Estimating and Valuing Morbidity in a Policy Context: Proceedings of June 1989 AERE Workshop


United States
Environmental Protection
Agency
Office of Policy
Planning and Evaluation
Washington DC, 20460
August 1989
EPA-230-08-89-065
Estimating and Valuing
Morbidity in a Policy
Context:
Proceedings of June 1989
AERE Workshop

-------
EPA 230-08-89-065
AERE Workshop
Estimating and Valuing Morbidity
in a Policy Context
Proceedings
Workshop Sponsors
The Association of Environmental and Resource Economists
U.S. Environmental Protection Agency
National Oceanic and Atmospheric Administration
Session Chairpersons
Session I
Chair: Ann Fisher, U.S. EPA
Session II
Chairs: Richard Williams, FDA
Allen Basala, U.S. EPA
Session III
Chair: William H. Desvousges, RTI
Research Triangle Ftark, North Carolina
June 8-9, 1989

-------
The information in this document has been funded in part by the
United States Environmental Protection Agency (EPA) under
Cooperative Agreements CR-812056 and CR-815869. It has been
subjected to the Agency's peer and administrative review, and
approved for publication as an EPA document. Mention of trade
names or commercial products does not constitute endorsement or
recommendation for use.

-------
CONTENTS
Session I: Estimating the Amount of Illness and Inury Associated with
Specific Causes
The Role of Epidemiology in Developing Useful Data for Public Health Policy
Daniel A Hoffman (to be provided at workshop)
Acute Health and Variable Air Pollutants
James C Murdoch, Mark A Thayer, William N Weirick
Estimating Skin Cancer (Melanoma) Deaths from Sunlight Exposure
Hugh M Pitcher
Session II: Valuation of Changes in Illness and Injury
Pricing Environmental Health Risks Survey Assessments of Risk-Risk and Risk-Dollar
Trade-offs W, Kip Viscus, Wesley A Magat, and Joel Huber
The Social Costs of Chronic Heart and Lung Disease
Maureen L. Cropper and Alan J. Krupnick
Estimating the Value of Avoiding Morbidity and Mortality from Foodborne Illnesses
Josephine A Mauskopf and Michael T French
Utility-Adjusted Impairment Years: A Low-Cost Approach to Morbidity Valuation
Ted R. Miller, Charles Calhoun, and W Brian Arthur
Valuing Nonmarket Goods: A Household Production Approach
Mark Dickie and Shelby Gerking
Valuation of Morbidity Reduction Due to Air Pollution Abatement Direct and Indirect
Measurements
Mordechai Shechter
Risk, Self-Protection and Ex Ante Economic Value
Jason F Shogren and Thomas D Crocker
The Economics of Quarantines: An Application to Pesticide Regulation
Erik Lichtenberg, Robert C Spear, and David Zilberman

-------
THE ROLE OR EPIDEMIOLOGY IN DEVELOPING
USEFUL DATA FOR PUBLIC HEALTH POLICY
by
Daniel A Hoffman

-------
The Role of Epidemiology in Developing
Useful Data for Public Health Policy
Daniel A. Hoffman, Ph.D.
Centers for Disease Control
Atlanta, Georgia
INTRODUCTION
In the last two decades, the role of epidemiology in
providing data for public health policy makers has become
more prominent than any time in the history of this
discipline. Its greatest advantage is that it provides
direct human evidence of the health outcomes from various
environmental exposurs, unlike animal models. However,
there are many caveats that need to be attached to these
data. The objective of this paper is to review some of
the basic limitations to the epidemiologic method, both
in study design and in interpretation of the data. The
perspective that I will present is that of an
epidemiologist in a public health agency. The Centers for
Disease Control does not engage in the development of
regulations or have a large program in risk assessment.
1

-------
Our principle function is to serve the state and local
health departments by offering advice and assistance when
necessary through field investigations of potential
public health problems-in our case,
environmentally-induced disease. This involves the
identification of study hypotheses, designing the study,
developing the necessary survey instruments, collecting
health data, analyzing this data, and finally offering
our interpretation and recommendations. Basically we
engage in the classical epidemiologic method of
hypothesis generation and testing through field
investigations. Consequently, my talk today is focused on
the techniques involved in acquiring these data, and
examining the strengths and weaknesses of these data as
they relate to various interpretations of their meaning
for use in risk assessments, regulatory actions, or
public health policy decisions.
DESCRIPTION OF EPIDEMIOLOGIC METHODS
The assessment of effects on humans of various
environmental exposures relies heavily upon the results
from testing of animal models and clinical and
epidemiological studies. However, the most important
advantage that epidemiological studies have over animal
investigations is that they provide direct evidence of
the effects of toxic exposures in humans. Conversely,
2

-------
human studies are difficult to conduct properly and the
interpretation of the results from these studies makes
life difficult for both regulators and policy makers.
Part of the problem in interpretation stems from the
design of these studies which can be very complex.
Another problem is dealing with the inherent biases which
inevitably creep into the interpretation of the data, no
matter how thoroughly these have been addressed either in
the study design or analysis. This stems from the fact
that, with the execption of clinical trials,
epidemiological studies are observational by nature, not
experimental . Not only do humans vary widely in their
response to toxic agents but they vary also in their
capacity for response as well as in their exposure to
factors such as alcohol and tobacco, which may may
greatly modify the nature or severity of their responses
to toxic exposures. One example would by the relation
between radon exposure and cigarette smoking which could
either be additive, submultiplicative or mulplicative
depending upon which data is reviewed and which model is
applied to that data.
Despite these difficulties, techniques for the evaluation
of data from human studies have been developed and
refined. The epidemiological method has matured to the
point that it has withstood the criticism that it is
3

-------
incapable of establishing the etiology of disease.
Epidemiological inferences have been sustained and
corroborated by the results of toxicological and
biochemical studies, and epidemiology has proven to be a
powerful tool for the exploration of both qualitative and
quantitative cause-and-effect relationships between
environmental exposures and human disease. However, there
is still much to be done, especially at the rather low
levels of exposures that most human populations
experience, to further refine the tools of epidemiology.
I would now like to briefly discuss some of the various
study designs used in epidemiology. Next I will address
some of the sources of bias in epidemiologic data, and
conclude with a discussion on interpretations of
causality based on data derived from epidemiological
studies. Two areas of study which I will not discuss in
any detail today are the appropriateness of animals models
as they apply to risks in humans and the use of
biomarkers as indicators of risk in epidemiologic
studies. The majority of our experience at CDC has been
concerned with collection and interpretation of
epidemiologic data so the principal focus of my talk will
be on that process.
The most commonly used designs in epidemiology are: 1)
case reports; 2) ecological or correlational studies; 3)
4

-------
cross-sectional studies; 4) case-control studies; and 5)
cohort studies.
1 . CASF. REPORTS
Case reports identify one or more cases of a disease that
have been detected by clinicians, by company or union
officials, or by through active surveillance or passive
reporting such as cancer registries. The first recorded
case studies of environmental disease were Sir Percival
Pott's observations of scrotal cancer among chimney
sweeps in London. Publication of such case reports often
constitutes the first recognition that a problem of
environmnetally induced disease exists, and subsequent
epidemiological assessment proceeds from this
recognition. A more recent example includes the first
recorded cases of AIDS by clinicians at UCLA medical
center in 1978. In a case series, an inference of causal
association between causation and an environmental agent
is based on the plausibility of the following
considerations: clustering of the cases in a limited time
frame; the relative rarity of the types of diseases
observed; a history of common environmental exposure; an
the apparent strength of the association. The most common
use of case reports are hypothesis generation,
surveillance, and case registries
5

-------
Survei 11ance
The case report has historically been an important"
surveillance tool, especially for recognition of
infectious diseases. Occupational case reporting has been
useful in terms of reporting occupational injuries for
workman's compensation, but not so much for occupational
diseases due to the long latency period between exposures
and disease. A more recent use of case reports as a
surveillance tool is for the identification of senital
health events. These are cases of disease associated with
well-characterized causes whose appearance signals a
breakdown in mechanisms for disease prevention. This
method has been applied with success in the reduction of
maternal and infant mortality and has been extended to
such environmental illnesses as lead poisoning.
Case Registries
Other surveillance systems relying on case reports
include case registries, such as the CDC Dioxin Registry
or workers suspected of having been exposed to dioxin.
These exposure registries perform the task of grouping
potentially high-risk populations for future
epidemiological studies.
An advantage of case reports over most other types of
epidemiological studies is their low cost. In addition, a
6

-------
short lag time between identification of cases and
dissemination of information is more typical of case
reports. However, relying on case reports as an early
warning system is less useful when:
1)	the cases are sporadic;
2)	the relative risk is low;
3)	the outcome is a common disease or a symptom
with multiple common etiologies such as lung
cancer or heart disease;
4)	there is a long latent period between exposure
and effect; and
5)	there is a continuum of disease and health and
no clear distinction between cases and noncases is
possible, for example, premalignant dysplasia and
carcinoma in situ.
In addition, case reports can provide only a rough
estimate of disease frequency, in that they give no
information on the size of the population at risk and
thus make it impossible to calculate a disease rate.
Finally, case reports are difficult to generalize to a
population since the population from which the cases are
identified is not usually well defined.
2. CORRELATIONAL OR ECOLOGIC STUDIES
Another type of descriptive tool used by epidemiologists
is the so-called correlational or ecologic study, which

-------
uses data from entire populations to compare disease
frequencies between different groups during the same
period of time or in the same population at different
points in time.
As an example of the former, correlational studies have
suggested that various dietary components, in this case
per capita meat consumption, may be risk factors for
colon cancer. Figure 1 shows the correlation between per
capita consumption of meat and rates of colon cancer in
women from a large number of countries. As apparent from
this figure, the rates of colon cancer are lowest in
countries with the lowest per capita meat intake and vice
versa.
Figure 2 illustrates the change in disease frequency
within the same population over time. In this slide, the
difference between tge approximately 820,000 deaths from
coronary heart disease that would have been expected in
the United States if the 1968 rates had continued to
apply and the approximately 630,000 deayhs actually
observed. Such data suggest two possible explanations: 1)
that the decline in deaths from coronary heart disease
could be due to prevention due to improvements in
life-style habits and consequent risk factor reduction,
and 2) that while the rates of CHD did not decline,
persons were surviving longer due to improvements in
medical management of CHD.
8

-------
While correlational studies are useful in developing
hypotheses for study, they cannot be used to test them
because of a number of imitations inherent in their
design.
1)	Correlational studies refer to populations
rather than to individuals. Therefore, it is not
possible to link an exposure to occurrence of
disease in the same person.
2)	The distribution of other risk factor's which
may account for different rates of a disease, may
be differentially distributed among populations.
This is known as the "ecologic fallacy".
3. CROSS-SECTIONAL STUDIES
Another type of descriptive study design is the
cross-sectional survey, in which the status of an
individual with respect to the presence or absence of
both exposure and disease is assessed at the same point
in time. For example, the Health Interview Survey is a
national cross-sectional study that periodically collects
extensive information by questionnaire from a sample of
over 100,000 persons throughout the United States. These
studies often rely on personal interviews or
9

-------
questionnaires to obtain demographic information,
symptomatic, and exposure data on clinical evaluations
based on physical examinations and laboratory and
environmental sampling data to identify the
characteristics of the sample population and to
quantitate exposure to potential risk factors. An
advantage of the cross-sectional survey is the rapid
estimation of numerator values for determining frequency
or prevalence rates of both exposure and effects.
Limitations of this method include the inability to
distinguish whether the exposure preceeded the
development of disease or whether the presence of disease
affected the individual's level of exposure, since
exposure and disease are assessed at the same point in
time. Cross-sectional approaches have limited usefulness
in cancer studies because of the usual low prevalence of
cases. It is also extremely difficult to quantify
exposure in cross-sectional studies. However, for factors
that remain unaltered over time, such as sex, race or
blood group, the cross-sectional survey can provide
evidence of valid associations.
Five common pitfalls can be found in the cross-sectional
method. These are:
1) Selection bias, in that a nonrepresentative
sample of the population may be surveyed, limiting
the generalizability of the survey results;
10

-------
2)	Confounding bias, which can result for factors
related to both exposure and outcome, such as age;
3)	Inadequate sensitivity of the survey
instruments. This includes specificity, which is
the ability to detect "true" negatives, and
sensitivity or the ability to detect "true"
positives;
4)	Lack of standardization of the instruments used
for data collection, which may prohibit the
pooling of data from multiple surveys; and
5)	Inadequate validation of either exposures of
health outcomes, resulting in misclassification of
either category.
Summarizing, in general, cross-sectional studies are
useful for raising the question of the presence of an
association rather than testing a hypothesis.
The next two types of epidemiologic studies are
observational in design. These are the case-control study
and the cohort study.
In theory, it is possible to test a hypothesis using
either design strategy. In practice, however, each design
offers certain unique advantages and disadvantages. In
general, the decision to use a particular design is based
11

-------
on the features of the exposure and disease, the current
state of knowledge^nd logistic considerations such as
available time and resources.
4 . CASE-CONTROL STUDIES
In the case-control study, a case group or series of
patients who have a disease of interest and a control or
comparison group of individuals without the disease are
selected for investigation, and the proportions with the
exposure of interest in each group are compared. Lung
cancer patients, for example, can be compared to persons
without that disease for differences in exposures, such
as cigarette smoking, occupational exposures, and radon
levels in the home.The relative frequency of
distribution of the exposure in the case and control
groups is usually evaluated by computing an odds ratio
which is defined as the product of the number of exposed
cases and unexposed controls divided by the product of
the unexposed cases and exposed controls. This is also
somtimes known as the cross-product odds ratio because of
the manner in which it is calculated.
Case-control studies can be conducted relatively rapidly.
Many simultaneous exposures can be evaluated in relation
to even the rarest disease. Howver the sequence of
12

-------
exposure-health event is often difficult to assess if the
case population includes patients selected from
historical records. If the disease studied is rapidly
fatal, interviews with surrogate respondents may be
required which may result in misclassification of
exposures. The individual exposure status is often
difficult to quantify with any precision, especially in
environmental studies, and control of possible
confounders may require a complex design or analysis.
Consequently, only environmental exposures with a high
prevalence and relative strong toxic effect are
effectively studied by the case-control method.
5-	COHORT RTTTDTF.S
In a cohort or follow-up study, the study population is
divided on the basis of exposure status. For example, in
a recent study of the health effects of volatile organic
compounds in Michigan, we assembled study cohorts on the
basis of whether or not VOC'S were detected in their well
water and if they had lived for a specified period of
time in the study area. Residents who had moved away
prior to the initiation of the study were still eligible
for inclusion in either the exposed or unexposed cohorts.
Once the exposure status of the study cohorts has been
determined, which is sometimes quite complex and can
13

-------
result in misclassification of exposure status thus
biasing the study outcome towards the null hypothesis of
finding no effect, the history of disease is determined
in both the exposed and unexposed groups. The rate of
disease in the exposed group is compared to that in the
unexposed group resulting in a relative risk of disease
which could be due to the exposure being studied. This is
also called the rate ratio since it is simply the ratio
of two incidence rates. Both of these measures of
association include a factor for follow-up time known as
the person-year. This is simply defined as the interval
from the time exposure began to the date of diagnosis of
disease, death, loss-to-follow-up or, if disease-free, an
arbitrary date.
The strengths of the cohort approach include the
following:
1) the sequence of exposure and health outcome can
be studied;
) many health outcomes can be evaluated with
regard to the one exposure of interest (although
this may have become a problem in some studies as
multiple comparisons inevitably lead to at least
one "significant" finding);
3) the initial exposures can be quantified through
historical records or even more so if there is a
14

-------
biologic marker of exposure such as blood or bone
lead levels;
4)	rare exposures can be studied;
5)	collection and analysis of potential
confounding factor is possible; and
6)	absolute risks may be calculated for use in
public health prevention strategies.
Some of the drawbacks to the cohort approach are the
expense and difficult logistics of these studies, the
potential for misclassification of exposure and disease
outcome resulting in a biased estinate of risk, and the
inability to study rare disease because of the very large
populations necessary for study. This latter drawback is.
important in studying the effects of low-level
environmental exposures. Because the anticipated risk of
these exposures is low, very large numbers of exposed
persons are required for study if the outcome is to have
any decent statistical power.
PROBLEMS IN CURRENT STUDY DESIGNS
From the previous discussions, four areas of major
problems become evident: 1) the assessment of the
exposure-response sequence; 2) quantification of
exposure; 3) recognition of bias and confounding; and 4)
15

-------
quality and validity of data. Clearly, a very complex
study design may be required to yield useful results.
Measures to improve the usefulness of human studies for
risk assessment purposes include the extension of the
duration of follow-up time, assessing the time component
in exposure and disease diagnosis, focusing on
potentially high-risk populations for study, and quality
assurance of information on exposure and disease. While
most of these measures are in the area of logistics and
funding, an important exception is improvement of the
quality of the exposure data.
In the past decade, development of environmental exposure
mesures has been very rapid. Detection limits for
chemicals in environmental media have dropped by three to
four orders of magnitude, and the progress of tests for
some chemicals in biological media is almost as
impressive. The detection limits for dioxin in sera, for
example, is now measured in parts per quadrillion.
Unfortunately, little progress to date has been found to
be of practical use in epidemiologic analysis and risk
estimation. For instance, issues of background levels,
biological persistence, adaption mechanisms, absorption
kinectics, saturation of metabolic pathways, and the
impact of an individual's characteristics on the
pathogenetic process have not been addressed in most
16

-------
epidemiologic study designs, and, for the most part, have
yet to enter the area of regulatory risk assessment.
There are other practical exposure issues that need to be
addressed such as noncontinuous or fluctuating exposures,
the cause of interspecies differences, and whether or not
an observed dose-response relationhip is stable over a
wide range of dose levels. We will also see an increasing
demand to incorporate quality assurance and quality
control in epidemiologic studies with regard to matters
other than laboratory work. For example, it is of utmost
importance to make certain that the disease of concern is
following and not pre-dating exposure. Finally, there is
the issue of the quality of the diagnostic criteria for a
case or a non-case.
The quality of diagnosis becomes a very central issue
when it comes to scenarios of localized environmental
pollution, for example, at a chemical dump site, and
residents with nonverifiable and subjective complaints,
which may be real to them, sucah as headache, fatigue,
nausea, chest pain, and loss of libido. Currently, there
is an inclination among epidemiologists to ignore or
disqualify this so-called "dump-site syndrome" from
serious study. However, such an attitude is usually
followed by a deterioration of a conflict situation
between citizens and authorities. There are many
17

-------
instances where eventually epidemiologists have been
forced by heavy and relentless public and political
pressures to conduct studies of such perceived illnesses.
In doing so, they will have to derive methods to cope
with non-verifiable health outcomes, while maintaining
scientific integrity and credibility. In theory, it
should be possible to either solve the problems with
statistical tools, or by developing tests for the kinds
of complaints often described as emotional or
behavioral. CDC staff are currently developing and
applying such tools to several large studies.
Statistical methods usually fail since the situation at a
dump site is inherently associated with an abundance of
negative publicity, usually in the direction of stating
the association of voiced complaints with exposure, or
even just living near a dump site, as a fact. This
scenario often results in serious reponse biases for
persons who perceive they may be exposed. I do not forsee
that behavioral toxicology, an exciting new field of
research, can provide us in the near future with the
appropriate scientific tools to address currently
nonconfirmable complaints.
DEVELOPMENT OF MOLECULAR EPIDEMIOLOGY
A special problem, both in animal and human studies, is
18

-------
that current designs deal with observed disease, which is
a more or less advanced stage of a toxic effect. In
animal studies, most diseases are observed in moribund or
sacrificed animals. In humans, disease detection is
usually in an earlier phase by virtue of man's ability
for detailed communication. However, even common diseases
such as cancer, arthritis, hypertension, and diabetes
still pose unresolved problems in assessing the date of
onset. Estimates of this date may differ by many years,
and this would offset greater accuracy in exposurement.
The logical response to this problem is to develop
techniques to diagnose the disease in the earliest
possible stage. But the question then arises: "What is
earliest possible?" An aggressive biopsy regimen for
diseases such as cancer and kidney disease may shift the
date of diagnosis from months to years earlier. Certain
inborn metabolic disorders can now be detected
prenatally. The use of electron microscopy has brought us
closer to the early onset of renal disease.
Unfortunately, these striking improvements in early
diagnosis require invasive procedures. This is a serious
handicap to epidemiologic studies, especially those
involving environmental rather than clinical or
occupational exposures. This explains the increasing
interest of epidemiologists and risk assessors in the use
of biomarkers indicating past exposures or early stages
of tissue dysfunction, for example, DNA-adducts.
19

-------
VALADIATION OF ASSUMPTIONS FROM EPIDEMIOLOGIC STUDIES FOR
REGULATORY PURPOSES AND PUBLIC HEALTH POLICY DECISIONS
Finally, I would like to discuss the interest of
epidemiologists in the validation of a number of
assumptions used in risk assessments for regulatory
purposes or public health policy decisions. One of these
is the assumption that the presence of a toxic chemical
in the environment automatically implies exposure, and
that that body dose is proportional to environmental
concentrations. This assumption leads to the often-used,
but nevertheless incorrect practice of assuming that the
concentration of a chemical in media such as soil, air or
water is a direct measure of the amount of chemicals
absorbed in the human body. Worse, without much thought
it is often considered identical to the challange to the
organ or tissue interest when determining acceptable
exposure levels. Studies into the relation between
environmental presence, human exposure, and
organ-specific dose are increasing in number.
The findings from these studies have sometimes been
contrary to expectation. For example, at the CDC, studies
have shown that the concentration of arsenic, PCB'S,
mercury, and lead in the soil of a neighborhood is only
partly related if at all to the levels in the biologic
specimens of residents. In this light, it is important to
recognize the importance of well-conducted research with
20

-------
negative findings. Such research is critical to our
understanding of the effects of toxicants on human
biology. Moreover, such findings help concerned
scientists to inform the public of true risks and allay
undue anxiety. Indeed, despite the abundance of available
data to date, the relation between environmental
concentrations of chlorinated hydrocarbons such as DDT,
dioxin, and PCB'S, and human sera or adipose samples,
remains unclear, and the relation of these levels of body
burdens to clinical disease remains uncertain.
To date, epidemiologic studies almost never prove cause
and effect, though in a few instances, reasonable people
would accept some of them as such. For example, in
looking at the pathway of exposure and body burden, the
association of the reduction of lead used in gasoline
production and the reduction of mean blood lead levels in
the U. S. population is striking. Over a 4-year period
when the lead phasedown in gasoline was occurring, we
were conducting a study of blood lead levels in the U. S.
population using data from the Second National Health and
Nutrition Examination Survey or NHANES-2, an example of a
cross-sectional study. Two things, declining blood lead
levels and lead used in gasoline production were highly
correlated. We removed over 100 potentially confounding
variables from this association in the analysis and the
coefficient of correlation did not appreciably change.
Yet many epidemiologists stated that this did not provide
21

-------
adequate evidence of cause and effect. The only way to
unequivocally prove cause and effect in this situation
would be to conduct an experimental study where children
were placed in chambers and breathed air with different
lead levels and then measure their blood lead levels.
This experiment, of course, would be entirely unethical
and would not be supported by society. Studies conducted
in humans must use only inadvertent exposure or natural
experiment s" such as that occurred with water
fluoridation and dental carries.
Proper use of epidemiologic data can lead to important
collective public health benefits. On the other hand, to
press such data into service to respond to causal effects
for an individual's disease holds high potential for
misuse of the data.
We will continue to respond to specific incidents of
human exposure to toxic or hazardous substances. We will
also continue our efforts, through epidemiologic
techniques, to measure both the immediate and long-term
health effects and to make sound recommendations for the
attenuation of these potential risks.
Although the results of such epidemiologic investigations
may not provide the conclusive answers about health risks
from environmental exposures, which are now in such
22

-------
demand and so prevalent in the media, we have hope that
we can study and detect these associations where they
exist, so that prudent public health actions can be
taken. Thus, we see the ultimate role of epidemiology as
one of prevention, which is the most effective public
health policy to implement.
23

-------
SUGGESTED READINGS
1.	Brunekreef, B., Noy, D., and Clausing, P.:
Variability of exposure measurements in environmental
epidemiology. Am J Epidemiol 1987; 125:892-898.
2.	Lilienfeld, A. M.: Practical limitations of
epidemiologic methods. Environmental Health Persp
1983;52:3-8 .
3.	Schneiderman, M. A.: Extrapolation from incomplete
data to total or lifetime risks at low doses.
Environmental Health Persp 1981;42:33-38.
4.	Dinman, B. D. and Sussman, N. B.: Uncertainty,
risk, and the role of epidemiology in public policy
development. J Occup Med 1983;25:511-516.
5.	Glass, R. I.: New prospects for epidemiologic
investigations. Sci ence 1986;234:951-956.
6.	Stein, Z. and Hatch, M. : Biological markers in
reproductive epidemiology: prospects and precautions.
Environmental Health Persp1987;74:67-75.
7.	Einarson, T. R., Leeder, J. Steven, and Koren, G.:
A method for meta-analysis of epidemiological studies.
Pharmacoepidemiology 1988;22:813-823.
24

-------
50r—
• NZ
- 30
10
• USA
'Can
Den
Nor
#Swe
• Net
Isr
'UK
Nig
• Jam
Yug.
~Jap Chi*
PR
Fin
:R°m***Hun
• FDP
_ • Ice
DDP
•Col
J	
40
80
120
160
200
240
280
320
Per caput daily consumption — grams
Figure 1. Colon Cancer Incidence Rates/100,000 by Per-Capita
Meat Consumption (grams) for selected countries

-------
800
800
Expected
191,500
Deaths
700
700 -
Observed
(/>
= 600
600
1968 69
70
72
73
71
74
75
76
77
Year
Figure 2. Difference Between Observed and Projected Mortality
Cases from Coronary Heart Disease, United States
(If deaths in 1977 occurred at the same rates as
in 1968)

-------
DRAFT
AERE Conference
June 8, 1989
Malignant Melanoma Death Rates,
Outdoor Recreation
and
Sun Screens
Hugh M. Pitcher
Economic Studies Branch
Office of Policy Analysis
US EPA
May 18, 1989
Disclaimer: This paper reports interim results on research which
is still in process. It presents the personal opinions of the
author. It has not been reviewed by EPA and does not represent an
official EPA position.

-------
Ab s t r a c t
Previous work has shown that there is a six fold increase in
the risk of death from melanoma for white males born in the 1940's
when compared to white males born in the 1880's and 1890's. For
women the same ratio is slightly less than three. Accepting the
hypothesis that most melanoma is caused by exposure to solar
radiation, an investigation of changes in residence patterns,
occupation, and outdoor recreation is made to see if the changes
in cohort risk can be explained by changes in factors related to
exposure patterns. Household access to automobiles turns out to
be the best potential measurable factor explaining outdoor
recreation patterns. While no conclusive findings are reached,
support is developed for the hypothesis that intense exposure of
skin which has not developed natural defenses under low to moderate
exposure is the primary risk factor for melanoma. The introduction
of sun screens" is associated with reductions in this risk. Since
lifetime incidence rates for white males in the 1940's cohorts will
approach 2.5 percent with death rates of about .6 percent, melanoma
is a significant public health problem. The paper's results
suggest that a risk communication policy should be aimed at
modifying sun exposure habits to reduce intensity of exposure.
The association of the automobile with problem exposure behavior
suggests a strategy of keeping sun screen in the glove compartment.
The payoff from such a policy could be a dramatic reduction in
melanoma incidence and death.

-------
INTRODUCTION
Cutaneous malignant melanoma is the most rapidly rising cause
of cancer death for white males and the second only to lung cancer
for white females.^" In response to this, there has been much
2
investigation of potential causes. Because melanoma cannot be
induced in small laboratory animals by ultraviolet radiation alone,
the ability of laboratory research to settle etiologic issues has
been sharply limited.	Epidemiologic results have been
inconsistent, with less melanoma observed on frequently exposed
parts of the body, and death rates increasing with latitude in
4 5
Europe. ' Occupations involving outdoor exposure have been found
to be mildly protective. ® It has been difficult to develop a model
which can comfortably explain all of these results. Thus, despite
all the work on melanoma to date, there is still a clearly
understood feeling by the research community that this is a disease
g
whose etiology is not at all well understood.
The first part of the paper reviews some basic biological and
epidemiological results. The next section reports on previous
results obtained in this research project. The project has been
centered around factors affecting melanoma death rates for US
whites between 1950 and 1984. County death rates have been
agregated into Standard Metropolitan Areas(1980 definition) and
merged with census data on sociodemographic characteristics of the
1980 population, weather data for each city, and model based
3

-------
4
g
estimates of exposure. This data set has been used to investigate
the response of death rates to potential exposure, the cohort
structure of death rates, and the response of death rates to
individual components of the ultraviolet spectrum. For this paper,
the data set is used to predict cohort levels of risk, which serves
as the basis for the analysis of changes in ecologic risk factors.
The third part of the paper then precedes to examine how
factors such as outdoor recreation, outdoor work, and residence
have varied over the period for which cohort risk of death of
melanoma can be inferred from the data set. First, some measures
of how these factors have changed are developed. These are then
compared to the summary risk measures for each cohort. Out of this
there emerges a fairly clear picture of the kinds of exposure
factors that can be related to the observed change in risk. These
factors can explain the rise and stabilization of the cohort risk
factors. They cannot explain the downturn in risk seen in the
youngest cohorts.
The next section looks at available data on sun screens to see
if they are a potential cause of the downturn. It is shown that
sunscreens can be an explanation of the decline only under the
hypothesis that it is control of intense exposure of skin which has
not developed natural protection which is important if risk is to
be reduced. Usage levels are too low for them to have been a factor
if control of all exposure is necessary to reduce risk. The last
section developes some of the potential benefits of a risk

-------
5
communication strategy developed along the lines suggested by the
results of the previous sections.
BACKGROUND
This section reviews some basic biology, some of the little
that is known about how the skin develops natural protection, and
some results about how exposure changes as a function of latitude,
time of year, and time of day.
Melanoma arises in the melanocyte, the cell which produces
melanin, the compound responsible for skin color. ^ The precise
process by which the transformation to a tumor takes place is not
known. ^ The tumor is normally highly antigenic--meaning the
immune system will attack it--and one of the clinical markers for
an early lesion is a red iritated area around the lesion. 12 since
UV radiation is known to suppress some aspects of the immune sytem,
immune suppression via this route is thought to play some role in
13
the disease.	However, this role remains to be worked out in
detail. The tumor metastizes readily once it penetrates the
surface of the skin and it is the metastases which are responsible
for the mortality associated with melanoma. ^ On the other hand,
five year survival rates for melanomas removed before the dermis
15
has been invaded are about 95 percent. Thus early diagnosis and
removal are critical to effective treatment of the disease.
Incidence and death rates from melanoma have been growing very
rapidly.16 Figure 1 shows death and incidence rates for whites in
the US. In 1984 total deaths from melanoma in the US were 5377.
of these deaths 5264 were whites and 113 were non whites. Age

-------
6
adjusted death rates were 3.11 for white males, 1.65 for white
females, .37 for non-white males and .41 for non-white females.
This is a world wide pattern, indicating that melanoma is primarily
a disease of white populations. For non-whites, melanoma almost
always arises in the non-pigmented portion of the body, either
under the nails or on the soles of the feet.^ Thus pigmentation
is protective. This is true even within the white population, with
southern Europeans such as Spanish and Portuguese much less likely
18
to get melanoma than those of northern European origin.
The hypotheses that melanoma might be solar related stems from
the fact that non-melanoma skin cancers seem to be clearly sun
19
related.	Non-melanoma cancers occur most frequently on the
exposed portion of the body, and are much more frequent on those
with lots of outdoor activity--thus they clearly are a function of
2 0
lifetime exposure.	Melanoma, on the other hand does not follow
21
this pattern. Less exposed parts of the body, such as the trunk
in males, and the legs in females, are the predominate place where
melanoma is found. This clearly indicates the need for some
modification of the solar hypothesis. The second problem stems
from the results for Europe, which show that the expected decrease
of melanoma incidence and death rates with latitude does not
22
occur.	Rather rates are lower m southern Europe than m
northern Europe. This may be due to the pigmentation variations
discussed earlier. Later results in the paper on the possible
role of recreation, occupation and residence patterns may also
help explain the anomaly.

-------
Existing results also point to the role of early exposure as
• , 23,24
being critical.	Again, the results are not unambiguous, and
in some cases depend on quite small samples. Finally, due to the
lack of an animal model, the exact portion of the spectrum
O R
responsible for carcmogenisis is not clear.	The hypothesis is
that the UVB part of the spectrum is responsible, since this is the
part of the spectrum where damage to DNA occurs. Due the lack of
a widely distributed network of instrumentation capable of
individual waveband measurement, there has been no confirmation of
this by epidemiological studies. Thus the potential role of
sunscreens as a protective device has been difficult to determine
since the major chemicals are effective only in the UVB part of the
,	26
spectrum.
Exposure to the sun elicits the production of melanin and the
development of a thicker stratum corneum, the outermost layer of
27
cells on the skin. . Both of these factors reduce penetration of
UV radiation to the growing layer of cells. While it is difficult
to determine the exact extent of the protection induced by these
factors, the tanning process does increase the length of time
necessary to produce erythema(sunburn) by at least a factor of
28
three.	Black skin reduces the level of radiation reaching the
29
melanocyte by about a factor of 10.	Perhaps obviously, the
incident angle of radiation is also very important, since radiation
entering the skin at a sharp angle must travel much further before
reaching the growing layer of cells. Thus, most work activities

-------
8
expose substantially less of the skin to intense doses than do
activities like sunbathing, where the body is prone.
Ultraviolet radiation present at ground level starts at about
290nm and increases by about 5 orders of magnitude in intensity by
325 rati. From this wavelength to 400 nm, the lower end of visible
spectrum, radiation is rougly constant in intensity and varies in
the same manner as visible light. The large variation in intensity
between 290nm and 325nm is due to absortion by ozone in the
stratosphere. Figure 2 shows variation in DNA weighted radiation
by latitude for a clear day in the peak month of the year and for
total radiation during the year. Note that there is little
variation in peak values between the equator and 30 degrees,
latitude. Figure 3 shows DNA and Erythema weighted radiation
measures during the year for Washington DC. Note that DNA weighted
radiation varies more than does erythema weighted radiation.
Figure 4 shows variation during speak day in July. Note again
that DNA radiation varies more during the day than does erythema.
The relevance of these differences in behavior will become clear
later in the paper.
PREVIOUS RESULTS FROM THIS PROJECT
The work already done on this data set bears on a number of
the open questions discussed above.	First, it shows that
variations in intensity of ultraviolet radiation are associated
with higher death rates.JU A one percent increase in peak(clear
summer day) DNA weighted radiation yields a .85 percent increase
in the death rate for males and a .58 percent increase in the death

-------
9
rate for females. Controls for socioeconomic variables do not
affect the results while including the effect Qf ethnic origin
reduces the responsiveness of death rates by about 20%.
The second area of work with this data set suggests that it is
exposure in the UVB part of the spectrum which is responsible for
• • 31
the carcmogenisis.	The exposure measures used in previous
epidemiology on melanoma have been simple latitude(which is a non-
linear function of exposure as figure 2 illustrates), hours of
sunlight, or an integrating meter(known as the Robertson Berger
meter) which gives a single measure of UV radiation. 32 The
exposure measure used in this study is developed from a model which
incorporates satilite measures of ozone into a radiative transfer
model to predict ground level UV radiation. These predictions can
either be in the form of wavelength weighted measures where the
weights are the inverse of the biological effectiveness of
different wavelengths, or as individual waveband energies. jn this
particular work, individual waveband energies from 295-299 through
330-334 for a clear day in June were used as exposure measures.
Table 1 presents the estimates for different wavebands. Deaths
were modeled as a poisson process and estimation was done using
iteratively reweighted least squares to get maximum liklyhood
estimates. The results show a positive relationship between
radiation below 320, with a negative and significant relationship
above 330 for males. For females that pattern is similar, but the
results are not significant above 330nm.	Because of high
correlations between different wavebands, it is not possible to

-------
10
introduce more than two wavebands simultaneously into the
equations. The second part of Table 1 shows the results using 295-
299 and 300-304 as the short waveband with various wavebands used
as the long waveband. This indicates the upper range for positive
response to radiation lies at about 315 nm or at the upper end of
the range where radiation damages DNA.
These results rely heavily on variations in specific parts of
the spectrum. Since the model has only been tested with aggregate
mesures produced by the Robertson-Berger meter, more work is needed
to baseline the model. However, the overall pattern of variation
is dependent only on variations in measured ozone and very basic
radiative principles. Thus while there may be measurement error,
it is unlikely to be systematic in nature, and thus, in this simple
model, the expected result would be to bias the estimated
coefficients toward zero.
The third set of analyses done with this data look at cohort
33
experiences.	As seen m Figure 5, there is a very systematic
structure to a plot of the log of the national cohort death rates
against age. Cohorts are defined as those who are 0 to 4 years of
age for a five calender year interval. This results in a median
birth year equal to the initial calender year of the period. This
definition was required because death data were only available in
five year age groups in the source data set. The labels on the
plot refer to the median birth year for each cohort. The parallel
slopes of the cohort death rate curves above the age of thirty
suggest that it is early exposure which is critical to the

-------
11
potential risk. Statistical analysis	confirms that the curves
above the age of 30 have equal slope.	For men this slope is 7
percent per year and for women it is	five percent per year. As
Figure 5 shows, there is no slope to	the death rate experience
before age 10. Clinical experience indicates these deaths are due
. 34
to congenital nevi.	Therefore, it seems reasonable to assume
that the rate for 0 to 9 year olds is constant, and all the
variation in cohort risk is due to variation in how the death rate
changes between age 10 and age 30. Table 2 presents estimates for
a model which includes DNA weighted exposure, individual cohort
estimates for 7 < Age < 32, and a common age effect above age 32.
These results suggest that variations in some aspect of exposure
across time for the age group less than 30 are at the root of the
varying coefficients for the cohort specific age variable.
Using a much simpler procedure, estimates at age specific
rates at age 32 can be made for chorts born between 1865 and 1970.
For the 35 years of data available, average ratios for each five
year differential are computed. These averages are used to
extrapolate from the nearest available death rate to the age 32
death rate. Table 3 gives the results of these forecasts for each
birth cohort. White males show marginally greater than a ten to
one variation while white females show about a five to one
variation. Interestingly, there is a predicted downturn in the age
specific rates for cohorts born after 1950. As seen in figures 5
and 6, these reductions are already seen in these cohorts at
35
younger ages.	Next to the differential rates for blacks and

-------
12
whites, this is the largest variation seen in experience with
melanoma. Thus any explanation of melanoma aetiology must deal
with this experience. The next section of the paper looks at some
potential explanations for these large cohort effects.
COHORT VARIATIONS IN DEATH RATES
Given the small number of degrees of freedom across cohorts
and the very limited quantity and quality of data on recreation in
particular, the analysis in this section is more qualitative in
nature that the analysis in the previous sections. The essential
question to be addressed is what changes have occurred in exposure
habits and opportunities between 1880-84, when the oldest cohort
in the study was 15-19 years old and the 1980-84 period, when the
1965 cohort was 15 to 19 years of age. There are a number of
hypothesis which could be suggested for the variation across this
period of time. Here only solar related hypotheses are considered
since there is little indication in the literature of any other
cofactor besides genetic predisposition as a potential cause of
melanoma. (This is not to say one might not exist--but only that a
creditable one has not been found so far) .
The first potential hypothesis is that changes in place of
residence during the critical exposure years might have changed so
that average intensity of exposure is higher. However, as Table
4 shows, DNA relevant radiation weighted by state populations
between 15 and 24 for every five years between 1890 and 1985
increases by only 2.8 percent in intensity(average exposure in 1980
is 3.25) . Since this would amount to only an 2.5 percent change

-------
13
in risk at age 32 for males and a 1.65 percent change in risk at
age 32 for females, this does not explain the very large changes
in lifetime risk seen in Table 4.
Likewise, occupational exposure is not the explanation. The
two major occupation groups with extensive sun exposure are farming
and construction. As Table 5 shows, these have fallen sharply in
relative size, and even in absolute size during the 1880-1985
period. Also occupation is less apt to be a risk factor for those
under the age of 20 since labor force participation rates are
relatively low and have been quite static in the 50 to 60 percent
range for white males between 14 and 19 and between 20 and 30
percent for white females in the same age group.
A third potential hypothesis centers around outdoor
recreational exposure. This can at best be a partial explanation
of the changes in melanoma risk. Around 1900, forty percent of the
population lived on farms and participation in outdoor recreation
was about 4 hours per capita per year(see Table 7), while the
lifetime risk of death from melanoma was only about 1 in 1000 for
both males and females. In 1960, eight percent of the population
lived on farms, per capita participation in outdoor recreation had
risen to almost 120 hours, and the lifetime risk of death from
melanoma had reached 6 per thousand for males and 2.7 per thousand
for females. From 1960 to 1985, farm population fell to about 2.5
percent of total population, per capital outdoor recreation hours
by 2 and one/half fold, but the risk of melanoma has decreased.
While the results between 1900 and 1960 are suggestive of a role

-------
14
for outdoor recreation, the 1960 to 1985 results suggest(as always
with melanoma it seems) that if there is a role for recreational
exposure it is not a simple one.
One can make sense of the role of recreation if what matters
is not the extent of participation, but simply participating at
all. Under this hypothesis one would expect to see a stabilization
in participation rates in sun intensive activities beginning in the
sixties. Unfortunately, data on a comparative basis does not
exist. What can be examined is a number of proxy variables for
participation. One proxy for recreation behavior is the percent
of the labor force not at work due to vacation (see Table 8) . While
comparable data is not available before World War 2, data given in
Clawson and Knetch indicate weeks of vacation per worker rose from
.37 in 1929 to 1.09 in 1959, suggesting percent participation rose
during the 1929-1946 interval also.
A second indicator of percent participation comes from noting
that over ninety percent of outdoor recreation involves automotive
^ • XX
transportation.	Assuming that most recreational activity
involving automotive transportation is family oriented, the
critical variable controlling access to outdoor recreation is
household motor vehicle ownership. Table 9 gives this data for
the post war period. Table 10 extends this back before WW II as
mean vehicles per household. Comparison with Table 9 indicates
that mean vehicles oer household is roughly double the percentage
of households owning at least one car. What is clearly interesting

-------
15
about this variable is the apparent saturation on a per household
basis which occurs in the early 1960's.
These two variables suggest that a case can be made that the
breadth of participation in outdoor activity stablized in the
1960's. Several surveys of participation in outdoor activities
were done between 1962 and 1982-83. While summary results from
these surveys are not in a format that makes comparison across time
possible, the latest survey(1982-83) does indicate that all but 11
percent of the general population participate in some form of
outdoor recreation. For those between 12 and 24, all but 3 percent
of the population participate. The next step in this process is
to get the source level documents and see if a more coherent
picture can be developed.
The other aspect of changes in recreational exposure which is
important to understand is that activities associated with intense
exposure have increased over time.	Swimming, especially
sunbathing, is typically associated with more intense exposure than
hiking or bicycling. Further, in the northern part of the country,
the outdoor swimming season does not begin until the intensity
level is within ten percent of the peak level it will achieve
during the year. One illustration of the increase in intensity is
that the number of muncipal swimming facilities per capita
increased more than seven fold between 1910 and 1965.
The issue of when expsoure begins, alluded to in the previous
paragraph, is also important. There have been very significant
changes in the time at which exosure begins for the critical age

-------
16
groups. In the 1880's, they typical student attended school for
only 80 days a year, and stoped school after the eighth grade.
Today, the typical student attends school twice as long and in
excess of 95 percent of the 5-17 year old population is in school.
Given the shortness of the typical school year before 1900, we can
suppose that the real pattern for farm children, especially those
in the early teens, was to be outside helping with farm work
beginning in the spring and continuing through the fall. This is
not a pattern of limited sun exposure. It is a pattern which leads
to the development of a tan prior to the period of peak intensity.
As Table lb shows, if the participation in farm work begins in
March or April, the exposures levels are much lower than those
found beginning in late May or June, the typical time at which
school closes in the modern era and outdoor recreation starts.
Thus we have a potential hypothesis explaining the increase in
melanoma as a function increase in the effective intensity at the
time when sun exposure begins for the season. Under this
hypothesis, the stabilization of rates occurring in the 1930's to
1950 for females and in the 1950's for males is explained by
stabilization in the percent of the population getting intense
exposure. The large increase in recreation behavior since the
sixties is one of more extensive participation by each individual,
rather than a broader participation. If this is in fact the case,
and more work is needed on this score, then a consistent pattern
can be found which explains the growth of melanoma by an increase
in the effective intensity of radiation brought about by changes

-------
17
in education and work patterns which delay the onset of outdoor
exposure until the period of peak insolation, and the spread of
activities such as swimming, which expose much of the body to
sunlight, especially parts of the body which rarely receive any
prior exposure.
This would explain the markedly lower rates seen for the head
neck and hands for melanoma, since these parts of the body are
exposed year round and thus have always developed some level of
natural protection. It does not however, explain the decrease in
melanoma rates seen in cohorts born since the 1950's. One possible
explanation, consistent with the solar hypothesis, is that sun
screens have played a role. This is discussed in the next section.
SUNSCREENS
Under the intensity hypothesis, to be effective in reducing
risk, sunscreens do not have to be used all the time, but only
during the period of initial exposure. The question is whether the
total use Of Sunscreens, given in Table 12, is sufficiently high
to have possible been effective in reducing risk. To provide
adequate protection at the rated level, about one ounce of product
X X
is required. Typical applications to	the entire body seem to be
at about half this rate. xx This rate	is still enough to produce
a very significant reduction in risk.	Thus the actual number of
applications available is about twice the number of ounces sold.
This yields about three applications	per individual, which is
probably a minimum level of protection	for one day on the beach,
but not enough to get all potentially	exposed individual through

-------
18
the period of developing a tan without getting an intense dose.
Thus it is unlikely that sun screens are the potential explanation
for the declining risk seen in the younger age cohorts. Only if
all sunscreen use were concentrated in the younger cohorts would
this be possible. In fact, anecdotal evidence suggests it is the
younger cohorts which are least apt to use sunscreens. Thus some
other explanation for the decline must be sought. This does not
mean that a policy of increased use of sun screens would be
ineffective. The potential of such a policy is discussed in the
next section.
FOUNDATION OF A RISK ASSESSMENT POLICY
The intensity hypothesis suggests that a policy is possible
which might be very effective in reducing melanoma. The primary
goal of the policy would be to limit exposure very carefully during
the period before the skin has a chance to develop its natural
defenses of thickenign of the stratum corneum and tanning. Since
these processes both take time, this would imply either the careful
use of sunscreens or a significant limitation on activity during
a vacation taken by somebody who starts exposure when natural
intensity levels are high or travels to a sunny area during the
winter.
How much might such a prevention program be worth. No
strategy could probably return us to the results of the 1900 era.
There is simply too much intense radiation present in modern
recreational activities and use of sun screens is unlikely to be
universal. However, it might be reasonable to reduce risk by 60

-------
19
percent. The results of a policy with this level of effectiveness
are illustrated in Table 12. Since we have a fairly detailed sense
of the death rate pattern for melanoma, the table looks at years
of life saved rather than reductions in mortality. The figures are
for a cohort group of 100,000. The total reduction in melanoma
mortality, in a given year, under steady state cohort behavior,
would be 368 lives per 100, 000 males, and 162 lives per hundred
thousand females. At currently typical white birth cohort sizes
of about 1.75 million each for males and females, this yields a
total reduction of better than 9000 melanoma related deaths. Total
associated incidence would be about four times these levels, giving
reduced incidence of about 36,000 cases. it should be emphasized
that these are long run numbers and do not take account of whatever
is currently acting to reduce death rates. They do suggest that
a policy to moderate sun exposure habits has a very high potential
public health payoff.

-------
Skin Melanoma Trends Among Whites in the United States
20-1
CL
O
CL
o 10
O
o
*
o
o
<
ai
'S
on
<
Q
z
<
(/)
(/)
ZD
o
CD
V)
ZD
—>
Q
<
I
LJ
O
<
1-
tncldence
Mortality
Legend
MALES
FEMALES
0.H i—"—f
1950
-i—|—«—>"
1955
-i	1	1	1	r-
1960
-I	1	1	1	I	r
1965
YEAR
—i—i—i—i—i—|—i—i—i—i—i—i—i—i—" i
1970	1975	1980	1985

-------
Skin Melanoma Mortality by Birth Cohort
Among WHITE MALES
100
10:
Q.
O
Q.
O
O
O
m
O
0
1
tKL
1:
0.1:
0.01:
1895—99
1925-29
1945-49
° 00'' £	&*'
n -\K <\°> rr
AGE GROUP

-------
Skin Melanoma Mortality by Birth Cohort
Among WHITE FEMALES
lOOq
10:
a.
o
a.
o
o
o
*
o
o
f
OH
h
0.1
0.01:
\ /
1925-29
1895-89
0 001 /<> Jb1 \K' \9' oN1 o?' ^	*'
AGE GROUP

-------
20
Table la
Ultraviolet Radiation Variation by Latitude
Latitude Clear Day	Annual
DNA Weighted	DNA Weighted
0 5.45	1162.5
10 5.39	1181.6
20 4.59	989.1
30 4.14	658.0
40 3.31	370.7
50 2.33	204.0
60 1.67	124.3
Table lb
Clear Day UV Radiation by Month
Month	DNA Weighted	Erythema Weighted
Jan 15
.25
31. 0
Feb 15
. 50
57 . 9
Mar 15
1.01
108 . 7
Apr 15
1. 85
186.8
May 15
2 .51
246.5
Jun 15
3 . 14
299. 6
Jul 15
3 .40
318 . 3
Aug 15
2 . 91
274 . 6
Sep 15
2 . 03
196.2
Oct 15
1.09
110 . 9
Nov 15
.46
51. 2
Dec 15
.24
28 . 9
Source: Model based estimates using satelite data on ozone.
Units are not comparable.

-------
21
Table lc
UV Radiation on July 15 (Clear Day)
Time of Day	DNA Weighted	Erythema Weighted
6-6:30 am
. 0055
.78
6:30-7 am
. 0129
1. 66
7-7:30 am
. 0259
3 .06
7:30-8 am
. 0456
5.03
8-8:30 am
. 0723
7 . 55
8:30-9 am
. 1050
10 . 53
9-9:30 am
. 1419
13 .77
9:30-10 am
. 1800
17 . 05
10-10:30 am
.2162
20 . 10
10:30-11 am
.2469
22 . 66
11-11:30 am
.2693
24.51
11:30-12 am
.2811
25.48
12-12:30 pm
.2811
25.48
12:30-1 pm
.2693
24.51
1-1:30 pm
.2469
22 . 66
1:30-2 pm
.2162
20 . 10
2-2:30 pm
. 1800
17 . 05
2:30-3 pm
. 1419
13 .77
3-3:30 pm
. 1050
10 . 53
3:30-4 pm
. 0723
7 . 55
4-4:30 pm
. 0456
5 . 03
4:30-5 pm
. 0259
3 .06
5-5:30 pm
. 0129
1. 66
5:30-6 pm
. 0055
.78
6-6:30 pm
. 0020
.30
6:30-7 pm
. 0006
.09
All radiation values from radiative transfer model
incorporating satelite measurements of ozone.

-------
22
Table 2a
Wavelength Specific Estimates
of Exposure Effects*
Mai e s
Wavelength
Coefficient
Standi
2 95-2 9 9nm
.142

. 0104
3 0 0-304 rati
.148

. 0108
3 0 5-309nm
.153

. 0115
310-314 rati
. 145

. 0125
315-319nm
.112

. 0132
32 0-324 ran
. 055

. 0139
32 5-32 9nm
. 0014

. 0143
3 3 0-334 nm
- . 028

. 0143
3 3 5-339nm
- . 044

. 0136
3 5 5-359nm
- .064

. 0138


Females

2 95-2 9 9nm
. 0817

. 0123
3 0 0-304 nm
. 0871

. 0128
3 0 5-309nm
.0923

. 0135
310-314 nm
.0912

. 0144
315-319nm
. 0765

. 0152
32 0-324 nm
. 0458

. 0157
32 5-32 9nm
. 0159

. 0158
3 3 0-334 nm
- . 00090

. 0158
3 3 5-33 9nm
- .0106

. 0149
3 5 5-359nm
- . 0204

. 0152

-------
23
2 95-2 9 9nm
315-319nm
2	95-2 9 9nm
32 0-324 rati
3	0 0-304 rati
32 0-324 ran
Table 2b
Multiple
Mai e
. 151
( .0182)
-. 0255
( .0197)
. 150
( .0114)
- .0298
( .0152)
.166
( .0124)
-.0472
( .0158)
Waveband
Female
. 0709
( .0219)
. 0197
( .0225)
. 0804
( .0136)
. 0012
( .0176)
.0908
( .0147)
- . 0091
( .0182)
* All units have been converted to standard deviation form so that
coefficients can be directly compared.

-------
24
Table 3

Age and Exposure Model
Coefficients
Vari able
Mai e

Female

Coef St.Dev.
Coef.
St. Dev.
Constant
-3.99 .118
-3 . 90
.117
DNA expos.
.263 .0125
. 188
. 0137
Age 65
.0730 .0430
.074
. 0405
Age 6 0
.146 .0153
. 126
. 0162
Age 5 5
.154 .00859
.143
.00874
Age50
.155 .00583
.142
.00597
Age4 5
.150 .00523
. 140
.00528
Age4 0
.149 .00493
. 138
.00496
Age35
.143 .00480
. 139
. 00476
Age30
.143 .00469
. 139
.00465
Age2 5
.139 .00463
.132
.00460
Age2 0
.131 .00462
. 126
.00458
Age 15
.125 .00463
. 120
.00460
Age 10
.117 .00467
. 117
. 00463
Age 0 5
.110 .00469
. Ill
.00466
AgeOO
.102 .00473
. 107
.00471
Age95
.0962 .00478
. 105
. 00475
Age 9 0
.0921 .00480
. 100
.00478
Age8 5
.0837 .00483
.092
8 .00482
Age 8 0
.0806 .00487
.0920 .00486
Age>32
.0653 .000817
.0538 .000882
Sum of Squares



Regression
4273 . 7

2655.3
Error
200 . 7

153 . 0
Total
4474 . 4

2808 . 4
About Mean
4199.7

2577 . 1
The agexx variables denote a variable
which for cohort xx
0	if Age < 7
age - 7 if 7 < Age < 32
25	if Age > 32
and 0 for all other cohorts. Cohorts are denoted by the last two
years of their median birth date and run from 1965 back to 1880.
The Age>32 variable is 0 if Age < 32 and Age - 32 otherwise.

-------
25
Table 4
Predicted Rates at Age 32
Median Birth Year White Males	White Females
1865	.146 .238
1870	.194 .234
1875	.222 .311
1880	.257 .330
1885	.309 .313
1890	.398 .424
1895	.419 .441
1900	.438 .460
1905	.553 .494
1910	.643 .593
1915	.907 .693
1920*	.986 .857
1925*	1.224 .969
1930*	1.555 1.109
1935*	1.482 1.116
1940*	1.509 .998
1945*	1.690 1.051
1950*	1.866 1.155
1955	1.690 .998
I960	1.330 .954
1965	.800 .807
1970	.795 .395
* denotes observed value for that cohort. Other values are
predicted from the nearest observed value and the average ratio
between age 32 and the observed value at that age.

-------
26
Table 5
Population(Age	15-24) Weighted Measures of Exposure
Year	DNA Exposure
1890	3.185
1900	3.201
1910	3.194
1920	3.205
1930	3.212
1940	3.212
1950	3.225
1960	3.239
1970	3.238
1980	3.263
1985	3.276
* Data on population from US Historical Statistics and various
issues of US Statistical Abstract. UV measures are mean of SMSA
spectific measures within each state.

-------
27
Table 6
Farming and Construction Employment
Year
Total
Farming
Construction
Perc
1880
17390
8920
900
56.5
1890
23320
9690
1510
48 . 0
1900
29070
11680
1665
45 . 9
1910
37480
11770
1949
36.6
1920
41610
10790
1233
28 . 9
1930
48830
10560
1988
25 . 7
1940
56290
9575
1876
20 . 3
1950
63377
9926
2364
19.4
1960
71489
7057
2926
14 . 0
1970
84889
4596
3588
9.6
1980
108544
3705
4346
7 . 4
1985
117167
2941
4673
6.5
* Data from 1950 are not strictly comparable to earlier data.
Data from 1880 to 1940 are from US Historical Statistics, US
Dept of Commerce, Washington DC 1975 (Series D167, D170 and
D17 3) . Data from 1950 to 1985 are from Economic Report of the
President, Council of Economic Advisors, Washington, DC,
1989(Tables B32, B43, and B98) .

-------
28
Table 7
Per Capita Hours of Recreational Activity
Year
Hours
1900
1910
1920
1930
1940
1950
1960
1970
4
7
20
43
59
80
1985
1980
116
211
272
304
*Data through 1960 are adapted from Clawson and Knetch, The
Economics of Outdoor Recreation. From 1970 to 1985, they are
extended by computing an index based on visits to National
Parks, National Forests, State Parks, Personal Consumption
Expenditures for Gardening, and Travel to the Carribean and
South America.

-------
29
Table 8
Percent Participating in Vacation
Not at Work by Reason of Vacation
(annual average, 1000's)
Year
On Vacation
Percent
1985
3338
34 . 2
1980
3320
36.7
1975
2815
35.4
1970
2341
33 . 1
1965


1960
1576
26.5
1955
1268
22 . 7
1950
1137
21. 5
1946
662
13.8
* Percent assumes everybody counted during the year as being not
at work due to vacation is distinct and multiplys not at work by
12 to get an estimate of total fraction of the work force which
takes a vacation.

-------
30
Table 9
Household Ownership of Motor Vehicles
Year Total Percent	Percent
One Car	Two or More
1948	54
1950 59 52	7
1955 70 60	10
1960 77 62	15
1965 79 55	24
1970 82 54	28
1977 84 47.5	36.5
Source: US Historical Statistics 1948-1970, Motor Vehicle
Facts and Figures, 1978.

-------
31
Table 10
Access to Motor Vehicles
Year
1890
1895
1900
1905
1910
1915
1920
1925
1930
1935
1940
1945
1950
1955
1960
1965
1970
1975
1980
1985
Cars
(1,000's!
0
0
8
77
458
2332
8132
17440
22973
22495
27372
25695
40191
51961
61420
74909
88775
106077
120866
129329
Total
(1,000's!
0
0
8
79
469
2491
9239
19941
26532
26230
32035
30638
48567
61949
72887
89090
106808
130919
153358
167342
Per
Capita
0
0
0.00011
0 . 00094
0.0051
0 . 025
0 . 087
0 . 17
0.22
0.21
0.24
0.22
0 . 32
0 . 37
0040
0.46
0.52
0 . 61
0 . 67
0 .70
House-
holds
(1,000's!
12690
14341
15992
17939
20183
22501
24467
27540
29997
31892
35153
37503
43554
47874
52799
57251
63401
71120
80776
86789
Per
HH
0.0
0.0
0.0005
0.0044
0 . 023
0 . 11
0 .38
0.72
0 .88
0.82
0 . 91
0.82
1.12
1
1.
1,
1.
1.
1
1.
29
38
56
68
84
90
93

-------
32
Table 11
Sales of Sun Protection Products
(Real 1988 $, 1,000,000's)
Year	Total	SPF > 8
Sales	Ounces	Sales Ounces
1960	85.6	57.1
1965	115.8	77.2
1970	155.3	103.6
1975	193.9	129.3
1980	238.9	159.3	80.0	53.3
1984	249.0	189.5	103.3	68.9
Source: I am indebted to Jim Murdoch and Mark Thayer for
providing the original data.

-------
33
Table 12
Potential Mortality Impacts
of Reducing Acute Exposure
Males-1940 Birth Cohort
Age	Baseline Reduced Difference
Mortality Mortality
0-4
.28
.07
.21
5-9
.26
.07
.19
10-14
.51
.20
.31
15-19
1.04
.42
.62
20-24
2 . 05
.82
1.23
25-29
4 .06
1. 63
2 .44
30-34
8 . 00
3.20
4 .80
35-39
11.27
4 .51
6.76
40-44
15 .77
6.31
9.46
45-49
22.39
8 . 96
13 . 43
50-54
30 . 93
12 . 37
18 .56
55-59
42 . 02
16.81
25.21
60-64
55 . 68
22 .27
33.41
65-69
71.09
28 .44
42 . 65
70-74
85 . 72
34 .29
51043
75-79
95 . 00
38 . 00
57 . 00
80-84
93 . 35
37 . 34
56.01
85 +
75 . 98
30 .39
45.59
Expected Total
Li fetime
Add.

Yrs .
70 . 7
14 .85
65 . 8
12 . 50
60 . 9
18 .88
56.1
35 .01
51. 4
63 .22
46.8
114.00
42 . 2
202.56
37 . 5
253.58
32 . 9
311.30
28 . 4
381.53
24 . 2
449.10
20 . 2
509.28
16. 6
554.57
13.3
567.30
10 . 5
540.04
8 . 0
456.00
6.0
336.06
4 . 5
205 . 15
Total
615
247
5025

-------
34
Table 12 Continued
Females-1940 Birth Cohort
Age
Baseline
Reduced
Di f ference
Expected
Tota^

Mortality
Mortality

Li fetime
Add.





Yrs .
0-4
.26
.10
.16
77 . 4
12 . 07
5-9
.24
.10
.14
72 . 5
10 .44
10-14
.45
.18
.27
67 . 6
18 .25
15-19
.85
.34
.51
62 . 7
31 . 98
20-24
1.57
.63
.94
57 . 8
54 .44
25-29
2 . 92
1. 17
1.75
53 . 0
92 . 86
30-34
5 .40
2 .16
3 .24
48 . 1
155 .84
35-39
6. 95
2 .78
4 . 17
43 . 3
180.56
40-44
8 .89
3 .56
5 . 33
38 . 5
205.36
45-49
11.51
4 . 60
6. 91
33 . 9
234.11
50-54
14 . 63
5 . 85
8 .78
29.4
258.07
55-59
18 . 43
7 . 37
11.06
25 . 0
276.45
60-64
22 . 95
9. 18
13 .77
21. 0
289.17
65-69
28 . 04
11.22
15 . 12
17 . 2
289.37
-J
0
1
-J
33 . 30
13 . 32
19. 98
13 . 6
271.73
75-79
37 .81
15 . 12
22 . 69
10 . 5
238 .20
80-84
39.81
15 . 92
23.89
7 . 7
183.92
85 +
36.05
14 . 42
21.63
5 . 6
121.13
Total
270
108

2924

-------
35
FOOTNOTES
Forthcoming at the Conference

-------
36
BIBLIOGRAPHY
Two basic citations--detailed bibliography available	at the conference.
1.	Longstreth, JD ed., Ultraviolet Radiation	and Melanoma, US
Environmental Protection Agency, EPA 400/1-87/001D,	December, 1987.
2.	Gallagher, RP ed., Epidemiology of Malignant	Melanoma, Springer
Verlag, New York, 1986.

-------
Acute Health and Variable Air Pollutants
By
James C. Murdoch
Auburn University at Montgomery
Mark A. Thayer
San Diego State University
William N. Weirick
Northeast Louisiana University

-------
2
I . Introduction
Previous epidemiological studies [4,5,6,8] have established
an empirical relationship between measures of urban air pollution
(the dose) and human illness (the response) . The results from
these studies are interesting to policy analysts because they
can, in principle, be used to predict the impacts of proposed air
pollution control policies on urban populations. These aggregate
dose-response predictions are credible to the extent that they
seem to confirm the association between air pollutants and
illness that has been found in clinical studies [2,7].
Nevertheless, the sensitivity or robustness of the predictions
from the estimated dose-response functions to alternative
specifications, datasets, and estimation strategies remains an
important issue [2] .
This paper addresses two methodological issues in estimating
air pollution dose-response functions. Both concern accuracy in
measuring the air pollution dose.
The first involves the intracity variation in the pollution
data and the location of respondents. Previous studies have not
assigned respondents different location codes over the day. Yet
many people, especially those working, get a different dose
during the day, when compared to the evening when they are at
home. The dataset analyzed here facilitates a work and home
assignment of air pollution to individuals within a city. This
allows us to compare the responses of individuals to air
pollution at their workplace with their responses to air

-------
3
pollution at home. By modeling more of the variation in the
pollution data, we measure the "real-world" dose; hopefully,
improving the accuracy of the estimates of the influence of air
pollution on human health.
The second issue concerns the intralocation variation in air
pollution. Since air pollution in an urban area can vary from
hour to hour, we hypothesize that the dose is more appropriately
modeled as variable over a day. Even when an individual does not
change locations, he or she will experience different doses as
the air pollution varies from hour to hour. Therefore, the air
pollution dose depends on where and when a person is exposed.
The air pollution doses used in previous studies have been based
on a periodic (either one year or two weeks) average. By
averaging the pollution data, the intraday variation in the data,
which may influence human health, has been ignored. This can
cause specification error bias, owing to left out and incorrectly
measured variables in the dose-response function.
The remainder of this paper is organized into four sections.
In Section II, we present a brief review of the relevant
literature on estimating the relationship between air pollution
and human morbidity. The empirical models, data, and basic
estimation methods are described in Section III. The results are
presented in Section IV, while the last section contains
concluding remarks.
II. Air Pollution and Morbidity: Previous Studies
The design of this paper is most closely related to the

-------
4
studies by Ostro [4,5], Hausman et al. [3], and Portney and
Mullahy [6] . These authors use the Health Interview Survey
(HIS), an annual health survey of people in various locations
throughout the U.S., to study the empirical relationship between
air pollution and human morbidity. Morbidity is measured by a
variable that reflects the changes in the normal activities,
owing to health impairments, of the survey respondents during the
2-week recall period of the HIS. Several pollutants are
analyzed, including measures of the atmospheric concentrations of
total suspended particulate (TSP), ozone, fine particulate, and
sulfates. In addition to including several socio-economic and
weather measures in their models; these authors examine numerous
subsamples based on sex, working status, and smoking status,
attempting to hold constant as many confounding influences as
possible.
In Ostro [4], the variation in the air pollution data comes
from the pooling of respondents from different cities. Doses
were measured by the annual average of total suspended
particulate (TSP) and sulfates (S04) and ignore the intrayear
and intraday variation in the pollution data. The morbidity
measures reflect the number of "work loss days" (WLD) and the
number of "restricted activity days" (RAD) that survey
respondents reported during the two-week recall period. The TSP
term is significant and has the expected relationship to RAD and
WLD. The S04 term is not significant, which may not be surprising
since S04 is more localized than TSP.

-------
5
In his follow-up piece, Ostro [5] uses a Poisson distribution
to model the relationship between the number of RADs and the
contemporaneous (with the survey) two-week average of fine
particulate. Fine particulate are estimated from airport
visibility and TSP data. The two-week average of fine
particulate is significant over several different samples and
years.
Hausman et al [3] concentrate on WLD and specifically
control for intrayear variation in pollution. Additionally, they
estimate models with alternative lags of the two-week (in
contrast to the annual) average of TSP, although no formal tests
to choose among the specifications are presented. Like the Ostro
studies, intracity variation in the pollution exposure (within a
period) is ignored and the S04 measure is not significant.
Hausman et al provide some empirical support for the Poisson
specification; i.e., the pollution coefficients were robust when
the Poisson assumption that the variance equal the mean was
relaxed and when a fixed effects model was estimated.
Portney and Mullahy [6] estimate a Poisson model with the
number of respiratory related RAD as the dependent variable and
various measures of ozone and sulfates for the exposure
variables. Portney and Mullahy's ozone exposure measures are
probably better suited to test acute health effects because they
average over the daily maximum of ozone during the two-week
recall period. Moreover, by matching respondents to the
pollution monitoring stations closest to their census tract (and

-------
6
within a 10 and a 20 mile radius of their census tract) , Portney
and Mullahy analyze some of the within city variation in the
pollution data. They find that the estimates for the ozone
coefficients do not vary greatly among these different assignment
strategies, meaning that the intracity variation is not
empirically important in their data. As in the aforementioned
studies, sulfates do not perform in an _a priori expected manner.1
III. Methodology and Data
The methodology and data used in this study were constructed
in order to examine the robustness of acute health predictions.
In particular, we propose to compare the predicted health
responses from a "traditional" specification to specifications
where the pollution measures and assignments more accurately
reflect real world exposures. This comparison exercise will
provide information for policy makers as well as future research
efforts.
Define the following notation:
H. = the health response of individual i in time period t.
X. = a vector of individual specific covariates.
W = the weather in time t.
POL.^(L) = the pollution exposure experienced by i in time
period t. The exposure is a function of i s location (L)
over the time period.
Then,
H = f (X. , W , POL . (L) )	(1)
it	i t	it

-------
7
is a hypothesized dose-response function.
To estimate a model like (1) , requires data on H^_ , X. , W^_ ,
and POL. (L) and a functional form for the model. The necessary
it
data were obtained from a health survey, the Weekly Weather and
Crop Bulletin, and the SAROAD system data tapes. The functional
form for the model was specified to be consistent with previous
studies.
The health survey data
During 1978-1980, Geomet Technologies, Inc. administered a
health survey to the members of 2,594 households in the greater
St. Louis area. Households were enrolled in the survey in groups
of about eighty per week beginning June 4, 1978 and ending May
27, 1979. The respondents maintained daily logs of their
activities, locations, restrictions in activities, and the
reasons for any restrictions in activities. The logs were kept
over four two-week periods; thus, the dataset includes the
restrictions on activities and the locations for each respondent
for 5 6 days.
The structure of the survey also facilitated the collection
of extensive data on socioeconomic conditions, lifestyle choices,
work environment, medical care, and health.2 A complete
description of the data and the datafiles are available from the
authors upon request.
The appropriate measure of the health response depends on
the focus of the study. Here, we are particularly interested in
acute respiratory health responses. As an empirical measure of

-------
8
H , we used the number of RADs reported by the respondent in the
it
time period, owing to a respiratory disorder or symptom (NRRAD^_ ) .
Given this type of limited dependent variable, a
reparameterization of the Poisson distribution is a particularly
attractive statistical model for equation (1) .
As shown elsewhere [6], the expected value of NRRAD^_ under
the Poisson model is given by
E (NRRAE^. ) = exp(X^fl + W V + P0Lit(L)5)	(2)
where the fl,	and 6 represent parameter vectors that are
estimated via maximum likelihood methods. Using equation (2), a
prediction for a small change in a POL (L) variable (or any
it
other) is a straightforward computation.
The variables in X. should include measures on i's age,
income, living arrangements, working conditions, personal health
habits, and personal health status. Since an incorrect
specification of the X could bias the estimates of the
i
relationship between H and POL (L) , we included several
it	it
covariates. Moreover, the data were limited to people between
the ages of 16 and 65 who are non-smokers and working outside of
the homes
A brief description and summary statistics of the X
i
covariates, the weather covariates, and NRRAD are presented in
Table 1.
Pollution Data
The pollution data were obtained from the U.S. Environmental
Protection Agency's SAROAD system. The data tapes contain hourly

-------
9
observations, collected at 14 monitoring sites, on numerous
pollutants in the St. Louis Air Quality Control Region.4 The
pollution data were matched to the survey respondents by time, as
described below, and location vis-a-vis the monitoring stations.
The respondents averaged about three miles from a monitoring
site.	The pollutants analyzed here, ozone and sulfur dioxide
(S02), were chosen for two reasons. First, the data for these
two pollutants were collected at all of the monitoring stations
over the time period of the health survey. For the other air
pollution measures, for example, total suspended particulate and
N02, the data are not available for several weeks during the
survey period or they were not collected at each site. Second,
S02 tends to be more localized than ozone. This contrasting
nature of the two pollutants provides a natural "laboratory" for
measuring the appropriateness of our measures and assignments of
pollution.
The pollution measures differ from the X. and the W because
an individual's exposure to pollution is not constant over t.
Pollution exposure can varyas individuals change locations over
the day. Even when an individual is stationary, their exposure
changes as the pollution varies over the course of the day.5
In defining the measures of air pollution dose, our
objective was to preserve as much variation in the pollution data
as possible. The format of the health survey means that, for
each enrollment week, there are four associated two-week periods.
Since we used the survey data from weeks 1 through 41, there are

-------
10
164 two-week periods. However, within each two-week period, we
grouped the data into "day-time" observations (the hours 11:00 am
through 5:00 pm) and "night-time" observations (the hours 5:00 am
through 10:00 am and 6:00 pm through 12:00 pm).* The pollution
data were, therefore, initially grouped into 328 subperiods.
Each day-time subperiod contains 98 observations, while the
night-time subperiods consist of 182 observations.
For each subperiod, we computed the following sets of
parameters:
(i)	The mean and standard deviation. if the data are normally
distributed, then these parameters fully describe the
distribution of the pollution.
(ii)	The mean and standard deviation of the natural logarithm of
the data. If the data are lognormally distributed, which may be
more plausible than normality, then these parameters characterize
the distribution of the dose.
Also, to facilitate a comparison with previous models, the two-
week mean and the average over the daily maximums were computed.
Two methods were developed for investigating the sensitivity
of the dose-response function to the individual's pollution
exposure. Both address the variability in air pollution doses.
The impact of locations changes on pollution measures
For an individual who lives in one location and works in
another, we are uncertain about the correct assignment of the
pollution dose. With the data analyzed here, each respondent has
two location codes; a home code and work code. The home location

-------
11
pollution, the work location pollution, or some combination of
the two could seemingly be used to assign pollution exposures to
the individuals. Moreover, the pollution data reflect different
times of the day; thus, the home code may be more appropriate for
night-time exposures and the work code more appropriate for day-
time exposures. Since we are uncertain about the correct
assignment, one possibility is to let the data determine it.
Let 9^ be the fraction of the total exposure time to air
pollution experienced during the day at work. Similarly, let 0^
be the fraction experienced at home during, our definition of,
the night-time. Finally, let 0^ be the fraction of exposure time
experienced at home during the day. We assume that 0^ + ^2 + ^3
= 1, implying that all of the exposure is experienced in the
manner hypothesized.
The 0's, if assumed to be unknown, can be estimated given
some criterion. the correct mean exposure experience by a
respondent is a weighted average of the means at each location
for each time period, where the weights are the 0's. Let
WDPOLMU = the mean of pollution calculated from the day-time
data at the work location code,
DPOLMU = the mean pollution calculated from the day-time
data at the home location code, and
NPOLMU = the mean pollution calculated from the night-time
data at the home location code.
Then,
POLAVE = 8 WDPOLMU + 8 DPOLMU + 0 NPOLMU	(2)
X	z	o

-------
12
is the weighted average pollution experienced by the respondent.
We use a grid search over the 0's to find the set that maximizes
the likelihood function.
As reference points, we alternatingly let each of the 8*s
have a value of 1. Additionally, we assumed 112 hours of
exposure per week; 40 at work, 56 at home during the night-time
hours, and 16 at home during the day-time hours. These
assumptions give 0^ = 40/112, 0^ = 56/112, and 0^ = 16/112.
The impact of intradav variation on pollution measures
Some intraday variation in the pollution data is captured by
using the day-time and night-time means. However, there does not
seem to be any theoretical reason for using the mean of the
pollution distribution to measure the dose. In fact, a simple
example illustrates that using the mean imposes a linear
restriction on the dose-response function. Assume that the
pollution exposure for some individual is variously POL1, POL2,
and POL3 over the time period. Let the probability of each level
occurring be fl, f2, and f3, respectively. Then, the mean equals
fl*POLl + f2*POL2 + f3*POL3. Next, let the dose -response
function be linear in parameter space. Write it as
H = 6q + 6 (fl*POLl) + 52(f2*POL2) + 63(f3*POL3)
for some individual in some time period. If 6. = 5,. = 6 , then
«L	M	W
using the mean is equivalent to entering the probability
distribution. On the other hand, different 5's would indicate
that a mean model incorrectly restricts the health response to
changes in average pollution; i.e., the response depends on how

-------
13
the mean changed. The problem with this method is that we have
no guidance on the number of 6's to specify. Still, this type of
analysis may provide insights into the health response of
distributions of air pollution.
IV. Empirical Results
The impacts intradav variability in Pollution exposures
In Table 2 we display the parameter estimates from several
models. Only data from the first follow-up period was used to
estimate the parameters. While the qualitative conclusions are
similar for the second follow-up period, we could not
statistically pool the data from the first two periods. The
estimates based on the third and fourth period observations were
quite different than those obtained from the first two. In
particular, most of the parameter estimates were sensitive to the
alternative specifications. This problem is apparently
symptomatic of some type of survey bias, perhaps because
respondents lost interest after the first two periods.
The specifications presented in Table 2 differ in the type
of pollution measures entered. Several other independent
variables could be selectively entered into the specifications.
Those presented here are representative of the literature. The
pollution coefficients are not particularly sensitive to any of
the measures, except the seasonal dummies and, the weather
variables. Selectively dropping these variables can change the
sign of the pollution measures in some of the specifications.
The weather variables and the seasonal dummies are statistically

-------
14
significant in each specification, however.
The first specification represents a "traditional" air
pollution dose-response function. The air pollution is assigned
to the respondents home location code. The variable labels
represent average ozone (OZMU) and average S02 (SOMU), where the
average is computed using data from the entire day. Fourteen of
the eighteen coefficient estimates exhibit p-values of less than
.05. As in other studies using a sulfur-oxide term, SOMU has a
negative sign but is insignificant.
The remaining coefficient estimates are remarkably stable
across different specifications. They show, all else equal,
that:
older respondents have fewer expected restricted activity
days due to respiratory symptoms and disorders [E(NRRAD)],
•	years of education do not affect E(NRRAD),
•	the E (NRRAD) is lower for males,
respondents in higher income classes have a lower E(NRRAD),
respondents with good perceived health have a lower
E(NRRAD),
previous smokers have a greater (or insignificant) E(NRRAD),
cooler temperatures and more rain increase E(NRRAD),
when respondents are exposed to irritants at work, E(NRRAD)
diminishes,
respondents who exercise regularly, reduce E(NRRAD), and
the greatest E(NRRAD) occurs during weeks 17-24, which is
from the end of September to the middle of November.

-------
15
One of the most interesting estimates is the coefficient on EXER.
If individuals can reduce their expected number of respiratory
related activity days by "expenditures" on regular exercise, this
may give analysts an avenue for assessing the benefits of a
cleaner environment.
Specifications (2), (3), (4), (5), and (6) illustrate the
impacts of using different assignment methodologies (or weights)
for the pollution terms. As indicated above, the actual dose is
hypothesized to be some combination of the air pollution at home
during the day (DOZLMU and DSOLMU) , the air pollution at home
during the night (NOZLMU and NSOLMU) , and the air pollution at
work during the day (WDOZLMU and WDSOLMU).
We used the mean of the natural logarithm of ozone and S02
(all variable end with "LMU") because the log of the data
appeared to be more normally distributed than the levels; thus,
using the mean of the logs is a better measure of "the central
tendency in the data.*
The impacts of changing the 0's are dramatic on the
estimated coefficients for ozone and S02. In specification (2),
0 =0=0 and 0 =1. The respondents are assigned the mean
JL	O	«
pollution at their homes computed over the day-time hours (11:00
am - 5:00 pm). The coefficient on DSOLMU remains insignificant,
but becomes positive, and the likelihood function rises (the
negative falls) slightly. Since the maximum reading usually
occurred during this time interval, this specification is similar
to [6], who used the average of daily maximums.

-------
16
The third specification shows that, when the respondents are
assigned a dose based on their home location pollution average
over the night-time hours (5:00 am - 10:00 am and 6:00 pm - 12:00
pm), the ozone influence remains stable and the S02 coefficient
remains insignificant. The S02 coefficient estimate jumps
noticeably in magnitude, however, and the likelihood function
continues to rise.
e
1 = 1 and ®2 = = ® ~"~n specification (4) . The S02 term
is significant and of similar size to the one exhibited in (3) .
The coefficient estimate on the ozone term is also significant
and about the same size as the estimate in (2) and (3) .
Specifications (5) and (6) show the impact of non-zero 8's.
The 0'S are constructed a priori in specification (5), while in
six they are estimated using a grid search. The log of the
likelihood functions are the same up to the second decimal point.
When comparing the coefficient estimates in (5) and (6) to the
estimates obtained with the other models, we see that the impact
of the pollution terms increases as the measures approach "real
world" exposures.
The empirical significance of the alternative models is
displayed in Table 3. Predictions of the expected value of NRRAD
from each specification for one thousand identical individuals
are shown. The predictions differ by the, type of change in ozone
and S02.

-------
17
A change in just ozone (Predict2) can change the prediction
on E(NRRAD) by from 1.928 per thousand to 10.671 per thousand; a
difference of over 400%. Similarly, a change in just S02
(Predict3) can change the E(NRRAD) per thousand from .356
(or -2.393, using specification 1) to 9.246; over 2000%.
Clearly, the choice of the pollution measure can have a dramatic
impact for policy analysts.
An analysis of non-mean models
As noted above, it is possible to test the mean
specification. Based on the mean and standard deviation
estimates of the logged data and assuming both pollutants were
lognormally distributed, we computed the probability that the
pollutants would fall into various categories. For ozone, we
chose four categories; 0-5, 5.01-20, 20.01-60, and greater than
60. For S02, we used 0-5, 5.01-10, 10.01-25, and greater than
25. The probabilities were computed for each of the
distributions used above (i.e., day-time work, day-time home, and
night-time home) and then averaged using the maximum likelihood
estimates for the 0's. The specifications with the probabilities
entered as dependent variables did not significantly improve the
model. * This was true for both ozone and S02, indicating that
the exposure time weighted mean model can not be rejected.
Evidently, the distributional aspects of the air pollutants are
adequately captured by specification (6) for the data analyzed
here.

-------
18
V. Conclusions
The primary conclusion of this paper is that predictions
from alternative estimated dose-response functions differ
substantially, depending on how pollution exposures are measured
and assigned to individuals. The exposure varies because
individuals" locations and air pollutants are not constant
throughout a day. Dose-response specifications that use a
weighted average of pollution experienced during the day at home,
during the day at work, and during the night at home
statistically outperform more traditional models. Moreover, the
weighted average models indicate that the pollutants adversely
affect human morbidity more than traditional models.
Our results indicate that sulfur-dioxide adversely affects
human health. This finding is different from previous studies.
The apparent reason for the difference is our treatment of the
variable nature of the pollution. This particularly appealing,
in the case of S02, because S02 is more localized than ozone.
Hence, too much aggregation in the pollution data would mask the
strength of the influence. By disaggregating the data, we have,
hopefully, uncovered the true relationship.

-------
References
1.	S. Gerking and L. Stanley, An economic analysis of air
pollution and health: the case of st. louis. Rev. Econ Statist.
68, 115-121, (1986).
2.	D. Hammer, et al, Los angeles student nurses study: daily
symptom reporting and photochemical oxidants, Arch. Environ.
Health. 28, 255-260 (1974).
3.	J. Hausman, B. Ostro, and D. Wise, Air pollution and work
loss, NBER Working Paper No. 1263 (1984).
4.	B. Ostro, The effects of air pollution on work loss and
morbidity, J. Environ. Econ. Management, 10, 371-382 (1983).
5.	B. Ostro, Air pollution and morbidity revisited: a
specification test, J. Environ. Econ. Management. 14, 87-98
( 1987) .
6.	P. Portney and J. Mullahy, Urban air quality and acute
respiratory illness, J. Urban Econ.. 20, 21-38 (1986).
7.	S. Vedal, et al, Daily air pollution effect's on children's
respiratory symptoms and peak expiatory flow, Amer. J. Public
Health. 77, 694-698 (1987).
8. A. Whittemore and E. Korn, "Asthma and air pollution in the
los angeles area, Amer. J. Public Health. 70, 687-696 (1980) .

-------
20
Endnotes
1.	Portney and Mullahy find nonlinearities in the dose-response
function. The marginal responses to increases in ozone are greater
when equations are estimated with the data reflecting higher
(greater than .05 pphm) ozone concentrations. They also find that
the elasticity of the expected value of the respiratory RAD is not
constant with respect to changing ozone as implied by the simple
poisson distribution.
2.	Survey respondents provided information on the number of visits
to a doctor, the travel time to the doctor, visits to emergency
rooms, and other "cost" measures. Gerking and Stanley [ ] use
these data to estimate a willingness to pay for reduced air
pollution expression that is based on an averting behavior model
of consumer choice.
3.	A statistical test indicated that the smokers could not be
pooled with the non-smokers using a dummy variable to reflect
smoking status.
4 . An independent benchmark, the Regional Air Monitoring System
(RAMS ) data, was used to assess the quality of the SAROAD data.
The RAMS data, which were collected in the late 1970s, but not
during the time of the health survey, were subjected to extensive
quality control and more accurately measure airborne pollutants.
During the time that both systems were operational, the ozone
readings between RAMS and nonRAMS data exhibit zero order Pearson
correlations in the range of . 3 to . 6 for the hourly data. These
correlations improve substantially as the data are aggregated to
the daily and weekly level. Hence, there is every reason to expect
that our use of the SAROAD data over a two-week period accurately
measures the air pollution dose.
5.	We did not model the weather as changing over the course of
the day. It probably should be. However, the methodology proposed
here facilitates a comparison to previous studies.
6.	We computed Chi-square tests of distributional independence for
all possible aggregations of the daily data by monitoring station.
In the vast majority of cases, we cannot reject the hypothesis that
the 11 am - 5 pm data come from the same distribution. Similarly,
we cannot reject the hypothesis that the 5 am - 10 am and 6 pm -
12 pm data come from the same distribution. The hypothesis that
all the daily .observations are generated by the same distribution
was rejected in most cases, however.
7.	The p-values are based on the variance-covariance matrix
computed directly from the maximum likelihood estimates. They do
not reflect a correction like in [6] or [3] .

-------
21
8.	To identify the appropriate distribution of the data, we
estimated the "transformed" mean and standard deviation and the
transformation parameter. These parameters are based on the
powernormal distribution, which utilizes the Box-Cox power
transformation. The Box-Cox transformation facilitates a test
between normal and lognormal distributions. In the majority of
cases for S02, the lognormal distributional assumption could not
be rejected. With respect to ozone, we found that, usually, both
the normal and lognormal distributions could be rejected. However,
the transformation parameter was closer to 0 ( indicating
lognormal), than to 1.
9.	We tested the probability model three ways. Firstly, by
entering the probabilities for just ozone. Then, by entering the
probabilities for just S02, and, finally, by entering both. None
of the Chi-squares, comparing twice the difference in the log
likelihood values, indicated rejecting the linear constraint
imposed by the mean specifications.

-------
Table 1.
Variable Descriptions and Summary Statistics
Non-smokers, First Follow-up Period.
(Observations = 597)
Variable
Description
Mean
StDev.
Minimum
Maximum
AGE
Age in years
38.88
13.37
18
65
EDUC
Years of school
13.32
2.93
0
24
SEX
1 if male
.49
.50
0
1
WHITE
1 if white
.78
.43
0
1
INCOME
Income category
6.06
1.48
1
8
PHEALTH
1 if perceived health good
.93
.30
0
1
PSMOKE
1 if previous smoker
.18
.39
0
1
TEMP
Average temperature
60.40
20.84
13
83
RAIN
Average rainfall
.63
.57
0
2.15
IRR
1 if irritants at work
.34
.47
0
1
EXER
1 if exercise regularly
.11
.31
0
1
SI
1 if weeks 1-8
.26
.44
0
1
S2
1 if weeks 9-16
.30
.46
0
1
S3
1 if weeks 17 - 24
.13
.34
0
1
S4
1 if weeks 25 - 32
.17
.38
0
1
NRRAD	Number of respiratory related restricted activity days during follow-up
period two.
Frequency for NRRAD
Value of NRRAD Frequency
0
558
1
39
2
9
3
7
4
6
5
2
6
3
7
4
8
1
9
0
10
0
11
1
12
0
13
0
14
0
Mean = .241

-------
Table 2.
Alternative Coefficient Estimates of the Dose-Response Function
Dependent Variable = NRRAD
Variable
(1)
(2)
(3)
(4)
(5)
(6)
AGE
-.033*
-.032*
-.036*
-.030*
-.034*
-.033*
EDUC
-.018
-.024
-.015
-.024
-.021
-.022
SEX
-.083*
-.121
-.048
-.118
-.060
-.074
WHITE
.617*
.676*
.798*
.734*
.704*
.692*
INCOME
-.203*
-.199*
-.188*
-.164*
-.192*
-.191*
PHEALTH
1.146
1.266*
1.201*
1.326*
1.268*
1.284*
PSMOKE
.219*
.238
.181
.129
. 185
. 188
TEMP
-.054*
-.042*
-.038*
-.037*
-.040*
-.041*
RAIN
.853*
.957*
.716*
.956*
.982*
1.031*
IRR
-.302*
-.319*
-.214
-.273
-.229
-.245
EXER
-.888*
-.863*
-.800*
-.819*
-.854*
-.861*
SI
-2.441*
-2.681*
-2.321*
-2.536*
-2.867*
-2.941*
S2
1.318*
.773
1.421*
.796*
.853*
.719*
S3
1.565*
1.268*
1.575*
1.107*
1.516*
1.446*
S4
.841*
1.036*
1.057*
1.086*
1.410*
1.424*
OZMU
.037*





SOMU
-.028





DOZLMU

.951*




DSOLMU

.065




NOZLMU


.989*



NSOLMU


.407



WDOZLMU



.905*


WDSOLMU



.367*


0ZLAVE1




1.692*

S0LAVE1




.587*

0ZLAVE2





1.708*
S0LAVE2





.571*
CONST
1.043
-2.442*
-2.858*
-3.568*
-5.560*
-5.705*
8
1
na
0
0
1
40/112
.4
e
2
na
1
0
0
56/112
.4
6
3
na
0
1
0
16/112
.2
-L like
270
269.5
268.8
269.4
264.5
264.5
~Indicates that the p-value is less than .05.

-------
Table 3.
Predicted Reductions in the Expected Value of NRRAD
Per 1000 People by Specification
(1)
(2)
(3)
(4)
(6)
Predictl
Predict2
Predict3
.637
2 .290
-2.393
2.197
1.928
.356
14.294
8.188
8.119
5.418
3.089
3.022
16.351
10.671
9.246
12.940
8.546
7.186
Notes: The predictions are based on the following initial values: AGE=40, EDUC=12, SEX=1,
WHITE=1, INCOME=6, PHEALIH=1, PSMOKE=0, TEMP=70, RAIN=.5, IRR=0, EXER=0, Sl=l, S2=0, S3=0,
and S4=0. The initial value for ozone is 40, while the initial value for S02 is 20. The
predictions are per 1000 people, where:
Predictl is based on reducing ozone to 30 and S02 to 10,
Predict2 is based on reducing ozone to 30 and maintaining S02,
Predict3 is based on reducing S02 to 10 and maintaining ozone.

-------
Pricing Environmental Health Risks:
Survey Assessments of
Risk-Risk and Risk-Dollar Trade-offs*
by
W. Kip Viscusi
Wesley A. Magat
and
Joel Huber
Revised, April, 1989
*This research was supported by EPA Cooperative Agreements to
Northwestern University (CR-814478-01-0) and Duke University (CR-
814388/01-0) . We would like to thank our contract officer, Dr.
Alan Carlin, and his colleagues at EPA for helpful suggestions
with respect to our study's focus and survey design.

-------
Ab s t r a c t
This study develops a methodology for measuring the values
that individuals place on morbidity risk reductions and applies
it to the measurement of the benefits from reducing the risks of
contracting chronic bronchitis. The survey methodology involves
the use of an iterative computer program that presents
respondents with a series of pairwise comparisons which are
individually designed to measure respondents' marginal rates of
substitution for chronic bronchitis risk reduction. The approach
is innovative in that it measures the rates of trade-off for
chronic bronchitis risk reduction in terms of the risk of an
automobile accident fatality, as well as in dollars. Since it
generates estimates for each individual, it can reveal
distributions of benefit measures rather than simply a population
mean estimate. The resulting rates of trade-off for chronic
bronchitis and auto fatality risks suggests that the risk of a
chronic bronchitis case is worth 32% of the comparable risk of
death, as measured by the median trade-off rate. When risk
reduction for chronic bronchitis is compared to a cost of living
increase, the median rate of trade-off is $457,000, whereas the
comparison between automobile fatality risk reductions and cost
of living increases yielded a median rate of trade-off of $2.29
million. The results across different risk-risk and risk-dollar
trade-offs were internally consistent.

-------
1. Introduction
Over the past decade economists have devoted substantial
attention to the implicit valuation of health outcomes. These
analyses of risk-dollar trade-offs have relied in large part on
market-based data.^" For example, wage-risk trade-offs have been
used to analyze the implicit value of fatalities and the average
nonfatal job accident risk. Similarly, economists have analyzed
the trade-offs implied by seat-belt usage decisions to infer a
value of life.2
Although studies using market data provide useful benchmarks
for health risk valuation, they do not resolve the issue of how
government agencies should attach benefit values to health
outcomes for which we do not have good market data. This
omission is particularly important for government agencies, such
as the U.S. Environmental Protection Agency (EPA), which
generally focus on policy contexts in which market forces are
believed to not be fully effective. For these situations, no
useful market trade-off data may be available. Nevertheless,
economic analysts would like to select the efficient project mix,
and some benefit measure is required to perform such an analysis.
In recent years, a large number of studies have addressed these
benefit issues using non-market techniques, thus greatly
expanding the range of benefit components that can be valued.
This study makes several contributions to the literature on
non-market techniques for benefit valuation. First, we develop a
methodology for measuring the benefits of reducing the risks from
various types of morbidity effects. The methodology uses an

-------
2
iterative computer program to ascertain the points of
indifference for consumers who are asked to trade off the reduced
morbidity risk with increases in other attributes of a location
decisions, such as an area's cost of living and the risk of an
automobile fatality.^
Second, we apply the methodology to an important health
benefit valuation problem, that of estimating the value of
reductions in the risk from chronic bronchitis, one central type
of chronic obstructive pulmonary disease alleged to be a major
adverse effect of ozone pollution exposure. Most previous
studies of health valuation focus on acute health effects, such
as accidental death, rather than chronic diseases whose effects
are more difficult to communicate to potential victims.
Third, our approach yields the entire distribution of
consumer values for chronic bronchitis risk reduction, rather
than just the mean valuations which can be derived from market-
based approaches to the problem. This information is important
for policy makers in situations where consumers place, widely
divergent values on reducing risk.
Fourth, because chronic disease effects are difficult to
communicate to potential sufferers, it is important to use a
methodology that adapts to whether subjects understand the
valuation task being asked of them. By administering the
questionnaires interactively on a computer, our approach allows
us to build in several tests of task comprehension that, if
failed, provide additional information before proceeding with the
questionnaire .

-------
3
Finally, our methodology produces values for morbidity risk
reduction in terms of trade-offs with several other metrics
besides money. In our chronic bronchitis application, we measure
trade-offs with the risk of automobile fatalities, as well as
with a dollar measure derived from changes in the cost of living.
Many policy-makers are hesitant to base decisions on benefits
denominated in dollars, and they may be more willing to
implicitly consider benefit values when measured in units of a
common risk such as death. Converting all health outcomes into
death risk equivalents facilitates cost-effectiveness analysis by
calculating the cost per statistical life equivalent saved, and
it addresses concerns with respect to dollar pricing. Even if
the morbidity valuations are elicited in terms of trade-offs
between risks, they can still be converted into dollar values by
using hedonic measures of the value of the comparison risk if
that comparison risk is death (with the appropriate application
of sensitivity analysis to the assumed values of life used to
make the translation).
There are reasons to suspect that consumers may have fewer
difficulties with the task of specifying rates of trade-off of
one risk with another, as opposed to trading off a risk with a
certain dollar amount. The risk-dollar trade-off task sometimes
produces alarmist responses from subjects who cannot envision
that they would voluntarily subject themselves to a higher risk
of a serious morbidity effect for a finite amount of additional
income. ^ Dollar valuation tasks also are difficult to design in
ways that subjects will find analogous to real choice situations,

-------
4
and they may offer biased responses to questions that do not
force them to pay for the risk reduction being valued. There is
a final reason to prefer the risk-risk trade-off approach. To
the extent that consumers are equally adverse to the risks from
different types of risks, asking them to trade off one risk
against another produces rates of trade-off which measure the
relative value to them of the two risks without regard to the
risk aversion which enters in trading off uncertain health risk
with certain dollars. in this sense the risk-risk trade-offs
provide values which are not as heavily influenced by the
consumers' attitudes towards facing risks per se.
The outline of this paper is as follows. Section 2 provides
an overview of the study design and the sample. Section 3
describes the risk-risk trade-offs whereby respondents put their
chronic morbidity valuations into auto death equivalents. In
Section 4 we describe the direct estimates of risk-dollar trade-
offs for chronic bronchitis obtained by asking respondents to
trade off chronic bronchitis risks with either the area's cost of
living or property damage from storms. As a check of the
validity of the approach, we provide evidence on auto fatality
risk-dollar trade-offs in Section 5. These implicit value of
life numbers are tested against those in the literature to assess
the valibity of the survey approach. In Section 5 we also
convert all of our results for the value of chronic bronchitis to
dollar equivalents. Section 6 concludes the paper.

-------
5
2. Study Design and Sample Description
Genera 1	Approach
We used a sample of 593 shoppers from a blue-collar mall in
Greensboro, North Carolina to measure willingness-to-pay values
for reducing the probability of contracting chronic bronchitis.
The subjects made four series of pairwise comparisons of
different locations where they could live with the locations
differing in two attributes. In most of these comparisons, one of
the locational attributes varied was the probability of
contracting chronic bronchitis.
The first series of questions yielded a rate of trade-off
between decreases in the risk of chronic bronchitis (CB) and
increases in the risk of an automobile fatality, thus providing
what we call a "risk-risk" trade-off. The second series of
questions determined a "risk-dollar" trade-off, where the
reduction in the risk of CB was achieved at the expense of a
location with a higher cost of living.
If subjects were found to more easily trade off a reduced CB
risk with a higher auto fatality risk than with a cost of living
increase, we wanted to sort out whether this result was due to
the fact that the cost-of-living differences were measured in
dollars or that they were given with certainty (that is, with no
risk involved over dollar gambles) . Thus , our third series of
questions asked subjects to trade off reductions in the CB risk
with increases in a lottery on dollar losses expressed as a risk
of storm damage, where if a storm were to occur, it would cause
$2,000 of damage to the subject's home and belongings. Finally,

-------
6
in order to compare the CB risk--auto fatality risk trade-offs
with the risk-dollar trade-offs, it was useful to obtain a dollar
measure of the value of reducing the risk of automobile
fatalities. This fourth series of questions provided a rate of
trade-off of risk reduction in automobile fatalities to increases
in a location's cost of living.
The results from these four series of questions allows us to
address the following questions:
*	What is the distribution of CB risk--death risk trade-offs?
*	What is the distribution of CB risk-- (certain) dollar trade-
offs?
*	What is the distribution of CB risk-- (uncertain) dollar trade-
offs?
*	Which of these three trade-offs is easier to elicit accurately
from consumers?
*	What is the distribution of death risk--(certain) dollar
trade-offs?
*	How does the distribution of CB risk-- (certain) dollar
trade-offs compare with the distribution of CB risk--dollar
trade-offs derived from combining the CB risk--death risk
trade-offs with the death risk--(certain) dollar trade-offs?
*	How does the distribution of CB risk-- (certain) dollar trade-
offs compare with the distribution of CB risk--dollar trade-
offs derived from combining the CB risk--death risk trade-offs
with the values of life derived from wage hedonic studies?
It should be noted that the first question is the most
important one to answer because it addresses the use of an

-------
alternative metric to dollars for measuring morbidity risk
willingness-to-pay values, that of another health risk, namely
death. For cost-effectiveness purposes, it is not necessary to
go beyond the death risk metric, as alternative policy
initiatives can be compared on the basis of this metric rather
than dollars. However, if the CB risk values measured in death
risk units translate closely to the direct dollar valuations of
reducing CB risks that we obtain, policy makers can be more
confident in the reasonableness of the risk-risk valuations.
In order to understand our empirical results that allow
responses to the questions above, it is first necessary to
carefully describe the design of the survey and sample.
Methodology
The task of eliciting individuals' valuation of chronic
bronchitis is not straightforward. The first problem is that few
individuals fully understand the health effects of chronic
bronchitis. Second, once given this information, they may not
have sufficient experience in dealing directly with such trade-
offs to give meaningful valuation responses. To accommodate
these difficulties, we developed an interactive computer program
that would inform consumers as well as elicit trade-off
information.
Three different questionnaires were used, but for
concreteness let us focus on what we will designate Questionnaire
A. After acquainting the respondent with the computer, the
program elicits information regarding the respondent's personal
characteristics (e.g., age) . A substantial portion of the

-------
8
questionnaire (about 40 questions) is then devoted to acquainting
the respondent with the health implications of chronic bronchitis
and the nature of the trade-offs that would be encountered.
These questions elicit the respondent's familiarity with chronic
bronchitis, information on smoking history, and provide a
detailed summary of the health implications of chronic
bronchitis .
The thirteen principal health implications of chronic
bronchitis are summarized in Table 1. The chronic bronchitis
disease classification includes a variety of illnesses of
differing severity. Our intent was not to value each possible
combination of systems, but rather to establish a methodology
that could be used to value this and other adverse health
effects. Consequently, our valuation procedure pertains to the
set of symptoms summarized in Table 1, but the broader purpose of
our analysis is to develop a methodological approach that is more
generally applicable to other patterns of chronic bronchitis, as
well as to different diseases such as cancer.
Since chronic bronchitis takes many forms, this study
focused on the most severe chronic morbidity effects. ^ Thus, the
survey's focus is on the adverse health outcomes at the extreme
and of the cluster of diseases within the chronic bronchitis
grouping. Because a quick overview of these effects may not be
fully comprehended by respondents, in each case subsequent
questions ascertain the respondents' assessed disutility ranking
of each outcome in a linear 49-point scale. The purpose of these
questions is not to establish attribute-based utilities, but to

-------
9
Table 1
Health Implications of Chronic Bronchitis
1.	Living with an uncomfortable shortness of breath for the
rest of your life.
2.	Being easily winded from climbing stairs.
3.	Coughing and wheezing regularly.
4.	Suffering more frequent deep chest infections and pneumonia.
5.	Having to limit your recreational activities to activities
such as golf, cards, and reading.
6.	Experiencing periods of depression.
7.	Being unable to do the active, physical parts of your job.
8.	Being limited to a restricted diet.
9.	Having to visit your doctor regularly and to take several
medications.
10.	Having to have your back mildly pounded to help remove
fluids built up in your lungs.
11.	Having to be periodically hospitalized.
12.	Having to quit smoking.
13.	Having to wear a small, portable oxygen tank.

-------
10
encourage respondents to think carefully about the health
implications of chronic bronchitis and their own view of the
effect of this disease on their well-being.
At this point in the questionnaire, the respondents confront
the first of two set of trade-off questions. Individuals are
presented with a choice of moving to one of two alternative
locations which differ in terms of their chronic bronchitis risk
and automobile accident risk. To ensure that respondents would
be willing to consider making such a move at all, they were told
that these two locales posed a lower risk of both outcomes than
their current place of residence.
Since risk levels differ across individuals, the program
elicits information regarding individual activities that are
likely to influence their person-specific risks, such as smoking
habits (for chronic bronchitis) and mileage driven per year (for
auto accident deaths). The program then informs the respondents
that the probabilities presented in subsequent questions are
calculated based on their responses to the earlier risk-related
activity questions, even though the same risks are actually
•	ft
presented to all subjects. This procedure increases the extent
to which the stated risk levels are taken at face value, while
facilitating the comparison of risk trade-offs across subjects
because they all responded to the same risks.
To ensure that respondents understand the task before
proceeding to questions in which one location is lower in one
risk but higher in the other risk, they are first presented with
a dominant choice situation. Let the notation (x,y) denote a

-------
11
locale where the chronic bronchitis probability is x/100,000 and
the automobile death risk is y/100,000. The actual survey did
not present the choices in such abstract terms, but this notation
9
makes the exposition of the survey structure simpler.
To ascertain whether respondents understand the task, they
are first asked whether they prefer Area A with risks per 100,000
population of (75, 15) or Area B with risks (55, 11) . Since each
of the Area B risks is lower, this alternative is dominant.
Respondents who do not comprehend the task and incorrectly answer
that they prefer Area A are sent through a series of questions
that explain the structure of the choice in more detail.
The performance with respect to the dominance question was
quite good. Over four-fifths of the sample gave a correct
response to the dominance questions on their initial attempt.
After being given additional information, fewer than one percent
of them gave an incorrect answer, and these respondents were
excluded from the sample since they did not understand the
interview task.
The program then proceeds with a series of pairwise
comparisons in which the attributes are altered based on the
previous responses until indifference is achieved. The computer
program used tabular summaries, but for expositional purposes we
will consider the abstract formulation of the trade-offs.
&	Model	Ql	State-Dependent	TTti 1 i ti es
Consider the following model of state-dependent utilities.
Let subscript a denote Area A and b denote Area B. Also, let
U (CB) be the utility of a case of chronic bronchitis, U(D) equal

-------
12
the utility of an auto accident death, and U(H) equal the utility
of being healthy (i.e., having neither CB nor an auto accident).
To simplify this exposition, we assume that contracting CB and
dying from an automobile accident are mutually exclusive events.
Also, let Xa denote the probability x/100, 000 for Area A and Ya
denote the probability y/100, 000 for Area A, and let Xb and yb be
defined similarly. The survey continually modifies the choice
pairs until subjects reached the situation where
(1)	XaU(CB) + YaU(D) + (l - xa - Ya)UtH)
= XbU(CB) + YbU(D) + (1 - Xb - Yb)U(H).
Our general objective is to establish the death risk
equivalent of chronic bronchitis. If we assume for concreteness
that Xa > Xb and Yb > Ya (no loss of generality) , then
(2)	(Xa - Xb)U(CB) - (Yb - Ya)U(D) + (Xa - Xfa + Ya - Yb)U(H),
"*b-*a	*b-Ya
(3)	U(CB) - 	 U(D) + (1 - 	)U(H) .
xa-xb	xa-xb
If we define the rate of trade-off between CD and D as	SO
that
(4)	tx -
Yb-Ya
xa~xb
we obtain the result that
(5)
U (CB)
= t-^U(D) + (1
- t1)U(H)

-------
13
The utility of CB cases has been transformed into an equivalent
lottery on life with good health and death, for which we have a
well-developed literature.
Survey Structure
Now consider the first set of paired comparison questions
presented in Questionnaire A after the dominant choice question
described above. In this case, respondents are given the choice
between Area A with risks (75, 15) and Area B with risks (55,
19). Suppose that Area B is preferred in this example. Area B
has the lower chronic bronchitis risk and higher auto accident
risk; therefore, in subsequent questions the program raises the
CB risk in the preferred Area B until indifference is achieved.
If in the original choice the subject prefers Area A, in
subsequent questions the program lowers the auto death risk in
Area B until the point of indifference is reached.
Suppose that after considering a series of such comparisons
the subject reaches indifference where he views the risk (75, 15)
as being equivalent to (65, 19) . Using equations 4 and 5 above,
this would imply that
19-15
t, = 	 = 0.4
75-65
and
U (CB) = 0 . 4U (D) + 0 . 6U (H) .
The second set of paired comparison questions in
Questionnaire A focuses on the more traditional risk-dollar
trade-off involving CB and cost of living. Area A has the same

-------
14
cost of living as the respondent's present residence, but Area B
has a cost of living that is $80 higher, yet poses a lower CB
risk Xb. If in the initial question Area B is preferred, Area
B's CB risk is increased until indifference is achieved.
Similarly, if Area A is preferred, Area B's cost of living is
reduced until reaching the point of indifference.
In the context of a state-dependent utility function with
two arguments, health status and income, we have
XaU(CB) + (1 - Xa)U(H) - XbU(CB,-$80) + (1-Xb)U(H,-$80).
If utility functions are additively separable in money and
health, then
XaU(CB) + (1 - Xa)U(H) - XbU(CB) + (l-Xb)U(H) + U(-$80),
which simplifies to
(Xa - Xb)U(CB) = U(-80) + (Xa - Xb)U(H),
or
U(-$80)
U(CB) - 	 + U(H) .
(Xa-Xb)
If we assume that utility is linear in money (with a coefficient
equal to one) in establishing our health valuation scale, then we
have
U (CB) = -L + U (H) ,
i.e., CB is equivalent to being healthy and suffering a financial
loss tantamount to L dollars, where

-------
15
-$80
L » 	 .
xa"xb
This procedure to establish a risk-dollar trade-off rate
involves two assumptions regarding the structure of utility
functions. First, we assume additive separability with respect
to money and health. Second, we assume that the dollar
magnitudes treated are sufficiently small that utility is
approximately linear in money. Since even risk-averse utility
functions meet this test for small monetary changes, ^ we
selected our health-risk levels so that the dollar magnitudes
involved be small.
The structure of Questionnaire B is similar to Questionnaire
A except the certain $80 loss in terms of living costs has been
replaced by a lottery on $2000 storm damage loss. -*-n this case,
respondents must specify the storm damage probability that
establishes an equivalent CB-storm damage pair. If we assume
that respondents are risk-neutral, then the storm damage loss can
be replaced by its expected value. The possible advantage over
the cost-of-living approach is that respondents may be able to
make more meaningful comparisons of two different lotteries
rather than having one attribute -- the dollar payoff -- being
non-stochastic. As with the first set of questions in
Questionnaire A, if the consumer prefers Area B in the initial
question, the program leads the consumer to indifference by
increasing the CB risk of Area B until indifference is achieved.
Similarly, when the consumer initially prefers Area A, the

-------
16
program reduces the storm damage risk in Area B until reaching
the point of indifference.
Questionnaire C repeats the first part of Questionnaire A,
and these samples are pooled in the analysis below. The second
set of questions addresses the more traditional death risk--
dollar trade-off using auto deaths and cost-of-living trade-offs.
The structure is similar to that of the second set of questions
in Questionnaire A except that CB has been replaced by auto
fatality risks so that respondents must reach the point that
U(D) = -L + U(H),
where
-80
h + 	
xa"xb
as before. This portion of the study provides a direct
comparability test with the literature on market-based values of
life. The fatality risk--dollar trade-offs will also be used in
conjunction with the chronic bronchitis--fatality risk trade-offs
to establish a chronic bronchitis--dollar trade-off rate.
Table 2 summarizes the structure of the 3 questionnaires
described above.
Sample Description
The interviews of the subjects were all done through an
interactive computer program, thus avoiding problems of
interviewer bias and promoting honest revelation of preferences.
Response rates to sensitive questions, such as income level, were

-------
1
Table 2
Summary of Survey Structure
Questionnaire A
Trade-Off
1. Chronic bronchitis
auto deaths
Units of Measurement
Auto deaths per chronic
bronchitis case
Procedure
In the area with the
higher auto accident
risk, increase the
bronchitis risk (to
make that area less
desirable) or reduce
the auto accident
risk (to make that
area more desirable)
until reaching in-
di f ference.
2. Chronic bronchitis
cost of living
Dollar value per
1/100,000 reduced
risk of bronchitis
In the area with
lower bronchitis
risk, increase the
bronchitis risk (to
make that area
less desirable) or
decrease the cost of
living (to make that
area more desirable)
until reaching in-
di f ference.
Questionnaire B
1. Chronic bronchitis
storm damage
Reduced probability of
$2000 storm damage
that is equivalent to
one bronchitis case
prevented
In the area with the
higher storm damage
risk, increase the
bronchitis risk (to
make that area less
desirable) or reduce
the storm damage
risk (to make that
area more desirable)
until reaching
indi f ference.

-------
18
Table 2 (cont'd)
Summary of Survey Structure
Questionnaire C.
1. Chronic bronchitis
auto deaths
Auto deaths per chronic
bronchitis case
!Same as Questionnaire A - Part 1]
In the area with the
higher auto accident
risk, increase the
bronchitis risk (to
make that area less
desirable) or reduce
the auto accident
risk (to make that
area more desirable)
until reaching in-
di f ference.
2. Auto accidents -
cost of living
Dollar value per
1/100,000 reduced
risk of an auto
accident
In the area with
lower auto accident
risk, (to make that
area less desirable)
or decrease the cost
of living (to make
that are more
desirable) until
reaching indif-
ference .

-------
19
much higher than those usually achieved with face-to-face
interviews. In addition, subjects were not concerned with
whether their responses impressed the interviewer. Use of a
computer also made it possible to ask a sequence of questions to
ascertain the appropriate marginal rates of substitution.
The sample was recruited for the study by a professional
marketing firm at a mall intercept in Greensboro, North Carolina.
This locale has a representative household mix and is used as a
test marketing site for many national consumer brands. This firm
and locale have been used successfully in two previous studies by
the authors. ^ Use of such a consumer sample also yields more
reliable responses to issues such as the valuation of property
damage from storms than would a student sample or a sample from a
city with an unrepresentative population, such as the college-
oriented cities of Evanston, Illinois, or Chapel Hill, North
Carolina.
Table 3 provides a glossary of the variables and the
associated sample statistics. Questionnaires A and C had a
similar mix of respondents, with a mean age in the low thirties,
a even split between males and females, two years of college
education, a 50 percent married rate, about 0.6 children under 8
years old, a household size of 2.7 - 2.8, and a household income
in the mid-range of thrifty to forty thousand dollars.
Questionnaire B has a somewhat different mix because of the
difference in the times at which the samples were recruited
(e.g., week-end shoppers differ from day-time weekday shoppers).
The Questionnaire B sample is about 10 years older, more likely

-------
Table 3
Summary of Sample Characteristics
Mean and Std. Deviations
Questionnaire
Demographic
Variables
A
AGE, in year
33 .74
(12.42)
43.47
(12.68)
33 . 07
(11.66)
MALE, sex dummy
vari able
0 .50
0 . 42
0 .51
EDUCATION, years of
schooling
14 . 02
(2.23)
14 . 32
(2.47)
13 .79
(2.66)
MARRIED, married
dummy variable
0.49
(0.50)
0 .79
(0.41)
0 .49
(0.50)
KIDS, number of
children under 8
0 .56
(1.00)
0.83
(1.04)
0 . 65
(1.07)
HOUSEHOLD, number of
people in household
2 .71
(1.25)
3 . 00
(1.16)
2 .80
(1.23)
INCOME, annual	35,386.60	45,367.65	37,153.85
household income	(19,009.95)	(20,335.54)	(21,333.80)
in dollars
194
204
195

-------
21
to be married, and with a household income about $10,000 greater.
As the last row of Table 3 indicates, each of the three samples
had about 200 respondents, with combined sample for the study of
593.
3. Risk-Risk Trade-Offs
Table 4 displays the means and standard deviations of the
trade-off rates implied by the indifference points of the subject
responses. To go beyond these summary statistics, consider first
set of trade-offs between CB and auto accident deaths. For this
analysis Questionnaires A-l and C-l are pooled since the
questions are identical.
Establishing a death risk metric for CB enables respondents
to think in risk terms, avoiding the comparability problems that
might be encountered if monetary attributes were introduced.
Similarly, for policy purposes EPA can establish a death risk
equivalent and establish cost-effectiveness ratios in terms of
the cost per statistical death prevented. As indicated in
Viscusi (1986), this cost-effectiveness index will provide a
comprehensive measure of the policy impact and also avoid the
political sensitivities of placing dollar values on all health
outcomes. Once a uniform health metric is established, one can
then compare the cost per life equivalent saved with various
value-of-life reference points and decide whether the policy
should be pursued if one wishes to take a benefit-cost approach.
Unlike market-based studies of the value of life, the survey
technique yields information on the entire distribution of the
valuations. Table 5 reports the deciles of the distribution for

-------
22
Table 4
Rates of Trade-off Implied by Indifference Points
Means and Std. Deviations
Part B
Trade-off Rates
CB-Auto (A-1 & C-l),
auto deaths per CB case
CB-Cost of Living (A-2),
dollar value per
1/100,000 CB risk
CB-Storm Damage (B-l),
number of $2,000 storms
equal to one CB case
Auto-Cost of Living (C-2), -
dollar value per 1/100,000
reduced auto accident risk
A
0. 68
(0.82)
8 . 83
(12.501
Questionnaire
852.60
'1064.20)
C
0 .70
(0.95)
81.84
'168 . 54'
Sample Size
194
204
195

-------
23
Table 5
Distribution of Chronic Bronchitis
Auto Death Trade-Offs
Auto Death Equivalents per Chronic
	Bronchitis Case	
Decile
.10	0.12
.20	0.20
.30	0.23
.40	0.27
.50	(median) 0.32
.60	0.40
.70	0.80
.80	1.00
.90	1.33
1.00	4.00
Mean	0.68
(St.	error of mean) (0.06)

-------
24
respondents who gave consistent answers that converged to a
particular trade-off value. Subjects whose responses indicated
that they did not fully comprehend the valuation task were
excluded from our sample.
Specifically, we excluded subjects who failed one of the
following consistency checks:
1)	they started the series of paired comparison questions by
preferring one area, say Area A, and as Area B was made
more desirable in subsequent comparisons they continued to
prefer Area A, even on the last question of the series in
which Area B dominated Area A on both attributes;
2)	like inconsistency #1, they continued to prefer Area A in
each comparison until the last one in which Area B dominated
Area A in both attributes, yet on this last question they
indicated indifference between Area A and Area B;
3)	they indicated preference for one area, say Area A, on the
first and all subsequent questions in the series (including
the last one in which Area B dominated Area A) , then when
confronted with this inconsistency and asked to repeat the
series of questions chose Area B in the first question
(despite have selected Area A the first time they were given
this question);
4)	they indicated preference for one area, say Area A, on all
questions in the series except the last one in the series (in
which Area B dominated Area A) but including the next-to-last
question (for which Area B easily dominated Area A on one
attribute and Area A just barely dominated Area B on the

-------
25
other attribute) , thus making it impossible to interpolate
between the trade-offs implied by the last two questions to
obtain an indifference point (because the last question
yields no rate of trade-off) ; or
5) they expressed indifference between all pairs of areas in the
series of questions, despite wide variation in their
attributes.
Individuals who failed one of these inconsistency checks either
did not understand the choice task, were not responding honestly,
attached no value to one of the two attributes, or have non-
monotonic preferences for one of the attributes. we assume that
neither of the last two preferences attributes are possessed by
any subjects, thus implying that answers which fail any of the
five inconsistency checks do not represent the subjects' true
preferences .
The requirement that the response pattern to the series of
paired comparisons be internally consistent will lead to more
meaningful estimates than if no such checks were imposed. About
two-thirds of the sample converged to an indifference situation
and had consistent responses, where this percentage was similar
across all questionnaires. ^ These consistency checks
distinguish our approach from the usual contingent valuation
method in which respondents' answers are taken at face value
without such formal tests of whether the subjects understood the
valuation task and displayed consistent choices.
In evaluating the distribution in Table 5, first consider
the respondent at the tenth percentile. This person viewed a

-------
26
chronic bronchitis probability as being just as severe as a risk
of an auto accident that was 0.12 as great. Thus, this
individual would view a chronic bronchitis risk of 100/100,000
risk of 100/100, 000 per year as being equivalent to the annual
chance of being involved in an auto accident of 12/100,000.
Now examine the respondent at the other end of the
distribution. This individual views a chronic bronchitis risk as
being four times as severe as a risk of death, so that a
100/100,000 risk of CB would be viewed as comparable to a
400/100,000 risk of death. He or she gave consistent responses
to the questions, but opted for the choice reflecting the highest
CB valuation.
Many studies in the survey valuation literature exclude the
tails of the distribution since they are tainted by extreme
respondents such as this. Rather than discard such information
altogether, we report the entire distribution, recognizing that
the top and bottom deciles may be affected by a lack of complete
understanding of the interview task. The reported distributions
enable readers to assess how important outliers are within the
context of the study and by focusing primarily on the median
responses rather than the mean we avoid the distortion of our
results by these outliers.
The response pattern in which CB was more highly valued than
auto death risks was exhibited by the top two deciles for each
questionnaire's response distribution. Such a pattern is not
necessarily implausible. In addition to possibly
misunderstanding the interview task, two explanations can be

-------
2
offered. First, individuals might legitimately believe that such
a severe chronic illness is a worse outcome than death. The
health outcome described in Table 1 is quite serious and will
have substantial duration. Their normal activities would be
curtailed, medical interventions including hospitalization and
possible reliance on a portable oxygen tank would accompany
severe cases of CB, other illnesses would be more likely, and
they would experience periods of depression.
The second possible explanation is that the respondents were
establishing equivalences between different average risks in an
area rather than different risks to themselves. The CB risk was
characterized as an involuntary risk not under their control
except for smoking, whereas the auto accident risk differs
depending on one's driving habits and skills. Other studies
suggest that individuals may have overly optimistic assessments
of risks influenced by their actions, such as auto death risks,
as discussed in Viscusi and Magat (1987) . If this were the case,
the perceived person-specific risk would be below the stated
risk, causing an upward bias in the results in Table 5.
The median CB valuation is equivalent to 0.32 auto deaths.
Because of the skewed nature of the responses, the mean value of
0.68 is more than double the median response. Regression
analysis of the CB-auto death trade-off rates indicate no
significant variation across subjects with respect to either
demographic factors such as age, income, and education, or
personal characteristics such as smoking habits. This result is
neither surprising nor disturbing. Most individual attributes,

-------
28
such as household income, should affect the CB valuation and the
value of life similarly, and thus be unrelated to variation in
the CB--auto death trade-off rates across subjects. Because
there are no systematic differences among individuals in their
risk-risk trade-offs, we can aggregate them into meaningful
summary measures such as medians and means without the risk of
drawing misleading conclusions from an unrepresentative sample.
The general implications of these results is a follows.
Most, but not all, people regard the risk of chronic bronchitis
as a less severe outcome than the risk of death. However, the
prospect of a sustained chronic illness is viewed as a very
severe outcome. Based on the median responses, the death risk
equivalent of CB is. 0.32, and based on the mean response it is
0.68. The general order of magnitude of both the median and the
mean is the same and is just below that of fatalities. As will
be indicated in Section 5, these statistics can be transformed
into dollar valuation equivalents using established value-of-life
statistics.
4. Risk-Dollar Valuations of Chronic Bronchitis
The second approach that we employed to value chronic
bronchitis was to establish risk-dollar trade-offs. The two
approaches used were to establish the chronic bronchitis risk
equivalent of a higher cost of living and to determine the
relationship between chronic bronchitis risks and storm damage
risks.
Consider first the cost-of-living results in Table 6. The
first column of Table 6 lists the decile of the distribution.

-------
29
Table 6
Distribution of Chronic Bronchitis
Cost of Living Trade-Offs
Decile
Trade-Off Levels
Dollar Value per
1/100,000 Reduced Risk
of Chronic Bronchitis
(A-2)
Implicit Dollar
Value per Case
of Chronic
Bronchitis
, 10
.20
.30
.40
.50 (median)
. 60
.70
.80
.90
1.0
Mean
(St. error of mean)
1.50
3.00
3.50
4 . 00
4.57
5.33
6.40
8 . 00
20 . 00
80 . 00
8 . 83
(1.14)
$150,000
$300,000
$350,000
$400,000
$457,000
$533,000
$640,000
$800,000
$2,000,000
$8,000,000
$883,000
($114,000)

-------
30
Column two presents the increased dollar value in the annual cost
of living that the respondent was willing to incur per 1/100,000
reduction in the annual probability of chronic bronchitis. If we
multiply the results in column 2 by 100,000, we obtain the
implicit dollar value per statistical case of chronic bronchitis.
As in the case of the risk-risk results, the response
pattern is skewed so that the upper tail of the responses
generates a mean valuation estimate in excess of the median. The
results here indicate the average dollar value of chronic
bronchitis is $883,000, with an associated standard error of
$114,000. The $457,000 median of the distribution is just over
half of the mean. Each of these values is below the usual
estimates of the implicit value of life, which are reviewed in
Viscusi (1986). These results follow the expected pattern, given
the CB--auto death risk trade-off results reported above.
As in the case of the risk-risk trade-offs, the upper bound
of the chronic bronchitis valuation estimates exceeds most
estimates of the value of a fatality, as $8 million exceeds some
but not all estimates of the value of life. More precise
comparisons of all of the results using a dollar metric will be
undertaken in Section 5.
The second set of CB risk-dollar trade-offs, which is
reported in Table 7, uses storm damage risks as the dollar
counterpart so that respondents must compare monetary lotteries
and health status lotteries rather than certain monetary (cost of
living) differences with health status lotteries. The first
column of results gives the value of y for which a storm causing

-------
31
Table 7
Distribution of Chronic Bronchitis
Storm Damage Trade-offs
Decile
Equivalent $2000
Damage Probability
(xlOO,000)
Implicit Dollar
Value per Case
of Chronic
Bronchitis
, 10
.20
. 30
.40
.50 (median)
.60
.70
.80
.90
1. 0
Mean
(St. error of mean)
175.00
228.57
266.67
266.67
400.00
533 . 33
800.00
1,333.33
2,000.00
4,000.00
852.60
(91.93)
$350,000
$457,140
$533,340
$533,340
$800,000
$1,066,660
$1,600,000
$2,666,660
$4,000,000
$8,000,000
$1,705,200
($183,860)

-------
32
damage of $2000 with a probability of y/100,000 is equivalent to
a chronic bronchitis probability of 1/100,000. A more meaningful
metric is the expected storm damage that is equivalent to each CB
case. This figure is obtained by multiplying the first column of
results by the $2000 damage per storm damage event. The second
column of results gives the dollar value per statistical case of
chronic bronchitis, where these dollar values have been obtained
using the storm damage costs.
A comparison of the distributions of implied CB valuations
in Tables 6 and 7 suggests that the subjects may have found the
storm damage lottery comparison to have been more difficult to
make than the comparison with a non-probabilistic cost-of-living
increase. The distribution derived from the storm damage lottery
comparison stochastically dominates the distribution from the
cost-of-living comparison, with both its median and mean almost
double that of the cost-of-living distribution. Based on a
comparison with the dollar values of avoiding automobile accident
fatalities reported in next section, the CB avoidance values
derived from the storm damage lottery questions appear to be
somewhat high. Further, the standard error of the mean is about
50 percent higher for the distribution derived from the storm
damage distribution than for the cost-of-lived based distribution
of CB values. In any event, these results do not suggest that
expressing dollar trade-offs in probabilistic form, as in the
storm damage lottery, aids people in making risk-dollar trade-
offs, which was our original hypothesis.

-------
33
5. Trade-Offs Between Auto Deaths and Cost-of Living
A useful check on the survey methodology is to ascertain the
implicit value of life using a direct fatality risk-dollar trade-
off. This is done using automobile accident risks and cost of
living in Questionnaire C-2, and the results of this exploration
are reported in Table 8.
The median response of $2,286,000 is quite reasonable in
view of the similar (in 1987 dollars) market-based estimate by
Blomquist (1979), but the mean value of $8,184,000 seems rather
large. The high mean estimate was generated by a portion of the
sample with value of life estimates as high as $80, 000, 000. Such
implausibly large estimates can occur because of the difficulty
of the comparison task. Respondents are being asked to establish
an equivalence between some annual chance of chronic bronchitis
x/100,000 that is equivalent to an $80 cost-of-living increase.
This is a difficult comparison to make. in contrast, the risk-
risk questions focused on chronic bronchitis--auto accident risk
comparisons of x/100,000 and y/100,000, where most respondents
did not believe that the severity of outcomes differed by more
than an order of magnitude.
The implicit dollar value of CB can be obtained by chaining
the responses to questionnaire part C-l, which gives the CB-auto
death trade-off, and part C-2, which gives the auto death--
dollars trade-off. These results appear in Table 9. The median
dollar value of each chronic bronchitis case is $800,000. The
mean is much greater because there is one outlier with a $320
million value. This individual expressed extreme responses on

-------
34
Table 8
Distribution of Auto Accident --
Cost of Living Trade-Offs
Dollar Value per	Implicit Dollar
1/100,000 Reduced	Value of
Decile	Risk of an Accident	an Accident
. 10
10 . 00
$1,000,000
.20
17 .50
$1,750,000
.30
17 .50
$1,750,000
.40
20 . 00
$2,000,000
.50 (median)
22.86
$2,286,000
. 60
26. 67
$2,667,000
.70
40 . 00
$4,000,000
.80
80 . 00
$8,000,000
.90
177 .78
$17,778,000
1. 0
800.00
$80,000,000
Mean
81.84
$8, 184, 000
(St. error of mean)
(14.40)
($1,440,000)

-------
35
Table 9
Implicit Valuation of Chronic Bronchitis
Implied by CB--Auto Death and Auto Death --
Cost of Living Trade-offs
Questionnaire C
Fracti 1e s	Inferred CB Value
.10	$200,000
.20	$350,000
.30	$522,449
.40	$646, 154
.50	$800,000
.60	$1,066,667
.70	$2,133,333
.80	$3,555,556
.90	$12,800,000
.99	$71,111,111
1.00	$320,000,000
Mean	$6,962,364
(Std. Error of Mean)	($2,977,373)
(N = 112)

-------
36
each component part, valuing each CB case at four times the
amount of each death and having an implicit value of an auto
fatality of $80 million. In each case, these were the highest
values in the sample and the highest permitted by the Program,
which indicates that this individual probably did not understand
the valuation task.
As instructive summary of the results is provided in Table
10. For the results creating CB/auto death risk equivalents, the
numbers have been transformed into implicit value-of-life terms
using three different reference points: a $2 million value of
life; a $3 million value of life; and a $5 million value of life.
The $2 million figure is comparable to the median auto death risk
valuation within the survey so that this estimate provides an
internal comparison of the results. The $3 million figure is
included since the recent estimates by Moore and Viscusi (1988)
indicate that the labor market value of life is in the $2-$3
million range using BLS data, and this was the "best estimate" of
the value of life in earlier work by Viscusi (1983) . The $5
million reference point is the value of life figure obtained
using new NIOSH data on job fatality risks, which Moore and
Viscusi (1988) view to be superior to the BLS data.
The pattern displayed by the results is fairly similar. In
each case mean valuations are at least double the value of the
median. Although one would not expect symmetry in a distribution
truncated at zero, the very high end responses observed appear to
be due to response errors.

-------
Table 10
Summary of Risk-Dollar Equivalents
Direct
Valuation
Estimate
CB Estimate
Using
$2 Million
CB Dollar
E s timate
Using
$3 Million
CB Dollar
Estimate
Using
$5 Million
Value of Life Value of Life Value of Life
CB/Auto Fatality:
A-1 & C-l (Median)
A-1 & C-l (Mean)
$640,000
$1,360,000
$960,000
$2,040,000
$1,600,000
$3,400,000
CB/Cost of Living:
A-2 (Median)	$457,000
A-2 (Mean)	$883,000
CB/Storm Damage:
B-l (Median)
B-l (Mean)
$800,000
$1,705,200
CB/Dollars (Derived from CB/Auto Fatality and Auto/Cost of Living)
C-l & C-2
C-l & C-2
$800,000
$6,962,364
Auto/Cost of Living:
C-2 (Median)
C-2 (Mean)
$2,286,000
$8,184,000

-------
38
The most clearcut divergence from plausible patterns is the
mean value of life of $8, 184, 000 for the auto death\cost-of-
living trade-off. Whereas the mean CB/auto values were roughly
double the median, the mean auto/cost of living values were
almost four times the size of the median, indicating a much more
skewed distribution. As noted in the discussion of Table 8, this
mean value was influenced in part by individuals with implied
values of life as high as $80 million. These outliers suggest
that for some People making meaningful trade-offs involving small
cost-of-living differences and low risks of auto accident
fatalities is a task they cannot handle effectively.
The valuation of chronic morbidity across the difference
questionnaire approaches is quite similar for the case in which
we use a $2 million value of life figure to transform the death
equivalent statistics into meaningful dollar estimates. The
median value for the CB/auto death risk trade-offs is $640,000,
as compared with a median value of $457, 000 for the CB/cost of
living trade-off and a median value of $800,000 for the CB/storm
damage results. These results are similar to the $800,000 median
CB value that was obtained by chaining the CB/auto and auto/cost
of living responses. Even with a higher value of life of $3
million, the CB/auto median of $960, 000 is not out of line with
the CB/cost of living and CB/storm damage results.
Once we move to the case where a $5 million value of life is
used, the median dollar valuation of each CB case prevented is
greatly increased to the $1, 600, 000. if EPA were to rely on, for
example, the CB/cost of living results to value CB and then use a

-------
39
value of life of $5 million without also using an appropriately
adjusted CB value, this procedure could potentially understate
the value of the CB cases prevented by a factor of three. By
converting all outcomes to a health risk equivalence scale using
a death risk metric, EPA avoids any distortion in the mix of
targeted illnesses that might otherwise occur if the value of
life number selected was incorrect.
6. Conclusion
Although market evidence remains our most reliable guideline
for assessing the shape of individual preferences, such evidence
is unavailable for many outcomes that are either not traded
explicitly in markets or traded implicitly but in a market for
which available data are not rich enough to identify the
pertinent trade-off rates. Analysis of risk-risk and risk-dollar
trade-offs using various types of simulated market choices
provides a useful mechanism for establishing such values.
This study has developed a methodology for deriving
morbidity valuation estimates based on the trade-off with another
well-known risk, rather than forcing individuals to express
trade-off rates between morbidity rate reductions and dollars, a
task which is unfamiliar to most people. We presented several
conceptual reasons why consumers should be able to more
accurately convey risk-risk trade-offs than risk-dollar trade-
offs, and the application of our methodology to the valuation of
reductions in the risk of chronic bronchitis indicate that most
individuals can make risk-risk trade-offs, even with a disease as

-------
complicated and unfamiliar to healthy people as chronic
bronchitis.
Although for the purpose of cost-effectiveness analysis
there is no need to measure risk reduction value in terms of
dollars, when we translated our risk-risk estimates into risk-
dollar estimates using either survey results on auto accident
risk reduction values or published value-of-life estimates, the
distributions compared favorably, thus providing additional
confidence in the reasonableness of the results derived from our
methodology. While this study applied the approach to the
valuation of only two risks, that of chronic bronchitis and an
auto accident fatalities, the favorable results suggest that the
methodology may be more widely applicable to other morbidity
risks, such as various forms of cancer.

-------
41
Footnotes
¦'•See Viscusi (1986) for a review of the market trade-off
literature.
See analysis by Blomquist (1979) for an inventive use of
seatbelt usage data to infer a value of life.
3
Survey studies of various health and environmental risks
include the seminal work by Acton (1973) as well as more recent
studies often grouped under the designation "contingent
valuation. " These recent analyses include: Brookshire, Thayer,
Schulze, and d'Arge (1982) ; Cummings, Brookshire, and Schulze
(1986);	Fischhoff and Furby (1988); Gerking, de Haan, and Schulz
(1988); Smith and Desvousges (1987); Viscusi and Magat (1987);
Viscusi, Magat, and Forrest (1988); and Viscusi, Magat, and Hube
(1987)	; and Fisher, Chestnut, and Violette (1989) .
^ In designing our survey, we used software from Sawtooth
Software, Inc.
5
For an important recent study of the valuation of health
risks rather than mortality, see Berger et al. (1987) .
®For example, see Viscusi, Magat and Huber (1987), pages
477-478 .
7
See Petty (1985) for a discussion of the distinction
between chronic bronchitis, the related disease emphysema, and
the broader disease category called chronic obstructive pulmona
disease. The authors selected the type of chronic bronchitis
described in Table 1 after consulting closely with two lung
specialists at Duke University Medical Center and visiting the
Medical Center rehabilitation program for patients with severe
lung diseases.
Q
At the end of the interview, subjects were carefully
debriefed about this use of average rather than person-specific
risks.
9
Our past studies suggest that presenting the risk in terms,
of the number of cases for a large base population is more
comprehensible than giving risk levels such as 0.00075.
10See Arrow (1971).
i:LSee Viscusi and Magat (1987) and Viscusi, Magat, and Hube
(1987) .

-------
42
12
Problt analysis was used to identify personal
characteristics that explain the division of subjects between
those giving consistent and inconsistent responses. The only two
significant variables in the equation are AGE and SMOKER, with
older respondents less likely to give consistent responses and
smokers more likely to respond consistently. These results may
reflect the difficulty that older subjects have with the new
interview technology (computers) and the greater thought that
smokers have given to the implications of chronic bronchitis.

-------
43
References
Acton, Jan, Evaluating Programs to Save Lives: The Case of Heart
Attacks (Santa Monica: Rand Corp., 1973) .
Arrow, Kenneth J. F.ssavs in the Theory of Risk-Rearing (Chicago:
Markham, 1971) .
Berger, Marck C., Glenn C. Blomquist, Don Kenkel, and George S.
Tolley, "Valuing Changes in Health Risks: A Comparison of
Alternative Measures," Southern Economic Journal, Vol. 53
No. 4 (1987), pp. 967-984.
Blomquist, Glenn, "Value of Live Saving: Implications of
Consumption Activity, " Journal of Political Economy, Vol. 87
(1979), pp. 540-558.
Brookshire, David, Mark Thayer, William Schulze, and Ralph
d'Arge, "Valuing Public Goods: A Comparison of Survey and
Hedonic Approaches, " American Economic Review, Vol. 72
(1982), pp. 165-177.
Cummings, Ronald, David Brookshire, and William Schulze, Valuing
Environmental	Goods :		Assessment	of	£il£	Conti nent
Valuation Method (Totowa, N.J.: Rowman and Allanheld,
1986) .
Fischhoff, Baruch, and Lita Furby, "Measuring Values: A
Conceptual Framework for Interpreting Transactions with
Special Reference to Contingent Valuation of Visibility,"
Journal of Risk and Uncertainty, Vol. 1, No. 2 (1988), in
press.

-------
Fisher, Ann, Lauraine Chestnut, and Daniel Violette, "The Value
of Reducing Risks of Death: A Note on New Evidence,"
Journal of Policy Analysis and Management. Vol. 8, No.
1 (1989), pp. 88-100.
Gerking, Shelby, Menno de Haan, and William Schulze, "The
Marginal Value of Job Safety: A Contingent Valuation Study,"
Journal of Risk and Uncertainty, Vol. 1, No. 2 (1988), in
press.
Moore, Michael, and W. Kip Viscusi, "Doubling the Estimated Value
of Life: The Implications of New Occupational Fatality
Data," Journal of Policy Analysis and Management. Vol. 7,
NO. 3 (1988), pp. 476-490.
	, and 	( "The Quantity-Adjusted Value
of Life," Economic Inquiry. vol. XXVI, No. 3 (1988), pp.
369-380.
Petty, Thomas L., Chronic Obstructive Pulmonary Pi sease (New A
York: Marcel Dekker, 1985).
Smith, V. Kerry, and William Desvousges, "An Empirical Analysis
of the Economic Value of Risk Changes,tt Journal of Political
Economy. Vol. 95 (1987), pp. 89-114.
Viscusi, W. Kip. Risk bv Choice: Regulating Health and Safety in
the Workplace (Cambridge: Harvard University Press. 19831,
	, "The Valuation of Risks to Life and Health:
Guidelines for Policy Analysis," in Bentkover, et. al. ,
Benefit Assessment: The State of the Art (Dordrecht,
Holland: Reidel Publishers, 1986), pp. 193-210.

-------
45
Viscusi, W. Kip and Wesley A. Magat, Learning about Risk:
Consumer and Worker Responses to Hazard Information
(Cambridge: Harvard University Press, 1987).
	f Wesley A. Magat and Anne Forrest, "Altruistic
and Private Valuations of Risk Reduction," Journal of Policy
Anal vsi s	and	Management. Vol. 7, No. 2 (1988), pp. 227-245.
, and Joel Huber, "An
	 I 	 '	'
Investigation of the Rationality of Consumer Valuations °f
Multiple Health Risks, " Rand Journal of Economics", Vol. 18,
No. 4 (1987), pp. 465-479.

-------
THE SOCIAL COSTS OF CHRONIC HEART
AND LUNG DISEASE
by
Maureen L Cropper
University of Maryland
and
Alan J. Krupnick
Resources for the Future

-------
THE SOCIAL COSTS OF CHRONIC HEART AND LUNG DISEASE
Maureen L. Cropper and Alan J. Krupnick
University of Maryland and Resources for the Future
INTRODUCTION
To value a program that reduces the probability of contracting a
chronic disease, one would like to know what a person who does not have the
disease would be willing to pay to reduce his probability of getting it.
The sum across individuals of these willingnesses to pay, plus the expected
costs of the disease that are not borne by these individuals, comprise the
theoretically correct measure of social benefits from reducing incidence of
the disease.
In this paper we measure the medical costs and lost productivity
associated with various chronic heart and lung diseases. Our justification
for focusing on these components of the social cost of illness is that
medical costs and lost earnings are often not borne by individuals
themselves and, hence, are unlikely to be reflected in willingness to pay
figures. Therefore, they must be added to willingness to pay estimates to
compute the total benefits of reducing the incidence of a disease.
Effects on Earnings
Our estimates of the effects of chronic illness on labor force
participation and on earnings differ in two respects from those available
in the literature (Bartel and Taubman, 1979; Salkever, 1985). First, our
dataset—the Social Security Survey of Disabled and Non-Disabled Adults-
-allows us to distinguish the effects of individual diseases (e.g.,
emphysema, chronic bronchitis) rather than disease categories (chronic
respiratory illness). As one might expect, there is significant
variation in the effects of individual diseases within broader categories:
Emphysema, for example, has a large negative effect on earnings whereas
chronic bronchitis does not. Hypertension has no significant effects on
1. The diseases studied are: allergies, asthma, chronic bronchitis,
emphysema, other chronic lung disease, arteriosclerosis, heart attack,
hypertension, other chronic heart disease and stroke.

-------
2
probability of participation or on earnings, whereas a heart attack
occurring between 45 and 54 reduces both.
Second, we examine how the effect of each disease varies with age of
onset and duration. It is generally believed (Bartel and Taubman, 1979)
that, other things equal, a person is more likely to participate in the
labor force at any age the earlier in life he contracts a chronic disease.
The argument is that the benefits of making adjustments to the disease
(retraining, changing occupations) are larger the earlier in life the
disease begins. Thus, the earlier the age of onset the more likely it is
that adjustments will be made. It is not, however, clear that the human
capital argument applies to the diseases examined here, most of which are
contracted later in life. Since one seldom witnesses changes in occupation
after age 45 it is unlikely that small variations in age of onset matter
after this age. Indeed, age of onset may have a positive effect on
participation if a disease is more serious when contracted at an earlier
age.
It is also of interest to see how the duration of a disease alters
labor market behavior. For two persons who contracted emphysema at age 45,
are effects on earnings greater for a person currently 50 or for a person
currently 60? Holding age of onset constant, this is equivalent to asking
whether the disease has a greater effect on participation and earnings when
one has had the disease for five years or for fifteen years. One might
hypothesize that the longer one has had a disease the longer he has had to
adjust to it; hence, labor market effects should diminish with duration.
On the other hand, for progressive diseases, e.g., emphysema, the longer
one has had the disease the more serious it is likely to be.
We find that the tendency of chronic disease to reduce labor force
participation and earnings does not increase with age of onset. Indeed,
for emphysema, heart attack, arteriosclerosis and stroke, an age of onset
between 45 and 54 significantly reduces the probability of working at all
future ages, but an age of onset between 55 and 65 does not. It might seem
that this result occurs because people who contract a disease earlier will,
on average, have had it for a longer time than persons who contracted it

-------
3
later in life. For emphysema this appears to be true. When duration is
held constant, it is having the disease for 6 or more years that affects
labor market behavior rather than contracting it at age 45. For heart
attack, arteriosclerosis and stroke, however, the duration of the disease,
holding age of onset constant, has no effect on participation.
Medical Costs
Our estimates of medical costs, which come from the National Medical
Care Expenditure Survey (NMCES), have two advantages over existing
estimates of medical expenditures (National Heart, Lung and Blood
Institute, 1982; Hartunian et al., 1981). The National Heart, Lung and
Blood Institute allocates aggregate costs, such as hospital costs and
doctor costs to diseases based solely on a disease's proportion of total
activities, e.g., hospital days and total doctor visits, respectively.
This approach has two shortcomings: (1) it assumes that the average cost
of, say, a hospital day or doctor's visit is the same for all diseases, and
(2) it does not allow one to examine the distribution of medical costs per
person. An alternative "engineering" approach is to multiply the number of
hospital days or doctor visits attributable to a condition by the typical
price for a hospital day or typical price for a doctor visit for that
condition (see e.g., Freeman (1976)). This approach circumvents the first
objection raised above but not the second.
By using individual data on medical costs, collected over a one-year
period for over 40,000 persons, we are able to examine the distribution of
medical costs per person by disease. Our most interesting results pertain
to the size distribution of medical costs. For the five diseases whose
medical costs we study—bronchitis, emphysema, hypertension, ischemic heart
disease and non-specific heart disease—the distribution of annual costs
per person is highly skewed. For emphysema, ischemic heart disease and
non-specific heart disease median expenditures are less than one-tenth of
mean expenditures. For bronchitis and hypertension median expenditures are
about one-fourth of mean expenditures.
Because NMCES contains information on source of payment, it is also
possible to see to what extent individuals and their families bear the

-------
4
medical costs of these diseases. For emphysema, ischemic heart disease and
non-specific heart disease only about 10% of aggregate medical costs are
borne by patients' families. The percentages are somewhat higher for
bronchitis (34%) and hypertension (23%). The percent of cost borne by the
patient's family differs, however, by size of cost. As noted above, the
majority of persons with the diseases studied here incur small annual
medical expenses. Averaging across individuals, the fraction of medical
costs paid for by one's family is 2/3 for hypertension and bronchitis and
half for emphysema, ischemic heart disease and non-specific heart disease.
This implies that, on average, individuals (or their families) pay a higher
fraction of small medical expenditures than of large ones.
THE EFFECT OF CHRONIC ILLNESS ON LABOR FORCE PARTICIPATION AND EARNINGS
The Model
In modelling the effects of various diseases on earnings it is
standard practice (Bartel and Taubman, 1979; Mitchell and Butler, 1986) to
distinguish the effects of each disease on participation from its effects
on earnings given that one participates. Debilitating diseases such as
emphysema and stroke may force a person to drop out of the labor force
because he is physically unable to work, or may reduce earnings to the
point where they fall below the reservation wage. If a person continues
working he may curtail hours (if free to do so) or suffer a drop in pay
because he changes jobs or because his productivity falls. This implies a
drop in earnings, conditional on working.
The decision to participate, and earnings, conditional on
participation, constitute a two-equation system. The individual
participates if the decision function, 1^, is non-zero. Earnings, Y^f
are observed only if the individual participates.
1^ - - et	Participation decision	(1)
Participate if 1^ > 0,
m	Earnings in labor market	(2)

-------
5
Yt observed if I > 0
Y not observed if It < 0.
Equation (1) can be viewed as a reduced-form equation that results
from comparing the utility received from income and leisure, conditional on
working, with the utility received from income and leisure given that the
individual does not work. If income and leisure in each state are replaced
by their exogenous determinants, one obtains equation (1).2
Because earnings in (2) are observed only for working persons,
estimation of (2) involves a classic selectivity problem: persons for whom
earnings data are available are in the lower tail of the error distribution
in equation (1). As long as the errors in equations (1) and (2) are
correlated, applying least squares to (2) results in inconsistent parameter
estimates since E(ut|Zt« > et) * 0.
To obtain consistent estimates of this system we follow the two-stage
approach outlined by Lee (1983) [see also Maddala (1983)]. We assume that
the error term in the participation equation has a logistic distribution
F(et) - l/[l+exp(-ZtS)], and estimate a logit model of labor force
participation. The error term can be transformed to an error term
e* with a standard normal distribution,
e* - J(et) - ~"1(P(et)),
-1
where ~ is the inverse of the standard normal distribution function.
Assuming that e* and u are bivariate normally distributed with
2
correlation coefficient p and V(u^) ¦ a , expected earnings are a linear
function of X plus a term that represents the density of e*
conditional on working,
E(Xt0+ ut|et
-------
6
Applying OLS to (3) yields consistent estimates of the parameters fi and
3
op.
The Data
The Sample. The data used to estimate our model come from the 1978
Social Security Survey of Disability and Work (U.S. Department of Health
and Human Services, Social Security Administration, 1981). The survey,
which was designed to examine issues relating to eligibility for disability
benefits and the effects of disabilities on labor force participation,
consists of two samples, a stratified random sample of 6,853 persons from
the 1976 Health Interview Survey, and a sample of 4,886 persons from the
population of recipients of Social Security Disability Insurance who were
declared eligible for benefits no earlier than 5 years before the survey.
Our sample consists of 2,218 men between the ages of 18 and 65 from the
4
Health Interview Survey portion of the Social Security Survey.
Earnings Equation. To avoid transitory fluctuations during the survey
week, earnings are measured as wages and salaries received from all jobs
during 1977. (All earnings are measured in 1977 dollars.) The independent
variables entering the earnings equation X , are listed in Table 1.
Earnings are assumed to depend on education (measured by a series of dummy
variables), experience (proxied by a series of age dummies), experience
squared, marital status, family size, race, locational dummies and the
health variables described below and in Table 2.
Labor Force Participation Equation. As with earnings, participation
is defined based on behavior throughout the 1977 calendar year. An
individual is considered to have been in the labor force if he worked 30 or
more weeks during the 1977. Men who did not work at all during 1977 are
classified as not participating in the labor force. Men working between
3. The two-stage estimation procedure, including asymptotic standard
errors (Maddala, 1983), was programmed by the authors using the SAS
matrix language.
4. There are a total of 2,626 men between 18 and 65 in the HIS portion of
the Social Security survey. 408 of them were eliminated because they
appeared to change labor force status during 1977, the year for which
participation and earnings were measured.
-------
7
one and 29 weeks were eliminated from the sample on the grounds that these
persons were either students or changed labor force status.
Since the decision to participate in the labor force is made by
comparing the utility of income and leisure when in the labor force with
income and leisure when out of the labor force, the variables in Z^
should include all those entering the earnings equation, plus variables
that would affect income conditional on not participating, and variables
that would affect the utility of leisure time. The only such variables
available in the survey that are not included in are (1) whether the
individual is aware of Social Security disability benefits and (2) whether
the individual is a veteran, both of which might affect income received if
the individual did not participate. A third variable included in Z^ to
capture motives for working is the size of the respondent's debt.
Health Variables. The survey contains two types of information about
chronic illness. Respondents were asked whether they had ever been
diagnosed by a doctor as having any one of the 35 chronic diseases listed
in Table 2, as well as when the disease first began to bother them (age of
onset) . They were also asked whether they were functionally limited by any
of the diseases. Functional limitation questions include whether the
respondent had difficulty walking, climbing stairs, lifting heavy objects,
etc. Respondents were also asked whether they experienced symptoms such as
pain, fatigue, swelling and shortness of breath.
In both the earnings and participation equations the severity of
chronic disease is measured by dummy variables that indicate the presence
of a chronic condition. Pleasures of functional limitation, while possibly
useful as indicators of the severity of disease, are not associated with
specific diseases and, hence, cannot be used to measure the severity of
individual diseases.^
5. In addition to collecting these measures of functional limitation, the
survey also asks respondents if they "have a disability that limits the
type or amount of work [they] can do?" This variable, which is
included in addition to the chronic disease dummies in Mitchell and
Butler's (1986) analysis of the labor market effects of arthritis, was
excluded from our analysis for two reasons. First, the answer to this
question is not an exogenous measure of health but reflects the
Footnote 5 continued on next page
-------
8
In measuring the effect of particular diseases on participation and on
earnings we would like to distinguish effects by age of onset and by
duration of the disease. The extent to which this is possible depends on
the disease studied. Table 3 gives the distribution of age of onset for
persons in our sample for each of the 10 respiratory and circulatory
diseases studied. In our sample few cases of emphysema, arteriosclerosis,
or stroke occur before age 45. For this reason these diseases are
represented by only two age of onset dummies indicating that the disease
was contracted between the ages of 45 and 54 or between the ages of 55 and
65.
Chronic bronchitis and other chronic lung disease occur earlier in
life than emphysema; however, the small numbers of persons in our sample
with these conditions restrict us to only two age of onset categories for"
each disease: before age 45 and after age 45. Allergies, asthma, heart
attack, hypertension, and other chronic heart disease occur frequently
enough and early enough in life that we can distinguish between 3 and 5 age
of onset categories for each disease, as indicated in Table 2.
We have attempted to distinguish between duration of disease and age
of onset only for those diseases that appeared to have a significant effect
6
on labor force participation when age of onset alone was measured. These
included emphysema, arteriosclerosis, heart attack, stroke and other heart
disease. Each disease was significant only when age of onset was 45 or
older. The fact that these diseases occur later in life, together with a
maximum sample age of 65, means that we can distinguish only two duration
categories: persons who have had the disease 0-5 years and persons who have
had the disease 5-10 years.^
Footnote 5 continued from previous page
decision to stop/continue working. Second, the variable may capture
effects of multiple diseases that we wish to capture using disease-
specific dummies.
6. Throughout the paper "statistically significant" means significant at
the 5% level, one-tailed test.
7. Chronic bronchitis beginning between ages 25 and 44 significantly
decreased the probability of labor force participation; however, there
were too few persons who had had chronic bronchitis for more than 10
years to permit using additional duration dummies for this disease.
-------
9
Results
Labor Force Participation. The more serious respiratory and
circulatory diseases examined—chronic bronchitis and emphysema;
arteriosclerosis, heart attack, stroke and other heart disease-
-significantly reduce the probability that a man participates in the labor
force, other things equal. Table 4 presents coefficients obtained from the
logistic participation equation for the respiratory and circulatory disease
variables listed in Table 2. [The coefficients of other variables in the
participation equation appear in the appendix to this paper.] The table
indicates that the less serious diseases—allergies, asthma, other chronic
lung disease and hypertension—have no significant effects on
participation. To calculate the effect of each disease on probability of
participation its coefficient must be multiplied by P(l-P), where P is
the probability of participation. Since P = 0.670 for our sample, the
coefficients in Table 4 imply that contracting emphysema between ages 45
and 54 reduces the probability of participating in the labor force by an
average of 23.3 percentage points. Arteriosclerosis reduces probability
of participation by 15.6 percent, while having a stroke between 45 and 54
reduces subsequent probability of participation by 57.3 percent.
What is somewhat surprising is the effect of age of onset on
participation. For emphysema, arteriosclerosis, heart attack and stroke,
an age of onset between 45 and 54 significantly reduces probability of
working at all future ages, but an age of onset between 55 and 65 does not.
Such a result runs counter to the standard argument that, the earlier the
onset of a disability, the more likely it is that the individual will
adjust to it by retraining and/or switching jobs. One reason that the
standard argument may not apply is that, for the diseases studied here, a
diagnosis at age 45 may indicate a more severe case of the disease than a
diagnosis at age 60 (a heart attack at age 45 is often more devastating
than a heart attack at age 60).
A second possibility is that for progressive diseases such as
emphysema and arteriosclerosis, persons who contract the disease earlier
will, on average, have had it for a longer time than persons who contract
it later in life. To the extent that severity increases with the duration
-------
10
of the disease, persons who have had the disease longer will be less likely
to work. The results in Table 4 may thus be due to the fact that age of
onset is directly correlated with the number of years the individual has
been bothered by the disease.
To test this hypothesis the age of onset categories in Table 2 were
subdivided to distinguish duration of disease from age of onset. Persons
with an age of onset between 45 and 54 were divided into two categories:
those who had had the disease for 0-5 years and those who had had the
disease for 6-10 years. For persons with an age of onset between 55 and 65
9
only the 0-5 year duration category was used.
The estimated coefficients of the age of onset/duration dummy
variables appear in Table 5. These coefficients suggest that controlling
for duration alters the effect of age of onset only in the case of
emphysema. For emphysema, when duration is held constant at 0-5 years, age
of onset has no effect on participation. Having the disease for 6-10
years, however, significantly reduces the probability of participation. In
the case of arteriosclerosis, heart attack and stroke, however, the main
effect on labor force participation is caused by age of onset, with onset
between 45 and 54 making participation less likely, and onset between 55
and 65 having no significant effect. These results suggest that the effect
of age of onset and duration are, in general, disease-specific.
Earnings. The results for our earnings equations suggest that, for
the respiratory and coronary diseases studied here, most labor market
effects occur through reductions in participation rather than reductions in
earnings. Table 6 presents coefficients of the disease dummies in an
earnings equation in which diseases are distinguished by age of onset and,
8. One could, of course, argue that persons with very severe cases of the
disease die soon after diagnosis; hence duration may not measure
severity.
9. Persons with an age of onset between 55 and 65 with duration greater
than 5 years thus had a value of zero for all health dummies, as did
persons without the disease.
-------
11
in the case of emphysema, by duration. The only respiratory and
circulatory diseases studied that significantly reduce earnings are
emphysema and heart attack. Having emphysema for 6-10 years reduces
earnings by 65%. Having a heart attack between the ages of 45 and 54
reduces earnings by 45%.
The Magnitude of Expected Earnings Losses. The expected loss in
earnings to a person who contracts a chronic disease is the sum of the
effects of the disease on probability of participation, and on earnings,
given that one participates. Specifically, the expected loss in earnings
is the sum of the change in probability of participation times pre-illness
earnings, plus the reduction in earnings caused by the disease times the
post-illness participation rate,
Expected Loss in Earnings * AP(EarningSQ) + P^( AEarnings). (4)
This loss begins at age of onset and continues until the age that
retirement would occur in the absence of the disease.
Tables 7 and 8 present estimates of the first term in (4), expected
earnings losses due to non-participation. The effect of each disease on
probability of participation, AP, is determined by multiplying the
coefficient of the disease in the participation equation, 6^, by P(l-P),
where P is the probability of being in the labor force. Table 7 presents
estimates of fiP, the fraction by which pre-illness earnings are reduced due
to non-participation. In the tablet P is estimated at each age from
Bureau of Labor Statistics data on labor force participation rates (U.S.
Department of Labor, Bureau of Labor Statistics, 1988) . In Table 8 AP
has been multiplied by average 1987 earnings of all male workers to produce
annual earnings losses, by age, due to non-participation.
In both tables earnings losses due to increased probability of not
working peak between 55 and 65, because P(l-P) is maximized in this
10. Because fewer chronically ill people appear in the earnings equation
than in the participation equation it was necessary to eliminate
certain age of onset categories from the earnings equations.
-------
12
interval. The maximum annual expected reduction in earnings ranges from
15.5% for heart attacks to 57.1% for strokes. Bronchitis and emphysema
each reduce expected earnings (through effects on participation) by at most
25%.
For emphysema, arteriosclerosis, stroke and other heart disease
earnings losses due to reduced probability of participation constitute the
total change in expected earnings. For emphysema and heart attack the
second term in equation (4) must be computed. This term, in $1977, appears
in Table 8 together with expected earnings losses due to non-participation.
Comparison with Previous Work. The only study of the labor market
effects of chronic respiratory and circulatory diseases of which we are
aware is Bartel and Taubman (1979). Using data from the NAS Twins Panel,
Bartel and Taubman examine the effects of each of several disease groups on
labor force participation and on earnings, conditional on participation.
Unfortunately the diseases groupings used by Bartel and Taubman do not
correspond exactly to the diseases used in our study. They combine
bronchitis, emphysema and asthma into a single disease category (BRON) , and
heart disease and hypertension into another category (HH). The effect of
each disease category, is examined for various ages of onset; however,
emphasis is placed on diagnoses that occurred between 1962-67, when
respondents were in their early forties. Because emphysema,
arteriosclerosis and stroke are rare at this age, it is unlikely that BRON
and HE capture these more severe diseases.
When they examine the effects of a diagnosis at age 40 on
participation at age 50 Bartel and Taubman do not find any significant
effects of respiratory or circulatory diseases on labor force
participation. This is in sharp contrast to the results presented in Table
7, which indicate that chronic bronchitis, emphysema, arteriosclerosis,
heart attack, stroke, and other heart disease reduce the probability of
labor force participation between 6 and 57 percentage points. The
difference in findings may be due in part to the relatively young age of
their sample. The disease variable used in the participation equation
represents the effects on participation at (mean) age 50 of a diagnosis
-------
13
that occurred at (mean) age 40. For the diseases we study the most
significant effects on participation correspond to an average age of onset
of 50.
Regarding effects on earnings, Bartel and Taubman find that a
diagnosis of respiratory illness (BRON) at age 40 reduces earnings by 25%
at age 50 and that heart disease/hypertension (HH) , diagnosed at age 40,
reduces earnings by 8.5% at age 50. By contrast, we find that having
emphysema for at least 6 years reduces earnings by an average of 65% for
persons who continue working. The corresponding reduction in earnings due
to having a heart attack between 45 and 54 is 45%. We thus find greater
effects on earnings than do Bartel and Taubman, but for more narrowly
defined diseases. The difference between our results and theirs reflects
the fact that their disease categories include less severe diseases, such
as bronchitis and hypertension, as well as more debilitating ones.
MEDICAL EXPENDITURES AND SERVICES UTILIZATION
The medical costs of a chronic disease to society are the costs of the
detection, treatment, and rehabilitation of the disease, as well as a
portion of research, training, and facilities costs. In this section we
present measures of medical expenditures for individuals for five target
diseases: hypertension, ischemic heart disease, non-specific heart disease,
chronic bronchitis, and emphysema. These measures were computed from self-
provided cost of treatment data for persons in the 1977-78 National Medical
Care Expenditure Survey (National Center for Health Services Research,
1981) .
There are three reasons why our measures of medical expenditures do not
measure the true social costs of medical treatment. First, medical
expenditures are computed using market prices, which may not reflect
marginal productivities due to the absence of competition in the market for
medical services. Second, because the data are specific to individuals
with chronic diseases, the costs of detection are not included. In
addition, because medical care providers are a minor source of research and
medical training, these cost components are likely to be greatly
underestimated (if included in overhead charges) or ignored completely.
-------
The National Medical Care Expenditure Survey
To estimate the medical costs of chronic respiratory and heart disease
we used the 1977-78 National Medical Care Expenditure Survey (NMCES).
NMCES presents data on health care utilization and expenditures for
approximately a one year period for 14,000 households (40,320 persons)
selected randomly from the civilian noninstitutionalized U.S. population.
Each of these households was provided with a calendar diary for recording
their use and cost of medical services. Each was interviewed six times
over this period, with responses in prior periods provided to the household
for verification.
Each time a person in the NMCES suffered an activity limitation,
disability day, visited or called a doctor, vent to the hospital or
purchased medication a record was created for an illness episode.
Information on the number and cost of illness episodes and on the cause of
each illness episode comes from the household survey. Medical costs are
thus self-reported costs. ^ The diseases associated with each illness
episode were reported by households, and translated into ICDA codes by
interviewers.
The five respiratory and circulatory diseases we examine, their ICDA
codes, and the number of persons reporting episodes involving each
condition appear in Table 9.
Allocation of Medical Costs Among Multiple Conditions
To calculate the costs associated with a target condition one must add
the costs associated with the condition across all illness episodes. This
would pose no problem if all episodes of illness were associated with only
a single disease. If, however, an illness episode is associated with more
11. To check on the accuracy of these costs, the household survey was
supplemented by a survey of physicians and facilities that provided
medical care to persons in the household sample period and by a survey
of employers and insurance companies responsible for the health
insurance coverage of responding households. A close correspondence
was found between reported and actual costs.
-------
15
than one condition, the cost of the episode must be allocated among
conditions.
Table 10 indicates the extent of the joint cost allocation problem.
The table indicates that of the 3,479 persons with at least one episode of
hypertension, 71% (2,476) had episodes that involved hypertension alone.
[In the language of NMCES an episode involving only a single condition is a
"simple" episode.] For these persons the problem of cost attribution does
not arise. Thirteen percent of persons (426 persons) with hypertension
episodes have "related to" episodes—episodes that involve hypertension and
some other condition. In these cases the respondent attempted to allocate
costs among the related conditions; however, in cases where no attribution
was possible, for example, the case of hospital room charges, the costs
were duplicated for each condition. "Same as" episodes, involving 7% of
all persons with hypertension, mean that the individual attributed the
episode to hypertension and a condition that was the "same as"
hypertension—although it was assigned a different ICDA code. In this case
no allocation of costs among the multiple conditions is possible; instead,
the total costs of the episode are associated with each condition. "Same
as " episodes thus lead to double counting of medical costs, and "related
to" episodes may involve some double counting.
The number of persons with "multiple episodes" are found by subtracting
those with 'single episodes from the total (e.g., for hypertension, 314
persons had multiple episodes). In general, persons with more than one
episode involving the same disease have other than "simple" episodes that
may involve double-counting problems.
Results
Magnitude of Expenses, by Disease. Table 11 shows the frequency
distribution of annual medical expenses for each of our target diseases, as
well as mean and median expenses. [All figures are in 1977 dollars. ] As
one would expect, the highest average expenditures are associated with
ischemic heart disease ($1256) and non-specific heart disease ($1041).
Emphysema is associated with a mean expenditures of $633. The average
-------
16
annual costs of hypertension and bronchitis are considerably less: $216 and
$97, respectively.
In each case the distribution of annual expenses is highly skewed:
median expenses are one-quarter of mean expenses for bronchitis and
hypertension and approximately one-tenth of mean expenditures for
emphysema, ischemic heart disease and non-specific heart disease. For all
diseases but ischemic heart disease at least half of all persons have
annual expenditures of $75 or less. [For ischemic heart disease 41% of all
persons have annual expenditures of $75 or less.]
Categories of Expenses. Table 12 shows how expenditures are
distributed across categories for each disease. NMCES allocates expenses
to three major categories: medical contacts (primarily doctor visits),
hospital expenses, and drugs. There are several minor categories that are
omitted from the table.
As would be expected, hospital expenses are the largest category of
expenses for all conditions, even when people with no hospital expenses are
included in the averaging computation. The maximum hospital expenses per
person exceed $20,000 for the heart diseases and are in the $10,000 range
for the other target diseases. Expenses on medical contacts are the next
largest category of expenses for all conditions.
Comparison With Other Studies. The NHLBI (1982) estimates annual
expenditures on chronic bronchitis and emphysema using the "top-down"
approach described above while Freeman et al. (1976) use an engineering
approach with aggregate data to estimate annual expenditures on emphysema.
Table 13 provides the NHLBI and Freeman estimates of total and per person
expenditures adjusted to 1977 dollars using the medical price index.
The NHLBI estimates of expenses per case, at $118 and $102 for chronic
bronchitis and emphysema, respectively, contrast sharply with ours, at $97
and $633. Nevertheless, because of the top-down nature of the NHLBI
approach, their estimates may differ from ours if different estimates of
disease prevalence are being used. In fact, the NHLBI prevalence estimates
-------
17
for these diseases (which are taken from the Health Interview Survey (HIS))
are 3.5 and 1.0 percent of the civilian, noninstitutionalized population of
the U.S. in 1979 (216 million people) for chronic bronchitis and emphysema,
respectively. Our estimates of prevalence, which are conditional on the
occurrence of some medical event (i.e., a restricted activity day, some
cost incurred, or some service used (including a phone call to the
doctor)), are far lower — 1.1 and 0.5 percent for chronic bronchitis and
emphysema, respectively, for 1977.
The underestimate of prevalence implied by this conditionality implies
that our sample would under-represent, relative to the NHLBI, people with
zero medical costs. This implies, in turn, that the NHLBI estimate of
expense per case should be lower than ours. Instead, the NHLBI estimate
for chronic bronchitis, the disease for which the highest proportion of
sufferers in our sample has zero costs, actually exceeds our estimate.
Freeman et al, using data on health care utilization and average prices
for 1970, estimate expenses on emphysema in 1977 dollars of $233.5 per case
annually. These estimates are over double those of the NHLBI but still are
far lower than ours.
Sources of Payment. NMCES provides information on five sources of
funding for medical expenses: family, medicaid, medicare, personal
insurance, and other. In addition to being of intrinsic interest,
information about sources of funding suggests the extent to which medical
costs are likely to be internalized in willingnesses to pay to avoid
disease. In theory, willingness to pay should take into account the
medical costs of the condition paid for by the family, but not those costs
borne by others. Thus, the portion of expenses paid by others should be
added to the bid as part of the social cost of each of the target
conditions.
Table 14 identifies these funding sources by condition for males 20
years of age and older, the group to which our labor market analysis
applies. For each disease the second row of the table gives the percent of
total costs paid for by each source. Even for hypertension and bronchitis,
-------
18
the least serious diseases studied, families pay a minority of total costs,
23% and 34%, respectively. For emphysema and the heart diseases families
pay less than 15% of total costs. What are the most important sources of
funding? Personal insurance is the most important source of funding for
ischemic heart disease (46 percent), reflecting the high proportion of
expenses for the hospitalization component and the high degree of coverage
afforded this type of expense by health insurance plans. The insurance
share for emphysema is large (28 percent) for much the same reason.
Coverage for non-specific heart disease, the condition with the least
family funding, is not dominated by insurance. Rather, because the
population with this condition tends to be older than that for ischemic
heart disease, the largest funding share comes from medicare (36 percent).
Finally, it is curious that medicaid funds less than one percent of
expenses for ischemic heart disease while funding from 7 to 17 percent of
the expenses for the other target conditions.
Although a minority of total medical costs are paid for directly by
patients and their families family funding is the q most important source of
payment for a majority of patients. This is because most patients incur
small expenses (see Table 11) and families bear a larger percent of small
expenses than of large expenses. For each disease the third row of Table
14 computes for each individual the percentages of funding received from
various sources and then averages these percentages across individuals for
each source. As can be seen, the average percentages for the family source
(in brackets) are much higher than the aggregate percentages for the family
source (in parentheses), the former ranging from 52 to 70 percent, while
the latter ranges from 13 to 36 percent. This difference implies that
relatively large numbers of people have episodes with small expenses that
they pay for themselves. This may reflect deductibility clauses, the
exclusion of drugs from coverage for some policies, or other factors.
Age Distribution of Expenses. To permit comparison of the labor
market effects of chronic respiratory and circulatory diseases with medical
costs, Table 15 presents average medical costs for males, by age. Mean
annual expenses appear generally to increase with age, up to the A60's or
-------
19
70's for bronchitis, emphysema and hypertension. Expenses for those with
heart disease (heart attacks), however, peak in the MO's.
A comparison of average medical expenses with the labor market effects
of each chronic disease (see Table 8) suggests that the labor market costs
of chronic respiratory and circulatory diseases are generally greater than
the medical costs. Exceptions to this result are hypertension, which has
no effect on labor force participation or on earnings, and heart disease
before the age of 45, which also appears to have no significant labor
market effects.
-------
20
Table 1. Non-health Variables Entering Earnings and Participation
Equations

Mean
Standard
deviation
Maximum
Minimum
Earnings, 1977*
14,362.
77045.
50000.
0
In labor force, 1977
0.670

1
0
Married3,
No. in household3
No. children < 5
No. children 5-18®
No. children > 18
0.718
3.294
0.190
0.670
0.184
0.45
1.732
0.512
1.174
0.482
1
15
5
8
3
0
1
0
0
0
Age dummies:
18-24
35-44
45-54
55-65
0.141
0.174
0.222
0.261
0.348
0.379
0.416
0.440
1
1
1
1
0
0
0
0
Highest educ. level:
Elementary school
High school
College
0.193
0.487
0.229
0.394
0.500
0.421
1
1
1
0
0
0
Non-white
0.124
0.330
1
0
Regional dummies3:
Northcentral
South
West
0.265
0.335
0.178
0.441
0.472
0.383
1
1
1
0
0
0
Lives in7Urban Area3
(Age-16)
0.679
888.25
0.467
730.23
1
2401
0
4
Veteran
0.452
0.498

0
Aware of disability
benefits
0.407
0.491
1
0
Debt3
2116.9
8858.00
200800 •
0.
~Average based on 1486 persons in labor force
Measured as of interview date
-------
21
Table 2. Health Variables in Earnings and Participation Equations
Each of the following variables assume a value of 1 if the respondent
contracted the disease at the age indicated and a value of 0 otherwise:
RESPIRATORY AND CIRCULATORY DISEASES
Age of Onset Categories (Sample Size)
Allergies
0-17
(35)
18-34
(37)
35-65
(18)
Asthma
0-17
(40)
18-34
(14)
35-65
(19)
Chronic Bronchitis
25-44
(18)
45-65
(21)

Emphysema
45-54
(49)
55-65
(23)

Other Chronic Lung Dis.
18-44
(17)
45-65
(26)

Arteriosclerosis
45-54
(55)
55-65
(24)

Heart Attack
25-44
(28)
45-54
(57)
55-65
(42)
Hypertension
25-34
(57)
35-44
(79)
45-54
(148
Other Chronic Heart Disease
0-34
(23)
35-44
(34)
45-54
(51)
Stroke
45-54
(17)
55-65
(20)

55-65 (66)
55-65 (22)
OTHER CHRONIC DISEASES
Sample Size
Arthritis or rheumatism
Other trouble with back or spine
Deformity of foot, leg, arm, hand
Nervous or emotional problems
Deformity of back or spine
Deafness
Stomach ulcer
Diabetes
Hernia or rupture
Difficulty reading (with glasses)
Kidney stones or kidney trouble
Other chronic stomach trouble
Tumor, cyst or growth
Hissing arms, hands or fingers
Gallbladder or liver trouble
Paralysis
Alcohol or drug problems
Cancer
Epileptic seizures
Mental illness
Blindness
Thyroid trouble or goiter
Hissing legs or feet
Tuberculosis
Multiple sclerosis
367
296
228
209
154
133
130
113
92
86
76
64
52
46
40
35
25
24
24
20
19
18
14
-------
22
Table 3. Distribution of Respiratory and Circulatory Diseases by Age of
Onset
Number of persons in sample with
age of onset
0-17 18-24 25-34 35-44 45-54 55-65
Allergies
35
18
19
10
4
4
Asthma
40
5
9
7
9
3
Chronic Bronchitis
15
2
13
5
15
6
Emphysema
0
1
4
3
49
23
Other Chronic Lung Diseases
1
4
7
6
20
6
Arteriosclerosis
0
0
7
11
55
24
Heart Attack
2
0
5
23
57
42
Hypertension
12
23
57
79
148
66
Other Chronic Heart Disease
18
5
10
34
51
22
Stroke
1
0
2
2
17
20
-------
23
Table 4. Effects of Chronic Diseases on Labor Force Participation
by Age of Onset
Age of
onset
Coefficient
t-Ratio
Asthma
0-17
18-34
35-65
00093
0.625
0.093
0.22
0.75
0.16
Allergies
0-17
18-34
35-65
-0.061
0.505
-0.565
0.13
0.95
0.91
Chronic Bronchitis
25-44
45-65
-1.229
-0.816
1.69
1.17
Emphysema
45-54
55-65
-1.053
-0.683
2.55
1.21
Other Chronic Lung Disease
18-44
45-65
-0.218
-0.528
0.29
0.95
Arteriosclerosis
45-54
55-65
-0.707
0.134
1.72
0.26
Hypertension
25-34
35-44
45-54
55-65
-0.435
-0.131
0.189
-0.112
1.16
0.38
0.78
0.34
Heart Attack
25-44
45-54
55-65
-0.463
-0.720
0.507
0.94
1.94
1.15
Stroke
45-54
55-65
-2.593
-1.530
2.38
1.41
Other Heart Disease
0-34
35-44
45-54
55-65
-0.393
-0.184
-0.896
-1.462
0.90
0.40
2.39
2.04
-------
24
Table 5. Effects of Chronic Diseases on Labor Force Participation
by Duration of Disease and Age of Onset
Duration Onset Coefficient t-Ratio
Asthma 0-17 0.017 0.04
18-34 0.780 0.92
35-65 0.029 0.05
Allergies 0-17 -0.040 0.09
18-34 0.542 1.02
35-65 -0.479 0.78
Chronic Bronchitis 25-44 -1.254 1.70
45-65 -1.013 1.46
Emphysema 0"5 45-54 -0.230 0.35
5-10 45-54 -1.299 2.04
0-5 55-65 -0.370 0.62
Other Chronic
Lung Diseases 18-44 -0.465 0.65
45-65 -0.670 1.19
Arteriosclerosis 0-5 45-54 -0.389 0.57
5-10 45-54 -0.252 0.41
0-5 55-65 0.659 1.11
Hypertension 25-34 -0.418 1.12
35-44 -0.151 0.44
45-54 0.084 0.35
55-65 -0.088 0.27
Heart Attack 25-44 -0.449 0.91
0-5 45-54 -1.003 1.70
5-10 45-54 -1.069 1.85
0-5 55-65 0.371 0.79
Stroke 0-5 45-54 -1.503 1.25
5-10 45-54 -7.551 0.38
0-5 55-65 -0.900 1.06
Other Heart Disease 0-34 -0.352 0.81
35-44 -0.165 0.36
0-5 45-54 -1.119 1.75
5-10 45-54 -0.007 0.01
0-5 55-65 -1.273 1.73
-------
25
Table 6. Effects of Chronic Diseases on Ln (Earnings) by Age of Onset
Age of
Onset Coefficient T-Ratio
Asthma
-
-0.232
1.020
Allergies
-
-0.061
0.318
Chronic Bronchitis
-
-0.023
0.065
Emphysema
0-5\
6-10
0.229
-1.038
0.641
2.009
Other Chronic Lung Disease

-0.511
1.294
Arteriosclerosis
45-54
55-65
0.279
-0.624
0.680
1.510
Hypertension
25-34
35-44
45-54
55-65
0.207
-0.041
0.193
0.311
0.916
0.188
1.211
1.167
Heart Attack
25-44
45-54
55-65
0.056
-0.590
-0.376
0.151
1.706
1.141
Stroke
-
0.843
1.386
Other Heart Disease
35-44
45-54
0.302
0.055
1.008
0.165
denotes duration of disease rather than age of onset.
-------
26
Table 7. Effect of Respiratory and Circulatory Diseases on Probability of Participation
by Age of Onset
Disease
Change in probability of participation at each age
Age of Onset 25-34 35-44
45-54
55-65
65 +
Chronic Bronchitis
25
45
-0.067
-0.067
-0.111
-0.084
-0.288
-0.218
-0.180
0.136
Emphysema
45
-0.099
-0.256 -0.159
45
-0.060 -0.157 -0.098
Heart Attack
45
-0.059
-0.155
-0.096
Stroke
45
55
-0.220 -0.571 -0.356
-0.327 -0.204
Other Heart Disease
45
-0.075 -0.196 -0.122
55
-0.324
-0.202
-------
i. /
Table 8. Annual Change in Expected Earnings at Each Age Due to Various Chronic Diseases ($1977)
Annual Change Due to Reduced Probability of Participation
(Change Due to Reduction in Earnings if Working)
Disease
Age of onset
25-34 3 5 - 4 4
45-54
55-65
65+
Chronic Bronchitis 25
45
$-870.2 $-1226.3 $-2229.1 $-4860.9 $-1680.4
-1689.6 -3684.4 -1273.7
Emphysema'
a
45
-1978.4
-4314.3
-10891.)
-1491.5
-6044 .7)
Arteriosclerosis
45
-1210.6 -2639.9
-912.6
Heart Attack
45
-1197.7
-8949. 6)
-2611.8
(-7515.8)
-902.9
-4171.2;
Stroke
45
55
-4415.8
-9629.5
-5511.0
-3328.9
-1905.2
Other Heart Disease 45
55
-1513.7 -3301.0 -1141.2
-5455.6 -1886.0
Effects on Earnings do not begin until duration is greater than or equal to 6 years.
-------
28
Table 9. Sample size by condition, NMCES.
Disease ICDA codes Persons
Total 4789
Hypertension 401-404 3479
Ischemic heart disease 410-414 378
Non-specific heart disease 429 884
Chronic bronchitis 490-491 430
Emphysema 492 222
-------
29
Table 10. Distribution of single vs multiple episodes types.
Number of persons with Percent
single episodes with
only
One* one
Total One One related-to/ single
Disease persons simple same-as stand-alone episode
Hypertension
3479
2476
227
462
91.0
Ischemic
378
195
34
80
81.7
Non-specific heart
CO
CO
501
104
166
87.2
Chronic bronchitis
430
272
49
63
89.3
Emphysema
222
130
21
42
86.9
*In each of these cases there is only one 'stand alone' episode to analyze
that is associated with our target disease.
-------

Table 11. Frequency Distribution of Annual Expenses per Person, by Condition.
Unweighted.
Total Expense
Percentage of Sample in Each Expense Category
Bronchitis Emphysema Hypertension
xpei
777
($19
$ 0
0-25
25-50
50-75
75-100
100-150
150-200
200-300
300-400
400-500
500-750
750-1000
1000-1500
1500-2000
2000-3000
3000-4000
4000-5000
5000-10000
10000-20000
20000+
N
Hean Expense
Median Expense
17.4
36.3
19.8
8.4
4.7
5.1
1.6
3.0
0.7
0.5
0.9
0.2
0.5
0.2
0.2
0.2
0.2
430
$96.74
$23.27
20.7
23.0
7.7
5.4
4.5
8.6
4.1
5.4
1.8
0.9
2.7
1.4
3.6
4.1
1.4
0.9
1.4
1.4
1.4
222
$632.76
$42.63
6.7
21.8
19.2
13.2
8.7
11.2
5.5
5.1
2.4
1.2
1.3
0.6
1.1
0.4
0.6
0.2
0.1
0.5
0.1
0.1
3479
$215.79
$53.51
Ischemic HD
9.0
15.9
10.3
5.6
6.3
9.5
6.9
9.3
2.6
2.4
2.4
2.6
2.9
2.4
1.3
2.9
0.3
3.4
2.6
1.3
378
$1257.55
$116.26
Nonspecific HD
12.9
19.2
11.5
6.8
5.1
7.0
4.6
7.6
3.2
1.6
2.5
1.5
2.6
1.9
2.6
3.1
1.4
2.9
1.1
0.9
884
$1041.26
$73.90
-------
31
Table 12. Average
: Expenses Per Person By Disease
and Category
($1977) .

Mean Expense
Std Dev
Maximum
CONDITION
Expenses