse and poststratification weighting
adjustments are applied to the final survey data to reduce the effects of non-sampling
error (variance and bias).

The endpoint selection, specific risk reduction, and follow up human health
questions are all randomized across the respondents. There are two human health
endpoints, two risk reductions, two ecological endpoints and associated risk reductions,
and two quality adjusted life year questions randomized across respondents. Each
respondent faces only one human health endpoint and associated risk reduction, one
ecological endpoint and associated risk reduction, and one QALY mortality risk (either
SG or TTO).

In the next section, we report the results of the surveys and discuss the
implications of the results.

3. Model Framework and Survey Results

Economic theory postulates that society is comprised of individuals who make

tradeoffs in order to satisfy their preferences, or, put another way, to maximize their
utility.

The statistical model for CV responses must satisfy both statistical and economic
criteria (Hanemann and Kaninnen, 2001). CV responses can be modeled as discrete
dependent variables with binary responses since respondents can either state "yes" or

-------
"no" to a particular bid value. An equivalent but alternative modeling form takes the bid
interval agreed to by an individual respondent as the dependent variable. In economic
terms, the statistical model for CV responses must be consistent with the theory of utility
maximization inherent in economic models. This assumes individuals show preferences
for market commodities (x) and nonmarket amenities (q) as represented by a utility
function U(x,q) which is continuous and non-decreasing (Hanemann, 2001). Individuals
face budget constraints based on income (y) and prices of the market commodities (p).
Individuals are assumed to be utility-maximizers given a budget constraint (e.g.,
disposable income). Willingness to pay, or the compensating variation (C) is the
maximum an individual is willing to pay to secure an increase to the nonmarket amenity.
In this case, the nonmarket amenity is expressed as a risk (r); therefore, a decrease in the
risk increases utility U(x, r).

Each respondent has an indirect utility function for which one can plot the
tradeoff between risk and income while maintaining utility as given by the slope of that
curve.

The economic measure of value is given as:

v(p, rlt y-C) = v(p, r0, y)	(3)

where C = the amount of money at which the individual is indifferent between a

lower probability of risk and higher income, and r0 and /'/ are different levels of:

•	Risk of a 6-point reduction in IQ to an unborn child given maternal exposure (IQ)

•	Risk of a 7-month deficit in reading comprehension given maternal exposure
(RC)

The assumption is that a smaller risk relative to baseline leads improves well-
being so compensating variation, or WTP, is positive. Expected utility is roughly

-------
proportional to risk; consequently WTP should be approximately proportional to risk, and
we test for this. As individuals spend more money, the utility loss increases. However,
WTP is likely small with respect to income and so an income effect is also likely to be
negligible.

All analyses are conducted using S-Plus 6.2 (Insightful Corporation, 2004) and
Microsoft Excel.

3.1 Descriptive Statistics

Table 4 presents the frequencies of response to the bid vectors across the surveys.
The proportion of yes responses decreases as the offered bid increases.

Table 5 provides a summary of the demographic characteristics of the sample, and
for comparison purposes, data from the 2000 census. This table shows that the sample is
representative of the US population. The median income differs, but this is primarily
attributable to the fact that income was provided in terms of ranges, and the median
income was estimated from the midpoint of the range provided for each individual. If one
compares the income distribution (shown in the table below the median and mean
income), it shows that survey samples are statistically indistinguishable from the
demographics of the US population.

The sample also shows a lower proportion of individuals with less than a high
school education as compared to the general public, and a higher proportion of
individuals with at least an associates degree. However, it is not clear that more
traditional survey methods (e.g., direct mail and/or telephone) would have reached a
higher proportion of this fraction of the population.

Table 6 provides the means for model covariates.

-------
3.2 Statistical Models

The double-bounded dichotomous choice elicitation format used here is
analogous to interval-censored survival data in medical and engineering settings which
model time to illness or failure of a component. In this case, we know the interval within
which WTP for any individual respondent lies; for example, for the yes-yes response, it is
known that the interval lies somewhere between the highest amount the respondent
agreed to and infinity. Table 1 shows the intervals for each bid vector based on the initial
bids for each survey, and Table 4 shows the proportion of respondents for each bid
interval.

The WTP model takes the form:

LNWTP1 = J30+ P] LN(ARisk) + P2LNIncome + j3xX + s	(4)

where

WTP for the z'th individual in the interval given in Table 1

ARisk - is the risk reduction (0.1 or 0.15)

Income - respondent household income

X- vector of respondent-specific attributes as given in Table 6

s - error term

The log likelihood function can be maximized assuming a particular parametric
distribution (e.g, lognormal) or by using the Turnbull nonparametric modification of the
Kaplan-Meier estimator, which makes no assumptions about the shape of the underlying
WTP distribution (Carson et al., 2003; Hanemann and Kanninen, 2001). We evaluated
several parametric forms (e.g., lognormal, weibull) and found the lognormal to provide
the best fit based on a Likelihood Ratio test. In addition, properties of the lognormal
distribution facilitate interpretation of the results. Figure 1 presents the visual goodness-
of-fit plots across distribution types.

-------
Parameter estimation is accomplished through maximum likelihood methods to
obtain the values of unknown statistical parameters that are most likely to have generated
the observed data. Figure 2 shows the WTP function for reading comprehension (IQ=0)
for two risk reductions (0 = small risk reduction, 1 = large risk reduction) and for IQ
(IQ=1).

Table 7 presents the results for several models based on the single endpoint
valuation results of the HHFirst survey only. Models 1 and 2, stratified by endpoint
(reading comprehension and IQ, respectively), include all covariates, while models 3 and
4 present the results for the reduced models. As shown in this table, the human health risk
reduction coefficient is positively related to WTP, and approaches statistical significance
for the IQ endpoint (p=0.14), but not for the reading comprehension endpoint. The only
significant predictors in the full models include behavioral and motivational variables,
including concern about PCBs in the environment (highly statistically significant across
all four models), and the response to the QALY question (used in the model as change in
QALY). As shown in Model 2, information received from scientists is positively
associated with WTP (p<0.1). WTP is proportional with respect to risk reduction
(coefficient = 1.0) for the IQ endpoint. Models with various interaction terms were not
significant and are omitted from the table.

Table 8 presents the results from a set of models using the EcoFirst survey results
for total WTP, which asks whether respondents would be willing to pay more into the
cleanup fund when considering human health endpoints in addition to ecological
endpoints. Models 1 and 2 are stratified by developmental endpoint for the total bid
amount. Under this model, there is a difference between the risk reduction coefficient

-------
(HHLNRR) for IQ as compared to reading comprehension as outcomes. For IQ, Table 8
shows the coefficient is 1.0 and approaches significance atp<0.18. For those respondents
who were asked about reading comprehension as an endpoint, the risk reduction
coefficient is statistically significant at -1.6 (p<0.03), indicating that respondents showed
a negative relationship between risk reduction and WTP for this endpoint.

Models 3 and 4 in Table 8 show the results for the full models including all
covariates for total WTP in the EcoFirst survey. For model 3, with reading
comprehension as the endpoint, statistically significant covariates include the risk
reduction coefficient, being female, concern about chemicals in the environment, whether
or not the respondent believes that PCBs can cause developmental delays as a result of in
utero exposures, and the QALY weight. All of these covariates are positively associated
with WTP, except for the risk reduction coefficient. Model 4, by contrast, stratified by IQ
as the endpoint, shows statistically significant covariates for the risk reduction variable,
concern about PCBs in the environment, whether or not the respondent believes that
PCBs can cause developmental delays as a result of in utero exposures, and the degree of
confidence in information received from industry scientists. The risk reduction
coefficient is positive, and only slightly more than proportional with respect to WTP, and
statistically significant, unlike for the reading comprehension subset. Concern about
PCBs in the environment generally and believing that PCBs can cause developmental
delays are both positively associated with WTP for the IQ subset of respondents.

The magnitude of the risk reduction coefficient is very similar across both the
HHfirst and Ecofirst surveys. Economic theory predicts that WTP should be

-------
approximately proportional with respect to risk reduction, and this hypothesis cannot be
rejected across these two datasets.

3.2.1	WTP per IQ Point

Cognitive ability, in addition to having an impact on later health status, also
influences productivity through an impact on earning potential as well as through years of
schooling and probability of employment. This relationship has been explored in the
literature through the relationship between childhood lead exposures and loss of lifetime
earnings by Grosse et al. (2002) and Salkever (1995). Grosse el al. (2002) evaluated three
different linear relationships between earnings and IQ, ranging from 1.76% to 2.37%
percentage earnings loss per IQ point. Based on this relationship, and the present value of
earnings of a two-year-old in 2000 dollars, results in values of a one point decrease in IQ
ranging from $12,700 to $17,200.

Estimates of WTP using these survey results represent WTP for a probability of a
6-point reduction in IQ, thus, WTP for a 100% probability of a 1-point reduction is
estimated by dividing WTP by 6 and dividing again by the risk reduction. This assumes
that WTP is linear in the probability of a reduction in IQ as a result of exposure and the
number of IQ points at risk. We evaluated WTP per IQ point using both the single
endpoint results from the HHFirst survey and the difference between the total valuation
and single endpoint valuation from the EcoFirst survey. The result for the HHFirst survey
is $466 (95% confidence interval = $380, $520) per IQ point.

3.2.2	WTP and QALYs

Table 9 shows the results of the models across surveys. The dependent variable
for the first model is the interval-censored WTP for the first set of questions from the
HHFirst survey, while the second model dependent variable is the total interval-censored

-------
bid amount from the EcoFirst survey. In both cases, covariates include whether the
endpoint was IQ (1) or reading comprehension (0), and a code for whether the elicitation
method for the QALY weight was standard gamble (0) or time-tradeoff (1). Finally, the
change in QALY for each respondent was calculated as described in section 3.4
(LNQALY). The resulting coefficients are very similar across the datasets, except for IQ.
For the HHFirst survey, there is no appreciable difference in the relationship between
change in QALY and WTP by developmental endpoint. But for the EcoFirst survey, the
IQ coefficient is negative and statistically significant. Respondents to that survey had a
33% lower WTP when asked about IQ as compared to reading comprehension.

The individual QALY weights (1 - mortality risk) range from 0.948 to 0.99975
for a 6-point reduction in IQ. This translates to a range of mortality risks that respondents
would accept on behalf of their (hypothetical) 10-year-old child of 2 in 10,000 to 9 in
1,000 per IQ point. Table 10 shows the mean, standard deviation, and number of
respondents by endpoint (IQ or reading comprehension) and elicitation method (standard
gamble or time tradeoff). There is no statistical difference by endpoint (%2 = 0.6, df=l,
p=QA\ while there is a statistically significant difference by elicitation method (%2 =
10.1, df=l,^=0.001).

We estimated WTP per QALY by dividing WTP by the expected change in
QALYs, where the change in QALY accounts for the probability of having the cognitive
deficit. The mean WTP per QALY is $109,000 (95% confidence interval = ($70,000,
$148,000). WTP per QALY has been proposed as a potential criterion for evaluating
efficacy of social programs (Baker etal., 2004; Gyrd-Hansen, 2003; Krupnick, 2004;
Van Houtven etal., 2003) based on cost-effectiveness. King et al. (2005) discuss

-------
standards for evaluating WTP/QALY ratios, and find that this ratio varies considerably
depending on the valuation methodology. In 2003 dollars, the median ratio from eight
CV studies based on (personal) safety was $184,200. In contrast, revealed preference
studies, based on safety, have a median value of $106,700. The results of this study are
consistent with these literature values.

4. Discussion

The importance of obtaining behavioral and motivational answers from
respondents in CV surveys has been shown (Heberlein etal., 2005; Nunes and
Schokkaert, 2003; Dubourg etal., 1997). In this case, concern about PCBs in the
environment and the respondent-specific QALY weighting are important, highly
statistically significant predictors of WTP. The QALY weighting indirectly addresses
perceived risk in that it elicits from respondents an indication of the perception the parent
has about the quality of life for the child if s/he has the cognitive deficit. It addresses the
issue more directly by asking about your hypothetical child, as opposed to how
significant do you think the risks are in general (e.g., risk.baby, PCBChild).

Interestingly, in responses to open ended questions, a number of respondents
indicated that because there were fish consumption advisories in place in their particular
State (indeed, most States), they felt the risks were lower than what had been portrayed in
the survey, although the survey does indicate that the risks are only to those individuals
who consume fish.

The risk reduction coefficients for IQ are both positive and approaching statistical
significance based on the responses to the single endpoint in the HHFirst survey (1.0,
p=0.14) and the EcoFirst total endpoint (1.1, p=0.14), providing greater confidence that
the surveys have captured the relationship between risk reduction and WTP for IQ. In a

-------
reduced model using just risk reduction as a predictor based on the single endpoint in the
HHfirst survey, the coefficient is 1.0 (p=0.15), a proportional result approaching
significance. The results for reading comprehension as an endpoint are not as robust.
These results suggest that survey takers were able to think about IQ as a developmental
endpoint and were indeed willing to pay for risk reductions, while this is not the case for
reading comprehension.

It is true that these risks are not experienced directly by the respondents
themselves. Women of childbearing age who are pregnant or thinking of becoming
pregnant and that consume freshwater fish are the only ones who would actually be
exposed, and even in that case, they do not experience the risk directly. The risk is to the
unborn child. This is the most immediate that the risk can be, but the proportion of
respondents who are pregnant (this question was not asked - the only information we
have is the number of women of child-bearing age and the number of children by age
group in the household) is itself likely a relatively small proportion of the overall
respondent population.

Respondents were willing to increase their stated bids between the single
ecological endpoint in the EcoFirst survey when asked about a total bid. This was not the
case in the HHFirst survey (respondents were not willing to increase their stated bids
when asked about ecological effects after they had already responded to human health
endpoints).

The estimated WTP values per IQ point from these surveys are orders of
magnitude lower than estimates based on future earnings. The estimates obtained here are
approximately $500 while the estimates from the earnings literature are in the $10,000 to

-------
$20,000 range. The results presented here represent the average WTP per IQ point from a
representative sample of the American general public. It is possible that respondents do
not realize (or do not think about) the implications of IQ on future earnings and so
underestimated the potential value of the loss. Another possibility is that respondents
recognize the effect on future earnings, but use higher discount rates in evaluating these
benefits than the rates used to calculate the estimates from the literature (consistent with
the idea that people discount the future too much).

The policy implications of these WTP values, however they are expressed, comes
in the context of a particular decision. One of the goals of this survey was to demonstrate
how stated preference methods might be used to develop economic values for risk
reductions within a particular regulatory framework. In a companion paper (von
Stackelberg, 2006), we develop an application based on the Hudson River Superfund site
to show how this might be done.

The survey results suggest that IQ represented a more meaningful endpoint for
respondents than reading comprehension. However, it is known that people have
difficulty evaluating and responding to numerical differences in the magnitude of risk
reduction, particularly for small risks or small effects (Hammitt and Graham, 1999; Corso
et al., 2001; Schwartz eial., 1997). Further, in this case, exposures are experienced by
one cohort while effects are experienced by another who also happen to be children and
therefore unable to make risk-based decisions for themselves. Women of childbearing
age who are pregnant or thinking of becoming pregnant and that consume freshwater fish
are the only ones who would actually be exposed, and even in that case, they do not
experience the risk directly. The risk is to the unborn child. This is the most immediate

-------
that the risk can be, but the proportion of respondents who are pregnant (this question
was not asked - the only information we have is the number of women of child-bearing
age and the number of children by age group in the household) is itself likely a relatively
small proportion of the overall respondent population. However, this is an issue that is
likely to arise time and again with significant policy implications given the increasing
evidence of in utero environmental exposures leading to significant and potentially
lasting health effects later in life. It is, after all, children who presumably still have most
of their lives in front of them and will be the ones who directly experience the
repercussions of decisions made today, ostensibly on their behalf.

-------
References

(ATSDR) Agency for Toxic Substances and Disease Registry. 2000. Toxicological
profile for polychlorinated biphenyls (PCBs). US Department of Health and
Human Services, Public Health Service. Prepared by Syracuse Research
Corporation, November.

Alberini, A. 1995. Efficiency vs bias of willingness-to-pay estimates: Bivariate and
interval data models. Journal of Environmental Economics and Management
29:169-180.

Alberini, A. 2005. What is a life worth? Robustness of VSL values from contingent
valuation surveys. Risk Analysis 25(4):783-800.

Baker, R., S. Chilton, C. Donaldson, M. Jones-Lee, H. Metcalf, P. Shackley and M.

Ryan. 2004. Determining the societal value of a QALY by surveying the public in
England and Wales: A research protocol. Submission to the National Institute for
Clinical Excellence and National Co-ordinating Centre for Research
Methodology, School of Population and Health Sciences, University of Newcastle
upon Tyne.

Bateman, I.J. andK.G. Willis, Eds. 1999. Valuing environmental preferences: Theory
and practice of the contingent valuation method in the US, EU, and developing
countries. New York: Oxford University Press. 645 pp.

Berrens, R., A. Bohara, H. Jenkins-Smith, C. Sliva and D. Weimer. 2003. The advent of
Internet surveys for political research: A comparison of telephone and Internet
samples. Political Analysis, 11(1).

Cannon, M., R. Kopp and A. Krupnick. 1996. Valuation of Developmental and

Reproductive Effects: A Scoping Study. Resources for the Future. October, 1996.

Carson, R.T. and R.C. Mitchell. 1995. Sequencing and nesting in contingent valuation
surveys. Journal of Environmental Economics and Management 28:155-173.

Carson, R.T., R.C. Mitchell, M. Hanemann, R.J. Kopp, S. Presser and P.A. Ruud. 2003.
Contingent valuation and lost passive use: Damages from the Exxon Valdez oil
spill. Environmental and Resource Economics 25:257-286.

Corso, P.S., J.K. Hammitt and J.D. Graham. 2001. Valuing mortality-risk reduction:

Using visual aids to improve the validity of contingent valuation. Journal of Risk
and Uncertainty 23:165-184.

Couper, M.P., M.W. Traugott and M.J. Lamias. 2001. Web survey design and
administration. Public Opinion Quarterly 65 (2) 230 -253.

Diamond, P. 1996. Testing the internal consistency of contingent valuation surveys.
Journal of Environmental Economics and Management 30:337-347.

Dickie, M. and S. Gerking. 2002. Willingness to pay for reduced morbidity. Manuscript
for presentation at the workshop, "Economic Valuation of Health for
Environmental Policy: Assessing Alternative Approaches," March 18-19, 2002,
Orlando, FL.

-------
Dockins, C., R.R. Jenkins, N. Owens, N.B. Simon and L.B Wiggins. 2002. Valuation of
childhood risk reduction: The importance of age, risk preferences, and
perspective. Risk Analysis 22(2):335-346.

Dubourg, W.R., M.W. Jones-Lee and G. Loomes. 1997. Imprecise preferences and
survey design in contingent valuation. Economica 64:681-702.

Gold, M.R., J.E. Siegel, L.B. Russell andM.C. Weinstein. 1996. Cost-effectiveness in
health and medicine. New York: Oxford University Press.

Grosse, S.D., T.D. Matte, J. Schwartz and R.J. Jackson. 2002. Economic gains resulting
from the reduction in children's exposure to lead in the United States.
Environmental Health Perspectives 110(6): 5 63-5 69.

Gyrd-Hansen, D. 2003. Willingness to pay for a QALY. Health Economics 12:1049-
1060.

Hammitt, J.K. and J.D. Graham. 1999. Willingness to pay for health protection:

inadequate sensitivity to probability? Journal of Risk and Uncertainty 18:33-62.

Hammitt, J.K. and J-T. Liu. 2004. Effects of disease type and latency on the value of
mortality risk. Journal of Risk and Uncertainty 28(l):73-95.

Hanemann, W.M. 1991. Willingness to pay and willingness to accept: How much can
they differ! American Economic Review 81(3): 635-647.

Hanemann, W.M., J.B. Loomis and B.J. Kanninen. 1991. Statistical efficiency of double-
bounded dichotomous choice contingent valuation, Am JAgEc 73:(4)1255-1263.

Hanemann, W.M. and B. Kanninen. 2001. Statistical analysis of discrete-response CV
data. In Valuing Environmental Preferences, Bateman, I.J. and K.G. Willis, Eds.
New York: Oxford University Press.

Heberlein, T.A., M.A. Wilson, R.C. Bishop andN.C. Shaeffer. 2005. Rethinking the
scope test as a criterion for validity in contingent valuation. Journal of
Environmental Economics and Management 50:1-22.

Insightful Corporation, Inc. 2004. Users Manual for S-Plus.

Jacobson, S.W., L.M. Chiodo and J.L. Jacobson. 1999. Breastfeeding effects on
intelligence quotient in 4- and 11-year-old children. Pediatrics 103:71-76.

Jacobson, J.L and S.W. Jacobson. 2002. Breast-feeding and gender as moderators of
teratogenic effects on cognitive development. Neurotoxicology and Teratology
24:349-358.

Jacobson, J.L. and S.W. Jacobson. 2002. Association of prenatal exposure to an

environmental contaminant with intellectual function in childhood. Clinical
Toxicology 40(4):467-475.

Jacobson, J.L., J. Janisse, M. Baneijee, J. Jester, S.W. Jacobson and J.W. Ager. 2002. A
benchmark dose analysis of prenatal exposure to polychlorinated biphenyls.
Environ Health Perspect 110(4):393-398.

-------
King, J.T., J.Tsevat, J.R. Lave and M.S. Roberts. 2005. Willingness to pay for a quality-
adjusted life year: Implications for societal health care resource allocation.

Medical Decision Making 25(6):667-677.

Krotki, K. and M.J. Dennis. 2001. Probability-based Survey Research on the Internet.

Presented at the August 2001 Conference of the International Statistical Institute.

Krupnick, A.J. 2004. Valuing health outcomes: Policy choices and technical issues.
Washington, D.C.: Resources for the Future, 104 pp.

Levin, E.D., S.L. Schantz and R.E. Bowman. 1988. Delayed spatial alternation deficits
resulting from perinatal PCB exposure in monkeys. Arch Toxicol 62: 267-273.

Longnecker, M.P., Wolff, M.S., Gladen, B.C., Brock, J.W., Grandjean, P., Jacobson, J.L.,
Korrick, S.A., Rogan, W.J., Weisglas-Kuperus, N., Hertz-Picciotto, I., Ayotte, P.,
Stewart, P., Winneke, G., Charles, M.J., Jacobson, S.W., Dewailly, E., Boersma,
E.R., Altshul, L.M., Heinzow, B., Pagano, J.J. and A.A. Jensen. 2003.

Comparison of poly chlorinated biphenyl levels across studies of human
neurodevelopment. Environ Health Perspect 11 l(l):65-70.

McFadden, D. and G. Leonard. 1993. Issues in the contingent valuation of environmental
goods: methodologies for data collection and analysis, In Hausman, J. A. Ed.
Contingent Valuation: A Critical Assessment. New York: North-Holland, 165-
215.

Salkever, D.S. 1995. Updated estimates of earnings benefits from reduced exposure of
children to environmental lead. Environmental Research 70:1-6.

Schantz, S.L., E.D. Levin, R.E. Bowman. 1989. Effects of perinatal PCB exposure on
discrimination-reversal learning in monkeys. Neurotoxicol Teratol 11: 243-250.

Schantz, S.L., J.J. Widholm and D.C. Rice. 2003. Effects of PCB exposure on

neuropsychological function in children. Environ Health Perspect 111(3):357-
376.

Schwartz, J. 1994. Societal benefits of reducing lead exposure. Environmental Research
66:105-124.

Schwartz, J., H. Pitcher, R. Levin, B. Ostro and A.L. Nicholas. 1985. Cost and Benefits of
Reducing Lead in Gasoline: Final Regulatory Impact Analysis. Washington,
D.C.: Office of Policy Analysis, Environmental Protection Agency.

Schwartz, L.M., S. Woloshin, W.C. Black and H.G. Welch. 1997. The role of numeracy
in understanding the benefit of screening mammography. Annals of Internal
Medicine 127:966-972.

Van Houtven, G., M. Rousu, J-C. Yang, C. Pringle, W. Wagstaff and J. DePlatchett.
2003. Valuation of morbidity losses: Meta-analysis of willingness-to-pay and
health status measures. Final Report, Prepared for the Food and Drug
Administration by Research Triangle Institute.

Van Houtven, G., Powers, J., Jessup, A. and J-C. Yang. 2004. Valuing avoided morbidity
using meta-analysis: What can health status measures and QALYs tell us about
WTP? Research Triangle Institute Working Paper 04-01, www.rti.org/enrepaper/.

-------
TABLE 1: Initial Bid Vectors and Followup Bids for the CV Surveys



Bid vectors based on final response in first section and are given as initial bid, upper, lower:

Initial Bid

Y-Y1

Y-N1

N-Y1

N-N

$25

C ($100, $200, $50)

B ($50, $100, $25)

A ($25, $50, $10)

random

$50

D ($200, $400, $100)

C ($100, $200, $50)

B ($50, $100, $25)

random

$100

E ($400, $800, $200)

D ($200, $400, $100)

C ($100, $200, $50)

random

$200

F ($800, $1000, $400)

E ($400, $800, $200)

D ($200, $400, $100)

random

$400
$800

G ($1000, $1500, $800)
H ($2000, $1500, $800)

F ($800, $1000, $400)

E ($400, $800, $200)

random
random

G ($1000, $1500, $800)

F ($800, $1000, $400)

Notes:

1 - It is possible, in the followup, to respond "no" to a value for the total that had already been agreed to

in the previous section. In that case, respondents are shown the following prompt: "You already agreed you'd be

willing to pay this amount for human health benefits alone. Now we're asking about the total you'd be willing to pay"

-------
TABLE 2: Risk Reductions in the Surveys

Endpoint

Context

Small Risk
Reduction

Large Risk
Reduction









Eagle

Probability of reproductive impairment
significant enough to affect viability of the
population

0.1

0.15

Species Sensitivity Distribution (SSD

Probability of reproductive significant
reproductive effects to 20% of all avian species in
a freshwater ecosystem

0.25

0.4

Reading Comprehension

Probability of reading at approximately 7 months
below grade level

0.1

0.15

IQ

Probability of a 6-point reduction in IQ

0.1

0.15

-------
TABLE 3: Mortality Risk and Longevity Reduction Questions to Determine QALYs

Initial Probability of

Death versus
Successful Treatment

Followup Probability
if "yes"

Followup Probability
if "No"

Initial Reduction in
Longevity (days)

Followup Reduction
if "yes" (days)

Followup Reduction
if "No" (days)

5 in 10,000

10 in 10,000

2.5 in 10,000

11

22

5

10 in 10,000

20 in 10,000

5 in 10,000

22

44

11

20 in 10,000

40 in 10,000

10 in 10,000

44

88

22

QALYcode = 1 if life expectancy reduction, 0 if mortality risk

-------
TABLE 4: Proportion of Respondents in Each Bid Interval for HHFirst (Single Endpoint) and Ecofirst (Total Across Endpoints)

HHFIRST — Single Endpoint

IQ (n=2C

>8)

RC (n=196)

Bid Amount

n

Y-Y

Y-N

N-Y

N-N

n

Y-Y

Y-N

N-Y

N-N

A ($25, $50, $10)

35

11%

3%

0%

3%

35

12%

2%

1%

4%

B ($50, $100, $25)

36

8%

4%

1%

5%

32

7%

5%

4%

2%

C ($100, $200, $50)

27

3%

3%

2%

5%

21

3%

1%

2%

3%

D ($200, $400, $100)

30

4%

3%

2%

4%

33

4%

4%

2%

7%

E ($400, $800, $200)

41

2%

5%

4%

8%

40

4%

5%

2%

10%

F ($800, $1000, $400)

33

4%

1%

1%

9%

32

3%

4%

2%

7%



ECOFIRST - Total Bid for Both
Endpoints

IQ (n=194)

RC (n=208)

Bid Amount

n

Y-Y

Y-N

N-Y

N-N

n

Y-Y

Y-N

N-Y

N-N

A ($25, $50, $10)

11

0%

2%

2%

2%

14

2%

0%

1%

3%

B ($50, $100, $25)

16

2%

3%

2%

3%

18

1%

3%

2%

2%

C ($100, $200, $50)

37

11%

5%

6%

3%

47

10%

5%

5%

3%

D ($200, $400, $100)

47

6%

8%

7%

4%

39

3%

8%

6%

2%

E ($400, $800, $200)

30

0%

7%

5%

4%

31

3%

2%

4%

5%

F ($800, $1000, $400

32

3%

1%

6%

8%

32

5%

1%

6%

3%

G ($1000, $1500, $800)

10

2%

2%

2%

0%

11

1%

1%

3%

0%

H ($1500, $2000, $1000)

5

2%

0%

1%

0%

9

3%

0%

1%

0%

-------
TABLE 4, continued: Proportion of Respondents in Each Bid Interval for the Combined Survey

COMBINED

Combined (n=204)

Bid Amount

n

Y-Y

Y-N

N-Y

N-N

A ($25, $50, $10)

37

11%

4%

0%

3%

B ($50, $100, $25)

41

9%

6%

0%

5%

C ($100, $200, $50)

23

4%

2%

1%

4%

D ($200, $400, $100)

34

5%

4%

2%

5%

E ($400, $800, $200)

35

2%

5%

1%

9%

F ($800, $1000, $400)

29

3%

3%

0%

8%

-------
TABLE 5: Demographics for each Subsurvey and the US Census

ECOFIRST	HUMANFIRST COMBINED



Eagle

SSD

RC

IQ

Combined

US Census

Demographic

(n=193)

(n=210)

(n=196)

(n=208)

(n=204)

Data1



Some high school, no diploma

7%

8%

19%

11%

16%

20%

High school

29%

30%

29%

35%

32%

29%

Some college, no degree

23%

20%

21%

24%

21%

21%

Associate degree (AA, AS)

15%

12%

7%

5%

6%

6%

Bachelor's degree

17%

19%

16%

19%

14%

16%

Master's degree

4%

7%

7%

5%

9%

6%

Other

5%

4%

2%

2%

1%

3%



Black, Non-Hispanic

10%

12%

12%

15%

12%

12%

Hispanic

9%

15%

17%

9%

11%

13%

Other, Non-Hispanic

5%

5%

4%

4%

5%

0%

White, Non-Hispanic

76%

68%

67%

72%

72%

75%



Female

57%

50%

48%

51%

52%

51%

Male

43%

50%

52%

49%

48%

49%

Income

Less than $10,000

12%

10%

12%

13%

13%

10%

$10,000 to $14,999

11%

5%

9%

8%

4%

6%

$15,000 to $19,999

5%

4%

5%

4%

8%

6%

$20,000 to $24,999

8%

10%

6%

8%

5%

7%

$25,000 to $29,999

8%

7%

10%

6%

5%

6%

$30,000 to $34,999

7%

7%

5%

4%

8%

6%

$35,000 to $39,999

4%

10%

10%

10%

9%

6%

$40,000 to $49,999

9%

11%

10%

6%

15%

11%

$50,000 to $59,999

10%

9%

7%

13%

7%

9%

$60,000 to $74,999

10%

9%

8%

12%

12%

10%

$75,000 to $99,999

11%

9%

12%

6%

7%

10%

$100,000 to $124,999

2%

3%

5%

5%

5%

5%

$125,000 to $149,999

1%

2%

1%

1%

2%

3%

$150,000 to $174,999

1%

1%

1%

0%

2%

2%

$175,000 or more

2%

2%

2%

0%

1%

2%



Divorced

12%

15%

13%

20%

14%

10%

Married

52%

50%

48%

46%

52%

54%

Separated

2%

2%

3%

4%

1%

2%

Single (never married)

26%

28%

28%

26%

29%

27%

Widowed

7%

5%

7%

4%

3%

7%

1: Data provided for males and females combined (except gender); therefore, percentages



may not equal 100 due to combining. Data from: factfinder.census.gov, 2000 Census



-------
TABLE 6: Means for the Covariates Across Subsurveys



Eagle (n=193)

SSD (n=210)

IQ (n=208)

RC (n=196)

Combined (n=204)





ECOFIRST

HHF

1RST

COMB

[NED

Parameter

Parameter Name

Mean

Stdev

Mean

Stdev

Mean

Stdev

Mean

Stdev

Mean

Stdev

























Education (1 for college and above, 0
otherwise)

EDUCAT

0.53



0.61



0.55



0.53



0.50



White (1 for yes, 0 otherwise)

WHITE

0.76



0.68



0.72



0.67



0.71



Black (1 for yes, 0 otherwise)

BLACK

0.09



0.12



0.15



0.12



0.22



Hispanic (1 for yes, 0 otherwise)

HISPANIC

0.09



0.15



0.09



0.17



0.14



Gender (1 if Female, 0 if Male)

MALE

0.57



0.50



0.52



0.48



0.52



Natural log of income

LNInc

10.36

0.86

10.46

0.83

10.41

0.86

10.41

0.89

10.38

0.89

Married (1 if yes, 0 otherwise)

MARRIED

0.52



0.50



0.46



0.48



0.52



Live in a metropolitan area (1 if yes, 0 if no)

METRO

0.83



0.82



0.83



0.84



0.79



Natural log of ecological risk reduction

LNEcoRR

-2.09

0.20

-1.17

0.23

-1.67

0.49

-1.60

0.52

-2.11

0.21

Natural log of human health risk reduction

HHLNRR









-2.09

0.20

-2.09

0.20

-2.09

0.20

Have you ever heard of PCBs (1 if yes, 0
otherwise)

PCBs

0.48



0.50



0.45



0.43



0.41



Confidence in response to single endpoint
valuation (scale of 1 to 5 where 1 is not
confident and 5 is very confident)

ConfWildlife

4.39



4.16

1.64

3.70

1.15

3.62

1.16

na



Confidence in total

ConfTotal

4.55

1.19

4.06

1.71

3.67

1.11

3.60

1.15

3.31

1.39

QALY code (0 if standard gamble, 1 if time
tradeoff

QALYcode





















Are you able to think about ecological
endpoints separately from human (1 if yes, 0 if
no)

eco.sep

0.78



0.72



0.71



0.77



na



Are you able to think about ecological benefits
separately from human health benefits? (1 if
yes, 0 otherwise)

eco.ben.sep

0.62



0.63



0.62



0.64



na



Concerned about chemicals in the
environment (1 if yes, 0 otherwise)

ChemConcern

3.12



2.96



3.04



2.89



3.03



-------
TABLE 6: Means for the Covariates Across Subsurveys

Concerned about PCBs in the environment (1























if yes, 0 otherwise)

PCBConcern

2.96



2.77



2.69



2.62



2.87



Do you believe PCBs can cause reproductive























effects in wildlife? (1 if yes, 0 otherwise)

PCBWildlife

0.66



0.59



0.59



0.60



0.60



Do you believe PCBs can cause























developmental effects in children exposed in























utero ? (1 if yes, 0 otherwise)

PCBChild

0.61



0.54



0.65



0.60



0.59



Rate the risks facing eagles in this state (0 =























not sure, 1 = not serious, 2 = somewhat























serious, 3 = very serious, 4 = extremely























serious)

risk.wldlf

2.14

1.17

2.04

1.20

1.94

1.17

1.94

1.19

2.08

1.13

Rate the risks facing unborn babies in this























state (0 = not sure, 1 = not serious, 2 =























somewhat serious, 3 = very serious, 4 =























extremely serious)

risk.baby

2.22

1.27

2.01

1.28

2.17

1.25

2.11

1.30

2.16

1.29

How often do you watch programs on























television about wildlife (1 = never, 2 = rarely,























3 = sometimes, 4 = often)

tv.wldlf

2.99

0.88

2.91

0.97

2.75

0.94

3.03

0.93

2.90

0.90

Do you live near freshwater (1 = yes, 0 = no)

live.fw

0.69



0.64



0.60



0.60



0.66



How much time do you spend on a river, lake,























or stream? (1 = never, 2 = rarely, 3 =























sometimes, 4 = often)

time.fw

2.60

1.03

2.65

1.02

2.49

0.97

2.61

1.03

2.62

0.99

How often do you eat recreationally caught























fish (0 = never, 1 = a few times a year, 2 = a























few times a month, 3 = a few times a week)

eat.fish

2.50

0.81

2.53

0.85

2.51

0.80

2.47

0.85

2.57

0.83

How much confidence do you have in























information you receive from government























sources (1 = none, 2 = some, 3 = a lot)

conf.gov

1.85

0.56

1.78

0.49

1.93

0.53

1.85

0.30

1.85

0.51

How much confidence do you have in























information you receive from industry























scientists (1 = none, 2 = some, 3 = a lot)

conf.sci.ind

1.88

0.58

1.82

0.54

1.85

0.62

1.81

0.60

1.86

0.58

-------
TABLE 6: Means for the Covariates Across Subsurveys

How much confidence do you have in
information you receive from university
scientists (1 = none, 2 = some, 3 = a lot)

conf.sci.univ

2.25

0.59

2.27

0.60

2.21

0.60

2.20

0.59

2.31

0.56

How much confidence do you have in
information you receive from television
sources (1 = none, 2 = some, 3 = a lot)

conf.tv

1.70

0.58

1.68

0.54

1.72

0.55

1.70

0.56

1.71

0.56

How much confidence do you have in
information you receive from government web
sites (1 = none, 2 = some, 3 = a lot)

conf.gov. web

1.87

0.50

1.78

0.53

1.87

0.55

1.83

0.54

1.81

0.50

How much confidence do you have in
information you receive from commercial web
sites (1 = none, 2 = some, 3 = a lot)

conf. comm. web

1.69

0.52

1.62

0.52

1.61

0.55

1.59

0.52

1.65

0.52

How much confidence do you have in
information you receive from nonprofit web
sites (1 = none, 2 = some, 3 = a lot)

conf.np.web

2.10

0.62

2.09

0.58

2.04

0.58

2.02

0.60

2.05

0.59

How much confidence do you have in
information you receive from university web
sites (1 = none, 2 = some, 3 = a lot)

conf.uni.web

2.21

0.59

2.20

0.54

2.12

0.64

2.06

0.62

2.15

0.58

How much confidence do you have in
information you receive from print media (1 =
none, 2 = some, 3 = a lot)

conf. print

1.86

0.56

1.88

0.40

1.84

0.53

1.81

0.51

1.88

0.54

-------
TABLE 7: Model Results for HHFirst Model for
Developmental Endpoints



Model 1

Model 2

Model 3

Model 3

Model 4



RC only

IQ only

across endpoints

RC only

IQ only

Intercept

1.3 (2.4)

1.6(2.9)

4 6(1 1)****

3.5 (1.5)**

5.9 (1.6)****

Risk Reduction

0.1 (0.7)

0.5 (0.7)

0.7(0.5)

0.4 (0.7)

1.0 (0.7)

Age

-0.01 (0.009)

0.002 (0.009)







Education

0.2(0.3)

0.6(0.3)*







Race (Ref = White)











Other

1.2(0.7)

0.6 (0.9)







Black

0.2(0.5)

0.1 (0.4)







Hispanic

0.07 (0.4)

0.2 (0.6)







Male

0.2(0.3)

-0.02 (0.3)







Income

-0.08 (0.2)

-0.01 (0.2)







Married

0.09 (0.3)

-0.1 (0.3)







Metro

0.9(0.4)**

0.1 (0.4)







PCB Concern

0.9 (0.2)****

0.4 (0.2)***

0.9 (0.1)****

1.1 (0.1)****

0.8 (0.2)****

QALY

0.2(0.1)**

0.3 (0.1)****

0.3 (0.1)****

0.2 (0.1)***

0.3 (0.1)***

risk.baby

0.08 (0.1)

0.3 (0.1)**







live.fw

0.3 (0.3)

-0.02 (0.2)







eat. fish

0.2 (0.2)

0.1 (0.2)







confgov

0.3 (0.3)

0.2(0.3)







conf.sci.ind

-0.2 (0.3)

0.5 (0.3)*







conf.sci.uni

0.4(0.3)

0.8 (0.3)**







conf.tv

0.03 (0.3)

-0.5 (0.4)







conf.print

0.3 (0.4)

0.2 (0.4)



















-2 *Log-Likelihood

423

460

942

444

492



n=192

n=206

n=398

n=192

n=206

* p<0.10, ** p<0.05, *** p<0.01, **** p<0.001







-------
TABLE 8: Model Results for EcoFirst Model for
Total WTP Based on Developmental Endpoints



Model 1

Model 2

Model 3

Model 4



RC only

IQ only

RC only

IQ only

Intercept

2.2(1.5)

7.3 (1.3)****

-0.008 (2.6)

4.4 (2.1)**

Risk Reduction

-1.6 (0.7)**

1.0 (0.6)

-1.3 (0.6)**

1.1 (0.6)**

Eagle





0.2(0.3)

-0.4 (0.2)

Education





-0.2 (0.3)

0.06 (0.3)

Race (Ref = White)









Other





0.4 (0.9)

-0.3 (0.4)

Black





0.4(0.5)

0.4 (0.4)

Hispanic





0.8 (0.4)

0.2 (0.4)

Male





0.5 (0.3)**

0.2 (0.2)

Age





-0.003 (0.009)

-0.001 (0.008)

Income





-0.04 (0.2)

-0.02 (0.2)

Married





0.3 (0.3)

-0.1 (0.3)

Metro





0.05 (0.4)

-0.1 (0.3)

PCBConcern





0.6 (0.2)****

0.4(0.2)**

risk.baby





0.3 (0.1)***

0.2(0.1)*

live.fw





0.1 (0.3)

-0.3 (0.3)

QALY





0.2 (0.06)***

0.06 (0.05)

eat.fish





-0.2 (0.2)

0.2 (0.2)

confgov





0.4(0.3)

0.3 (0.3)

conf.sci.ind





-0.3 (0.3)

-0.1 (0.2)

conf.sci.uni





0.2 (0.4)

0.5 (0.2)**

conf.tv





-0.01 (0.3)

-0.1 (0.3)

conf. print





0.1 (0.3)

0.4(0.3)











-2 *Log-Likelihood

704

658

635

569



n=208

n=194

n=205

n=188

* p<0.10, ** p<0.05, *** p<0.01, ****

p<0.001





-------
TABLE 9: WTP versus QALY Across Surveys



HHFirst Single

Ecofirst Total



Endpoint

Endpoint

Intercept

5.6 (0.3)****

5.8 (0.2)****

IQ

0.08 (0.2)

-0.4 (0.2)**

QALYcode

0.1 (0.2)

0.1 (0.2)

LNQALY

0.3 (0.07)****

0.1 (0.04)***

-2 *Log-Likelihood

1034

1345



n=398

n=397

* p<0.10, ** p<0.05,

*** p<0.01, **** p<0.001

-------
TABLE 10: Mean (Standard Deviation) QALY Weights by
Endpoint and Elicitation Method

Elicitation Method

IQO)

n

RC (0)

n

Standard Gamble (Mortality Risk) (0)

0.993 (0.016)

192

0.993 (0.016)

215

Time Tradeoff (Decrease in Longevity) (1)

0.987 (0.021)

204

0.989 (0.019)

183

QALYweightby QALYcode, Kruskal-Wallis%2=10.3,/> =0.001
QALYweight by Endpoint, Kruskal-Wallis %2=0.6, p =0.4

-------
.98
.7
.3

£ 1

I .03

I 01

1.003

.0005
.0002

.00005
.00002

.9999
.998
.98

>, 9

1	7

E -4

2	.2

CL

.05
.005
.0005
.00005
.000003

.9999
.9995
.998
.99
.95
£ .8
¦n .5
ra ~
.£2 .2

2.05
.01
.002
.0005

.00005

.000005

weibull

extreme

1b 50 100 5(to 1000

WTP

0 2l5o 4&0 6(50 8(Jo 1oEo

WTP

lognormal

0

normal





1b 5b 160 560 iobo

WTP

0 2^0 400 6(50 8&0 1ob6

WTP

"W

log logistic

logistic

....... ... —	inula,

50 100

"sSo "10I0 0

_U1UI_

460

660

WTP	WTP

FIGURE 1: Probability Plots for the HHFirst Single Endpoint

¦¦¦F" 1

1000

-------
00

o

0

Q_

CO
-------
Parental Decision-Making and Children's Health

Ann Bostrom, Sandra Hoffmann, Alan Krupnick and Wictor Adamowicz1
With Robin Goldman, Michael McWilliams, and Jeremy Varner

INTRODUCTION

Recent interest in valuation of children's health has raised many questions for how stated
preference studies are conducted (EPA 2000, 2003, OECD 2006). One of the most pressing
methodological questions is how parental willingness to pay (WTP) should be elicited in a
stated preference survey. Typically, stated preference surveys randomly sample households
and then either randomly sample adults within the household or, where pre-screened panels
are used, rely on the person in the panel. The responding adult is asked to report household
willingness to pay. These study designs assume that stated household WTP is invariant to
who reports it or, at least, that there is no systematic bias between respondents on the basis of
gender or other observable demographic characteristics.

This approach is consistent with a unitary model of household decisionmaking, which
assumes that the household acts as a single decisionmaking unit, with a single set of fixed
preferences and a single budget constraint (Samuelson 1956, Becker 1974). Since the 1970s,
this view has been augmented by the view that household level consumption and labor
supply decisions are the outcome of a bargaining process between adult decision makers in
the household (Ashworth and Ulph 1981, Manser and Brown 1980, McElroy and Horney
1981). The empirical literature on alternative household models has focused on
identification of departures from the unitary model using secondary household level data
(Browning et al. 1994, Lundberg et al. 1997, Browning and Chiappori 1998).

Stated preference surveys, by their nature, collect individual level data. As a result, it is
critical to understand the relationship between individual statements and household level
choice. For example, Bateman (2005) shows that unless adults in a multi-adult household
fully pool income, the standard approach of asking one adult to provide household WTP will
not give an accurate estimate of household WTP. It is unclear at present whether
respondents are providing their own preferences or their appraisal of the outcome of a
household decision process, whether unitary or bargained. This problem may be particularly
important in valuing children's health outcomes. Differences between parents' risk
perceptions, risk attitudes, knowledge about and responsibility for children's health and care,

1	The authors are respectively from the Georgia Institute of Technology, Resources for the Future,
Resources for the Future, and University of Alberta. This project is funded by a grant from the EPA STAR
program. We would like to thank Will Wheeler and the ORD staff, other staff at the USEPA and reviewers
for the EPA STAR program for their support of this research. We would also like to thank James Bason
and his staff at the University of Georgia's Survey Research Center for their work in survey administration.

2	Research Assistants, RFF, RFF and Georgia Institute of Technology, respectively.

1

-------
and control over household budget could affect individual parents' responses about their
WTP to reduce their children's health risks.

These concerns suggest that elicitation of household WTP in a stated preference study, in
particular eliciting parents' WTP for reductions in children's health risks, may be more complex
than typically assumed in stated preference studies. To begin to sort out this complexity and
ultimately help design a WTP survey of households, we conducted a study examining parental
decision-making about a variety of decisions, including reducing children's health risks in the
context of lead paint exposure. This paper reports on some of the findings of this study,
focusing on family decision processes, leaving to another paper analysis of how parents perceive
and react to decisions about reducing lead paint risks to their children.

Section one of this paper provides a review of the economics literature on household
decisionmaking and of the "mental models" literature related to eliciting decision models. In
section two we set out the methodology used in this study. In section three, we present results.
The implications of these results for design of stated preference surveys is discussed in section
four.

1. LITERATURE
Household Economics Literature

A fundamental problem for economics in studying family decision making is that modern
microeconomics has a subjective, individualistic theory of value, but data are typically collected
at the household level (Vermeulen 2004). As a result, even though households are micro-
societies, it is difficult to infer the role that individuals play within the household. Effectively,
what modern household economics attempts to do is to infer the relationship between
preferences of individuals within the household and the decisions reached by the household from
household level revealed preference data.

Early models aggregate individual utility into a unitary household-level social welfare function.
Samuelson (1956) does this by assuming the family acts as if it were maximizing a weakly
separable household welfare function that is increasing in individual household members' utility,
Wh = W(ui(qi), W2G/2), W3O/3), • • •)• The family is assumed to allocate income across family
members by consensus, Y = y\ +y2+ .... Samuelson (1956) shows that if one can assume that
income is distributed within the family "so as to keep each member's dollar expenditure of equal
ethical worth, the family can be said to act as if it maximizes such a group preference function."
Becker (1974) assumes the household acts as if a benevolent family dictator were allocating total
household purchasing power among family members to maximize a weakly separable and
increasing in the household head's own consumption and other family members' utility, Wh =
Ui(qi u2(112),	...). Unitary models imply an income pooling hypothesis, namely that only

aggregate household income and not individuals' income affects resource allocation within the
household.

2

-------
An alternative approach to modeling household decisions uses game-theoretic models that
explicitly take the behavior of individual household members into account. One major class of
models assumes non-cooperative bargaining (Leuthold 1968, Ashworth and Ulph 1981,
Browning 2000). Household members maximize their own utility taking other household
members' behavior as given. The resulting intra-household allocations may not be Pareto-
efficient. These models imply restrictions on observable household behavior that are not implied
by the unitary household models. The second major class of models assumes cooperative
bargaining (Manser and Brown 1980, McElroy and Horney 1981). Household members bargain
over division of the gains of cooperation that accrue from living as a family. The bargaining
power of household members and assumptions regarding which bargaining strategy is used
determines the specific intrahousehold allocation resulting from the bargaining (McElroy and
Horney 1981, Manser and Brown 1980). These bargaining models allow for the possibility that
the source of non-labor income affects allocation of household resources, i.e., that income is not
pooled. Using this modeling framework, empirical studies have shown that children's health and
welfare outcomes can differ depending on whether mothers or fathers are given transfers of
income (Lundberg et al. 1997, Phipps and Burton 1996, Hoddinott and Haddad 1995, Doss 1996,
Strauss et al 2000). This finding is not explainable by unitary models.

A major criticism of the cooperative and non-cooperative bargaining models has been that it is
not possible empirically to tell whether the household is rejecting a particular choice and or
whether the assumed bargaining structure does not fit the data (Vermeulen 2004).3 More
recently, an alternative class of models that avoids this problem, called collective household
models, has gained acceptance (Bourguignon and Chiappori 1992, Browning and Chiappori
1998). As in other bargaining models, individuals in collective household models maximize
their own utility. But unlike cooperative and non-cooperative bargaining models, collective
models assume only that the outcomes of the bargaining process are Pareto-efficient. In general,
one individual in the household maximizes their own utility from the household allocation of
consumption, leisure and a public good subject to similarly defined utility of other household
members being greater than or equal to their reservation utility. Household allocation of
resources is also assumed to be influenced by the reservation utility (Apps and Rees 1997).
Reservation utility is usually suppressed in formal presentation of collective models because they
are a function of wage and unearned income and are unobservable (Apps and Rees 1997).
Similarly, factors in addition to price, wage and non-labor income, that are recognized to affect
individual utility are generally suppressed in these models formal notation (Apps and Rees
1997). Utility is maximized subject to a pooled budget constraint with income including both
labor and non-labor income. Assuming that individual utility functions are concave and the
budget constraint is convex, the household's problem can be characterized as maximization of a
weighted Utilitarian social welfare function subject to a unified full income budget constraint:

3 This discussion of collective household models draws heavily on Vermeulen's (2004) review of collective
household models.

3

-------
W(p,w,y) = ^ ma^S{p,w,y)ua(qa, qb ,la, lb,q)+(\- S(p,w, y))ub (qa, qb,la, lb ,q)

s.t.

p'q + wala + wblb 
-------
income in empirical studies using revealed preference data to test alternative household models
(Lundberg et al. 1997; Doss 1996). Stated preference studies may be able to directly estimate
relationships implied by alternative household models that are unobservable in revealed
preference data.

Stated preference surveys, by their nature, collect individual level data. As a result, it is critical
to understand the relationship between individual statements and household level choice. For
example, Bateman (2005) shows that unless adults in a multi-adult household fully pool income,
the standard approach of asking one adult to provide household WTP will not give an accurate
estimate of household WTP. It is unclear at present whether respondents are providing their
own preferences or their appraisal of the outcome of a household decision process, whether
unitary or bargained. This problem may be particularly important in valuing children's health
outcomes.

The above models provide a framework for deciding when it is appropriate to ask one member of
the household a stated preference question or when both members need to be asked, as well as
how to ask these questions and what types of supplementary information to request.

If there are differences in preferences, then it may not be adequate to survey a single household
member. One likely way in which preferences may differ between spouses regards their
preferences over health risks. Many studies show gender differences in risk perceptions (e.g.,
Finucane et al. 2000) and some in risk taking (Byrnes et al. 1999). In most cases males are found
to have lower concerns about risk, or perceive risks as being smaller, than females (Davidson
and Freudenburg 1996; Flynn, Slovic and Mertz 1994). In some cases differences male-female
risk attitudes depend on the type of risk being examined or on more complex relationships
between the risk and the individual (Finucane et al. 2000). Nevertheless, differences between
men and women in their risk attitudes appear to be robust findings across various risk categories
and analytical methods. Since risk attitudes affect the form of individual's utility function,
gender differences in risk attitudes could lead to different responses to questions about WTP to
reduce risk to children's health.

Division of responsibility for household production activities may also affect household
decisions affecting children's health. There are several ways in which this could result in
individual preferences mattering in a WTP study, whether the model is a collective model or a
Samuelson type unitary model. Responsibility for a certain class of activity may influence the
weight placed on a person's utility. One possibility might be a domain-specific dictator, as in the
traditional case where, "my wife makes all the decorating decisions." It may also result in
greater weight being placed on the utility of the person with responsibility for a particular
activity in decisions related to that activity, perhaps because the person has gained greater
knowledge about that domain. Finally, in a model with household production of a non-market
good, differences in responsibility for provision of that good would lead to individuals' time
constraints being affected differently. For example, it is possible that if one person has primary
or sole responsibility for home repairs, that the tradeoffs that person is willing to made on
removal of lead paint might differ from those of a partner who has little responsibility for home
repair. Existing collective household models have assumed that bargaining weights are

5

-------
invariant to the decision domain. We hypothesize here that weights vary by decision domain.
Thus the bargaining weight is actually a vector of weights. The value of scalars in this vector
may change with the type of decision being made. This allows for the possibility of
specialization within the household. We also hypothesize that knowledge or skill in household
tasks affects these domain-specific bargaining weights.

The household models also suggest that supplementary information, includes good measures of
income and variables to estimate weights, needs to be collected. Welfare weights are a function
of exogenous variables affecting the standing of individuals' preferences in family decisions
including: information on individuals' unearned income, relative wages, education and other
variables affecting individuals' prospects in the labor market. Because stated preference surveys
rely on individuals' subjective evaluations and because bargaining power depends on both
party's evaluation of their own and the other party's position, it is also important to know
whether individuals differ in their subjective estimates of these variables.

Mental Models Literature

Choice decisions involving multiple parties, like those in a family, are more complex than
individual decisions and may involve hierarchies of choices. Mental models research offers a
systematic way to investigate the structure of individual and group decisions and can provide a
sounder scientific basis on which to design a valuation survey.

Two decades of work in cognitive and decision science has begun to show how people represent
knowledge about their decision environment in mental models (Gentner and Stevens 1983,
Langan-Fox 2000). Craik (1943, p 61) described mental models as "small-scale model[s] of
external reality" that people invoke and 'run' in their heads to see how to understand and explain
the world. These models are associations that exist within long-term or short-term memory and
strongly influence how information is retained, recalled and used in decision settings (Bainbridge
1991).

Recent studies have examined how mental models of decisionmaking in a team setting differ
from those of individuals (Orasanu and Salas 1993, Adelman et al. 1986). A marriage can be
seen as a team, with differentiated roles and responsibilities. Team mental models research
provides a methodological foundation for eliciting mental models of joint decisionmaking
from couples (Rouse, Cannon-Bowers, and Salas 1992, Daniels, de Chernatony and Johnson
1995). Researchers have long assumed that teams work better if members share mental
models of team tasks and processes, and that members' mental models of both task and team
process become more similar - that is, more shared - over time. Levesque et al (2001) found
instead that mental models of team tasks and processes diverged over time, as team members
specialized. Literature on group decisionmaking indicates that individuals in groups often
defer decisionmaking power to those perceived to have more knowledge or experience in the
decision context (Sorkin et al 2001).

Langan-Fox et al. (2000) found cognitive interviewing techniques, including open-ended
questions followed with prompts asking respondents to elaborate, and visual card sorting, to be
useful in eliciting mental models of team decisionmaking. These same methods have been used
successfully to elicit mental models of individual decisions to engage in risky activities, like

6

-------
smoking in adolescents (Lynch 1995) and lay mental models of indoor radon risk and risk
mitigation (Bostrom et al. 1992).

In this study we elicit both individuals' and couple's mental models of lead hazards and of the
couple's (dyadic) decision-making process. We elicited task-specific knowledge (i.e., about lead
hazards), task-related knowledge (i.e., about the couple's risk decision-making), individuals'
risk-related attitudes and beliefs, and knowledge of their partner's risk-related attitudes (cf.
Cannon-Bowers and Salas 2001). The approach extends mental models research used in other
risk domains (e.g., Bostrom et al. 1992, Morgan et al. 2001) by building on team mental models
research (Levesque Wilson and Wholey 2001, Mohammed and Dumville 2001).

2. METHODOLOGY

We conducted in-person interviews with thirty-five couples (70 individuals). Samples of this
size have been found adequate to capture much of the conceptual variability in a substantive
domain (Morgan et al. 1992). Each spouse was first interviewed individually (all couples in the
sample happened to be married); spouses were then brought together and interviewed as a
couple. Finally spouses were again separated and asked to complete a written questionnaire,
which characterized their decisionmaking styles, took sociodemographic information, asked
numerous questions about their relationship and attitudes towards risks in general and lead paint
exposure, in particular.

This survey included three strategies to assess parental decisionmaking: characterization of direct
statements by parents of how they make decisions; analysis of responses to closed-ended
questions about decisionmaking and factors hypothesized to affect decisionmaking in the
literature, and finally; examination of hypothetical decisionmaking about lead paint mitigation.

The study drew from the population of two-parent households in Atlanta, Georgia, with children
under the age of 7, living in housing built before 1979. We limited the population to owner-
occupied housing. Including rental housing would increase the heterogeneity of the sample by
raising additional issues of control over abatement interventions, and by changing the relevance
of control options. Given a small sample size, a decision was made to control for family
structure to reduce heterogeneity. U.S. Census of Housing data was used to identify
neighborhoods in the Atlanta, Georgia Metropolitan Statistical Area with housing stock built
before 1979. Households were sampled from phone number lists by the Survey Research Center
at the University of Georgia and screened for appropriate characteristics in initial phone contacts.
The first fifteen interviews were conducted by research assistants at a central location at Georgia
Institute of Technology. Because of difficulty in recruiting couples to travel to the interviews,
the final fifteen interviews were conducted in couples' homes, by the Survey Research Center
interview staff.

A semi-structured interview protocol was used to investigate parental decision-making behaviors
and their mental models of lead paint risks. Prior to the interviews, each spouse was asked to
write down three recent major children's health decisions. Interviewers selected the highest-
ranking jointly mentioned decision as a focus for the first part of the individual interviews. The

7

-------
individual interviews began with open-ended questions exploring the decisionmaking process
involved in the couples' most recent major children's health decision. Follow-up prompts were
used to assure that issues such as what the problem was, what decision was reached, who
identified the problem, who was involved in the decision, whether prior discussion took place,
who initiated the discussion, what factors were considered, how the respondent felt about the
decision, how their spouse felt about it, and whether this was a typical decision. Open-ended
questions were used to ask about differences between this decision and more routine purchase or
home repair decisions. The next section of the interview dealt with children's environmental
health problems and focused on parental awareness and level of concern about lead paint hazards
compared to other environmental hazards. Finally, each spouse was presented with a
hypothetical lead paint decision scenario and asked to talk through what they thought their
family would do. Follow-up prompts were used to assure that information on the information
desired, factors considered and role of cost in the family decision was collected. After a break,
spouses were interviewed as a couple.

The couple's interview followed much the same protocol as the individual interviews, except that
in the hypothetical lead paint decision, instead of eliciting possible health effects and mitigation
options, the couple was given a list of specific effects and options. They were asked to sort these
by seriousness of concern, effectiveness and likelihood that a mitigation option would be
selected.

Finally, the spouses were again separated and asked to fill out a written questionnaire (see
Appendix I). This questionnaire included questions about household decisionmaking styles in
various domains (e.g., home decorating and home repair), basic demographic information,
homeownership, education, employment and commitment to the labor market, income,
household financial management, time spent in various household production activities, division
of responsibility for specific types of family decisions, beliefs and attitudes about children's
environmental health risks, and knowledge about impacts of lead on children's health.

The written questionnaire also included a set of questions on marital adjustment, the Dyadic
Adjustment Scale (DAS). The DAS is a 32-question instrument developed by Spanier (1976)
to assess the quality of the relationship perceived by married or cohabiting couples. The DAS
remains the most frequently used instrument with different groups of participants and
cultures for assessing the quality of married life (Casas & Ortiz, 1985; Crane, Allgood,

Larson & Griffin, 1990; Shek, 1994). The items for the DAS were those chosen out of an
initial pool of 100 that (a) were normally distributed; (b) discriminated between married and
divorced people; and (c) loaded highly on one of four factors (Dyadic Consensus; Dyadic
Cohesion; Dyadic Satisfaction; and Affectional Expression). Response scales differ across
the questionnaire, with the consensus items including verbally anchored response scales that
represent the extent of agreement or disagreement between the spouse and his or her partner
for each item (from always agree, to always disagree; or from all of the time to never). The
total score is the sum of scores on all items, ranging from 0 to 151. The scale scores have
been found to have good content and construct validity (Spanier, 1976). Spouses with scores
below 98 are classified as discordant (Eddy et al., 1991; Jacobson et al., 1984).

8

-------
Spanier built the DAS on four subscales, one of which, the dyadic consensus subscale, is
particularly relevant to decisionmaking. The dyadic consensus subscale consists of thirteen
items assessing spousal agreement on issues ranging from, for example, handling family
finances, household tasks and amount of time spent together, through friends, ways of
dealing with parents or in-laws, religious matters, major decisions, and philosophy of life.
Two of these items (on major and career decisions) are sometimes used as an alternative
consensus subscale (Busby et al., 1995).

3. RESULTS

In this section, we provide both qualitative and quantitative results concerning
decisionmaking processes across the couples. The former are drawn from the open-ended
oral parts of the interviews; the latter from the written survey. The former as of this writing
cover 19 couples. The latter cover 35 couples.

Qualitative Results

To initiate our personal interviews with parents, we asked each parent to list three recent
major child health decisions, or family decision affecting their child. We then selected the
most important of these listed by both parents independently. The individual and couple
interviews each opened with a request that the parents describe this decision: "Could you tell
me about [the most recent major child health] decision that your family made?" The health
decisions discussed by the nineteen couples included vaccination decisions (4 couples),
toothache, earache or ear surgery (4 couples), accidents (skiing, falling through a window),
illnesses (asthma, fever and cold, food poisoning), what to do about a bleeding birthmark,
and choices about summer camp, high school, and speech therapy. A fourth of the couples
had made the decision in question within the previous six months, another fourth within the
previous year.

Several features of their responses are of interest, including how they structured the decision,
and what kinds of factors they took into account in making it. To learn more about how they
structured the decision, we asked whether they had discussed the decision at the time it was
made and/or prior to that time, and who had initiated those discussions. All couples said that
they discussed the decision, and most had also discussed it previously, for an hour or less. In
almost all cases, couples reported that the mother had initiated the discussion that led to the
decision. The two exceptions were a sole father-initiated discussion, and one couple who
initiated the discussion mutually. In the couple interviews, the couples also
reported that the mother usually initiated such discussions. All of the couples reported
having agreed with the decision.

When asked what factors they took into account in making a major health care decision, in
this case regarding a severe, acute onset ear ache, one couple [4C] responded as follows:

Mother: "just wanted to be sure that she was, that we took care of it. We
wanted to be sure that... She could not go in pain. We had to do
something. We had a fear of long term effects of all these burst eardrums."

9

-------
Mother: "That's about it"

Interviewer: Are there any other factors?

Father: "I am sure we thought about the financial part of it"

Another couple described their child's seizures and epilepsy, and the difficult decision they
had to make whether or not to give her medicine to control them. When asked by the
interviewer to talk about how the decision was made, the father responded first:

Father: "Well, she [mother] discussed it with me, she did the research on
the internet.

Found out exactly what the medicine could do and how it would help her
[daughter]... so"

Mother: "That's after the, a, the neurologist, you know, discussed it with
me. I went home and looked it up, you know... the internet is a great
thing!"

As these conversations illustrate, parents' reports of these decisions emphasize the urgency
of many child health decisions, the empathy parents feel with their children when they are in
pain, that information is usually incomplete, but both the internet and a variety of experts and
friends can be called on to fill in gaps. However, the data suggest that the majority of couples
chose the plan of action that was initially considered or most common.

In terms of learning about the viability of a WTP survey of parents about their children's
health, we were concerned that cost would not be a factor for a significant share of spouses.
The following comment from one mother illustrates our concern:

A couple discussed a decision to take their child to see a specialist about their child's
persistent cough, which was not clearing up. When the interviewer asked the couple "What
were the factors considered in discussing this decision?" the mother [21C] replied: "When it
comes to your kids, there aren't any factors. Their health is the most important thing. Cost,
nothing, that doesn't matter to me."

Yet, the majority of couples said that they considered the quality and effectiveness of the
decision alternatives, for example, the quality of the hospital to which they could take their
child, as well as cost. Eleven couples mentioned costs in their unprompted description of the
decision process they nominated in the beginning of the survey and one mentioned it after
being prompted. However, no couple reported having considered borrowing ability.

Later in the survey, spouses were asked whether cost would play a role in what to do about
lead paint, assuming they found high levels of lead dust in their house. Most who responded
to this question answered affirmatively - 15 wives and 15 husbands said yes, 2 wives and 2
husbands said no. Further, when asked if there were conditions under which they would
choose a cheaper and less effective option, 9 of the 15 wives and 10 of the 13 husbands
answering said yes, suggesting that a majority but not all of the spouses are willing to think
about tradeoffs.

10

-------
Descriptive statistics

For each question in the written questionnaire, there are six sets of statistics: the husband
answering for himself, the husband answering about his wife, the wife answering for herself,
the wife answering about her husband, the answers of the husband for himself and the wife
for herself averaged across the couple, and a variety of statistics at the couple, rather than the
individual level (Table 1). These latter statistics permit us to look at the degree of agreement
in answers across the spouses. Disagreement in responses from spouses of some form is a
necessary condition for it to matter which spouse responds in a stated preference survey.
Many disagreements are what might be termed "mild." The husband says his wife does most
of an activity (like helping the child with homework), the wife says she does all of it. Other
disagreements are more substantial, for example, if the wife were to maintain that she does
all the helping and the husband were to say he does all of it. For factual questions at the
spouse level, we assume the husband's (wife's) answers for himself (herself) are true or
reliable. For questions at the couple level, we will use the average answers in further
analyses.

Demographics. The sample is younger and more educated than the general population. The
average age of men respondents was 36; the average for women was 35. Respondents' ages
ranged from 26-45 years old (table 1). On average respondents had 16 years of education.
African Americans, but not other minorities are well represented. About two-thirds of the
couples were white and the rest were black. Only six percent of respondents were previously
married. Most (57%) have two children, with up to six children (in one family). Because the
interview protocol required families to have at least one child 7 and under, 73% of children
in the study fit this criterion. All were homeowners, in homes built 1979 or earlier,
consistent with the sampling protocol, with average tenure 6 years. Six percent of the
couples had been married previously. In general, there were minor disagreements among
couples on virtually every demographic question except having been divorced. Most of these
are of a level that would qualify as measurement error, but it is interesting to see that this
kind of error is present even on basic factual information about the families.

Employment and Income. Employment and income patterns can affect the weight of
individual preferences in family decision. Ninety-one percent of husbands and 63% of wives
in our sample were employed (table 1). Most husbands (64%) said they worked more than
40 hours a week. Most wives who worked, reported working 20-39 hour per week range.
Not surprisingly, the husbands' contribution to family income was far higher than the wives:
73%) vs. 27%), although seven wives (of 34 answering) contributed over 50% of household
income. Median pre-tax, household income (in 2004) was between $60,000 and $74,000.
However, the wives thought mean family income (in 2004, before taxes) was a bit lower than
the husbands did: $79,860 vs. $83,290.

Spouse's perceptions of their own and their spouse's relative contribution to family income
are also theorized to affect household decisions. Husbands and wives were each asked what
percent of household income they and their spouse contributed. We see from the table 2 and
figure 1 that on average, husbands and wives have the same perception of the amount of
income the wife is contributing. This happy average state of affairs masks significant

11

-------
differences in perception. From the husband's perspective, the worst cases in this study are a
husband who thinks his wife is contributing 40% less than she thinks she is and a husband
who thinks his wife is contributing 20% more than the wife thinks she is. From the wife's
perspective, one wife thinks her husband contributes 75% less than he thinks he does and
another wife thinks her husband contributes 55% more than he thinks he does.

Some of these disagreements may simply be lack of knowledge about what total household
income is. On average, husbands think household income is $3,400 greater than wives do.
But again, at the extremes, one husband thinks their combined income is $45,000 less than
the wife thinks at the other extreme, one husband thinks total household income is $30,000
greater than what the wife thinks it is.

Because attachment to the labor market figures heavily in the empirical literature on
household bargaining, four additional questions were commonly asked to gauge degree of
desire to working outside the home: whether the spouse would prefer to stay at home with the
children, whether the respondent would prefer that their spouse stay home with the children,
whether the spouse's career is more important than the respondent's, and whether the spouse
feels he or she should be the breadwinner in the family. These questions evoke very different
responses in husbands and wives, while the husband and wife generally agree with one
another's assessments. In general, wives want to stay home with their children and do not
want their husbands to do so. Husbands want their wives to stay home with the children, but
have a range of feelings about themselves, not strongly skewed against staying home. Both
wives and husbands generally agree that the husband should be the breadwinner. However,
there is close to indifference about whose career is most important, with an edge to the
husband's, given by both the husbands and wives.

Decisionmaking. This study focuses on financial and health decisions because these are
relevant to children's health valuation. Couples in the study exhibit three general approaches
to household financial management: joint management, separate management, and allocated
or assigned management. In allocated or assigned management, one spouse has a
housekeeping or personal spending allowance and the other spouse manages the rest of the
household money. Within couples' there is general agreement about which model fits. Most
(73%-79%) of the couples managing their money jointly. Most of the rest are in the
assignment mode (table 3). Later in the survey, respondents were asked to make a general
characterization of who makes decisions in the household and then were ask about division
of decisionmaking responsibility about in specific decision contexts. Self-reporting on
decision style may lead to an over-reporting of "joint" decisionmaking because people may
want to view themselves as conforming to a norm that family decisions should be made
jointly.

In their general characterization of who makes decisions in the household, both 88% of wives
and 88%) of husbands said that decisions were made jointly, although there was some
disagreement at the couple level, as discussed below. Only 9% of the husbands said they
made more of the decisions. Once the context was made specific, these percentages
sometimes changed. Most couples make financial decisions jointly but some wives (23%)
and husbands (18%) said that the husbands make more of these decisions. No men said their

12

-------
wives made more of the financial decisions. For decisions involving major purchases, which
is another way of describing financial decisions, 79% of wives and 82% of husbands say
these decisions are made jointly. The remaining respondents are split equally in saying the
wives or themselves make more (or all) of these decisions.

In their characterization of specific decision domains, there was evidence of specialization.
This was particularly prominent in the context of children's health, where only 26% of wives
and 32% of husbands say that decisions are jointly made. 71% of wives say they make the
decisions about doctor visits for their kids, for instance, with only one saying her husband
makes more of these decisions. As a group, the husbands generally agree with their wives on
this issue.

Couples differ in the extent to which they agree about how they make specific decisions or
manage finances. To aid this discussion, we define the following terms: Joint (spouses agree
that they make decisions jointly), Agree (spouses agree that one or the other makes the
decision), Disjoint (where one thinks they make decisions jointly and the other thinks the
situation is different), and Disagree (one thinks one makes the decision and the other thinks
the other makes the decision). Considering the general decisionmaking question first, 26
couples agreed that decisions are made jointly. None agreed that one spouse or the other
makes all, most or more of the household's decisions. The rest of the responses can all be
classified as disjoint. Childcare is one of the domains with the most disagreements: 13
couples agree that they make joint decisions, and 7 couples agree that the wife makes more
decisions. The remaining 14 couples that answered this question are disjoint.

Allocation of time. The amount of time different individuals spend on certain activities may
affect help explain patterns of decisionmaking. It is clear (table 4) that financial tasks are
shared fairly equally in nearly all households, while husbands dominate only home repair and
renovation in terms of the time spent on these tasks. For all other tasks, wives spend more
time than their husbands do and the spouses generally agree on this. In particular, wives
spend more time than their husbands caring for children. One interesting area of
disagreement (or disjointedness) between couples concerns time spent helping children with
their homework. Husbands think they do more of this activity than their wives think they do.
There is also disagreement between spouses about who takes the kids to the doctor. While 22
couples agree that the wife spends more time taking children to the doctor, 13 couples are
disjoint. A similar situation with is found for homework, cleaning the house, spring-
cleaning, decorating, major purchases and financial management.

Marital Adjustment. To test if marital adjustment affects decision making in couples, we use
the 32-item DAS, as described above. Scoring rules differ by question (see Appendix I).
Unhappy couples have been normed to be those with a score of 98 or less. In this study both
spouses completed the DAS questionnaire. Husbands' scores range from 61 to
138 with an average of 113 (table 5, figure 2). 12% of the husbands rate their marriage with
a 98 or lower. Wives' scores have a wider range (59-143), but the average is the same as the
husbands at 113. Only 9% of the wives scored their relationship 98 or less. In some couples,
spouses have different scores on this 32-item scale. In eight couples, the wife's marital
adjustment score is 10 points or more above her husband's and for another six couples the

13

-------
husband's score is ten or more points greater than the wife's. All told there is only one couple
with both spouses rating their relationship at or below the cutoff score of 98.

There is a subset of 13 questions in the DAS in which respondents are asked how often they
agree or disagree on specific decision areas such as handling family finances, religious
matters, recreation, dealing with parents or in-laws, etc. Among these thirteen decision areas,
there is almost no area in which one spouse says that they always or almost always agree and
the other spouse says they always or almost always disagree. We do find some serious for
which both spouses acknowledge disagreements: in-laws, the amount of time the couple
spends together, leisure interests and activities, and career decisions.

In this series of questions, there are also more factual questions, such as "How often do you
and your spouse quarrel?" Differences in spouses' responses on these questions could be
problematic because they indicate different perceptions about the quality of the marriage.
For these questions, serious differences in couples responses are defined as two or more
points of difference on the five-point scale in which 0 indicates poor marital adjustment and
5 indicates high. The questions "How often do you engage in a stimulating exchange of
ideas?", "how often do you calmly discuss something?", and "how often do you work on a
project together" provoked some serious differences between spouses' responses. There
were also serious differences in responses on yes/no questions including the question about
whether being too tired for sex has caused problems. Twelve couples had one spouse say
Yes and the other say No. Six couples both said Yes and fourteen couples both said No.

Attitudes Towards Risk. Another factor that could influence decisionmaking is attitudes
towards risk. To gauge such attitudes about lead exposure, in the oral section of the survey
we asked spouses whether they were worried about lead paint. Fewer husbands (6 yes, 13
no) said they had worried about lead paint than wives (11 yes, 8 no).

In the written survey, we placed asked respondents to rank eight health risks, including lead
paint, according to various dimensions of qualitative and quantitative risks. These other
health risks included air pollution, climate change, radon, small pox, small pox vaccine,
anthrax and influenza.

The results are voluminous, but the main ones are: (i) flu and air pollution are viewed as the
most risky with lead in the middle of the group, assessed equally by the wives and husbands
and viewed by both parents as a bigger risk to children than to the overall population.

Climate change was the most "unknown" risk, anthrax the most "serious," and climate
change had the longest lead time. Air pollution is viewed as the risk causing the most
exposure. For lead, wives think exposures are more widespread than husbands do, as we saw
in oral responses.

In table 6, we show detailed results for the qualitative risk dimension "controllability" across
the eight risk categories for husbands and wives. Here, we supply the percentage of each
gender ranking each risk as most controllable down to the least controllable. We find that
lead paint is seen as the most controllable risk (not surprising, given the alternatives) and that
there is more disagreement about this across the husbands than the wives.

14

-------
Knowledge About Lead. As noted above, knowledge about a topic of concern may help
explain patterns of specialization in decisionmaking. It may also be an indicator of ability to
or interest in searching out or absorb information relevant to decisionmaking. This ability
may help explain patterns of specialization. We looked at this issue both in the oral and
written parts of the survey. In the oral section, we asked spouses separately and the couple
together to consider a hypothetical decision concerning lead paint mitigation: "How much do
you know about the health risks from lead paint (own knowledge); how much do you think
your spouse knows?"

Interestingly, both husbands and wives thought their spouses knew more about health effects
than they did, on average. On this question, 7 couples agreed on how much the wife knows.
The average score of women on their own knowledge was 2.8, with the husbands giving their
wives an average score of 3.3. Their responses were positively correlated, r = 0.62. There
was somewhat less agreement on how much husbands know, with only 5 couples giving the
same the estimates for the husband's level of knowledge, with husbands rating their own
knowledge at 2.5, and wives rating their husbands' knowledge at 3.2 on average, (one-tailed
paired /-test, p < 0.05), r = 0.11.

In the written part of the survey, we asked thirteen true-false-no opinion questions to test for
knowledge about lead and its effects. Overall, the wives as a group are more often right than
the husbands (if we simply sum up right answers over all 13 questions) (table 7, figure 3).
On average, the wives got 10 questions right, the husbands nine. The questions most
frequently missed by both groups are whether lead absorption is greater when a person has
iron deficiency (TRUE), and whether lead exposure can lead to hypertension (TRUE).

It is plausible that when one spouse has more knowledge about a problem than another, that
spouse might take a greater role in the decision over what to do about that problem. We
therefore tallied up the number of times a husband had a different answer to the lead
knowledge questions than the wife did, and in which direction. We found that for six couples
the wife outperformed the husband, being correct on four or more questions her husband
missed. Correspondingly, we only found two couples where the husband outperformed the
wife on four or more (four) questions.

Regression Analysis

Below, two types of regression analyses are presented. The first explains couples'
decisionmaking in each of five decision domains specifically relevant to children's health
valuation: child doctor visits, childcare, paying bills, family income management and
household purchase decisions. The second pools responses to all ten decisionmaking
domains examined in the survey to explain spouses perceived decisionmaking. Both
analyses feature observations at the spouse level (rather than the couple) because husbands
and wives may disagree. The data set includes responses from 35 couples, or 70 spouses.
Table 8 presents definitions of the dependent and explanatory variables.

15

-------
Hypotheses. As discussed above, recent developments in the economic theory of household
decisions focus on the weight that different household members' preferences have in
household decisions, the determinants of that weighting, and whether decisions outcomes are
the result of bargaining. The fundamental feature of these models is that factors that affect
spouses' options outside the marriage influence their bargaining power in household
decisions. On this basis, we would expect that relatively exogenous factors related to
employment decisions such as: relative income or wage rates, relative education levels, the
level of commitment to work, unemployment spells and the extent to which both spouses
work full time will influence household decisions. We would also expect that factors that
reflect the quality of communication in the marriage, here measured by the DAS, could affect
household decisions. Finally, literature on group decisions suggests that relative levels of
knowledge about a problem affects who influences group, i.e., family, decisions.

A fundamental empirical problem for the household literature is that the primitives of this
model are unobservable. Most commonly, empirical work testing for the appropriateness of
alternative models has taken a revealed preference approach relying on household level
consumption and labor supply outcomes as measures of the outcomes of household decisions
(Phipps and Burton 1998, Lundberg et al. 1997, Strauss 2000). Several studies have used
purchases that benefit only specific members of the household, like women's or children's
clothing purchases, as a measure of the influence of those individuals' preferences in family
purchase decisions. Dosman and Adamowicz (forthcoming) elicit individual and couples
choices in a conjoint stated preference survey and use this to estimate implied household
welfare weights. In the study presented here, the observable outcome is who plays a role in
household decisionmaking. We define the dependent variable as whether a decision is made
jointly or by one of the spouses alone. To explain variation in this dependent variable, we
use data collected on a wide range of independent variables that household economic models
and mental models suggest could influence the role of individual preferences in family
decisions.

Assuming that an individual's preferences play a greater role in the household decision if
they are involved in the decision, we can use the distinction between joint and individual
decisionmaking as an indicator of whose preferences carry weight in household decisions.
Obviously this conclusion might not hold if altruism plays a strong role in the way families
make decisions. For example, Becker's family dictator takes the utility of other household
members into account. This conclusion also may not hold if the couple agrees to specialize
in decisions over various domains. As a result, this work should be viewed as a means of
getting at stylized facts about household decisions that will be tested more rigorously in our
planned stated preference survey research.

We have specific hypotheses about how some of our independent variables affect the
likelihood that decisions are made jointly or by an individual spouse. For instance, we
expect that where income contributions of the spouses are more equal we are more likely to
see joint decisionmaking. For other variables, we do not have hypotheses about the direction
of an effect but do expect that an effect could be present, for example, for race, income,
education or age. Note also that some of the explanatory variables, such as time allocation,
are themselves endogenous. At this point in the analysis, we have not attempted to estimate

16

-------
more complex models to account for this. We also recognize that, ideally, we should use
multinomial logit techniques to analyze these data, as decisionmaking could be joint, the
husband's lead or the wife's lead. However, there are not enough observations about all
three options for any decision variable to justify this more complex approach.

The fundamental question we seek to address is whether it matters for stated preference
survey research which adult in the household is interviewed. Even if a unitary model
properly describes household behavior, it would matter who researchers interview in a stated
preference study if there is specialization of responsibility for and knowledge about particular
household decisions. Responsibility could vary by domain. This would suggest that the less
the difference in the amount of time spouses spend on a household task, the more likely it is
that decisions about that domain would be made jointly. We hypothesize that the larger the
number of children, the more likely it is that spouses will specialize between home and
market labor and the less likely it will be that decisions about children will be made jointly.
Another way in which such specialization might arise is if one of the spouses specializes in
information gathering (Sorkin et al. 2001). As a proxy for knowledge levels we use correct
responses to a set of knowledge questions about lead paint hazards as an indication of
information gathering performance. The less the difference in this variable, the more likely
that decisions will be made jointly.

Regression results. As noted, due to small sample size we construct a bivariate dependent
variable from the multivariate variables on who makes decisions. We restrict decision

outcomes to a dummy variable taking a value of zero for joint decisions and one for "	

makes more of the decisions." For many decision domains there is a strong gender bias in
decision responsibility across the 35 couples. So for example, for childcare decisions no
respondents said husbands made this decision, while 28 said the wife made the decision.

Forty respondents said childcare decisions were made jointly. In this case we dropped the
observations for "husband makes most child care decisions" and constructed a binary
variable with 0 for joint decisions and one for "wife makes more of the decisions." A similar
pattern was followed for decision domains where few wives made more of the decisions. For
decision domains that did not exhibit strong gender bias, we addressed the small sample
problem by constructing a dummy in which zero indicates a joint decision and one indicates
that one spouse or the other makes more decisions. In this case no observations are dropped.
For income pooling, the dependent variable is defined as 0 if the couples manage their
financial accounts jointly and 1 if they do not.

Table 9 provides the results from logit regression on two child-related decisions and three
financially-related decisions: taking children to the doctor (1 = wife; 4 husband decision
makers dropped); childcare decisions (1= wives make decisions; none dropped), paying bills
(1 = either husband or wife makes decision; none dropped), financial decisions (1 = husband;
2 wife decision makers dropped), and household income management (1 = either husband or
wife makes decision; none dropped).

Household income and whether one or both spouses spend time on a task are significant in
explaining both child-related and financial decisions. The higher household income, the less
likely it is that decisions about taking children to the doctor, childcare and finances are made

17

-------
jointly. For decisions about taking children to the doctor, paying bills, and managing
finances, decision are more likely to be specialized when one or the other spouse spends most
of the time on this task. Years of education is significant in explaining some of the finance
decisions, but not the child-related decisions. The more years of education, the less likely it
is that couples pay bills jointly and the more likely it is that they pool their income. Oddly,
the greater the difference in education, the more likely childcare decisions will be made
jointly. The more children there are in the family the more likely it is that decisions on
whether to take children to the doctor and bill paying are made jointly. Number of children
is not significant for any other decisions.

There are a cluster of independent variables related to employment and income. Total
household income affects both child-related decisions as well as general financial
management decisions. The higher the income, the more likely it is that couples will not
make child-related decisions and general financial decisions jointly. The wife being
employed is associated with joint childcare decisions. However, given that the wife is
employed, the greater the wife's share of the household income, the less likely it is that
childcare decisions will be made jointly. Also the greater the wife's share of the household's
income, the more likely it is that financial decisions will be made jointly. The lower the
index of commitment to working in the labor market, the less likely it is that decisions to take
children to the doctor will be made jointly. The implication is that where wives are at home
and happy about it, they are more likely to specialize in making child medical decisions.

There are a number of results that are difficult to explain. It is not clear why a high score on
the marital consensus subscale would be associated with joint decisions about taking children
to the doctor, but would not be significant for other decisions. We use knowledge about lead
gained through earlier parts of the survey (or from prior knowledge) as an indicator of an
interest in and ability to obtain knowledge in general. It is not clear why this would be
associated with joint decisions on taking children to the doctor and paying bills, but not other
decisions. The index of beliefs about controllability of risks to children is associated with
non-joint decisions about paying bills. It is conceivable that there is some correlation
between a sense of being able to control risks with a willingness not to have as tight joint
control over bill paying, but this is speculation on our part.

Table 10 presents results of logit regressions on the pooled set of decisions in all ten
decisions domains. As noted above, much of the household decisionmaking literature
assumes that decisionmaking models are invariant to the decisionmaking context or domain.
With our data we can test this proposition by lumping together all the decision-making
responses (10 domains per survey) to create a 700-observation dataset. As before, dummy
variables were created to indicate whether the decisionmaking model was classified by the
respondent as joint or other.

First, decision domain matters. All dummies for domains (but one) are significant (against
the childcare default dummy and the show significant differences in some instances with one
another, clustering in two groups, one where joint decisionmaking is more likely and the
other where either spouse specializing is more likely. Once domain is controlled for, then the
effect of gender on decisionmaking style (which we see in chi-square tests) is eliminated.

18

-------
A number of patterns identified in the individual decision analysis become even clearer in the
pooled analysis, controlling for domain. As in the individual decision domains, higher
income is associated with specialization, however, the effect could only be detected when
income is included as a categorical variable (above or below median income) and not when it
is included as a continuous variable. More years of education are also associated with
specialization. Age is associated with a higher likelihood of joint decisions, but only once
we use the income dummy variable. The more children a couple has, the more likely they
specialize. Similarly, holding constant household income, age, number of children,
education and marital consensus, households in which wives are employed also specialize.
Finally, all else constant, the higher the DAS subscale score for marital consensus, the more
likely it is that decisions are made jointly.

4. CONCLUSIONS AND IMPLICATIONS

The major concern of this paper is whether asking WTP questions of one parent in a two-
parent household will lead to an accurate representation of household willingness to pay.

This paper does not directly address this issue, in the sense that we do not ask different
parents their household WTP and compare their responses. That is our next step. This paper
addresses a prior question, although one that has implications for a WTP survey, which is
whether decisions in a variety of domains are made jointly or there is specialization by
spouses and what factors drive this difference. We infer that if spouses in the household
specialize in decisionmaking then asking different spouses a WTP question is more likely to
lead to different answers.

This investigation is informed by both the economics literature on household behavior and
the mental models literature on group decisionmaking. It looks to the economics literature
for variables that are expected to influence the relative role of different spouses in household
decisions. It looks to the mental models literature both for factors that influence the role of
individual's in-group decisions and for methodology to systematically study couples'
decisionmaking processes.

In general, decisionmaking style varies by domain and is affected by variables that are
expected from theory to contribute to power (welfare weights) in the relationship, such as
income share, wife employment status, work commitment, and differences in education.
From the literature on mental models, as well as from an interpretation of the household
behavior literature, we also expect and find evidence for effects of domain knowledge, time
spent in the domain, and marital consensus.

We also learned several lessons for a future WTP study. The most important is that the
majority of couples appear to consider cost in their major decisions about children's health
and, specifically, in response to our hypothetical question about decisions in response to a
finding of high lead levels in the home. They were willing to make tradeoffs with
effectiveness and cost. Another lesson is that there are gender differences in risk attitudes
and risk perceptions, e.g., wives think lead exposure is more widespread than husbands and

19

-------
are more worried about its impact on children. However, they both view the lead paint
problem as equally controllable. A further lesson is that it is a viable strategy to administer a
survey to couples and spouses separately.

There are many caveats to these conclusions. The sample is too small to do a more thorough
test of decisionmaking styles using MNL techniques to capture the three styles: husband
decides, wife decides, joint decision. With a larger sample we could remove the ambiguity
of the "other" answer. Future work will involve almost ten times this sample size. In
addition, we will need to address endogeneity issues associated with some of our explanatory
variables. Further work is also needed on understanding differences between spouses
responses - both on opinion and facts - across couples. For instance, spouses seem to have
significant differences in perceptions about what household income is and different
perceptions about their own and their spouses' income share.

20

-------
Table 1. Descriptive Statistics of Selected Variables.

Husbands

Wives

Total

Age

Race

(Non-White=l)

Previous Marriage (%)

Number of Children

Age of Children

Age of Oldest Child
Education (Years)

Education Difference
(Husband - Wife)

Absolute Value

House Built (Year)
Years in House

Employed (%)

Full-time (%)

Part-time (%)

Total Household
Income ($000)*

Contribution to
Income (%)*

Dyadic Scale of
Marital Adjustment
(32-Item)	

13-Item DAS
Scale

Index of Work
Commitment

Mea
n

Std

Dev

Min

Max

36.21

6.08

30.50

50.50

0.29

0.46

0.00

1.00

0.14

0.36

0.00

1.00

2.26

1.04

0.00

1.00

5.79

4.84

0.25

28.00

7.68

5.85

0.58

28.00

16.06

2.45

12.00

19.00









-

-

-

-

1954

22.24

1903

1979

5.50

3.15

0.08

11.50

0.91

0.29

0.00

1.00

0.74

0.44

0.00

1.00

0.14

0.36

0.00

1.00

83.29

35.15

22.00

142.0
0

72.61

27.52

0.00

95.50

113.0

6

15.95

61.00

138.0
0

48.78

8.25

25.00

63.00

2.32

1.36

0.00

4.00

Mean

Std

Dev

Min

Max

34.56

6.38

21.50

50.50

0.31

0.47

0.00

1.00

0.06

0.24

0.00

1.00

2.26

1.04

1.00

6.00

5.58

4.89

0.00

28.00

7.52

5.90

0.58

28.00

16.09

2.23

12.00

19.00









-

-

-

-

1954

21.32

1907

1979

6.40

5.26

0.08

30.00

0.63

0.49

0.00

1.00

0.17

0.38

0.00

1.00

0.46

0.51

0.00

1.00

79.86

35.49

22.00

142.0
0

26.51

32.56

0.00

95.50

112.9
4

15.25

59.00

143.0
0

49.73

4.36

37.00

61.00

0.32

0.53

0.00

2.00

Mean

Std

Dev

Min

Max

35.39

6.25

21.50

50.50

0.30

0.46

0.00

1.00

0.10

0.30

0.00

1.00

2.26

1.03

1.00

6.00

5.69

4.85

0.00

28.00

7.60

5.84

0.58

28.00

16.07

2.32

12.00

19.00

0.00

1.97

-4.00

5.00

1.23

1.53

0.00

5.00

1954

21.61

1903

1979

5.96

4.34

0.08

30.00

0.77

0.43

0.00

1.00

0.46

0.50

0.00

1.00

0.30

0.46

0.00

1.00

81.57

35.11

22.00

142.0
0

49.90

37.85

0.00

95.50

113.0
0

15.49

59.00

143.0
0

49.26

6.54

25.00

63.00

1.32

1.44

0.00

4.00

*Total Household Income and Contribution to Household Income values are created by
taking the midpoint of respondent-selected income intervals.

21

-------
Table 2. Income Related Disagreements.

Husband's
Contribution to
Income (%)

Wife's Contribution
to Income (%)	

Total Household
Income ($000)

Husbands View Wives View

Mean

Std

Dev

72.61

27.52

28.68

31.61

83.29

35.15

Mean

Std

Dev

74.00

28.46

26.51

32.56

79.86

35.49

Difference Within Each Couple
(Husband-Wife)

Mean

Std

Dev

Min

Max

-0.44

18.52

-55.00

75.00

0.77

10.87

-40.00

20.00

3.43

14.59

-45.00

30.00

22

-------
Table 3. Decisionmaking and Income Pooling Statistics.

Decisionmaking

DM: General

DM: Home Repairs

DM: Decoration

DM: Childcare

DM: Paying Bills

DM: Kids to Doctor

DM: Kids Clothing
DM: Car

DM: Major Purchases
DM: Finance

Husbands View
(%)

Husband

Wife

Joint

DM

DM

DM

9

3

88

50

3

47

6

67

26

0

41

59

38

35

26

9

59

32

3

74

24

18

0

82

9

9

82

18

0

82

Wives View
(%)

Husband

Wife

Joint

DM

DM

DM

3

9

88

47

3

50

6

70

24

0

41

59

35

29

35

3

71

26

0

67

33

18

6

76

12

9

79

23

6

71

Couple-Level Agreements
(%)

Agree
Joint

Agree
Non-
Joint

Disagree

Disjoint

76

0

0

24

38

41

0

21

12

61

3

24

38

21

0

41

21

59

0

21

15

53

3

29

18

61

0

21

68

6

3

24

68

6

0

26

65

12

0

24

Income
Pooling

Pooled Income?

Husbands View
(%)

Pooled
Income

Separate

Assigned
Manage-
ment

73

0

27

Wives View
(%)

Pooled
Income

Separate

Assigned
Manage-
ment

77

6

17

Couple-Level Agreements
(%)

Agree
Pooled

Agree
Assign

Disagree
Pool vs.
Assign

Disagree
Pool vs.
Separate

61

9

27

3

23

-------
Table 4. Time Allocation Statistics.

Decisionmaking

TA: Infant Care

TA: Sick Kids

TA: Kids to Doctor

TA: After School Care

TA: Kids Homework

TA: Teacher Meetings

TA: Cleaning House
TA: Seasonal
Cleaning	

TA: Decoration

TA: Home Repairs

TA: Home
Renovation

TA: Paying Bills

TA: Finance

TA: Major Purchases

Husbands View
(%)

Husband
Spends
More
Time

Wife
Spend
s More
Time

Joint
TA

3

84

13

9

83

9

3

74

23

6

68

26

0

52

48

4

44

52

6

54

40

17

37

46

9

54

37

71

3

26

59

6

34

40

40

20

37

29

34

23

17

60

Wives View
(%)

Husband
Spends
More
Time

Wife
Spend
s More
Time

Joint
TA

4

82

14

6

83

11

3

83

14

0

75

25

0

78

22

7

54

39

6

69

26

12

50

38

6

68

26

69

6

26

60

6

34

34

40

26

37

31

31

21

29

50

Couple-Level Agreements
(%)

Agree
Joint

Agree
Non-
Joint

Disagree

Disjoint

7

79

0

14

6

83

3

9

0

63

0

37

15

63

4

19

24

57

0

19

28

32

0

40

20

54

0

26

26

32

9

32

15

44

6

35

14

63

0

23

25

56

0

19

17

71

0

11

17

48

3

31

38

23

6

32

24

-------
Table 5. Descriptive Statistics for the Dyadic Adjustment Scale (32 Questions).

Handling Finances

Recreation

Religious Matters

Showing Affection

Friends

Sex Relations

Proper Behavior

Philosophy of Life

Dealing With In-laws
Aims and Goals

Time Spent Together
Major Decisions

Household Tasks

Leisure Time

Career Decisions

Discuss Divorce

Leave After a Fight

Things Going Well

Confide in Mate

Regret Marriage

Quarrel

Get on Nerves

Kiss your mate

Engage in Interests

Exchange of Ideas

Laugh Together

Calmly Discuss

Project Together

Too Tired For Sex

Not Showing Love

Overall Happiness

Husbands

Future of Relationship

Mean

Std

Dev

Min

Max

3.56

0.82

1.00

5.00

3.79

0.73

2.00

5.00

3.82

0.80

2.00

5.00

3.85

0.99

1.00

5.00

4.03

0.87

1.00

5.00

3.76

0.94

1.00

5.00

3.76

0.96

2.00

5.00

3.81

0.78

2.00

5.00

3.68

0.84

1.00

5.00

4.12

0.77

2.00

5.00

3.74

0.90

1.00

5.00

4.03

0.80

2.00

5.00

3.21

0.98

0.00

5.00

3.56

0.99

0.00

5.00

3.94

1.01

1.00

5.00

4.44

0.70

2.00

5.00

4.56

0.66

3.00

5.00

3.32

1.32

0.00

5.00

3.65

1.57

0.00

5.00

4.62

0.70

2.00

5.00

3.45

0.75

1.00

5.00

3.32

0.84

1.00

5.00

3.56

0.86

1.00

4.00

2.56

0.75

1.00

4.00

3.24

1.23

0.00

5.00

4.27

0.72

3.00

5.00

3.97

0.98

2.00

5.00

2.70

1.33

0.00

5.00

0.61

0.50

0.00

1.00

0.76

0.44

0.00

1.00

3.97

1.45

0.00

6.00

4.55

0.51

4.00

5.00

Wives

Mean

Std

Dev

Min

Max

3.62

0.65

2.00

5.00

3.94

0.34

3.00

5.00

3.85

0.86

1.00

5.00

3.91

0.67

2.00

5.00

3.94

0.55

3.00

5.00

3.68

0.81

1.00

5.00

3.68

0.64

2.00

5.00

4.00

0.65

2.00

5.00

3.85

0.66

3.00

5.00

4.03

0.68

2.00

5.00

3.94

0.60

3.00

5.00

3.79

0.64

2.00

5.00

3.38

0.65

2.00

5.00

3.82

0.52

3.00

5.00

3.88

0.54

3.00

5.00

4.36

0.99

0.00

5.00

4.61

0.56

3.00

5.00

3.45

1.25

0.00

5.00

3.61

1.69

0.00

5.00

4.58

0.71

2.00

5.00

3.39

0.75

1.00

5.00

3.12

0.78

0.00

4.00

3.61

0.75

1.00

4.00

2.58

0.75

0.00

4.00

3.55

1.03

1.00

5.00

4.21

0.82

2.00

5.00

4.09

0.80

2.00

5.00

2.73

1.40

1.00

5.00

0.61

0.50

0.00

1.00

0.79

0.42

0.00

1.00

4.25

1.22

2.00

6.00

4.16

0.68

2.00

5.00

Difference Within Each
Couple
(Husband-Wife)

Mean

Std

Dev

Min

Max

-0.06

0.69

-2.00

1.00

-0.15

0.70

-2.00

2.00

-0.03

0.80

-1.00

2.00

-0.06

0.89

-2.00

2.00

0.09

0.93

-2.00

2.00

0.09

0.98

-2.00

3.00

0.09

0.93

-1.00

2.00

-0.16

0.85

-2.00

1.00

-0.18

0.87

-2.00

1.00

0.09

0.77

-1.00

2.00

-0.21

0.91

-2.00

2.00

0.24

1.02

-2.00

2.00

-0.18

1.03

-3.00

2.00

-0.26

1.16

-4.00

2.00

0.06

1.10

-3.00

2.00

0.09

1.07

-3.00

4.00

-0.06

0.70

-1.00

1.00

-0.12

0.65

-1.00

1.00

0.00

1.25

-2.00

5.00

0.03

0.73

-2.00

1.00

0.09

0.59

-1.00

2.00

0.21

0.70

-2.00

1.00

-0.03

0.77

-3.00

2.00

0.00

0.87

-2.00

3.00

-0.31

1.53

-5.00

3.00

0.09

0.89

-1.00

2.00

-0.06

1.11

-2.00

2.00

-0.06

1.66

-3.00

4.00

0.00

0.62

-1.00

1.00

0.00

0.44

-1.00

1.00

-0.23

1.48

-5.00

3.00

0.35

0.95

-1.00

3.00

Sum of Responses
(32-Item Scale)



113.06

15.95

61.00

138.00



112.94

15.25

59.00

143.00



0.12

15.14

-23.00

47.00

Sum of Responses
(13-Item Scale)



48.78

8.25

25.00

63.00



49.73

4.36

37.00

61.00



-0.97

7.49

-14.00

18.00

25

-------
Table 6. Ranking of Hazard Controllability by Spouse - % of Respondents for each
Possible Ranking (l=Most Controllable).

Rank

Air
Pollution

Climate
Change

Radon

Lead
Paint

Smallpox

Smallpox
Vaccine

Anthrax

Influenza

H

W

H

W

H

W

H

w

H

W

H

w

H

W

H

W

1

9

3

3

3

35

20

69

54

12

17

54

71

9

3

14

14

2

9

0

3

0

24

23

20

34

3

11

3

6

6

3

9

6

3

17

6

11

6

21

31

0

11

15

11

20

3

17

9

17

34

4

14

34

11

17

9

14

3

0

21

29

11

12

20

23

34

23

5

23

31

20

14

6

9

6

0

29

14

9

3

20

20

11

17

6

9

11

6

14

3

3

3

0

9

9

3

0

14

26

11

3

7

17

14

31

26

3

0

0

0

12

9

0

3

14

11

3

0

8

3

0

14

20

0

0

0

0

0

0

0

3

0

6

0

3

26

-------
Table 7. Descriptive Statistics for Questions About Lead Knowledge.

Lead can be found throughout the
environment. (TRUE)	

Lead-based paint is rarely found in pre-
1978 housing. (FALSE)	

Young children are less vulnerable to
lead poisoning. (FALSE)	

Young children are more likely to come
into contact with lead if it is in their
environment. (TRUE)	

Iron deficiency may increase
vulnerability to lead poisoning. (TRUE)

Children absorb and retain relatively
less lead than adults. (FALSE)	

Lead poisoning can decrease a person's
1Q. (TRUE)	

Lead poisoning can cause respiratory
problems. (TRUE)	

Lead can be found in the blood, brain,
and bones. (TRUE)	

Lead poisoning can lead to lower
school performance. (TRUE)

Lead does not contribute to
hyperactivity in children. (FALSE)
Lead dust can be found in windowsills
in houses that are contaminated.
(TRUE)	

Cleaning can help minimize lead dust.
(TRUE)	

Husbands

Number of Correct Answers

Mean

Std

Dev

0.86

0.36

0.71

0.46

0.80

0.41

0.69

0.47

0.29

0.46

0.74

0.44

0.80

0.41

0.71

0.46

0.89

0.32

0.94

0.24

0.23

0.43

0.91

0.28

0.80

0.41



9.37

2.38

Wives

Mean

Std

Dev

0.97

0.17

0.77

0.43

0.91

0.28

0.83

0.38

0.54

0.51

0.66

0.48

0.80

0.41

0.74

0.44

0.91

0.28

0.91

0.28

0.40

0.50

0.91

0.28

0.60

0.50



9.97

2.15

Difference Within Each
Couple
(Husband-Wife)

Mean

Std

Dev

Min

Max

-0.11

0.40

-1.00

1.00

-0.06

0.59

-1.00

1.00

-0.11

0.40

-1.00

1.00

-0.14

0.55

-1.00

1.00

-0.26

0.61

-1.00

1.00

0.09

0.56

-1.00

1.00

0.00

0.49

-1.00

1.00

-0.03

0.62

-1.00

1.00

-0.03

0.38

-1.00

1.00

0.03

0.38

-1.00

1.00

-0.17

0.57

-1.00

1.00

0.00

0.24

-1.00

1.00

0.20

0.63

-1.00

1.00



-0.60

2.75

-7.00

4.00

*l=Question answered correctly; 0=Question answered incorrectly.

27

-------
Table 8. Variables Used in Regression Analyses.

Explanatory Variables	Definition

Non-White

0 = White; 1 = Non-White.

Male

0 = Female; 1 = Male.

Age

Age in Years (spouse variable as well).

Educational Difference

Husband's Age minus Wife's Age (with Absolute Value option).

Education

Years of Education (spouse variable as well).

Total Household
Income

Income in ($000). Constructed from the midpoint of respondent-selected income
intervals. Also used a dummy variable (1= Above median income)

Share of Household
Income

Respondent's share (%) of household income. Constructed from the midpoint of
respondent-selected income intervals.

Children

Number of Children.

Wife Employed

0 = Wife not employed; 1 = Wife Employed in household of the respondent.

Wife Fulltime Job

0 = Wife does not work fulltime; 1 = Wife works fulltime in household of the
respondent

Interactions with
Income & Income
Shares

Wife Employed and Wife Fulltime Job Dummy interacted with Total Household
Income and Wife's Share of Income.

Index of Work
Commitment

Index of commitment to the work force generated from responses regarding
desires for staying at home, careers, and who should be the breadwinner. Values
range from 0 to 4 with higher values implying greater desired commitment to
work.

Previous Marriage

0 = No previous marriage; 1 = Married Previously.

Age of Oldest Child

In years

Dyadic Scale of Marital
Adjustment

Total Score on the 32-item "Dyadic Adjustment Scale" to assess marital
adjustment. Possible range from 0 to 151, with higher numbers indicating better
adjustment

DAS Subscale of
Marital Consensus

13-item subscale of above

Performance on Lead
Knowledge Questions

Number of correct answers about lead-health knowledge (true-false).

Index for Beliefs About
Children's Risk Levels
for Hazards

Index identifying beliefs about current risk levels over 8 hazards. Values range
from 0 to 8 with higher values implying higher perceived risk.

Index for Beliefs on
Controllability of
Children's Risks

Index identifying beliefs about controllability of children's risks from 8 hazards.
Higher values imply greater controllability.

Time Allocation Series

Variables for each time allocation domain (see Table 4). 0 = Joint Time
Allocation; 1 = Non-Joint Time Allocation (i.e., either wife spends more time or
husband spends more time in this domain).

Dependent Variables

Definition

DM: Kids to Doctor

Who decides when the children go to the doctor? 0 = Joint; 1 = Wife.

DM: Childcare

Who makes childcare decisions? 0 = Joint; 1 = Wife.

DM: Paying Bills

Who makes decisions about paying bills? 0 = Joint; 1 = Other.

DM: Financial

Who makes major financial decisions? 0 = Joint; 1 = Husband.

Income Pooling

Arrangement for managing household income. 0 = Pooled; 1 = Not Pooled.

DM: Total

Used in domain regressions (Table 10). All 10 DM domain variables are pooled so
that there are 10 observations per respondent. 0 = joint; 1 = Other

28

-------
Table 9. Selected Logit Regression Results for Different Specifications of the
Decision Variables at the Individual Level (N=70 or less).

Dependent Variables

Independent
Variables

DM:
Doctor

DM:
Childcare

DM:
Paying Bills

DM:
Finance

Income
Pooling

Wife vs. Joint

Wife vs. Joint

Other vs. Joint

Husband vs.
Joint

Other vs.
Pooled

Non-White



Joint*







Educational Difference
(Abs Value)

Joint

Joint*







Years of Education





Other§
(dummy)

Husband
(dummy)

Pooled*

Total Household
Income

Wife§
(Dummy for
above median
income)

Wife*

Joint

Husband§



Number of Children

Joint**



Joint§





Wife's Share of
Household Income



Wife**



Joint§
(dummy)



Wife Employed



Joint**





Other§

(Wife Employed) x
(Wife Inc Share)









Pooled

Index of Work
Commitment

Joint§







Other

Dyadic Scale of Marital
Adjustment (13
Questions)

Joint*









# of Lead Questions
Answered Correctly

Joint§



Joint**





Index for beliefs about
children's risk levels for
hazards





NA

NA

NA

Index for Beliefs on
Controllability of
Children's Risks



Wife

Other*
(dummy)





Time Allocation for
Related Tasks

Wife§
(TA: Doctor)



Other**
(TA: Bills)

Husband**

(TA:
Purchases)

Other*
(TA: Finance)

Note: To interpret this table, of a sample of couples where decision about childcare are
made either jointly or predominately by the wife, non-white respondents are more likely
to report that such decisions are made jointly.

Note: § indicates significance at the 10% level; * indicates significance at the 5% level;
** indicates significance at the 1% level.

29

-------
Table 10. Logit Regressions on Pooled Decision-Making Domains
(Joint = 0; Other = 1)



Specification A

Specification B

Specification C

Specification D

Variable



Std



Std



Std



Std



Coefficient

error

Coefficient

error

Coefficient

error

Coefficient

error

Intercept

0.36

0.25

-0.34

0.26

-3.18**

0.99

0.36

1.49

Domain Variables:

















DM: General

-1.66**

0.45

-1.66**

0.45

-1.93**

0.48

9i**

0.50

DM: Home Repairs

0.42

0.35

0.42

0.35

0.40

0.36

0.30

0.39

DM: Decoration

\ 44**

0.37

\ 44**

0.37

1.51**

0.39

1.54**

0.42

DM: Paying Bills

1.16**

0.36

1.16**

0.36

1.22**

0.38

1.15**

0.40

DM: Kids to Doctor

1.23**

0.36

1.23**

0.36

1.30**

0.38

I 4^**

0.41

DM: Kids Clothing

1.28**

0.37

1.28**

0.37

1.36**

0.39

1.30**

0.41

DM: Car

-0.99*

0.39

-0.99*

0.39

-1.08**

0.40

-1.23**

0.44

DM: Major Purchases

-1.09**

0.39

-1.08**

0.39

-1.18**

0.41

-1.12**

0.43

DM: Finance

-0.82*

0.38

-0.82*

0.38

-0.90*

0.39

-0.92*

0.42

Male





-0.04

0.17

-0.05

0.18





Age









-0.20

0.02

-0.04*

0.02

Household Income ($000)









0.00

0.00





HH income dummy (> 80,000)













0.54*

0.23

Education (years)









0.21**

0.05

0.19**

0.05

Number of Children









-0.13

0.13

-0.27*

0.14

Wife Employed













-0.52§

0.28

Wife Employed x wife income













-0.00

0.00

Dyadic scale of marital

















consensus (13 Questions)













-0.04*

0.02

-2*Log Likelihood

-768

-768

-709

-634

Number of Observations

678

678

668

608

Note: § indicates significance at the 10% level; * indicates significance at the 5% level;
** indicates significance at the 1% level.

30

-------
Figure 1.

Income Related Disagreements
Husband's View minus Wife's View



























m - r#l

rv fhTI -

d<-40	-40? d<-20	-20?d<0	0	0
-------
Figure 2.

Dyadic Adjustment Scale Differences
Husband's Score minus Wife's Score

















1

1













. n













II 1

L

i





1

L.rn. . 1

d<-20 -20?d<-15 -15?d<-10 -10?d<-5 -5?d<0	0?d<5 5?d<10 10?d<15 15?d<20 20?d<25 25?d<30 30? d

Difference Amount (Husband - Wife)

~ 32-Question Scale ¦ 13 Question Scale

32

-------
Figure 3.

Differences in the Number of Correctly Answered Lead Questions
Husband's Score minus Wife's Score

-3	-2	-1	0

Difference Amount (Husband - Wife)

33

-------
REFERENCES

Adelman, L., D.A. Zirk, P.E. Lehner, R.I. Moffett, and R. Hall. 1986. "Distributed

Tactical Decisionmaking: Conceptual Framework and Empirical Results." IEEE
Transactions on Systems: Man and Cybernetics 16: 794-805.

Apps, P. and R. Rees. 1996. "Labour Supply, Household Production and Intra-family
Welfare Distribution." Journal of Public Economics 60: 199-219.

Apps, P. and R. Rees. 1997. "Collective Labor Supply and Household Production."
Journal of Political Economy 105: 178-190.

Ashworth, J. and Ulph, D. 1981. "Household Models. In C. Bown (ed.), Taxation and
Labour Supply. London: George Allen and Unwin, 117-133.

Bainbridge, L. 1991. "Mental Models in Cognitive Skill," in A. Rutherford and Y.

Rogers (eds.), Models in the Mind. Academic, New York.

Bateman, I. and A. Munro. 2005. "An Experiment on Risky Choice amongst
Households." Economic Journal 115(502): C176-89.

Becker, G. 1974. "A Theory of Social Interactions." Journal of Political Economy 82:
1063-93.

Bostrom, A., B. Fischhoff, and M. Granger Morgan. 1992. "Characterizing Mental

Models of Hazardous Processes: A Methodology and an Application to Radon."
Journal of Social Issues 48(4): 85-100.

Bourguignon, F. and P. A. Chiappori. 1992. Collective Models of Household Behavior:
An Introduction, European Economic Review, 36, 1-10.

Bourguignon, F. 1999. "The Cost of Children: May the Collective Approach to

Household Behavior Help?" Journal of Population Economics 12: 503-521.

Browning, M., F. Bourguignon, P.-A. Chiappori, and V. Lechene. 1994. "Income and
Outcomes: A Structural Model of Intrahousehold Allocation," Journal of
Political Economy. 102: 1067-96.

Browning, M. and P.-A. Chiappori. 1998. " Efficient Intra-Household Allocations: a
General Characterization and Empirical Tests", Econometrica. 1241-78.

Browning, M. 2000. "The Saving Behavior of a Two-person Household," Scandinavian
Journal of Economics, 102: 235-251.

Busby, D.M., C. Christensen, D.R. Crane, and J.H. Larson. 1995. "A Revision of the
Dyadic Adjustment Scale For Use With Distressed and Nondistressed Couples:

34

-------
Construct Hierarchy and Multidimensional Scales." Journal of Marital and
Family Therapy 21(3): 289-308.

Casas, J. M., and S. Ortiz. 1985. "Exploring the Applicability of the Dyadic Adjustment
Scale for Assessing Level of Marital Adjustment with Mexican-Americans."
Journal of Marriage and the Family 47: 1023-1027.

Chiappori, P.-A. 1997. "Introducing Household Production in Collective Models of
Labor Supply." Journal of Political Economy 105: 191-209.

Craik, K. 1943. The Nature of Explanation. Cambridge University Press: Cambridge,

UK.

Crane, D.R., S.M. Allgood, J.H. Larson, and W. Griffin. 1990. "Assessing Marital
Quality with Distressed and Nondistressed Couples: A Comparison and
Equivalency Table for Three Frequently Used Measures." Journal of Marriage
and the Family 52: 87-93.

Daniels, K., L. de Chernatony, and G. Johnson. 1995. "Validating a Method for Mapping
Managers' Mental Models of Competitive Industry Structures." Human Relations
48: 975-991.

Dosman, D. and W. Adamowicz. Forthcoming. "Combining Stated and Revealed
Preference Data to Construct an Empirical Examination of Intrahousehold
Bargaining." Review of Economics of the Household.

Doss, C. 1996. "Testing among Models of Intrahousehold Resource Allocation," World
Development. 24, 1597-1609.

Eddy J.M., R.E. Heyman, and R.L. Weiss. 1991. "An Empirical Evaluation of the Dyadic
Adjustment Scale: Exploring the Differences Between Marital 'Satisfaction' and
'Adjustment'." Behavioral Assessment 13: 199-220.

Gentner, D. and A. Stevens (eds.). 1983. Mental Models. Lawrence Erlbaum Associates:
Hillsdale, New Jersey.

Hoddinott, J. and L. Haddad. 1995. "Does Female Income Share Influence Household
Expenditure Patterns?," Oxford Bulletin of Economics and Statistics, 57(1): 77-
96.

Jacobson, N.S. 1984. "A Component Analysis of Behavioral Marital Therapy: The
Relative Effectiveness of Behavior Exchange and Communication/problem-
solving Training." Journal of Consulting and Clinical Psychology 52: 295-305.

Langan-Fox, J. and S. Code. 2000. "Team Mental Models: Techniques, Methods and
Approaches." Human Factors 42(2): 242-271.

35

-------
Lesvesque, L.L., J.M. Wilson, and D.R. Wholey. 2001. "Cognitive Divergence and
Shared Mental Models in Software Development Project Teams. Journal of
Organizational Behavior 22: 135-144.

Leuthold, J. 1968. "An Empirical Study of Formula Income Transfers and the Work
Decision of the Poor," Journal of Human Resources, 3: 312-323.

Lundberg, S.J. and R.A. Pollak. 1993. "Separate Spheres Bargaining and the Marriage
Market," Journal of Political Economy. 101 (6): 988-1010.

Lundberg, S.J., R.A. Pollak, and T. Wales. 1997. "Do Husbands and Wives Pool

Resources?: Evidence from the UK Child Benefit," Journal of Human Resources.
32(3): 463-480.

Lynch, C. 1995. "The Use of Repertory Grid Technique to Obtain the Views of Residents
in a Continuing Care Setting." International Journal of Therapeutic and
Supportive Organizations 18: 111-122.

Manser, M. and M. Brown. 1980. "Marriage and Household Decision-Making: A
Bargaining Approach," International Economic Review 21(1): 31-44.

McElroy, M. 1990. "The Empirical Content of Nash-Bargained Household Behavior,"
Journal of Human Resources. 25(4): 559-83.

McElroy, M. and M.J. Horney. 1981. "Nash-bargained Household Decisions: Toward a
Generalization of the Theory of Demand," International Economic Review 22(2):
333-349.

Mohammed, S. and B.C. Dumville. 2001. "Team Mental Models in a Team Knowledge
Framework: Expanding Theory and Measurement across Disciplinary
Boundaries," Journal of Organizational Behavior 22: 89-106.

Morgan, M.G., B. Fischhoff, A. Bostrom, and C.J. Atman. 2001. Risk Communication: A
Mental Models Approach. Cambridge University Press: Cambridge, UK.

Orasanu, J. and E. Salas. 1993. "Team Decisionmaking in Complex Environments," in
G.A. Klein, J. Orasanu, R. Calderwood, and C Zsambok (eds.), Decisionmaking
in Action: Mental Models and Methods. Ablex: Norwood, New Jersey.

OECD (Organisation for Economic Co-operation and Development). 2006. Economic
Valuation of Environmental Health Risks to Children. OECD Publishing: Paris,
France.

Phipps, S. and P. Burton. 1998. "What's Mine is Yours? The influence of Male and
Female Incomes on Patterns of Household Expenditure," Economica 65: 599-613.

36

-------
Rouse, W.B., J. A. Cannon-Bowers, and E. Salas. 1992. "The Role of Mental Models in
Team Performance in Complex Systems." IEEE Transactions on Systems, Man,
and Cybernetics 22: 1296-1308.

Samuelson, P. 1956. "Social Indifference Curves." Quarterly Journal of Economics 70:
1-22.

Shek, D.T.L. 1994. "Psychometric Properties of the Chinese Version of the Dyadic
Adjustment Scale." Psychologia 37: 7-17.

Sorkin, R.D., C.J. Hays, and R. West. 2001. "Signal Detection Analysis of Group
Decisionmaking." Psychological Review 108: 183-203.

Spanier, G.B. 1976. "Measuring Dyadic Adjustment: New Scales for Assessing the
Quality of Marriage and Similar Dyads." Journal of Marriage and the Family
38(1): 15-28.

Spanier, G.B., and L. Thompson. 1982. "A Confirmatory Analysis of the Dyadic
Adjustment Scale." Journal of Marriage and the Family 44(3): 731-738.

Strauss, J., Mwabu, G. and Beegle, K. 2000. "Intrahousehold allocations: A Review of
Theories and Empirical Evidence," Journal of African Economies, 9, AERC
Supplement 1: 83-143.

U.S. EPA. 2000. National Center for Environmental Research Proceedings:

EPA/NIEHS/CDC Centers for Children's Environmental Health and Disease
Prevention Research Progress Review Workshop. November 5-7: University of
California, Berkeley.

U.S. EPA. 2003. National Center for Environmental Economics National Workshop:
Valuing Environmental Health Risk Reductions to Children. October 20-21:
Washington, DC.

Vermeulen, F. 2004. "Collective Household Models: Principles and Main Results."
Journal of Economic Surveys 16(4): 533-564.

37

-------
Value of Reducing Children's Mortality Risk:
Effects of Latency and Disease Type

James K. Hammitt
Center for Risk Analysis, Harvard University
IDEI and LERNA-INRA, Universite de Toulouse
jkh@harvard.edu

Kevin Haninger
Center for Risk Analysis, Harvard University
haninger@fas. harvard. edu

March 2006

-------
Abstract

Despite research showing children may be differentially susceptible to various environmental
health hazards, and that risks to children may be of greater social concern than risks to adults,
there have been relatively few studies that estimate the economic value of reducing risk to
children's health. We propose to design and conduct a contingent valuation (CV) survey to
estimate household willingness to pay (WTP) to reduce mortality risk from pesticides in food,
and to compare WTP to reduce risks to children and risks to adults. We will examine how WTP
depends on latency (the length of the period between exposure and development of symptoms),
noting that childhood exposure may lead to childhood or adult disease and fatality, depending on
latency. We will also evaluate how WTP depends on disease type, comparing terminal cancer
and non-cancer illnesses that present similar symptoms and prognosis.

We will elicit values for risk reductions that vary across the following characteristics: whether
the pesticide exposure is to a child or to an adult, whether the disease is latent or acute, whether
the disease is cancer or not cancer. We will vary the level of detail provided about the disease to
determine whether differences in WTP to reduce risks of cancer and non-cancer disease reflect
differences in information. We will also vary the magnitude of risk reduction, and use sensitivity
of WTP as a diagnostic criterion for validity of the results. Survey respondents will include both
parents and non-parents to allow comparison with prior studies of the value of reducing risks to
adults, and we will measure a variety of demographic variables that may influence WTP. By
comparing estimated WTP between and within respondents, it will be possible to estimate the
relative value of reducing health risks to children versus adults. The survey will be administered
over the World Wide Web, which will facilitate the presentation of visual aids to assist in
communicating the magnitude of risks to survey respondents.

This project is anticipated to provide estimates of the value of reducing food-borne pesticide risk
to children versus adults, as well as analysis of how age, latency, and disease type influence the
valuation. Policymakers can use such estimates to evaluate the benefits of programs aimed at
reducing risks to children.

Keywords: Willingness to pay, health risk, stated-preference, children, cost-benefit analysis

-------
1. Introduction

Despite evidence that children and adults differ significantly in their exposure and vulnerability
to toxic substances, and observations that individuals may systematically place a different value
on child health than they do on adult health (US EPA, 2001), most of the existing valuation
estimates pertain to risks to adults. Moreover, most previous studies have focused on risks of
traumatic fatality, such as workplace or transportation accidents, which differ qualitatively from
the risks of cancer and other disease that are more often associated with environmental
contaminants (Savage, 1993; Revesz, 1999; Sunstein, 1997).

This study is intended to complement previous studies by estimating household willingness to
pay (WTP) to reduce environmental health risks to children, and by examining how the value of
reducing risks to children compares with the value of reducing similar risks to adults. In addition,
the study will investigate the effects of two risk characteristics that are particularly important in
valuing environmental health risks to children: latency (the period between exposure to an
environmental contaminant and development of adverse health effects) and disease type (cancer
versus other degenerative, fatal diseases).

Many environmental risks are characterized by a latency period between exposure to the
environmental contaminant and adverse health effects. The duration of the latency period can
determine whether childhood exposure to a contaminant manifests as disease or death of the
child, or of the adult. In contrast, adult exposure necessarily manifests as disease or death of the
adult. We will investigate the effects on WTP of latency and of whether the exposure and/or
disease manifestation occur to children or adults.

We propose to use contingent valuation (CV) to estimate the effects of age, latency, and disease
type on WTP to reduce mortality risk. In particular, we will elicit parents' WTP to reduce fatal
risks to their children associated with exposure to pesticides in food, and we will compare these
values with parents' and other adults' WTP to reduce similar risks to themselves. In both cases,
the risks presented will vary in latency, whether they cause cancer or another disease, and other
attributes.

In the following section, we describe the theoretical and empirical background for the study. In
Section 3, we describe the survey instrument and sample. In Section 4, we report the results of
regression models relating WTP to the severity and duration of illness, reduction in its
probability, other risk attributes, and to demographic and preference characteristics of the
respondents.

2. Background

In this section, we describe the theoretical and empirical background for this study. First, we
briefly review the literature on the value of reducing health risks to children. Second, we
describe the reasons for selecting health risks of pesticide residues on food as the hazard whose
reduction we will value. Third, we describe the economic theory and prior empirical results
concerning the effects of latency and disease type on risk to adults and the implications for
children's risk. Fourth, we justify the use of household WTP for valuing children's health.

1

-------
2.1.	Prior Work on Valuing Children's Health

There are two strands of prior work that relate to valuing children's health: estimates of altruistic
WTP to protect another individual's health, and estimates of household spending on children's
health (Becker, 1981; Johansson, 1994). Viscusi et al. (1987) used CV to estimate WTP to
prevent the risk of injury associated with household pesticides. They found that WTP to reduce
risks to one's children exceeds WTP to reduce risks to oneself, but could not distinguish between
the effects of parental altruism and injury severity. Viscusi et al. (1988) examined household
WTP to reduce risks of injury associated with household insecticides, for injuries to adults and
children within and outside the household. They found that household values for a statistical case
of child inhalation poisoning were about 75 percent larger than for a statistical case of adult skin
poisoning. Unfortunately, this research does not allow estimation of the relative value of adult
and childhood risks of the same injury.

In the same study, Viscusi et al. (1988) elicited WTP to reduce these risks to people in other
households, both in the same state (North Carolina) and in the United States as a whole. Viscusi
et al. found that altruistic WTP to reduce risks to other households was substantial and was
greater for reducing risks to children than for reducing risks to adults. In particular, the
probability of contributing to a program to reduce risks in the state was 79 percent for a program
that reduced risks to children, and 57 percent for a program that reduced risks to adults. Average
contributions to each program, accounting for the probability of contributing, were $11.53 for
reducing risks to children and $8.75 for reducing risks to adults.

Agee and Crocker (1996) estimated parental WTP to reduce the risk of neurological impairments
from childhood exposure to lead using a revealed-preference approach based on the parents'
decision to obtain chelation therapy for their child. They did not examine WTP to reduce risks of
neurotoxicity to adults, which are much smaller than the risk to children.

A more recent study by Liu et al. (2000) used CV to estimate mothers' WTP to protect
themselves and their children from suffering a cold. WTP was positively associated with the
severity of symptoms and the duration of illness. In addition, mothers' WTP to protect their child
from a cold was nearly twice as large as their private WTP to protect themselves from a cold of
equivalent severity and duration, an indication that mothers value their children's health more
than their own.

2.2.	Pesticide Risks

We propose to study WTP to reduce health risks from residual pesticides on food for a variety of
reasons. First, pesticide contamination of food is a topic of major public concern. Opinion polls
show that pesticides consistently rank as one of the greatest concerns about food safety in the US
(Buzby et al., 1995; Bruhn et al., 1992; Ott et al., 1991). In part as a result of this concern, the
market for "organic" or foods grown without use of synthetic pesticides has grown to
approximately 2% of the US food market (US Department of Agriculture, 1997).

Second, to compare WTP to reduce risks to children and adults, we require a hazard that allows
us to distinguish actions that reduce risks to different members of the household. Exposures to
many environmental health risks are similar to all household members (e.g., air, drinking and
bathing water). Even though some household members are more highly exposed to certain
environmental media (e.g., children may be more exposed to dust and soil than adults), it is

2

-------
difficult to construct plausible scenarios for a CV study that reduce risks to children, or to adults,
but not to both. In this respect, foodborne risks are attractive because it is often the case that
children and adults in a household will consume different foods (at least in part), and so it is
plausible to imagine reducing pesticide concentrations on a food that only the children eat, or a
food that only the adults eat.

2.3. Theoretical Background

The economic approach to valuing mortality risk was developed by Schelling (1968) in an article
suggestively entitled "The Life You Save May Be Your Own." Several years earlier, Dreze
(1962) proposed a similar approach in a French operations research journal, but his work has
received little attention among English-speaking economists. Schelling observed that for
environmental regulations and other life-saving programs, one cannot know whose life will be
"saved." The question is not how to value prevention of a specific death, but how to value small
changes in mortality risk across a population.

The value per statistical life (VSL) is defined as an individual's marginal rate of substitution
between mortality risk and wealth. VSL is not a universal constant but varies by individual and
circumstance. The standard economic model of preferences for wealth and mortality risk (Jones-
Lee, 1974; Weinstein et al., 1980; Dreze, 1962) assumes that an individual's welfare can be
represented as:

EU(p,w) = (\- p)ua(w) + pud(w)	(1)

where p is the individual's chance of dying during the current period and ua (w) and ud (w)
represent his utility as a function of wealth conditional on surviving and not surviving the period,
respectively. The function ua(w) incorporates the individual's preferences for bequests and can
incorporate any financial consequences of dying (such as medical bills or life-insurance
benefits). In this one-period model, wealth and income are treated as equivalent, but the
difference between them can be important in multiple-period models.

The individual's VSL is derived by differentiating Equation (1) holding expected utility constant
to obtain

VST = — = Ua	= Mw)	(2\

dp {\-p)u'a{w) + pu'd{w) Eu'(w)	K)

where prime indicates first derivative.

The numerator in Equation (2) is the difference in utility between surviving and dying in the
current period. The denominator is the expected marginal utility of wealth, i.e., the utility
associated with additional wealth conditional on surviving and dying, weighted by the
probabilities of these events. Assuming that life is preferred to death and that greater wealth is
preferred to less, both numerator and denominator are positive and so VSL is positive. If the
marginal utility of wealth is non-negative, and greater in the event of survival than death (i.e.,
u'a (w) > u'd (w) > 0 ), then VSL increases in mortality risk p. Weak risk aversion with respect to
wealth, conditional on survival and on death (i.e., u"{w) < 0, ud(w) < 0), is a sufficient condition
for VSL to increase with wealth.

3

-------
In the following subsections, we describe what the theory tells us about how VSL depends on
age, latency, and disease type.

Effect of Age on VSL. Theoretical and empirical studies of VSL have generally focused on own
WTP for own risk, treating the individual as the economic agent. Some of these studies have
evaluated the effect of age, but only within adults, and have not considered the valuation of risks
to children. Nevertheless, it may be informative to consider extrapolating results from young
adulthood to childhood.

Theoretical models (e.g., Shepard andZeckhauser, 1984; Rosen, 1988; Ng, 1992) representee
individual's lifetime utility as the expected present value of his utility in each time period. Utility
within a period depends on consumption, which is limited by current income, savings and
inheritance, and ability to borrow against future earnings. The individual seeks to maximize
lifetime utility by allocating his wealth to consumption, savings, and reductions in current-period
mortality risk.

Two factors influence the life-cycle pattern of VSL. First, the number of life years at risk
declines as one ages, so the benefit of a unit decrease in current-period mortality risk declines.
Second, the opportunity cost of spending on risk reduction also declines with age as savings
accumulate and the investment horizon approaches. The net effect may cause VSL to fall or rise
with age.

In models that assume an individual can borrow against future earnings, VSL declines
monotonically with age. For example, Shepard and Zeckhauser (1984) calculate that VSL for a
typical American worker falls by a factor of three from age 25 to age 75. If individuals can save
but not borrow, VSL rises in early years as the individual's savings (and earnings) increase
before it ultimately declines. In this case, Shepard and Zeckhauser find that VSL peaks near age
40 and is less than half as large at ages 20 and 65.

Ng (1992) argues that the rate at which individuals discount their future utility is likely to be
smaller than the rate of return to financial assets, whereas Shepard and Zeckhauser (1984)
assume these rates are the same. If the utility-discount rate is less than the rate of return,
individuals should save more when they are young and consume more when old. Under these
conditions, VSL may not peak until age 60 or so (Ng, 1992). Even if individuals discount future
utility at the rate of return, if they are prudent (Kimball, 1990), younger people might be
anticipated to save more, and spend less on reducing mortality risk, because of the greater range
of future financial contingencies they face.

Although many CV studies include age as one of several covariates in a regression model
explaining WTP for risk reduction, these studies have not typically focused on estimating the
effect of age on VSL. The results of these studies are somewhat contradictory, with several
finding VSL increases with age (Gerking et al., 1988; Johannesson et al., 1997; Lee et al., 1997)
and others finding VSL decreases with age (Buzby et al., 1995; Hammitt and Graham, 1999).
Jones-Lee et al. (1985) included both linear and quadratic age terms in their regression models
and concluded that VSL peaks at about the mean age in their sample (which is not reported).

Several studies have attempted to empirically estimate the effect of age on the benefits of public
life-saving programs, by asking respondents to choose between hypothetical lifesaving programs
that protect people of different ages at different dates. These results do not necessarily reflect
individual WTP to reduce different risks to oneself, since it is implausible to assume that survey

4

-------
respondents compare programs solely in terms of their own private benefits. Cropper et al.
(1994) asked survey respondents about programs to save people of different ages. Their results
suggest that respondents most prefer to protect people in young middle age. Lives of 30 year olds
were valued about 11 times more highly than lives of 60 year olds. For comparison, lives of 20
and 40 year olds are valued as equal to about 8 and 7 60 year olds, respectively. Risks to children
were not evaluated explicitly, but extrapolating the relations found for other ages suggests that
risks to children would be valued as less than risks to young adults. Interestingly, these results
were not sensitive to the age of the respondent.

Two recent empirical studies are specifically directed toward estimating the effect of age on
VSL. Krupnick et al. (2002) conducted a CV study of WTP for a hypothetical intervention that
would reduce the respondent's risk of dying in the next 10 years by either 1 in 1,000 or 5 in
1,000. The sample was restricted to individuals aged 40 years and above. Krupnick et al.
estimate that VSL is roughly constant for ages 40-69, and is about 30 percent smaller for
individuals aged 70 and above. Smith et al. (2001) estimate compensating-wage differential
estimates using data from the Health and Retirement Survey. Their estimates of VSL for
individuals aged 51-65 are not sensitive to age and are comparable to standard estimates for
younger populations.

Accounting for Latency. In Equation (2), VSL is defined in terms of wealth and mortality risk in
a single period. Many environmental risks are characterized by a latency period between the time
an individual is exposed to an agent and the time when he may die from its toxic effect. Since
preventive measures must be undertaken before the exposure occurs, there is often a need to
determine WTP now to reduce the risk of fatality in a future period.

Standard economic theory suggests that the appropriate procedure to account for latency is to
value the risk change using the VSL representing the individual's value when the risk manifests,
and to adjust for the time-value of money and the chance that the individual will die before then
(Cropper and Sussman, 1990; Cropper and Portney, 1990). The adjustment is made by
discounting the future value of the risk reduction back to the time when the expenditure must be
incurred (at the individual's rate of interest). For example, assume that pollution-control
equipment that could be installed today would reduce an individual's risk of dying from cancer
by 1 chance in 100,000, that the cancer would prove fatal 20 years after exposure, that his VSL
in 20 years will be $8 million, and that the individual can earn a 5 percent annual return on
investments. In 20 years, he would be willing to pay $80 to reduce a contemporaneous fatality
risk of 1 in 100,000. The amount he would be willing to pay now is the present value of $80,
about $30 (= $80 x 1.05"20). This amount should be multiplied by the probability that the
individual will survive the intervening 20 years, since the cancer-risk reduction is of no benefit in
the event that he dies of other causes before the environmental pollutant could have killed him.
In many cases, this survival factor is much less important than the discount factor. For the
average American, the probability of surviving 20 years is greater than 0.7 if the individual is
younger than 55 (National Center for Health Statistics, 1998).

The effect of calendar time on VSL has received relatively little attention in the literature, except
to observe that if economic welfare grows over time, VSL would be expected to increase. The
United States Environmental Protection Agency (EPA) has sometimes accounted for the
anticipated growth of income and VSL in regulatory impact assessments, especially when
benefits extend across generations. For example, in evaluating the effects of restrictions on use of

5

-------
CFCs to protect stratospheric ozone, EPA assumed that VSL would grow at annual rates of 0.85-
3.4 percent (U.S. EPA, 1987).

The rate at which VSL increases with income growth (the income elasticity1) is not well
estimated. The primary source of VSL estimates, compensating-wage-differential studies,
usually do not provide information about the income elasticity, because the wage rate is the
dependent variable and so income cannot be used as an explanatory variable.

The income elasticity can be estimated by meta-analysis of compensating-wage-differential
studies where the study populations differ in income, risk, and other factors, but these studies
lack power. Liu et al. (1997) estimated the relationship between VSL, income, and workplace-
fatality risk for a sample of 17 compensating-wage-differential studies in the US and other
industrialized countries. Their point estimate for the income elasticity is 0.54, with a standard
error of 0.85. Mrozek and Taylor (2002) expanded on this approach by including multiple VSL
estimates from each of 33 wage studies and controlling for the average wage, risk, and other
factors. They report four specifications yielding estimated elasticities of VSL with respect to the
wage rate between 0.36 and 0.49 with standard errors of 0.20 and above.

CV studies elicit WTP directly and can be used to estimate the income elasticity of VSL. Typical
estimates range from 0.2 to 0.5. For example, Jones-Lee et al. (1985) estimated values of 0.25 to
0.44, Mitchell and Carson (1986) estimated 0.35, and Corso et al. (2001) estimated 0.41.

Subramanian and Cropper (2000) asked respondents to choose between different public
programs to reduce health risks, and then asked how much more effect (in terms of lives saved)
the less preferred program would need to be to make the respondent indifferent between
programs. In each case, the risks presented the same health endpoint but differed in delay until
benefits would be achieved, voluntariness, controllability, and other factors. Using a multivariate
regression to control for the effects of various factors, Subramanian and Cropper (2000) found
that people discounted for delay. They estimated a marginal rate of substitution of-0.15, which
implies that a 1.5 percent increase in the number of lives saved would compensate for a 10
percent increase in delay.

Hammitt and Liu (2004) use CV to test for the effect of latency on WTP to reduce the risk of a
fatal disease from environmental pollution in Taiwan. The authors find that respondents discount
for the latency period between exposure to environmental contaminants and development of any
resulting disease at a rate of 1.5 percent per year, and that WTP depends on the payment
mechanism, affected organ, and environmental pathway.

WTP to reduce exposure to environmental pollution was not sensitive to the latency period
between exposure and manifestation of disease. The insensitivity of WTP to latency suggests that
respondents anticipate that their VSL will grow over time at a rate about equal to their discount
rate.

In summary, the effects of latency on WTP to reduce own mortality risk are unknown. In theory,
latency increases WTP if individual VSL increases faster than the interest rate, and decreases
WTP otherwise. Empirical studies have not resolved this ambiguity.

1 Carson et al. (2001) note that the income elasticity of demand and income elasticity of WTP are fundamentally
different. The former describes how the quantity demanded increases with income while the latter describes how
WTP for a fixed quantity of a good changes as income increases.

6

-------
Magnitude of Cancer Premium. The value of preventing a fatal cancer is often considered to be
greater than the value of preventing a fatal trauma in a workplace or transportation accident.
Cancer is also frequently viewed as more threatening than other degenerative conditions, such as
heart disease. A striking example is provided by the controversy over whether to encourage
hormone replacement therapy for postmenopausal women. Therapy reduces risk of heart disease
and hip fracture but increases the risk of breast and endometrial cancers. Because heart disease is
five times more likely to kill a woman than is breast cancer, the net effects of treatment are
substantial with gains in life expectancy as large as three years (Col et al., 1997).

There are a number of differences between cancer and accidental fatalities that might affect
relative WTP to reduce each risk, including the often protracted suffering from cancer before
death and the knowledge with cancer that one's condition will deteriorate and lead to death.
Despite the plausibility that there may be a "cancer premium," the empirical literature supporting
this supposition is limited. There are a few studies that provide information about the relative
value of reducing risks of cancer and of acute trauma (e.g., motor vehicle fatality) but no studies
of which we are aware have compared the value of reducing risks of cancer and of other fatal
disease.

Jones-Lee et al. (1985) asked respondents to choose between public programs that would reduce
the number of people dying in the next year by 100 from one of three causes (motor-vehicle
accidents, heart disease, and cancer), and to indicate how much they would voluntarily contribute
to reducing the number of deaths from the cause they selected. A large majority of respondents
(76 percent) chose to reduce cancer deaths, and the mean voluntary contribution was larger for
cancer than for the other causes. Interpreting the mean contributions as estimates of WTP yields
a VSL of £23 million for cancer, £13 million for heart disease, and £7 million for motor vehicle
accidents.

Savage (1993) asked survey respondents to allocate a hypothetical $100 contribution to research
intended to reduce risks of stomach cancer, household fires, commercial-airplane accidents, and
automobile accidents. He found that respondents would allocate the largest amount to stomach
cancer ($47) with much smaller amounts ($15-$21) to the other risks. Although this study
suggests greater WTP to reduce cancer risks, it does not measure individual WTP to reduce own
risk. The value of research on methods to reduce risk of cancer (or the other fatality risks)
depends on the probability that the research will identify interventions to reduce the risk, the
magnitude of the risk reduction produced by the interventions, and the cost of implementing
them. None of these parameters were specified, and so we cannot know what assumptions
respondents made about them. In addition, the pattern of responses seems inconsistent with a
measurement of WTP. The optimal response is to allocate all $100 to whichever risk the
respondent believes will benefit most, since significant diminishing marginal efficacy of
spending is implausible for contributions of $100.

McDaniels et al. (1992) conducted a CV study with only 55 respondents to estimate WTP for
programs to reduce a wide range of health risks. The programs were described as public goods
that would reduce risks to the relevant populations, not only to the respondent. The authors also
elicited risk-perception variables, such as dread. They found that WTP to reduce risk was
positively associated with dread.

Magat et al. (1996) used a risk-risk survey to elicit preferences for reductions in the risk of fatal
automobile accidents and three chronic diseases: terminal lymph cancer, curable lymph cancer,

7

-------
and non-fatal nerve disease. The latency periods for the diseases were not specified in the survey
instrument. The median respondent was indifferent between equal reductions in the probability
of terminal lymph cancer and of fatal automobile accident, suggesting that there is no cancer
premium or that any cancer premium is offset by an assumed difference in latency. The loss in
utility due to curable lymph cancer and non-fatal nerve disease were estimated as 58 percent and
40 percent as great as the loss from a fatal automobile accident, respectively, which suggests that
the utility loss from lymph cancer morbidity is 45 percent larger than the loss from nerve disease.

Hammitt and Liu (2004) also examined whether respondents were willing to pay more to reduce
liver cancer versus liver disease associated with contaminated drinking water, as well as lung
cancer versus lung disease associated with industrial air pollution. The authors estimate that
WTP to reduce the risk of cancer is about one-third larger than WTP to reduce risk of a similar
chronic, degenerative disease.

2.3. Household WTP as a Measure of the Value of Children's Health

There are a variety of reasons why children's own WTP for health and safety initiatives are not
appropriate measures of the value of these goods to children. One obvious issue is that society
does not generally view children as autonomous economic agents. Most children do not earn
income or make economic choices regarding their health and well-being. Children also differ
from adults in their view of death, and may exhibit higher degrees of risk-taking behavior,
perhaps because of their undeveloped cognitive abilities and limited practical experience
(Harbaugh, 1999). Young children often have difficulty imagining and understanding death in
the same way that adults do. They may instead view death as a type of sleep or as an event that
happens only to bad people (Carey, 1985). Another difference from adults is that both children
and adolescents have shorter time horizons, discount the future at higher rates, and often
underestimate the value of future consumption (Krause and Harbaugh, 1998; Harbaugh, 1999).
In short, all of these observed differences present problems for the standard economic
assumptions of informed and rational behavior.

While children's own WTP may be an inappropriate measure of value, household WTP is an
appropriate starting point. Understandably, parents know and care about their children's health,
and they are accustomed to making economic decisions that will affect their children. To some
extent, economists may view parental choices as altruistic behavior, but they may also regard
households as unitary economic agents, with preferences and behaviors that are the result of
some intra-household decision-making process.

Indeed, although most of the literature on the value of statistical life treats the concept as
measuring an individual's rate of substitution between income and mortality risk, in both theory
and practice it seems equally tenable to interpret this literature as measuring household WTP for
changes in mortality risk. In some cases, the change in mortality risk is to a defined individual
(e.g., the worker in studies of compensating wage differentials). In other cases, the risk change
may benefit the entire household (e.g., studies valuing the risk of residential proximity to
hazardous-waste sites, Smith and Desvousges, 1987). In all cases, the opportunity cost of a
mortality risk reduction is smaller household income. Depending on how households allocate
consumption among their members, some or all of them may have lower consumption as a result.

8

-------
3. Survey

We will design and conduct a stated-preference survey to elicit values for reductions in mortality
risks that vary in the baseline probability of illness, reduction in probability, latency of
symptoms, disease type, symptom detail, and whether the exposure occurs to a child or to an
adult. This section describes the survey instrument and sample.

3.1. Survey Instrument

The survey includes a dichotomous-choice experiment in which respondents decide whether to
purchase a safer but more expensive food. The survey instrument is organized as follows. First,
respondents are asked about their knowledge of foodborne pesticide risk and their perception of
how common it is compared with other health and safety risks. Second, respondents assess their
current health using a visual analogue scale (VAS) and the Health Utilities Index Mark 3 (HUI).
The VAS is a numbered line with endpoints of 0 and 100 labeled "equivalent to dead" and
"perfect health," respectively. The HUI is a generic, preference-based, multiattribute health-
status classification system and index that is widely used as a measure of HRQL in clinical
studies, population health surveys, and economic evaluation (Feeny et al., 2002). The HUI
classifies health according to the degree of function on eight dimensions: vision, hearing, speech,
ambulation, dexterity, emotion, cognition, and pain. For each dimension, there are five or six
levels of functional impairment that range from complete function to severe impairment.

Third, respondents complete a tutorial designed to help them practice making tradeoffs between
the price and safety of food. The tutorial also familiarizes respondents with a visual aid that
communicates the probability of risks (Corso et al., 2001). The visual aid contains red and white
areas that represent 10,000 apples, where the fraction of the area that is colored red equals the
probability that an apple contains unsafe levels of pesticide.

Fourth, respondents are asked to consider buying food for a meal that only they will eat.
Respondents are asked whether they eat a type of food randomly selected from the set {apples,
grapes, lettuce}. If they do not eat the selected food, respondents are asked about another
randomly-selected food. After answering questions about how often they eat the food and how
much they typically eat, respondents are presented with a description of the symptoms of a fatal
disease caused by consuming pesticide in the food. Respondents are then told their baseline
probability of illness (either 2 in 100,000 or 4 in 100,000 per year) and informed that they could
reduce their risk to 1 in 100,000 per meal by purchasing a safer but more expensive brand of
food. The baseline probability of illness and reduction in probability are communicated using the
visual aid described above. The risk reduction is described as produced by a stringent pesticide
safety program established and monitored by the United States Government. Respondents are
told that while the food produced by the pesticide safety program is safer to humans than
conventional food, the program is not an organic farming practice, nor does it affect other
animals or the environment any differently than conventional farming. WTP to reduce the
probability of illness is elicited using double-bounded, dichotomous-choice questions. Each
respondent is asked if he would purchase the safer food if the extra cost per year were a
randomly selected amount from the set {$10, $20, $50, $80, and $100}. There is one follow-up
question, in which the bid is equal to twice the initial bid if the respondent is willing to pay the
initial amount, and equal to half the initial bid otherwise. Finally, respondents answer follow-up

9

-------
questions about their food-handling practices, acceptance of the hypothetical scenario, and
relevant personal characteristics.

Each respondent is asked to value three health-risk reductions that vary in baseline probability of
illness, reduction in probability, severity and duration of symptoms, conditional probability of
mortality, and type of food affected. Using a full factorial design, the risk attributes are randomly
assigned so that each of the possible combinations is asked of some respondents. Table 1 shows
the risk attributes, which we describe in more detail below.

Table 1. Risk Attributes (Full-Factorial Design)

Individual
Exposed

Annual Risk
Reduction

Latency

Disease Type

Symptom
Detail

Type of Food

Self
Child

Other Adult

1 in 100,000
3 in 100,000

1 year
10 years
20 years

Cancer
Non-Cancer

Brief
Detailed

Apples
Grapes
Lettuce

Person Exposed. Depending on their household composition, respondents are asked about
reducing risks to their own health, the health of a child, or the health of another adult.
Respondents who live in a household with at least one child under the age of 18 and at least one
other adult are asked about reducing one risk to their own health, one risk to the health of a
randomly-selected child from their household, and one risk to the health of a randomly-selected
adult from their household (in random order). Respondents who live in a household with at least
one child under the age of 18 and no other adults are asked about reducing one risk to their own
health and two risks to the health of a randomly-selected child from their household (in random
order). Respondents who live in a household with at least one other adult and no children under
the age of 18 are asked about reducing two risks to their own health and one risk to the health of
a randomly-selected adult from their household (in random order). Respondents who live alone
are asked about reducing three risks to their own health, but are not presented with the same food
twice.

Latency. The risks presented will differ in latency, defined as the period between the time when
an individual is exposed to an environmental contaminant and the time when he or she develops
symptoms of disease or is diagnosed. Three latency periods (1 year, 10 years, and 20 years) will
be considered. In the short latency case, respondents will be told that, if they develop the stated
disease, symptoms will begin within a year and they will live only about two years longer. In the
long latency cases, respondents will be told they will not know if they were sufficiently exposed
to develop the disease until they experience symptoms about 10 years (or 20 years) in the future.
After developing symptoms, the prognosis is identical to the short latency case.

Disease Type. WTP will be elicited for one or more disease pairs that consists of a specific form
of cancer and a non-cancer disease that affects the same organ and has similar symptoms and
prognosis. All diseases will be terminal. The symptom descriptions presented to respondents will
be identical except for the name of the disease.

Symptom Detail. The symptom descriptions will be varied to provide different levels of detail.
Our hypothesis is that the cancer premium may be sensitive to the comprehensiveness of the

10

-------
symptom description. When respondents are given little or no information about the symptoms
and prognosis of a disease other than its name, they may have a higher WTP to reduce the risk of
"cancer," if cancer is generally perceived to lead to more severe morbidity than other fatal
diseases. In this case, we might observe a substantial cancer premium. Alternatively, when the
respondent is given extensive information about the symptoms associated with a disease, the
additional information associated with knowing that the disease is a form of cancer rather than
another fatal disease may have less impact, and so the magnitude of the cancer premium may be
much smaller or non-existent.

Magnitude of Risk Reduction. The magnitude of the risk reduction will be varied across
valuation tasks to provide information about whether the CV instrument produces WTP
estimates that are sensitive to scope. Under conventional economic theory, WTP for a small
reduction in mortality risk is nearly linear in the magnitude of the risk reduction. The sensitivity
of estimated WTP to magnitude of risk reduction can be used as a diagnostic test of the
performance of the survey instrument (Hammitt and Graham, 1999; Hammitt, 2000; Corso et al.,
2001). If WTP is not proportional to the magnitude of risk reduction, then estimated VSL is
sensitive to the arbitrary magnitude of the risk reduction offered.

Inadequate sensitivity of estimated WTP to magnitude of risk reduction has been a substantial
problem in almost all CV studies of health risks. Hammitt and Graham (1999) identified 14 CV
studies published from 1980 through 1998 that either reported a test of sensitivity to magnitude
or provided enough information to enable them to conduct such a test. They found that although
estimated WTP was sensitive to the magnitude of risk reduction (i.e., the estimated value of a
larger reduction exceeded the estimated value of a smaller reduction) in 11 cases, WTP was
inadequately sensitive (i.e., less than proportionate to magnitude of risk reduction) in all cases.

To test whether inadequate sensitivity to magnitude is a result of difficulties in communicating
small risk changes to survey respondents, Corso et al. (2001) asked respondents to value
reductions in automobile fatality risk. Corso et al. presented respondents with one of three visual
aids (a field of 25,000 dots, a logarithmic risk ladder, or a hierarchical linear risk ladder) or no
visual aid, and then elicited values for reducing annual risk by 5 or 10 in 100,000 from separate
sub-samples. Corso et al. found that estimated WTP was sensitive to risk reduction for
respondents presented with any of the visual aids, but not for the control group. Moreover, the
hypothesis that estimated WTP was proportionate to the risk reduction could not be rejected for
the groups of respondents presented with either the dots or the logarithmic risk ladder. The study
by Corso et al. suggests that CV can be used to estimate WTP for small risk reductions that are
consistent with economic theory, and hence that near-proportionality of estimated WTP to risk
reduction may be used as a test for the validity of CV estimates (Hammitt, 2000).

For the valuation tasks, we anticipate using two magnitudes of risk reduction: 1 in 100,000 per
year and 3 in 100,000 per year. These risk reductions are small enough to be relevant to the
pesticide risks of concern, yet are sufficiently far apart that WTP should differ substantially (by a
factor of three). The risk reductions will be accompanied by visual aids that were found to work
well by Corso et al. (2001). In addition, describing risks using a common denominator is
anticipated to assist respondents in recognizing differences between the two risk magnitudes.

WTP will be elicited using double-bounded discrete-choice questions (Hanemann et al., 1991).
Each respondent will be randomly assigned to one of five initial bid values ($10, $20, $50, $80,
and $100) that represent the additional cost of meals made with food containing reduced

11

-------
pesticide levels. There will be one follow-up question, where the bid is equal to twice the initial
bid if the respondent indicates he would be willing to pay the initial amount, or equal to half the
initial bid otherwise. Respondents will receive different initial bids for the first and second
valuation questions to minimize follow-up effects (e.g., giving the same "yes" or "no" response
to the second valuation question as given to the first).

Discrete-choice questions are often preferred to open-ended questions because they appear to be
easier for respondents to answer. The referendum format is incentive-compatible and was
recommended by the NOAA panel (Arrow et al., 1993). In addition, dichotomous-choice
questions are often considered superior to open-ended, bidding-game, and payment-card formats,
because they do not create anchoring effects. The double-bounded format provides substantially
greater information per respondent than a single-bounded format. The corresponding double-
bounded or interval-data models of WTP have been shown to produce more efficient estimates
than those obtained using only the single-bounded payment format (Hanemann et al., 1991;
Alberini, 1995). Although the initial bid may influence responses to the follow-up question
(Alberini et al., 1997), we will calculate single-bounded estimates using only the response to the
first valuation question to investigate the magnitude of any follow-up effect.

3.2. Sample

The survey will be fielded to members of a demographically representative panel maintained by
Knowledge Networks. Households are recruited to the panel using random digital dialing and
provided free Internet access and hardware, such as MSN® TV, as a participation incentive. In
total, 2,000 interviews will be completed. We plan to over-sample households with children so
that we have sufficient responses about reducing risks to children's health.

4. Analysis

Using theory to inform model specification, we will develop an empirically estimable model
relating WTP to health risk attributes, the respondents' socioeconomic characteristics, and
variables characterizing risk attitudes. For the purposes of illustration, consider the following
model:

log (WTP) = a + + y,R, +s	(3)

where X, is a vector of covariates describing the respondent (e.g., age, sex, health, education,
marital status, household income) and the person at risk (e.g., age, sex, health), R, is a vector of
risk characteristics (e.g., latency, disease type, magnitude of risk reduction), and e is an error
term.

Because WTP is elicited using double-bounded binary choice questions, individual WTP is
interval censored. We observe only the upper and lower bounds on an individual's WTP (which
may be infinite and zero, respectively). Equation (3) will be estimated using maximum
likelihood methods (Alberini, 1995) implemented in standard statistical software (e.g., SAS).
Estimates will be obtained using alternative parametric assumptions regarding the distribution of
the error term, including a "mixed model" which allows for the possibility that a finite fraction of
respondents have WTP equal to zero (Werner, 1999).

12

-------
In order to test for differences in WTP to reduce risks to children versus adults, we will include a
dummy variable indicating whether the individual who benefits from the reduced pesticides is a
child or adult. The effects of age will also be evaluated using dummy variables; e.g., the "child"
dummy may be replaced by a series of dummy variable for age category (e.g., 0-5, 6-12, and 13-
18 years). Adult age will be represented using dummy variables for age categories and,
alternatively, using simple polynomial functions (e.g., age, age2).

To determine how WTP depends on other characteristics of the health risks, we will estimate a
regression that includes dummy variables for various risk characteristics, such as the degree of
latency, whether the risk causes cancer or not, and the level of detail of the symptom description
provided. We will also interact the dummy variable for long-latency with the child age dummy
variables, to determine whether the valuation of latent risks (where exposure occurs to a child but
the risk only manifests to the adult), is sensitive to the age of the child at time of exposure.

We will incorporate several methods to test the validity of estimated WTP. First, we will
estimate the coefficient on risk magnitude to determine how WTP depends on the magnitude of
risk reduction. Under standard economic theory, WTP should be almost exactly proportional to
the magnitude of risk reduction for small risk reductions (where income effects are negligible)
(Corso et al., 2001; Hammitt, 2000). Hence, if we can reject the hypothesis that WTP for the 3 in
100,000 risk reduction is not three-times WTP for the 1 in 100,000 risk reduction, this will
provide evidence suggesting that respondents did not accurately report their WTP for risk
reduction. Given the difficulties in communicating and comprehending small risk changes, this
proportionality test is quite demanding and has only once been satisfied, to our knowledge
(Corso et al., 2001). A weaker test is to require that estimated WTP be statistically significantly
larger for the larger risk reduction. Even this test is frequently not satisfied by prior studies,
perhaps because of inadequate attention to communicating the magnitude of risk changes
(Hammitt and Graham, 1999).

Additional evidence regarding the validity of estimated WTP will come from use of follow-up
questions and examining the relationship between individual WTP and covariates that are
anticipated to be associated with it. Follow-up questions will include some addressed to accuracy
of risk perception (e.g., asking respondents if they believe they are more likely to get sick or
injured, or to die, from, e.g., pesticides on food, microbial contaminants on food, heart disease,
or other causes). Previous studies have found some ability to accurately rank these risks
(Williams and Hammitt, 2001), and we anticipate that respondents with a better sense of the
relative probabilities of these events would give more valid answers about WTP. Other questions
will address respondents' health habits both for themselves (e.g., dietary choices, smoking,
drinking, exercise, preventive care, seatbeltuse) and for their children (e.g., dietary choices,
preventive care, seatbelt and child seat use, bicycle helmet use, childproofing home by storing
hazardous materials carefully and covering electrical sockets). We anticipate that people who
adopt healthier habits may also have greater WTP for reductions in pesticide-related risk. There
is some collaborating evidence that those with poorer health habits (smokers and those who do
not use automobile seatbelts) have smaller WTP to reduce risk of workplace injury (Hersch and
Viscusi, 1990; Hersch and Pickton, 1995; Viscusi and Hersch, 2001).

13

-------
References

Agee, M.D., and T.D. Crocker, "Parental Altruism and Child Lead Exposure: Inferences from
the Demand for Chelation Therapy," Journal of Human Resources 31: 677-91, 1996.

Alberini, A., "Efficiency vs Bias of Willingness-to-Pay Estimates: Bivariate and Interval-Data
Models," Journal ofEnvironmental Economics and Management 29: 169-180, 1995.

Alberini, A., B. Kanninen, and R.T. Carson, "Modeling Response Incentive Effects in

Dichotomous Choice Contingent Valuation Data," Land Economics 73: 309-324, 1997.

Arrow, K., R. Solow, P. Portney, E. Learner, R. Radner, and H. Schuman, "Report of NOAA
Panel on Contingent Valuation," Federal Register 58(10): 4601-4614, 1993.

Becker, G., "Altruism in the Family and Selfishness in the Market," Economica 48: 1-15, 1981.

Bruhn, C.M., K. Diaz-Knauf, N. Feldman, J. Harwood, G. Ho, E. Ivans, L. Kubin, C. Lamp, M.
Marshall, S. Osaki, G. Stanford, Y. Steinbring, I. Valdez, E. Williams, and E.

Wunderlich, "Consumer Food Safety Concerns and Interest in Pesticide-Related
Information," Journal of Food Safety 12: 253-262, 1992.

Buzby, J., R. Ready, and J. Skees, "Contingent Valuation in Food Policy Analysis: A Case Study
of a Pesticide-Residue Risk Reduction," Journal of Agricultural and Applied Economics
27: 613-625, 1995.

Carey, S., Conceptual Change in Childhood, MIT Press, Cambridge MA, 1985.

Carson, R.T., N.E. Flores and N.F. Meade, "Contingent Valuation: Controversies and Evidence,"
Environmental and Resource Economics 19: 173-210, 2001.

Col, N.F., M.E. Eckman, R.H. Karas, S.G. Pauker, R.J. Goldberg, E.M. Ross, R.K. Orr, and J.B.
Wong, "Patient-Specific Decisions About Hormone Replacement Therapy in
Postmenopausal Women," Journal of the American Medical Association 277: 1140-1147,
1997.

Corso, P.S., J.K. Hammitt, and J.D. Graham, "Valuing Mortality-Risk Reduction: Using Visual
Aids to Improve the Validity of Contingent Valuation," Journal of Risk and Uncertainty
23: 165-184, 2001.

Cropper, M.L, and P.R. Portney, "Discounting and the Evaluation of Lifesaving Programs,"
Journal of Risk and Uncertainty 3: 369-379, 1990.

Cropper, M.L., and F.G. Sussman, "Valuing Future Risks to Life," Journal of Environmental
Economics and Management 19: 160-174, 1990.

Cropper, M.L., S.K. Ayded, and P.R. Portney, "Preferences for Life Saving Programs: How the
Public Discounts Time and Age," Journal of Risk and Uncertainty 8: 243-265, 1994.

Dreze, J., "L'Utilite Sociale d'une Vie Humaine," Revue Frangaise de Recherche Operationelle
6: 93-118, 1962.

Fu, T.-T., J.-T. Liu and J.K. Hammitt, "Consumer Willingness to Pay for Low-Pesticide Fresh
Produce in Taiwan," Journal of Agricultural Economics 50: 220-233, 1999.

Gerking, S., M. De Haan, and W. Schulze, "The Marginal Value of Job Safety: A Contingent
Valuation Study," Journal of Risk and Uncertainty 1: 185-199, 1988.

14

-------
Gray, G.M., and J.K. Hammitt, "Risk/Risk Tradeoffs in Pesticide Regulation: An Analysis of the
Public Health Effects of a Ban on Organophosphate and Carbamate Pesticides," Risk
Analysis 20: 665-680, 2000.

Hammitt, J.K. and J.-T. Liu, "Effects of Disease Type and Latency on the Value of Mortality
Risk," Harvard School of Public Health, 2002.

Hammitt, J.K., "Consumer Willingness to Pay to Avoid Pesticide Residues," Statistica Sinicah:
351-366, 1993.

Hammitt, J.K., "Risk Perceptions and Food Choice: An Exploratory Analysis of Organic- Versus
Conventional-Produce Buyers," Risk Analysis 10: 367-374, 1990.

Hammitt, J.K., "Using Willingness to Pay to Value Food Safety: Theory and Empirical

Methods," in Setting Food Safety Priorities: Toward a Risk-Based System, Resources for
the Future, Washington, DC, in press, 2002.

Hammitt, J.K., and J.D. Graham, "Willingness to Pay for Health Protection: Inadequate
Sensitivity to Probability?" Journal of Risk and Uncertainty 18: 33-62, 1999.

Hammitt, J.K./'Evaluating Contingent Valuation of Environmental Health Risks: The

Proportionality Test," Association of Environmental and Resource Economists Newsletter
20(1): 14-19, May 2000. Reprinted in Stated Preference: What Do We Know? Where Do
We Go? (Proceedings). Report number EE-0436, U.S. Environmental Protection Agency,
October 2000.

Hanemann, W.M., J. Loomis, and B. Kanninen, "Statistical Efficiency of Double-Bounded
Dichotomous Choice Contingent Valuation," American Journal of Agricultural
Economics 73: 1255-1261, 1991.

Harbaugh, W.T, "Valuing Children's Health and Life: What Does Economic Theory Say about
Including Parental and Societal Willingness to Pay?" EPA workshop on Valuing Health
for Environmental Policy with Special Emphasis on Children's Health Protection, March
1999.

Hersch, J., and T.S. Pickton, "Risk-taking Activities and Heterogeneity of Job-Risk Tradeoffs,"
Journal of Risk and Uncertainty 11: 205-217, 1995.

Hersch, J., and W.K. Viscusi, "Cigarette Smokers, Seatbelt Use, and Differences in Wage-Risk
Tradeoffs," Journal of Human Resources 25: 202-227, 1990.

Johannesson, M., P-0 Johansson, and K-G Lofgren., "On the Value of Changes in Life

Expectancy: Blips Versus Parametric Changes," Journal of Risk and Uncertainty 15:
221-239, 1997.

Johansson, P-O., "Altruism and the Value of Statistical Life: Empirical Implications," Journal of
Health Economics 13: 111-118, 1994.

Jones-Lee, M., "The Value of Changes in the Probability of Death or Injury," Journal of
Political Economy 82: 835-849, 1974.

Jones-Lee, M.W., M. Hammerton, and P.R. Philips, "The Value of Safety: Results of a National
Sample Survey," The Economic Journal 95: 49-72, 1985.

15

-------
Comments on "Use of Contingent Valuation to Elicit Willingness-to-Pay for Benefits of

Developmental Health Risk Reductions"
by Katherine von Stackelberg and James K. Hammitt

•	Why not ask questions about the household, or if the respondent is a parent? This would
impact how well the respondent could identify with questions about a hypothetical child.

•	It is an important result that respondents were willing to increase their bids from their
initial ecological bid when asked for a total bid (ecological and health), but not when the
health bid was asked for first (especially since 63-74% indicated that they could separate
the two endpoints).

•	The standard gamble and time tradeoff questions seem like they would be difficult for
respondents to truly understand and answer. Could a parent of a real 10 year old child
really answer a question that trades off a small probability of death (or weeks of
longevity - this one might be easier) to a reduced cognitive deficit that is relatively mild?

Those types of questions may possibly be easier for a non-parent to answer, however
a non-parent, or maybe even to some extent a parent of only a baby, may not fully
understand the implications of the trade-off
- Because the QALY questions turn out to be significant in most of the models, I think
the responses could be viewed as representing respondents' perceptions about how a
cognitive deficit would affect a child's quality of life.

•	Overall, I found the paper interesting and could be a useful approach in getting values for
mild developmental effects.

Comments on "Parental Decision-Making and Children's Health," by Ann Bostrom,
Sandra Hoffmann, Alan Krupnick and Wictor Adamowicz with Robin Goldman and
Michael McWilliams

•	Well-written and very fun to read - certainly made me reflect on decision making in my
own household.

•	Results highlight that there is a lot of disagreement in marriages/households about factual
information as well as about how household decisions are made.

- Even factual information provided by couples separately contained differences.
Couples not knowing exact percentages of contributions to household income,
spouse's income, or total household income didn't surprise me.

It makes sense that if spouses specialize in decision domains such as paying bills or
managing finances that the "specialist" would know more (e.g. I pay the bills in our
house and my husband doesn't know exactly how much I make)

Spouses may have different concepts of income (e.g., I would answer an annual
household income question assuming just my husband's salary however when he
answers, he includes bonuses and extra fees). Respondents having jobs in sales

1

-------
where a significant part of salary is based on commission, could introduce answers
that vary between spouses.

•	The survey collected a lot of information about decision-making behavior with a section
specifically geared towards marital adjustment, but I'm wondering if you could ask
questions to reveal the personality traits of each spouse as well. Personality could
influence decision-making in household.

•	The results imply that it does matter who in a household is interviewed for a survey. It
would be nice to see more discussion on how the couples separately and together dealt
with the hypothetical lead paint decision scenario and did it correspond to the results
from the rest of the survey. Does a respondent consider other household members'
preferences when answering individually?

•	I'm excited to see results of the future WTP survey and answers to the questions: How
does separate WTP for each spouse compare to each other and to a jointly arrived at
WTP? What are some questions that could be asked of individuals to determine how
representative of household preferences their own answers are?

Comments on "Value of Reducing Children's Mortality Risk: Effects of Latency and
Disease Type," by James K. Hammitt and Kevin Haninger

Paper did not yet include results so I only have a few comments.

Nice survey of the literature on several different dimensions of WTP for mortality risk (exposure
to child or adult; exposure to self or other household member; the fatality from disease is
immediate or latent; the fatal disease is a cancer or non-cancer; the amount of information
provided about the fatal disease).

How much information are respondents given about the pesticide safety program? Are they told
specifics about how it works? For example, if they are told that there is a special wash applied to
produce after it is harvested, there is clearly no ecological benefit. But if the program is less or
different pesticide use, then respondents may still confer an ecological benefit to the program
even if you state there isn't any.

Are respondents asked about organic food purchases?

I'm not sure how able respondents will be at comprehending a risk reduction to only one member
of the household - most food brought into a household is consumed by everyone in the
household (with some exceptions). Could you also ask a question about reducing the risk to the
entire household?

2

-------
Childlike Values:

Measurement Strategies for Children's Health Values

F. Reed Johnson

Senior Fellow
Research Triangle Institute

Discussant Remarks
U.S. EPA NCER/NCEE Workshop
April 2006

1

-------
The goal of the STAR program is to support research that translates existing methods
and findings into policy-relevant research and to fill in gaps in knowledge that limits our
ability to assess the efficiency of environmental regulations. After three decades of
environmental and health valuation research, we have acquired some respect for the
difficulties inherent in nonmarket valuation. These difficulties are magnified when we
attempt to estimate willingness to pay to reduce risks to health and safety.

From an individual's point of view, most environmental regulations reduce relatively
small risk exposures by relatively small amounts. We thus encounter various
impediments to obtaining valid and reliable values for such risk reductions, including
among other challenges, respondent innumeracy, sensitivity to risk framing, sensitivity to
features of the risk that are independent of probability or health endpoint, poor
descriptive power of the standard expected utility model. As evidence accumulates
regarding the differential sensitivity of children to environmental hazards, demand has
increased for valid and reliable estimates of the value of reducing such risks. The
papers presented in this workshop evaluate the extent to which people are willing to
accept tradeoffs between money and children's health risks and what methods are likely
to give us valid, policy-re levant estimates.

The three papers in this session offer different strategies for answering such questions.
Von Stackelberg and Hammitt compare classic contingent valuation, standard gamble,
and time tradeoff elicitation formats. They obtain estimates of $466 per IQ point for
developmental impairment, or $109,000 per QALY. Hammitt and Haninger offer a
research prospectus to evaluate risks from pesticide contamination of food using classic
contingent valuation, visual analog scale, and health utilities obtained from the Health
Utilities Index Mark 3 health-related quality of life instrument. They propose to evaluate
the effect of outcome latency, disease type, and information treatment on values
measured in each way. Finally, Bostrom, Hoffman, Krupnick, and Adamowicz offer
some preliminary results from a survey of household decision patterns. They find that
about 32% of surveyed couples' preferences were disjoint for major purposes and for
financial decisions generally. They also find that most couples were willing to consider
cost-efficacy tradeoffs for lead exposure.

2

-------
Von Stackelberg and Hammitt, "Use of Contingent Valuation to Elicit Willingness
to Pay for the Benefits of Developmental Health Risk Reductions"

The authors set out to determine whether WTP is proportional to risk reduction and to
obtain WTP per QALY. The standard-gamble (SG) elicitation format is relatively
unfamiliar to environmental economists. This method obtains a von Neumann-
Morgenstern utility index scaled between death, assumed to have utility equal to zero,
and perfect health, assumed to have utility equal to one. The utility index is the
probability for a lottery between perfect health and instantaneous, painless death that
makes respondents indifferent between the lottery and a sure outcome—in this case a
specified developmental disability. The elicitation generally is assumed to be
independent of the usual factors we generally use to condition utility such as income,
demographic factors such as age and gender, duration of the certain condition,
treatment options, and other context factors. Moreover this approach requires assuming
preferences conform to the expected-utility model that generally performs poorly in
describing actual behavior under risk. While SG is popular among (mostly non-
economist) health researchers, it is hard to justify suspending so many considerations
that guide preference research in virtually every area of applied economics other than
health.

The authors follow the environmental economics convention of using a double-bounded
format for both the standard gamble and CV questions. The convention in health
economics is to use a bidding game for standard gamble elicitations. It is likely that the
two methods would yield different utility weights. The authors acknowledge known
problems with double-bounded CV formats. It isn't clear later whether they found no
significant anchoring bias and used the double bounded estimator or appealed to
Alberini's finding and pooled the first and second bids. The strategy for the second-bid
starting point conditions on the first-bid starting point. While logical, it also imposes
some degree of monotonicity and consistency in responses that might not have resulted
from randomization.

Economical administration of stated-preference surveys conflicts with OMB requirements
that for high response rates and validated claims of representativeness. OMB appears
uncompromisingly opposed to using web panels to collect data in support of regulatory
decisions. Nevertheless, the authors assert that the Knowledge Networks panel "is the

3

-------
only available method for conducting internet-based survey research with a nationally
representative probability sample." It is worth noting that (1) random-digit dialing no
longer ensures reaching a representative sample; (2) there is selection bias in the
sample that agrees to join the KN web panel once contacted; (3) there is selection bias
in attrition from the panel. That doesn't mean we shouldn't use web panels, however.
Both Knowledge Networks and other web panels use sophisticated weighting techniques
to correct for possible selection bias. It is difficult to imagine any other alternative that is
consistent with the actual resources available to conduct stated-preference studies.

The assertion that the estimated WTP is approximately proportional to risk reduction
appears to rely on a weak test. In fact, there are competing hypotheses to support an
expectation that WTP is nonlinear in probability. One possibility is that risk preferences
follow rank-dependent utility axioms rather than expected-utility axioms. Rank-
dependent utility overweights small probabilities and underweights large probabilities.
Figure 1 indicates the possible effect of such weighting. Expected utility dictates that
WTP at risk level 1 be at point A. However, if probabilities between 0 and level 1 are
weighted more heavily than probabilities between levels 1 and 4, then WTP at risk level

Figure 1. Nonlinear Effect of Risk on WTP

1 will be at some point B. Alternatively, if the risk levels 1-4 are very small probabilities,
respondents may find it difficult to discriminate between absolute differences. They may

4

-------
simply recode 0 as "low", 2 as "medium", and 4 as "high" and set the utility differences to
be equal. That would yield WTP at the medium level at point C. If plotted against
nominal risk, C would look like B. If plotted against equally spaced categories, C would
lie on a straight line. If preferences follow either B or C, estimating WTP as a linear
function, as shown by the solid line, might not detect the kink and fail to reject a
hypothesis of linearity. A better practice is to estimate the model using categorical risk
levels and test whether utility differences are proportional to nominal risk values or not.

Cost per QALY is widely computed in health economics to evaluate the relative
efficiency of alternative interventions. However, knowing that the cost per QALY for one
policy is less than that for another policy does not provide any guidance about whether
either policy is worth adopting. I am troubled by using WTP/QALY to solve the lack of a
cost-effectiveness threshold. Lack of a threshold is the result of resistance to monetizing
benefits to facilitate a real cost-benefit analysis in health economics, much as
environmentalists have resisted monetizing environmental benefits for environmental
policy analysis. Practitioners argue QALYs avoid all the equity baggage of WTP. If
QALYs are all we need, why try to find a WTP value to do the analysis in QALY terms?
Doing so combines incompatible conceptual models (Johnson, 2005).

The authors perpetuate a common confusion in comparing their WTP per QALY
estimates with calculations reported in the literature based on the value of a statistical
life (Hirth, 2003). Apart from the well-known problems in obtaining valid VSL estimates,
it is inappropriate to divide VSL by life expectancy and interpret that as WTP per QALY.
A statistical life year is not the same as a year of life, much less the same as a year of
life in perfect health. That is exactly the misinterpretation that scandalizes non-
economists when they hear us argue about the dollar value of a (statistical) year of life.

While the analysis in this paper is carefully done, there are several puzzling results that
might warrant additional thought. For example, the significant negative sign on the
reading-comprehension health endpoint is counter-intuitive and would benefit from some
explanation. The statement that WTP was 33% lower for IQ compared to reading
comprehension seems inconsistent with the wrong sign on reading comprehension.
Furthermore, the significance of the IQ endpoint parameters is weaker than expected
and values per unit IQ loss are an order of magnitude lower than the expected income

5

-------
loss. The authors speculate that respondents are discounting the effect on expected
lifetime income inappropriately, but there may be other explanations.

Hammitt and Haninger, "Value of Reducing Children's Mortality Risk: Effects of
Latency and Disease Type"

Jim Hammitt conducted a well-conceived study for EPA in 1986 entitled "Organic
Carrots: Consumer Willingness to Pay to Reduce Food-Borne Risks." I was interested in
seeing how this plan to conduct a study on a similar topic reflected how much his and
our understanding of risk-preference elicitation methods has evolved over the
intervening 20 years. I think he would agree that we have not progressed as far as we
would have liked.

The authors propose a repeated-CV design, along with visual analog scale and HUI-
Mark 3 QALY weights to obtain QALY estimates. They propose to evaluate the
insensitivity to latency noted in previous studies, although they appear to be unaware of
the latency results reported in papers by Cameron and DeShazo. The proposed risk
reduction from 2 or 4/100,000 to 1/100,000 may invite respondents to recode such small
numbers into low, medium, and high categories. It might be prudent to include a scope
test to see whether respondents are paying attention to absolute risk levels.

Asking only 3 repeated CV questions doesn't impose much of a cognitive burden on
respondents. It is likely they could answer 10 or 12 questions, which would greatly
increase the power of the sample. With careful attention to the experimental design, the
data might provide enough information to estimate hierarchical Bayes individual-level
estimates of WTP.

Bostrom, Hoffmann, Krupnick, and Adamowicz, "Parental Decision Making and
Children's Health"

This study is an interesting first start at understanding how to interpret household
preferences based on responses from one member of the household. This work is long
overdue. The standard practice in stated-preference research is to administer the
survey to one household member. The preference-elicitation question may or may not
explicitly ask the respondent to indicate household preferences. In any case, in the
absence of data or theory to help discriminate among household members, we simply

6

-------
assume the observation represents an aggregation of household values. However, if
spouses in the household specialize in decision making, then asking different spouses a
WTP question is likely to lead to different answers.

In the next draft of the paper, it would help to be more explicit about what insights were
obtained from the data about the basic research problem and how the results will be
used to develop a better stated-preference instrument. For example, how might one
adapt the standard time-to-think experiment? One possible explanation for differences
between an immediate and a "considered" response is that the respondent takes the
extra time to consult other decision makers in the household and construct a value that
is a better aggregation of household preferences. Could the decision questions in this
survey be adapted to measure what preference-aggregation process was used during
the time to think?

There are several published studies on income-pooling experiments. (See, for example,
Bateman and Munro, 2005.) Such experiments rely on actual decisions on lotteries with
payoff rules designed to reveal how income is controlled within the household. It may be
possible to extend these methods to explore how responsibility for expenditures in
particular categories is allocated within a household.

The authors attribute the allocation of responsibilities on the basis of utility and
bargaining power and thus the locus of decision making authority reveals the implicit
weights attached to household members' utility functions. However, suppose spouses
are highly altruistic and have good information about each other's preferences. Then
allocation of decision making responsibilities might reflect comparative technical
advantages—i.e. production-function factors—not welfare weights. A common example
of the separation of preferences and allocation of responsibility is the "honey-do" list,
suggesting that the wife's preferences dominate prioritizing household tasks, but the
husband has responsibility for actually doing the tasks.

The introduction to this draft promises to employ a mental-models framework, but the
focus is primarily on the cooperative household decision model. It is not clear to what
extent these two frameworks are complements or substitutes. In any case, people may
not be good at explaining decision processes after the fact. Well-known problems with

7

-------
recall bias are likely to be even more serious in reconstructing subjective thought
processes. Curiously, the authors average couple's answers in many cases, implying
equal utility weights, which their conceptual framework suggests that is unlikely. The
inverse correlation between income and joint decision making may simply indicate that
joint decisions are time-intensive and the opportunity cost of time rises with income.

The evidence on disjoint reporting of supposedly factual data may be the most
interesting feature of this study. I would have liked to see more effort to explain the
direction and magnitude of disjoint responses. It is curious that 88% said decisions were
made jointly, which isn't consistent with evidence on specific decisions. How do these
results relate to the theoretical material? How might disjoint perceptions affect
household decision making? It might be interesting to ask how responsibilities have
changed over time. Suppose decision-making responsibility evolves over time as family
circumstances change or couples gradually specialize. It is possible that disjoint
responses are partly explained by husbands and wives averaging over different time
periods. Perhaps the wives are recalling recent history and husbands are averaging
over a longer period.

The main result from the quoted interview material seems to be that "a majority" were
willing to consider tradeoffs. Of course, that result should be evident in a pretest of the
instrument. Some of the quotes may reflect socially acceptable attitudes. We're actually
less interested in their willingness to trade in the abstract than whether they are willing to
accept tradeoffs in the specific context of a preference elicitation. I look forward to
seeing how insights obtained from this study influence the design of a stated-preference
survey to obtain true household values, including values for children's health.

References

Bateman I. and Munro A. 2005. An Experiment on Risky Choice Amongst Households,
The Economic Journal. 115, C176-C89.

Hirth, R.A. et al. 2000. Willingness to pay for a quality-adjusted life year: In search of a
standard. Medical Decisionmaking 20: 332-42.

Johnson, F.R. 1005. Einstein on Willingness to Pay per QALY: Is There a Better Way?
Med Decis Making25: 607-608.

8

-------
Summary of the Q&A Discussion Following Session IV

J.R. DeShazo, (UCLA)

NOTE: Dr. DeShazo's comments/questions were inaudible at first and are picked up
here toward the end.

"I think, very importantly, people come to choices with subjective expectations that arise
out of information they collected based on mental models they currently use. So, very
often, subjective expectations about risk levels and risk reductions associated with
different hazards and different programs are brought into the survey environment, and we
have no idea really what's going on there.

Finally, in terms of the parent-child relationship, whether the parent is practicing
altruistic paternalism or not is probably going to be a function of the age of the child. I
can force my five-year-old to eat her vegetables, but I probably won't feel a
responsibility to do that for my 25-year-old daughter. So, understanding the nature of the
parental responsibility comes from understanding how they represent their role as a
parent in their child's health."

Sandra Hoffmann, (Resources for the Future)

"One comment I'd like to make is on the relationship between the hazard and the health
outcome: This is a classic way in which mental models are used. We didn't discuss this
in our presentation today, but that's a major focus of the mental model study that we
conducted. We structured what is called an "expert mental model" of the relationship
between the environmental hazard and the risk that was peer-reviewed by a number of
leading experts on children's lead hazards. That is being used as a basis to compare the
parents' understanding of the relationship between lead exposure and health outcomes—
and between mitigation and health outcomes. Our intention is to use that to help refine
the way the risk is presented, and it's been used that way to improve risk communications
in the past."

Alan Krupnick, (Resources for the Future)

Dr. Krupnick added, "Of course, we're planning on getting into the decision-making
process mental model," and noted that they would be refining the work that was
presented at the workshop. Addressing Dr. DeShazo's comments more directly, he
stated, "I like the idea of asking perhaps some direct questions to try to get at their mental
model for parental responsibility for the child. We thought we could get at that by just
asking decision-making questions with respect to children's health and so on, but it's not
enough. We can maybe get at it more directly."

Bryan Hubbell, (U.S. EPA)

Addressing his questions to Dr. Hammitt, he commented, "When we're dealing with the
IQ evaluation, one of the things that struck me is when you asked the parents for their
willingness to pay, and the reason it might be different than the cost of illness, is that
you're essentially asking them to be able to project the relationship between IQ loss and
future earnings. If they don't actually know that relationship, you're asking them to

Session I Q&A

1

-------
somehow figure out what that six-point difference means. A question I have is: Could it
instead be offered in showing them the information that's in the epidemiological
literature relating the two?" He went on to phrase the question another way also: "If
they're really not giving you their expectations of earnings loss, should this willingness to
pay actually be additive to the cost of illness—so that there is some kind of estimate of
utility loss beyond earnings?"

Still addressing Dr. Hammitt, Dr. Hubbell continued by saying he was also concerned
about "the payment vehicle, in that you had it be a one-time payment in a particular year
for what is essentially a lifetime impact." His question was: "If you would ask them
instead what they would be willing to pay annually up through their child's eighteenth
birthday in order to prevent this kind of exposure, would you be able to get a different
value per IQ point? Again, this would reflect a lifetime impact rather than just a one-time
payment, because you start getting into budget constraint issues and current trade-offs
versus future earnings potential and future impacts."

Dr. Hubbell continued, "On your pesticide questionnaire one thing I'm really concerned
about is the payment vehicle, again." He cited a study done by Kerry Smith and
colleagues back in 1994 (he believes), in which they looked at the willingness to pay for
avoiding risks from pesticides, focusing on grapefruit. Dr. Hubbell stated, "If you
calculate a VSL based on their results, you get something like $80,000 or perhaps
something even lower. Part of the reason for this is because it's tied to the specific
product or to a particular sub-category of your budget. In those cases, in order to get a
VSL that is more typical of what we get for environmental policy, you would have had to
pay something like a hundred times the price of a grapefruit. Clearly, people are going to
reject that. They're either going to hit the reservation price, or they're going to substitute,
or something else." He closed by saying that his concern is that "you're going to run into
the same problem here. While it still may be good to test the latency question, I wouldn't
want to be able to use that VSL for anything—it's not really a VSL. The other related
question is: While you say that you're not going to focus this on organics, people use
organics as sort of a reference point. They know what organic foods cost and they've
already made the decision one way or the other, so you can see that as a bounding on
their willingness to pay extra for products. In fact, what they may do if you tell them a
price that is higher than the organics is decide just to go to organics to get the health
benefits plus the eco-benefits. Again, there's a bounding question there."

James Hammitt, (Harvard University)

Saying that those were "all good points," Dr. Hammitt first addressed the willingness to
pay per IQ questions. He stated, "Clearly, I don't mean to suggest that EPA should use
our value instead of the cost of illness. I think it's clear that people don't appreciate how
much IQ apparently contributes to lifetime earnings. Whether some CV value should be
added to the cost of illness value, I don't know—it might be that some part of the cost of
illness is already in the CV. That's a good question."

Session I Q&A

2

-------
Turning to the one-time payment issue, Dr. Hammitt clarified: "The willingness to pay
amounts, the bids we offered people, were not extraordinarily high—they were a few
hundred dollars. If you think of that as part of a tax payment, I don't think the income
effect is going to be really important there. The one-time payment is consistent with the
intervention as a one-time cleanup that will provide a long-stream term of benefits. So,
asking about a one-time payment is not unreasonable on its face. These one-time
cleanups of course could be financed by bonds, thereby spreading the cost to the
taxpayers over many years, so one could do it many ways."

Reiterating that "everything matters," Dr. Hammitt continued, "I mentioned in the ERS
study we asked about paying per meal or paying per month, where we had information on
the frequency with which people consumed the various foods. So, we told them what the
risk reduction would be on a per-month basis as well. I think our estimates of willingness
to pay per meal are implausibly high—I think they're off by a couple of dollars per meal.
That may be due to error in the sense that we tell them that the risk of getting sick from
this one particular meal. . . -so there's a huge amount of salience there and maybe that's
why they're paying a lot." He summarized that a $3 per meal increase over a month
period really adds up to some money, but the gauge also involves "much bigger risks—
these microbial illness risks are huge. So, as it turns out, our willingness to pay per unit
of risk reduction is actually a little bit higher on the per-month basis than on the per-meal
basis. But this is a general issue—how we allocate the timing of payments and what the
benefits are, I think, is going to matter to our results."

Susan Chilton, (University of Newcastle, United Kingdom)

Addressing her comment to Alan Krupnick and Sandra Hoffmann, Dr. Chilton said, "The
issue about whether the mother's and the father's willingness to pay is the same—if it
follows some empirical work that I've just completed—they won't be. In my study, they
were asked separately and there were differences. Another interesting thing we found
was that for an injury of low severity the mother's willingness to pay was higher than the
father's in the same household. As the injury became more severe—this was in the
context of child farm safety—the father's willingness to pay became higher than the
mother's willingness to pay. It may be that the major decision maker in a household
changes across the scope of an injury or illness, so that may be something to bear in
mind."

Mary Evans, (University of Tennessee)

Stating that she had "just a quick clarification question" for Drs. Krupnick and
Hoffmann, Dr. Evans asked, "Can you talk a little bit about the level of information of
respondents when they go into the initial interview? For example, are they aware of the
fact that they will first be interviewed separately and then jointly—or are they expecting
only to be interviewed by themselves?"

Session I Q&A

3

-------
Alan Krupnick, (Resources for the Future)

Dr. Krupnick responded that the participants are aware of the format of the interview. He
went on to clarify: "Actually, before the interview starts they are brought in together and
asked to write down three recent decisions they've made regarding their children's
health. The interviewers then get together and look at the responses and find one that's
the same (or if not, they go back to the participants). Then, they use that common
decision as the basis for the discussion of the decision making styles in the separate
interviews. They know that they will then be coming back together to complete a second
interview, so, yes, there is full information on that."

Dr. Krupnick added, "I'm not sure what your concern was—why don't you go a little
further on that one?"

Evans

Dr. Evans clarified, "I guess I was just thinking about the broader implications of the
question on who should we survey? Even in that context, if you find willingness to pay's
to be equal, it still is not surprising that in a context where they're interviewed separately
and those answers will never be rectified that we can see differences."

Lauraine Chestnut, (Stratus Consulting)

Addressing Dr. Hammitt, Ms. Chestnut said, "Maybe I need to see how you get from the
question you asked about the IQ to the dollar per IQ to clarify this, but weren't you
asking people about their willingness to pay for a cleanup program that's going to reduce
risks to somebody's children but not necessarily their own? How many children were in
the community? I guess I'm fuzzy about how we get from that to dollar per IQ—is that
per one kid or per the community? Are we comparing apples and oranges?"

James Hammitt

Dr. Hammitt answered, "There is potentially a little ambiguity on that, but the idea is:
What would you pay to reduce the risk that your child has this? So, it's one child—and
then it's a reduction in the risk of suffering the six-point IQ deficit. So, it's willingness to
pay divided by the change in probability divided by the six IQ points."

Chestnut

"So, it's: Suppose you had a child, and then . . ."

Hammitt
"Yes, right."

Sylvia Brandt, (University of Massachusetts)

Dr. Brandt asked this question of Drs. Krupnick and Hoffmann: "How are you going to
connect your theoretical model to an empirical study? The reason I ask is because I have
a concern. In building your theoretical model, you're working with a group of
homogeneous, very traditional households. I understand why you wanted that group to
be homogeneous. However, when I think about the population that we worry about when

Session I Q&A

4

-------
we think about lead, I think about two things. One is housing structures of poor quality,
typically in inner-city, lower-income neighborhoods. The second thing is poor nutrition,
because the lower the iron level in your blood, the more likely it is that lead will bond to
red blood cells. Both of these are more likely to occur in low-income, non-white
populations. I know from personal experience in the Springfield, Massachusetts area,
where we have a lead paint problem, eighty percent of our group were single-parent
households. They were typically female, but they varied from being an aunt to a foster
parent to a grandparent, so there was a lot of variation in the household structure. I
wonder how you're going to make that leap from a model built on what I think of as a
suburban setting to where the real problem is." Dr. Brandt went on with a second
comment related to how participants were asked to rank health effects. She stated,
"Again, building on my experience in Springfield and Oakland, when we ask households
to rank health effects or health risks, they all might be ranked pretty low. For example,
asthma morbidity, which in the suburb we may think is just outrageously out of control,
may not be ranked as a high stress in inner-city households because they have competing
stressors that are more basic than improved health—maybe it's making the rent payment
or dealing with spousal abuse or kids' school issues, whatever. So, I would encourage
you in asking about what are concerns to include, along with the health issues, also other
things that may be important and that may completely dominate any health-related
concerns in those settings where lead is a real problem."

Sandra Hoffmann

"In response to the first question, the focus of the study is really to try to get at the
methodological question about whether we're taking the right approach in stated-
preference surveys when we're trying to get at parental willingness to pay. The sample
size that we can do, given the grant size, is fairly small, so it's always been conceived of
as a pilot study that is focused on trying to examine this household modeling question.
So, no, I don't think we're going to get really good measures of willingness to pay for
reduction in neurotoxins that are representative of the entire population. That said,
twenty -five percent of children in our country do live in homes that have lead paint as a
potential hazard. I know in interviewing physicians in the Washington, DC area, they say
that while one would expect that the risk is going to be highest in low-income
households, they also see a lot of problems still in middle- and higher-income housing.
So, what we're looking for are housing settings in which it could be a problem and family
settings that raise a scenario in which we can test the alternative household hypotheses.
Further work will have to be done to get more representativeness in income on
neurotoxin hazards."

Alan Krupnick

Dr. Krupnick added, "Your second point is well taken, and we'll think about how to do
that. On the first point I just wanted to add that we have no intent of generalizing these
results beyond the group that we're targeting. We do find, however, that race has a
significant effect on decision-making style—but, in our data it's correlated with income,
so it's hard to know which is doing what."

END OF SESSION IV Q&A

Session I Q&A

5

-------
Morbidity and Mortality: How Do We Value the Risk of

Illness and Death?

PROCEEDINGS OF SESSION V: EMPIRICAL ISSUES ASSOCIATED WITH

MORTALITY RISK VALUATION

A WORKSHOP SPONSORED BY THE U.S. ENVIRONMENTAL PROTECTION
AGENCY'S NATIONAL CENTER FOR ENVIRONMENTAL ECONOMICS AND
NATIONAL CENTER FOR ENVIRONMENTAL RESEARCH

April 10-12, 2006

National Transportation Safety Board
Washington, DC 20594

Prepared by Alpha-Gamma Technologies, Inc.
4700 Falls of Neuse Road, Suite 350, Raleigh, NC 27609

ACKNOWLEDGEMENTS

This report has been prepared by Alpha-Gamma Technologies, Inc. with funding from
the National Center for Environmental Economics (NCEE). Alpha-Gamma wishes to
thank NCEE's Maggie Miller and the Project Officer, Cheryl R. Brown, for their
guidance and assistance throughout this project.

DISCLAIMER

These proceedings have been prepared by Alpha-Gamma Technologies, Inc. under
Contract No. 68-W-01-055 by United States Environmental Protection Agency Office of
Water. These proceedings have been funded by the United States Environmental
Protection Agency. The contents of this document may not necessarily reflect the views
of the Agency and no official endorsement should be inferred.

-------
Table of Contents

Session V: Empirical Issues Associated With Mortality Risk Valuation

Session Moderator: Nathalie Simon, U.S. EPA, National Center for Environmental
Economics

Update on Mortality Risk Valuation at EPA

Kelly Maguire, U.S. EPA, National Center for Environmental Economics

Eliciting Risk Tradeoffs for Valuing Fatal Cancer Risks

Chris Dockins, U.S. EPA, National Center for Environmental Economics; George
Van Houtven, Research Triangle Institute; and Melonie Sullivan, Institute for
Family Centered Services

Update on Mortality Risk Valuation at EPA

Kelly Maguire, U.S. EPA, National Center for Environmental Economics

Eliciting Risk Tradeoffs for Valuing Fatal Cancer Risks

Chris Dockins, U.S. EPA, National Center for Environmental Economics; George
Van Houtven, Research Triangle Institute; and Melonie Sullivan, Institute for
Family Centered Services

Questions and Discussion

-------






\7aki!ngR)rtality Rfelc Reductions







^ifsentation at EPA Workshop



(^Morbidity and Mortality:



How Do We Value iiie Risk of Illness and Death?



April 11, 2006



Kelly Maguire



US EPA



National Center for Environmental Economics

i



£e,(VSL)1stimate used in
risk reductions

fstimafie used since 1999

_ Konomic Guidelines and

H Literatuf^Jrowiffionsiderably since default

estimate was derived
™ EPA commissioned reports have raised issues with
underlying literature

- Recently published meta-analyses provide new
means of combining estimates

2

-------
BK—illanrp

estimate of^.'2 nnl^ion ($1999) as an
~ estimate ojjthe value"of statistical life

studies
-2l| hedonic wage studies

Studies elate from 1974-1991)

- Values range from $0.7 million to $16.9 million
($1999); $0.8 mil|on to $18.6 million ($2004)

• One value from each study was used to fit a
Weibull distribution

BifflElEB

CentrkL^allfe framffie range of values
suffe^d ij^CenSmeta-analyses

DismDutibf1) 'nas a confidence interval from $1
to $13 million^

™ $CTffl|fi|l ls lower end of interquartile range from
Mrozek and Taylor (2000)

- $10 million is upper end of interquartile range
from Viscusi and Aldy (2003)

4

-------
EPllE® estimate to all

; rameWstimite to alji types

VSL s adjusted for timing

- Discounted for risk reductions in future
years

*liSflatecrto account fojgrowth in real
income over time

5

BMm

lortality risk valuation

ded tiree'stud ies to examine
vanouf se^mejt^df tftelteratu re

^5§Hck, et ai. (2002): Ww literature
t AlberS (2004): CV literature
- Blomquist (2004): AB literature

6

-------
;SJL (coot)

^iXieta-ajjalyses of mortality risk

1^1 rozgkIrljjaylCfr (2000jTjHW only
^vIsolli^anSKl d y ^003): only
Koichi, et aj. (forthcoming): SP and HW

7

Proegss fS^Reyisipg.VSL:
-aitalfSfe

SABsEEAC expressed aiwiterest in

mata-a na lysis

PaneKof ©cperts Si#tpDecember 2005

-	Dllcuss the rissues aid challenges in
conducting meta-analysis for mortality risk
literature

-	Prepare summary report

-------
Revising VSL:
wfl&l £AB-BEAC

: meta-analysis report by

Othgjj^sues to-iwse
^Pqfllijp# sSu es

^Types oretudies
- Raevarft ifiea.su res
™CovaMtes

Guidance on revising VSL estimate

16 optiQns

jse feurrefit esti mate
Srorffexistng meta-

• DerWem^N estimate

-nfa distribfition
* Con duct a new meta-analysis
- Other?

10

-------
report for

Prepay Vfnite Piper fofpresentation to
SaPeem m gtd&y selection criteria
anPmeEf^dology to combine estimates

SAWeK meeting jffi^uly 2006

Final guidance to be completed n 2007

-------
Eliciting Risk Tradeoffs for Valuing Fatal Cancer Risks

George Van Houtven*

RTI International

Melonie B. Sullivan

Institute for Family Centered Services, Inc.

Chris Dockins

US Environmental Protection Agency

May 2006**

Paper prepared for presentation at

U.S. EPA NCER/NCEE Workshop:

"Morbidity and Mortality: How Do We Value the Risk of Illness and Death "

Washington, DC
April 10-12 2006

*Send all correspondence to: George Van Houtven; RTI International; 3040 Cornwallis Road; P.O. Box 12194;
Research Triangle Park, NC 27709; Voice: (919) 541-7150; Fax: (919) 541-6683; e-mail: gvh@rti.org.

** We dedicate this paper to the memory of our courageous and passionate colleague Elizabeth McClelland, formerly
of the National Center for Environmental Economics, U.S. EPA, whose efforts were integral to the genesis and
early design of this project. Financial support for this research was provided by the U.S. Environmental Protection
Agency under Cooperative Agreement C R 824861-01-0. Thanks are due to John Bennett, Rebecca Allen, Mark
Dickie, James Hammitt, and Alan Krupnick, Clark Nardinelli, Maureen Cropper, and Mary Evans for their helpful
comments and suggestions. We also acknowledge research assistance provided by Catherine Corey and Jui-Chen
Yang. Any opinions, findings, conclusions, or recommendations expressed in this paper are those of the authors
and do not necessarily reflect the views of the U.S. EPA.

-------
Abstract:

Reductions in cancer risks are among the most important and tangible benefits resulting from a
variety of environmental, food safety and other public health initiatives; however, relatively little
is known about how individuals value reducing cancer risks compared to other types of risks.
Most of the existing empirical research on the valuation of mortality risks has focused on
accidental (occupational and/or automobile) fatalities. It is often argued, however, that differences
between the characteristics of cancer risks and accidental risks may lead to significant differences
in how they are valued. In particular, the time lag between exposure to carcinogens and its
physical manifestation (i.e., the latency period), as well factors such as the fear, dread, pain and
suffering may affect individuals preferences for avoiding cancer risks. To address this issue, we
conducted a national survey of adults that elicits their relative preferences for avoiding
automobile fatality and fatal cancer risks. We specifically examine how strongly individuals
prefer avoiding one type of risk over the other, how this strength of preference is affected by the
length of the morbidity and latency periods, and how preferences differ across different types of
cancer. Our results indicate that individuals generally have a strong preference for avoiding fatal
cancer risks relative to automobile fatality risks; however, as expected, this preference is
inversely related to the length of the cancer latency period.

2

-------
Introduction

Reductions in cancer risks are among the most important and tangible benefits resulting
from a variety of environmental, food safety, and other public health initiatives.
Nevertheless, little is known about how individuals value reducing cancer risks relative to
other types of risks. Although a large empirical literature exists that has generated
estimates of willingness to pay (WTP) to reduce mortality risk, this literature has focused
almost exclusively on accidental (occupational and/or automobile) fatalities. It is often
argued, however, that differences between the characteristics of cancer risks and
accidental risks may lead to significant differences in how they are valued (as measured
by WTP). Specifically, the time lag between exposure to the cancer risk and its physical
manifestation (i.e., the latency period) may lower WTP for cancer risk reductions relative
to accidental risk reductions. In contrast, the pain and suffering associated with the
morbidity period that precedes a cancer death may increase WTP for reducing cancer
risks. If a fatal cancer engenders more fear and dread than an accidental fatality, then
WTP to reduce cancer risks may be higher.

Although a few studies have examined the empirical effects of risk characteristics on
preferences for risk reductions, there has been little attempt to specifically and
systematically test for how variation in latency and morbidity associated with cancer
affects preferences.! To address this research gap, we have designed and implemented a
national survey of adults that elicits their relative preferences for avoiding two types of
potentially very different mortality risks—risk of automobile fatality and risk of
contracting a fatal cancer.

The objective of this study is to use stated preference methods to assess individuals'
tradeoffs between the two types of risks. In particular, we estimate how strongly
individuals prefer avoiding one type of risk over the other, how this strength of
preference is affected by the length of the morbidity and latency periods, and how
preferences differ across different types of cancer. In addition to informing the debate on
how individuals perceive different types of risks, and how these perceptions may affect
preferences to reduce different risks, the results will indicate that additional research on
valuing different types of fatal risks is warranted.

The analysis finds that individuals have a strong preference for avoiding cancer risks
relative to automobile fatality risks of the same magnitude; however, as expected, this
preference decreases as the cancer latency period increases. Individuals' preferences for
avoiding future cancer risks are also, as expected, positively related to their chances of
surviving until the age of onset of illness. The details of these results are discussed below.

1 The only exception appears to be recent work by Trudy Cameron and J.R. DeShazo described in section 1, below.

3

-------
1. Background

Although some anecdotal evidence exists that preferences for reducing fatal cancer risks
may depend on characteristics of the risk, the U.S. Environmental Protection Agency's
(EPA's) Science Advisory Board Environmental Economics Advisory Committee (SAB-
EEAC) (EPA, 2000) concluded that "there is not sufficient theoretical and empirical basis
for ... accounting for these differences [in characteristics]." On the specific question of
how cancer risk valuation differs from risks of accidents, the SAB-EEAC noted that "the
value of reductions in cancer risks should include both the value of the reduced risk of
death and the value of reduced risk of the morbidity, fear, and dread that precedes the
death incident" but added that "existing studies provide little reliable information as to
the magnitude of this premium." The purpose of this study is to begin to provide a basis
for this empirical research by exploring specific aspects of cancer risk valuation in a
limited sample setting.

The existing empirical literature on mortality valuation focuses largely on safety rather
than on cancer-related mortality risks.2 There are a few exceptions, but even these
studies provide relatively little evidence on how specific risk characteristics may affect
preferences for reducing fatal cancer risks. For example, Smith and Desvousges (1987),
Hammitt (1990), and duVair and Loomis (1993) use contingent valuation (CV) to
estimate individuals' WTP to reduce environmental/food safety risks of death; however,
in none of these cases were deaths described or necessarily interpreted as cancer risks. In
contrast, Carson and Mitchell (2000) used a CV survey, conducted in 1985, to elicit WTP
to reduce carcinogenic risks from trihalomethanes in drinking water. The survey did not
specify the type of cancer or related health outcomes, nor did it discuss latency effects;
therefore, it is difficult to establish whether and how these factors affected respondents'
stated preferences. The Carson and Mitchell results were based on a relatively small (n =
237) and very localized sample (Herrin, IL).

Two recent surveys from Asia have begun to examine cancer risks relative to other
mortality risks. Hammitt and Liu (2004) employ CV in Taiwan to estimate WTP to
reduce risks of cancer and non-cancer illness (liver disease). The design characterizes
risks as either acute (beginning within a few months with death to follow 2 or 3 years
later) or latent, where symptoms begin about 20 years in the future. Results suggest that
WTP for cancer is about one-third larger than WTP to reduce the risk of a comparable
chronic disease, but the estimate is not statistically significant at the 10% level.
Individuals appear to discount for latency at about 1.5% per year, but the survey does not
vary latency periods, morbidity times, or consider accidental fatalities.

2 Some of these studies address risk dimensions associated with cancer, such as latency (see, for example, Krupnick, et
al. 2005). There is also a body of literature on how public preferences for risk reduction programs vary by type of
illness and other risk attributes (e.g., Subramanian and Cropper, 2000).

4

-------
Tsuge, Kishimoto, and Takeuchi (2005) use a choice experiment to estimate marginal
WTP for reduced mortality risks from cancer, heart disease, and accidents. Results from
a Tokyo-area sample (n=400) show a small, but significant preference for reducing a
generalized cancer risks relative to heart disease and accidents. This difference is not
sensitive to risk magnitude. Results also suggest that respondents exhibited a strong
preference for earlier risk reductions, with implicit discount rates estimated at
approximately 20%. The survey does not vary morbidity periods or examine specific
cancer types.

Cameron and DeShazo (2004) use a choice experiment to evaluate several aspects of
health and risk valuation, including cancer, morbidity, and latency effects. Draft results
suggest WTP generally diminishes over latency periods, with the specific effect
contingent upon age, wealth, and the "illness profile," including length of morbidity and
whether or not the effect is ultimately fatal.

Magat, Viscusi, and Huber (MVH) (1996)) used a computer-based survey to explore
individuals' tradeoffs between automobile fatality and specific cancer risks. Using a mall
intercept recruitment in Greensboro, NC, MVH administered the survey to 727 adults.
The survey asked individuals to choose between two hypothetical residential locations
that differed only in terms of the risks of automobile death and risk of lymph cancer. By
varying the risks in the two locations, MVH estimated the lymph cancer "risk
equivalents" for auto death—the risk ratio at which respondents were indifferent between
the two locations. They found that for terminal lymph cancer the median respondent
viewed the two risks as equivalent (risk equivalent = 1), and on average, nonfatal lymph
cancer risks were "valued" at roughly two-thirds the rate of auto death risks. Although
respondents were provided with information about the consequences of the disease, it is
not clear which attributes of the disease or its risk primarily affected individuals' relative
preferences for avoiding the two types of risks. In particular, the role of latency periods
for cancer risks was not addressed in this study.

This study builds on the work by MVH, using a similar preference elicitation method.
However, our survey is specifically designed to examine how individuals' risk
equivalence rates between auto death and fatal cancer risks are affected by latency
periods, morbidity outcomes, and types of cancer. This type of information is not
available from existing research.

3. Conceptual Model

To model risk preferences we use an approach similar to MVH (1996). We assume that
respondents make choices to maximize expected lifetime utility, E(U), which is defined
in the following way:

5

-------
E(U) = PdU(DJ) + PcU(CJ) + (1 - PD - PC)U (H,Y)	(1)

According to this expression, lifetime utility is determined by health outcomes (D, C, or
H) and wealth (Y). Individuals are assumed to face probabilities of three mutually
exclusive lifetime health profiles. The first is dying in the very short term (e.g., within a
year) in an auto accident (D) with probability Pd, the second is contracting an eventually
fatal cancer (C) with probability Pc, and the third "normal health" or all other health
outcomes (H), with probability 1 - Pc - Pd^-

Totally differentiating Eq.(l) and setting dE(U) = 0 and dPc = 0 results in the following
expression for the marginal rate of substitution between income and risk of automobile
death, which is also commonly referred to as the value of a statistical life (VSL) (see, for
example, Hammitt [2000]).

dPD	my)

OY

Alternatively, setting dPD = 0 results in the following expression for the marginal rate of
substitution between income and risk of cancer, which can be interpreted as the value of a
statistical cancer case avoided (VSC):

dY U(H,Y)-U(C,Y)
dPc	dE(U)	( }

dY

Combining Eqs. (2) and (3), and assuming that U(D,Y) = 0, the relationship between VSL
and VSC can be rewritten as

MER__vs£Jl_m£j1\

VSL y U(H ,7) J

We refer to the term in brackets as the "mortality equivalence ratio" (MER) for avoided
fatal cancer risks, which translates avoided fatal cancers into equivalent avoided
accidental deaths. In other words MER = VSC/VSL. Therefore, if MER is less (greater)
than 1 this implies that avoided fatal cancer risks are valued less (more) than avoided
immediate mortality risks from car accidents. If, for example, MER equals 2, this implies
that an avoided fatal cancer is "equivalent" to 2 avoided car fatalities.

3 Since latent risks of cancer are only relevant if one does not die from immediate automobile fatality risks, Pc in this
model should more accurately be replaced by Pc(1-Pd) in Eq.(l); however the second order interaction of the two
risks is small enough relative to Pc that excluding the interaction has little effect on the analysis.

6

-------
The conceptual framework outlined above defines the general relationship between VSL
and VSC; however, it does not specifically address how VSC is expected to vary with
respect to characteristics of the cancer risks. In particular, it does not address the effects
that latency, t, may have on the expected utility of the fatal cancer profile, U(C, Y) and
therefore on VSC and MER. Including latency in equation (4) we express MER as:

MERjt) = (1 -	(5)

U{HJ)

In this expression, the lifetime utility of the cancer health profile can be expressed as the

discounted sum of utilities in future periods.

f t-1	'N

£ 0, )Wh 0;.) + (st )(dt )uc (yt)	(6)

V j=°	J

U(C(t),Y) =

where:

Sj= probability of survivingy periods into the future from the present (j=0)
dj = time preference factor, discounting utility in period j to the present
yj = consumption in period j

uk(yj) = state dependent utility in period j, with k=c referring to cancer state and
k=h referring to healthy state.

Similarly the lifetime profile for H, which is independent of the latency factor t, can be
expressed as:

CO

U(H,Y)	(7)

j=0

Based on this framework it is possible to formulate and examine specific hypotheses
regarding the effects of latency period and perceived survival probabilities on preferences
for avoiding cancer risks. In particular, as demonstrated in Appendix A,

7

-------
•	an increase in the cancer latency is expected under most circumstances to have a
negative effect on MER, so the relative preference for avoiding cancer risks
declines^, and

•	for any specified cancer latency t, an increase in the perceived survival probability
to that period (st) is expected to have a positive effect on MER.

4. Empirical Methods

In our survey, which is described in more detail below, respondents are faced with a
choice between two locations, A and B, where the only difference between the locations
is the rate (i.e., risk) of fatal cancers (Pc^ vs. Pc®) and auto deaths (Pd^ vs. pdb).
Location A has fewer auto deaths and Location B has fewer cancers than the respondent's
current location. In effect, they are presented with a pair of lotteries and asked to choose
the one they prefer.

This choice is also illustrated in Figure 1, where Location A has fewer auto deaths and
Location B has fewer stomach cancers than the respondent's current location. To
compare the two options, we define the risk difference ratio (RDR) between A and B as:

PB _ pA

RDR = S.	(8)

PA _ PB
1 c 1 c

The RDR therefore represents the slope (in absolute value) of the line between the A and
B risk combinations. The respondent is assumed to choose the location that lies on the
indifference line that is closer to the origin (i.e. the risk combination that provides the
highest expected utility).

In Figure 1, both locations A and B are shown on the same indifference line. Indifference
between the two areas (lotteries) implies that these areas offer the same expected utility:

P£ U(D, Y) + PCA U(C, Y) + (1 - PA - PCA )U(H, Y) =

P*U(DJ) + PcBU(C,Y) + (1 - P* - PCB)U(H,Y)

Assuming again that U(D, Y) = 0, and rearranging terms

4 Using a somewhat different framework, Hammit and Liu (2004) also conclude that under most conditions,

individuals' willingness to pay for reducing latent risks will be lower than for reducing current risks by the same
amount.

8

-------
MER = \ -

U(C,Y)

PB

1 D

~ PA

1 D

U(H,Y)

PA

1 C

~ PB

1 C

= RDR"

(10)

This equation show that the value of RDR that equates E(U) between two location —
RDR* — is also equal to MER. In other words, MER represents the negative slope of the
indifference curves in Figure 1; therefore, it defines the RDR that is consistent with
indifference between the two locations.

By varying the values of Pc® and Pd^ across respondents, the survey presents location
choices that entail different RDRs. By observing how choices vary with respect to this
variation in RDR, the survey responses can be used to estimate the average/expected
value of MER. Equally important, they can be used to estimate how MER varies
according to the characteristics of the cancer risks and the characteristics of respondents.

To model and interpret results using the discrete choice approach, we assume that MER
varies in both systematic and stochastic ways across respondents. This assumption is
formally expressed as

MERt = a + PXi + si.	(11)

The systematic component of this expression describes MER as a function of X., which is
a vector that includes both survey variables and characteristics, as well individual
characteristics. The random component (e,) captures factors that are unobservable to the
analyst and are assumed to vary randomly, identically, and independently across
respondents.

In the discrete choice context, one does not observe MER, for each respondent but rather
a latent variable mi*, which can be characterized as

mi* = 0 ifRDRi >MERi	(12a)

mi* = 1 ifRDRi < MERi.	(12b)

In this case, mi* can be represented by a dummy variable, which is equal to 1 if the
respondent prefers Location B (the location with fewer cancers) and 0 otherwise. In
other words, the lower (higher) the value of RDRt, the larger (smaller) is the reduction in
cancers relative to auto deaths and the more likely that respondent i chooses Location B.

Assuming that et is normally distributed N(0, crj, a. probit model can be used to analyze
the discrete choice responses and to estimate coefficients of the MER function (Eq. [4.9])
and o. The results of the probit analysis are discussed in Section 7 below.

9

-------
A stated preference by individual i for Location B (PREFERBi=l) indicates that she
prefers the location offering a reduction in cancer risk over the location offering a
reduction in auto death risk. It also implies that MERj>RDRj. Given the probability
distribution of sj, the probability of preferring Location B can be expressed as:

Pr (PREFERB1 = 1)

= Vr(MERt > RDRt)	(13)

= Pr( pXt - RI)Ri > Sj)

The last equality holds due to the symmetry of the distribution. By defining 6 = s! a ,
we define a standard normal random variable, 0 «iV(0,l), which implies that

Pr (PREFERB1 = 1)

= Pr(f|V - fi-W, >
-------
The survey instalment presented respondents with information on two hazards: death
from a car accident and death from cancer. Respondents were randomly assigned one of
three different types of cancer: stomach, liver, or brain cancer. The survey provided
information on national averages and ranges of risk of each hazard, as well as
information on cancer symptoms, treatment, and side effects. Importantly, the survey
further explained that, although death from an automobile accident usually occurs almost
immediately, cancers take years to develop before they are diagnosed (the latency period)
and that the individual is typically sick for some time before death occurs (the morbidity
period). Respondents were asked to assume for the purposes of the survey that the
latency period has a length of t years (where respondents are randomly assigned values of
t equal to 5, 15, or 25 years) and the morbidity period has a length of m years (with
randomly assigned values of 2 or 5 years). Time lines, which are individualized to the
respondent's reported age, were used to illustrate the differences in timing of exposure
and death from the two hazards. Specifically, the time lines show that auto accidents and
death are typically simultaneous occurrences, while demonstrating that exposure to the
carcinogen occurs in year 1, diagnosis occurs in year t + 7, and death occurs in year t +
m+1.

Respondents were then presented with a sequence of similar choice scenarios. They were
asked to imagine that they have a job that requires them to move to one of two areas (A
or B) for a period of 1 year. They must choose between the two areas, which differ only
with respect to their exposure to the two hazards. They were asked to assume that their
annual baseline risks from the two hazards— the risk of dying in auto accident and the
risk of dying of a specific cancer in their current area of residence - were both
represented by 100 deaths per million people. They were then asked to choose between
moving to Area A which has fewer auto accident deaths per million than their current
location or Area B which has fewer fatal cancer deaths per million than their current
location. Respondents were first introduced to the choice task with a few simplified
practice questions. They were then presented with a choice scenario where they faced a
tradeoff between avoiding risks of fatal auto accidents or avoiding risks of fatal cancers.
An example choice scenario is shown in Figure 2.

Several aspects of the questionnaire design were randomly varied across respondents to
test for their effects on a respondent's choices regarding risk reductions. These
treatments were selected to test for scope effects and question-framing effects. They
include the following:

•	three different types of fatal cancers (stomach, liver, or brain cancer, each
compared to fatal auto death risks);

•	three different assumed latency periods for the cancer (5, 15, or 25 years);

•	two different assumed morbidity periods for the cancer (2 or 5 years);

11

-------
•	two different formats for "introductory" choice questions, one in which Location
A was clearly superior in the introductory scenarios and the other in which
Location B was superior (included to test for framing effects in choice responses);
and

•	five different choice scenarios, each corresponding to a different RDR -

Thus, all together, there are 180 (3x3x2x2x5) different versions of the survey that are
randomized across respondents.

Several other design characteristics are noteworthy. First, restricting the time frame to 1
year allows us to focus on risks from 1 year of exposure to the carcinogen and avoids the
confounding issue of cumulative exposure to the carcinogen. This implicitly assumes an
underlying dose-response model in which a single exposure can cause the cancer, as
opposed to a model in which there is no risk of cancer until some threshold of cumulative
exposures is reached. Second, emphasizing that the two new areas are exactly the same
in every way but the risk exposure controls for the effects of other perceived location
characteristics on reported preferences. Third, providing a baseline and maintaining new
risk levels at or below the baseline controls for scenario rejection. Finally, after the
practice questions and before the choice task, respondents were reminded of their
individual time line for cancer exposure, diagnosis, and death.

6. Survey Data

The on-line WebTV survey was sent to a total of 1,351 households participating in the
KN panel. To ensure proper functioning of the instrument, a subset of this sample—125
households—was initially contacted via email, and responses were acquired from about
half this sample. After reviewing these responses and making minor adjustments to the
instrument, email invitations were sent to the remainder of the sample.

By the end of March, 1,010 individuals (each from a different household) had submitted
completed surveys to KN—a 73.7 percent invitation response rate. To achieve this rate
of response, several of the 1,351 households were sent email and telephone reminders
throughout the survey administration period.

To analyze responses to the main choice question, we excluded 136 respondents who
"failed" the practice choice question. That is, if respondents did not indicate a preference
for the "dominant" location (with fewer auto deaths and fewer fatal cancers), even after
being given a chance to revise their response, it was assumed that they did not understand
or were not willing to accept the choice scenario. An additional 17 respondents were
dropped because, when presented with an automated follow-up description of their
response to the first (nonpractice) choice question, they did not agree with the description

12

-------
but also were not willing to revise their answer. We also excluded 69 respondents who
did not have a preference for either location. Consequently, the size of the analysis
sample is 788 respondents.

To investigate whether there were systematic differences between the analysis sample
and (N=788) initial recruitment sample (N=1351), we conducted a probit analysis with
respect to demographic characteristics. This analysis revealed that age, race, education,
and household size were all significant determinants of whether respondents were
included in the analysis sample. However, when this process was included as the first
stage of a Heckman sample selection model, with the probit analyses described in Section
7 as the second stage, there was no evidence that the selection process led to biased
estimates of the coefficients in the second stage model.

Descriptions and summary statistics for all the variables used in the analysis are provided
in Tables 1 and 2 respectively. Overall, over half of the respondent (65 percent)
preferred to location with lower cancer risks. The average age of the sample was 45.4
years, ranging from 18 to 93, average income was $50,600, and the average number of
years of education was 12.5. Nineteen percent of the sample classified themselves as
from a minority group.

The analysis also includes variables describing respondents' experience with and
perceptions of cancer and automobile fatality risks. A relatively small percentage of the
sample had experienced cancer themselves (CANCYOU, 8 percent) or had a close friend
or relative who had experienced the cancer described to them in the survey
(CANCFRIEND, 12 percent). A somewhat larger percentage had experienced a serious
autoaccident (CARYOU, 18 percent) or had lost a close friend or relative to a car
accident (CARFRIEND, 17 percent). On average, respondents believed that they had
lower risk of dying of cancer or a car accident than others in their area; however, a large
majority indicated that, for the purposes of the survey, they were able to assume that their
risks were the same. Twenty five percent of respondents indicated that, in choosing
between Locations A and B, they considered the possibility that a cure for cancer might
be found.

Finally, to account for how differences in perceived survival probabilities affect
preferences for avoiding cancer risks, we included data from the survey where
respondents were asked: "How likely do you think it is, in percentage terms, that you
will live for another X years or more?" The value for X corresponded to the cancer
latency period that was presented to the respondent later in the survey. As expected, the
average perceived survival rate was higher for X=5 (87 percent) than for X=15 (75
percent), which was also higher than for X=25 (68 percent). However, contrary to
expectations and evidence from life tables, the perceived survival rate declined more
rapidly from 5 to 15 years than from 15 to 25 years.

13

-------
A main objective of the analysis is to evaluate how preferences for avoiding fatal cancer
risks relative to auto death risks vary with respect to the relative size of the risk
reductions and the length of the cancer latency period. Figure 3 provides a first look at
this issue, by graphing the percent of respondents who preferred the location with lower
cancer risks in relation to RDR and latency period. As expected, this percent generally
declines with the RDR (i.e., larger relative reductions in auto death risks reduce the
preference for the lower cancer risk location) and it declines with latency. The statistical
significance of these results and their implications for calculating MERs are examined in
the next section.

7. Model Results

Based on this framework, we estimated several probit specifications, all using PREFB as
the dependent variable. In the simplest model specification, we assumed only random
(no systematic) heterogeneity across respondents:

MERl=P0+sl	(16)

This model was estimated using probit specification (1) in Table 3. Xt in this case is
simply the constant term, and, using Eq. (15), expected MER is estimated to be 2.3. In
other words, without accounting for systematic heterogeneity across respondent
characteristics or across survey versions, individuals were estimated to value avoided
fatal cancer risks at somewhat more than twice the rate of fatal auto risks.

More complex models, which allow and control for heterogeneity in various ways are
reported in specifications (2) through (5) in Table 3. All of these additional
specifications control for and measure the effects of latency period on MER. These
models consistently find that respondents' choices are significantly affected by
differences in the latency period. As expected, individuals' preferences for avoiding fatal
cancer risks (relative to automobile risks) decrease as the length of the cancer latency
period increases. These specific results are described and discussed in more detail below.

Specifications (2) through (5) also control for the type of cancer (STOMACHC and
BRAINC), the duration of cancer morbidity (MORB5), and the framing of introductory
"practice" questions (INTROFORMAT), all of which were varied randomly across
respondents. The brain cancer coefficient is consistently negative and statistically
significant, whereas the coefficient for stomach cancer is never statistically significant.
These results suggest that individuals have a significant preference for avoiding stomach
and liver cancer risks compared to brain cancer risks.

Differences in the duration of cancer morbidity (MORB5) never have a significant effect
on stated preferences in any of the model specifiations. The lack of an observed

14

-------
morbidity duration effect on preferences may be because any negative effect of
increasing the length of illness prior to death is offset by a corresponding delay in the
time of death (for a given latency period, which in the survey is defined from the current
period to the time of diagnosis). Alternatively, the difference between two and five years
of morbidity may not have been large enough to influence respondents' choices.

In contrast, the framing of the introductory questions does have a significant effect on
respondent choices in all of the model specifications. Respondents who received the
format in which Location A (Location B) was clearly superior in the introductory
scenarios were also less (more) likely to prefer Location B in the choice question
involving a tradeoff between cancer and automobile risk reductions. Therefore, although
these questions were included to help respondents understand the choice framework, they
also appear to have created somewhat of a starting point bias for respondents..

Specification (3) also includes several demographic characteristics such as age, health
status, education, and race, as well as the variables characterizing respondents'
experience with and perceptions of cancer and automobile fatality risks. Of these
variables, the only ones that have consistently significant (at a 10% level or less) effect
on stated preferences are household income and whether they live in an MSA.

Individuals in higher income households are less likely to prefer reducing cancer risks,
whereas urban residents are more likely to do so. The effect of age on the relative
preferences for avoiding cancers is negative, but it is not statistically significant in any
specifications. To the extent that age affects preferences through perceived survival
probabilities, these effects are explored in specification (4) and are discussed in more
detail below.

Individuals' experience with and perceptions of cancer and automobile risks are also
explored in specifications (2) to (4). Although individuals were asked to assume, in
answering the choice questions, that their own risks were the same as others in their area,
perceptions of higher than average cancer risks for themselves made them more likely to
prefer the area with lower cancer risks. Similarly, perceptions of higher than average
automobile risks for themselves made them more likely to prefer the area with lower
automobile risks. These results suggest the respondents may have implicitly adjusted the
risk reductions presented to them to fit their own circumstances. Individuals who had
experienced cancer themselves were less likely to prefer reducing future cancer risks.

This effect is not statistically significant, but it may reflect some adaptation to the illness.
Also, respondents who had close friends or relatives die from cancer or automobile
accidents were more and less likely, respectively, to prefer avoiding these risks. These
effects are not statistically significant either, but they may be a sign of individuals'
heightened fear or dread of these outcomes through indirect personal experience. Finally,
individuals who had considered the possibility of a cancer cure were significantly less
likely to prefer avoiding latent cancer risks. This finding is consistent with individuals

15

-------
expecting to derive higher utility from a future cancer health state, if the cancer has a
lower chance of being fatal.

To evaluate the effects of latency on preferences, specifications (2) and (3) estimate
separate coefficients for the 15 year and 25 year latency period dummies {dx for LAT15
and a2 for LAT25), with the 5 year latency period as the reference condition. Both
coefficients are negative, significantly different from zero, and significantly different
from one another. Therefore, as expected, longer latency periods reduce the relative
preference for avoiding cancer risks. To specifically explore the effect of latency on
MER, we define latency-specific MER as MER(t), such that.

MER(t\=pX(t\+Sl	(17)

Using this definition and equation (15) and setting all variables in X set at their sample
means (except for LAT 15 and LAT 25), we estimate separate expected MERs for the 5,
15, and 25 year latency periods^. The predicted values range from 3.23 for the 5 year
latency to 1.54 for the 25 year latency.

If MER declines linearly with respect to latency, adapting equation (17) we then have:
MER(t)i = MER(0),. -t + ei	(18)

Testing the linearity restriction in specifications (2) and (3) is therefore equivalent to
testing whether a2-al = a]. Applying a Wald test to the estimated coefficients, we
found that in both cases the linearity restriction cannot be rejected (at a 5% level of
significance).

In specification (4), we impose the linearity assumption by replacing the latency dummy
variables with a continuous variable (LATENCY = t). With this model and equation
(18), it is also possible to extrapolate the results and estimate:

•	E[MER(0)]—the implied expected MER if latency were zero and the onset of
cancers, like auto deaths, were immediate^ and

•	t* —the length of the latency period that would be required to make expected
MER equal 1 (i.e., to make individuals indifferent between reducing fatal cancer
and auto death risks).

To estimate E[MER(0)] we set LATENCY=0 and the other explanatory variables in X at
the sample mean. The results, which are reported Table 3 indicate that on average,
avoided fatal cancer risks without latency would be valued at over three times avoided
automobile death risks.

To estimate t*, we define the following condition

5 For the MER calculations, the values of CANCERRISK and CARRISK were set at 3 - equal to the same risk as the

average individual in their area - rather than at the sample means, which were somewhat smaller.

6In the case of cancer, death would still be delayed by the duration of morbidity.

16

-------
E[MER(t*)i ] = PX(0),. -
-------
directly comparing risk reductions of the same magnitude, is there evidence of a "cancer
premium"? That is, do individuals systematically prefer avoiding cancer risks and, if so,
by how much? Second, to what extent does cancer latency modify differences in
preferences for the two types of risks?

To address these issues we administered a web-based preference elicitation survey to a
general population sample of adults in the US. The focal point of the survey was a choice
task that asked respondents to choose locations that offered either lower automobile
fatality or cancer risks. The relative risk reductions, as well as the characteristics of the
cancers, were varied randomly across respondents.

The main findings of the survey are that individuals made choices that revealed (1) a
significant cancer premium and (2) a cancer premium that declined with the length of the
cancer latency period. On average, to make individuals indifferent between avoiding the
two types of risks, they required risk reductions for fatal cancers that were two to three
times larger than for fatal automobile risks. Preferences for avoiding cancer risks were
also significantly reduced by longer latency periods; however, the survey results indicate
that latency periods greater than 30 years were generally required to offset the effects of a
cancer premium.

Our analysis also finds that the effect of latency periods on preferences is itself affected
by individuals' perceived survival probabilities. The lower the chance of survival for a
given latency period, the less individuals preferred avoiding cancers with that latency.
Our results indicate that perceived survival probabilities (less than 100 percent) are one
reason that individuals discount future cancer risks; however, this discounting persists to
some extent even after accounting for survival.

Our results suggest that using current estimates of VSL based mainly on data from
accidental death risks may not be appropriate when evaluating the benefits of avoided
cancer risks. Unless cancer latency periods exceed 30 years, these VSL estimates are
likely to understate the true benefits of reduced cancer risks. Further research using
different preference elicitation and measurement approaches is needed to confirm these
findings; however, they provide more evidence that policy analyses would benefit from
VSL estimates that are better tailored to the risk reductions contexts in which they are
applied.

18

-------
References

Cameron, T.A. and J.R. DeShazo. 2004. "An Empirical Model of Demand for Future
Health States when Valuing Risk-Mitigating Programs." University of Oregon,
Economics Department Working Paper 2004-11.
(http://ideas.repec.Org/p/ore/uoecwp/2004-l 1 .html)

Carson, R.T., and R.C. Mitchell. 2000. "Public Preference toward Environmental Risks:
The Case of Trihalomethanes." University of California San Diego, Discussion
Paper.

du Vair, P and J. Loomis. 1993. "Household's Valuation of Alternative Levels of

Hazardous Waste Risk Reductions: An Application of the Referendum Format
Contingent Valuation Method." Journal of Environmental Management. 39(2):
143-155.

Hammitt, J. K. 2000. "Valuing Mortality Risk: Theory and Practice." Environmental
Science and Technology 34(8): 1396-1400.

Hammitt, J. K. 1990. "Risk Perceptions and Food Choice: an Exploratory Analysis of
Organic- versus Conventional-produce Buyers." Risk Analysis, 10(3):367-74.

Hammitt, J. K. and J. Liu. 2004. "Effects of Disease Type and Latency on the Value of
Mortality Risk." Journal of Risk and Uncertainty 28(1): 73-95.

Krupnick, A., A. Alberini, N. Simon, and M. Cropper. 2005. "Willingness to Pay for
Mortality Risk Reductions: Does Latency Matter?" Resources for the Future
Discussion Paper (www.rff.org/documents/rff-dp-04-13.pdf)

Magat, W. A., W. K. Viscusi, and J. Huber. 1996. "A Reference Lottery Metric for
Valuing Health." Management Science 42(8): 1118-1130.

Smith, V. K. and W. H. Desvousges. 1987. "An Empirical Analysis of the Economic
Value of Risk Changes." The Journal of Political Economy, 95(1):89—114.

Subramanian, U., Cropper, M. 2000. "Public choices between life saving programs: The
Tradeoff Between Qualitative Factors and Lives Saved," Journal of Risk and
Uncertainty 21(1): 117-149.

Tsuge, T., A. Kishimoto, and K. Takeuchi. 2005. "A Choice Experiment Approach to the
Valuation of Mortality." Journal of Risk and Uncertainty 31(1): 73-95.

U.S. Environmental Protection Agency (EPA). 2000. "An SAB Report on EPA's White
Paper Valuing the Benefits of Fatal Cancer Risk Reduction. " Letter to the EPA
Administrator from the EPA Science Advisory Board, July 27, 2000. EPA-SAB-
EEAC-00-013.

19

-------
Figure 1. Preference Map for Two Categories of Risk

Number of



Stomach



Cancers per

\

Million

\ \ \



\ \ Current \

100

\ \l_ocation \



LocationX j\ \



A \ ! \ \



Location B T \

1 \

1 \

1



100 Number of



Automobile Deaths



per Million

20

-------
Figure 2. Example Choice Task in the Risk Tradeoff Survey

The table below summarizes the only differences between Location A
and Location B.





Location A

Location B





Car accident
deaths

(per year)

50

per million people

100

per million people





Fatal stomach
cancers

(caused per year)

100

per million people

50

per million people



If you had to move to one of these locations, which one
would you prefer?



Location A
9

Location B
•





No Preference Between Location A and
Location B
•









21

-------
Figure 3. Preferences for Avoiding Fatal Cancer Risks Relative to Auto
Death Risks

~	25 year latency

~	15 year latency
¦ 5 year latency

Percent
Preferring
Cancer Risk
Avoidance

0.43 0.71	1	1.4 2.33

Risk Difference Ratio (auto:cancer)

22

-------
Table 1. Descriptions of Analysis Variables

Variable Name

Description

PREFERB

= 1 if choose "Prefer Location B"

RDR

= Risk difference ratio presented in choice table

LAT15

= 1 if 15-year latency period

LAT25

= 1 if 25 year latency period

LATENCY

= Latency period (5, 15 or 25)

SURVIVERATE

Self-reported (perceived) probability of surviving for duration of latency period



= Interaction between 5-year latency period and perceived chance of survival

LAT5SURV

during latency period (LAT5*SURVRATE)



= Interaction between 15-year latency period and perceived chance of survival

LAT15SURV

during latency period (LAT15*SURVRATE)



= Interaction between 25-year latency period and perceived chance of survival

LAT25SURV

during latency period (LAT25*SURVRATE)

M0RB5

= 1 if 5-year morbidity period (= 0 if 2-year morbidity period)

INTROFORMAT

= 1 if Location A dominates Location B in introductory choice questions

STOMACHC

= 1 if stomach cancer version

BRAINC

= 1 if brain cancer version

AGE

= respondent's age

HEALTHNOW

= Self assessment of respondent's current health status

GENDER

= 1 if male

MINORITY

= 1 if race non-white

EDUC

Number of years of education

HHINCOME

Household income ($'000)

HHSIZE

Household size

MSA

= 1 if respondent lives in an MSA

CANCERYOU

= 1 if respondent ever had cancer

CANCFRIEND

= 1 if friend or relative had experienced the cancer described in the survey

CANCERCURE

= 1 if considered possibility of cure for cancer during latency period

CANCERRISK

Self-rated fatal cancer risk compared to average (1= much lower, 5 = much higher)

CARYOU

= 1 if hospitalized because of a car accident

CARFRIEND

= 1 if friend or relative died in car accident in last 10 years

CARRISK

Self-rated fatal car risk compared to average (1= much lower, 5 = much higher)

23

-------
Table 2. Summary Statistics for Analysis Variables

Variable

N

Mean

SD

Min

p25

p50

p75

Max

PREFERB

788

0.65

0.48

0

0

1

1

1

RDR

788

1.16

0.65

0.43

0.71

1

1.4

2.33

LAT15

788

0.32

0.47

0

0

0

1

1

LAT25

788

0.36

0.48

0

0

0

1

1

LATENCY

788

15.36

8.21

5

5

15

25

25

SURVIVERATE

788

0.77

0.27

0

0.6

0.9

1

1

M0RB5

788

0.50

0.50

0

0

1

1

1

INTROFORMAT

788

0.49

0.50

0

0

0

1

1

STOMACHC

788

0.34

0.47

0

0

0

1

1

BRAINC

788

0.33

0.47

0

0

0

1

1

AGE

788

45.39

16.97

18

31

44

58

93

HEALTHNOW

787

2.54

0.92

1

2

3

3

5

GENDER

788

0.48

0.50

0

0

0

1

1

MINORITY

788

0.19

0.40

0

0

0

0

1

EDUC

788

12.45

3.22

6

12

14

14

16

HHINCOME

788

50.56

36.49

2.5

22.5

45

67.5

187.5

HHSIZE

788

2.61

1.24

1

2

2

3

8

MSA

788

0.84

0.37

0

1

1

1

1

CANCERYOU

788

0.08

0.27

0

0

0

0

1

CANCFRIEND

782

0.12

0.33

0

0

0

0

1

CANCERCURE

783

0.25

0.43

0

0

0

0

1

CANCERRISK

782

2.56

0.84

1

2

3

3

5

CARYOU

788

0.18

0.38

0

0

0

0

1

CARFRIEND

785

0.17

0.38

0

0

0

0

1

CARRISK

783

2.47

0.95

1

2

3

3

5

24

-------
Table 3. Analysis of Risk Tradeoffs: Probit Results for Location Choice

Dependent Variable = PREFERB





(1)

(2)

(3)



(4)



(5)

Variable

Coef.

z-

statistics

Coef.

z-

statistics

Coef.

z-

statistics

Coef.

z-

statistics

Coef.

z-

statistics

CONSTANT

0.816

8.56

1.262

8.28

1.788

4.01

1.947

4.31

0.302

0.57

RDR

-0.355

-5.03

-0.350

-4.89

-0.374

-4.92

-0.373

-4.91

-0.382

-4.98

LAT15





-0.316

-2.60

-0.300

-2.34









LAT25





-0.634

-5.39

-0.632

-5.06









LATENCY













-0.032

-5.10





LAT5SURV

















1.239

4.99

LAT15SURV

















1.061

4.09

LAT25SURV

















0.710

2.57

MORB5





0.031

0.33

-0.042

-0.43

-0.042

-0.42

-0.039

-0.39

INTROFORMAT





-0.173

-1.83

-0.207

-2.10

-0.208

-2.11

-0.234

-2.36

STOMACHC





0.111

0.95

0.096

0.78

0.097

0.80

0.071

0.57

BRA INC





-0.215

-1.86

-0.235

-1.97

-0.235

-1.96

-0.253

-2.10

AGE









-0.005

-1.57

-0.005

-1.56

0.002

0.53

HEALTHNOW









-0.046

-0.82

-0.046

-0.80

0.004

0.07

GENDER









-0.053

-0.54

-0.053

-0.54

-0.022

-0.22

MINORITY









-0.100

-0.79

-0.101

-0.79

-0.106

-0.82

EDUC









-0.008

-0.47

-0.008

-0.47

-0.013

-0.83

HHINCOME









-0.002

-1.79

-0.002

-1.79

-0.003

-2.15

HHSIZE









0.006

0.14

0.006

0.14

0.008

0.20

MSA









0.251

1.89

0.250

1.89

0.238

1.83

CANCERYOU









-0.296

-1.49

-0.296

-1.48

-0.192

-0.94

CANCFRIEND









0.249

1.53

0.250

1.55

0.268

1.67

CANCERCURE









-0.401

-3.60

-0.400

-3.58

-0.409

-3.63

CANCERRISK









0.239

3.73

0.240

3.75

0.240

3.71

CARYOU









-0.027

-0.21

-0.027

-0.21

-0.033

-0.26

CARFRIEND









-0.170

-1.27

-0.170

-1.28

-0.152

-1.15

CARRRISK









-0.207

-3.69

-0.206

-3.69

-0.189

-3.33

Number of obs

788



788



775



775



775



Calculated Values





















E[MER]

2.30



















E[MER(0)]













3.67







E[MER(5)]





3.31



3.23



3.24







E[MER(15)]





2.41



2.43



2.39







E[MER(25)]





1.50



1.54



1.54







t*













31.40







25

-------
Appendix: Proofs of latency and survival probability effects on MER

A. 1 The effect of cancer latency on MER

As shown in equation (5), the effect on MER of increasing latency (t) depends on how it
affects the lifetime utility of the cancer profile. Expanding on equation (6), the effect on
U(C(t), Y) of increasing latency by one period, from t to t+1, can be written as

U(C(t + l)J)-U(C(t)J)

=f S ){d<)u h 0>.-)] + (st )(dt)" \yt) + (sM )(dM)"c (yt+1)

r"	a	(A1)

- Z fa )(<*,> *(>,¦) "(st)(dt)uC(yt)

V i=1	J

(st)(dt )\u\yl) - uc(yt)]+ (X+1 )(dM)uc(yM )

For simplicity, the duration of each time period, as indexed by i, is the same as the
duration of cancer morbidity (i.e., between diagnosis and death). The first term in
equation A.l will be positive as long as the utility of a period in normal health, uh(y), is
greater than with cancer, uc( y). Even if the utility of a period (t or t+1) with cancer is
negative, the second term in this expression will also be less in absolute value terms than
the first, as long as uh(y) is positive (and yt+i is not substantially less than yt).
Consequently, an increase in latency should increase U(C(t),Y) and decrease MER. This
result is consistent with the intuition that extending cancer latency will reduce aversion to
cancer risks.

A.2 The effect of survival probability on MER

In contrast to a change in cancer latency, an increase in survival probability, st, will affect
MER through both the cancer and the normal health utility profiles. To examine the
effect of increasing st on the MER, we must examine its effect on the ratio
U(C.Y)/U(H.Y). To do this we first define the following expressions:

2 = 0

B = (dt)u°(yt)

00	(A.2)

c = S(^Xd>*0/)

;/(C(o,r)=i+^

t/(//,7) A+s,C

26

-------
Sij = probability of surviving to period j conditional on surviving to period t
Differentiating the lifetime utility ratio R with respect to st, we get:

— = A(B~C"> < o	if A>0 and B
-------
P4

w

April 2006 ¦ RFF DP 06-19

CT)
CT)

u

en

Q

Adjusting the Value
of a Statistical Life for

Age and Cohort
Effects

O 	

Joseph E. Aldy and W. Kip Viscusi

1616 P St. NW
Washington, DC 20036
202-328-5000 www.rff.org

FOR THE FUTURE

-------
Adjusting the Value of a Statistical Life for Age and Cohort Effects

Joseph E. Aldy and W. Kip Viscusi

Abstract

To resolve the theoretical ambiguity in the effect of age on the value of statistical life (VSL), this
article uses a novel, age-dependent fatal risk measure to estimate age-specific hedonic wage regressions.
VSL exhibits an inverted-U shaped relationship with age. In the year 2000 cross-section, workers' VSL
rises from $3.2 million (ages 18-24), to $9.9 million (35-44), and declines to $3.8 million (55-62).
Controlling for birth-year cohort effects in a minimum distance estimator yields a peak VSL of $7.8
million at age 46 and flattens the VSL-age relationship. The value of statistical life-year also follows an
inverted-U shape with age.

Key Words: value of statistical life, job risks, hedonic wage regression, VSLY
JEL Classification Numbers: J17,112

© 2006 Resources for the Future. All rights reserved. No portion of this paper may be reproduced without
permission of the authors.

Discussion papers are research materials circulated by their authors for purposes of information and discussion.
They have not necessarily undergone formal peer review.

-------
Contents

I.	Wage-Risk Tradeoffs over the Life Cycle	4

II.	Hedonic Wage Methods and Results	7

A.	Data	7

B.	Hedonic Wage Regression Framework	8

C.	Estimated Age Group VSLs	10

D.	Minimum Distance Estimator and Cohort Effects	11

IV.	Implications for the Value of a Statistical Life-Year	14

V.	Conclusion	15

References	17

Tables and Figures	19

-------
Resources for the Future

Aldy and Viscusi

Adjusting the Value of a Statistical Life for Age and Cohort Effects

Joseph E. Aldy and W. Kip Viscusi*

A strident controversy with respect to the value of life has been whether the benefit of
reducing risks to the old are less than for younger age groups. In particular, should there be a so-
called "senior discount" when assessing the value of reduced risks to life? This question has
drawn the attention of policymakers in a number of countries. In 2000, Canada employed a value
of statistical life (VSL) for the over-65 population that is 25 percent lower than the VSL for the
under-65 population (Hara and Associates 2000). In 2001, the European Commission
recommended that member countries use a VSL that declines with age (European Commission
2001). In 2003, the U.S. Environmental Protection Agency (EPA), which has traditionally
employed a constant value of a statistical life to monetize mortality risk reductions irrespective
of the age of the affected population, conducted analyses of the Clear Skies initiative that
included a "senior discount."1 This effort to apply such a discount in its Clear Skies initiative
analyses generated a political firestorm and ultimately led to abandonment of any age
adjustments in benefit values assigned by the Agency.2

Intuitively one might expect that older individuals may value reducing risks to their lives
less because they have shorter remaining life expectancy. The commodity they are buying
through risk reduction efforts is less than for younger people. Carrying this logic to its extreme,

* Aldy: Resources for the Future, 1616 P Street NW, Washington, DC 20036 (email: aldv@rff.org'): Viscusi:
Harvard Law School, 1575 Massachusetts Avenue, Cambridge, MA 02138 (e-mail: kip@law.harvard.edu'). Aldy's
research is supported by the U.S. Environmental Protection Agency STAR Fellowship program and the Switzer
Environmental Fellowship program. Viscusi's research is supported by the Harvard Olin Center for Law,
Economics, and Business. The authors express gratitude to the U. S. Bureau of Labor Statistics for permission to use
the CFOI fatality data. Neither the BLS nor any other government agency bears any responsibility for the risk
measures calculated or the results in this paper. Antoine Bommier, Gardner Brown, David Cutler, Bryan Graham,
Caroline Hoxby, Seamus Smyth, and James Ziliak and seminar participants at Harvard University and the AERE
Summer Workshop provided constructive comments.

1	In the "senior discount" analyses, the EPA provided two alternatives to account for age. One approach was based
on a standard value of a statistical life-year approach that explicitly accounts for life expectancy. The second
approach assumed that individuals over age 70 had a value of statistical life equal to 63 percent of the value for
those under 70.

2	For a sense of the political reaction and EPA's decision to discontinue the use of an age-based value of statistical
life, refer to "EPA Drops Age-Based Cost Studies," New York Times, May 8, 2003; "EPA to Stop 'Death Discount'
to Value New Regulations," Wall Street Journal, May 8, 2003; and "Under Fire, EPA Drops the 'Senior Death
Discount,'" Washington Post, May 13, 2003.

1

-------
Resources for the Future

Aldy and Viscusi

the VSL would peak at birth and decline steadily thereafter. For models in which consumption is
constant over the life cycle, Jones-Lee (1989) showed that the VSL should decrease with age.
Whether consumption will in fact be constant over time depends critically on the presence of
perfect capital and insurance markets.

Numerous theoretical studies have shown that the age variation in VSL becomes more
complex once changes in consumption over time are introduced into the analysis. Changes in
consumption levels and wealth over the life cycle influence risk-money tradeoffs in a complex
manner. Johansson (2002) concluded that the theoretical relationship between the VSL and age is
ambiguous and could be positive, negative, or zero. Often theoretical studies, however, have
imposed additional structure on the analysis, implying that there is either an inverted U-shaped
relationship between the value of statistical life and age or that VSL decreases with age. The
simulations by Shepard and Zeckhauser (1984) show a steadily declining value of life if there are
perfect annuity and insurance markets, and an inverted-U VSL-age relationship in an economy
with no borrowing or insurance, as do Johansson (1996) and Ehrlich and Yin (2004). Rosen
(1988), Arthur (1981), and Cropper and Sussman (1988) also present simulation results with
VSL decreasing with age.

Empirical evidence based on labor market data may be instructive in resolving the
theoretical ambiguity in the VSL-age relationship. Viscusi and Aldy (2003) review eight studies
of labor markets in Canada, India, Switzerland, and the United States that included an age-
mortality risk interaction term in their hedonic wage analysis. Five studies estimated statistically
significant coefficient estimates on the age-risk interaction and all find a negative effect
indicating that older workers value risks to their lives less.3 These results imply implausibly low
VSL levels with negative VSL amounts beginning at ages ranging from 42 to 60. The failure of
labor market evidence to resolve the age variation issue may stem in part from data limitations.
All these labor market studies use fatality risk data that are based on industry averages rather
than age-specific values, causing potential biases, where the magnitude of the bias varies with
age. If, for example, average industry fatality risks for workers of all ages overstate the risks
faced by older workers, the estimated implied VSL amounts for older workers will understate the
wage-risk tradeoffs that are actually being made.

3 These studies are reviewed in Section 8 of Viscusi and Aldy (2003). In contrast, a recent study by Smith et al.
(2004) has found that the value of statistical life is increasing with age and risk aversion for workers 51-65 years of
age.

2

-------
Resources for the Future

Aldy and Viscusi

All previous papers assessing how the compensating differential for job mortality risk
varies with age have employed cross-sectional survey data. By using a single cross-section, such
approaches confound the cohort-specific influence and age-specific effects on the estimated
compensating differential. The cohort influence based on the year of birth should have an
unambiguous effect on VSL. Lifetime incomes are rising over time, and the VSL has a positive
income elasticity of 0.5 to 0.6.4 Because older workers belong to an earlier cohort with lower
lifetime incomes, they will tend to be willing to pay less for a given risk reduction, implying a
lower VSL. The pure age effect is less clear-cut. As a worker ages, there are fewer years of
remaining life expectancy, implying lower benefits for a given risk reduction, which should
reduce the worker's willingness to pay to reduce risk. This effect is unambiguous if capital
markets are perfect. In a world with imperfect capital markets, however, lower income younger
workers will not be able to borrow against higher future expected earnings. This will depress
their VSLs at young ages until borrowing constraints become less stringent, resulting in an age-
related VSL trajectory similar to the inverted-U shape of life-cycle consumption patterns.
Extending the traditional analysis to a pooled series of cross-sections will enable us to
distinguish age effects from cohort effects. Two separate, but both policy-relevant, questions can
then be considered: (1) How does the value of life vary with age across the population? and (2)
How do differences in cohorts influence this relationship?

This article extends the previous literature in several respects. Because our focus is on
risky labor market decisions, we make job risk decisions a choice variable in a life-cycle
consumption model in Section I, deriving an expression for VSL in this context. In Section II, we
present empirical estimates how the VSL varies over the life cycle through conventional hedonic
wage equations and a minimum distance estimator. These results reflect two innovations to this
literature: (1) we employ age-specific job mortality and nonfatal injury risks in our hedonic wage
analyses; and, (2) we estimate how the VSL changes over the life cycle by pooling eight years of
cross-sectional data and by using a minimum distance estimator that controls for cohort effects
based on year of birth. In these empirical approaches, the VSL rises and then falls across the
population and over the life cycle. In the cross-sectional analysis, the VSL peaks at age 39 and
subsequently declines so that the VSL for workers in their early 60s have values of about $2
million. In the cohort-adjusted analysis, the VSL peaks at age 46, and experiences a more modest

4 See Viscusi and Aldy (2003) for a meta-analysis of the VSL income elasticity value.

3

-------
Resources for the Future

Aldy and Viscusi

decline to about $5 million by age 62.5 In Section III, we calculate age-specific values of
statistical life-years (VSLY) from our age-VSL profiles and find that VSLYs also take an
inverted-U shape with a peak at an older age than the VSLs. In the cross-sectional analysis, the
VSLY peaks at $375,000 at age 45 and subsequently declines to about $150,000 in workers'
early 60s. In the cohort-adjusted analysis, the VSLY peaks at $401,000 at age 54, and
experiences a more modest decline to about $350,000 by age 62. Section IV concludes the paper.

I. Wage-Risk Tradeoffs over the Life Cycle

The standard approach in the life-cycle VSL literature employs a time-separable utility
function in one consumption good, integrated over the life-cycle subject to a discount function
and a survival function, as in Shepard and Zeckhauser (1984), Rosen (1988), Johansson (1996,
2002), and Johannesson et al. (1997). The only choice variable is the level of consumption over
time. In these analyses, the value of statistical life is given by a representative agent's expected
present value of consumer surplus conditional on having achieved a given age. For example,
Shepard and Zeckhauser represent this as the ratio of expected remaining lifetime utility to the
marginal utility of consumption.

To motivate our empirical work, we provide a model of wage-risk tradeoffs in a life-
cycle setting. We modify and extend the standard life-cycle approach to explicitly account for
the choice of job fatality risk on the survival function and the worker's wage. Since a change in
job fatality risk affects both the worker's wage and life expectancy, our approach provides an
alternative illustration of the VSL varies over the life cycle by characterizing the wage-risk
tradeoff given the impacts of both on future consumption. By incorporating a compensating
differential framework in our model, we can demonstrate how the wage-risk trade-off varies over
the life cycle, which is what we will estimate in our empirical work presented below.

Our simple model indicates variations in VSL, but the linkage is ambiguous. This life-
cycle model can illustrate the influences - especially the life-cycle variation in consumption -
that can generate an inverted U-shaped relationship between VSL and age. The worker's
problem can be characterized by maximizing discounted expected remaining lifetime utility:

CO

(1)	ma xEU(t)= \u[c(t)]a[t;z,p(t)]ertdt,

p,C	J

5 All VSL estimates are presented in year 2000 dollars in this paper.

4

-------
Resources for the Future

Aldy and Viscusi

subject to

(2a)	k(t) = rk(t) + w[t, p(t)] - c(t) + f(t),

(2b)	k(t) > 0,

and

(2c)	Y\mk(t)ert = 0,

» co

where

p represents the probability of dying on the job,

u(c) represents the utility of consumption, c, and u'(c) > 0, u"(c) < 0,

k represents assets,

w represents labor income,

e~rt represents the discount function,

o\t\ r, /;(/)] represents the survival function, i.e., the probability of surviving to age t,
given that the individual has reached age r ,6

r represents the return on assets, and

/(t) represents the net amount received through an actuarially fair annuity represented
by the condition:

co

Je~rtcr|7;0, p(t)\f(t)dt = 0 7

0

The worker's expected utility is represented in (1) as the sum of period utilities weighted
by a discount factor and the probability that the worker will survive to that period conditional on
the worker's current age. The worker maximizes this expected utility expression subject to the
constraints: (2a) represents the dynamic budget constraint, and it allows for the worker's assets

6	This expression of the survival function follows Johansson (1996): cr[/; r, p(i)\ = cr[/; p(/)\ / cr(r).

7	To simplify notation, we have followed Shepard and Zeckhauser and assumed that the rate of time preference in
the discount function is equal to the rate of return on assets, and that this rate is time-invariant. Allowing for the rate
of time preference to differ from the return on assets would not substantively influence the primary conclusion of
this analysis that the age-VSL relationship is ambiguous.

5

-------
Resources for the Future	Aldy and Viscusi

to change over time based on capital income rk(t), labor income w\t, pit)], consumption c(t),
and net annuity receipts /(t); (2b) provides a no debt condition; and (2c) is the standard no
Ponzi game condition. The actuarially fair annuity envisioned here is similar to that in Shepard
and Zeckhauser's (1984) perfect markets case, and the annuity allows for the worker to borrow
against human capital during early years of life to provide for consumption smoothing.

The present value Hamiltonian, conditional on having lived to age x, is given by:

(3)	H(t) = u[cit)\c>[t; r, p(t)]ert + A it)[rkit) + w[t, pit)] - c(t) + fit)]

where Ail) represents the present value costate variable. The first-order conditions for
the Hamiltonian are:

(4)	^- = ucae~rt -/t = 0,8
oc

(5)	^r- = uc>pe'rt +Awp= 0,

dp

and

(6)	-— = X^X = -rA.

dk

To see more generally how the value of a statistical life varies with age, we rearrange (5),
differentiate with respect to time, where time derivatives are denoted by a dot over the variables
in question, and substitute into (6), yielding:

W r, U & v

(7)	-JL = H + ^L
wp u a p

The percentage change over time in the compensating differential for job fatality risk is
equal to the percentage change over time in utility and the percentage change over time in the
change in the survival function with respect to job fatality risk.9 This expression holds
irrespective of the assumption of actuarially fair annuity markets, although the assumption
regarding these markets clearly influences the change in utility over the life cycle. The sign on

8	This is essentially identical to equation 12 of Shepard and Zeckhauser (1984).

9	Note that the survival function, cr[/; r, pit)], and the discount function, e " , implicitly enter equation (7)
through their influence on the optimal consumption and job fatality risk paths.

6

-------
Resources for the Future

Aldy and Viscusi

equation (7) is ambiguous without imposing restrictions on the survival function and specifying
the assumptions regarding annuity markets. This ambiguity is consistent with the life-cycle
model provided by Johansson (2002) and the simulation results based on the life-cycle model in
Shepard and Zeckhauser (1984). This theoretical ambiguity motivates our interest in resolving
empirically how the value of a statistical life varies over the life cycle.

II. Hedonic Wage Methods and Results

To assess empirically the age-VSL relationship, we have expanded the standard hedonic
wage framework in two ways. First, using our new and more refined age-specific job-related
mortality and injury data, we estimated hedonic wage regressions that allow for the
compensating differential for these risks to vary among five age groups. These results indicate
how the VSL varies with age across the population. Second, we develop a minimum distance
estimator that incorporates age-specific hedonic wage regressions in the first stage and controls
for cohort effects in the second stage. This analysis, based on eight years of pooled cross-
sections, indicates how the value of life varies with an individual's age.

A. Data

To characterize the fatality risks faced by workers of different ages more precisely than is
possible using average risk values by industry, we constructed a novel risk measure conditional
upon age and the worker's industry rather than using an industry basis alone, which is the norm
for all previous studies of age variations in workers' VSL. The source of the fatality measures is
the Bureau of Labor Statistics (BLS) Census of Fatal Occupational Injuries (CFOI), for the 1992-
2000 period. We structured the mortality risk cells by 2-digit SIC industries and these six age
groups specified in the CFOI data: 16-19, 20-24, 25-34, 35-44, 45-54, and 55-64. To construct
the denominator for the mortality risk variable, we used the 1992-2000 Current Population
Survey Merged Outgoing Rotation Group files to estimate worker populations for each cell in the
mortality data. The annual mortality risk measures are averaged to minimize any potential
distortions associated with catastrophic mortality incidents in any one year and to have a better
measure of the underlying risks for industry-age groups with infrequent deaths. Our injury risk
measure, the probability of a lost-workday injury, also varies by age, and we constructed it in an
identical manner for each 2-digit industry and for each of the age groups listed above. While

7

-------
Resources for the Future

Aldy and Viscusi

injury risk decreases with age across most industries, mortality risk increases monotonically with
age in all industries, except for in mining.10

We have matched these constructed mortality risk and injury risk measures by age and
industry with data on adult workers in the Current Population Survey Merged Outgoing Rotation
Group data files for 1993-2000. We employed a number of screens in constructing our sample
for analysis. The sample excludes agricultural workers and members of the armed forces. We
have excluded workers younger than 18 and older than 62, those with less than a 9th grade
education, workers with an effective hourly labor income less than the minimum wage, and less
than full-time workers, which we defined as those working at least 35 hours per week.

B. Hedonic Wage Regression Framework

The standard hedonic wage model estimates the locus of tangencies between the market
offer curve and workers' highest constant expected utility loci. The age variation in the wage-
mortality risk tradeoff simultaneously reflects age-related differences in preferences as well as
age-related differences in the market offer curve. If older workers are more likely to be seriously
injured than are younger workers because of age-related differences in safety-related
productivity, then the market offer curve will reflect that, given that age is a readily monitorable
attribute. Because workers' constant expected utility loci and firms' offer curves each may vary
with age, there is no single hedonic market equilibrium. Rather, workers of different ages will
settle into distinct market equilibria as workers of different ages select points along the market
opportunities locus that is pertinent to their age group.11

Conventional hedonic wage analyses of job risks specify the natural logarithm of the
hourly wage or some comparable income measure as a function of worker and job
characteristics, mortality risk, and, in more comprehensive specifications, injury risk and a
measure of workers' compensation. Our base specification takes the following form:

(8)	In(wt ) = oc + H\P + YxPx + r2Vi + r^WCr + *,¦,

where

10	Refer to Aldy and Viscusi (2004) for more details about the construction of this age-specific job mortality risk
measure.

11	This analysis generalizes the hedonic model analysis for heterogeneous worker groups using the model developed
for an evaluation of smokers and nonsmokers by Viscusi and Hersch (2001). Their worker groups differ in their
safety-related productivity and in their attitudes toward risk.

8

-------
Resources for the Future

Aldy and Viscusi

w'( is the worker i's hourly after-tax wage rate,

H is a vector of personal characteristic variables for worker i,
pi is the fatality risk associated with worker i's job,

qi is the nonfatal injury risk associated with worker i's job,

WCi is worker i's compensation replacement rate for a job injury, and

si is the random error reflecting unmeasured factors influencing worker i's wage rate.

We calculated the workers' compensation replacement rate on an individual worker basis
taking into account state differences in benefits and the favorable tax status of these benefits. We
use the benefit formulas for temporary total disability, which comprise about three-fourths of all
claims, and have formulas similar to those for permanent partial disability.12 The terms a, P, yi,
72, and 73 represent parameters to be estimated.

All wage regression specifications used in this paper include the following controls:
demographic indicator variables (race and ethnicity, gender of head of household, marital status,
union membership, public sector employment, and resident of urban area); educational
attainment; indicator variables for one-digit occupation and region of residence; and job
mortality risk, job nonfatal injury risk, and expected workers' compensation replacement rate.13

The estimated regression then yields a measure of the average value of a statistical life
for the sample:

(9)	VSL = yl *w* 2,000 * 100,000 .

This equation normalizes the VSL to an annual basis by the assumption of a 2,000-hour
work-year and by accounting for the units of the mortality risk variable. As a preliminary check
on our age-industry risk variables, we estimated equation (8) with the 1997 CPS MORG and
compared this with the results for industry risk variables merged with the 1997 CPS MORG

12	The procedures for calculating the workers' compensation benefit variable are discussed in more detail in Viscusi
(2004), which also provides supporting references.

13	The workers' compensation expected replacement rate represents the interaction of a worker's injury rate and that
worker's estimated workers' compensation wage replacement rate based on the worker's wage, state of residence,
state benefit formulas, and estimated state and federal tax rates. Given the endogeneity of the wage, we have also
estimated instrumental variables regressions. IV estimation does not qualitatively influence determinations of
coefficient magnitudes or statistical significance for the mortality risk variable of interest in this study. Refer to Aldy
and Viscusi (2004) for additional details.

9

-------
Resources for the Future

Aldy and Viscusi

dataset presented in Viscusi (2004). We estimated a mean VSL of $4.5 million (1997$), which is
virtually indistinguishable from the Viscusi (2004) estimate of $4.7 million, and both studies fall
within the range of VSLs from hedonic wage regression studies of the U.S. labor market reported
in Viscusi and Aldy (2003).14

C. Estimated Age Group VSLs

As an initial assessment of how the value of life varies with age across the population, we
modified (8) so that the estimated compensating differentials can vary by age. We interacted five
age group indicator variables - for age groups 18-24, 25-34, 35-44, 45-54, and 55-62 - with
the various risk measures, and included the first four age group indicator variables in the
specification:

4	5	5	5

(8a) \n(wi) = a + H\(5 + Z 5jaSej + Hri iageiPi +Z^2 jage}-qi +Z^3 iagei
-------
Resources for the Future

Aldy and Viscusi

regressions reveal similar patterns of the VSL with respect to age: an inverted-U shape with the
VSL peaking for the 35-44 age group in six of the eight years. As an illustration, consider the
results for the year 2000 cross-section. The coefficient estimate on the 18-24 age group mortality
risk variable is 0.0021, and it increases substantially to 0.0039 for the 25-34 age group. The
mortality risk coefficient then declines with age: 0.0036 for the 35-44 age group, 0.0028 for the
45-54 age group, and 0.0014 for the 55-62 age group. The five age-group-specific job mortality
risk coefficient estimates are individually statistically significant at the 1 percent or 5 percent
level. The estimated VSLs for each age group depend on these coefficient estimates as well as
age-group-specific average wages, which follow an inverted-U shape over the life cycle. The 35-
44 age group has the largest VSL of $9.85 million, more than triple the 18-24 VSL of $3.16
million and nearly triple that of the 55-62 VSL of $3.77 million.17

To show how these differences in magnitudes are often statistically significant, we focus
on the results for the year 2000 cross-section, which we report again at the top of Table 2. We
conducted a series of pairwise Wald tests on the estimated VSLs, and the table presents the F-
statistics associated with these tests. The first row of these tests shows that the 18-24 VSL of
$3.16 million is statistically different from the VSL estimates for the next three age groups, but
does not differ significantly from the 55-62 VSL of $3.77 million. The last column,
corresponding to the 55-62 age group, shows that the estimated 55-62 VSL differs significantly
from the VSL estimates for the 25-34 age group, the 35-44 age group, and the 45-54 age group.
These results indicate that the VSL takes an inverted-U with respect to age across a population.
The VSL pattern is relatively flat in the middle age groups as there is no statistically significant
difference among the age 25-34, 35-44, and 45-54 categories for the 2000 cross-section.

D. Minimum Distance Estimator and Cohort Effects

We have extended this age-specific regression analysis in subsection C through a two-
stage minimum distance estimator using VSL estimates for each year rather than age bands. This
approach allows us to infer information about the VSL with respect to age based on a larger
number of regressions based on more narrowly defined age bands for each year. While these
individual regressions will provide less precise estimates of the compensating differential for risk

17 Refer to Jones-Lee et al. (1985) for an example of a stated willingness to pay for safety study that also finds an
inverted-U shaped VSL-age relationship.

11

-------
Resources for the Future

Aldy and Viscusi

than broader age groups, it will then be possible to estimate VSLs as a function of age if age-
specific VSLs follow a systematic pattern over the life cycle.

In the first stage, we estimate age-specific hedonic wage regressions of the form
expressed in equation 8 and use the mortality risk coefficient estimates to construct age-specific
VSL. We estimated age-specific compensating differentials for 45 age levels from age 18 to 62
and eight cross-sections from 1993-2000, yielding 360 separate regressions. With the exception
of the youngest and oldest birth-year cohorts, every cohort has eight observations in our
constructed panel.18 We estimated the VSL using the mean real wage for that respective age and
year. Based on these first stage regressions, we construct a panel of cohort-specific and age-
specific VSL estimates. Each VSL estimate is assigned to a birth-year cohort. For example, the
estimated VSL for a 40-year old in 1993 is assigned to the 1953 birth-year cohort; the estimated
VSL for a 41-year old in 1994 is also assigned to the 1953 birth-year cohort, and so on. We
followed this procedure for all 360 VSL estimates.

In the second stage, we specify these VSLs by age. To characterize how the VSL
estimates from the first stage, VSL, vary with age across a population, the second stage includes
a polynomial in age, a{9). To characterize how the VSL varies over the life cycle, we account

for the differences across cohorts by including a vector of birth-year indicator variables, c, in
addition to the age polynomial. We also employ V, the inverse of a diagonal matrix of the
variance estimates of these VSLs, as a weight matrix based on Chamberlain's (1984) analysis of
the minimum distance estimator and the choice of the inverse of the variance-covariance matrix
as the optimal weight matrix.19'20

18	Refer to Deaton (1985) and Deaton and Paxson (1994) for the advantages of such a constructed panel based on
birth-year cohorts.

19	Because of the potential small sample bias in the optimal minimum distance estimator, we also evaluated the
equally weighted minimum distance estimator (Altonji and Segal 1996). To address concerns about the small sample
bias, we have presented the results for the equally weighted minimum distance estimator in Figures 1 and 2. The
choice of weight matrix has no qualitative impact on our conclusions.

20	We have employed a test of overidentifying restrictions to assess the appropriate order of the polynomial in age. If
we assume that 6 is a Kxl vector, then a restricted parameter vector, a , which is Rxl where R
-------
Resources for the Future

Aldy and Viscusi

For the cross-sectional analysis, the minimum distance estimator solves:

(10)	min[VSL - a{e)]'[V]1 [VSL - a{0)\.

<9e©

For the life-cycle (cohort-adjusted) analysis, the minimum distance estimator solves:

(11)	min [iVSL-a{e)-c'5][VYl[VSL-a{e)-c'5].

0e&,SeA

where 6 and 8 represent parameters to be estimated. We specified a(0) in a variety of
analyses as a polynomial in age of order one to order eight.

The solid curve in Figure 1 presents the fitted age-VSL functions based on a third-order
polynomial in age specification (cross-section VSL), while the dashed line presents the
relationship based on a third-order polynomial in age with birth-year cohort indicator variables
(cohort-adjusted VSL).21 In the pooled cross-sections, the value of statistical life increases with
age from age 18 with a VSL of $4.87 million through age 39, at which the VSL peaks at $8.27
million. The value of a statistical life then declines with age to a minimum of $1.67 million at the
highest age in the sample, which is 62. The cohort-adjusted function, also yields a VSL that
follows an inverted-U shape over the life cycle. It starts at $3.39 million at age 18, peaks at $7.79
million at age 46, and then declines to $5.09 million at age 62. Across the population and along
the life cycle, the value of statistical life increases, peaks, and then decreases with age. While not
presented, the birth-year indicator variables follow a general trend of increasing values with year
of birth, consistent with the proposition that the value of life has increased with temporal
increase in lifetime income.

The cohort adjustment affects the age-related pattern of VSLs in several ways. The peak
of the age-VSL curve is seven years later when accounting for date of birth. The high VSLs for
younger age groups is due in part to their higher lifetime wealth, as their cross-section VSLs lie
above those in the cohort-adjusted values. For older age groups the pattern is reversed. While
there is a steep drop in VSL levels with age in the cross-section results, this decline is due in part
to cohort effects. Accounting for cohort differences attributable to changes in lifetime income
more than doubles the estimated VSLs for the older age groups and flattens their VSL trajectory.
Finally, the counter-clockwise pivoting of the VSL function from the cross-sectional analysis to

21 Based on the specification test presented in footnote 17, we could not reject the hypothesis that a third-order age
polynomial fit the data as well as higher-ordered polynomials. We could, however, reject the hypothesis that lower-
ordered polynomials fit the data as well as a third-order polynomial.

13

-------
Resources for the Future

Aldy and Viscusi

the cohort-adjusted analysis also illustrates the importance of accounting for lifetime income,
implicitly through the birth-year indicator variables, in estimating the age-VSL relationship over
the life cycle.

We also tested two economic propositions that are prominent in current policy
applications of the value of life. First, many analyses assume that the VSL remains constant,
irrespective of age.22 To assess this proposition, we employed our cohort-based minimum
distance estimator and specified the age polynomial function as a constant. We then tested this
restriction versus the more flexible, higher-ordered polynomials and we reject the hypothesis that
the VSL is constant over the workers' life cycle at the 1 percent level in comparison with all age
polynomials of order two or higher. Second, other analyses have assumed that the value of a
statistical life is always decreasing with age.23 To test this proposition, we specified the age
polynomial function as linear, but such an approach yielded a negative coefficient estimate that
clearly could not be distinguished from zero. The test of overidentifying restrictions rejected the
linear specification in comparison to all higher-ordered polynomials. It should also be noted that
all order two through order eight polynomials resulted in similar inverted U-shaped relationships
between the value of a statistical life and age.24

IV. Implications for the Value of a Statistical Life-Year

The preceding section illustrates the estimated age-VSL profile consistent with the theory
model presented in Section I and with previous simulations published in the literature. The
implicit assumptions underlying the value of a statistical life-year (VSLY) approach, which
requires the value of life to be decreasing with age at all ages, are rejected by our data. In light of
the common application of VSLYs in evaluations of medical interventions and government
regulations, such as those promulgated by the U.S. Food and Drug Administration and the U.S.
Environmental Protection Agency in their sensitivity analyses, we have estimated age-specific
VSLYs based on our age-specific VSLs.

22	For example, most U.S. Environmental Protection Agency benefit-cost analyses, including September 2003
revisions to its assessment of the Clear Skies initiative, make this assumption.

23	For example, this is consistent with the European Commission's proposed position and the life-year approach
used by the U.S. Food and Drug Administration.

24	We also evaluated whether the higher VSLs for individuals in the 25-44 age range reflect major life-cycle events
such as marriage or having children, and not variations in age, but find no evidence to support this notion. Refer to
Aldy and Viscusi (2004) for more details.

14

-------
Resources for the Future

Aldy and Viscusi

To construct values of statistical life-years, we have annuitized age-specific VSLs based
on age-specific years of life expectancy L and an assumed discount rate r of 3 percent:25

(12)	VSLY = —rVSL .

l-(l + r)

Figure 2 presents these calculations for the cross-section and cohort-adjusted VSLs
derived from the minimum distance estimator. The average VSLY is $296,000 for the cross-
section and $302,000 for the cohort-adjusted estimates. VSLYs follow a similar inverted U-
shaped relationship over the life cycle as depicted for VSL. The increase in VSLY is clearly
expected for young workers because VSL is increasing and life expectancy is decreasing. The
monotonic decrease in VSLY after its peak indicates that age-specific VSLs are decreasing at a
faster rate than life expectancy. The peak in the VSLY occurs at a higher value and at a much
higher age for the cohort-adjusted measure. It peaks at a value of $401,000 at age 54 for the
cohort-adjusted measure, as compared to a peak of $375,000 at age 45 for the cross-section
measure. The cohort-adjusted VSLY declines at a much slower rate than the VSLY after the
peak for the cross-section measure. The influence of cohort adjustments has an even greater
relative effect on the VSLY levels for the older workers in the sample than they did on VSL.
Interestingly, the VSLY for those age 62 is higher than for all age 39 or younger.

V. Conclusion

The implications of wage-risk tradeoffs for the dependency of VSL on age is consistent
based on both age group-specific estimated VSLs and a minimum distance estimator derived
from age-specific VSLs. We find that the VSL rises and then falls with age across the population
and over the life cycle, displaying an inverted U-shaped relationship. The minimum distance
estimator results are perhaps most instructive, as they can more flexibly represent the age
relationship while controlling for cohort effects. Failing to account for the secular increase in
incomes with birth-year indicator variables yields much lower VSLs for older individuals and
higher VSLs for younger individuals in cross-section analysis. Including cohort effects results in
a much flatter age-VSL function over the life cycle, and older individuals have a higher value of
a statistical life.

25 We have also calculated VSLYs based on a 7 percent discount rate (the current preferred rate by the U.S. Office
of Management and Budget for evaluating government regulations). The higher discount rate yields larger VSLYs
and a more pronounced inverted U-shaped age-VSLY relationship.

15

-------
Resources for the Future

Aldy and Viscusi

The result that the VSL rises and falls with age is of both theoretical and policy interest.
Theoretical analysis of VSL over the life cycle suggests such a relationship may exist,
particularly in situations in which there are insurance and capital market imperfections. The
results are supportive of these models rather than those that generate steadily declining VSL with
age, such as some models with perfect annuity and insurance markets. VSL is not steadily
declining with age even though the amount of expected lifetime at stake steadily declines with
age. As the life-cycle models indicate, this result is not surprising since the age-VSL linkage
depends on factors such as the life-cycle consumption pattern, which also displays a similar age
structure.

These estimates may help inform policymakers as they consider policies that would
simultaneously reduce mortality risk for individuals of various ages. In terms of the appropriate
"senior discount," in the cross-section analysis workers in their early 60s have a VSL of about
$1.7-$2.0 million, which is between one-fifth and one-fourth the size of the VSLs for prime-
aged workers. Understanding how the value of statistical life varies over the life cycle can inform
policymakers as they consider government interventions that would reduce mortality risks posed
to individuals over multiple stages of their life. The cohort-adjusted VSL levels for older workers
are much higher than in the cross-section analysis, with a VSL of about $5 million for workers in
their early 60s. While below the peak VSL over the life cycle, these older workers' VSLs are
above the VSLs for very young workers. This analysis does not provide support for approaches
that focus only on the remaining quantity of life as the valued attribute. Both the value per life-
year approach and the quality-adjusted life year methodology yield a steadily decreasing VSL
with age, whereas the revealed preferences of workers' risk decisions indicate a quite different
relationship that rises and then declines with age. Explicit construction of age-specific values of
statistical life-years from our age-VSL profiles show that the value of a statistical life-year varies
with age. Likewise, there is no support for the standard practice of transferring VSLs from
studies based on the average of the labor market to risk contexts specific to the elderly
population. Individuals make decisions over risk and income that clearly indicates that the value
of their life varies with age, but the relationship is not a simple one.

16

-------
Resources for the Future

Aldy and Viscusi

References

Aldy, J.E., and W.K. Viscusi. 2004. Age Variations in Workers' Value of Statistical Life. Olin
Center for Law, Economics, and Business Discussion Paper no. 392. Cambridge, MA:
Harvard Law School.

Altonji, J.G., and L.M. Segal. 1996. "Small-Sample Bias in GMM Estimation of Covariance
Structures." Journal of Business and Economic Statistics 14(3): 353-366.

Anderson, R.N. 1998. United States Abridged Life Tables, 1996. National Vital Statistics
Reports 47(13).

Arthur, W.B. 1981. "The Economics of Risks to Life." American Economic Review 71(1): 54-
64.

Chamberlain, G. 1984. "Panel Data." In: Z. Griliches and M.D. Intriligator, (eds.), Handbook of
Econometrics, Volume II, pages 1247-1318.

Deaton, A. 1985. Panel Data from Time Series of Cross-Sections. Journal of Econometrics 30:
109-126.

Deaton, A., and C. Paxson. 1994. Intertemporal Choice and Inequality. Journal of Political
Economy 102(3): 437-467.

Cropper, M.L. and F.G. Sussman. 1988. "Families and the Economics of Risks to Life."
American Economic Review 78(1): 255-260.

Ehrlich, I., and Y. Yin. 2004. "Explaining Diversities in Age-Specific Life Expectancies and
Values of Live Saving: A Numerical Analysis." NBER Working Paper 10759.

European Commission. 2001. "Recommended Interim Values for the Value of Preventing a
Fatality in DG Environment Cost Benefit Analysis." Internet:

http://europa.eu.int/comm/environment/enveco/others/recommended interim values.pdf

Hara Associates Inc. 2000. Benefit/Cost Analysis of Proposed Tobacco Products Information
Regulations. Prepared for Health Canada and Consulting and Audit Canada. Ottawa,
Ontario. June 5, 2000.

Hersch, J. 1998. "Compensating Differentials for Gender-Specific Job Injury Risks." American
Economic Review 88(3): 598-627.

17

-------
Resources for the Future

Aldy and Viscusi

Johannesson, M., P.-O. Johansson, and K.-G. Lofgren. 1997. "On the Value of Changes in Life
Expectancy: Blips Versus Parametric Changes." Journal of Risk and Uncertainty 15:
221-239.

Johansson, P.-O. 2002. "On the Definition and Age-Dependency of the Value of a Statistical
Life." Journal of Risk and Uncertainty 25(3): 251-263.

	 1996. "On the Value of Changes in Life Expectancy." Journal of Health Economics 15:

105-113.

Jones-Lee, M.W. 1989. The Economics of Safety and Physical Risk. Oxford: Basil Blackwell.

Jones-Lee, M.W., W.M. Hammerton, and P.R. Philips. 1985. "The Value of Safety: Results of a
National Sample Survey." Economic Journal 95: 49-72.

Rosen, S. 1988. "The Value of Changes in Life Expectancy." Journal of Risk and Uncertainty 1:
285-304.

Shepard, D.S., and R.J. Zeckhauser. 1984. "Survival Versus Consumption." Management
Science 30(4): 423-439.

Smith, V.K., M.F. Evans, H. Kim, and D.H. Taylor. 2004. "Do the Near-Elderly Value Mortality
Risks Differently?" Review of Economics and Statistics 86(1): 423-429.

Viscusi, W.K. 2004. The Value of Life: Estimates with Risks by Occupation and Industry.
Economic Inquiry 42(1): 29-48.

Viscusi, W.K., and J.E. Aldy. 2003. "The Value of a Statistical Life: A Critical Review of

Market Estimates Throughout the World." Journal of Risk and Uncertainty 27(1): 5-76.

Viscusi, W.K., and J. Hersch. 2001. "Cigarette Smokers as Job Risk Takers." Review of
Economics and Statistics 83 (2): 269-280.

18

-------
Resources for the Future

Aldy and Viscusi

Tables and Figures

Table 1. Age Group-Specific Values of a Statistical Life, Annual Cross-Sections, 1993-2000a

Year



18-24
Age Group

25-34
Age Group

35-44
Age Group

45-54
Age Group

55-62
Age Group

1993

Mortality
Risk

0.00040
(0.00045)
[0.00046]

0.00434
(0.00040)***
[0.00077]***

0.00308
(0.00041)***
[0.00084]***

0.000728
(0.00041)*
[0.00068]

0.00089
(0.00066)
[0.00087]



Mean VSL

$0.64

$9.92

$8.36

$2.04

$2.36

1994

Mortality
Risk

0.00238
(0.00047)***
[0.00064]***

0.00329
(0.00038)***
[0.00064]***

0.00277
(0.00038)***
[0.00078]***

0.00132
(0.00043)***
[0.00078]*

0.00176
(0.00065)***
[0.00083]**



Mean VSL

$3.97

$7.73

$7.75

$3.86

$4.87

1995

Mortality
Risk

0.00298
(0.00051)***
[0.00064]***

0.00313
(0.00039)***
[0.00063]***

0.00223
(0.00039)***
[0.00079]***

0.00174
(0.00042)***
[0.00078]**

0.00162
(0.00059)***
[0.00080]**



Mean VSL

$4.87

$7.31

$6.16

$5.02

$4.46

1996

Mortality
Risk

0.00319
(0.00077)***
[0.00089]***

0.00350
(0.00043)***
[0.00069]***

0.00310
(0.00044)***
[0.00084]***

0.00163
(0.00043)***
[0.00070]**

0.00124
(0.00056)**
[0.00069]*



Mean VSL

$5.13

$8.08

$8.45

$4.67

$3.39

1997

Mortality
Risk

0.00288
(0.00058)***
[0.00075]***

0.00348
(0.00043)***
[0.00071]***

0.00329
(0.00043)***
[0.00076]***

0.00196
(0.00043)***
[0.00073]***

0.00162
(0.00061)***
[0.00080]**



Mean VSL

$4.60

$8.08

$8.98

$5.64

$4.47

1998

Mortality
Risk

0.00346
(0.00064)***
[0.00086]***

0.00283
(0.00045)***
[0.00068]***

0.00305
(0.00044)***
[0.00076]***

0.00159
(0.00045)***
[0.00072]**

0.00158
(0.00058)***
[0.00078]***



Mean VSL

$5.65

$6.76

$8.61

$4.69

$4.55

1999

Mortality
Risk

0.00154
(0.00052)***
[0.00059]***

0.00359
(0.00050)***
[0.00071]***

0.00355
(0.00050)***
[0.00089]***

0.00337
(0.00048)***
[0.00082]***

0.00162
(0.00063)***
[0.00086]*



Mean VSL

$2.18

$7.18

$8.41

$8.35

$3.95

2000

Mortality
Risk

0.00211
(0.00060)***
[0.00073]***

0.00391
(0.00049)***
[0.00074]***

0.00356
(0.00047)***
[0.00088]***

0.00277
(0.00046)***
[0.00074]***

0.00135
(0.00059)**
[0.00086]



Mean VSL

$3.16

$9.03

$9.85

$7.97

$3.77

' VSLs are expressed in millions of year 2000 dollars based on age-specific wages. Dependent Variable: natural

logarithm of hourly labor income. Each specification includes 9 1 -digit occupation indicator variables, 8 regional
indicator variables, demographic variables, nonfatal injury risk, and expected workers' compensation replacement
rate. Robust (White) standard errors are presented in parentheses, and standard errors accounting for within-group
correlation are presented in brackets. ***, **, * Indicates statistical significance at 1 percent, 5 percent, and 10
percent levels, two-tailed test.

19

-------
Resources for the Future	Aldy and Viscusi

Table 2. Age Group-Specific Values of a Statistical Life, 2000a



18-24
Age Group

25-34
Age Group

35-44
Age Group

45-54
Age Group

55-62
Age Group

Mortality Risk

0.00211
(0.00060)***
[0.00073]***

0.00391
(0.00049)***
[0.00074]***

0.00356
(0.00047)***
[0.00088]***

0.00277
(0.00046)***
[0.00074]***

0.00135
(0.00059)**
[0.00086]

Mean Age Group VSL
(millions 2000$)

$3.16

$9.03

$9.85

$7.97

$3.77

Aae Group

H0: Pairwise Tests of Equality of VSL Estimates, F-Statistics, F(l, 118,639)

18-24

-

16.16

17.52

8.89

0.10

25-34

-

-

0.22

0.36

6.94

35-44

-

-

-

1.01

8.39

45-54

-

-

-

-

3.97

"N= 118,762. R =0.56. Dependent Variable: natural logarithm of hourly labor income. Specification includes 9
1-digit occupation indicator variables, 8 regional indicator variables, demographic variables, nonfatal injury risk,
and workers' compensation expected replacement rate. Robust (White) standard errors are presented in parentheses
and standard errors accounting for within-group correlation are presented in brackets. ***, ** Indicates statistical
significance at 1 percent, and 5 percent levels, two-tailed test.

20

-------
Resources for the Future	Aldy and Viscusi

Figure 1. Cohort-Adjusted and Cross-Section Value of Statistical Life, 1993-2000

VSL (millions 2000$)

NOTES: Both series are based on equally weighted minimum distance estimator with a third-order polynomial in age. The cohort-adjusted VSL also
includes indicator variables for year of birth.

21

-------
Resources for the Future	Aldy and Viscusi

Figure 2. Value of a Statistical Life-Year Based on Cohort-Adjusted and Cross-Section Value of Statistical Life,

1993-2000

VSLY (2000$)

NOTES: Value of statistical life-years based on an assumed 3 percent discount rate and average age-specific life expectancy and derived from the age-
specific VSLs presented in Figure 1.

22

-------
EPA NCER/NCEE Workshop

Morbidity and Mortality: How Do We Value the Risk of Illness and Death?

Empirical Issues Associated with Mortality Risk Valuation

Joseph Aldy and Kip Viscusi, "Adjusting the Value of a Statistical Life for Age
and Cohort Effects"

George Van Houtven, Melonie Sullivan, and Chris Dockins, "Eliciting Tradeoffs
for Valuing Fatal Cancer Risks"

Comments by Clark Nardinelli
April 11, 2006

The two papers in this session both explore differences in people's willingness to
pay for a small reduction in the risk of death, or what economists call the value of a
statistical life. Joseph Aldy and Kip Viscusi look at differences in the values of statistical
life across workers by ages, whereas George Van Houtven, Melonie Sullivan, and Chris
Dockins look at the difference between the value of reducing fatal cancer risks and the
value of reducing fatal accident risks. Because of the way I approached the papers, I will
first discuss Aldy and Viscusi.

1

-------
Aldy and Viscusi's work on the compensating differentials associated with
occupational risks forms the basis for the most widely-used estimates of the value of a
statistical life. Most regulatory economists use their estimates and will continue to do so.
Their new study is especially welcome because it deals with the highly controversial
topic of the so-called "senior discount", or more generally the effects of age on the value
of a statistical life. The recent dust-up over regulatory analyses that used a lower value of
statistical life for older persons, ensures that economists who analyze regulatory policies
will pay close attention to the results of this study.

Aldy and Viscusi introduce age-specific fatal and non-fatal occupational risks to
estimate the relationship between age and the value of a statistical life. The cross-
sectional hedonic regression based on age-specific fatal accident risks generates an
inverted U-shaped curve. The curve shows a steep decline from the peak value of $9.9
million per statistical life at age 39 to $3.8 million at age 62. Much of that decline,
however, apparently reflects the rise in real incomes over time. Older workers come from
generations with lower real earnings and therefore have lower values of statistical life.
Aldy and Viscusi show that the steep decline after age 39 comes from the combination of
the changes over time within an age cohort and the differences in lifetime earnings across
cohorts. Including a cohort adjustment in the regressions substantially flattens the slope
of the inverted U, particularly at older ages. In the cohort-adjusted estimates, the value
peaks at $7.8 million at age 46, and then declines to $5.1 million at age 62. Indeed, the
difference between maximum value and the value at age 62 in the cohort-adjusted
estimates is not large enough to be raise serious doubts in someone who believes that the
value of statistical life does not vary substantially across age groups. I would like to see if

2

-------
other adjustments would flatten the curve even more. Knowing what those adjustments
were might help illuminate the difference between these results and the stated preference
experiments that generate a constant value of a statistical life across ages.

As I read this paper and looked at the inverted U-shaped curve, I thought the
curve looked familiar. I thought about it for a while and finally realized that the value of
statistical life curve generated by their hedonic wage regressions has the same basic
shape as a traditional lifetime age-earnings or age-productivity curve. Because they have
the data on age and wage rates, I would like to see Aldy and Viscusi compare the age-
earnings curve with the age-value of statistical life curve. Do they have similar or
different shapes? Do they peak at the same age as earning or at a different age? By doing
the comparison, they can show how adding age-specific fatal accident rates to the
estimating equations alters the shape of the values of statistical life from that implied by
the age-earnings profile alone.

As a regulatory economist I was asked to consider the implications of these
results for my work. My first reaction was mild elation because we now have some
empirical estimates on how the value of statistical life changes with age. The Aldy-
Viscusi estimates cover ages 18 through 62 and most of the illnesses dealt with by my
agency affect people younger than 18 and older than 62, but I thought that perhaps we
could extend the estimating polynomial equations backward and forward out of sample to
generates estimates for our analyses. But then I immediately got discouraged as I read
Aldy and Viscusi's warning that "there is no support for the standard practice of
transferring VSLs from studies based on the average of the labor market to risk contexts
specific to the elderly population."

3

-------
Feeling a little depressed because Aldy and Viscusi would not let me use their
results, I turned to the paper by Van Houtven, Sullivan, and Dockins. I immediately felt
better because Van Houtven, Sullivan, and Dockins promised to show me how to use a
stated preference survey to derive a relationship between the willingness to pay to reduce
the risk of fatal cancer and the willingness to pay to reduce the risk of a fatal automobile
accident.

The paper's literature review surprised me. I did not know that some economists
have failed to find a premium when estimating willingness to pay to reduce cancer risks
compared with other risks. I am more familiar with the risk analysis research on cancer
risks; that literature always finds a cancer effect - including more dread, higher aversion,
steeper trade-offs, and skewed risk rankings when cancer is one of the risks. In studies of
risk perception, cancer always generates a response that cannot be explained by the
actuarial risk. Based on risk analysis literature, I did not think that the direction of the
cancer effect was in doubt, just the size.

Van Houtven, Sullivan, and Dockins indeed find a cancer premium, which they
model as the ratio of the value of a statistical cancer to the value of a statistical life as
estimated by the willingness to pay to reduce the risk of a fatal automobile accident. The
contrast between the near-instant death through automobile accident and the prolonged
illness and other unpleasantness that accompanies death from cancer allows them to
identify the full difference in willingness to pay as a cancer premium. They also show
how latency (death from cancer occurs years after exposure) and the types of cancer
influence the value of a statistical cancer.

4

-------
Van Houtven, Sullivan, and Dockins extract the value of a statistical cancer and
the value of a statistical life from the results of a probit analysis of stated preferences
between two hypothetical locations with different risk characteristics. The survey
respondents chose between moving to an area with fewer automobile accident deaths per
million than their current location and moving to an area with fewer cancer deaths per
million than their current location. Using a dichotomous choice to elicit preferences and
then deriving the willingness to pay from the results makes fewer demands on
participants than many stated preference methods but still gives the full continuous range
of results. I like the method and was pleased to hear that the Office of Management and
Budget has approved it in principle.

In the description of the participants, however, I noticed something strange in the
reported personal experiences of the survey participants. Among the participants, 18
percent reported knowing someone who had died in an automobile accident and 12
percent reported knowing someone who had died of cancer. But the number of annual
cancer deaths is 13 times larger than the number of annual automobile accident deaths.
Unless people who die of cancer die without friends or relatives, there's something not
quite right here. My guess is that the sample, drawn from a web-based national panel, is
not in fact made up of persons with unusual life experiences. Instead, the result tells us
something about the perceptions and recall biases associated with the risks of fatal cancer
and fatal automobile accidents. I suggest that Van Houtven, Sullivan, and Dockins
explore this result in later research. It may help them to explain more of the cancer
premium itself.

5

-------
Van Houtven, Sullivan, and Dockins find that, compared with the base case of a fatal
automobile accident, the value of a statistical cancer ranges from 2 to 3 times the value of
a statistical life, depending on the latency period - a large premium by any yardstick.

This finding alone makes the study of value to those of us who must assess the benefits of
policies designed to reduce the risks of cancer and other illnesses.

For policy analysis, however, it is important to identify the source of the cancer
premium. Van Houtven, Sullivan, and Dockins do not find a strong effect from the
period of morbidity that precedes death from cancer. This negative finding may reflect
the difficulty of teasing out an estimate of morbidity's contribution to the cancer
premium. For policy analysts, the more disturbing possibility is that the cancer premium
is not due to morbidity but to something else. Does the cancer premium reflect some fear
based not on actual but on imagined outcomes or superstition? The responses of people
with some experience of cancer imply that much of the fear associated with cancer may
stem from unfamiliarity with the illness.

We need some way to separate the real from the imaginary parts of the cancer
premium. Morbidity must account for the real part of the premium. Morbidity can cause
real losses above the value of a statistical death directly through the effects of illness on
victims and perhaps indirectly through its effects on friends and relatives. Whatever
generates the large cancer premium found here has to be related to something that
happens before death, during the morbidity phase of the cancer. Once death occurs, the
cause ceases to matter.

Many researchers apparently believe that we should treat different causes of death
differently based solely on differences in willingness to pay, a practice difficult to justify

6

-------
for regulatory analysis. Suppose that we could survey a representative sample of dead
people. If we asked them whether they found it worse to be dead because of cancer or
worse to be dead because of an automobile accident, I doubt that we would find a
significant difference between the cancer dead and the accident dead. Unless we have
some survey evidence from dead persons saying that yes, it is far worse to be dead
because of cancer than anything else, we have to find something that occurs before death
that makes cancer worse than a fatal accident. The cancer premium derived from stated
preference alone does not justify regulatory analysts placing a higher value of preventing
statistical cancers.

To make these results of practical use for policy analysis, Van Houtven, Sullivan,
and Dockins need to tease the mortality and morbidity effects out of the value of
statistical cancers. Doing so would make it possible to de-compose the cancer premium
into a realized utility loss and the superstition, stigma, or dread that generates the rest of
the premium. I am not suggesting that superstition plays no role in real-world market
valuation, only that it should play no role in the valuations used by public health
agencies. Public health agencies exist partly because the general public does not always
have adequate information on true actuarial probabilities and true severities. In valuing
policy alternatives, regulators should ignore superstition, stigma, and irrational fear. As
public health economists, we should measure human welfare with actuarially correct risks
and real measures of severity, not with dread, superstition, or other imaginary effects.
The Food and Drug Administration, for example, came into existence to reduce the
consumption of snake oil. To assess the effects of that agency's regulations based on
imaginary effects would be the equivalent of introducing snake oil into regulatory

7

-------
analyses. Public health agencies should stick to concrete, measurable health effects when
assessing regulatory policies.

Let me conclude by saying that I am grateful for the opportunity to read and
comment on these two fascinating papers.

8

-------
Morbidity and Mortality:
How Do We Value the Risk
of Death and Illness?



Comments on VSL Papers

• ••
• •••



• • •• B



• •••

Maureen L. Cropper

• ••••

University of Maryland and World Bank

• •••
• •••

April 11, 2006

1

Eliciting Risk Tradeoffs for •
Valuing Fatal Cancer Risks :

•	Why This Paper is Important:

•	SAB's Environmental Economics Advisory
Committee determined in 2000 no valid estimates
exist of the value of a fatal cancer case

•	This paper fills this void by estimating the ratio of
the value of a statistical fatal cancer (VSC) to the
VSL associated with immediate, accidental death in
an auto accident (VSL)

VSC = -dPD
VSL dPc|E(U)=k

-------
Answers Seem Generally
Reasonable

•	78% of respondents pass probability choice quizzes

•	In choice between City B (fewer Cancer deaths) and City
A (fewer Auto deaths), proportion choosing City B FALLS
as the relative death ratio (RDR)—the number of auto
deaths saved for every additional cancer death—rises

•	Latency reduces proportion choosing City B, holding the
RDR constant

BUT:

•	Percent choosing City B never falls below 50% even with 25
year latency

•	No sensitivity to length of morbidity preceding death

Effect of Risk Difference Ratio on :•••
P(Choose City B)

~	25 year latency

~	15 year latency
B 5 year latency

Percent 100
Preferring
Cancer Risk
Avoidance

0.43 0.71 1	1.4 2.33

Risk Difference Ratio (auto:cancer)

-------
• ••
• ••

How to Value Non-Fatal Cancers?

• Risk-risk tradeoffs for non-fatal cancers
effectively produce a QALY weight for
cancer:

vsc

VSL

(

1 U(C,Y)
U(H,Y)

•	How does this compare with other elicitation
methods for obtaining QALY weights?

•	Does this effectively monetize QALYs?

5

• ••
• ••
• •

Adjusting the Value of a Statistical
Life for Age and Cohort Effects

Why this paper is important:

•	Examines how hedonic wage function shifts in wage-risk
space with worker age

•	Estimates how MWTP for a change in risk changes with
age

•	Uses multiple cross-sections to disentangle age and cohort
effects

• However, to use these estimates for policy, one
must believe that hedonic wage equations provide
unbiased estimates of a change in risk on the wage.

6

-------
• ••
• ••

Should Hedonic Wage Equations Be
Interpreted as Causal Relationships?

•	How do we interpret a cross-sectional regression of infant mortality on
air pollution levels?

•	How do we interpret a regression of property values on air pollution
levels using a single cross section of data?

•	Due to omitted variable bias problems both results would be suspect:
Need to find a natural experiment that causes an exogenous change in
air quality (see e.g. Chay and Greenstone, QJE, August 2003; JPE,

April 2005).

•	Dan Black et al. (2003) raise similar concerns about hedonic wage
equations: risk is likely to be correlated with the error term, causing
results to be suspect

•	Perhaps a natural experiment involving changes in road safety could be
used to measure the impact of changes in fatal risk on wages of
transport operators.

-------
Summary of the Q&A Discussion Following Session V

JR. DeShazo, (UCLA)

Directing his comments to Dr. Joe Aldy, Dr. DeShazo stated, "We do have some stated
preference estimates that reflect almost exactly the same pattern of age-adjusted VSLs
that you revealed—they're slightly lower, on average." Referring to comments made by
discussant Clark Nardinelli, he added that because he and his colleagues had a large
number of seniors in their sample, these age-adjusted VSLs "actually can be used to
evaluate the senior population." He also stated that "the general pattern is very much the
same, and I think the only difference is that our VSL estimate is about $2 million less, on
average."

Saying that he had two questions, Dr. DeShazo posed the first: "Why do we see this
change with age?" He cited Ehrlich's work as "the best theoretical study to date." He
said that Ehrlich "shows that there are a variety of reasons that we might value reductions
in health risks as we age. First and foremost, we value health in the current period, and
as we age the marginal utility of that is going to increase as our health state declines."
Other factors Dr. DeShazo identified include "changes in the remaining expected
lifespan—changes in the marginal utility of income or consumption—changes in
individuals' discount rates that you would, consistent with theory, expect to increase. Of
course, for any given risk that you're focusing on, the background risk profile is
changing, so that the other risks that you face are going up." Addressing the presenters,
he asked, "Out of those things that vary with age, what do you think explains this
decline?" He added that in their analysis he and his colleagues "were able to identify
changes in the marginal utility of consumption, changes in the discount rate, and changes
in the background risks that people face as they age, all of which might explain this
decline."

The second question from Dr. DeShazo, which he classified as "much more fundamental"
than the first, was "whether or not we should be applying hedonic estimates which give
us current period values for a mortality risk reduction in efforts to value the types of
health risk reductions that FDA and EPA focus on primarily, which follow, typically,
years of chronic or severe morbidity. The basic question is: Is the marginal mortality
risk reduction the same today, if you're perfectly healthy, as it would be if you've
suffered for 10 years from chronic morbidity or maybe 3 years from severe morbidity?"
He added that preliminary results from his studies "suggest that that's not the case—that
the marginal value of a risk reduction is highly context dependent, and your willingness
to trade off morbidity and mortality health states is such that your value of mortality
reductions falls as you experience more prior morbidity." In closing, Dr. DeShazo asked,
"So, what's the best argument for transferring the sort of hedonic wage analysis?"

Joe Aldy, (Resources for the Future)

"To get to the first question about why we think we see this kind of age profile with
respect to the value of a statistical life—I think when we look at the young workers that
there are two things driving that result. One is that for the 18- to 24-year-olds we're

Session II Q&A Summary

1

-------
focusing on just the full-time workers. That group is going to have a disproportionally
large sample of full-time workers who do not have a college degree relative to older age
groups in our sample. To the extent that those are people who never have a college
degree, they have lower lifetime income, and we would expect them to have a lower risk-
income tradeoff in the labor market. I think the other thing that's driving that is that you
can't really borrow against future income. With the exception of college loans, it's really
hard to be able to do that. This is why we see very little savings behavior among most
households until they get into their late 30's or early 40's. You see a little bit of
precautionary savings, but other than that, very little saving occurs early in life, and I
think that's one reason we see the lower value for the younger workers." He added that
the impact seen with older workers is, he feels, "being dominated by life expectancy."
He also said he felt it would be great to get a sense of what's driving "stated preference
results that show relatively small declines or no decline in the value of life," and he asked
"is it because we see changes in discount rates as they get older, and we see differences in
risk attitudes? Are there ways in which we can try to structure future surveys to try to get
at those questions explicitly so we can have a better understanding of why we see that
impact?" Citing existing literature and specifically naming Zeckhauser, Rosen, Ehrlich
(two papers), and Johannson (several papers), Dr. Aldy stated that most of this work
shows that the value of life is going to decline as people get older. He added, however,
that almost all of these studies "have assumed that attitudes toward risk are constant
across the lifecycle and the discount rate is constant over the lifecycle. Also, with the
exception of the Ehrlich stuff, health doesn't really enter in at all. As we get more
complex, there's the question: Are we able with a richer model and a richer theory to
explain the fact that the value of life may not decline much with age? With what we're
finding among the workers, though, we don't have any basis for saying it's because of
health—their health can't be changing that much because they're still full-time workers.
It might be declining some, but I think at the end of the day it's life expectancy that's
really going to be driving the results on that."

Dr. Aldy then turned to the other question of why we should be using hedonic wage
VSLs when we look at policies that have latent impacts. He commented that "part of the
point we made at the end of our paper is that if most of those you are looking at are
elderly who are enjoying the benefits—whether it's clear skies, whether it's the tier-2 rule
that reduced sulfur in gasoline and had a lot of PM reduction benefits—in that context it's
difficult to reconcile a VSL where the average age of the worker is 35-40 years old. You
bring up the latency issue; I think the age issue makes it difficult." He went on to
explain, "There's a question of whether or not in ours you say: Well, you come up with
this value of life for someone who is 62—that still doesn't really help me much if the
average age of the beneficiary is 70. I haven't even really gotten to that person yet, and if
you think, for example, there may be differences in attitudes after one leaves the
workforce. One can come up with plausible stories about how one's attitudes toward risk
would vary from what they were before and raise questions about whether or not one
should be trying to transfer a hedonic wage estimate over. So, I'm not going to come out
and forcefully say that hedonic wage is the way to go. As I already mentioned, I have a
personal bias towards revealed preference. My co-author may have a slight bias towards
that, but he's done a good number of CV studies, too."

Session II Q&A Summary

2

-------
He closed by saying, "I think we're getting closer to doing a better job in the hedonic
wage literature of trying to get to interesting questions. For about 15 years, there wasn't
much that was very interesting going on. I think if you look at the last 5 years, it's
getting more interesting to try to better understand the heterogeneity in the value of life. I
think that's something that if you work through theory models, you start seeing that there
should be a lot of heterogeneity in the value of life, not just with respect to age but to
other issues and attributes as well. We're moving in that direction, but I'm not standing
up here and saying you should ignore all the stated preference stuff and go with hedonic
wage, because I don't think we're really cracking that nut yet."

Bryan Hubbell (U.S. EPA)

Dr. Hubbell stated, "Just for the record, we're actually not using VSLY either anymore,
even in sensitivity analyses," and added "but what I really actually wanted to raise a
question about is: One of the things I find most intriguing about your results were the
graphs that showed that fatal risks increase with age." Saying that this issue has bugged
him for a number of years, Dr. Hubbell asked, "How well do we really understand the
wage trajectory over the lifespan of an individual? I think about a person entering an
occupation, and with that occupational choice they're making a decision at that point
about the level of risk they want to accept in the wage tradeoff. From that point forward,
however, how much are they actually able to renegotiate based on their own individual
age-level risk with their wage trajectory? . . . Say, for example, that risk didn't change
over your lifespan for the particular occupation that you're in. You would then actually
see an increase in VSL over time, simply because your skill level and wages are going
up, and so forth, but your risk level is going to stay the same. So, you're getting an
implied VSL that perhaps would seem large if you didn't adjust for individual specific
factors well enough. So, one question that arises is: Without a panel study—if you're
not controlling random effects in panel fashion—are we getting confounding with
individual-level effects in terms of the wage-risk relationship? The other question is:
How would individuals actually understand how those risks change over time? Certainly,
it was unexpected to see that change, and the question is how much is that information
actually out there so there's difference between perceived risk and actual risk in those
particular cases."

He closed by mentioning, "One of the things I think is interesting is that there are not a
lot of studies out there looking at things other than hedonics as a way to identify these
marginal values using the labor market decisions," and he questioned whether hedonics
were the only way to use this information or "are there other revealed-preference
methods, such as discrete choice type models which look at occupational or job choices
and job switching, that could help capture some of this information." He stated that
"hedonics tend to assume that there is no bundling of attributes, that you can have any
kind of combination of experience with risk and everything else in a very continuous
fashion so that you can get these derivatives—and if that doesn't occur, you can actually
end up with some biased results. Hence, the question is: How much have we explored
the bundling of attributes and jobs and whether we can disentangle that."

Session II Q&A Summary

3

-------
Joe Aldy

Dr. Aldy answered, "There at the end, Bryan, you started addressing how I would
respond to the very first question that you raised, which is whether or not over their
lifecycle workers can really adjust their wage or salary in response to changes in the risk.
Clearly when one applies the model, we're making the assumption that the labor market
has enough freedom and mobility so that one can move—so that you can get these kind
of equilibria. In the case of what we've done, these equilibria that are age specific, we
looked at the relationship between what the firms offer in terms of a combination of
safety and labor compensation and what the workers are demanding in terms of that
combination of risk and income. We so see that labor income does increase over much of
the lifecycle, but then it does start to decrease for some older workers in their 50's. We
actually see in our sample a slight decline in labor income for our oldest group." He
explained that this could be partly explained by "the standard story that for those who
stay within one firm for a long time there's sort of an agreement that they will be paid
less than their marginal product when they are young and then will tend to be paid more
than their marginal product when they're older." He added that "there's some concern
that because of this we're not really getting the right measures. Having said that, we still
should be seeing among workers that if they really don't like what's being offered to
them they should be going to a different job. That does raise the question of whether
workers are really that mobile when they're at older stages of their lives." This led back
to the Dr. Hubbell's comment regarding "if there isn't sort of a continuous set of job
market characteristic bundles, then you could have some potential problems with this."

Dr. Aldy summarized that "unfortunately there's not much one can do when looking at
these cross sections. The benefit of using the CPS is that it's a massive cross section; the
downside is that it's just a cross section. You could construct a quasi-two-year panel, but
that's actually very problematic with how they've designed the CPS. We're actually
thinking about trying to explore the PSID, where we could have a pretty long panel." He
added that he has not "seen any evidence of what perceived risks are and how they vary
with age" and he said that he wasn't sure "if in academia professors know how their risk
profile changes with their age, but if you're in a blue-collar job where there is a good
number of injuries, there's probably a decent sense of that." He said he's sure that there's
a sense of that within the firms, "because they're the ones who have to pay workmans'
compensation premiums to the state governments, so they should have a sense of how the
more serious incidents that can lead to either long-term hospitalization or fatality can
vary with age."

Mary Evans, (University of Tennessee)

Dr. Evans said that she first "wanted to applaud your efforts in this paper and other
papers in refining the occupational job risk measure that we're able to use—I think that's
an important contribution to the literature." She then picked up on something that was
mentioned in the response to J.R. DeShazo's earlier question, which was "the issue of
possible selection effects within the sample." She said, "You focused on the lower age

Session II Q&A Summary

4

-------
range, the 18- to 24-year-olds, but my concerns are more about the upper end of the age
distribution, the 55- to 62-year-olds. My question is: Are you able to estimate some sort
of selection model?—Have you thought about doing that?—and how do you think doing
so might impact your results?"

Joe Aldy

Dr. Aldy responded, "Again, the problem here is in using the CPS—it doesn't really give
you anything to identify the decision to work or not." He went on to say that although
they had not looked at non-labor income in the household, they did "try to look at non-
head-of-household labor income" and added that "it yielded virtually no impact on our
estimates, but it wasn't a good instrument—it clearly was not exogenous." Dr. Aldy
closed by saying, "As I mentioned, we're thinking about trying to go with the PSID,
where we can use a panel. There we can definitely use asset measures to try to identify
that... it's a much, much richer data set that might enable us to identify selection. It's
one reason why we decided to cut the age at 62, the age of early social security
retirement. We recognize that there's still a potential problem there, but there are these
tradeoffs with the sample that we wanted to use, because at the end of the day when you
want to cut this thing by a specific age, it helps to have a 100,000 observation sample.
You know, if you want to get the VSL for people who are just age 60, you're not going to
get much if you're using the PSID. In the end, that's the tradeoffs one has to make when
using a really large data set."

Greg Poe, (Cornell University)

Addressing his comment "to George [Van Houtven] and to the audience," Mr. Poe stated,
"It really looks desirable that we're doing risk-risk tradeoffs, and that pulls out the
money, and we immediately think that this is going to be a much simpler type decision
framework and it might be better. However, a long body of literature from marketing to
psychologists to political choice to even birds and bees has shown that these two choices,
these tradeoffs, are not that stable. You can add a third choice, which would pass some
sort of dominance test such as Maureen [Cropper] referred to, which nobody would
choose, but it greatly changes the proportions of people who choose both of those. So,
just because we're getting rid of money doesn't mean that we've solved everything and it
immediately makes it a preferred technique—it's just another technique, and we need to
investigate that."

George Van Houtven

"I would agree that it certainly doesn't solve everything, and I wouldn't want to make
that claim. I do think, though, that the framework helps in terms of cognitive burden the
way we've set it up—in terms of not having to spend as much time explaining the
absolute value of the risks but rather the relative risks. It's easier for people to trade
off—you know, when the denominators are the same, they sort of get canceled out of the
equation as long as we're willing to assume and expect the utility framework or
something close to that. That's the sense in which I think it maybe offers some
advantage, but otherwise I agree—it's not a silver bullet."

Reed Johnson, (RTI)

Session II Q&A Summary

5

-------
Dr. Johnson commented that "both Joe [Aldy] and Maureen [Cropper] made passing
reference to possible connections between the value of a statistical life year and
willingness to pay for QALY. I just want to emphasize Alan's [Krupnick] point
yesterday that it's important to be vigilant about keeping separate welfare theoretic
measures from quality-adjusted health indices like QALY. We get hit over the head all
the time by non-economists about this term "value of a statistical life." It's a bad choice
of terms, and I think Trudy [Cameron] has advocated something that really describes
what we're getting at, which is the willingness to pay for a small change in risk. I think
we deserve to be ridiculed if we make the same mistake they make in thinking of the
value of a statistical life year as the value of a life year, which of course is what
willingness to pay for QALY is supposed to get at."

He went on to say that he's "not sure, though, about George's [Van Houtven]
manipulation of ratios of values of a statistical life, whether there's some way of backing
out some QALY-like measure out of that. I think it's worth thinking about that, but let's
not make the same mistake non-economists make about dropping the statistical part of
these value measures."

DavidRisley, (U.S. EPA)

Starting with the disclaimer that he just started with the Clean Air Markets division about
a month ago, so this is all very new to him, Mr. Risley addressed this comment/question
to Joe Aldy: "Maybe I'm just offended that my VSL seems to be lower than most
peoples' but. . . I'm three years out of undergraduate studies; I have no savings; I moved
to the most expensive part of D C.; and I'm about to start grad school. I know that my
debt will be growing, but I hope that in the future I'll have earnings potential. I was just
wondering if there's any thought of perhaps adding my current VSL to some fractional
VSL that represents the likelihood that I'll get to be 40 make more money and have a
family."

Joe Aldy

Dr. Aldy replied, "The good news is that your value is going to go up for a while. When
Kip [Viscusi] and I were working on this, he didn't like the fact that the peak was always
at an age younger than his current age. . . When I think about whether or not the younger
population should have a lower value of life, I sometimes wrestle with the question of—
you know, this is reflecting imperfections in the labor market, in the capital markets, and
that's why they have this lower value. Having said that, this is still what people are using
to make actual decisions. Your compensating differential for mortality risk in your job at
EPA I presume is probably pretty low—I actually haven't looked at the data closely
enough to know how risky it is to work at EPA—I hope it's low. What one is able to
infer from that suggests that you're making what is, for you, a rational decision. The idea
here is that you're valuing your current consumption enough that you're willing to take
less compensation for that probability of dying on the job right now—that's what's
implicit in the modeling framework that we have."

Wrapping up, Dr. Aldy commented, "You can take some offense. I can tell you that my
father, whom I've talked to about this, takes probably greater offense, for the same reason

Session II Q&A Summary

6

-------
Kip does. At the end of the day, what we're trying to say here is that these are the values
through which people are trying to reveal their preferences about labor income and risk in
their labor market decisions. There are a host of issues in terms of how we try to estimate
this and interpret it, as we've already discussed, but at the end of the day that's what
we're trying to achieve empirically. It would be nice if you could go to a bank right now
and say, "Hey, I'm about to go to graduate school and I'm going to make a lot of
money—why don't you give me a lot of money right now?" They're probably not going
to do that, but if they did, then you would probably be demanding a larger compensation
for the probability of dying at EPA sometime over the next couple of years."

END OF SESSION V Q&A

Session II Q&A Summary

7

-------
Morbidity and Mortality: How Do We Value the Risk of

Illness and Death?

PROCEEDINGS OF SESSION VI: VALUING MORBIDITY AND MORTALITY:

DRINKING WATER

A WORKSHOP SPONSORED BY THE U.S. ENVIRONMENTAL PROTECTION
AGENCY'S NATIONAL CENTER FOR ENVIRONMENTAL ECONOMICS AND
NATIONAL CENTER FOR ENVIRONMENTAL RESEARCH

April 10-12, 2006

National Transportation Safety Board
Washington, DC 20594

Prepared by Alpha-Gamma Technologies, Inc.
4700 Falls of Neuse Road, Suite 350, Raleigh, NC 27609

ACKNOWLEDGEMENTS

This report has been prepared by Alpha-Gamma Technologies, Inc. with funding from
the National Center for Environmental Economics (NCEE). Alpha-Gamma wishes to
thank NCEE's Maggie Miller and the Project Officer, Cheryl R. Brown, for their
guidance and assistance throughout this project.

DISCLAIMER

These proceedings have been prepared by Alpha-Gamma Technologies, Inc. under
Contract No. 68-W-01-055 by United States Environmental Protection Agency Office of
Water. These proceedings have been funded by the United States Environmental
Protection Agency. The contents of this document may not necessarily reflect the views
of the Agency and no official endorsement should be inferred.

-------
Table of Contents

Session VI: Valuing Morbidity and Mortality: Drinking Water

Session Moderator: John Powers, U.S. EPA, Office of Water

Combining Psychological and Economic Methods To Improve
Understanding of Factors Determining Adults' Valuation of Children's
Health

Cheryl Asmus, Paul Bell, John Loomis, Byron Allen, and Helen Zita Cooney,
Colorado State University

Economic Valuation of Avoiding Exposure to Arsenic in Drinking Water

Kathleen Bell, University of Maine, and Kevin Boyle, Virginia Polytechnic Institute

Perceived Mortality Risks and Arsenic in Drinking Water:Preliminary
Research

Douglass Shaw, Texas A&M University; Paul Jakus, Utah State University; Klaus
Moeltner and Mark Walker, University of Nevada-Reno; and Mary Riddel,
University of Nevada-Las Vegas

Willingness To Pay To Reduce Community Health Risks from Municipal
Drinking Water: A Stated Preference Study

Alan Krupnick, Resources for the Future; Vic Adamowicz, University of Alberta;
and Diane Dupont, Brock University

Discussant: Trish Hall, U.S. EPA, Office of Ground Water and Drinking Water
Discussant: Greg Poe, Cornell University
Questions and Discussion

-------
Combining Psychological and Economic Methods to Improve Understanding of Factors
Determining Adults' Valuation of Children's Health: The Case of Nitrates and Infants

Cheryl Asmus, John Loomis, Helen Cooney, Paul Bell and Bryon Allen (Colorado State

University)

May 11, 2006

Abstract

The objective of this research is to evaluate the gain in explanatory power from adding
independent variables from the a psychology model of predicting behavior, the Theory of
Planned Behavior (TPB) to an economic model, conjoint analysis for determining adults'
willingness to pay (WTP) to protect children's health, with the method to be adapted for policy-
making. For the development of this method, nitrate in drinking water will serve as the risk
factor because it only affects children's health. A questionnaire is used to assess knowledge,
attitudes, beliefs, norms, and perceived control with respect to the risk factor, as well as the
components of TPB. Respondents also complete a choice task for a conjoint analysis to assess
their preferred choices of behavior for averting this risk. One half of the groups are told the
choice is hypothetical. The other group is told that one of their four choices will be binding and
they will actually buy the amount of bottled water using the money given to them at the
beginning of the experiment. We test whether the behavioral responses of these two groups are
equivalent or not. The majority of the data collected to date have been in the English-speaking
(88%) and hypothetical (76%) treatments.

There was a statistically significant difference in the real/cash cost coefficient and when the costs
were hypothetical. The real/cash cost coefficient was far more negative (price sensitive) than the
hypothetical cost coefficient, although the hypothetical cost coefficient was still negative.

A household would pay $2.64 in the real cash treatment and $18 in the hypothetical treatment for
bottled water that would result in a .0001 (1 in a thousand) reduction in the chances of an infant
going into shock from nitrate in water. A household would pay $5.25 in the real cash treatment
and $36 in the hypothetical treatment for bottled water that would result in a .0001 (1 in a
thousand) reduction in the chances of an infant experiencing permanent brain damage from
nitrate in water.

Dividing the coefficient on infants present in the household by the cost coefficients allows us to
calculate to investigate the extent of altruism of households without children in terms of their
willingness to pay to buy bottled water for households with infants. While willingness to pay
(WTP) rises by $49 with real money and $332 for hypothetical payment for households with
infants at risk, WTP is still positive for households without an infant. This suggests there is some
measure of altruism reflected in our WTP results.

The Theory of Planned Behavior (TPB) variables were only significant at between the .14 and
.19 levels and added about 2.5% to the explanatory power of the logistic regression model. Of
the variables in the TPB, attitudes about infant health issues were not significant p = .48, health
perceived control (one can protect an infant from environmental contaminants) was significant at
p = . 19, water perceived control (one can control the quality of one's drinking water) was

1

-------
significant at p = . 14, and water norms (subjective norms for being concerned about drinking
water quality) was significant at p = . 15. The results indicate that perhaps perceived control and
community norms would be most useful for a policy maker.

Legal Background

Increasingly federal agencies are being called upon to explicitly factor children's health into their
regulatory decisions and benefit cost analyses. For example Executive Order 13045 issued by
President Clinton on April 21, 1997 required making children's health a high priority in federal
agency decision making. In that same year, EPA established the Office of Child Health
Protection to give increased emphasis on children health in the agency's many programs. See
U.S.E.P.A. (2003) for more details on the Executive Order.

Study Objective

There are two basic issues when valuing children's health. One is selecting the appropriate risk-
reducing policies and actions and the other is the value of reducing these risks. Although it is
important to economically put a value on the reduction of an environmental health risk to a child,
doing so does not necessarily give public and private stakeholders the information they really
need to decide upon the appropriate policies or actions.

The kind of information that is most useful to these stakeholders would not necessarily be a
dollar figure. It may be an understanding of how and if knowledge, education, belief systems,
cultural or societal norms and general attitudes actually lead to the decisions each individual
makes when they put a value on a child's health.

To that end, the objective of the proposed research is to test a combining of the explanatory
variables from the Theory of Planned Behavior (TPB) with conjoint analysis for determining
adults' willingness to pay (WTP) to protect children's health, with the method to be adapted for
policy-making.

Study Design

The overall study design focuses on:

(a)	Deriving adults' willingness to pay to reduce their infants' risk of shock, brain damage
and death from nitrate in drinking water during their first year of life;

(b)	Deriving these values using a choice experiment, which involves a hypothetical WTP for
bottled water.

(c)	Using a consequential treatment in which adults will be asked to pay real money for the
bottled water, with a pre-paid coupon for the bottled water provided to those agreeing to
pay.

(d)	Using the Theory of Planned Behavior to see if attitudes, beliefs, knowledge, norms, and
perceived control increase the predictability of adults' WTP choices.

(e)	Testing for whether there is altruism toward children's health by testing whether people
without infants at risk would pay for bottled water for other households with infants at
risk.

2

-------
Hypotheses:

1.	In tests of the internal validity of the choice experiment, adults' demand for
children's health will be reduced at higher prices (i.e., negative own price), and
positive with respective to the amount of risk reduction.

2.	In tests of the external validity of the experiment, marginal value for risk
reduction i from the traditional hypothetical choice experiment (MVi(h)) will
equal the marginal value for risk reduction i from the consequential (real money)
choice experiment (MVi(c)).

3.	In tests of the predictive power of the Theory of Planned Behavior, regression
analyses of WTP for bottled water as a function of risk reduction and cost will
show increased predictability by adding beliefs, attitudes, knowledge, subjective
norms, and perceived control as predictors and which of those predictors may be
more relevant to stakeholders and policy-makers as they make decisions around
education or potential mitigation.

Literature Review

Agee and Crocker provide an evaluation of the available methods for valuing children's health.
They suggest that stated preference methods such contingent valuation are one of two methods
that are most theoretically tenable and analytically tractable. Stated preference methods are not
only able to measure parents' willingness to pay for their children, but may also allow elicitation
of community public good values toward children's health as well.

While there is a rising demand for children's health information, there have been very few
primary valuation studies of children's health issues using stated preference methods. One of the
first was Viscusi, et al. (1987) where adults are asked their WTP to reduce adverse health effects
to children (in this case pesticide poisonings). Dickie and Messman (2004) perform a very
thorough stated preference study of parents' WTP to reduce their own acute illnesses versus
those of their children. They used WTP for a medicine that would treat the acute respiratory
symptoms such as cough, chest pain, shortness of breath, fever and the untreated duration of
these symptoms. For severe acute illness parents are WTP about $217 to reduce one symptom
day (Dickie and Messman, 2004: 1167). The values for younger children (age three) is nearly
double that of children ages 12 to 17.

WTP of parents to reduce latent skin cancer chances were studied by Dickie and Gerking based
on parents WTP for a sunscreen product. Liu, et al. (2000) studied mother's WTP to reduce their
own and their child's multiple day, multiple symptom episodes of colds in Taiwan. Converting
WTP into U.S. dollars average WTP was $71, and upwards of $121 if adjustments made for
differences in income levels and a mid-range income elasticity of WTP.

Valuation Methodology

The methodological approach used in this study is based on the conjoint or choice experiment
approach (Holmes & Adamowicz, 2003). This is a stated preference method, in which a
respondent makes a series of contingent choices. These choices are contingent upon the
characteristics in the choice set. Our choice set has cost as one attribute, and risk of the child
going into shock, risk of the child suffering brain damage and risk of death as the key variables

3

-------
we wish to value. By dividing the attribute coefficient by the cost coefficient the marginal value
of a one unit change is monetized.

Following theoretical foundation of Hanemann (1984) on utility difference from random utility
models and Roe et al. (1996)'s application to conjoint, we make the first choice a "no action" or
baseline risk level associated with no cost. Then the action alternative that reduces the three
health risks to the child is offered at a one time cost of X, that varies across the sample. We do
this in pairwise fashion, whereby each choice task or choice set is a no action and a single action
alternative. As Carson et al. suggest, having just two choices increases the likelihood that the
choice will be incentive compatible (even in the hypothetical treatment).

The probability a respondent will choose the action alternative should be related to the expected
gain in the parents' well being obtained from their infant receiving the health risk reduction, over
and above the satisfaction lost due to paying higher cost. To be more specific, a state-dependent
utility function is posited focusing just on the risk of death, to keep the notation simple. Thus UL
and UD is the utility to the parent when the child is alive and dead, respectively. Further let PD
be the baseline probability of the child dying with and without the risk reduction intervention
(e.g.., bottled water). Baseline expected utility (EU) to the parent can be defined as:

EU = PD[UD(I)] + (1-PD)[UL(I)],

where I is income.

The parents' purchase of bottled water reduces the probability of premature death from PD to
P'D, but at a proposed cost to the respondent of $X each year. If the reduction in the probability
of premature death from PD to P'D yields more expected utility than the loss of $X in income,
the parent will select the action alternative in the choice question. Specifically, the expected
utility difference (EUD) is given by:

EUD= {P'D[UD(I-$X)]+ (1 -P'D)[UL(I-$X)]} - {PD[UD(I)]+ (1-PD)[UL(I)]}

If this expected utility difference is linear in its arguments, and if the associated additive random
error term is distributed logistically, then the probability a respondent will select the action
alternative to a question asking him or her to pay $X for the bottled water that would reduce the
risk of the child's death from PD to P'D is:

Probability of buying bottled water = P(Y) = 1 - [ 1 + eBo B1($x)]1

Maximum likelihood statistical routines such as logistic regression can be used to estimate a
transformation of this equation in the form of:

Log {P(Y)/[1-P(Y)]} = Bo - B1($X) +B2 (Reduction in Risk of Death)

The marginal value to the parent of reducing a child's risk of death (or parental WTP) is:

B2/B1.

4

-------
Theory of Planned Behavior

Besides deriving the adults' value of each type of risk reduction, we wish to explore whether the
Theory of Planned Behavior (TPB) adds explanatory power to this model.

According to TPB, there are certain factors that influence behavior. Attitudes toward behavior,
knowledge, subjective norms (beliefs about whether the behavior is appropriate), and perceived
control have a combined influence on behavioral intentions (whether the individual intends to
engage in the behavior or not). In this study, the choices made in the contingent valuation task
served as a measure of behavioral intentions. Attitudes, beliefs, knowledge, and perceived
control were measured via a questionnaire.

Theory of Planned Behavior

Some sample items are: "I am not aware of any potential negative health effects for children
caused by drinking water contaminated with nitrate" (knowledge); "Overall, the children in my
community are healthy" (beliefs); "Children's health is an important issue" (attitudes); "Most of
the people I know would take steps to ensure that their drinking water is safe" (subjective
norms); and "I can ensure that my children are healthy" (perceived control). Behavioral
intentions will be assessed via a contingent valuation task. Actual Behavior will be assessed in
the experimental conditions via a consequential choice treatment in which the participants will
be instructed that the decision they make on one of the choice tasks will be binding.

Choice Experiment Design

The choice experiment involves four attributes (cost, risk of shock, risk of brain damage and risk
of death). There were four levels of the risk attributes and seven levels of the cost attribute. We
utilized a main effects design to develop an orthogonal choice set with ten different survey
versions.

Peer Review of Study Design

The overall study design evolved with numerous discussions with water quality specialists and
economists. Several versions of the survey were reviewed by economists that were experts in the
area of contingent valuation and choice experiments.

Key Elements of the Survey Design

The key elements of the choice task involves the information provided the respondent and the
nature of the alternatives before them.

5

-------
Figure 1. Key Elements of the Choice Task Given in the Survey

Section 5 This section contains a choice task for you to complete. We have listed below
some important information, which you may or may not be aware of, about nitrate in
water. Please read this information before you continue.

•S Your community is one of many in Colorado that is at risk for nitrate contamination of its
drinking water.

•S Both public water supplies and private wells can be affected.

•S Because infants do not have fully developed digestive systems, drinking nitrate
contaminated water can have negative effects on infants' health, but it will not affect
adults.

•S Consuming nitrate contaminated drinking water places infants at risk for a condition

called "blue baby syndrome" that is caused by depleting the oxygen in the blood.
•S Symptoms of "blue baby syndrome" include a bluish tint to the infant's skin, shortness of

breath, shock, brain damage, coma, and death.

•S Using bottled water or water that has had the nitrate removed to prepare formula will
eliminate negative health effects caused by nitrate contaminated drinking water for
infants, but will not reduce risks from other sources.

What follows is some information concerning different choices you have to reduce
health risks to infants associated with exposure to nitrate contamination of drinking
water. Please read through the following information and for each pair of options,
choose the option that you feel is best.

* Option B may have other potential benefits in addition to reducing exposure to nitrate.

Option A

Use tap water

Options for Preparing Infant Formula
Option B

Use bottled water

Effects of Over-exposure to Nitrate Contaminated Drinking Water

Risk of Temporary

Risk of Permanent

Cost

Total, one-time
cost of the option
in dollars

Shock

Brain Damage
Risk of infant
experiencing

Risk of Death
Risk of infant
dying

Risk of infant
experiencing

decrease in blood
pressure and a
weak, rapid pulse

damage to the brain

6

-------
Adults with infants were told the following in the Non Consequential Treatment:

In the next part of the survey you will be asked whether you would purchase or not
purchase various amounts of bottled water. This water would help to reduce your infant's
exposure to water with excessive levels of nitrate.

If you purchased the water, the health risks to your childfrom nitrate contaminated
drinking water (as well as other potential drinking water contaminants) would be reduced. The
amount by which these risks would go down for a given amount of water is presented on the
sheet for each choice. Purchasing the bottled water would not reduce risks to your child to zero
because she would still face all of the normal risks that do not come from drinking contaminated
water.

If you would not purchase the water, your child would continue to face the risks
associated with drinking contaminated water (either by drinking the water by itself or by
drinking formula that was prepared with contaminated water). The total risk that your child
wouldface if you chose not to purchase the water is also presented on the sheet for each choice.
You will be asked to make 4 choices in total.

Households without children were told the following in order to allow for investigation into
altruism:

In the next part of the survey you will be asked to imagine (pretend) that you have to
choose between purchasing or not purchasing various amounts of bottled water for a needy
family in your community to help reduce their infant's exposure to water that may contain
excessive levels of nitrate.

If you purchased the water, the health risks to the infant from nitrate contaminated
drinking water (as well as other potential drinking water contaminants) would be reduced. The
amount by which these risks would go down for a given amount of water is presented on the
sheet for each choice.

If you chose not to purchase the water, the infant would continue to face the risks
associated with drinking contaminated water (either by drinking the water by itself or by
drinking formula that was prepared with contaminated water). The total risk that the infant
wouldface if you chose not to purchase the water is also presented on the sheet for each choice.
You will be asked to make 4 choices in total.

7

-------
CONSEQUENTIAL SURVEY TREATMENT

Adults with infants were told the following in the consequential survey treatment.

In the packet containing this survey, you were also given a voucher for $	. In the

next part of the survey you will be asked whether you would purchase or not purchase various
amounts of bottled water. This water would help to reduce your infant's exposure to water with
excessive levels of nitrate.

If you purchased the water, the health risks to your childfrom nitrate contaminated
drinking water (as well as other potential drinking water contaminants) would be reduced. The
amount by which these risks would go down for a given amount of water is presented on the
sheet for each choice. Purchasing the bottled water would not reduce risks to your child to zero
because she would still face all of the normal risks that do not come from drinking contaminated
water.

If you would not purchase the water, your child would continue to face the risks
associated with drinking contaminated water (either by drinking the water by itself or by
drinking formula that was prepared with contaminated water). The total risk that your child
wouldface if you chose not to purchase the water is also presented on the sheet for each choice.

You will be asked to make 4 choices in total. Choosing between Option A and Option B
will allow you to either: actually purchase bottled water for your infant using money provided by
Colorado State University or keep the money that it would take to purchase the water.

At this time, look over the voucher that was attached to your survey. You will see that it is
good for a dollar amount that matches the highest cost given for bottled water on the four choice
tasks. Once you have completed the survey, send the completed survey along with the signed
voucher back to us in the self-addressed postage-paid envelope that we have provided. Once we
have received the surveys and vouchers back, we will randomly select one of your four choices
between A and B in Section 5. If on that particular task you chose "Do Nothing, " you will
receive a check for the full amount listed on the voucher. If, on the other hand, you chose
"Purchase Bottled Water, "you will receive a pre-paid punch-card to obtain the bottled water
from a local grocery store. If the value of the punch-card is less than the dollar amount given on
the voucher, you will be sent a check for the difference.

Adults without infants were told the following in the consequential survey treatment.

In the packet containing this survey, you were also given a voucher for $	. In the

next part of the survey you will be asked whether you would purchase or not purchase various
amounts of bottled water. This water would go to a needy family to help to reduce their infant's
exposure to water with excessive levels of nitrate.

If you purchased the water, the health risks to the childfrom nitrate contaminated
drinking water (as well as other potential drinking water contaminants) would be reduced. The
amount by which these risks would go down for a given amount of water is presented on the
sheet for each choice. Purchasing the bottled water would not reduce risks to the child to zero
because she would still face all of the normal risks that do not come from drinking contaminated
water.

If you would not purchase the water, the child would continue to face the risks associated
with drinking contaminated water (either by drinking the water by itself or by drinking formula
that was prepared with contaminated water). The total risk that the child wouldface if you chose
not to purchase the water is also presented on the sheet for each choice.

8

-------
You will be asked to make 4 choices in total. Choosing between Option A and Option B
will allow you to either: actually purchase bottled water for an infant in a needy family using
money provided by Colorado State University or keep the money that it would take to purchase
the water.

At this time, look over the voucher that was attached to your survey. You will see that it is
good for a dollar amount that matches the highest cost given for bottled water on the four choice
tasks. Once you have completed the survey, send the completed survey along with the signed
voucher back to us in the self-addressed postage-paid envelope that we have provided. Once we
have received the surveys and vouchers back, we will randomly select one of your four choices
between A and B in Section 5. If on that particular task you chose "Do Nothing, " you will
receive a check for the full amount listed on the voucher. If, on the other hand, you chose
"Purchase Bottled Water, " a needy family with an infant will receive a pre-paid punch-card to
obtain the bottled water from a local grocery store. If the value of the punch-card is less than the
dollar amount given on the voucher, you will be sent a check for the difference.

9

-------
ACTUAL CHOICE TASK

For this task, we want you to compare Option A to Option B and choose the option you
would actually pick if you had to pay the cost shown. *Risk information is presented in the
number of infants in your community out of1,000 who will be affected.	

Effects

Option A
Do Nothing

Option B
Buy Bottled Water for an Infant in
Your Household

Cost

$0

$300

Risk of
Temporary
Shock*

100/1000

80/1000

Risk of
Permanent

Brain
Damage*

40/1000

30/1000

Risk of
Death*

9/1000

6/1000

Which option do you choose?

10

-------
Data Collection

The survey was pilot tested with two groups, one English-speaking and one Spanish-speaking, in
the San Luis Valley area of Colorado. Due to pilot results, the survey was revised to decrease its
length and to improve clarity. Data collection was to take place through in-person sessions with
participants conducted at various recruitment sites (day care, childbirth classes, etc.). However
both participants and sites proved reluctant to participate in this manner. As a result, the data
collection methods were altered to include a mail survey mode and "hosted sessions," as well as
recruiting from a broader range of areas in Colorado.

For the mail surveys, the survey packets were sent to five early childhood sites, such as Head
Start, family centers, or preschools. The packets include a self-addressed stamped envelope for
the participants to return the survey. From the time the surveys were mailed to the sites to the
time the first participants picked up surveys was approximately three weeks. Participants
complete a contact sheet when they pick a packet up at the site and the contact sheets are sent
back to the experimenters. Participants are asked to date the slips so that the experimenters know
when to begin the reminder phone calls. Using this survey tracking method, the experimenters
call participants who have not returned the survey within two weeks and remind them to mail
back the survey or send them a new one if necessary. If respondents have simply forgotten to
return the survey, they are reminded to do so. If they have lost the survey and are still interested
in participating, they are mailed another. In another two weeks they are contacted by phone again
and if they don't return the survey, they are counted as a non-respondent and dropped from the
study.

To date, information on hosting a session has been disseminated via word of mouth. Starting
May 15th, fliers for hosted sessions will be given to individuals who attend in person sessions.
For the "hosted" sessions, individuals who are interested in being a host set up a time when they
can meet with any friends, family, or acquaintances who are in the demographic groups of
interest. An experimenter attends and conducts an in-person session with the guests. Individuals
participating in an in-person session received $25 for their participation and those completing the
survey via mail receive $15. In the case of "hosted" sessions, participants receive $25 and the
host receives $5 for each completed survey.

The target number of participants is 280 (see Table 1). To date, data have been collected from 92
individuals. About one third of them in the adults with infants or expecting category and most
have been non-consequential (Non).

Table 1



Expecting

Child(ren)

Child(ren) 1 to

No Children







under 1



3







English-

Non

Con

Non

Con

Non

Con

Non

Con

Speaking



















2

0

20

4

27

6

21

1

Spanish-

Non

Con

Non

Con

Non

Con

Non

Con

Speaking



















0

0

0

5

0

4

0

2

11

-------
RESPONSE RATE

Over the last two months 216 survey packets have been sent to five sites and at least one survey
has been returned from each site for a 100% recruiting site response rate. Of the 216 surveys sent
to the five sites, 55 participants (25%) have completed a contact card. Of those 55, 23 or 42%,
have returned a survey. In addition to the mail recruitment, there have been 2 hosted sessions and
19 surveys have been completed there. There have also been two in-person sessions, both at
family centers in southern Colorado.

Response rates to health surveys tend be lower than other types of valuation surveys. For
example, Dickie and Messman (2004) who did a parental health survey regarding themselves and
their children obtained response of 7.5% of eligible households (those with children). This is on
a par with other health valuation surveys such as Johnson, et al. (1997) obtained about 8.8%. So
our response rate to date is on a par with these other surveys.

ECONOMIC MODEL RESULTS

Table 2 provides the basic economic model that focuses primarily on the cost and risk reduction
variables.

Table 2 Logistic Regression of the Binary Choice Model

Dependent Variable: YESPAY

Method: ML - Binary Logit (Quadratic hill climbing)

Date: 05/08/06 Time: 21:03

Sample(adjusted): 3 370

Included observations: 363

Variable

Coefficient

Std. Error

z-Statistic

Prob.

CONSTANT

-0.610367

0.503933

-1.211206

0.2258

COST

-0.010853

0.002351

-4.616077

0.0000

HYPCOSTDUM

0.009256

0.001889

4.899394

0.0000

SHOCK RISK REDUC

0.028697

0.010585

2.711134

0.0067

BRAIN DAM RR

0.056937

0.024426

2.331006

0.0198

DEATH RISK REDUC

0.026172

0.081679

0.320430

0.7486

INFANT

0.530914

0.271195

1.957684

0.0503

Mean dependent var

0.707989

S.D. dependent var

0.455315

S.E. of regression

0.431550

McFadden R-squared

0.094395

Sum squared resid

66.29976







Log likelihood

-198.5367 LR statistic (6 df)



41.38877

Restr. log likelihood

-219.2311 Probability(LR stat)

2.43E-07

Obs with Dep=0

106

Total obs



363

Obs with Dep=1

257







Where:

Cost is the one time cost to you.

12

-------
HypCostDum is whether the survey is hypothetical-consequential dummy variable (Hypothetical
equals 1) times the one time Cost.

Shock Risk Reduc is the reduction in risk of shock to your child
BrainDamRR is the reduction in risk of brain damage
Death Risk Reduc is the reduction in risk of death to your child

Infant is whether the respondent has an infant (ages 0-1) that would be at risk from drinking
water with nitrates in it.

Note that the one time cost is negative and statistically significant at the 1% level. However, the
HypCostDum is positive and significant. Thus, when the cost is hypothetical (not actual or
consequential), then the net or overall price coefficient becomes much less price sensitive,
although still negative suggesting that the higher the price the less likely households are to
purchase the risk reduction through bottled water. The difference in the real cash cost coefficient
and the hypothetical cost coefficient, provides results of our hypothesis test regarding whether
there is a statistical difference in responses of people facing a hypothetical cost and an actual
cost. There is quite a difference, with households facing the hypothetical cost being much less
sensitive to the cost than households that face an actual monetary opportunity cost. For purposes
of comparing marginal values calculated using the actual monetary cost versus the hypothetical
cost treatment, we set the HypCostDum to one for hypothetical and adding its coefficient to the
Cost coefficient results in a net Cost coefficient of -.001597. Thus to calculate marginal values
for the real cost, we divide the attribute coefficient by Cost variable of -.010853, while for the
hypothetical cost we use the -.001597.

The positive signs on Brain Damage Risk Reduction, Shock Risk Reduction and Death Risk
Reduction make sense. People are willing to pay more the greater the reduction in risk of shock
and brain damage is provided by using bottled drinking water. However, the Death Risk
Reduction coefficient is not statistically significant and therefore we will not calculate marginal
values for this coefficient.

The coefficient on Infant is positive and statistically significant, indicating individuals with an
infant in their household are more likely to pay, than those without.

Calculating Marginal Values of Risk Reduction

Marginal Value is Shock or Brain damage risk reduction coefficient divided by the absolute
value of the cost coefficient. It is the willingness to pay to reduce shock or brain damage by 1 per
1000 infants. Performing such calculations with our data yields the following results.

A household would pay $2.64 in the real cash treatment and $18 in the hypothetical treatment for
bottled water that would result in a .0001 (1 in a thousand) reduction in the chances of an infant
going into shock from nitrate in water. A household would pay $5.25 in the real cash treatment
and $36 in the hypothetical treatment for bottled water that would result in a .0001 (1 in a
thousand) reduction in the chances of an infant experiencing permanent brain damage from
nitrate in water.

Dividing the coefficient on Infant by the cost coefficients allows us to calculate to investigate the
extent of altruism of households without children in terms of their willingness to pay to buy

13

-------
bottled water for households with infants. While WTP rises by $49 with real money and $332 for
hypothetical payment for households with infants at risk, WTP is still positive for households
without an infant. This suggests there is some measure of altruism reflected in our WTP results.

Comparison of Economic Results to Results with the Theory of Planned Behavior

Table 3. Logistic Regression of the Binary Choice Model with Theory of Planned Behavior
Variables.

Dependent Variable: YESPAY

Method: ML - Binary Logit (Quadratic hill climbing)

Date: 05/08/06 Time: 21:05

Sample(adjusted): 3 370

Included observations: 343

Variable

Coefficient

Std. Error

z-Statistic

Prob.

C

-2.204187

1.452501

-1.517511

0.1291

COST

-0.012265

0.002641

-4.644250

0.0000

HYPCOSTDUM

0.010640

0.002184

4.870890

0.0000

SHOCKRISKREDUC

0.022966

0.011058

2.076906

0.0378

BRAIN DAM RR

0.038686

0.026311

1.470324

0.1415

DEATHRISKREDUC

0.015765

0.085622

0.184118

0.8539

INFANT

0.749895

0.301004

2.491310

0.0127

HEALTH ATTITUDES

0.149765

0.212629

0.704350

0.4812

HEALTH PERCEIVED

0.263524

0.199599

1.320263

0.1867

CTRL









WATER PERCEIVED

0.461788

0.312784

1.476383

0.1398

CTRL









WATER NORMS

-0.240546

0.167906

-1.432628

0.1520

Mean dependent var

0.720117

S.D. dependent

var

0.449598

S.E. of regression

0.420073

McFadden R-squared

0.121784

Sum squared resid

58.58504







Log likelihood

-178.5810 LR statistic (10 df)

49.52848

Restr. log likelihood

-203.3452 Probability(LR stat)

3.26E-07

Obs with Dep=0

96

Total obs



343

Obs with Dep=1

247







Health Attitudes is positive or negative evaluation of health-related behaviors.

Health Perceived Control is perceived control over means of reducing risks to infant health.

Water Perceived Control is perceived control over drinking water safety.

Water Norms is subjective norms for being concerned about drinking water quality.

14

-------
Evaluation of the Contribution of the Theory of Planned Behavior Variables

Comparison of the McFadden R square of the standard economic model at .094 and the . 12
McFadden R square of the full model with the inclusion of the Theory of Planned Behavior
suggests these Planned Behavior variables add a small amount of explanatory power (roughly
2.5%). Health Attitudes variable was not significant (p=.48) and Health Perceived Control is
significant at the 19% level. The Water Perceived Control and Water Norms were significant at
the 14% and 15% levels, respectively.

The Health Attitudes items were scored so that a high score indicates an orientation toward
viewing infant health issues as a community problem. Health Perceived Control items were
scored so that a high score indicates a strong feeling that one has control over keeping infants
free from harm caused by environmental contaminants. Water Perceived Control items were
scored such that a high score indicates a strong feeling of personal control over drinking water
quality. Water Norms items were scored such that high score indicates a strong subjective norms
for being concerned about drinking water quality.

Conclusions

The results support the first hypothesis, indicating that respondents' WTP was negatively
correlated with one time cost for bottled water and positively correlated with risk reduction. The
second hypothesis was not supported, with respondents in the consequential treatment being
more cost sensitive than respondents in the hypothetical treatment. The third hypothesis was
partially supported with TPB components accounting for a very small amount of the variance.

The fact that respondents were willing to pay more in the hypothetical treatment than in the
consequential treatment makes sense and indicates that for such choices there is a hypothetical
bias. The data indicate that individuals who believe infant health is a community issue and have
a high degree of perceived control over both infant health issues and water quality issues are
more likely to choose to purchase the bottled water, which makes intuitive sense. On the other
hand, individuals who perceive strong subjective norms for being concerned about water quality
were less likely to choose the bottled water option. It is possible that such individuals feel that
the norms are extreme and to a certain extent are reacting against them.

It is hoped that with a complete data set the TPB components will have better explanatory power.
A complete data set will also allow more in-depth analyses, including testing for differences
between English-speaking and Spanish-speaking participants and a more detailed test of the
differences between the different demographic groups (expecting parents, parents of infants,
parents of children 1-3 years old, and adults with grown children or no children). The difficulties
initially encountered with participant recruitment are informative. Despite the offer of
compensation, both sites and participants were generally unwilling to participate in in-person
data collection sessions. Sites were much more receptive to distributing mail surveys and the
response rate for this targeted quasi-mail was much higher than that obtained in previous
research on this topic.

15

-------
References

Agee, M. and T. Crocker. 2002. On Techniques to Value the Impacts of Environmental Hazards
on Children's Health.

Carson, Richard, Theodore Groves and Mark Machina. 2000. Incentive and Information
Properties of Preference Questions. Dept. of Economics, University of California-San Diego.

Dickie, M. and V.L. Messman. 2004. Parental Altruism and the Value of Avoiding Acute Illness:
Are Kids Worth More than Parents? Journal of Environmental Economics and Management 48:
1146-1174.

Dickie, M. and S. Gerking. 2003. Parent Valuation of Latent Health Risks to their Children. In J.
Wessler, H. Weikard, R. Weaver, Eds. Risk and Uncertainty in Environmental Economics. Elgar
Publishing. Pages 251-278.

Hanemann, M. 1984. Welfare Evaluations in Contingent Valuation Experiments with Discrete
Responses. American Journal of Agricultural Economics 66(3): 332-341.

Holmes, T. and V. Adamowicz. 2003. Attribute Based Methods. In Champ, P., K. Boyle and T.
Brown. A Primer on Non Market Valuation. Kluwer Academic Publishers.

Johnson, R., M.R. Banzhaf and W. Desvousges. 2000. Willingness to pay for Improved
Respiratory and Cardiovascular Health. Health Economics 9: 295-317.

Liu, J-T, J.K. Hammitt, J-D. Wang, J-L Liu. 2000. Mother's Willingness to Pay for Own and Her
Child's Health: A Contingent Valuation Study in Taiwan. Health Economics 9: 319-326.

Roe, B., K. Boyle, and M. Teisl. 1996. Using Conjoint Analysis to Derive Estimates of
Compensating Variation. Journal of Environmental Economics and Management 31: 145-159.

U.S. Environmental Protection Agency. 2003. Children's Health Valuation Handbook. EPA 100-
R-03-003. Washington DC.

Viscusi, K., W. Magat, and J. Huber. (1987). An Investigation of the Rationality of Consumer
Valuations of Multiple Health Risks. Rand Journal of Economics 18: 465-479.

16

-------
APPENDIX A - Sample Mail Survey

Valuation of Infant Health Survey Directions

The survey you are going to be completing contains questions concerning water quality, infant
health, nitrate, environmental attitudes, and some demographic questions such as age and gender.
Part of the survey will also ask you to make a series of choices between two different options for
averting risks to infant health that are associated with unsafe levels of nitrate in drinking water.

Please answer all the questions honestly. There are no right or wrong answers to any of the
questions. We are only interested in your opinion and attitudes. Your responses will be
completely confidential. Even though we have your names, they will not be associated with your
responses in any way.

Please feel free to contact Helen Cooney at (970) 491-2119 if you have any questions. Your
participation is voluntary and you may quit at any time without any negative consequences.

Please remember to return a signed copy of our informed consent form along with your survey.

Thank you for your participation!

17

-------
Section 1 This section asks some general questions about you and your drinking water.

Note: "Tap water" means water that comes out of the faucet in you kitchen.

1)	How long have you lived in	County, Colorado?	

2)	a) Overall, how would you rate the taste of your tap water?

O Poor O Below Average O Average O Above Average O Excellent

b)	Overall, how would you rate the smell of your tap water?

O Strong unpleasant smell O Somewhat unpleasant smell O Noticeable smell O No smell

c)	Overall, how would you rate the appearance of your tap water?

O Colored (brown, red, yellow) O Very Cloudy O Cloudy O Slightly cloudy O Clear

d)	Overall, how would you rate the safety of your tap water?

O Poor O Below Average O Average O Above Average O Excellent O Don't Know

3)	List any problems that you think your tap water has.

4)	Do you use a water filter system at home to purify your tap water?

O Always O Often	O Sometimes O Never (Go to question 5)

If you use a filter system in your home, what type is it?

O Filter Pitcher O Faucet Mounted O Under-sink O Refrigerator

5)	How much money do you spend on each of the following over the course of a typical month?

Bottled Water (for use at home only)

O None O $1-$10 O $11-24 O $25-$49 O $50 or more
Filter System at home (maintenance or replacement filters)

O No System	O Less than $25	O More than $25

6)	Does the water in your home come from a well on your property?

O Yes	O No (if "No" skip to question 7)

6a) Do you have your well water tested?

O Yes	O No (if "No" skip to question 7)

6b) How often do you have your well-water tested?

O Once a year O Once every two years	O Every five years

6c) Does your well water meet standards when tested?

O Yes	O No

18

-------
7)	Check any of the items below that you think can be a source of nitrate contamination in drinking
water.

O Fertilizer Runoff O Natural Deposits O Decaying Plant Matter
O Fossil Fuels	O Sewage	O Landfill Runoff

O Steel Factories	O Discharge from Coal-burning Factories

O Leaching from Ore-processing Sites	O Leaching from Septic Tanks

8)	Check any of the items that you think can help you avoid drinking water with high levels of nitrate.
O Under-sink Filter O Faucet-mounted Filter O Filter Pitcher (e.g., Brita™ filters)

O Bottled Water O Boiling Tap Water

9)	Have you heard about the quality of your community's drinking water?

O Yes	O No

10)	Do you read the water quality information included in your water bill?

O Always O Sometimes	O Never O Don't receive a water bill

11)	Do you prepare formula for an infant (a child under one year old)?

0	Yes	O No (if "No" skip to question 12)

11a) How old is the infant?	

1	lb) Do you use bottled water to prepare infant formula?

O Always O Often	O Sometimes O Never

12)	Have you or a woman in your household been pregnant in the last three years?

O Yes	O No (if "No" skip to question 13)

12a) While pregnant, how often did you or a woman in your household buy bottled water to drink
at home?

O Always O Often O Sometimes O Never
12b) While nursing, how often did you or a woman in your household buy bottled water to drink
at home?

O Always O Often O Sometimes O Never O Didn't Nurse

13)	Do you have health insurance?

O Yes	O No (If no, skip to question 14)

13 a) Does your insurance cover emergency room care?

O Yes	O No

13b) Is your family (spouse and/or children) covered?

O Yes	O No

19

-------
14) If you have children, how much does a visit to the doctor for your child usually cost you?
o $0 o $5-$20 O $21 - $30 O$31-$50 O$51-$70

O $71 - $90 O $91-$100 O$100 +

14a) Does an adult in your household have to miss work in order to take a child to the doctor or
hospital?

O Yes	O No

20

-------
Section 2 This section asks about your beliefs regarding infants' health (consider infants to be
children under 1 year of age).



Strongly
Agree

Agree

Don't
Know

Disagree

Strongly
Disagree

1) If drinking water is safe for
adults, it is also safe for infants.

O

o

O

o

O

2) If infants consume water
contaminated with nitrate, it can be

o

o

o

o

o

harmful to their health.











3) If adults consume water
contaminated with nitrate, it can be

o

o

o

o

o

harmful to their health.











4) It is natural for infants to become

o

o

o

o

o

ill more often than adults.











5) The infants in my community are
never ill due to pollution.

o

o

o

o

o

6) My friends and family are
concerned with infants' health

o

o

o

o

o

issues.











7) The parents I know are worried
about the health of their infants.

o

o

o

o

o

8) It is possible to reduce the
exposure infants have to pollution.

o

o

o

o

o

9) It is possible to prevent infants
from becoming seriously ill due to
environmentally caused illnesses.

o

o

o

o

o

10) Only people with infants living
in their home need to be concerned

o

o

o

o

o

about pollution.











11) Parents, not the public, have the
sole responsibility for protecting
their infants from harm.

o

o

o

o

o

12) More state and community
resources need to be devoted to

o

o

o

o

o

infant health issues.











13) There is too much emphasis
placed on issues regarding infants'
health.

o

o

o

o

o

If you are NOT currently caring for an infant, skip to question 1 of Section 3.

14) My infant(s) are not exposed to
dangerous environmental

o

o

o

o

o

contaminants.











15) I can ensure that my infant(s)
do not become ill due to

o

o

o

o

o

environmental contaminants.











16) I can afford to take my infant(s)
to the doctor when they are ill.

o

o

o

o

o

21

-------


Strongly
Agree

Agree

Don't
Know

Disagree

Strongly
Disagree

17) I can prevent my infant(s) from
becoming seriously ill.

O

O

O

o

O

Section 3 This section asks what you think about the quality of your drinking water. Please fill in



Strongly
Agree

Agree

Don't
Know

Disagree

Strongly
Disagree

1) My community has safe drinking
water.

O

o

O

o

O

2) My home's drinking water
(straight from the faucet) does not
have unsafe levels of nitrate.

o

o

o

o

o

3) My friends and family are
worried about our drinking water.

o

o

o

o

o

4) Most of the people I know would
take steps to ensure that their
drinking water is safe.

o

o

o

o

o

5) Nitrate in drinking water is an
unavoidable occurrence.

o

o

o

o

o

6) It is important to me to test the
quality of my home's drinking
water.

o

o

o

o

o

7) It is the government's
responsibility to ensure that my
drinking water is safe.

o

o

o

o

o

22

-------
Section 4 We are now going illustrate some risk information for you to help you get used to the
way in which risk information is presented as pie charts. Please read the information and then
choose which chart represents the greatest risk.

In the first example, the gray pie wedge represents the fraction or proportion of 1000 accidents which
involve Car A and Car B. The larger the gray slice, the greater the risk. As long as the bottom numbers in
the fractions (as in this case, 1000) are the same, the larger the top number, the larger the risk.
1) The following charts represent the risk (in number of accidents out of 1000) of being involved in a
fatal car crash in two different types of car.

Car A

150/1000

Car B

60/1000

Which car poses the greatest risk?

2) The following charts represent the risk (in number of park visitors out of 1000) of being attacked by a
mountain lion in two different national parks.

Park A

15/1000

Park B

6/1000

Which park poses the greater risk?

1)	The correct answer is A. The top number for A (150) is greater than the top number
for B (60).

2)	The correct answer is A. The top number for A (15) is greater than the top number
for B (6).

23

-------
Section 5 This section contains a choice task for you to complete. We have listed below some
important information, which you may or may not be aware of, about nitrate in water. Please read
this information before you continue.

•S Your community is one of many in Colorado that is at risk for nitrate contamination of its drinking
water.

•S Both public water supplies and private wells can be affected.

•S Because infants do not have fully developed digestive systems, drinking nitrate contaminated

water can have negative effects on infants' health, but it will not affect adults.

•S Consuming nitrate contaminated drinking water places infants at risk for a condition called "blue

baby syndrome" that is caused by depleting the oxygen in the blood.

•S Symptoms of "blue baby syndrome" include a bluish tint to the infant's skin, shortness of breath,

shock, brain damage, coma, and death.

•S Using bottled water or water that has had the nitrate removed to prepare formula will eliminate
negative health effects caused by nitrate contaminated drinking water for infants, but will not
reduce risks from other sources.

What follows is some information concerning different choices you have to reduce health risks
to infants associated with exposure to nitrate contamination of drinking water. Please read
through the following information and for each pair of options, choose the option that you feel is
best.

Option A

Use tap water

Options for Preparing Infant Formula
Option B

Use bottled water

* Option B may have other potential benefits in addition to reducing exposure to nitrate.

Effects of Over-exposure to Nitrate Contaminated Drinking Water

Risk of Temporary

Risk of Permanent

Cost

Total, one-time
cost of the option
in dollars

Shock

Brain Damage
Risk of infant
experiencing

Risk of Death
Risk of infant
dying

Risk of infant
experiencing

decrease in blood
pressure and a
weak, rapid pulse

damage to the brain

24

-------
CR

In the packet containing this survey, you were also given a voucher for $	. In the next part of

the survey you will be asked whether you would purchase or not purchase various amounts of bottled
water. This water would help to reduce your infant's exposure to water with excessive levels of nitrate.

If you purchased the water, the health risks to your child from nitrate contaminated drinking water
(as well as other potential drinking water contaminants) would be reduced. The amount by which these
risks would go down for a given amount of water is presented on the sheet for each choice. Purchasing the
bottled water would not reduce risks to your child to zero because she would still face all of the normal
risks that do not come from drinking contaminated water.

If you would not purchase the water, your child would continue to face the risks associated with
drinking contaminated water (either by drinking the water by itself or by drinking formula that was
prepared with contaminated water). The total risk that your child would face if you chose not to purchase
the water is also presented on the sheet for each choice.

You will be asked to make 4 choices in total. Choosing between Option A and Option B will allow
you to either: actually purchase bottled water for your infant using money provided by Colorado State
University or keep the money that it would take to purchase the water.

At this time, look over the voucher that was attached to your survey. You will see that it is good
for a dollar amount that matches the highest cost given for bottled water on the four choice tasks. Once
you have completed the survey, send the completed survey along with the signed voucher back to us in
the self-addressed postage-paid envelope that we have provided. Once we have received the surveys and
vouchers back, we will randomly select one of your four choices between A and B in Section 5. If on that
particular task you chose "Do Nothing," you will receive a check for the full amount listed on the
voucher. If, on the other hand, you chose "Purchase Bottled Water," you will receive a pre-paid punch-
card to obtain the bottled water from a local grocery store. If the value of the punch-card is less than the
dollar amount given on the voucher, you will be sent a check for the difference.

25

-------
For this task, we want you to compare Option A to Option B and choose the option you
would actually pick if you had to pay the cost shown.

*Risk information is presented in the number of infants in your community out of 1,000 who will be
affected.

Effects

Option A
Do Nothing

Option B
Buy Bottled Water for an Infant in
Your Household

Cost

$0

$300

Risk of
Temporary
Shock*

100/1000

80/1000

Risk of
Permanent

Brain
Damage*

40/1000

30/1000

Risk of
Death*

9/1000

6/1000

Which option do you choose? 	

Why did you choose that option?

26

-------
For this task, we want you to compare Option A to Option B and choose the option you
would actually pick if you had to pay the cost shown.

*Risk information is presented in the number of children in your community out of 1,000 who will be
affected.

Effects

Option A
Do Nothing

Option B
Buy Bottled Water for an in Your
Household

Cost

$0

$450

Risk of
Temporary
Shock*

100/1000

60/1000

Risk of
Permanent

Brain
Damage*

40/1000

20/1000

Risk of
Death*

9/1000

3/1000

Which option do you choose?	

Why did you choose that option?

27

-------
For this task, we want you to compare Option A to Option B and choose the option you
would actually pick if you had to pay the cost shown.

*Risk information is presented in the number of children in your community out of 1,000 who will be
affected.

Effects

Option A
Do Nothing

Option B
Buy Bottled Water for an Infant in
Your Household

Cost

$0

$400

Risk of
Temporary
Shock*

100/1000

60/1000

Risk of
Permanent

Brain
Damage*

40/1000

30/1000

Risk of
Death*

9/1000

6/1000

Which option do you choose?	

Why did you choose that option?

28

-------
For this task, we want you to compare Option A to Option B and choose the option you
would actually pick if you had to pay the cost shown.

*Risk information is presented in the number of children in your community out of 1,000 who will be
affected.

Effects

Option A
Do Nothing

Option B
Buy Bottled Water for an Infant in
Your Household

Cost

$0

$500

Risk of
Temporary
Shock*

100/1000

80/1000

Risk of
Permanent

Brain
Damage*

40/1000

20/1000

Risk of
Death*

9/1000

3/1000

Which option do you choose?	

Why did you choose that option?

29

-------
Section 6 This section asks for some general demographic information. Your responses will be
confidential. No information about your identity (name, SSN, etc.) will be associated with your data.

Only researchers on this project will have access to your data.

1)	Age	

2)	What is your gender? O Male	O Female

3)	Occupation	

4)	Number of Years of Schooling:	

5)	Ethnicity (Check all that apply)

O African American

O American Indian
O Asian American
O European American
O Hispanic/Latino
O Native Hawaiian/Pacific Islander
O Other (	)

6)	Do any of your children (under the age of 18) live in your community?

O Yes	O No	O I have no children.

7)	Do any of your grandchildren (under the age of 18) live in your community?

O Yes	O No	O I have no grandchildren.

8)	Do any of your nieces or nephews (under the age of 18) live in your community?

O Yes	O No	O I have no nieces or nephews.

9)	Yearly Household Income from all Sources

O $0 - $10,000	o $10,001 - $20,000 o $20,001 - $30,000	O $30,001 - $40,000

O $40,001 - $50,000 O $50,001 +

30

-------
Appendix B - Sample Risk Group Voucher

Colorado State University
Family and Youth Institute Study on Valuation of Infant Health

Voucher

$250

Sign this voucher where indicated and return with your completed survey. Once the
choice has been randomly selected, you will be sent one of three things:

—A check for the full amount of this voucher (you chose "Do Nothing" on the
selected choice)

—A pre-paid punch-card for bottled water worth the dollar amount listed as the
cost for the choice (you chose "Purchase Bottled Water" and the randomly selected choice
was the one with the highest dollar amount)

—A pre-paid punch-card for bottled water worth the dollar amount listed as the
cost for that choice and a check to make up the difference between the worth of the punch
card and the amount listed on this voucher (you chose "Purchase Bottled Water" and the
randomly selected choice was not the one with the highest dollar amount)

Staff Signature	Participant Signature

31

-------
Economic Valuation of Avoiding
Exposure to Arsenic in Drinking
Water

Kathleen P. Belli
University of Maine
Kevin 3. Boyle

Virginia Polytechnic Institute

U.S. EPA Morbidity and Mortality Workshop
April 10-12, 2006

Research Team

¦	Kelly Maguire (US EPA)

¦	Andrew E. Smith (Maine Bureau
of Health)

¦	Laura Taylor (Georgia State
University); Tom Crocker
(University of Wyoming); Anna
Alberini (University of Maryland)

-------
Research Objectives

¦ Economic Valuation of Avoiding
Exposure

-	Scrutinize behavioral response of
households to information regarding levels
of arsenic in private wells

¦	private actions at home

¦	transactions of residential properties

-	Examine public support for government
programs aimed at reducing arsenic levels
in drinking water

¦	coverage (public and private water supplies)

¦	level of reduction

Central Research Questions

¦	What will be the relationships among
valuation estimates derived using
different valuation methods?

-	averting behavior

-	hedonic property value

-	hybrid conjoint / contingent valuation

¦	Do household composition and location
factors influence behavioral responses?

-	children, age, gender, health status

-	household location - proximity to arsenic
"cluster" areas

-------
Multiple Valuation Methods

¦	Revealed Preference

¦	Hedonic Property Value

¦	Averting Behavior

¦	Stated Preference

¦	Hybrid Conjoint / Contingent Valuation

5

Study Area: Maine

¦	Upwards of 50 % of Maine
Households Rely on Private Wells
for Drinking Water

¦	Assessment of Risks (Loiselle,
Marvinney, and Smith 2001)

-	10% exceed 10 micrograms per liter

-	6% exceed 20 micrograms per liter

-	2% exceed 50 micrograms per liter

-------
Arsenic concentrations inatleast

Source: Citation—

Ryker, S.J., Nov. 2001,	: Geotimes v.46 no.11, p.34-36.

Accessible at:: http://webserver.cr.usgs.gov/trace/arsenic/.

Sample Selection

¦	Town Sample

-	1,000 randomly selected households from
arsenic "cluster" towns

¦	Buxton, Hollis, North port, Standish

¦	State Sample

-	1,000 randomly selected households

¦	split - general population (500) versus private
well / prior arsenic test (500)

¦	Property Sample

-	Sales data from arsenic "cluster" towns

-------
Comparative Approach

¦	Samples

-	Town Sample

¦	averting behavior

¦	hybrid conjoint / contingent valuation

-	State Sample

¦	hybrid conjoint / contingent valuation

-	Property Sample

¦	hedonic property value

¦	Permits joint estimation

¦	Facilitates comparison and contrast of
valuation estimates

Relevant Literature

¦	Hedonic Property Value Studies

¦	Contamination of Private Wells (McCormick 1997;
Ma lone and Barrows 1990)

¦	Health Risks/Stigmas (Gayer et al., 2000; Gayer et
al. 2002; McCluskey and Rausser 2001; Kiel 1995;

Kiel and McClain 1995)

¦	Conjoint and CV Studies of WTP for State
Programs

¦	Safe Drinking Water Supplies (Boyle et al. 1994;
Edwards 1986; Bergstrom et al. 2001; Poe et al.

2001)

¦	Averting Behavior

¦	Gerking and Stanely 1976; Dickie and Girkie 1991;
Shogren and Crocker 1999; Abdalla 1994; Abdalla et
al. 1992; Bartik 1998	10

-------
Relevant Literature (continued)

¦	Health and Risk Communication

¦	Lead in Tap Water (Griffin and Dunwoody 2000)

¦	Risk Communication (Fischhoff 1995; Slovic 1987;
NRC 1989; Covello et al. 1989)

¦	Environmental and Health Economics

¦	Hazardous Waste Sites (Gayer, Hamilton, and
Viscusi 2002; Gayer, Hamilton, and Viscusi 2000)

¦	Smokers (Smith et al. 2001)

¦	Radon (Smith and Johnson 1988; Smith and
Desvousges 2001)

¦	Chemical Industry Workplace (Viscusi and
O'Connor 1984)

11

Hedonic Property Value Study



¦ Objective



- Examine evidence of impacts on



property values of arsenic levels



¦ "elevated"



¦ spatial spillovers



¦ Valuation estimates



- Marginal WTP to avoid exposure

12

-------
Averting Behavior Study

¦	Objective

-	Examine evidence of relationships between
averting expenditures/decisions and
potential causal factors

¦	household composition

¦	household location

¦	arsenic level in drinking water

¦	Valuation estimates

-	WTP to avoid exposure

-	Value of a statistical life

-	Value of a statistical cancer

Hybrid Conjoint / Contingent
Valuation Study

¦	Objective

- Examine evidence of relationships between
support for State Programs and potential
causal factors

¦	household composition

¦	household location

¦	arsenic level in drinking water

¦	household drinking water source

¦	program coverage

- private wells, public supplies, both private and
public

¦	program scope (level of protection)

¦	Valuation estimates

	- WTP for State Programs	

-------
Progress

¦	Hedonic Property Value Study V

¦	Averting Behavior Study

-	Focus Group V

-	Survey Design/Approval V

-	Survey Implementation

-	Analysis

¦	Hybrid Conjoint / Contingent Valuation
Study

-	Survey Design/Approval V

-	Survey Implementation

-	Analysis

¦	* Risk Communication Study V

Results

¦	Hedonic Property Value Study

-	Devanney(2005)

¦	Risk Communication - Aggregate
Analysis of Household Testing
Decisions

-	Huang (2005)

¦	Focus Group Research on Averting
Behavior

-------
Hedonic Results
(Devanney 2005)

¦	Sample

-	Buxton and Hollis
-1991 to 2003

-	2,212 transactions

¦	Arsenic level

¦	Other explanatory variables

-	acreage, structures (age, sqft), time

Measurement of Arsenic

¦	continuous or discrete

-	elevated levels (> 50 ppb)

¦	property and test

-	1 to 1 correspondence

-	"closest" test

-	average test result within a radius of
1/4 mile, 1/2 mile, or 1 mile

-------
Estimated Parameters



(Arsenic Variables)



¦ 1 to 1 correspondence



- insignificant (Buxton)



- significant (0.1) and negative (Hollis)



¦ Closest test result > 50 ppb



- significant and negative (Buxton)



- insignificant (Hollis)



¦ Average test result in buffer



- 1/4 mile



¦ significant and negative (Buxton)



¦ insignificant (Hollis)



- 1/2 and 1 mile



¦ insignificant

19

Marginal WTP (Devanney 2005)

!

T able* 4.6 .** Estimated' implicit- prices* of* arsenic* concentrations- in* groundwater .f



Buxton:!

HoIUsk

Variables

Marginal- implicit' price- ($)•:¦:



MLk

OLSk

MLk

OLSk

ASHE1L

-254.15k

-334.03s

-1.4P7.37*®

-1,134.41k

ASSQk

-400.73**13

-337.fi7*a

-62.92k

-26.60k

AS50AVGQT"

-216.7d*cs

-195.68**a

-200.91s

-125.83k

AS5QAV GHALFk

-31.08k

-36.25k

-91.63k

29.61k

ASSOAVGONEa

-16.07si

-19.19k

-158.39k

-278.3'7k

Notes-'-* **»**, * • denotes - sigiific anc e -at-the -0.01, -0.0 5, - and-0.1 -level, -re sp e ctiveiy. • • Ars enic •
c oncenfcrations • are -me asured-in-p arts v er billionfppb") .U

-------
Sales Price Effects
(Devanney 2005)

Table- 4.7.- Sales-price-effect- of arsenic- contamination- of groundwater.!

, Variables:!

A-in-Arsenic-

Buxtoni::

Hollis:

concentration"

Sales-price- effect- ($)<



(ppb)"

ML=i

OLS=i

MLh

OLS=s

.ASHETLh

90-to-50<:s

6,318.21!=

7,715.17=

15,764.57=

15,422.50=



50-tol0a

7,064.88=

8,935.10=

29,131.23=

24,557.76=

AS50=s

90-to-50<:s

9,049.29=

7,430.47=

1,664.40=

732.56=



50-tol0a

10,791.36=

8,619.16=

1,707.89=

740.58=

AS50AVGQT=s

90-to-50<:s

5,508.58c

5,044.03=

5,403.47=

3,479.40=



50-tol0a

6,059.81=

5,497.01=

5,867.52=

3,663.66=

AS50AV GHALFo

90-to-50<:s

887.21=

1,029.54=

2,695.1 1=

-936.61=



50-tol0a

900.06=

1,046.06=

2,797.57=

-925.32=

AS50AV GONEes

90-to-50<:s

465.32=

550.42=

4,264.64=

3,315.02=



50-to-10=i

469.29=

555.75=

4,550.82=

3,482.20=

H

Risk Communication Study
(Huang 2005)

¦ Panel Count Regression Model of
Annual Arsenic Test Requests by
"town"

- Household-level Data on Tests

¦	Maine State Testing Lab (HETL)
. 1990 - 2003

¦	Final sample size of 16,854 tests(private
residential) over 520 towns

-------
Mean Test Result

Median Test Result

Maximum Test Result

| No Data
|Less than 10
10 to 50
PM Greater than 50

| No Data
| Less than 10

¦	10 to 50

¦	Greater than 50

| | No Data
| Less than 10
D 10 to 50
¦ Greater than 50

Descriptive Statistics -Private
Residential Tests (1990-2003)

Explanatory Variables

¦	Demographic Characteristics

-	Census of Population and Housing
(1990 and 2000)

. WELLS, GENDER, CHILDREN,
EDUCATION, INCOME

¦	Newspaper Coverage of Arsenic in
Drinking Water

-	Number of total articles
-"town" referenced in articles

-------
Results (Huang 2005)

¦	Newspaper coverage (general and
town-specific)

-	positive and significant

¦	Household Composition (at "town"
level)

-	education, gender, and income

¦	positive and significant

-	proportion of town under age 3 and age 17

¦	negative (?) and significant

Focus Group (Averting Behavior)

¦	Joint Production

¦	Uncertainty / Misinformation
-treatment methods in place

¦	Share Information with Neighbors

-------
Reflections on Current Results



¦ Hedonic Property Value Analysis



- further exploration of measurement of



arsenic concentration



- past mitigation



- timing of sample



¦ Household test decisions



- role of test



- sample selection



- perverse incentive - disclosure





27

EXTRA / BACKGROUND SLIDES

-------
Background - Arsenic in Drinking
Water: Federal Policy

¦	1976 SDWA

-	MCL of 50 micrograms per liter (1942)

¦	1999 NRC Report

-	Proposed MCL of 5 micrograms per liter (5
ppb)

¦ Evaluated 3, 5,10, and 20

¦	2001 SDWA

-	MCL of 10 micrograms per liter (10 ppb)

-	Public Water Supply Systems Must Comply
by 2006

Health Effects (NRC 2001)



¦ Cancer effects



¦ skin, bladder, lung



¦ Non-cancer effects



¦ diabetes, high blood pressure



¦ * adverse reproductive outcomes,



respiratory effects





30

-------
Variation in Exposure to Arsenic
in Drinking Water

¦ Ground water sources of drinking
water

-	Public Water Supply Systems

¦	100 million persons (2000)

-	Private Wells*

¦	40 million persons (2000)

-------
U.S. EPA NCER/NCEE Workshop
Morbidity and Mortality: How Do We Value the Risk of Illness and Death?

Washington, DC
April 10-12, 2006

Session VI: Valuing Morbidity and Mortality: Drinking Water

"Perceived Mortality Risks and Arsenic in Drinking Water: Preliminary Research"
Presenter: Douglass Shaw, Texas A&M University

Dr. Shaw opened by stating that he had "way too much to say and was going to just
launch right in." He went on to put a disclaimer on the upcoming discussant comments,
admitting that what he was about to present had evolved since the submission of the
paper and, therefore, there might be little correlation between the two. In anticipation of
"running out of time" and perhaps not getting to his scripted wrap-up to the presentation,
he set out the following summary in advance.

"Here's what we're trying to do—I think it's way different than anything you've heard at
this conference. There's been a little talk about what to do with people who look like
they're irrational, or what do you do with people who don't get the probabilities, and that
kind of thing. There's a sense that you either ignore that and you don't know about it at
all or that you throw those people out [of your study]. We're not going to do either one
of those things. What we're trying to do is to really bridge the gap between what the
decision theorists and the psychologists have been saying about risk and uncertainty for
the last 25 or 30 years, but which economists, to some extent, are ignoring—and I don't
mean the theoretical economists. The theoretical economists are loaded with that stuff—
they all know it. If you ask theorists in risk and uncertainty what they think about the
expected utility model, they'll say, "It doesn't work, and we have 6 billion experiments to
show that it doesn't work."

Now, our task then is to determine how we adopt a more-general framework in an
empirical setting. Can we do something to bring some of what they're telling us is true
into an empirical model? So, our agenda is to try and develop an empirical model of one
of the non-expected utility models, make it work with survey data, and derive a formal
derivation of an [unintelligible word] welfare measure that's consistent with that
generalized or non-expected utility model.

So, if this sounds like mumbo jumbo, take a look at the paper that we have coming out in
the Journal of Risk and Uncertainty, technically coming out this month although it
probably won't be. That one's on nuclear waste. We're going to try to do it better when
we do this with arsenic in drinking water. So, I won't talk a lot about arsenic in drinking
water, but I'm going to talk more about the theory, and I'll hope you're awake enough to
catch it, because I think a lot of this is important.

D. Shaw Presentation, 4/11/06

1

-------
So, let me jump to the important things. First of all, what we're interested in is called
"ambiguity," and ambiguity goes back a long way. Daniel Ellsberg thought of this in
1961 and talked about how when he did experiments, people were averse to ambiguity.
So, what is ambiguity? Ambiguity is uncertainty about the risks. So, if you think that
you know the risks but people say, "I heard what you said, and I saw all your visuals and
everything you communicated to me, and I still don't get it—I still am not certain about
these risks"—that's ambiguity. In a lot of conventional settings, you'd say that we can't
do anything with ambiguity. Well, we can, so that's what we're trying to do.

When you introduce ambiguity, what happens is that the conventional expected utility
model is not going to be that useful. It's going to provide a benchmark, so we're not
going to throw it out—we're going to try and use it—but we're going to try to expand on
that and let it provide a benchmark for us. So, Kathleen [Bell] told you all about arsenic
risks, but let me get a little bit of the nitty-gritty of where the problem is [he refers to
slides.] If you go to the reports that were done for the arsenic rule, it's as Kathleen said:
there were some thresholds. The old threshold was, of course, 50 parts per billion (ppb);
the new one is 10. What do we know about 50? Well, that in itself the experts don't
agree on. Even one of the members of the Science Advisory Board for arsenic said, when
I asked him if 10 was safe, "Well ... we didn't all agree about that." When I asked then
why they set it at 10, he responded, "Cost—the cost of compliance. If you go below 10,
it's going to be a real problem, particularly for rural areas."

So, is 10 safe? Well, we're going to say that 10 is safe, but the reality is if we tell people
that 10 ppb is safe (which may or may not be true, but let's assume for now that it is), the
real problem for people, as Kathleen suggested, is between the 10 and wherever they are.
So, if someone is at 22 ppb, and they're wondering if they're safe, what exactly are the
health risks? We know some things from the Science Advisory Board. For instance,
we're looking at 30-to-60 times higher than baseline risk for the incidence of lung and
bladder cancers. That's good that we know that, but we have no exact credible
relationship that's been mapped out between 0 and 10, or 0 and 22, or 0 and 50. Now,
there was extrapolation done in the EPA report that looks pretty good, but a lot of the
physical scientists are arguing and disputing with that—there are some papers out in
some of the science journals that say it was just an extrapolation and if we apply a
different approach, we get different results.

So, this is a perfect setting to allow for ambiguity, because if the experts don't even know
what the risks are, then how can we expect the respondents to know what the risks are?
Then we add to that that you've got all of these complicating factors. It matters hugely in
a drinking water setting that risk is completely endogenous. It can be completely
controlled by averting behavior, either through your drinking water behavior itself or
adopting a treatment—you can solve the arsenic problem pretty quickly if people are
willing to install a reverse osmosis or distillation treatment method in their homes or if
they're willing to support a program that brings the public drinking water system into
compliance. But, that means that the risk is endogenous and there is going to be a lot of
averting behavior that we have to watch out for.

D. Shaw Presentation, 4/11/06

2

-------
Okay, I'll tell you what I know so far about arsenic, and then I'll try to do a little more of
the theory to give you an idea about where we're going with this. We did a pilot study
for USD A. If you want the papers, there are two of them out that are already published—
you can ask me about those—I have a couple of copies of them. Here's the bottom line,
and I think Kathleen suggested that they're finding the same thing. Drinking water
behavior is very complicated. When we started the pilot study for USD A, we thought it
was really simple. We thought that all you have to do is ask people, "Do you drink tap
water?" Wrong! What we found out in the pilot study is when you ask people this
question, a whole bunch of them say, "No." Then when you follow that up and ask what
they do, they might say they drink bottled water. Then when you ask them if they use
bottled water in all their cooking and to make tea and to fill up the ice trays and all those
kinds of things, they say, "Oh no, we don't do that. We drink bottled water if we're at
work, and then when we're at home we use the tap water . . ." They said, "No," but the
reality, and what the published studies report, is that they do use tap water. So, you have
to ask them very detailed questions about what their drinking water behavior is. People
who live in two-story houses will ask whether you mean in the kitchen or whether you're
asking about the glass of water they drink when they get up in the middle of the night. If
you're really going to get a detailed report on drinking water, you have to ask all of that.

[Again, referring to slides, he continued] People don't treat because of cost. We have a
relationship in the data that we have, where we were looking at a rural area of Nevada.
They were completely aware of the arsenic in the drinking water of the rural area
community that we studied. They had been studied to death—the CDC had come in
there—Hillary Clinton had gone there—everybody had gone there—they had a very well
publicized cancer cluster—they were all very aware, and they're still drinking it. We
have people in this rural area who are drinking water with arsenic at 100 ppb—we have
some with 500 ppb—the risks are very, very high. We were astonished to see that. So,
when you ask them why they're doing that, they say, "Well, you know, the government
doesn't know what they're talking about. This stuff is safe—I've lived here all my life,
and I don't have cancer." That's a common response that we received. So, we learned a
lot from that. We learned that we have to expand the surveys that we are doing to
substantially rethink and rework the kinds of questions about tap water that we're going
to ask.

Trudy [Cameron] has a paper that I don't think many people know about. It's in the
Journal of Risk and Uncertainty the year before ours. It's very different from our paper,
but she does have a measure of ambiguity, and she finds that it is important in explaining
behavior.

Here's another way to think about ambiguity that might be useful. The decision theorists
say that you can view a lot of complex problems as a two-stage lottery. In normal
expected-utility frameworks, we think that everybody can take a compound lottery and
they can do the multiplication and they can reduce that to an easy single-lottery problem.
That's called the "reduction of compound lottery axiom," and that's something you
would adhere to and believe in if you believe in expected-utility theory. With ambiguity,
that's not so. What we think, in fact, is that under ambiguity, people cannot do that. In

D. Shaw Presentation, 4/11/06

3

-------
experiment after experiment after experiment—many of these get reported on in the
Journal of Management Science, which is a big decision theory journal, and some other
ones—people can't do it. And, in conjunction with Reed's comments this morning, [F.
Reed Johnson] they are much better able to do it in the context of financial gambles
because when you talk about flipping a quarter and getting a head or a tail, they can
figure that out. So, if you say there's going to be an expected outcome, they'll say, "Oh
yeah, I get that. I know what an expected outcome is"—because it's simple. What we've
found in the work that we're doing is that they can't do the same thing when you talk
about mortality. It's emotional, and it's very difficult for them to do.

So, what does the utility function look like with ambiguity in it? [again, referring to a
slide] There's one right there. They're complicated. You can put in two different
probabilities, and you can put in what's called an "absolute risk aversion coefficient." In
empirical work, we don't want to assume that people are averse to ambiguity—we want
to test for it in an empirical model and see if it turns out to be true. In that utility function
there [referring to a slide], there are two states, and in state 1, the person doesn't know—
the probability could be equally likely to be very small or very large. We see that in
behavioral experiments all the time.

So why do people do this? They do this for a lot of reasons. One, psychologists think
they do this because they get confused. You give them too many sources of information,
and those sources conflict with one another. For instance, in climate change, gee, it
might get hotter or it might get cooler—it might get wetter or it might get drier. So their
perception is that the questioners don't even know, so they get totally confused and that
creates ambiguity. It's more pronounced when they're not confident and, if it can be
overcome, if they're quite confident. Chip Heath and Amos Tversky have a paper on
that.

[referring to a slide] Here's the mess that you get into: If you allow for ambiguity, then
the probabilities that you're dealing with may not do the things that you're hoping they'll
do. We all think, for example, that probabilities are supposed to sum to 1 if we've given
people a complete space of probabilities. With ambiguity they may not, and the decision
theory people are well aware of that. David Schmeidler has this whole new mathematical
technique called [unintelligible word] integration which allows for that. You can also
have violations of stochastic dominance—and then something I'm working on as a side
paper to all of this: What is the welfare measure in all of this? The decision theory
people don't care about welfare measures. If we went to them right now and we told
them that we're running around estimating willingness to pay for this and that kind of
risk reduction, they'd wonder what we're talking about. They'd say, "Do you mean like
Pratt's Risk Premium?"—because that's the only thing they know about. And if we say
"option price"—some of us in this room know that if you send out a paper to a journal
and you say "option price" they send it to a finance person because they think you're
talking about pricing financial options. So, there's sort of a gap between them and us on
all of that.

D. Shaw Presentation, 4/11/06

4

-------
Okay, here's the key thing: In all of the alternatives to the expected utility model, the
heart, foundation, and soul of these approaches is a probability-weighting function. What
we know from observed behavior is that people can over-weight very low probabilities
and they can under-weight high ones, and each person can do something different. So,
that weighting function there has an inverse-S shape—that's the one that Kahneman and
Tversky thought that we would most often observe when we observe behavior. So, for
everybody in the sample that we're going to try to go after, we're going to try to get their
probability weighting function. The question is "How do you get it?"—and that turns out
to be very hard to answer. So, in the focus groups we have been trying to uncover this
probability weighting function. Again, the risk and decision theory people have done this
in experimental settings, but they've primarily done it with incentive-compatible
financial gambles. They'll tell people, "We'll give you $100 to participate in this
experiment—we're going to try to do these things on your probability weighting
function—if you get it right, you can win a whole bunch of money"—and people try
very, very hard to get it right. This is where a lot of this stuff about the trade-offs comes
from, so when you talk to people about doing kinds of risk/risk trade-offs or life trade-
offs, this is really coming from Peter Fokker's work on whether we want to ask people
about the difference between a certainty equivalent and risk measures or whether we
want to ask them about something different than that. That led people to wonder, "Gee,
could we do this with mortality risks?"

So, that's what we've been up to for the past six months—running experiments and doing
focus groups to try to see if we could get at this probability weighting function. We're
not having much luck, unfortunately. It's turning out to be quite hard. There is one paper
in the entire literature that I know of (and I'm probably wrong—somebody's probably
got another one somewhere) where they did it for mortality risks and they were
convinced that it was right.

The last little bit that I wanted to talk about is this willingness to pay issue. It has come
up time and time again over the past two days, and I wanted to put this slide up about
what is this welfare measure that we're trying to get? I want to remind people: We're
trying to get the option price. The reason we're trying to get the option price is because
Daniel Graham and all that work that he did was trying to help all those great people
(Rich Bishop among them) who had struggled with trying to figure out "What do we
want?—Do we want an option value?—Do we want an option price?—Can we use the
expected surplus?—and all of those kinds of things.

The bottom line is that it's a state-dependent concept. We have two states. We've got
the balance of the left-hand side with the right-hand side. What I wanted to make a plug
for is that if you look at Trudy's [Cameron] paper in the Journal of Public Economics last
year and you look at ours that's coming out, you'll see that we're very careful about the
derivation of that option price. In ours it's even more complicated because we have
ambiguity in the model and we derive the "quasi-" option price—because I don't know
for sure what it is yet. But, we take the expected utility difference, and when you solve
an equation like this one [referring to a slide], and you set it equal to zero in the top, the
expected utility difference, and you solve for some sort of payment that balances these

D. Shaw Presentation, 4/11/06

5

-------
two things, that's how you're going to get the formal expression for what the willingness
to pay is. What I'm interested in finding out is whether a lot of the really good people
who have been here the last couple of days are doing that.

Okay, that's kind of the theoretical stuff that I wanted to talk about, and if I lost you, I'll
be happy to talk about it afterwards. Here's what we're thinking, and it sounds as if
we're thinking a lot of what Kathleen [Bell] and them are thinking: With public systems,
we're trying to do both public and private wells—the private wells are not regulated by
the federal government, whereas the public systems are. What we have figured out from
our focus group work so far is that with public systems people can certainly choose to
treat within their own homes, and many people often do because they don't like the taste
of the drinking water that comes out of the public system. For example, in the town that I
live in the water is considered by EPA to be perfectly safe, yet thousands of households
have water treatment systems because they can't stand the taste. If you've ever been to
Texas, you know that the tap water has a lot of salt in it and tastes terrible. Another
interesting health issue is whether people's blood pressure is going up because of
drinking the tap water down there, so people are working on that down there too.

Now, their rates may increase if they're on the public system, or they may have already,
to pay for getting into compliance, and if it was an [unintelligible word] framework
before it happened, of course we would want to find out whether they'd support that rate
increase. We're going to also be trying to tackle the child vs. adult health issue, so thank
you for all those great papers—I've learned a lot on that, because we haven't been sure
how we're going to deal with that. We're going to try to allow for ambiguity for the
public system. For the private, we have to ask them if they treat. And remember, when
you ask this, it is a very complicated question, because in our focus groups and in the
pilot study we did for USDA people say they don't know what is meant by that. So, you
have to explain to them very carefully what treatment you mean and which types of
treatment that they can adopt actually get rid of arsenic. I never thought in a million
years to ask this or to explain this to people, but we've been asked in every focus group
so far: What about our refrigerator—is the water that comes out of the door dispenser
treated? They do actually put a little filter on the back of many refrigerators to filter the
drinking water, but as it turns out, it's a charcoal filter and it's not the kind that actually
gets rid of arsenic. But when you tell people that they need a reverse osmosis filter, they
don't know what that is. Or, they may respond that they treat, but it turns out to be a
Brita water pitcher, which doesn't do anything to get rid of arsenic.

Now, we were thinking of doing complete private welfare measures for those on private
wells and then we thought: These people could have a public-goods-related welfare
measure. So, again, similarly to what Kathleen and them are doing, we also are going to
try to take advantage of different valuations approaches. So, we can have a revealed
preference value that comes out of adopting a treatment method, which is going to be
averting behavior and can reveal their value for protecting themselves, but there's no
reason to think that those people in private homes don't also have some sort of value for a
public good. If we can get both out of the same household, we'll be able to cross-validate
and look at the preference functions in both cases, which would be really nice. That's

D. Shaw Presentation, 4/11/06

6

-------
one of the things that—those of you who have never cared about recreation demand
modeling, you missed a good thing. The good thing is that those guys figured out that
you can use stated preference and a revealed preference in the same model and then do
preference restriction tests, which are quite nice, actually.

So, we're going to tackle trying to get all of that—both a public and a private value—out
of the people on private wells, and we'll try to get the averting behavior if they've done
it.

I'll only share one of the focus groups so far and what came of that, [again, referring to
slides] For this group, which was on a public system—this was Eagle Mountain, Utah—
we knew that their drinking water was 26 ppb arsenic, so they're not in compliance with
the new arsenic rule. We did two focus groups there, and we thought that in the first one
we would not tell them. Well, in the first five minutes, when they started to figure out
that it was all about arsenic in drinking water, they began to demand that we tell them
what their level was. We realized that once you open this door, you have to tell them
pretty quickly what their arsenic level is. That's going to be a challenge for our survey
team. We're using a risk ladder, and J.R. [DeShazo] and I had a really good discussion
on the phone one day about risk ladders vs. grids. We have tested both risk ladders and
grids. Surprisingly, our folks are doing a lot better with the risk ladder than they are with
the grid, even though some people are saying that that grid is the way to go now. It may
be that you want to do baseline risks with the ladder and changes with the grid. But then
that raises the issue of "Are you overloading them with two different kinds of risk
communication devices?" In all of the risk experiments that we've done so far, when we
communicate to them that we think the risks are different for children and that they're
probably higher, they get that. So, when they come back with subjective risks, we let
them mark on the ladder what they think the risks are after we tell them what the experts
think the risks are. When they come back and mark them on the ladder, they mark very
different points than the experts did, but they get that relative difference between an
adult's and a child's risk.

[referring to slides] Here are some summary results from this particular focus group:
There were eleven subjects in the focus group and all but one said that the child's risks
were much higher. On the risk ladder, they all say very different orders of magnitude
from what we told them, which is interesting. That's borne out in the paper on nuclear
waste that's coming out in the JRU. Jim Hammitt was talking about having very, very
low risks of 1 in 100,000. Ours for nuclear waste were 2 in 10,000,000/ There is no way
that people can understand what 2 in 10,000,000 means—they just cannot do it. So,
when we get their subjective risks in that study, and we were looking at people that live
along the proposed transportation corridor for shipments to Yucca Mountain, they come
back with risks thousands of times higher than the DOE says that those risks will be. I
told that to Paul Slovik a couple of years ago and he laughed and said, "I told them that.
I've been telling DOE that for 20 years and they won't listen to me." But, again, if you
have very low risks, in the scheme of things all of this becomes much more important.

So, that's good enough—I'll stop there. Thanks.

D. Shaw Presentation, 4/11/06

7

-------
Willingness to Pay to Reduce Community Health Risks
from Municipal Drinking Water: A Stated Preference Study

Vic Adamowicz, Department of Rural Economy, University of Alberta
Edmonton, Alberta, T6G 2H1
Phone: 780.492.4603; Fax: 780.492.0268; Email: vic.adamowicz@ualberta.ca

Diane Dupont, Department of Economics, Brock University
St. Catharines, Ontario, L2S 3A1
Phone: 905.688.5550 ext 3129; Fax: 905.688.6388; Email: diane.dupont@brocku.ca

Alan Krupnick, Senior Fellow and Director,

Quality of the Environment Division, Resources for the Future
1616 P Street NW, Washington, DC 20036-1400
Phone : 202.328.5107; Fax : 202.939.3460; Email : krupnick@rff.org

With Assistance from Spencer Bahnzaf and Michael Batz, Resources for the Future, Lori
Srivastava and Jing Zhang, University of Alberta, and Paul De Civita and Andrew Macdonald,

Health Canada

October 2005
Please do not quote without permission.

ABSTRACT

This paper examines the value of health risk reductions to Canadians in the context of clean and
safe drinking water. The health risks we examine pertain both to microbial illnesses and/or
deaths and bladder cancer illnesses and/or deaths. The cancer risks arise because chlorine, the
most common disinfectant used to remove microbial contaminants, has been implicated in the
production of Trihalomethanes (a disinfection by-product) that are linked to increases in bladder
cancer cases. We evaluate results from an panel-based Internet survey of 1,600 Canadians
conducted in the summer of 2004. The survey included text and graphical information regarding
risk changes and employed contingent valuation and attribute-based stated choice benefit
valuation techniques. The valuation questions were designed to elicit consumer preferences for
public programs to reduce health risks associated with improved tap water. Our analysis of the
stated preferences of consumers reveals several types of values that are of interest to policy
makers. These include: the value of mortality risk reduction and the value of morbidity risk
reductions for both microbial contaminants and cancer. In addition, the value of reducing cancer
risks versus microbial risks in a public context is revealed. Our results suggest that reducing
mortality risks from microbial illness has greater value than reducing mortality risks from cancer.
Similarly, overall microbial risk reductions programs (mortality and morbidity) have higher
value than cancer risk reduction programs in this context. In addition, we provide separate
estimates of the value of statistical life associated with cancer and microbial risks, and the value
of statistical illness cases associated with these two risks. The results also include a host of

1

-------
comparisons between contingent valuation and attribute-based methods, as well as different
formats within each of these classes of methods. The values estimated in this study can be used
to evaluate investment decisions associated with water treatment, or as estimates of mortality and
morbidity value in benefit transfer cases.

We would like to acknowledge financial support from our partners on this project: the Canadian Water
NetworkZReseau canadien de I'eau, a federally funded Network of Centre of Excellence, the United States
Environmental Protection Agency, National Center for Environmental Economics, the Water Quality and Health
Bureau, Healthy Environments and Consumer Safety Branch of Health Canada and the Office of the Chief Scientist,
Health Canada. We would also like to thank Pierre Payment and the following people for their assistance in
preparation and development of the questionnaires: Spencer Bahnzaf Michael Batz, Lorie Srivastava, Anne
Huennemeyer, Jing Zhang, Paul De Civita and Andrew Macdonald.

2

-------
Background

Ninety percent of Canadians receive their tap water from public water systems (Environment
Canada, 2004). With assistance and scientific input from Health Canada, the Federal-Provincial
Subcommittee on Drinking Water (DWS) has developed a set of national drinking water
guidelines. The publication Guidelines for Canadian Drinking Water Quality lists substances
found in drinking water that are known or suspected to be harmful. The most recent summary
was published in 2004 (Federal-Provincial-Territorial Committee on Drinking Water, 2004).
Substances include both pathogens (microbes such as E. coli, Cryptosporidium, giardia, etc.) and
potentially carcinogenic chemical by-products (such as Trihalomethanes or THMs). These are
formed when chlorine - used for disinfecting water to destroy bacterial and viral contaminants -
reacts with other chemicals present in the water.

Provincial regulations require municipal water utilities to provide tap water that is as free as
possible from pathogenic micro-organisms called microbes. While many people are familiar
with the harm caused by the bacteria, E.coli 0157:H7 in Walkerton, it is not the only microbe of
concern. Over the last 10 years communities all across Canada have experienced problems with
other microbes including: Cryptosporidium and giardia. Microbes are generally transported into
surface water through agricultural runoff. While most municipalities employ both primary and
secondary disinfection technologies - typically chlorine-based - to remove microbes, recent work
shows that some microbes are present, even in disinfected tap water (Payment, Berte, Prevost,
Menard, and Barbeau, 2000).

3

-------
Concern has been expressed about the predominant use of chlorine for disinfection (Carson and
Mitchell, 2000). It is implicated in the production of a number of disinfection by-products
commonly called Trihalomethanes.1 These are considered to be potentially carcinogenic. Health
Canada convened an expert workshop in 2000 to look into the health risks of drinking water
chlorination by-products (Mills, Bull, Cantor, Reif, Hrudey, and Huston, 2000). After reviewing
the available evidence, the experts noted that, five epidemiological studies show a statistically
significant positive association of chlorinated by-product exposure with risk of bladder cancer.
The expert panel concluded "... that it was possible (60% of the group) to probable (40%) that
chlorination by-products pose a significant risk to the development of cancer, particularly
bladder cancer." Furthermore, they stated that "... this is a moderately important public health
problem."

For each substance, the Guidelines establish the maximum acceptable concentration (MAC)
permitted in tap water used. A change in any MAC level generally means that water suppliers
must improve disinfection techniques in order to meet more stringent requirements. In general,
these new methods are more expensive than the traditional chlorine-based methods. While they
may or may not be as effective at removing microbial contaminants, they are generally
considered to produce fewer THMs. Thus, it is possible that there is a tradeoff between reducing
THMs and reducing microbial contaminants.

To inform such tradeoffs, public preferences towards reducing bladder cancer from THMs and
microbial disease from pathogens in water must be gauged. Thus, the main research question
examined in this paper is how much Canadians are willing to pay on their municipal water bills

4

-------
in order to reduce these types of health risks from drinking tap water in their community. To our
knowledge, ours is the first effort to elicit tradeoffs from individuals between reducing microbial
risks and reducing cancer risks within the context of publicly supplied water quality. From a
methodological perspective, ours is also one of the few attempts to ask for mortality risk and
morbidity risk preferences in the same survey (see Cameron and DeShazo for another example),
albeit in a public goods, rather than private goods context. In addition, our study examines the
performance of two stated preference techniques within the same basic survey, i.e., contingent
valuation and choice experiments. While there have been several comparisons between CVM
with ABSCM (see e.g. Adamowicz et al 1998 or Hanley et al, 2001 for a survey) our comparison
includes controls for various context factors including information provision, number of
alternatives presented, and a referendum approach.

Using data from an Internet-based survey conducted across Canada during the summer of 2004,
the paper presents estimates of the value of reducing one more death or one more illness in the
overall population. Values such as these can be used to actually inform choices of technologies
for treating drinking water at the plant level and may also be used to help evaluate policy options
at the Provincial or Federal level. For instance, on the one hand, the status quo disinfection
technology implies a set of baseline risks for microbial illnesses and deaths and cancer illnesses
and deaths. On the other hand, alternative disinfection programs using ozone or ultra-violet light
are expected to reduce the health risks associated with cancer illnesses and deaths and with
microbial illnesses and deaths. However, these programs are more costly to the household (US
EPA, 1999). From the point of view of the public, the decision problem is whether it is worth the

5

-------
additional cost to have reduced risks of both morbidity and mortality effects and whether effort
should be focused more on microbial illness reduction versus reductions in cancer cases.

The next section discusses a number of methodological issues addressed in the surveys. This is
followed by a description of the survey versions employed in this study. Survey administration
and a brief description of the data are presented next. After this, the models and empirical
results, along with a number of statistical tests, are described in detail. A discussion of how
these results can be useful in a policy context follows. Conclusions and suggestions for future
research directions complete the paper

Methodology

Our goal is to obtain information about consumer preferences and tradeoffs relating to household
water bill increases and the morbidity and mortality health risks associated with the consumption
of tap water. Given the inefficient pricing structure adopted by water utilities and the absence of
competitive markets for the sale of tap water, virtually no information exists that yields the value
of potable water to Canadians, or indicates which aspects of water are subject to potential
tradeoffs according to preferences. In order to obtain this information we constructed a
hypothetical market, which allows respondents to express their preferences. We discuss below
in detail some important aspects of what we did: preference elicitation methods, presentation of
health risks, and public versus private risks.

Preference Elicitation Methods

We employ two non-market valuation methods for eliciting information about consumer
preferences for the public good "tap water" — contingent valuation methods (CVM) and the

6

-------
Attribute Based Stated Choice Method (ABSCM) (Adamowicz, Louviere, and Swait, 1998).
CVM requires the researcher to describe in detail the characteristics of the good to be valued
(scenario). Respondents then answer choice questions (we used a double-bounded dichotomous
choice format) about whether they would be willing to pay for the described good in its entirety
at a stated price. The researcher constructs the willingness-to-pay for the good, where the
expressed WTP is for the good in its entirety as described in the scenario, from the pattern of
responses. In the ABSCM framework a good is described expressly as a bundle of
characteristics or attributes. Each attribute provides valuable services to the consumer. While the
individual attributes have value, they cannot be purchased separately but are acquired by the
consumer at some stated price for the entire good. With this approach, then, the price paid for a
particular bundle of characteristics becomes itself an attribute. In contrast to the CVM method,
which provides an overall willingness-to-pay for the bundle of attributes, the ABSCM approach
permits us to determine separate willingness-to-pay values for each identified attribute, as well as
to examine tradeoffs between individual attributes. For the purposes of this project, the relevant
tap water attributes are household water costs and morbidity and mortality health risks from
microbial and bladder cancer.

Describing the Health Risks

In presenting the program choices to survey participants we need information about the health
effects (a description of each health risk in terms of the symptoms) and baseline risk levels (the
likelihood of contracting the disease and or dying from it), as well as changes in risk levels and
costs of different programs. A range of reasonable program cost increases was estimated from

7

-------
information on alternative disinfection technologies (US EPA, 1999). These were presented as
dollar increases per year in one's household water bill effective January 2005. Information
describing symptoms of microbial and bladder cancer illnesses is readily available from a
number of sources including Health Canada and the United States Centers for Disease Control.
(See Appendix 1 for descriptions used in the survey.)

Baseline information for the number of microbial illnesses and deaths attributable to waterborne
microbes is needed for our survey but difficult to ascertain. While outbreak data are collected by
regional health officials, they are generally considered to be lower bound estimates of endemic
health risks (Mead et al. 1999). This is for three reasons. First, some people become ill prior to
general knowledge of an outbreak and are not tested by the doctor for the presence of the
microbe, so these cases are not counted. Second, symptoms are often attributed to another cause
such as food poisoning or flu. Third, some microbial illnesses are not considered "notifiable"
diseases, so doctors are not required to report cases. A second source of data for water-based
microbial illnesses is from medical practice cases, which are generally considered to better
present the endemic risks (Wheeler et al. 1999; De Wit et al. 2001). A third source of data, which
presents the highest estimates of health risks from microbial illnesses, is from microbiological
studies examining water supplies for presence of pathogenic micro-organisms.2 These represent
the high end because they assume a dose-response model that links the number of organisms to
the number of affected persons (Payment and Riley, 2002). Data on the number of deaths
attribute to waterborne microbes are even scarcer. However, Ronchi and Wald (1999), writing in
the OECD Observer, claim that "in the United States about 900,000 cases of illnesses and 900
deaths occur every year as a result of microbial contamination of drinking water".

8

-------
Determination of the baseline risks of becoming ill and/or dying from bladder cancer from
consuming water that contains elevated levels of THMs also poses problems. While there is
some disagreement in the scientific/medical literature about the relationship between chlorine in
water and the incidence of bladder cancer, there are a number of studies that show an association.
Recent work under the auspices of Health Canada reports on a study of individuals living around
the Great Lakes. The research shows a link between the presence of THMs in drinking water and
increased cases of bladder cancer. These results suggest that long-term exposure (on the order of
20-35 years or more) to THMs in water may cause between 14-16 % of all bladder cancer cases
in Canada (King and Marrett, 1996). Similar numbers from the United States EPA are between
2-17 % (Mills, Bull, Cantor, Reif, Hrudey, and Huston, 2000). Cancer statistics are available
from Health Canada (Cancer Surveillance on-line) by site. Status quo bladder cancer cases
attributed to water consumption can be estimated by applying the attribution rates to all bladder
cancer cases. Mortality rates are also presented on the Health Canada Cancer Surveillance web
site.

With our baseline numbers established (See Appendix 1), we review the engineering and
microbiological literature for estimates of anticipated reductions in microbes and/or THMs
associated with improvements to water disinfection systems. Numbers from US EPA (1999),
Havelaar et al. (2000) and Barbeau et al. (2000) form the basis for our estimates of changes in
baseline risks presented to survey respondents (See Appendix 1).

9

-------
Appendix 1 shows three pages of health risk information presented to our survey respondents.
They review this information prior to answering the preference elicitation questions. The first
page describes potential health effects associated with using chlorine for water treatment. In
particular, it describes symptoms of bladder cancer and clearly identifies the potential tradeoff
between the beneficial aspects of reducing microbial contaminants and the potential adverse
effects in terms of enhanced risks of contracting bladder cancer. The second page places the
baseline microbial and cancer risks together and shows typical linkages between illnesses and
deaths for each health condition. In addition, it puts health risks from tap water consumption
into a more general perspective. It is important to present the contextual setting to respondents,
so that tap water health problems are not viewed in isolation from other health risks. The third
page summarizes the baseline health risks from the four health outcomes: microbial illness, death
from microbial illness, bladder cancer illness, and death from bladder cancer. Again, the
magnitude of health risks from tap water consumption are contrasted with all health risks for
each of these health outcomes.

Since we are asking our respondents to assess these health risks, we need to ensure that they are
able to evaluate changes in health risks in a meaningful way. Some respondents find numerical
representations difficult to interpret. There is a large literature on how best to communicate risk
and researchers have used visual aids such as graphs, pie charts, risk ladders, and tables (Jones-
Lee et al., 1985; Hammitt, 1990; Corso, Hammitt, and Graham, 2001). We adapt probability
communication techniques from Krupnick et al. (2002). After experimenting with a number of
options we use what we call our "snake in the sand" design. This begins with a blue rectangle
representing a population of 100,000. To this rectangle we add yellow squares representing

10

-------
individuals who get microbial illnesses from drinking tap water and red squares representing
individuals who get bladder cancer from drinking tap water. We superimpose black squares onto
either the red or yellow squares in order to illustrate the deaths arising from either microbial
illnesses or cancer illnesses. An example of this graphic is shown in Appendix 2 for a CVM
format question and in Appendix 3 for an ABSCM format question.

After reviewing the background information (in Appendix 1), the survey respondent is presented
with a discussion about changes to water disinfection methods that can alter health risks. The
respondent is told that he/she will be faced with a series of choices regarding alternative
municipal water treatment programs for his/her community. Each choice includes a status quo
(do nothing) option. Alternative programs presented generally lower the health risks and involve
an annual increase in the existing water bill for the household. A given respondent answers
questions either in the CVM format (example in Appendix 2) or the ABSCM format (example in
Appendix 3).

Private versus Public Risks

An issue arising from the approach adopted in this research is whether this particular problem
should be treated as an individual (private) decision or a social (public) decision. The private
decision context would readily yield individual specific measures of value (e.g. values of
statistical life or VSLs) that could be compared to other private good estimates (e.g. Krupnick et
al, 2002). A public context, however, is more realistic in this setting since drinking water is

11

-------
consumed by an individual at home as well as at other places (office, school, etc.) and most
people view drinking water treatment as a municipal or public responsibility. Therefore, the
decision context chosen for this case is a public or social decision. Carson and Mitchell (2000)
make the same choice in their open ended CVM survey to obtain willingness to pay for carbon
filtration to reduce the risks associated with trihalomethanes,

Thus, with our approach respondents are asked to indicate their preference for one program for
drinking water improvement over another (or the status quo). A potential drawback of this
approach, however, is that the resulting estimates of the willingness to pay for water quality
improvement and for the specific attributes of reduced microbial and cancer risks may contain
elements of altruism. That is, when individuals make their choices they may be thinking about
their family members, friends, and others in the affected community who will benefit from this
program in addition to themselves. Thus, we elicit the individual's preferences including, at
once, that for their own health and for the rest of their community. While, in principle, we would
like to have these "total social values" to make policy decisions, summing altruistic values from
all individuals can introduce an unknown, possibly large degree of double-counting, as opposed
to the summing of individuals' values for their own risk reductions, where the latter provides,
perhaps, a reasonable lower bound to social value. While this is a challenge, it may also provide
us with important and interesting information. Since so many individuals in certain provinces
and areas of Canada rely on tap water substitutes, and thus may believe the benefits of such
programs will be enjoyed wholly by others, ultimately we hope to be able to sort out altruistic
and individual values. This is a topic for future papers.

12

-------
Outline of Survey Versions

We developed two versions employing the CVM format and 6 versions employing the ABSCM
format. In this paper, only four versions of the ABSCM format are referenced. Details of each
version follow. Table 1 describes some of the key features of these different versions.

Versions 1 and 2: Contingent Valuation Methodology Format (CVM)

The CVM format (example shown in Appendix 2) presents the respondent with the option to
choose status quo (no increase in water bill, no reduction in health risks) or a new municipal
water treatment program (increase in water bill, reduction in some or all health risks).
Regardless of the versions, each respondent was presented with three separate double-bounded
dichotomous choice questions. For Version 1, when compared with the status quo, the first
question presented a reduction in bladder cancer illness (from 100 to 50) and a proportional
reduction in the risk of death (from 20 to 10), holding constant microbial illness and death risks
at their status quo levels. For the second question, respondents were asked to consider a
reduction in microbial illnesses from 23,000 to 7,500 and a proportional reduction in the risk of
death from 15 to 5, holding constant cancer illness and death risks at their status quo levels. For
the third question, the reductions in health risks pertained to all four risks and were the same as
those in questions one and two. The payment vehicle was additional costs to the household
water bill. Payment levels ranged between $25 per year to $350 per year.3

13

-------
For Version 2 the ordering of the first and second WTP questions was reversed; however, the
risk reduction and payment levels are the same as those in Version 1. The third question was
identical to that in Version 1.

Versions 3 to 6: Attribute Based Stated Choice Format (ABSCM)

The ABSCM approach begins with the determination of a number of attributes that characterize
the good to be valued, along with the setting of the number of separate levels of those attributes.
For each choice task the respondent compares a status quo option of no change (risks or
household water bills) with either one (Versions 3 and 6) or two (Versions 4 and 5) alternative
municipal water treatment programs, where attribute levels for these programs are varied
systematically according to the experimental design. Each combination of attributes/levels
represents a unique bundle of the good to be valued (Cochran and Cox, 1957). A fractional
factorial experimental design procedure is needed to identify those combinations that best reveal
the underlying consumer preferences (Louviere, J., D. Hensher and J. Swait, 2000). We identify
32 combinations and divide these into 8 blocks of four questions each. In order to avoid
respondent fatigue, each respondent is randomly chosen to face a particular block of 4 choice
tasks only. Appendix 3 presents an example of one of the choice tasks faced by survey
respondents who received the ABSCM format of the questionnaire.

In order to facilitate a direct comparison of the results from Versions 1 and 2 with the ABSCM
format, Versions 3 and 6 present respondents with a status quo option, along with a single
alternative program choice. Programs describe three attributes: cancer cases, microbial cases and

14

-------
household water bill. Cancer and microbial cases are each defined to have four levels of
attributes, while household water bill has five levels (including a status quo level of zero increase
in a water bill). These values are the same as those used in the CVM questions. We maintain the
fixed proportions ratio between morbidity and mortality effects.

One problem with assuming a fixed proportions relationship between illnesses and deaths is that
it does not permit us to disentangle the willingness to pay for cancer or microbial morbidity risk
reduction from that for cancer or microbial mortality risk reduction. While we could have
created a large number of sub-samples of CVM questions using varied proportions, this approach
would have been costly since it would have required at least 100 respondents per sample in order
to have confidence in the statistical properties of the estimates. A solution is to employ a
desirable feature of the alternative ABSCM format to obtain separate WTP values for each of the
health risks of interest.

Versions 5 and 6 relax the assumption of proportionality between morbidity and mortality health
effects. This requires us to specify five attributes: cancer illnesses, cancer deaths, microbial
illnesses, microbial deaths and household water bill. Version 5 presents the respondent with a
status quo option, along with two alternative programs. Version 6 is similar to versions 3 in that
the respondent may choose only between status quo and one alternative program.

Regardless of the version the ABSCM formats share a common framework. Each respondent
provides us with four separate choices across a number of different levels of each attribute.

15

-------
These choices are pooled and added to choices made by other respondents in order to obtain
WTP estimates for the various attributes, along with the tradeoffs between these attributes.

We identify results from the 6 versions as follows. We call results from Versions 1 and 2 our
CVM results. We call results from Versions 3 and 4 our Proportional (ABSCM) results and we
call results from Versions 5 and 6 our Non-Proportional (ABSCM) results. In all cases we first
estimate separate models using data from each version without covariates and follow this with
estimates that include covariates. Furthermore, we estimate all models using firstly the full
sample of data and secondly a reduced sample of data that removes individuals whose responses
identified them as "yea-sayers." (See discussion in next section of how these individuals were
identified.) Finally, we also estimate pooled versions of the models: CVM (using data from
versions 1 and 2), Proportional ABSCM (using data from versions 3 and 4) and Non-
Proportional ABSCM (using data from versions 5 and 6). We also perform series of statistical
tests to determine whether these data can be pooled.

• The CVM results are used to obtain a willingness to pay estimate for reductions in cancer
risks, a willingness to pay for reductions in microbial risks, and a willingness to pay for
reductions in both types of risks together. In the discussion of the empirical results we
examine the role played by question order upon these willingness to pay values. We also
examine whether the results support both a weak adding-up test (the willingness to pay
from question 3 is greater than either the willingness to pay from question 1 or question
2) and a stronger form of the same test (the sum of the individual willingness to pay
values from questions 1 and 2 is equal to the willingness to pay value obtained for both

16

-------
items in question 3). We also examine the effects of screening out for yea-saying
responses and, finding such effects, do most of our analyses with these respondents
removed.

•	The Proportional ABSCM results are used to calculate both an overall willingness to pay
for the same health risk reductions as described in the CVM scenario for the three items
(microbial risk reduction alone, cancer risk reduction alone, and combined cancer and
microbial risk reduction). In addition, we present results on the marginal willingness to
pay for a one unit reduction in either of these items. As for the CV analysis, we remove
respondents who answered questions leading to categorizing their answers as "yea-
saying."

•	We compare the results from the CVM approach with those using the Proportional
ASBCM format in order to determine whether question format has an impact upon the
estimated willingness to pay values.

•	The Non-Proportional ABSCM results are used to calculate overall willingness to pay
values for the same health risk reductions described in the CVM scenario. However, we
can now separate out the WTP for the cancer deaths from that associated with cancer
illnesses. Similarly, we calculate the WTP for reductions in microbial deaths, as separate
from the WTP for microbial illness risk reductions.

•	We compare results from the Proportional ABSCM and Non-Proportions ABSCM
versions to determine whether a relaxation of the fixed proportions assumption of deaths
to illnesses has an impact upon the estimated WTP.

17

-------
Survey Administration and Data Description

Survey Administration

We employed Ipsos-Reid, a marketing and public research agency to administer and put the
survey onto a secure on-line website. Respondents were solicited from amongst a panel of
Internet users maintained by Ipsos-Reid. The panel consists of over 100,000 members and
reflects an accurate, balanced representation of Internet-enabled Canadians, recognizing that this
does not necessarily mean that the panel is representative of all Canadians. These households
have been recruited primarily to the panel over the telephone using random digit dialing. After
focus groups and pilot testing to refine the survey, we implemented the final version in two
waves during the summer of 2004. The waves are only important because they gave us an
opportunity to make "mid-course corrections" to the survey, of which there were virtually none.
As, after analyzing the data, we have found no reason to distinguish responses by wave, we drop
this distinction from here on out. On our behalf Ipsos sent out 4,563 email invitations to its panel
of Internet users, of which 2,520 respondents began the survey. Of these 1,633 completed the
survey and 419 individuals quit the survey before completion. Additionally, 466 were dropped
because they did not obtain any of their tap water from a local municipal water supplier. Finally,
2 responses were deleted after errors arose when the Ipsos server went down in the middle of
completing a survey. Assuming that ineligibles are found in the same proportions to those
contacted as to those responding (466/2520 = 18.5%), the overall response rate is 46%
(1,633/3,536). As we utilize only six of the 8 versions of the survey in this paper, the sample
size is 1219.

18

-------
Table 2 presents summary statistics from the data for all the variables used in this paper, both for
the full samples (by stated preference approach — CVM or ABSCM) and samples that remove
"yea-saying" observations. (See below for discussion of how this was done). The Table also
reports 2001 Census average statistics for the Canadian population. For most characteristics,
average values for survey respondents are virtually the same as these average values. The only
socio-demographic characteristic that differs in any appreciable way from that of the general
Canadian population is the percentage of individuals educated beyond high school. The 2001
Census estimate is 55 per cent, while the corresponding value for our sample, collected in 2004,
is 79.1 per cent. In the previous five years, the percentage of people educated beyond high
school increased 5 points. So, the 2004 percentage is likely to exceed 55 percent.

The most important implication of our overeducated sample is in the implication for non-
response bias because of the Internet nature of the panel. Statistics Canada (2004) notes that two
thirds of Canada's 12.3 million households have at least one family member who regularly used
the Internet in 2003. Thus the degree of bias suggested by the education level in our sample may
not be very large.

Beyond the issue of sample representativeness is the issue of what is called in the stated
preference literature "warm glow" (see, for instance, Kahneman and Knetsch, 1992, "Valuing
Public Goods: The Purchase of Moral Satisfaction," JEEM 22 57-70; and Andreoni, J. 1989.
"Giving with Impure Altruism: Applications to charity and Ricardian Equivalence," Journal of
Political Economy 97(6), 1447-1458) and "yea-saying" (see, for instance, R. K. Blarney, J. W.

19

-------
Bennett, M. D. Morrison. (BBM) 1999. "Yea-Saying in Contingent Valuation Surveys," Land
Economics, Vol. 75, No. 1 (Feb.) , pp. 126-141), the former being a more narrowly defined
phenomenon than the latter. The warm glow issue is that, when asked to value a public good,
people may derive satisfaction, and be willing to pay something, just from the act of giving. The
latter is defined by BBM as "the tendency to subordinate outcome-based or 'true' economic
preferences in favor of expressive motivations... " (pg. 126). There are two implications for our
results. First, we might expect that people's responses to either CV or CE questions will be
insensitive to the commodities or attributes being put up for purchase, and that, therefore, if
present in large numbers in the sample, such responses will make it less likely for various tests of
sensitivity to scope to be passed. Second, such people are likely to be insensitive to the money -
health tradeoffs being posed and are therefore likely to drive WTP estimates inappropriately
upwards. We will use the term "yea-saying" to describe this phenomenon.

To address these issues, we used responses to one or two questions to remove "yea-saying"
respondents. For the CV analysis, we removed people who said that they would pay anything for
health risk reductions and who answered Yes-Yes (YY) to all three dichotomous choice
questions with follow-up posed to them. This amounted to 44 respondents (11% of the sample
of 407). Interestingly, 10 people who said they would pay anything actually did not answer YY
to all three WTP questions. For the ABSCM analysis, we could not use the "YY to all three
WTP questions" condition because in a choice experiment set-up there is no equivalent to the
YY condition. Even if we had identified those respondents choosing the alternative with the
largest health improvement all six times, we would still not necessarily be removing yea-saying
effects since attributes are able to differ across programs. Thus,, in very few cases are there

20

-------
clear and unambiguous "yea-saying" answers. Therefore, we used only the first criterion (a
statement that the respondent was willing to pay anything for health risk reductions) to screen
out respondents. This amounted to dropping 86 respondents out of 812 (10.6 %). The remaining
respondents are distributed among 4 ABSCM versions, 361 in the two versions that are directly
comparable to the CV approach and 366 in the two approaches that varied morbidity and
mortality attributes, permitting separate valuation of these endpoints for both cancer and
microbial cases.

As shown in Table 2, the various samples have similar demographic and other characteristics and
responses (variables are defined in table 3). The exceptions include URBAN, where a higher
percentage of the respondents in Versions 5 and 6 live in urban areas, and ASSETS, where
wealthier people were randomly slotted into ABSCM versions (We generally did not use this
variable in further analyses because it is missing for too many respondents).

Models

The econometric model used to analyze the CVM survey data is the one appropriate for interval
data. We use this model to obtain estimates of WTP

log WTP*=Xij3 + s1 (1)

In this equation, WTP* is the underlying willingness to pay for a selected risk reduction; X
denotes a vector of age, health, and other attributes; /? is a vector of coefficients; and s is an
extreme value Type I error term. Effectively, equation (1) describes a survival time model based
on the Weibull distribution. The log-likelihood function for this model is

21

-------
n

log L = Y^\og{I'\(\o§¥TI>" - XiP)l o~\- ^[(log WTPtL - Xifi) I 
-------
An individual application of the method involves the generation of a number of bundles of
attributes, and these are presented to respondents in series of choice tasks. Thus, the attributes of
each alternative offered in a task comprise the Z vector and the sets of alternatives in each task
comprise C, the choice set.4 In our case, respondents are required to answer four choice tasks. In
Versions 3 and 6 the choice tasks consist of comparing the status quo and one alternative
program and in Versions 4 and 5 the choice tasks consist of comparing the status quo and two
alternative programs. The resulting information is viewed as four individual choices from either
a binary or a trinary universe. The econometric analysis (maximization of the likelihood
employing the probabilities derived from the equation above) provides the estimates of the
marginal utilities associated with the attributes and allows for their use in welfare measures.

Empirical Results: Contingent Valuation Method

There are three sets of results in this section. The first presents the most assumption-free results
behind the estimates of willingness to pay - the percent of sample voting Yes to the first bid they
are given. The second presents estimates of WTP and subjects these estimates to a variety of
validity tests. The third presents regression results to examine construct validity and estimate
marginal effects of covariates explaining WTP.

% Yes results

Figure 1 shows "percent Yes" responses by bid separately for each of three public goods being
valued (cancer risk reductions, microbial risk reductions, and both types of risk reductions
together). Note that each bar in the chart refers to a separate subsample. In general, there is

23

-------
concern for ordering effects, in that the answers to the microbial reduction program that follow
answers to the cancer reduction program may be different than answers to the microbial
reduction program when it appears first in the order. We use a likelihood ratio test to show that
the two samples may be combined, therefore, we present only the results of pooling the % Yes
responses across versions. Our expectation is that % Yes should fall with the size of the bid,
and that % Yes should be greater for the third WTP question (the one that combines cancer and
microbial reductions offered in the first two WTP questions) than either the responses to cancer
or microbial reductions alone. Figure 1 generally and visually supports these expectations,
which are confirmed statistically using a Wald Test in Table 5. The figure also reveals that the
% Yes for a microbial risk reduction program are generally larger than for a cancer risk reduction
program.

CVM WTP Results

Mean and median WTP results appear in Table 4. They are presented for a variety of
combinations of cancer and microbial endpoints and for two assumptions about the underlying
error distribution (lognormal and Weibull). Estimates are shown both for the full sample and for
the sample that removes yea-saying observations. In addition, we present separately results from
the two CVM versions (1 and 2) in order to examine issues related to question ordering. Thus,
the mean household willingness to pay from the full sample for a reduction in 50 cancer cases of
which 10 would have resulted in death (both over a 35-year period) in a community of 100,000 is
$535 Cdn per year taken from the responses to the first WTP question in Version 1. This WTP
translates into a VSC (a case being the above mortality/morbidity combination) of $14.4
million.5

24

-------
Because the Weibull outperforms the lognormal distribution in a variety of ways and because we
believe the yea-saying observations should be deleted, refer to the last two columns but one of
the table. The most reliable comparisons are for Cancer asked first (Cancer VI), Microbial V2
(microbial asked first in Version 2), and Both Cancer and Microbial Pooled. The mean WTP are
$182, $200 and $294, respectively. Using the pooled versions for added power, mean WTP for
reductions in cancer is $157 per household per year, while that for microbial cases is $211 and
that for these changes combined is $294. Median WTP is about half that of the mean.

Are any of these differences statistically significant? Table 5 presents results of both Wald tests
and likelihood tests. The tests of whether question order matters, or alternatively, whether the
answers to the cancer questions can be pooled (and the same for microbial questions and "both"
questions) show that they can be pooled by both types of tests. The next relevant comparison is
whether the WTP for cancer risk reduction and that for microbial risk reduction are statistically
different. Comparing WTP Cancer Pooled to WTP Microbial Pooled we find that the Wald
statistics is 3.685, slightly lower than the 95% Chi-squared value of 3.84. Thus, we barely reject
the hypothesis that the microbial WTP is larger. Finally, there are two types of "adding up tests"
— what may be termed the weak and the strong adding up tests. The weak test asks whether the
WTP for both risk reductions when asked together (in Question 3) is greater than that for either
risk reduction separately. Comparing the Cancer Pooled to Both Pooled, we see that mean WTP
value for both risk reduction changes exceeds that for cancer alone. However, comparing
Microbial Pooled to Both Pooled, we reject this symmetric finding (barely). However, if we
compare the more reliable WTP for microbial risk reduction, i.e., when it is the first question
asked, to the WTP for the third question (Both) pooled, we find that the Wald statistic exceeds

25

-------
the target value and that therefore the WTP for both changes exceeds that for microbial risk
reductions alone.

The strong test asks whether there is a summation relationship, i.e., whether the sum of the risk
reductions for cancer (pooled) and microbial disease (pooled) is significantly the same as the
combined risk reductions asked in Question 3 (pooled). In fact, this hypothesis cannot be
rejected. In an absolute sense, however, the sum of mean WTPs ($157 + $211) exceeds the WTP
for both risk reductions ($294), which could indicate declining marginal utility of health
improvements.

26

-------
CVM Regression Results

Table 6 presents regression results assuming a Weibull distribution (the lognormal results are
similar) explaining variables affecting the pooled responses to the cancer risk reduction question,
the microbial risk reduction question and the question with both reductions. These results are
representative of results with many other specifications and with many variables tried. In all
regressions, whether respondents believe in the health information we give them is a robust
variable, where those who believe are willing to pay more than those who do not. Household
income is negative and significant for cancer, but positive and insignificant in the other
regressions. This result is somewhat surprising and may arise from correlation between income
and other factors in the model, including education and/or belief in the scientific information.
Those from larger households and who are older, who have a college education, and who live in
more rural areas (but are served by municipal water supplies) are willing to pay more.
Interestingly, those who do not engage in averting behaviour are willing to pay less than those
who do not. This variable could be hypothesized to take either sign. On the one hand, those who
do not engage in averting behaviour may feel tap water has few risks, so might be willing to pay
less. On the other hand, those who do not engage in averting behavior may have stronger
preferences for good water quality, so would be more willing to pay for improvements. The
former hypothesis is the one that appears to be closer to the mark.

Finally, for cancer only, there appears to be an ordering effect, where those who answered the
cancer question first were willing to pay more than those who answered the cancer question
second. This is shown by the significant coefficient on the variable VIQ1 in Table 6. This is a

27

-------
dummy variable coded 1 for respondents who answered Version 1 questions (cancer risk
reduction followed by microbial risk reduction). This is in contrast to the findings in table 5
using the Wald and Likelihood tests. However, the other two comparisons show no difference in
responses across versions. The insignificant coefficient on V2Q1 (dummy for version two
responses where microbial risk reduction is asked first) indicates no ordering effect for
microbials. Similarly, the insignificant coefficient for the dummy variable VIQ3 indicates that
there is no significant difference in responses to the third question (microbial plus cancer risk
reductions) in either Version 1 or Version 2 responses. Other variables, such as for health status,
were not significant.

Empirical Results: ABSCM Method

We estimated six models from the Proportional data (versions 3 and 4), each version
independently, and a pooled model, for each of a full sample and a smaller sample that removed
yea-saying observations. The parameter estimates are presented in Table 7. These models include
only the attributes and status quo constant. All parameters are highly significant and of the
expected sign. There are significant status quo effects as illustrated by the positive status quo
constant. Tests of pooling Versions 3 and 4 (likelihood ratio tests) indicate that these versions
can be pooled and the joint model used for further analysis.

Table 8 provides estimates of the WTP for Microbial Deaths and Illnesses, Cancer Deaths and
Illnesses, and the sum of these two WTP values. Since these are proportional models a single
WTP amount is presented for both the mortality and morbidity reduction similar to the

28

-------
presentation of the CVM cases. The removal of yea-saying observations generally reduces the
size of WTP. For example, a reduction in microbial risks in the full sample pooled model are
valued at $219 while, for the yea-saying removed sample, they are valued at $175. Overall the
yea-saying removed sample exhibits WTP reductions of approximately 20 to 35% relative to the
full sample. Table 8 also illustrates the effect of the status quo parameter. When excluded the
WTP measures are significantly higher (on the order of 50%). Welfare measures with the status
quo effect excluded rely on the attributes in the model to capture all of the welfare effect of the
change while welfare measures with the status quo included discount changes from the status
quo (or changes in attribute levels) by the amount of the status quo preference parameter. There
is little guidance in the literature on how to treat this difference, thus we present both measures.
If a more conservative measure is desired the WTP with status quo effect included is appropriate.
A further finding in Table 8 is that the microbial programs appear to be more highly valued than
the cancer programs. This is a finding similar to that obtained in the CVM responses. This policy
relevant result carries through many of our findings.

Table 9 presents the parameter estimates for the Non-Proportional versions. As in the
Proportional case the parameters are highly significant and the signs are as expected. In this case,
however, the test of pooling is rejected. The version with one alternative (Version 6) is
statistically different than the version with two alternatives (Version 5). The most significant
difference can be seen in the size of the status quo effect. A much lower proportion of
respondents chose the status quo in the two alternative version. The reason for this difference is
unclear, but shows up in other work by the authors and is a topic for further research. The
inability to pool results from these Non-Proportional versions results in our conducting many of

29

-------
the tests discussed below on Version 5 and 6 individually, as well as on the joint version for
comparison.

Willingness to pay measures for the Non-Proportional versions are presented in Table 10. Values
are provided for microbial deaths and illness reduction programs (jointly) and cancer deaths and
illness reduction programs (jointly) to parallel the Proportional versions and the CVM analysis.
In addition, the marginal values per cancer and microbial death and illness case are presented.
Separate measures of the mortality and morbidity values are made possible by the Non-
Proportional design. In Table 10, the yea-saying removed results are generally lower than the
full sample results for the programs, but the difference is not as pronounced as in the
Proportional versions. For example, the WTP for microbial deaths and illnesses, with the status
quo effect, in Versions 5 is $306 in the full sample and $288 in the removed sample. The cancer
program provides values of $110 for the full sample and $80 for the yea-saying removed sample
for this same version. The status quo effect, however, is of the same magnitude as in the
Proportional case. The difference between status quo included and excluded for a microbial
program is $443 versus $306 in Version 5, and $336 versus $163 for Version 6. WTP amounts
from Version 6 are generally smaller than those from version 5, although the size of the
difference varies. The two alternative version provides smaller WTP values, regardless of
whether the status quo effect is included or not. Table 11 provides Wald test statistics examining
the differences in the WTP measures in the Proportional and Non-Proportional models.
Examining the tests of differences within versions we find that in the Proportional version there
is no significant difference between cancer and microbial WTP, but there are differences
between WTP for other programs (the critical value for a 5% level of significant is

30

-------
approximately 5). Examining the Non-Proportional versions, only in Version 6 is the difference
between Cancer and Microbial WTP not significantly different. In all cases the WTP values
between Cancer and Both are significantly different. The differences between Microbial and
Both, however, are generally not statistically different, although they are close to the critical
value.

The middle panel of Table 11 tests difference across versions. Interestingly the tests of WTP
across version are all insignificant. That is, the WTP value for the Cancer and Microbial
programs are not different across the Proportional and the Non-Proportional versions. This is a
very powerful result suggesting that the different elicitation strategies do not generate widely
different WTP values.

Finally, the bottom panel of 11 provides tests of the adding-up of deaths and illness WTP in the
Non-Proportional version (where these two effects are separated) against the WTP in the
Proportional version. In all cases the null hypothesis is accepted, indicated that response format
did not significantly alter WTP.

Table 12 presents measures of the value of statistical life (VSL) and the value of statistical illness
(VSI) for cancer and microbial deaths and illnesses. This summary table is based on the Non-
Proportional versions. Several pieces of information emerge from the table. First, in most cases
VSLs for microbial mortality are higher than for cancer mortality. This is particularly the case
when we use data that has removed yea-saying observations. Second, the VSL values themselves
are "high" relative to those in the published literature, however, they are not too far outside the

31

-------
range of accepted values. Recall that these are values for public reductions in mortality and thus
one would expect them to be larger that private WTP values.

The values of statistical cases of cancer are in the $2M to $4M per case range, and range
between 20% to 50% of the value of cancer mortality reduction. The value of a statistical case of
microbial disease is in the vicinity of $20,000, which is the product of an estimated WTP of
$0,018 per case per household and 100,000 people (38,500 households) in the community over
35 years. The value per case appears to be quite high and in our view results from the inability of
respondents to register preferences in a choice format that would lead to WTP estimates of a
fraction of a cent.

To provide an analysis of the effect of demographic factors on WTP, Table 14 presents two
sample sets of parameter estimates that included interactions with demographic factors. These
are examples of similar models for the various versions of the Proportional and Non-Proportional
models. In general the most robust findings are significant impacts of income (higher income
respondents are more likely to choose an alternative program to the status quo and higher income
respondents are less sensitive to cancer), Male (less likely to choose a program), Urban (less
likely to choose program), and those who believe scientists ( are more sensitive to cancer
deaths).

Comparison of Results from Two Formats

32

-------
Results from statistical tests on selected WTP values from the ABSCM and CVM versions of the
survey are shown in Table 13. The joint models from the choice experiment are compared to the
CVM WTP distributions. The tests of the joint Proportional model reject the hypothesis of
equality with the CVM values except for the case of the microbial program and the status quo
effect included. However, the values of the Wald tests are not that much above the critical value.
The tests comparing the Non-Proportional WTP with CVM are accepted for cancer with the
status quo effect excluded, and rejected for all other cases. Recall that the WTP measures with
the status quo effect included are considerably lower that with the effect excluded. The WTP
from the models with status quo effects includes tended to be lower than the CVM values while
the WTP from the models with the status quo effects excluded tended to be higher than the CVM
models. Thus, the choice experiment results appear to bracket the CVM results or provide an
upper and lower bound.

Using Results to Assist in Policy Making

The values calculated using the approaches discussed in this paper can be used to inform
decisions regarding drinking water infrastructure renewal and enhancement, as well as being
useful for cost-benefit analysis of drinking water rulemaking. In addition, it is possible that the
WTP and regression estimates can be used in various kinds of benefit transfers, to the extent that
such values are insensitive to the cause of the health effects (in this case drinking water and its
treatment).

33

-------
The WTP estimates from the non-proportional versions of the survey can be used to derive
estimates of the value of statistical life (VSL) or value of a statistical illness (VSI) both for
cancer and microbial disease. These values were reported in Table 12. In order to put these
results into context, Viscusi and Aldy (2003) examine a number of studies that have produced
estimates for the VSL. They report that the range of values is fairly broad between $3.9 - $21.7
million US dollars (2000). While our estimates fall within the upper range of these values, we
must note that the majority of studies that calculate a VSL do so using a WTP for a reduction in
the risk of death to oneself (that is, a private mortality risk). In contrast, our VSL estimates are
based on the WTP to avoid public mortality risks. We would expect that altruistic WTP values
might be higher than private WTP values since the former would include the willingness-to-pay
to avoid the deaths of members of one's community (including family members). Further, our
estimates are for deaths from two specific causes. The fact that the VSL for deaths from
microbial disease is somewhat greater than that for cancer is a big surprise. This may be related
to previous experience with contamination of municipal water systems in Walkerton, Ontario,
and North Battleford, Saskatchewan by microbial contaminants in 2000 and 2001, respectively.
More research on this point is needed to verify if this is an artifact of our survey or a true
representation of preferences.

In a similar fashion we can calculate the Value of a Statistical Illness as presented in Table 12.
Previously, cancer morbidity costs have typically been expressed using costs of illness. Our
estimates express costs for cancer in welfare terms.

34

-------
In addition, we can use the estimated willingness-to-pay values from either the CVM or the
ABSCM formats to obtain estimates for the composite value of a statistical case of illness, which
includes deaths for a small proportion of cases. Such estimates actually integrate morbidity and
mortality in one number so may prove even more useful for policy analysis than the VSLs or
VSIs if the policy reduces the source of cases (such as a pollution reduction policy) rather than
alters the ratio of illness to death (such as would occur with a health care policy).

Conclusions and Future Directions

This report presents findings from an Internet-based survey designed to elicit preferences relating
to tap water quality and health risks. These values show that Canadians are willing to pay in
order to reduce the public risks for a number of different water-related health conditions and that
they may have a mild preference for reducing microbial contamination over cancer cases. The
numbers pertaining to cancer appear reasonable and accord with prior work; however, there are
no comparable estimates available for microbial illnesses. We would argue that respondents
appear to have trouble with the large number of illnesses presented in the microbial case. This
results in small values per illness per respondent ($0.02), but large values per illness when added
up over the community. This is clearly an area requiring future study.

A few caveats are in order. Firstly, the numbers in this report, while typical of what we have
found, are first round estimates. We are still working to incorporate respondent heterogeneity
and to adopt non-linear indirect utility functions. These are the next steps. Secondly, our results
show how difficult it is to collect values based on very small probability changes. This is future
work. Finally, values collected in this fashion are for public goods, rather than private goods.

35

-------
.With few studies of this nature and with concerns about double-counting when one adds up
public values in the presence of altruism, caution is in order. Attempts to purge our estimates of
altruism effects are in our plans for future research.

36

-------
Table 1: Key Features of 6 Versions of Survey

Version

Question

Number of

Question

Number of

Relationship between



Format

Questions/Tasks

Ordering

programs (status

mortality and





Per Respondent



quo included)

morbidity

1

CVM

3

Cancer,

2

Proportional







microbial, both





2

CVM

3

Microbial,

2

Proportional







cancer, both





3

ABSCM

4

Na

2

Proportional

4

ABSCM

4

Na

3

Proportional

5

ABSCM

4

Na

3

Non-Proportional

6

ABSCM

4

Na

2

Non-Proportional

37

-------
Table 2: Descriptive Statistics by Version and Sub-sample

Variables

Canadian

CVM and

CVM

CVM

ABSCM

ABSCM

ABSCM

ABSCM



Population

ABSCM

(VI and V2)

(VI and V2)

(V3, 5, 6, 7)

(V3, 5)

(V6)

(V7)



Values

(VI,2,3,5,6,7)

full sample

yea-saying

full sample

yea-saying

yea-saying

yea-saying





full sample



observations
removed



observations
removed

observations
removed

observations
removed

INCOME

58360

57458.52

58734.17

58080.27

56819.12

54743.30

57796.83

60289.60





(35650.89)

(35562.79)

(35501.41)

(35699.68)

(35012.21)

(35865.36)

(37069.08)

MALE

49.9%

52.75%

54.55%

54.96%

51.85%

49.86%

50.81%

56.35%





(0.50)

(0.50)

(0.50)

(0.50)

(0.50)

(0.50)

(0.49)

AGE

45.8

46.55

44.93

44.18

47.36

47.58

46.15

47.42





(15.02)

(15.02)

(15.25)

(14.97)

(15.07)

(15.30)

(14.31)

HHSIZE

2.6

2.59

2.63

2.61

2.57

2.65

2.52

2.53





(1.31)

(1.26)

(1.26)

(1.34)

(1.35)

(1.37)

(1.25)

EDUCATION

55 %

61.77%

61.18%

62.04%

62.07%

61.77%

64.86%

62.98%





(0.49)

(0.49)

(0.49)

(0.49)

(0.49)

(0.48)

(0.49)

ENGLISH

73%

76.13%

75.92%

75.64%

76.23%

76.73%

75.68%

74.03%





(0.43)

(0.43)

(0.43)

(0.43)

(0.42)

(0.43)

(0.44)

URBAN

80% d

65.14%

61.67%

63.17%

66.87%

64.82%

70.27%

71.27%





(0.48)

(0.49)

(0.48)

(0.47)

(0.48)

(0.46)

(0.45)

ASSETS

na

89417.54

79677.19

79464.06

94483.83

93916.12

94999.78

93749.79





(82906.93)

(74853.48)

(75809.54)

(86427.92)

(85986.44)

(85166.08)

(90464.05)

BELIEFMS

na

74.82%

73.96%

71.39%

75.25%

72.85%

80.00%

70.17%





(0.43)

(0.44)

(0.45)

(0.43)

(0.45)

(0.40)

(0.46)

NO AVERT

na

45.37%

45.95%

46.46%

45.07%

42.38%

49.19%

45.86%





(0.50)

(0.50)

(0.50)

(0.50)

(0.49)

(0.5)

(0.50)

N



1219(906*)

407(310*)

353(266*)

812(596*)

361(263*)

185(135*)

181(130*)

Notes: a Standard deviations are in brackets.

b Yea-saying data is identified in CVM samples when YY for all three CVM questions and respondents indicate are willing to pay

anything for health risk reductions. Yea-saying data is identified in CE samples when latter condition is true.
c* denotes number of observations for ASSETS.

d The Census definition is more encompassing than ours. It includes an individual as being in a rural area if the population is less
than 1000. We used 10,000 to better capture locations with municipally supplied water.

38

-------
Table3: Definition of variables

Variables

Definition

INCOME

Average household income in Canadian $

MALE

Percentage of respondents who are men

AGE

Average in years

HHSIZE

Average size of a household

EDUCATION

Percentage of respondents with more than Some Community
College/CEGEP/Trade School

ENGLISH

Percentage of respondents whose first language is English

(This is indicated by whether respondents completed the survey in their

choice of English or French)

URBAN

Percentage of respondents live in a city in which the population is over
10,000

ASSETS

Total value of household's financial assets in Canadian $

BELIEFMS

Percentage of respondents who believe scientists are certain about microbial
illnesses arising from drinking tap water. (Highly correlated with other
belief variables relating to certainty of scientific community about risks
associated with cancer and microbial deaths and cancer illnesses.)

NO AVERT

Percentage of respondents who undertake no averting behavior against
drinking water related health risks

39

-------
Table 4. Mean and Median WTP Estimates by Endpoint. by Assumed Distribution. Full and Clean Sample



Lognormal

Weibull





Full Sample

Yea-saying

Full Sample

Yea-saying









Observations





Observations

Weibull







Removed





Removed

Mean



Mean

Median

Mean

Median

Mean

Median

Mean

Median

Test

WTP Cancer VI

535

110

289

84

266

119

182

91

2.458



ri99^

(]6)

C82^

nn

(46)

(44)

(27)

(27)



WTP Cancer V2

393

72

201

55

200

79

133

60

3.097



ri49^

nn

f55^)

(%)

f33^)

f3n

( N)

( N)



WTP Microbial VI

1075

130

532

95

332

137

226

104

1.883



(567)

(22)

(220)

n.v>

(67)

(61)

(3l>)

(3X)



WTP Microbial V2

552

118

344

92

265

127

200

101

1.548



a 94)

(17)

no(y>

ri3^

(43)

(43)

(21))

(3D)



WTP Both Cancer

1187

198

667

149

404

200

293

156

1.191

ft nil Mirrnhinl VI

(597)

r35^

(273)

(24)

(%6)

r82^

(54)

(55)



WTP Both Cancer

758

179

514

143

345

186

276

153

0.849

ft nil Mirrnhinl V7

(278)

(26)

(162)

(20)

(60)

(60)

(44)

(45)



WTP Cancer Pooled

464

90

244

68

232

97

157

74

5.480



(124)

(9)

(48^)

(7)

(28^)

(27)

(10)

(10)



WTP Microbial

739

123

416

94

294

132

211

103

3.453

Pnnlcrl

(223)

(14)

non

(10)

(37)

(36)

(24)

(24)



WTP Both Cancer

922

187

739

123

370

192

294

132

1.478

nnil Mirrnhinl Pnnlnl

(277)

(21)

(223)

(14)



(49)

(37)

(30)



Note:a Standard errors in parentheses.

40

-------
Table 5: Likelihood Test and Wald Tests



Lognormal

Weibull

Full Sample

Yea-saying
Observations Removed

Full Sample

Yea-saying Observations
Removed

Mean

Median

Mean

Median

Mean

Median

Mean

Median

Likelihood Ratio Tests









Cancer VI, V2 vs. Cancer
Pooled

4.217

4.632

4.205

4.663

Microbial VI, V2 vs.
Microbial Pooled

1.221

0.887

0.662

0.294

Both VI, V2 vs. Both Pooled

0.552

0.288

0.304

0.059

Wald Tests









Cancer VI vs. V2

0.327

3.815

0.787

4.325

1.362

0.555

2.232

0.894

Microbial VI vs. V2

0.760

0.170

0.609

0.022

0.714

0.019

0.292

0.004

Both VI vs. V2

0.426

0.192

0.234

0.041

0.316

0.018

0.059

0.002

Cancer Pooled vs. Microbial
Pooled

1.171

4.161

2.387

4.744

1.710

0.590

3.685

0.964

Internal Consistency Tests

Weak Adding Up Test:
Cancer Pooled
vs. Both Pooled

2.290

17.812

4.727

13.398

5.697

2.861

11.269

2.138

Weak Adding Up Test:
Microbial Pooled
vs. Both Pooled

0.264

6.472

1.745

3.167

1.478

0.991

3.453

0.465

Microb V2 vs. Both Pooled

1.198

6.510

2.627

2.790

2.499

0.996

3.897

0.434

Strong Adding Up Test:
(Cancer pooled + Microbial
pooled)

= Both Pooled

0.558

0.938

0.102

4.592

5.161

0.309

2.507

0.971

Related Tests:

Cancer VI & Both VI

1.074

5.257

1.766

6.267

1.983

0.739

3.389

1.121

41

-------
Cancer V2 & Both V2

1.337

13.907

3.315

16.818

4.451

2.482

9.032

3.635

Cancer VI & Both Pooled

1.288

8.340

3.603

5.071

2.312

1.211

5.879

0.803

Microb VI & Both VI

0.019

2.726

0.149

3.661

0.434

0.373

1.001

0.605

Microb V2 & Both V2

0.367

3.750

0.796

4.605

1.149

0.639

2.086

0.927



















External Scope Test

Cancer VI & Both V2

0.423

4.902

1.530

6.726

1.070

0.799

3.355

1.372

Microb V2 & Both VI

1.024

4.286

1.242

4.494

2.069

0.613

2.291

0.782

42

-------
Table 6: Regression Results - Weibull Distribution (Using Data that Removes Yea-saying

Observations)

Variable



Cancer



Microbia

r/

Cancer pi
Microbia

us

a

Constant



4.348



4.252



4.564

"kitit



(0 394^1



(0 398">



ro 379">



Household Income



-5S70F-06

**

3.640f-06



1 600F-06





r? 63 OF.-



(7 770F.-



r?. 770F.-



Male



-0.302

*

-0.260



-0.355

*



(0 1 8?">



(0 1 87">



ro i9o^>



Household Size



0.135

*

0.108



0.070





(0 074^1



(0



ro 07M



Age 65 or Older



0.514

*

0.256



0.061





(0



m 7.67^



ro ?64">



English



0.302



0.143



0.248





(0 ?Q4\



ro ?09^



ro ?06">



College



0.164



0.540



0.525

"kifk



(0 1 8?">



ro 1



ro 1 88">



Urban



-0.407

**

-0.397

**

-0.189





(0 1 86^>



ro i9n



ro i9o^>



Believelnformation



0.577



0.710



0.774

"kifk



(0 1 96">



ro 1QQ^>



ro 196^1



Do Not Avoid Tap

TVntov



-0.285



-0.337

*

-0.266





ro i 80^>



ro 184">



ro i8M



V1Q1



o.3so

**











(0 176">











V2Q1







-0.132













ro i am







V1Q3











0.051













ro 180">



Scale



1.343



1.330



1.232





(0 096^1



(0 09M



ro 09?">



































N



363



363



363



Log Likelihood



-432.83



-460.80



-440.59



Notes: a Standard errors in parentheses.
b significant at 10% level, ** is 5% and *** is 1%.

43

-------
Table 7: ABSCM: Estimated Parameters Proportional Versions

- ——

Full Sample

Yea-saying Observations Removed

Variable

Version 3

Version 4

Pooled Version 3

Version 3

Version 4

Pooled Version 3



(proportional)

(proportional)

and 4

(proportional)

(proportional)

and 4

Status Quo

0.837*

0.618

0.720*

0.850*

0.664*

0.759

Constant

(.175)

(.129)

(.102)

(.190)

(.136)

(.109)

Microbial

-0.163*

-0.156

-0.159*

-0.154*

-0.164*

-0.163

deaths

(.020)

(.013)

(.011)

(.022)

(.014)

(.012)

Cancer deaths

-0.125*

-0.122

-0.121*

-0.121*

-0.122*

-0.120



(.017)

(.013)

(.010)

(.018)

(.014)

(.011)

Program cost

-0.004*

-0.004

-0.004*

-0.005*

-0.005*

-0.005



(.001)

(.001)

(.000)

(.001)

(.001)

(.001)

Observations
(choice sets)

824

800

1624

111

732

1444

Log-likelihood

-507.22

-780.19

-1288.76

-428.70

-700.68

-1131.35

Notes

a Microbial deaths and illnesses are proportional, statistics are reported for deaths in the model. Similarly, cancer deaths and
illnesses are proportional.

b Standard errors in parentheses. Asterisk indicates significance at the.01 level.
c Test of pooling version 3 and 4, chi-squared 2.70, critical value 11.07.
d Test of pooling version 3 and 4, chi-squared 3.95, critical value 11.07.

44

-------
Table 8: ABSCM: Estimated Mean Willingness to Pay Values

" ——

Full Sample

Yea-saying Observations Removed

Variable

Version 3

Version 4

Pooled Version

Version 3

Version 4

Pooled Version



(proportional)

(proportional)

3 and 4

(proportional)

(proportional)

3 and 4

Microbial deaths













and illnesses













(including SQ

202.420

237.820

219.270

142.580

196.830

175.400

effectf

(40.69)

(29.63)

(21.48)

(31.59)

(21.89)

(17.17)

Cancer deaths and













illnesses













(including SQ
effect)"

101.720
(26.90)

149.290
(26.03)

123.550
(18.26)

69.401
(26.66)

113.150
(22.16)

88.427
(16.31)

Microbial deaths













and illnesses













(excluding SQ

429.660

396.310

404.230

321.090

334.740

332.590

effect)''

(107.29)

(65.87)

(53.59)

(75.18)

(51.75)

(38.44)

Cancer deaths and













illnesses

326.080

309.800

306.920

250.300

247.690

244.150

(excluding SQ
effect)b

(81.03)

(51.36)

(40.33)

(58.21)

(38.87)

(29.07)

Cancer deaths and













illnesses













(including SQ

528.350

548.450

524.760

393.590

445.430

419.130

effect)a

(99.03)

(68.81)

(53.677)

(64.87)

(47.296)

(36.612)

Microbial and













cancer deaths and
illnesses
(excluding SQ
effect)b

752.56

709.80

707.100

575.20

580.88

574.22

(176.13)

(111.14)

(92.11)

(123.89)

(81.14)

(65.40)

Notes:

a Welfare calculations include consideration of the status quo constant.
b Welfare calculations do not include consideration of the status quo constant.
c Standard errors in parentheses, based on Krinsky Robb simulation using 1000 draws.

45

-------
Table 9: ABSCM: Estimated Parameters Proportional Versions

———

Full Sample

Yea-saying Observations Removed

Variable

Version 5
(non-
proportional)

Version 6
(non-
proportional)

Pooled Version
5 and 6

Version 5
(non-
proportional)

Version 6
(non-
proportional)

Pooled Version
5 and 6

Status Quo
Constant

0.523*
(.123)

1.067*
(.172)

0.728*
(.097)

0.528*
(.127)

1.132*
(.183)

0.753*
(.102)

Microbial deaths

-0.054*
(.011)

-0.074*
(.017)

-0.056*
(.009)

-0.053*
(.011)

-0.081*
(.018)

-0.058*
(.009)

Microbial illness

-7.621E-05*
(.000)

-8.662E-05*
(.000)

-8.040E-05*
(.000)

-7.552E-05*
(.000)

-8.769E-05*
(.000)

-7.974E-05*
(.000)

Cancer deaths

-0.058*
(.011)

-0.046*
(.015)

-0.055*
(.009)

-0.048*
(.011)

-0.041*
(.016)

-0.045*
(.009)

Cancer illness

-0.008*
(.002)

-0.022*
(.003)

-0.011*
(.002)

-0.008*
(.002)

-0.020*
(.004)

-0.010*
(.002)

Program cost

-0.004*
(.001)

-0.006*
(.001)

-0.004*
(.000)

-0.004*
(.001)

-0.006*
(.001)

-0.005*
(.001)

Observations
(choice sets)

812

812

1624

740

724

1464

Log-likelihood

-786.502

-458.291

-1262.62

-716.16

-398.46

-1130.95

Notes:

a Microbial and cancer deaths and illnesses are non-proportional.
b Standard errors in parentheses. Asterisk indicates significance at the.01 level.
c Test of pooling version 3 and 4, chi-squared 32.62, critical value 14.45.
d Test of pooling version 3 and 4, chi-squared 35.65, critical value 14.45.

46

-------
TableJO^XBSCM^EstimatedJ/ViNhic^ies^

			

Full Sample

Yea-saying Observations

Removed

Variable

Version 5
(non-
proportional)

Version 6
(non-
proportional)

Pooled
Version 5 and
6

Version 5
(non-
proportional)

Version 6
(non-
proportional)

Pooled
Version 5 and
6

Microbial deaths and illnesses
(including SQ effect)a

306.180
(44.14)

163.870
(31.53)

246.110
(27.46)

288.190
(41.81)

161.780
(31.36)

237.190
(26.88)

Cancer deaths and illnesses
(including SQ effect)a

110.440
(37.30)

81.127
(20.21)

81.927
(20.13)

80.136
(38.35)

43.977
(23.89)

40.421
(24.30)

Microbial and cancer deaths and
illnesses (including SQ effect)b

554.580
(74.78)

422.660
(52.42)

496.570
(47.02)

501.150
(67.49)

386.240
(49.83)

447.040
(44.43)

Microbial deaths and illnesses
(excluding SQ effect)b

443.490
(73.25)

336.750
(55.28)

415.200
(51.24)

426.580
(70.95)

343.660
(58.59)

404.61
(53.58)

Cancer deaths and illnesses
(excluding SQ effect)b

246.539
(52.26)

255.637
(42.25)

248.650
(35.31)

214.596
(47.99)

226.663
(43.86)

209.57
(33.82)

Microbial and cancer deaths and
illnesses (excluding SQ effect)b

690.029
(111.04)

592.387
(89.77)

663.850
(79.21)

641.176
(103.50)

570.323
(94.22)

614.180
(78.96)

Marginal value of microbial death

13.616
(3.37)

11.789
(2.99)

12.659
(2.44)

12.940
(3.32)

12.601
(3.17)

12.825
(2.50)

Marginal value of microbial illness

0.019
(0.00)

0.014
(0.00)

0.018
(0.00)

0.018
(0.00)

0.014
(0.00)

0.018
(.002)

Marginal value of cancer death

14.661
(3.02)

7.403
(2.38)

12.525
(2.10)

11.763
(2.86)

6.354
(2.46)

10.011
(2.06)

Marginal value of cancer illness

1.944
(0.62)

3.581
(0.69)

2.442
(0.47)

1.864
(0.62)

3.170
(0.68)

2.176
(0.47)

Notes:

a Welfare calculations include consideration of the status quo constant. Welfare measure for 10 fewer cases MICD and CAND, and 15500 cases
fewer MICI, and 50 cases fewer CANI)

b Welfare calculations do not include consideration of the status quo constant. Welfare measure for 10 fewer cases MICD and CAND, and
15500 cases fewer MICI, and 50 cases fewer CANI).

c Standard errors in parentheses, based on Krinsky Robb simulation using 1000 draws.

47

-------
Table 11: Wald Tests of Differences in Willingness to Pay In ABSCM Models



Tests of Difference within Version



Cancer v.s Microbial

Cancer vs Both

Microbial vs Both

Joint Proportional (JP)

3.367

34.924

15.687

Joint Non-Proportional
(JNP)

9.47

22.19

4.82

Version 5 (V5)

6.12

13.98

2.92

Version 6 (V6)

2.55

10.93

4.17



Tests of Difference Across Versions



JP vs JNP

JP vs V5

JP vs V6

Cancer

0.601

0.277

0.110

Microbial

1.193

1.357

0.025

Both

0.152

0.299

0.001



Test of Non-proportional version sum of death and illness vs Proportional

effect



Including status quo
effect

Excluding status quo
effect



Microbial

3.750934

1.193



Cancer

2.691023

0.601



Note: Using willingness to pay measures that remove yea-saying observations and exclude status quo effects unless
otherwise noted.

48

-------
Table 12: Value of Statistical Life and Case Calculations from Non-Proportional Versions Based on Marginal Values

(no status quo effect)



Full Sample

Yea-saying Excludec





Version 5

Version 6

Pooled Version

Version 5

Version 6

Pooled Version



(non-
proportional)

(non-
proportional)

5 and 6

(non-
proportional)

(non-
proportional)

5 and 6

Microbial death

18,887,000

16,352,000

17,359,000

17,498,000

17,135,000

17,634,000



(4,684,700)

(4,198,300)

(3,326,200)

(4,510,100)

(4,333,800)

(3,585,000)

Microbial illness

26,567

19,103

24,815

25,188

18,591

24,013



(4,519)

(3,404)

(3,036)

(4,291)

(3,322)

(3,124)

Cancer death

20,157,000

10,092,000

16,980,000

16,021,000

8,538,000

13,559,000



(4,281,100)

(3,334,200)

(2,917,000)

(4,057,400)

(3,261,100)

(2,785,800)

Cancer illness

2,676,000

4,933,000

3,335,000

2,539,900

4,330,900

2,952,400



(876,650)

(992,250)

(652,050)

(903,860)

(943,830)

(624,130)

Note: Standard deviations are in brackets. Results are generated using 1,000 draws in a Krinsky-Robb procedure.

49

-------
Table 13: Wald Tests of Differences in Willingness to Pay



Tests of Difference CVM versus ABSCM



JP - SO excluded v.v
CVM Pooled

JP - SO included v.v
CVM Pooled

JNP - SO excluded v.v
CVM Pooled

JNP - SO included v.v
CVM Pooled

Cancer

6.92

9.01

1.97

16.05

Microbial

7.20

1.45

10.87

14.45

Note: Weibull distribution forms used. All measures use data with yea-saying observations removed.

50

-------
Table 14: Sample Models with Demographic Interactions and Attributes



Non-Proportional, Pooled Version 5 and 6
Yea-saying Observations Removed

Non-Proportional, Pooled Version 5 and 6
Full Sample

Variable

Coefficient

Coefficient/St. Er.

Coefficient

Coefficient/St. Er.

so

0.301

0.725

0.290

0.727

MICD

-0.026

-0.605

-0.026

-0.631

MICI

-7.017E-05

-2.388

-6.232E-05

-2.215

CAND

-0.022

-0.550

-0.030

-0.757

CANI

-0.028

-3.247

-0.025

-2.971

SQ*INCM

-5.284E-06

-1.949

-5.852E-06

-2.238

MD*INCM

-2.002E-07

-0.674

7.512E-08

0.266

MI*INCM

5.286E-11

0.261

-5.176E-11

-0.269

CD*IN CM

1.855E-07

0.691

1.838E-07

0.720

CI* IIVCM

1.141E-07

1.997

1.075E-07

1.982

SQ*MALE

0.528

2.766

0.542

2.972

MD *MALE

0.027

1.336

0.025

1.316

MI*MALE

-1.264E-06

-0.092

-4.300E-06

-0.334

CD "MALE

0.003

0.158

-0.005

-0.287

CI*MALE

-0.002

-0.434

-0.001

-0.377

SQ*HHSZ

0.086

1.074

0.110

1.479

MD *HHSZ

-0.007

-0.814

-0.011

-1.409

MI*HHSZ

-5.894E-06

-1.023

-8.080E-06

-1.523

CD*LLHSZ

-0.010

-1.351

-0.005

-0.727

CI*HHSZ

0.002

1.365

0.001

0.977

SQ*ENGL

-0.233

-1.003

-0.188

-0.836

MD*ENGL

0.009

0.370

0.015

0.618

MI*ENGL

2.083E-05

1.237

2.193E-05

1.355

CD*ENGL

0.019

0.845

0.010

0.452

CI*ENGL

0.003

0.643

0.000

0.089

SQ*EDU

0.094

0.460

0.237

1.222

MD*EDU

0.023

1.055

0.006

0.313

MI*EDU

7.282E-06

0.493

4.903E-06

0.355

CD*EDU

0.013

0.684

0.007

0.368

CI*EDU

0.006

1.510

0.005

1.342

SQ*1LL

-0.018

-0.322

-0.038

-0.714

51

-------
MD *ILL

-0.003

-0.519

-0.003

-0.543

MI* ILL

4.375E-06

1.099

5.065E-06

1.368

CD*ILL

0.002

0.347

0.004

0.763

CI* ILL

3.477E-04

0.301

5.473E-04

0.509

SQ*URBAN

0.443

2.144

0.385

1.966

MD*URBAN

-0.032

-1.454

-0.029

-1.376

MI*URBAN

-1.757E-05

-1.185

-1.370E-05

-0.986

CD*URBAN

-0.002

-0.083

0.003

0.139

CI*URBAN

0.005

1.178

0.005

1.290

SQ*BLIEF

0.131

0.578

-0.007

-0.029

MD *BLIEF

-0.018

-0.743

-0.018

-0.783

ML*BLLEF

-3.107E-05

-1.887

-2.753E-05

-1.732

CD*BLIEF

-0.038

-1.706

-0.049

-2.290

CL*BLLEF

-0.004

-0.845

-0.003

-0.676

SQ*AVERT

0.116

0.595

0.156

0.839

MD*AVERT

-0.008

-0.397

-0.006

-0.297

MI*A VERT

1.565E-05

1.123

1.308E-05

0.990

CD*AVERT

-0.014

-0.724

-0.012

-0.690

CI*A VERT

-0.001

-0.326

-0.003

-0.840

SQ*AGE65

-0.278

-0.963

-0.342

-1.246

MD*AGE65

0.025

0.812

0.026

0.902

ML*AGE65

0.000

-0.349

0.000

-0.131

CD*AGE65

-0.010

-0.350

0.004

0.157

CI*AGE65

-0.005

-0.718

-0.004

-0.676

BILL

-0.005

-9.068

-0.005

-9.313











Number of observations

1464



1624



Log likelihood

-1082.634



-1206.9



52

-------
90

80

70

60

50

40

30

20

10

0

Figure 1. Percentage of "Yes" Responses by Bid Value (Pooled

versions)

78.65





68.54 68.63





U /



t

58.82 57 14





¦9.o;

)





47.62 ac oo ac nn











4





39.58











32.1



!9.1

7



27.50













o
o

)





























25	75	150	250	350

~	Cancer and Microbial Question ¦ Cancer Question	Bid Value (Can $)

~	Microbial Question

53

-------
References

Abt Associates Inc. Arsenic In Drinking Water Rule Economic Analysis Office of Ground Water and
Drinking Water. U.S. Environmental Protection Agency. USEPA 815-R-00-026. (December 2000). 257

pp.

Adamowicz, W., P. Boxall, M. Williams and J. Louviere (1998) Stated preference approaches to
measuring passive use values. American Journal of Agricultural Economics (80):64-75.

Adamowicz, W., J. Louviere, J. Swait. Introduction to Attribute-Based Stated Choice Methods. Final
Report to Resource Valuation Branch, Damage Assessment Center, NOAA, U.S. Department of
Commerce. (January 1998) 44pp.

Arrow, K., R. Solow, P. Portney, E. Learner, R. Radner and H. Schuman (1993)"Report of the NOAA
Panel on contingent valuation" Federal Register. US Department of Commerce. 58(10), 4602-4614.

Barbeau, B.; P. Payment; J. Coallier; B. Clement, and M. Prevost "Evaluating the Risk of Infection from
the Presence of Giardia and Cryptosporidium in Drinking Water" Quantitative Microbiology 2 (1): 37-54,
March 2000.

Carson, Richard T.; Mitchell, Robert C.; Hanemann, W. Michael; Kopp, Raymond J.; Presser, Stanley;
and Ruud, Paul A., A Contingent Valuation Study of Lost Passive Use Values Resulting From the Exxon
Valdez Oil Spill (Anchorage: Attorney General of the State of Alaska, November 1992).

Carson, R. T. and R. C. Mitchell "Public Preference Toward Environmental Risks: The Case of
Trihalomethanes" Department of Economics, University of California, San Diego Discussion Paper 2000-
21. (August 2000). 37 pp.

Cochran, W. G., and G. M. Cox. Experimental Designs. New York: John Wiley & Sons. 1957.

Corso, P., J. Hammitt, and J. Graham, "Valuing Mortality-Risk Reduction: Using Visual Aids to Improve
the Validity of Contingent Valuation." Journal of Risk and Uncertainty. 23 (2) (2001): 165-184.

de Wit, M; M. Koopmans; L. Kortbeek; W. Wannet; J.Vinje; F. van Leusden; A. Bartelds and Y. T. van
Duynhoven "Sensor, a Population-based Cohort Study on Gastroenteritis in the Netherlands: Incidence
and Etiology" American Journal of Epidemiology Vol. 154, No. 7 : 666-674. 2001.

Dillman, D.A. Mail and Telephone Surveys: The Total Design Method. John Wiley and Sons, New York.
1978.

Dillman, D.A. Mail and Internet Surveys: The Tailored Design Method. John Wiley and Sons, New York.
2000.

Fabian, Robert. "The Qualy Approach." In Valuing Health for Policy: An Economic Approach. Edited by
G. Tolley, D. Kenkel, and R. Fabian.. Chicago, University of Chicago Press. 1994.

54

-------
Federal-Provincial-Territorial Committee on Drinking Water. Summary of Guidelines for Canadian
Drinking Water Quality (April 2004). Safe Environments Program, Health Canada. Government of
Canada publication, http://www.hc-sc.gc.ca/ewh-semt/alt formats/hecs-sesc/pdf/pubs/water-eau/doc-sup-
appui/sum guide-res recom/summary-sommaire e.pdf.

Garber, A. C., et al. "Theoretical Foundations of Cost Effectiveness Analysis." In Cost-Effectiveness in
Health and Medicine: Report to U.S. Public Health Service by the Panel on Cost-Effectiveness in Health
andMedicine. Edited by Marthe R. Gold, et al., U.S. Department of Health and Human Services. 1996.

Haab, T. C. and McConnell, K. E. Valuing Environmental and Natural Resources. Edward Elgar
Publishing, London. 2002.

Hammitt, J. K. "Risk Perceptions and Food Choice: An Exploratory Analysis of Organic- Versus
Conventional-Produce Buyers," Risk Analysis. 10(1990): 367-374.

Hammitt, J.K., and J.-T. Liu, "Effects of Disease Type and Latency on the Value of Mortality Risk,"
Journal of Risk and Uncertainty 28: 73-95, 2004.

Hanley N., Mourato S. and Wright R.E. (2001) Choice modeling approaches: a superior alternative for
environmental valuation? Journal of Economic Surveys 15(3).

Havelaar, A., A. De Hollander, P. Teunis, E. Evers, H. Van Kranen, J. Versteegh, J. Van Koten, and W.
Slob. "Balancing the Risks and Benefits of Drinking Water Disinfection: Disability Adjusted Life-Years
on the Seale." Environmental Health Perspectives. 108(4): 315-321. 2000.

Health Canada. Cancer Surveillance On-Line. http://cythera.ic.gc.ca/dsol/cancer/index_e.html (20 March
2003).

Holmes, T. and W. Adamowicz. "Attribute Based Methods". A Primer on the Economic Valuation of the
Environment edited by P. Champ, T. Brown and K. Boyle. Kluwer, Amsterdam. 2002.

Jardine, C.G. N. Gibson and S. Hrudey. "Detection of Odour and Health Risk Perception of Drinking
Water." Water Science and Technology Vol 40(6) (1999): 91-98.

Jones-Lee, M. W., M. Hammerton, and P. R. Philips. "The Value of Safety: Results of a National
Survey " Economic Journal. 95 (1985): 49-72.

King W. "Great Lakes Water And Your Health. A Summary Of "Great Lakes Basin Cancer Risk
Assessment : A Case Control Study Of Cancers Of The Bladder, Colon And Rectum" Health Canada
publication. December 1995.

King W.D., L. D. Marrett. "Case-Control Study Of Bladder Cancer And Chlorination By-Products In
Treated Water (Ontario, Canada)." Cancer Causes and Control 7 (1996): 596-604.

Krewski, D. et al "Health Risk Perception In Canada. II: Worldviews, Attitudes, And Opinions," Human
and Ecological Risk Assessment. Vol. 1(2) (1995): 53-70.

55

-------
Krewski, D. J., Balbus, D. Butler-Jones, C. Haas, J. Isaac-Renton, K. Roberts, and M. Sinclair Managing
Health Risks from Drinking Water Commissioned Paper 7 for The Walkerton Inquiry.

Queen's Printer for Ontario, Ontario Ministry of the Attorney General. 2002.

Krupnick, A., A. Alberini, M. Cropper, N. Simon, B. O'Brien, R.Goeree, and M. Heintzelman, "Age,
Health, and the Willingness to Pay for Mortality Risk Reductions: A Contingent Valuation Survey of
Ontario Residents." Journal of Risk and Uncertainty. 24(2002): 161-186.

Louviere, J., D. Hensher and J. Swait. Stated Choice Methods - Analysis and Application. Cambridge,
Cambridge University Press. 2000.

McFadden, D. "Conditional Logit Analysis Of Qualitative Choice Behavior." Frontiers in Econometrics
edited by P. Zarembka. New York, Academic Press. 1974. 105-142.

Mead, P.; L. Slutsker; V. Dietz; L. McCaig; J. Bresee; C. Shapiro; P. Griffin, and R. Tauxe "Food-Related
Illness and Death in the United States" Emerging Infectious Diseases 5(5):607-625. 1999

Mills, Christina J., Bull, Richard J., Cantor, Kenneth P., Reif, John , Hrudey, Steve E., Huston, Patricia
and an Expert Working Group "Workshop Report Health Risks of Drinking Water Chlorination By-
products: Report of an Expert Working Group" Chronic Diseases in Canada 19(3).

Mitchell, R.C. and R.T. Carson. Using surveys to value public goods : the contingent valuation method.
Resources for the Future, Washington, DC. 1989.

Payment, P., Berte, A., Prevost, M., Menard, B., and Barbeau, B. "Occurrence of pathogenic
microorganisms in the Saint Lawrence River (Canada) and comparison of health risks for populations
using it as their source of drinking water" Canadian Journal of Microbiology 46 (2000):565-576.

Payment, P., Franco, E., Siemiatycki, J., Richardson, L., Renaud, G., Franco, E., and Prevost, M. "A
Prospective Epidemiological Study Of Gastrointestinal Health Effects Due To Consumption Of Drinking
Water. Int. J. Environ. Health Res. 7 (1997): 5-31.

Payment, P., Richardson, L., Siemiatycki, J., Dewar, R., Edwardes, M., and Franco, E. "A Randomized
Trial To Evaluate The Risk Of Gastrointestinal Disease Due To The Consumption Of Drinking Water
Meeting Currently Accepted Microbiological Standards" Am. J. Public Health 81 (1991): 703-708.

Payment, P. and M. Riley "Resolving the Global Burden of Gastrointestinal Illness: A CALL TO
ACTION" American Academy of Microbiology (2002)

Ronchi, E. and S. Wald "New molecular technologies for safe drinking water" OECD Observer No. 215 -
January 1999.

Torrance, George W "Measurement of Health State Utilities for Economic Appraisal: A Review." Journal
of Health Economics. 5(1)1986: 1-30.

United States Environmental Protection Agency Alternative Disinfectants and Oxidants. April 1999.

56

-------
J. Wheeler; D. Sethi; J. Cowden; P. Wall; L. Rodrigues; D. Tompkins; M. Hudson, and P. Roderick
"Study of infectious intestinal disease in England: rates in the community, presenting to general practice,
and reported to national surveillance" BMJ 1999;318:1046-1050 ( 17 April ). 1999.

57

-------
Appendix 1

Below are the descriptions of the health effects from microbial illnesses and bladder cancer
illnesses relating to the drinking of tap water presented in the survey.

Health Effects of Chlorine

When tap water is disinfected with chlorine, various by-products including Trihalornethanes
(THMs) are produced. Scientists believe that THMs are an indicator for substances in the
tap water that are linked to increased cases of bladder cancer when water is consumed
over long periods of time.

~	Symptoms of bladder cancer:

-	Urgent and frequent need to urinate, blood in your urine, pain during urination, and
pain from the tumour.

-	Symptoms for this cancer do not occur immediately after drinking tap water, rather
they take years to show since it takes years for this cancer to develop.

-	For about one in five cases, death occurs within five years from diagnosis.

~	Medical Treatment of Illness:

-	Surgery, radiation, and chemotherapy are used to treat bladder cancer.

-	Side effects from surgery may include a long recuperation period and the need for
colostomy (bag for body wastes).

-	Side effects of chemotherapy include loss of hair, change in taste or smell, mouth
sores, possible loss of fertility, fatigue and less ability to deal with infections.

~	Sensitive Groups:

-	Occurs most frequently in male smokers over the age of 70, but other older people
can also get this cancer.

~	Tap Water Treatment:

-	Providers of tap water can lower the chlorine levels in the water supply.

-	Less chlorine lowers cancer risks but raises microbial risks.

-	More expensive water treatment technologies are available to reduce both cancer

s E-mail: questions @i-say. com

Phone: 1-866-893-1133

ri^U-^ anrl mirrnhi'al riQU'Q

58

-------
BlHi-Ml

Bl E-mail: questjons@i-sav.com	1* Phone: 1-866-893-1138

Health Effects of Microbes and THMs in Tap Water

You won't need to remember these numbers. We just want to give you some idea of the risks
people face.

First we list effects from all causes, then we list effects from drinking tap water only.

Microbial Health Effects in Numbers

Cancer Health Effects in Numbers

From all causes of microbial disease

~ Scientists estimate that for every 100,000
people:

-	Over a 35-year period, microbes from
all sources (food, tap water and direct
contact such as swimming), lead to 2.5
million cases of microbial infection. This
means that a person may likely suffer
multiple episodes of microbial illness over
this period.

-	Over a 35-year period, about 100
deaths occur from microbes from all
sources.

1

From all causes of cancer

~ Scientists estimate that for every 100,000
people:

-	Over a 35-year period, 27,000 people will
contract cancer of all types.

-	Of these 27,000 people, 7,000 deaths are
due to cancer of all types.

I

From drinking tap water

~ Scientists estimate that for every 100,000
people drinking tap water:

-	Over a 35-year period, 23,000 people
will get some sort of microbial infection.

-	Of those infected, 15 will die over the
35-year period. Death often occurs soon
after infection.

From drinking tap water

~ Scientists estimate that for every 100,000
people drinking tap water:

-	Over a 35-year period, 100 people will
contract bladder cancer.

-	Of these, approximately 20 persons will
die within 5 years as a direct consequence
of the cancer.

-	Out of the 80 who do not die, some will
be fully cured; others will experience cancer
symptoms, and require medical interventions
and drugs over their remaining lifetime.

This information is summarized in the following screen.

Sources for Health Effects Estimates

59

-------
(§)lpsosi-Say

El E-mail: questions@1-say. com	I4 Phone: 1-S66-S93-11SS

For a community of 100,000 people, over a 35-year period, illnesses and deaths from
microbial disease and cancer will be approximately...

MICROBIAL DISEASE

CANCER



Illnesses

Deaths

Illnesses

Deaths

From all
Causes

2,500,000

100

27,000

7,000



f

t

t

t

From









Prinking Tap

23,000

15

100

20

Water









On the next two screens, this situation is shown with pictures.

60

-------
Appendix 2: Example of CVM Question Format (Version 2)

(§) ,psosi-Say

^ E-mail: questions@i-say.CQm	Phone: 1-866-893-1188

Here's the first program we want you to vote on.

THE BENEFITS OF MUNICIPAL WATER TREATMENT PROGRAM A

Based on current water drinking patterns in your community this program would have the following benefits to every 100,000 people:

•	15,500 fewer people will develop microbial illness over a 35-year period. Another way to say this is that the average person in a community of 100,000 people will
see their risk of getting microbial illness from drinking the waterfall from 23,000 in 100,000 to 7,500 in 100,000

•	With fewer people developing microbial illness, 10 fewer people will die from getting the disease. Another way to say this is that the average person in this
community will see their risk of dying from microbial illness reduced from 15 in 100,000 to 5 in 100,000

•	Bladder cancer illness and deaths will not be affected by the program

Here is a table showing these benefits:

For every 100,000 people, the
NUMBER who would...

CURRENT SITUATION

PROPOSED PROGRAM A

Get sick from microbial illness
in a 35-year period

23,000

7,500

Die from microbial illness in a
35-year period

15

5

Get sick from bladder cancer in
a 35-year period

100

100

Die from bladder cancer in a
35-year period

20

20









Out of 100,000 people...

Peopte who would get
microbial ilness

m People wtio would get
— bladder cancer

People who would die from
• microbial ilness or
bladder cancer

Remainrig
population









THE COST OF THE MUNICIPAL WATER TREATMENT PROGRAM A

If the majority of voters support this program your household will share in the cost starting January 2005 by paying an additional amount on your household water bill.
PLEASE VOTE NOW:

BfSSfcfl If the estimated addition to your household's water bill was $25 per year ($2.08 per month) starting in January 2005,
and a vote were held today, would you vote FOR or AGAINST the proposal?

FOR
r AGAINST



61

-------
Appendix 3: Example of ABSCM Question Format (Version 5)

This is the second scenario we want you to vote on.

For every 100,000 people, the
NUMBER who would...

CURRENT SITUATION

PROPOSED
PROGRAM A

PROPOSED
PROGRAM B

Get sick from microbial illness
in a 35-year period

23,000

23,000

7,500

Die from microbial illness in a
35-year period

15

15

5

Get sick from bladder cancer in
a 35-year period

100

50

75

Die from bladder cancer in a
35-year period

20

10

15

Change to your water bill
starting in January, 2005

No Change

Increase $350 per year
($29-17 per month)

Increase $25 per year
($2.08 per month)













Out of 100,000 people...

People who would get
microbial ilness
mm People who would get
bladder cancer

People wtio would die from
microbial ilness or
bladder cancer

Remaining
population











If there were a referendum, I would vote for...

CHECK ONE ONLY

f Current Situation
Proposed Program A
Proposed Program B

62

-------
Endnotes

1	The chlorine demand of the water is defined as: the amount of chlorine that reacts with the other chemicals in the water plus
the amount required to achieve disinfection. In addition, however, utilities add extra chlorine added to the water to account for
length of time in the distribution network. This is called free chlorine and is the culprit in the production of disinfection by-
products such as Trihalomethanes.

2	Two epidemiological studies suggest that drinking water from water treatment plants following standard treatment processes
could be responsible for half of the cases of gastrointestinal illnesses in the receiving population (Payment et al. 1991, 1997).

3	The typical range of annual household water bills in Canada is between $300 and $500.

4	Attribute levels for microbial illnesses were 7500, 15000, 23000 and 30000. Attribute levels for microbial deaths were
5,10,15 and 20. Attribute levels for cancer illnesses were 50,75,100 and 125. Attribute levels for cancer deaths were 10,15,20
and 25. All were defined for a population of 100.000 and over a 35 year period. Annual increases to household water bills
ranged between $25 and $350.

5	$14.4 million = $535/50 cases* 100,000*35 years/2.6 persons per household.

63

-------
Trish Hall's Discussant Comments for
Session IV: Valuing Morbidity and Mortality: Drinking Water

• Policy Implications: Adamowicz, Dupont, Krupnick

o Preliminary Conclusion:

¦	Higher VSL values for microbial illness

¦	Significant value for illness avoided

¦	Altruism: results can't be used at this point
o Interpretation:

¦	Canadians more aware of waterborne disease outbreaks?

•	Was also surprise by this result but agreed with the authors
that Walkerton etc... may have had an impact

•	Impacts on kids/sensitive sub-populations

o Did folks know from outbreaks that these groups
are more adversely impacted by microbial illness?

•	Bladder cancer description

o Average age of onset described as 70 years old

¦	WTP values for illness very useful

•	WTP for Mircrobial illness was significant but did not
seem unreasonable but will be heavily scrutinized if we
were to use

¦	Combined case avoided valuations could also be beneficial

•	Combined value for per case avoided would avoid the need
to estimate mortalities and also severities

¦	Altruism: Can it be sorted out

•	But how to tease out... almost everyone drinks from a
public supply at some point so how do we figure this out
(Shaw brings this up)

o Points to Consider

¦	Canada vs. USA

•	Make sure tables and text clearly indicate CAD $

•	Benefit transfer issues

¦	Description on TTHM impacts could perhaps change results

•	Would results be different with these additional
descriptors?

o Routes of exposure: dermal and inhalation
o Other health impacts: other cancers and repro and

developmental
o Exposure varies through out distribution system

•	However, it could make it more difficult to conduct benefit
transfer to other contaminant where this is not the case
(such as arsenic).

¦	Latency vs. Cessation lag (i will talk about this at the end)

1

-------
•	Policy Implications: Shaw et al.

o Conclusion: None yet but...

¦	Potentially useful for understanding this issues... even if valuation
proves elusive

o Potential implications

¦	Averting behavior: does it also correlate with greater WTP?

•	Averting behavior:? If valuation can be obtained, do the
results match those in the Adamowicz paper?

-Do folks who take averting action also have higher WTPs?

¦	Altruism vs. self-interest: benefits from transient water supply
regulation for chronic contaminants

•	Currently, the Safe Drinking Water Act exempts transient
water supplies (e.g. restaurants, truck stops) from chronic
contaminant regulation (such as arsenic). Are folks
concerned about these outside the home exposures even if
they have their own well?

¦	How can we improve risk communication?

•	Focus group shows that we need to work on putting risk in
context

¦	Potential valuation estimates

•	Would always be welcome

•	Points to Consider: Shaw et al.

o Addressing ambiguity in risk

¦	Some Clarity: Do you use tap water? —very good that researchers
clarified uses for cooking, making ice, etc... many people don't
realize how much they actually ingest.

¦	BTW: fountain sodas are also made with tap water!

o Lots of new arsenic risk information that could address some ambiguity

¦	Arsenic inhibits DNA repair

•	Arsenic may be a "promoter" of cancer and prevent the
body from making repairs to damaged DNA

•	Would describing this process help people understand the
risk?

¦	In utero Arsenic exposure and lung disease

•	UC Berkeley: Arsenic Health Effects Research Program (in
utero study)

o Sources of data

¦	Would not recommend using Burnett/Hahn report for benefit
estimates or risk data

•	many inaccuracies regarding the Arsenic Rule
o Latency vs. Cessation Lag

¦	Same issue as with the Adamowicz paper

¦	Agency prefers cessation concept

¦	Some examples of it use can be found here:

2

-------
•	See SAB report and Stage 2 DBPR EA for more
information

•	http://www.epa.gov/sab/pdf/ecQ1008.pdf

•	http://www.epa.gov/OGWDW/disinfection/stage2/regulatio
ns.html

•	I'll briefly describe the differences between latency and
cessation next

• Note about Cessation Lag

o Outlined in EPA's Science Advisory Board's Arsenic Rule Benefits
Review Panel

¦	Benefits analysis based only on latency greatly underestimates
actual benefits

¦	A good example of this is smoking:

•	Latency: initial exposure and increase in lung cancer risk is
~ 20 years

•	Cessation: risk of lung cancer declines quickly with
reduced exposure

•	Smoking probably both imitator and promoter of
carcinogenic effects.

Promoters should see more rapid decline in risk (i.e.
late stage actor not the one that started the problem)
Arsenic seems to be a promoter

Perhaps does not cause DNA damage but
inhabits DNA repair

•	The Final Stage 2 DBPR expands on the work of the SAB
and includes cessation models for: smoking/lung cancer,
smoking/bladder cancer, and arsenic/bladder cancer.

3

-------
Valuing Reductions in Health Risks from Drinking Water: Discussion

Gregory L. Poe
Associate Professor
Department of Applied Economics and Management
Cornell University
GLP2 @cornell. edu

It is a distinct pleasure to participate in this workshop, and to have the opportunity to
focus my attention on a group of research efforts directed toward exploring methods of
conceptualizing and measuring the economic benefits of reducing health risks from
drinking water. Individually and collectively the presentations in this session meet what
I view to be the objective of EPA, and more specifically EPA STAR, funded research and
collaboration: to make methodological contributions while remaining policy informative.

While I enjoyed and learned from each of the four presentations in this session, and
appreciate that they are at varying levels of completion, I have been asked to center my
present discussion on the presentation by Vic Adamowicz, Diane Dupont, and Alan
Krupnick (hereafter ADK). Of the four research efforts comprising this session, the work
by this research group is the furthest along and the only one in a position to provide a
manuscript to accompany the oral presentation.

Although it is not funded through the STAR program, the ADK research is clearly in the
spirit of EPA STAR objectives ascribed above. ADK does offer a methodological
contribution to a contemporary debate in non-market valuation by comparing
willingness-to-pay value estimates obtained from a contingent valuation (CV) study with
those obtained from an Attribute Based Stated Choice Method (ABSCM). More
colloquially this latter method is referred to a choice modeling or a variant of conjoint
analysis. Although the research was conducted in Canada, ADK's findings are relevant to
water quality policy in the United States. The tradeoff between microbial contamination
and the cancer risks associated with byproducts of chlorination (i.e., Trihalomethanes) are
fundamental to the Surface Water Treatment Rule, the Disinfectant/Disinfection
Byproducts Rule, and the Groundwater Rule (http://www.epa.gov/safewater/dwa/
electronic/ematerials.html#npdwr). The apparent high quality of this research suggests to
me that KDM will make a notable and lasting contribution to both the literature on
research methods and applied policy analysis.

The remainder of my comments is organized around central themes raised in ADK. With
an eye toward addressing ADK's (p. 33) expressed concern that the value of statistical
lives (VSL) that they find in their research "falls in the upper range of [previously
estimate VSL] value," and the "fact that VSL for deaths from microbial disease is
somewhat greater than that for cancer is a big surprise", the following sections discuss
issues related to risk communication, the valuation of private versions public risks, and
ADK's design and comparisons of stated preference methods.

1

-------
Risk Communication:

Communicating drinking water risks in a manner that induces reasonable protective
behavior when appropriate and reasonable inaction when exposure levels are well within
safety levels is not a simple task. For instance, a recent arsenic risk communication study
that endeavored to bring together concepts "of information processing, mental models
and health behavior" into a single model of health behavior theory identified 45 possible
variables in the path from arsenic exposure level to protective behavior (Severtson,
Baumann, and Brown).

Economists, however, are more parsimonious in their characterization of risk updating
with respect to new information. One such model treats an individual's subjective
posterior risk assessment (Rp) as a function of prior risk perceptions (Ro) and the
subjective risk associated with the information message (Ri) (see Smith and Johnson). A
simple form of this relationship, which is consistent with many updating models, is a
weighted linear average:

Rp = wo Ro + (1- wo)Ri

where w0 is the weight placed on the prior risk perceptions. In turn 'I' contains general
information about contaminants and their effects and exposure information. Past research
using this simple updating framework has demonstrated that in making informed risk
assessment, individuals place significant weight on both prior perceptions and new
information for various health risks (e.g. radon, Smith and Johnson; chemical labeling,
Viscusi and O-Connor; nitrates in groundwater, Poe and Bishop).

The above relationship has implications for ADK's analysis and conclusion. Of
overarching importance, it implies that Rp ^ Ri. Related to this is the supposition that
individuals likely have, and place weight on, prior perceptions of exposure and health
risks from drinking water in characterizing their reference risk. Hence the respondents'
subjective assessment of how the proposed program would affect the risk that they
(individually or collectively) face will typically not align with the "objective" change
presented (and modeled) in the research.

I posit that these implications shed light on ADK's "upper range" VSL finding indicated
previously. Specifically, prior perceptions of health effects may be artificially large
because of high profile microbial contamination events in Walkerton Ontario and North
Beettleford, Saskatchewan (p. 3, p. 33, ADK). If Rp is elevated relative to the exposure
and risk information provided by the researchers, and supposing that respondents take the
target exposure level at face value, then respondents will be valuing a larger change than
indicated. Dividing this larger value by the smaller change in "objective" risk conveyed
in the survey materials would engender upwardly biased VSL estimates.

Whilst I find it innovative, I worry too that the "snake in the sand" communication
approach that presents both microbial and cancer risks in the same diagram could lead to
disproportional focus on the change in microbial risk relative to that associated with

2

-------
cancer. In examining the question formats in Appendices 2 and 3,1 was taken by the fact
that microbial exposure risks were, in essence, represented by an area, and cancer risks
by a line. Although the changes in risks are proportional, to me the change in area
associated with microbial risks loomed much larger. Should this optical "illusion" carry
over to respondents, it would cause a further deviation between the change in objective
risks communicated in the survey and the subjective risks utilized by the respondents.

In identifying these issues of subjective and prior risks, I do not mean to imply that ADK
somehow failed in their efforts to communicate risk and risk changes. Indeed, I would
argue the opposite. I am genuinely impressed with ADK's efforts to accurately
understand and communicate the risks facing individuals, and would rate their work quite
high relative to previous valuation work in groundwater risks. Nevertheless, I do believe
that more could (can still?) be done with respect to understanding the subjective risks that
individuals used as a base for formulating their willingness-to-pay values. Enhanced
understanding of subjective risk, perhaps gleaned from a much smaller, shorter follow up
survey or other auxiliary information, would provide an informative step toward better
understanding the reported values and their relationship to prior work on groundwater
and more general VSL studies. It is in this area of understanding what the respondents
are valuing that I particularly commend the preliminary work presented by Douglass
Shaw in this same session.

Altruism and Public Values:

ADK are correct in highlighting the fundamental difference between private and public
valuation exercises and its impact on how we are to interpret value estimates, particularly
with respect to comparisons with VSL estimates. Whereas groundwater quality is a
public good, "best" estimates of the value of a statistical life derive largely from
individual choices made in wage or market place studies (although CV and averting
behavior studies have also been conducted and utilized in VSL estimates). As one moves
from the private to public arena, other-regarding preferences enter into an individual's
valuation equation, leading, potentially, to incomparable value estimates between public
and private risk valuation exercises.

While fairness, reciprocity and other concerns are key elements of the set of other-
regarding behaviors, ADK limit their concerns to "elements of altruism". That altruistic
preferences are a concern in the valuation of safety is made evident in Viscusi, Magat and
Forrest's work which compared willingness-to-pay values for personal risk reductions
with willingness-to-pay values for programs that reduce the risk to others. They report
that the sum of altruistic values for the risk reductions of other individuals are as high as
six to seven times the value of reduction placed on an equivalent reduction in individual
personal risk.

Economists have classified at least three types of altruistic preferences, each with a
differing economic-theoretic role in benefit-cost analysis. The first is deemed "pure"
altruism, reflecting the fact that I care for the well being or utility of others (Bergstrom,
2006). A second form is paternalistic altruism, which refers to the fact that I derive

3

-------
utility from how you consume (eat your peas! and don't take drugs!) and derive your
utility (Jones-Lee, 1992). The third is Andreoni's "impure" altruism (or warm glow
giving) in which I derive egoistic utility simply from the act of giving, independent of the
particular good in question (Andreoni). ADK's paper implies that they interpret
economic-theoretic benefit discussions of the role of the various forms of altruism in
welfare assessments to imply that it is appropriate to "purge our estimates" (p. 35) of
impure and pure altruistic motives1. But to the extent that altruistic preferences are
paternalistic or safety oriented, they should be accounted for in benefit-cost analyses of
risk reductions. I concur with this assessment.2

I do, however, dispute ADK's interpretation that pure altruism necessarily inflates values
relative to private values. As Bergstrom (2006) reminds us "we should not forget... to
count sympathetic losses each bears from the share of its costs paid by the other" (p.
339). The potential for such costs is of particular concern in the discrete choice
framework employed in the stated preference elicitation formats utilized in ADK.
Johannesson et al. argue that the coercive nature of voting and taxation raises the
possibility that some people who are pure altruists will vote "no" on a project that would
provide them private net benefits for risk reduction, narrowly defined, because they
desire not to impose costs on others for whom costs exceed the benefits.

Let us assume that [an individual] is willing to pay $t for a ceteris paribus
increase in his own safety. His total WTP for a uniform public risk
reduction of the same magnitude will fall short of $t if he believes that
others are willing to pay less than $t but will still be forced to pay that
amount ($t) for the project. This is because other individuals, for whom
he cares will experience a lower utility if the program is implemented. In
turn, this decrease in the utility of others reduces the pure altruist's WTP
for the public safety project, (p. 264)

In other words, purely altruistic behavior may in some instances lower the
proportion of affirmative votes relative to a self-interested model. Johanneson et

1	There is continuing debate in the economic literature regarding the role of pure altruism in benefit-cost
analyses. Conventional economic wisdom suggests that the optimal provision of public goods should be
based solely on selfish preferences (Bergstrom, 1982; Jones-Lee, 1991, 1992; Milgrom; Johansson) in
social benefit-cost analyses for small projects evaluated close to a social welfare optimum. However, as
Flores notes, public projects are rarely, if ever, financed under such conditions: most typically the funding
for specific public projects imposes coercive costs that result in utility gains and losses. Moreover, projects
evaluated tend to be discrete, and the initial allocation of public goods is inefficient. Under these conditions
the extrapolation of Bergstrom's (1982)result for marginal changes at the optimum do not carry over to the
"more modest problem [of benefit-cost analysis], determining whether a specific project can lead to a
Pareto improvement" (Flores, p. 304). While Bergstrom (2006) does not dispute Flores' argument he
concludes that "for a broad class of economies, a comparison of the sum of private values to the cost of a
project is the appropriate test for determining whether it can lead to a Pareto Improvement" (p. 348).

2	In the spirit of full disclosure, I should not that Professor Richard C. Bishop, an attendee at this
conference, Professor Emeritus at the University of Wisconsin, longstanding leader in non-market
valuation research, and the Chair of my dissertation committee indicated, after my presentation,
disagreement with the exclusion of warm glow giving from benefit-cost analyses. At the time of this
writing we have not yet had the opportunity to determine our point of departure on this issue.

4

-------
al. demonstrate this outcome in a dichotomous choice contingent valuation study
of safety. In an experimental study of willingness to pay for protection against
financial risks in coercive tax settings, Messer, Poe and Schulze further
demonstrate this result.

With respect to warm glow, I agree with ADK that warm glow should be removed
from value estimates for use in benefit-cost analyses, but disagree with their
method of doing so. To isolate warm glow respondents in the CV format, ADK
"removed people who said that they would pay anything for health risk reductions
and who answered Yes-Yes" (p. 20: for ABSCM they simply removed individuals
who said that they would pay anything). These types of people are best
categorized as yea- sayers, not warm glow respondents. Warm glow need not be
large. And it could be a small or large element of every respondent's values.
Hence, it appears the removal of selected yea-sayers from the data set bears little
relation to removing warm glow values from the entire data set of respondents.

In sum, I concur with the intent of the last sentence in ADK's paper, "Attempts to
purge our estimates of altruism effects are in our plans for future research," and
heartily urge the authors to undertake this effort. In doing so, however, they must
take care to do so in a manner consistent with the underlying economic-theoretic
construct.

The CV and ABSCM studies:

Overall the survey implementation and the analyses seem to be, as already
suggested, of high quality (as I would expect from this set of co-authors). My
comments on the survey design tend to be of a more specific rather than general
nature, and hence, I shall rely on a bulleted format to convey my impressions.

¦	Both modes: The 46% response rate is relatively low by contemporary
stated preference standards for established methods of survey research
such as mail, telephone or in-person contacts. Web-based survey
research is still fairly nascent and it is not clear at this time what
response rate expectations and non-response implications for this
mode. Nevertheless, it is a concern for any policy research when the
response rate falls below 50%.

¦	CV format:

- Valid comparisons of adding up should account for the likely
positive correlation in DC-CV responses (or more specifically
error terms) across risk scenarios. Failing to account for this
difference will lead to biased estimates of the significance of the
difference between the parameters being compared (see Poe,

Welsh and Champ)

There appears to be a "fat tails" problem at upper bids (45% yes to
the joint treatment of Cancer and Microbial) which should be
accounted for/addressed in estimating the mean WTP values.

5

-------
¦	ABSCM format:

- As ADK note, a fundamental question arises with the observation
that the inclusion/exclusion of the status quo (a difference on the
order of 50%). This is a concern, in part, because the authors
provide little guidance about which of the two measures is
appropriate.

¦	Comparing CV and ABSCM:

ADK note that the "WTP measures from the [ABSCM] models
with the status quo effect included tended to be lower that the
CVM values while the WTP form the models with the status quo
effects excluded tended to be higher than the CVM models" (p.
32). Either result is of interest as well as of concern. The former
result is of interest because it is not consistent with previous
comparisons of CVM and ABSCM that have found that ABSCM
values are not significantly different or are statistically higher than
CVM (see Boyle, Morrison and Taylor). In contrast the latter
results are consistent with the previous literature, but that is a
concern. Here, ADK use a dichotomous choice CV format, which
has been demonstrated to engender the highest deviations between
hypothetical and actual values in simulated market studies (e.g.,

Brown et al). It would thus be disappointing to find that ABSCM
provides higher values than the most upwardly biased CV format.

Concluding Thoughts:

My sense is that ADK have designed a study that provides one of the fairest and
competent comparisons of CV and ABSCM. I believe that this research will make
a notable contribution to the stated preference literature. The research also has
high potential for informing policy. As I see it the only shortcoming of this
research is that, I suspect, the present statistical analyses are far from final and
that there are several issues, some of which I have raised above, that merit closer
consideration as this research is brought to completion. I do look forward to
reading revised and updated analyses of this work, and maintain that the EPA and
the other agencies that have funded this research have made a solid investment
that will, in time, make a lasting contribution to the dual objective of policy and
methods in the valuation of drinking water risks.

References:

Adamowicz, V. D. Dupont, and A. Krupnick, 2006. "Willingness to Pay to Reduce
Community Health Risks from Municipal Drinking Water: A Stated Preference
Study." Conference Draft for US EPA NCER/NCEE Workshop on Morbidity and
Mortality: How Do We Value the Risk of Illness and Death? Apr. 10-12, Washington
DC.

6

-------
Andreoni, J., 1990. "Impure Altruism and Donations to Public Goods: A Theory of
Warm Glow Giving," Economic Journal 100: 464-467.

Bergstrom, T.C., 1982. When is a Man's Life Worth More Than His Human Capital?" in
M. Jones-Lee ed. The Value of Life and Safety. Amsterdam:North-Holland, pp. 363-
398.

Bergstrom, T.C., 2006. "Benefit-Cost in a Benevolent Society " American Economic
Review, 96(1):339-351.

Boyle, K., M. Morrison, and L. Taylor, 2004. "Why Value Estimates Generated Using
Choice Modeling Exceed Contingent Valuation: Further Experimental Evidence."
Paper Presented at the Australian Agricultural and Resource Economics Society
Conference, Feb. 11-13, Melbourne.

Brown, T.C., P.A. Champ, R.C. Bishop, and D.W. McCollum, 1996. "Which Response
Format Reveals the Truth about Donations to a Public Good?" Land Economics,
72:152-66.

Flores, N.E., 2002 . "Non-Paternalistic Altruism and Welfare Economics." Journal of
Public Economics 83(2):293-305.

Johannesson, M., P.-O. Johansson, and R.M. O'Conor, 1996. "The Value of Private
Safety Versus the Value of Public Safety," Journal of Risk and Uncertainty 13:263-
275.

Johansson, P.-O. ,1992. "Altruism in Cost-Benefit Analysis." Environmental and
Resource Economics 2: 605-613.

Jones-Lee, M.W., 1991. "Altruism and the Value of Other People's Safety." Journal of
Risk and Uncertainty 4(2): 213 -219.

Jones-Lee, M.W., 1992. "Paternalistic Altruism and the Value of Statistical Life."
Economic Journal 102(410):80-90.

Messer, K.D., G.L. Poe, and W.D. Schulze. 2006. "The Value of Private Risk Versus the
Value Of Public Risk: An Experimental Analysis." In K.D. Messer, L.E. Piatt, G.L.
Poe, D. Rondeau, W.D. Schulze, and C.A. Vossler AEM Research Bulletin - 06-02
Cornell University, Final Report: March 31, 2006 Ecosystem Values and Surface
Water Protection: Basic Research on the Contingent Valuation MethodNSF GRANT
SES- 0108667.

Milgrom, P. (1993). "Is Sympathy an Economic Value? Philosophy, Economics, and the
Contingent Valuation Method." In J. A. Hausman, ed., Contingent valuation: A
Critical Assessment, Amsterdam, North-Holland, pp. 417-442.

Poe, G.L, and R.C. Bishop, 2001. "Chapter 3: Information and Contingent Values for
Groundwater Protection: A Case Study of Wisconsin." In J.C. Bergstrom, K.J. Boyle,
and G.L. Poe eds. The Economic Value of Water Quality, Cheltenham: Edward Elgar
Publishers.

Poe, G.L., H.M. van Es, T.P. VandenBerg and R.C. Bishop, 1998. "Do Households
Update Their Health Risk and Exposure Perceptions with Water Testing?" Journal of
Soil and Water Conservation 53(4):320-325.

Poe, G.L., M.P. Welsh, and P. A. Champ, 1997. "Measuring the Differences in Mean
Willingness to Pay when Dichotomous Choice Contingent Valuation Responses Are
Not Independent." Land Economics, 73(2):255-267.

7

-------
Severtson, D.J., L.C. Baumann, and R.L. Brown, 2006. "Applying a Health Behavior
Theory to Explore the Influence of Information and Experience on Arsenic Risk
Representations, Policy Beliefs and Protective Behavior." Risk Analysis 26(2): 353-
368.

Smith, V.K., and F.R. Johnson, 1988. "How do Risk Perceptions Respond to

Information? The Case of Radon." Review of Economics and Statistics 70(1): 1-9.

Viscusi, W.K., W.A. Magat, and A. Forrest, 1988. "Altruistic and Private Valuations for
Risk Reduction." Journal of Policy Analysis and Management 7(2):227-245.

Viscusi, W.K., and C.J. O'Connor, 1984. "Adaptive Responses to Chemical Labeling:
Are Workers Bayesian Decision Makers." American Economic Review 74:942-956.

-------
Questions and Discussion section
was not conducted for Session VI

-------