6OOR96159
A Particulate
Matter Risk
Assessment for
Philadelphia
and Los Angeles
3 July 1996,
Revised November 1996
Prepared for
Office of Air Quality Planning arid
Standards
US. Environmental Protection
Agency
Research Triangle Park. NC
Prepared bv
Leland Deck
Ellen Post
Matthew Wiener
Kathv Cunningham
Work funded through
Contract No 68-W4-0029
Work AvM.enments s>.
-------
REVISED REPORT
This is a revised version of the report submitted to EPA on July 3, 1996. The
revisions correct minor errata discovered in the tables presented in the July 3rd report. No
new analyses are presented in this report. Only minimal revisions have been made to the text
where necessary to reflect the changes in the tables. The pagination has been changed
somewhat to incorporate these changes. Revised version submitted to EPA November, 1996
by Abt Associates Inc.
-------
DISCLAIMER
This report is being furnished to the U.S. Environmental Protection Agency by Abt
Associates Inc. in partial fulfillment of Contract No. 68-W4-0029, Work Assignment No. 15-
1. Some of the preliminary work for this report was completed under Work Assignments 9
and 9-1 of the same contract. The opinions, findings, and conclusions expressed are those of
the authors and are not necessarily those of the Environmental Protection Agency. Inquiries
should be addressed to Mr. Harvey Richmond, U.S. EPA, Office of Air Quality Planning and
Standards, MD-15, Research Triangle Park, North Carolina 27711.
-------
ACKNOWLEDGEMENTS
The authors gratefully acknowledge the assistance of Brad Firley, Robert Funk, Kevin
Petrik, and Lily Valle, of Abt Associates Inc. in the production of this report.
The members of and consultants to the Clean Air Scientific Advisory Committee of
EPA's Science Advisory Board provided suggestions and comments on drafts of this report.
Dr. Kinley Laratz, in particular, provided valuable assistance on the statistical techniques used
in the uncertainty analysis.
The Work Assignment Manager, Eric Smith, as well as Harvey Richmond, Karen
Martin, and John Bachmann, all of the U.S. Environmental Protection Agency, provided a
variety of constructive suggestions and comments and technical direction at all stages of work
on this report.
-------
Table of Contents
1. Introduction 1
2. Methods 11
2.1. Overview 11
2.2. Modeling attainment of alternative (or current) standards 12
2.3. The concentration-response function and estimation of health effect
incidence changes 19
2.4. Calculating the aggregate health effects incidence on an annual basis from the
changes in daily health effects incidence 22
2.5. Baseline health effects incidence data 24
2.6. Sensitivity analyses , . 24
3. Assumptions and Caveats 27
3.1. Concentration-response functions 27
3.1.1. Accuracy of the estimates of concentration-response functions ..... 27
3.1.2. Applicability of concentration-response functions in different locations
30
3.1.3. Extrapolation beyond observed air quality levels 32
3.2. The air quality data 33
3.2.1. Appropriateness of the PM indicator 33
3.2.2. Adequacy of PM air quality data 33
3.3. Baseline health effects incidence rates 34
3.3.1. Quality of incidence data 34
3.3.2. Lack of daily health effects incidence rates 35
3.4. Further caveats 36
3.4.1. Highly susceptible subgroups 36
3.4.2. Possible omitted health effects 37
4. Air Quality Assessment: The PM Data 39
4.1. The Philadelphia County PM data 39
4.2. The Southeast Los Angeles County PM data 44
5. Concentration-Response Functions 47
5.1. Concentration-response functions taken from the literature 47
5.2. Estimation of a distribution of P's, estimation of p in any given location, and
characterization of the uncertainty surrounding that estimate 49
5.2.1. Estimation of the distribution of p's 51
5.2.2. Estimating P in a given location 54
5.2.3. Pooled estimates of p 55
5.2.3.1. Pooled analyses of mortality PM-10 concentration-response
functions 56
5.2.3.2. Pooled analyses of mortality PM-2.5 concentration-response
functions 58
-------
5.2.3.3. Pooled analyses of morbidity concentration-response functions
60
5.2.4. Quantitative assessment of uncertainty surrounding P's applied to
Philadelphia and Los Angeles: results 61
5.2.5. Translating a 90 percent credible interval for p into a 90 percent
credible interval for avoided health effect incidence 64
6. Baseline Health Effects Incidence Rates 67
6.1. Sources of incidence data 68
7. Assessment of the Health Risks Associated with "As Is" PM Concentrations Above
Background 71
7.1. Results and sensitivity analyses : . 71
7.2. Uncertainty Analyses 103
7.2.1. A Monte Carlo analysis: propagation of uncertainties from several
sources 104
8. Assessment of the Health Risk Reductions Associated with Attainment of Alternative PM
Standards 112
8.1. Results and sensitivity analyses 112
8.2 An assessment of the plausibility of linear rollbacks and associated sensitivity
analysis 112
8.3. Uncertainty analyses 126
9. Characterization of Risk Associated with PM Pollution: Interpreting the Results of the Risk
Analysis 129
9.1. Variability of predicted health risks 130
9.2. Uncertainty surrounding predicted risks 130
9.3. Importance of the measure of risk 132
9.4. Importance of the indicator of PM: PM-10 vs. PM-2.5 133
9.5. Risk predictions based on concentration-response functions from long-term
exposure studies versus those from short-term exposure studies . 134
9.6. Dependence of results on the assumption that the concentration-response
relationship is applicable at low concentrations
136
Appendix 1: The Relationship Between the Ambient Concentration-Response Function and the
Individual Exposure-Response Function 137
Appendix 2: Pooling the Results of Different Studies . . 141
Appendix 3: The Concentration-Response Function and Relative Risk 147
Appendix 4: A Generalization of the Basic Concentration-Response Function 151
-------
Appendix 5: Adjustment of Means and Standard Deviations of Distributions for
Location-Specific P's 155
References 156
-------
A PARTICULATE MATTER RISK ANALYSIS FOR
PHILADELPHIA AND LOS ANGELES
1. Introduction
Assessing the impacts of ambient paniculate matter (PM) on human health has been a
concern of epidemiological research and government policy for many years. Because PM is
identified as a "criteria pollutant" by the Environmental Protection Agency (EPA) under the
Clean Air Act, PM standards must be reevaluated periodically. An assessment of the current
health risks due to PM and the reduction in health risks associated with achieving alternative
PM standards is part of this process. This document reports the method and results of
analyses to assess the risks associated with current levels of ambient PM hi two selected
locations and to estimate the risk reductions that might be achieved in those locations by
attainment of alternative PM standards.
The Criteria Document (EPA, 1996a) and Staff Paper (EPA, 1996b) evaluate the
scientific evidence on the health effects of PM, including information on exposure routes, the
physiological mechanisms by which PM might damage human health, and concentration-
response components of risk assessment. The risk analysis described in this report builds on
that work. It draws on the hazard identification and concentration-response information
provided in the Criteria Document in order to estimate the incidence of health effects
associated with "as is" ambient PM concentrations and the incidence of health effects that
might be avoided by the attainment of alternative PM standards or sets of standards.
The relationship between a health response and ambient PM concentration is referred to
in this report as the (ambient) concentration-response relationship. It is the relationship
between the average ambient concentration of PM (in /xg/m3) and the population response
(number of individuals exhibiting the health response). In contrast, the relationship between a
health response and individual exposure to PM is referred to as an individual exposure-
response relationship. This is the relationship between the actual exposure to PM (in ^g/m3)
experienced by the individual and the probability that that individual will exhibit the health
response.
Both the individual exposure-response relationship and the ambient concentration-
response relationship are of interest. The individual exposure-response relationship is of clear
scientific interest. This is the relationship that epidemiological studies would presumably
estimate if data on individual exposure were available for each member of a population.1
'Because of the lack of individual exposure data, epidemiological studies typically use ambient
concentration as a surrogate for individual exposure, effectively estimating the ambient concentration-response
function rather than the individual exposure-response function. While doing this may result in a biased estimate of
the individual exposure-response function, it does not result in a biased estimate of the concentration-response
function, which is what is actually estimated. This is explained more fully in Appendix 1.
Abt Associates, Inc. p. 1 July 3, 1996, Revised
-------
However, for the National Ambient Air Quality Standards (NAAQS), which influence ambient
concentrations of PM (through environmental policies and regulations that lead to meeting the
standards), the ambient concentration-response relationship is of primary regulatory interest.
That is, it is important to predict the risk reduction associated with changing ambient
concentrations, rather than the risk reduction associated with changing individual exposure
(which is not directly controlled by the NAAQS). It is therefore the concentration-response
functions which are appropriate to use in the PM risk analysis. The relationship between the
individual exposure-response function and the (ambient) concentration-response function is
examined formally in Appendix 1.
The risk analysis considers two different PM indicators. The indicator for the current
air quality standard is defined as those particles of diameter less than or equal to 10 microns
and is denoted as PM-10. The Staff Paper (EPA, 1996b) recommends consideration of an
indicator measuring fine particles, defined as those of aerodynamic diameter less than or equal
to 2.5 microns and denoted as PM-2.5. Both PM-10 and PM-2.5 are examined in the risk
analyses.2
There are two major phases of the risk analysis. The first phase assesses the risks
associated with "as is" PM concentrations in a specified location.3 If the location is not in
attainment of current standards, risk analyses are carried out in two ways: (1) daily PM
concentrations are left unadjusted, and (2) daily PM concentrations are first adjusted to
Simulate attainment of the current standards prior to the analysis. The method of adjustment is
described in Section 2.2 below. The basic question addressed in the first phase of the risk
analysis is of the following form:
For a given human health endpoint (mortality, hospital admissions,
etc.), what is the estimated incidence of the health endpoint that may be
associated with "as is " PM concentrations ?
The second phase of the risk analysis estimates the risk reductions that would be
associated with the attainment of alternative PM standards as opposed to attainment of current
standards. Annual average PM-2.5 standards of 15 and 20 /xg/m3 are each considered alone as
well as in combination with daily PM-2.5 standards of 65, 50, and 25 ^g/m3, respectively.
Attainment of a standard or set of standards is simulated by adjusting "as is" daily PM
2While "risk analysis" is used to refer to each separate analysis (e.g., for a particular location under
particular assumptions), the entire collection of risk analyses is referred to interchangeably as (the set of) "risk
analyses" or as a "risk analysis."
3Risk is assessed for PM levels down to the lowest level observed in the study reporting the
concentration-response function, but not lower than background level in the sample location. If the lowest
observed PM level was not reported, risk is assessed down to background level in the sample location.
Background PM level is the PM concentration in the absence of controllable anthropogenic sources in North
America. Background concentrations are treated in a manner consistent with the Criteria Document.
Abi Associates, Inc. p. 2 July 3, 1996, Revised
-------
concentrations to daily PM concentrations that would just meet the standard(s). The impact on
human health is assessed by comparing the health risks associated with PM concentrations that
attain the alternative PM-2.5 standards with the health risks associated with the "as is" PM
concentrations that attain the current (PM-10) standards. The basic question addressed is of
the following form:
For a given reduction in PM concentrations and a given human health
endpoint (mortality, hospital admissions, etc.), what is the estimated
reduction in incidence of the health endpoint associated with the
reduction in PM concentrations?
As in the first phase of the risk analysis, if the location is not in attainment of current
standards, daily PM concentrations are adjusted to simulate attainment of current standards
prior to the analysis.
The PM risk analysis described in this report is not a national risk assessment, nor does
it model micro-environment exposure (as was done as part of the risk assessment prepared for
the recent review of the ozone NAAQS4). Extensive risk analyses are instead carried out for
two sample locations by applying concentration-response functions from epidemiological
studies to data on daily ambient PM-10 and PM-2.5 levels in each location (consistent with the
general approach taken in the ozone risk analyses involving risk estimates based on
epidemiology studies).
The two locations chosen for risk analysis are Philadelphia County, Pennsylvania and
Southeast Los Angeles County, California. The geographic region comprised, the population
encompassed within the region, and the placement of air quality monitors used in the risk
analysis are illustrated in Exhibit 1.1 for Philadelphia County and Exhibit 1.2 for Southeast
Los Angeles County. A portion of southeastern Los Angeles County selected for use in the
analysis includes the portion of the county with the highest PM-10 levels. The region included
in this analysis approximates the portion of the county reported to have an annual average
PM-10 level above 40 jug/m3 in 1994 (from "Air Quality Standards Compliance Report,"
South Coast Air Quality Management District, 1995). The size and age distribution of the
population living within the selected region was estimated by totaling the population of U.S.
Census block groups falling within the region. A block group is considered to be within the
region if the population-weighted centroid of the block group is within the boundary of the
region.
""Review of National Ambient Air Quality Standards for Ozone: Assessment of Scientific and Technical
Information" (EPA, 1996c), and "A Probabilistic Assessment of Health Risks Associated With Short-Term
Exposure to Tropospheric Ozone" (Whitfield et al., 1996).
Abt Associates, Inc. p. 3 July 3, 1996, Revised
-------
Exhibit 1.1
Monitor Locations Used for PM Analyses
in Philadelphia County
N
w *5K^ E
s
I 1 Philadelphia County
j, PM Monitor Station Used in Analyses
Abt Associates, Inc. p. 4
1990 Population, Philadelphia County = 1,585,577
July 3, 1996, Revised
-------
Exhibit 1.2
Geographic Region Used for PM Analyses
in Southeast Los Angeles County
"Pasadena
Central
Los Angeles
West Los
Angeles
Diamond Bar
I I Los Angeles County
/\ Boundary of PM Region
j, PM Monitor Station Used in Analyses
i Other Monitoring Stations
1990 Population, LA County = 8,863,164
1990 Southeastern County Population
within Indicated Boundary = 3,635,984
Ahl Associates, Inc.
p. 5
July 3, 1996, Revised
-------
Philadelphia County has virtually complete daily air quality data for both PM-10 and
PM-2.5. For a one-year period from September 1992 through August 1993, monitor data for
PM-10 are available for 98.6 percent of the days in the year, and monitor data for PM-2.5 are
available for 96.4 percent of the days in the year. In addition, Philadelphia County has been
the site of extensive investigation of the health effects of air pollution.
Southeast Los Angeles County, a western location, provides a contrast to Philadelphia
County in type of particulate matter. In addition, as in Philadelphia County, substantial air
quality data for both PM-10 and PM-2.5 are available for Los Angeles. Finally, several health
studies have been carried out in this city.
Numerous epidemiological studies are used in the risk analyses. Most studies focus on
adverse effects associated with elevations in PM levels during short time periods. These
studies, referred to as "short-term exposure" studies, draw current incidence levels primarily
from hospital and vital statistics records. The "long-term exposure" studies, on the other hand,
evaluate mortality or morbidity in relation to long-term air quality, characterized by annual
mean levels of PM. These studies used large cohorts of adults with specifically defined
characteristics who were followed over years of observation. The health endpoints for which
the largest number of studies are available are mortality, hospital admissions for pneumonia,
hospital admissions for Chronic Obstructive Pulmonary Disease (COPD), and "total
respiratory" hospital admissions. (The exact set of ICD codes included in "total respiratory"
admissions varies from study to study.)
In some cases, most notably in the case of mortality and PM-10, concentration-
response relationships have been estimated by several studies in the literature. For those
health effects for which associations with PM have been estimated in several studies, .ideally,
the data sets from these studies could be combined and re-analyzed to produce a more robust
estimate of PM health effects. When it is impossible to combine the data, however, there are
various ways to pool the results of the studies to derive a concentration-response function.
One method, which is used in this risk analysis, is described in Appendix 2.
In the first phase of the risk analysis, assessing the risk associated with "as is" ambient
PM concentrations, the number of separate analyses is determined by the number of health
endpoint-PM indicator combinations for which concentration-response functions have been
estimated. In the second phase of the risk analysis, assessing the risk reduction associated
with attaining possible alternative (sets of) standards, the number of separate analyses is
determined by the number of health endpoints and the number of (sets of) alternative standards
considered. If there are N health endpoints and M sets of PM standards of interest, there is a
maximum of N*M analyses possible.
Abt Associates, Inc. p. 6 July 3, 1996, Revised
-------
An overview of the major components of the PM risk analysis discussed in this report
is presented in Exhibit 1.3. Each separate analysis in the first phase of the risk analysis
depends on the following four basic components:
air quality information,
information on the concentration-response relationship between the health endpoint of
interest and the PM indicator of interest in the location of interest;
baseline health incidence information for the location of interest; and
the size of the population living in the location of interest.
If the location is not in attainment of current PM standards (as in Los Angeles), the first phase
of the risk analysis requires, prior to risk analysis,
the simulation of attainment of current standards.
Each separate analysis in the second phase of the risk analysis depends on an additional
component:
the simulation of attainment of a set of alternative PM-2.5 standards.
There are substantive issues surrounding each of these components. These issues and
approaches in the absence of complete information on any one or more of these risk analysis
components are discussed at length in the sections that follow.
The basic methods used in all the analyses, and methodological issues specific to
particular parts of the risk analysis (e.g., to one phase or the other), are discussed in Section
2. Because the risk analyses were carried out in the face of incomplete information, it was
necessary to make assumptions at several points in the analysis process. These assumptions
and the various sources of uncertainty surrounding risk estimates are discussed in Section 3.
Section 4 discusses the Philadelphia County and Southeast Los Angeles County air quality data
used in the analyses, that is, the ambient PM-10 and PM-2.5 data from these locations. The
PM-10 and PM-2.5 concentration-response functions used in the analyses are discussed in
Section 5. Concentration-response functions in the epidemiological literature that were
considered in the risk analysis are given in Section 5.1. The estimation of a distribution of
concentration-response functions (across locations), the estimation of a concentration-response
function in any given location, and the characterization of the uncertainty surrounding
concentration-response functions is discussed in Section 5.2. Section 6 presents baseline
health effects incidence rates for each of the locations from vital statistics sources. These are
the health effects incidence rates associated with "as is" PM levels.
In both phases of the risk analysis, there is substantial uncertainty. The results of the
analyses depend on a number of analytic choices and will change if different choices are made.
Abt Associates, Inc. p. 7 July 3, 1996, Revised
-------
Exhibit 1.3 Major Components of Particulate
Matter Health Risk Analysis
Ambient Population-
oriented Monitoring
for Selected Cities
Air Quality Adjustment
Procedures
Alternative Proposed
Standards
Human Epidemiological
Studies (various health
endpoints)
City-specific (or National)
Baseline Health Effects
Incidence Rates (various
health endpoints)
"As is" Analysis
Changes in
Distribution
of PM Air
Quality
Concentration
Response
Relationships
Health
Risk
Model
Risk Estimates:
"As is"
"Alternative
Scenarios"
r v
S / = Sensitivity Analysis: Analysis of effects of alternative assumptions, procedures or data occurs at these points.
Abt Associates, Inc.
p. 8
July 3, 1996, Revised
-------
One way to assess the impact of particular analytic choices on the results of the risk analyses is
through sensitivity analyses, in which the impact of different analytic choices on the results of
the risk analysis is assessed.
The assessment of the risk associated with "as is" PM concentrations in Philadelphia
County and Southeast Los Angeles (the first phase of the risk analysis) is presented in Section
7. The results and associated sensitivity analyses are presented in Section 7.1. Monte Carlo
propagation of uncertainty analyses, considering the uncertainty from several sources, are
presented hi Section 7.2.
The assessment of the risk reduction associated with attaining possible alternative (sets
of) standards hi these two locations (the second phase of the risk analysis) is presented in
Section 8. The results and associated sensitivity analyses are presented hi Section 8.1. Other
sensitivity analyses concerning the rollback method are presented in Sections 8.3. Alternative
forms of PM standards are considered hi Section 8.3. Monte Carlo propagation of uncertainty
analyses, considering the uncertainty from several sources, are presented hi Section 8.4.
Finally, issues of interpretation of the results of the risk analysis are discussed in Section 9.
Abt Associates, Inc. p. 9 July 3, 1996, Revised
-------
Abt Associates, Inc. p. 10 MV 3, 1996, Revised
-------
2. Methods
This section describes the basic methods of the risk analysis. Conducting a risk analysis
requires substantial information. For each of the elements of the risk analyses described below
complete and certain information is not available, resulting in a significant degree of
uncertainty. The sources of uncertainty and the assumptions made to perform risk analyses hi
the face of incomplete information are discussed in Section 3.
2.1. Overview
Each separate analysis hi either the first phase or the second phase of the risk analysis
can be characterized as estimating the change in the incidence of a given health effect resulting
from a given change in PM concentrations. In the first phase, risk analyses consider the health
effects incidence associated with "as is" PM above either the lowest PM level observed hi the
study or background level. (In Los Angeles County, where "as is" concentrations do not meet
current PM standards, the health effects incidence associated with unadjusted "as is" PM
concentrations and the health effects incidence associated with "as is" PM concentrations
adjusted to simulate attainment of current standards are both considered.) This is equivalent to
assessing the potential change hi health effects incidence associated with a reduction in PM
concentrations from "as is" levels (in or out of attainment with current standards) to the
specified lower PM level (either the lowest observed in the study or background level).
In the second phase, risk analyses consider the change from "as is" PM concentrations
to those concentrations that would just attain a specified set of alternative PM standards. The
method used in both phases of the risk analysis is therefore basically the same. The important
difference between the two phases is in the specified alternative (lower) PM levels: in the first
phase the alternative air quality is either the lowest PM level observed in the study or
background PM level, whereas in the second phase the alternative air quality is based on
attainment of a set of alternative PM-2.5 standards. The first phase therefore requires either a
reported lowest observed PM level or an estimate of background PM (PM-10 and PM-2.5)
level; the second phase requires that a method be developed to simulate attainment of the
specified standard(s). This method is applied to the first phase as well to simulate attainment
of current PM-10 standards where appropriate prior to risk analyses.
To estimate the change in the incidence of a given health effect resulting from a given
change in ambient PM concentrations in a sample location, the following elements are
necessary:
(1) air quality data from the sample location to estimate both "as is" PM
concentrations and, for the second phase of the risk analysis, the concentrations
associated with attainment of proposed PM standards;
Abt Associates, Inc. p. 11 July 3, 1996, Revised
-------
(2) a concentration-response function estimating the relationship between ambient
PM concentrations and the health endpoint in the sample location (preferably
derived in the same location, although functions estimated in other locations can
be used at the cost of increased uncertainty see Section 3.1.2); and
(3) an estimate of the baseline health effect incidence (rate) corresponding to "as is"
PM levels (since most of the available concentration-response functions give a
percent change in incidence rather than an absolute number of cases).
The change in the health endpoint may be measured as a daily change, corresponding
to changes in daily average ambient PM concentrations from "as is" levels to some alternative
levels (e.g., either background or those levels corresponding to attainment of a set of
standards). Alternatively, the change in the health endpoint may be measured as an annual
change, corresponding to a change in the annual average PM concentration. When
concentration-response functions from short-term exposure studies are used, it is appropriate
to assess daily effects. When concentration-response functions from long-term exposure
studies are used, it is appropriate to assess annual effects. When daily effects are calculated,
these daily changes are aggregated, and, in the absence of PM data for all 365 days of the
year, adjusted to reflect the total for the entire year, as described below. All changes in health
effects, whether calculated on a daily or annual basis, are therefore aggregates for an entire
year. The risk analysis procedure described in more detail below is diagramed in Exhibit 2.1
for analyses based on short-term exposure studies and Exhibit 2.2 for analyses based on long-
term exposure studies.
Because there are substantive methodological issues involved in simulating the
attainment of a set of standards (either current PM-10 standards or alternative PM-2.5
standards), this is discussed separately in Section 2.2 below. The functional form of the
concentration-response relationships used in the risk analyses, and the prediction of changes in
health effects incidence associated with changes in ambient PM concentrations using these ,
concentration-response functions is described in Section 2.3. Issues involved in the calculation
of annual health effects incidence are discussed in Section 2.4. A brief discussion of issues
involved in attaining baseline incidence rates is given in Section 2.5. Finally, the sensitivity
analyses carried out in both phases of the risk analysis, and any methodological issues
pertaining to them, are described in Section 2.6.
2.2. Modeling attainment of alternative (or current) standards
Predicting the change in risk due to a change in air quality from an "as is" annual
mean to meet a lower annual standard when using a concentration-response function from a
long-term exposure study is straightforward: the "as is" mean is simply reduced to the
standard level. In this case, simulating attainment of a standard does not involve generating an
alternative set of daily PM concentrations, because the concentration-response function
Abt Associates, Inc. p. 12 July 3, 1996, Revised
-------
Exhibit 2.1
Flow Diagram of Typical Risk Analysis for Short-Term Exposure Studies
Air Quality Data
Obtain individual
monitor PM
data
Compute dally
average PM
concentrations
for available days
4
Compute change
In PM on each day
PM concentration
Is available
i
Concentration-Response Functions
Convert RR
toll '
i (If necessary) !
1 Calculate
I pooled
I function
(If necessary)
I
Compute % change
In health effects
associated with
changes In PM for
each day on which
PM concentration
Is available
Baseline Health Incidence
Compute total %
change In health effects
by summing dally
results, with missing
day corrections
Adjust to dally
baseline incidence
I
Compute
annual
#
of cases
associated
with
change In PM
r Identify rollback"
method or I
cutpolnt levels |
I (for certain
i analyses) '
Result:
Percent change in
total incidence
Result:
PM associated
incidence
Abt Associates, Inc.
p. 13
July 3, 1996, Revised
-------
Exhibit 2.2
Flow Diagram of Typical Risk Analysis for Long-Term Exposure Studies
Air Quality Data
Obtain Individual
monitor PM
data
/once
^-
Con
av
con
for a\
ipute daily
erage PM
:entratlons
rallable days
ntration-Response Function
Identify
location-
specific P
studies
Identify functional
form
Identify
Relative Risk
**" (RR)orslope
coefflcents (B)
Iconve
kh 1 In
(if nec
Compute annual
average
with missing day
corrections
s Identify range
In study
/
i : Calculate 1
rt RR j pooled ,
B ^^"| function ^^
Bssary) ; ' (If necessary) 1
Compute Chang*
In annual
average PM
1
r
Compute %
change
In health effects
associated with
changes In PM
I Identify rollback''
method or '
cutpolnt level |
< (for certain
I analyeee) '
Baseline Health Incidence
Compute #
of cases
associated
with
change In PM
Result:
Percent change in
total incidence
Result:
PM associated
incidence
Aht Associates, Inc.
p. 14
July 3, 1996, Revised
-------
estimated in a long-term exposure study is based on annual, rather than daily PM
concentrations.
When a concentration-response function from a short-term exposure study is used,
however, attainment of an alternative standard or set of standards is best simulated by
changing the distribution of daily PM concentrations. This section discusses the methods used
to change daily PM concentrations in a sample location to simulate the attainment of a new
standard or set of standards. The methods described below are also applicable to the
simulation of attainment of current standards when a location is not already in attainment, as
discussed below.
An area is considered in attainment of a standard if all PM monitors in the area are in
attainment. An area is in attainment of an annual standard if the annual average PM
concentration at each monitor in the area is at or below the standard. An area is in attainment
of a daily standard (which currently allows one exceedence) if no more than one monitor-day
exceeds the daily standard. Although it is possible to change the daily PM concentrations at
each monitor separately (to separate degrees) to simulate attainment, this would require
extensive analysis that is beyond the scope of this risk analysis. Therefore, although the
amount or percent of reduction on a given day might be determined by the PM concentration
at a single monitor on a single day, attainment is simulated by changing daily concentrations
averaged over all monitors.
There are many different methods of reducing daily PM levels that would result in
attainment of a given PM standard or set of standards. Preliminary analyses of historical PM
data found that year-to-year reductions in PM levels in a given location tended to be roughly
linear. That is, both high and low daily PM levels decreased proportionally. (This is
discussed more fully in the discussion of the sensitivity of results to the rollback method, in
Section 8.) This suggests that, in the absence of detailed air quality modeling, it is reasonable
to simulate PM reduction to bring a sample location into attainment of new proposed standards
by proportional rollbacks (i.e., by decreasing PM levels on all days by the same percentage).
Proportional (linear) rollback is only one of many possible ways, however, to create an
alternative distribution of daily concentrations to meet new PM standards. One could, for
example, reduce the high days by one percentage and the low days by another percentage,
choosing the percentages so that the new distribution achieves the new standard. At the
opposite end of the spectrum from linear rollbacks, it is possible to meet a daily standard by
"peak shaving." The peak shaving method would reduce all daily PM concentrations above a
certain concentration to that concentration (e.g., the standard) while leaving daily
concentrations at or below this value unchanged. While a strict peak shaving method of
attaining a standard is unrealistic, it is illustrative of the principal that patterns different from a
proportional rollback might be observed in areas attempting to come into compliance with
revised standards.
Abt Associates, Inc. p. 15 July 3, 1996, Revised
-------
If the short-term exposure concentration-response functions were exactly linear, then
the overall estimated change in health effects associated with short-term exposure would
depend only on the total change in PM concentration (above the lowest level at which PM
pollution causes health effects). Because the concentration-response functions being
considered are almost linear, the method by which daily PM concentrations are reduced to
meet annual standards makes almost no difference.
However, the method by which daily concentrations are reduced to meet daily
standards may make a sizeable difference, since it is the distribution of all the daily changes in
air quality concentrations (above the lowest observed level at which PM pollution is associated
with health effects) that results in the aggregate annual risk estimates. If one rollback method
results in an air quality distribution with considerably more days with large changes in air
quality than another method that also attains a given daily standard, the two methods will
estimate significantly different health risks.
The estimated change in health effects based on short-term (daily) exposure
concentration-response functions, then, is sensitive to the reduction method only to the extent
that different reduction methods result in different total amounts of PM being removed from
the atmosphere. Therefore, when the annual mean standard is the controlling standard (so a
given total amount of PM must be removed), results should be relatively insensitive to
different reduction methods (and would be totally insensitive to the reduction method if the
concentration-response functions were exactly linear). When a daily standard is the
controlling standard, however, results will be sensitive to different reduction methods.
Attainment of a set of standards was simulated by proportional rollback. That is,
average daily PM concentrations were reduced by the same percentage on all days. Because
pollution abatement methods are applied largely to anthropogenic sources of PM, rollbacks
were applied only to PM above estimated background levels. (Rollbacks were estimated only
for PM-2.5. Background PM-2.5 concentrations were estimated as 3.5 /ig/m3 in Philadelphia
County, and 2.5 ^g/m3 in Southeast Los Angeles County. This is consistent with the approach
of the Criteria Document.) The percent reduction was determined by the controlling standard.
For example, suppose both an annual and a daily PM-2.5 standard are proposed. Suppose pa
is the percent reduction required to attain the annual standard, i.e., the percent reduction of
daily PM above background necessary to get the annual average at the monitor with the
highest annual average down to the standard. Suppose pd is the percent reduction required to
attain the daily standard with one exceedence, i.e., the percent reduction of daily PM above
background necessary to get the second highest monitor-day down to the daily standard. If pd
is greater than pa, then all daily average PM concentrations above background are reduced by
pd percent. If pa is greater than pd, then all daily average PM concentrations are reduced by pa
percent. Information on controlling monitors and percent rollbacks necessary to simulate
attainment of alternative PM-2.5 standards in Philadelphia County and Southeast Los Angeles
County is given in Exhibits 2.3 and 2.4.
Abt Associates, Inc. p. 16 July 3, 1996, Revised
-------
Because the reduction method to attain a daily standard could have a significant impact
on the risk analysis results, sensitivity analyses were conducted on different rollback methods
for meeting proposed standards. The results of these analyses are presented in Section 8.
Exhibit 2.3. Controlling Monitors for Rollbacks to Attain Alternative PM-2.5 Standards
Monitor Site
Philadelphia County
N/E
PBY
TEM
Weighted Annual
Average PM-2.5
Concentration*
Second Daily
Maximum 24-Hour
PM-2.5
Concentration*
15.5
16.7
17.1
65.1
72.2
70.0
Controlling Monitor
For daily standard
For annual standard
Southeast Los Angeles County
Central LA
Diamond Bar
24.1
21.9
91.1
101.7
For annual standard
For daily standard
All concentrations are given in ^g/m3 .*Both weighted annual averages and second daily maximum concentrations
at the two monitors in Southeast Los Angeles County were adjusted to reflect attainment of the current PM-
lOannual standard of 50 /*g/m3 and the current PM-10 daily standard of 150 ^g/m3. These standards are currently
attained in Philadelphia County.
Abt Associates, Inc.
p. 17
July 3, 1996, Revised
-------
Exhibit 2.4. Controlling Standards and Percent Rollbacks Necessary to Attain
Alternative PM-2.5 Standards
Alternative PM-2.5
Standards
Annual Avg.
Standard
20 alone
20
20
20
15 alone
15
15
15
24-Hour
Standard
65
50
25
65
50
25
Philadelphia County
Controlling Standard and
Percent Rollback*
Daily - 10.4%
Daily - 32.3%
Daily - 68.7%
Annual - 15.5%
Annual -- 15.5%
Daily -32.3%
Daily - 68.7%
Southeast Los Angeles County
Controlling Standard and
Percent Rollback**
Annual - 18.8%
Daily - 37.0%
Daily -52.1%
Daily - 77.3%
Annual - 42.0%
Annual -- 42.0%
Daily -52.1%
Daily - 77.3%
All concentrations are given in
*Based on controlling values for Philadelphia County of 17.1 /ig/m3 for the annual standard and 72.2 /xg/m3 for
the daily standard.
** Based on controlling values for Southeast Los Angeles County of 24.1 /xg/m3 for the annual standard and 101.7
/ig/m3 for the daily standard.
The linear rollback methods described above to simulate attainment of alternative sets
of PM-2.5 standards are also used to simulate attainment of current PM-10 standards prior to
analyses in both the first and second phases of the risk analysis for Southeast Los Angeles
.County, which is out of attainment for current standards. Analyses for PM-10 in the first
phase of the risk analysis use exactly the rollback method described above. Analyses for PM-
2.5 in the first and second phases of the risk analysis assume that PM-2.5 is rolled back
proportionately to PM-10 rollbacks. If, for example, daily PM-10 concentrations are
decreased by 10 percent to simulate attainment of current PM-10 standards, it is assumed that
daily PM-2.5 concentrations decrease by 10 percent as well. It is these adjusted PM-2.5
concentrations, the "as is" PM-2.5 concentrations in attainment of current standards, that are
then reduced again by proportional rollback methods to simulate the attainment of alternative
PM-2.5 standards in the second phase of the risk analysis. The adjustment of PM-10 and PM-
2.5 concentrations in Southeast Los Angeles County to simulate attainment of current PM-10
standards is summarized in Exhibit 2.5.
Abt Associates, Inc.
p. 18
July 3, 1996, Revised
-------
Although some epidemiological studies estimated linear concentration-response
functions, most of the studies used a method referred to as "Poisson regression" to estimate
exponential concentration-response functions in which the natural logarithm of the health
endpoint is a linear function of PM6:
y = B e* , (1)
where x is the ambient PM level, y is the incidence of the health endpoint of interest at PM
level x, P is the coefficient of ambient PM concentration, and B is the incidence at x=0, i.e.,
when there is no ambient PM. The relationship between a specified ambient PM level, XQ, for
example, and the incidence (rate) of a given health endpoint associated with that level (denoted
as y0) is then
3> =Be^ . (2)
Because the exponential form of concentration-response function (equation (1)) is by far the
most common form, the discussion that follows assumes this form. However, because the
coefficients estimated by the epidemiology studies are extremely small, these exponential
functions are nearly linear. The consequences of this near-linearity are discussed below.
Ambient PM levels may be based on any averaging time, e.g., they may be daily
averages or annual averages, as long as the health effect incidence corresponds to the PM
averaging time. For example, the concentration-response function may describe the
relationship between daily average ambient PM concentrations and daily mortality, or it may
describe the relationship between annual average ambient PM concentrations and annual
mortality. Some concentration-response functions were estimated by using moving averages of
PM to predict daily health effects incidence. Such a function might, for example, relate the
incidence of the health effect on day t to the average of PM concentrations on days t and (t-1).
(This may be considered a variant on the short-term, or daily concentration-response function.)
The discussion below does not indicate averaging times and simply assumes that the measure
of health effect incidence, y, is consistent with the measure of ambient PM concentration, x.
The change in health effects incidence, Ay = y0 - y, from y0 to the baseline incidence
rate, y, corresponding to a given change in ambient PM levels, Ax = XQ - x, can be derived
from equations (1) and (2) (as shown in Appendix 3) as
A>> =y(e^ - 1] . (3)
6Poisson regression is essentially a linear regression of the natural logarithm of the dependent variable on
the independent variable, but with an error structure that accounts for the particular type of heteroskedasticity that
is believed to occur in health response data. What matters for the risk analysis, however, is simply the form of
the estimated relationship, as shown in equation (1).
Abt Associates, Inc. p. 20 July 3, 1996, Revised
-------
Alternatively, the change in health effects incidence can be calculated indirectly using
relative risk. Relative risk (RR) is a well known measure of the comparative health effects
associated with a particular air quality comparison. The risk of mortality at ambient PM level
KO relative to the risk of mortality at ambient PM level x, for example, may be characterized
by the ratio of the two mortality rates: the mortality rate among individuals when the ambient
PM level is Xo and the mortality rate among (otherwise identical) individuals when the ambient
PM level is x. This is the relative risk for mortality associated with the difference between the
two ambient PM levels, XQ and x. Given a concentration-response function and a particular
change hi ambient PM levels, Ax, the relative risk associated with that change in ambient PM,
denoted as RR^, is equal toepA*. The change in health effects incidence, Ay, corresponding
to a given change hi ambient PM levels, Ax, can then be calculated based on this relative risk:
by = y(RR^ - 1) . (4)
Equations (3) and (4) are simply alternative ways of writing the relationship between a given
change in ambient PM levels, Ax, and the corresponding change in health effects incidence,
Ay. The derivation of equation (4) is shown hi Appendix 3. These equations are the key
equations that combine air quality information, concentration-response information, and
baseline health effects incidence information to estimate ambient PM health risk.
Given a concentration-response function and air quality data (ambient PM values) from
a sample location, then, the change in the incidence of the health endpoint (Ay = y0 - y)
corresponding to a change in ambient PM level of Ax = XQ - x is determined. This can either
be done with equation (3), using the coefficient, (J, from a concentration-response function, or
with equation (4), by first calculating the appropriate relative risk from the concentration-
response function.
Because the estimated change in health effect incidence, Ay, depends on the particular
change in PM concentrations, Ax, being considered, the choice of PM concentration change
considered is important. These changes in PM concentrations are generally reductions from
the current levels of PM ("as is" levels) to some alternative, lower level(s).
If a location is not in attainment of current PM standards, as is the case in Southeast
Los Angeles County, current levels may be characterized in two ways. It may be of interest to
compare health effects at "as is" PM concentrations (that do not attain the current standards)
with health effects at some alternative PM level(s). Alternatively, it is also of interest to
compare health effects at those PM concentrations that just attain the current PM-10 standards
with health effects at those PM concentrations that just attain some alternative (PM.2.5)
standards. This is an appropriate context for examining the potential risk reductions associated
with revising the current standard.
The first and second phases of the risk analysis are distinguished primarily by the
choice of lower PM level(s). The second phase considers the changes in health effects
Abt Associates, Inc. p. 21 July 3, 1996, Revised
-------
incidence associated with changes from PM concentrations that meet the current standards to
PM concentrations that just meet alternative PM-2.5 standards.
When possible, the choice of lower PM level(s) in an analysis in the first phase of the
risk analysis is the lowest PM concentration observed in the study that estimated the
concentration-response function used in the risk analysis. This is the lowest PM concentration
at which the concentration-response function is supported. However, many of the short-term
exposure PM studies do not report the lowest observed PM concentration. (For example,
many studies report the lowest decile or quartile values.) When the lowest observed PM
concentration is not reported (or if it is lower than background level), analyses in the first
phase of the risk analysis consider the range of "as is" PM concentrations in the sample
location down to background PM concentration in that location.
In contrast to most short-term exposure studies, long-term exposure studies routinely
report the lowest observed annual average PM concentration. Risk analyses that use long-term
exposure concentration-response functions therefore consider the range of "as is" annual
average PM in the sample location to the lowest annual average PM level observed in the
study.
In both phases of the risk analysis, the ambient PM concentrations to which "as is"
ambient PM concentrations are compared are generally lower than or equal to "as is"
concentrations. Therefore Ax = XQ - x is negative (or zero), and so the corresponding change
in incidence of health effects, Ay, is also negative (or zero). That is, there are fewer cases of
any given health effect at lower ambient PM levels. Alternatively, -Ay may be interpreted as
the health effects attributable to PM concentrations between XQ and x.
Because different epidemiological studies report different estimated concentration-
response functions for a given health endpoint, predicted changes in health effects incidences
depend on the concentration-response function used. The uncertainty introduced into the risk
analyses by this is assessed both through sensitivity analyses and through Monte Carlo
methods (see Section 9).
2.4. Calculating the aggregate health effects incidence on an annual basis from
the changes in daily health effects incidence
To assess the daily health impacts of daily average ambient PM levels above
background or above the levels necessary to achieve a given standard, concentration-response
functions from short-term exposure studies were used together with estimated changes in daily
ambient PM concentrations to calculate the daily changes in the incidence of the health
endpoint. Adding these changes over all the days in a year yields the annual change.
(Alternative assumptions about the range of PM levels associated with health effects were
explored in sensitivity analyses. When a minimum concentration for effects is considered,
Abt Associates, Inc. p. 22 July 3, 1996, Revised
-------
reductions below this concentration do not contribute attributable cases to the calculation.
Only reductions down to this concentration contribute attributable cases to the calculation.)
After daily changes in health effects are calculated, an annual change is calculated by
summing the daily changes. However, there are some days for which no ambient PM
concentration information is available. The predicted estimated risks, based on those days for
which air quality data are available, must be adjusted to take into account the full year.
In Philadelphia County, there are very few missing days, and these are distributed
evenly throughout the year. In this case, the adjusted health effects incidence is the original
incidence multiplied by the number of days in a year and divided by the number of days for
which data are available; that is, the figure is simply scaled for the fraction of days on which
there are data. In Philadelphia County, for example, PM-10 data are available for 358 days
evenly distributed throughout the year. Suppose the sum of the daily premature deaths
associated with PM on those 358 days is 600, then the adjusted figure is 612 (i.e., 600 x
365/358). This reflects the assumption that the distribution of PM concentrations on those
days for which data are available accurately reflects the distribution of ambient PM
concentrations for the entire year, and that the concentration-response functions were estimated
using data from the entire year.
In Southeast Los Angeles County, however, the distribution of missing days varies
significantly in different periods of the year. During the first quarter of 1995, air quality
monitoring was done on roughly one in six days; during the second quarter, it was done on
roughly one in three days; and during the third and fourth quarters, it was done almost every
day. Because of this, adjustments were made separately in each quarter, and the results
added. Adjustment of health effects incidence within a quarter in Southeast Los Angeles
County was done with the same method used to adjust health effects incidence throughout the
year in Philadelphia County.
Some concentration-response functions are based on average PM levels during several
days. When these concentration-response functions are used, the air quality data are averaged
for the same number of days. For example, a function based on two-day averages of PM
would be used in conjunction with two-day averages of PM in the sample location to predict
the incidence of the health effect in that location. In some cases, intervals of three or more
consecutive days in a given location are missing data, and so no multi-day average is available
for use with multi-day concentration-response functions. These cases were treated by multi-
day functions just as individual missing days were treated by single-day functions: they
contributed no cases to the risk analysis, and figures were adjusted for the days on which
multi-day averages were missing.
Concentration-response functions from long-term exposure studies were used to assess
the annual health impacts of changes in annual average ambient PM concentrations. In this
case, the "as is" annual concentration is simply the average concentration for those days on
Abi Associates, Inc. p. 23 July 3, 1996, Revised
-------
which data are available, if missing days are evenly distributed throughout the year (as hi
Philadelphia County), or a composite of quarterly averages if missing days are not evenly
distributed throughout the year (as in Southeast Los Angeles County).
Note that while the long-term exposure studies use annual average PM concentration as
the PM indicator, the studies were conducted in such a way that they may have detected effects
due to PM exposure over some longer period. For example, average PM concentrations over
the course of five years might be the appropriate measure. It is therefore possible that the full
benefits of reducing PM predicted by these studies would not appear in the first year after
reductions to attain a standard, but would be "phased in" gradually as concentrations during
successive years were also reduced. If average PM concentrations over five years is the
appropriate measure, for example, the benefits of a standard would gradually increase to their
full level over the course of the five years after the new standard had been attained. The risk
analysis makes no attempt to determine the appropriate exposure period for long-term
exposure studies. The estimated annual benefits of reduced long-term exposure are assumed to
be completely achieved by the future year for which attainment of the new standard is being
modeled. The issue is partially addressed, however, in a sensitivity analysis which examines
the effect of altering the "slope" parameter in the long-term exposure concentration-response
function.
2.5. Baseline health effects incidence data
Baseline health effects incidence rates (e.g., death rates) and population sizes (to
calculate baseline incidence levels) for the selected locations were obtained from vital statistics
sources. Location-specific information was used whenever possible. However, location-
specific baseline incidence data for hospital admissions and other morbidity endpoints are not
as readily available as for mortality from national data sources. Where possible, local sources
of data (e.g., from city, county or state health agencies) were obtained. However, such data
are not uniformly available, and alternative procedures were used in some instances. For
respiratory symptom or illness health endpoints, routine surveillance and reporting is not
generally conducted in metropolitan areas, in contrast to the data gathered on mortality and
hospital admissions. For these endpoints, estimates of baseline incidence were derived from
the studies themselves to provide what should be viewed as only a rough estimate of
magnitude of potential effects, given the much greater degree of uncertainty concerning
baseline incidence information for these endpoints. The baseline health effects incidence data
are presented and discussed more fully in Section 6.
2.6. Sensitivity analyses
The predictions of the risk analyses depend on the input components discussed above.
Changes in the values of these input components change the predictions of the analyses. This
Abt Associates, Inc. p. 24 July 3, 1996, Revised
-------
is an important issue in risk analysis because the true values of parameters necessary for such
analyses, e.g., the location-specific concentration-response relationships and the location-
specific baseline health effects incidence rates, are often not known exactly and must be
estimated. The sensitivity of the results of an analysis to changes in the values of the input
components (or in assumptions or procedures that affect these values) is therefore an important
consideration.
The uncertainty associated with having to estimate parameter values can be assessed by
uncertainty analyses, in which the probability distribution of values for an input component is
estimated, and the resulting distribution of possible outcomes is assessed. Uncertainty analyses
to assess the uncertainty associated with key parameters of the risk assessment model, focusing
primarily on the concentration-response function, are discussed in Section 9.
Alternatively, sensitivity analyses can be used to illustrate the sensitivity of analysis
results to different possible input values or to different assumptions or procedures that may
affect these input values. Although a sensitivity analysis is not as comprehensive as an
uncertainty analysis, selecting only a few possible alternative values of an input component
rather than characterizing the entire distribution of these values, it is precisely the simplicity of
a sensitivity analysis that makes it preferable for illustrating the impact on results of using
different input component values. Exhibit 2.6 lists the sensitivity analyses carried out for each
of the two phases of the risk analysis. The results of those sensitivity analyses pertaining to
the first phase of the risk analysis (the "as is" analyses) are presented in Section 7; the results
of those sensitivity analyses pertaining to the second phase of the risk analysis (the alternative
standards analyses) are presented in Section 8.
Abt Associates, Inc. p. 25 July 3, 1996, Revised
-------
Exhibit 2.6. Sensitivity Analyses
Sensitivity analyses associated with the "as is" risk analyses:
1. Sensitivity analysis of the effect of alternative assumed background levels on predicted
health effects associated with "as is" PM (PM-10 and PM-2.5) above background.
2. Sensitivity analysis of the effect of using alternative "hockey stick" models on
predicted short-term exposure health effects associated with "as is" PM concentrations
above specified model outpoints.
3. Sensitivity analysis of the effect of alternative cutpoints (PM levels below which health
effects incidence is not considered) on predicted long-term exposure health effects
associated with "as is" PM above cutpoint.*
4. Sensitivity analysis of the effect of combining different averaging times in pooled
short-term exposure mortality concentration-response functions on predicted health
effects associated with "as is" PM-10 concentrations above background.*
5. Sensitivity analysis of the effect of using concentration-response functions for short-
term mortality from different individual studies on predicted health effects associated
with "as is" PM-10 and PM-2.5**
6. Sensitivity analysis of the effect of copollutants in the concentration-response model on
the predicted relative risk for a change in PM-10 concentration of 50 /xg/m3 and a
change in PM-2.5 concentration of 25
7. Sensitivity analysis of the effect of copollutants in the concentration-response model on
the predicted health effects associated with "as is" PM above background.
8. Sensitivity analysis of the effect of historical previous air quality on estimated
mortality associated with long-term exposure to PM-2.5.*
Sensitivity analyses associated with the alternative standards analyses:
9. Sensitivity analysis of the effect of different background levels on rollbacks required to
simulate attainment of alternative PM-2.5 standards.
10. Sensitivity analysis of the effect of different rollback methods to simulate attainment of
alternative PM-2.5 standards.*
*Sensitivity analysis done for Philadelphia County only. With the exception of sensitivity analysis number 6,
which is not specific to any location, other analyses were done for both Philadelphia County and Southeast Los
Angeles County (see Exhibit 7.5).
**A preliminary (unpublished) study of short-term exposure mortality has been conducted in Philadelphia for PM-
2.5 (Dockery et al., Abstr., 1996). The results of this study are compared with the results obtained by using the
pooled analysis function separately in Section 7. This study is not included, however, among the studies in this
sensitivity analysis and is not included in the main results because it is not yet published.
Abt Associates, Inc. p. 26 July 3, 1996, Revised
-------
3. Assumptions and Caveats
To carry any risk analysis to completion in the face of incomplete information, it is
necessary to make a variety of assumptions. The necessity of making simplifying assumptions
characterizes most scientific analyses, because analysis is usually performed with only limited
information. Some of the assumptions necessary in the risk analyses are assumptions generally
made in scientific analyses (for example, that the model used to describe the relationship
betw<;en variables does accurately describe this relationship). Other assumptions are specific
to these risk analyses. (Assumptions may be characterized instead as caveats: the validity of
the results of the analysis depend in part on the extent to which the underlying assumptions are
valid.)
The risk analyses discussed in this report are only as good as the inputs to the analyses
that is, the concentration-response functions, the air quality data, the health effects incidence
rates, and the population sizes. The quality of each component is stated as an assumption or,
alternatively, discussed as a caveat. Other assumptions/caveats concerning how each of the
three analysis components are used in the risk analyses are discussed below in turn. For many
of the uncertainties, it is not known whether the factors discussed might lead to over- or
under-estimates of risk. Exhibit 3.1 summarizes some of the key uncertainties in the risk
analysis, which are discussed in more detail below.
3.1. Concentration-response functions
The concentration-response function is a key element of risk assessment. The quality
of the risk analysis depends, in part, on (1) how well the concentration-response functions
used in the risk analyses have been estimated (e.g., whether they are unbiased estimates of the
relationship between the health response and ambient PM concentration in the study locations),
(2) how applicable these functions are to locations and times other than those in which they
were estimated, and (3) the extent to which these relationships apply beyond the range of the
PM concentrations from which they were estimated. These issues are discussed in the
subsections below.
3.1.1. Accuracy of the estimates of concentration-response functions
The adequacy of the estimation of the relationships between PM and various health
endpoints in epidemiological studies has received considerable attention. A significant portion
of this attention has focused on the issue of using average ambient PM concentration as a
measure of actual exposure to PM. Although they might prefer to estimate the individual
exposure-response relationship, such studies are actually estimating the concentration-response
relationship, as discussed in Section 1. Concern that this practice may produce biased
estimates of individual exposure-response relationships may be valid. However, because the
risk analysis examines the association between changes in health effects incidence and changes
in ambient PM concentrations, (ambient) concentration-response functions, rather than
Abt Associates, Inc. p. 27 July 3, 1996, Revised
-------
Exhibit 3.1. Key Uncertainties in the Risk Analysis
Uncertainty
Empirically estimated
concentration-response
relationships
Functional form of
concentration-response
relationship
Transferability of concentration-
response relationships
Extrapolation of concentration-
response relationships beyond
observed data range
Adequacy of PM
characterization
Accuracy of PM mass
measurement
Adjustment of air quality
distributions to reflect attainment
of proposed alternative standards
Baseline health effects data
Sensitive subgroups
Omitted effects
Direction of
Potential
Error
?
? -
?
+
?
7
?
7
?
-
Comments
Statistical association does not prove causation. Because
concentration-response functions are empirically estimated,
there is uncertainty surrounding these estimates. Omitted
confounding variables could cause upward bias.
Statistical significance of coefficients in an estimated
concentration-response function does not necessarily mean
that the mathematical form of the function is the best model of
the true concentration-response relationship.
Concentration-response functions may not be valid in times
and places other than those in which they were estimated.
A concentration-response relationship estimated by an
epidemiological study may not be valid at concentrations
outside the range of concentrations observed during the study.
Only particle mass per unit volume has been considered, and
not, for example, chemical composition or any other panicle
characteristics.
Possible differences in measurement error, losses of particular
components, and measurement method between the two risk
analysis locations and between these locations and the original
studies would be expected to add uncertainty to quantitative
estimates of risk.
There is uncertainty in the pattern and extent of reductions in
daily PM concentrations that would take place to attain
proposed standards.
Data may not be exactly appropriate for a variety of reasons.
For example, location- and age-group-specific baseline rates
may not be available in all cases. Baseline incidence may
change over time for reasons unrelated to PM.
Populations in the sample locations may have more or fewer
members of sensitive subgroups than locations in which
functions were derived. Thus functions might not be
appropriate. (This is a subset of the uncertainty associated
with transferring concentration-response functions from one
location to another (see above).
Some health effects caused by PM may have been omitted.
Abt Associates, Inc.
p. 28
July 3, 1996, Revised
-------
individual exposure-response functions, are relevant to the analysis discussed in this report.
The important question here, then, is whether epidemiological studies have produced accurate,
unbiased estimates of ambient concentration-response functions.
The accuracy of an estimate of a concentration-response function reported by a study
depends on the study design. The Criteria Document has evaluated the substantial body of PM
epidemiological studies. In general, critical considerations in evaluating the design of an
epidemiological study include the adequacy of the measurement of average ambient PM, the
adequacy of the health effects incidence data, and the consideration of potentially important
health determinants and causal factors such as:
copollutant ah-quality;
exposure to other health risks, such as smoking and occupational exposure;
demographic characteristics, including age, sex, socioeconomic status, and
access to medical care; and
population health status independent of PM air quality.
Other specific characteristics of concern depend on the health endpoints hi the studies. Study
selection for the risk analysis was guided by the evaluations in the PM Criteria Document
(EPA, 1996a).
Concentration-response functions may not be identical for all members or all subgroups
of a population; however, the concentration-response functions used in the risk analysis reflect
overall population responses at air quality levels similar to those found in the sample locations
(see Section 3.1.3).
To the extent that the studies did not address all critical factors, the concentration-
response functions may be limited. They may result in either over- or underestimates of risk
associated with ambient PM concentrations in the locations in which the studies were done. It
is possible, then, that their application to the sample locations in the risk analyses might also
have resulted in biased estimates of risk in those locations.
One possible source of bias in the estimation of concentration-response functions
warrants special note. A concentration-response function could be biased if the measurement
of average ambient PM concentration is inaccurate in a systematic way. Most epidemiological
studies use the average PM levels reported at some number of PM monitors as the measure of
the average ambient PM concentration (which, for the purposes of the studies, is itself a
surrogate for individual exposure to PM). This may or may not yield accurate measurements
of the actual daily average ambient PM concentrations in the study city. Depending on how
the monitors are placed, it could yield systematically inaccurate, or biased measurements.
What is important for the purpose of the risk analysis, in this case, is that the measurement of
daily average ambient PM concentrations in the sample location be biased in the same way as
in the study city. That is, a systematic bias in the measurement of daily average ambient PM
concentrations in the study city is not a problem for the risk analysis if there is the same
Abi Associates, Inc. p. 29 July 3, 1996, Revised
-------
systematic bias in the measurement of daily average ambient PM concentrations in the sample
location. Whether this is the case, however, is unlikely to be known.
3.1.2. Applicability of concentration-response functions in different locations
The method described in Section 2 combines PM data from a single year in a specific
location (the sample location) with a concentration-response relationship estimated by an
epidemiological study to predict the change in incidence of a given health endpoint associated
with a given change in ambient PM concentrations in that location. Preferably, this
concentration-response function is obtained from a study conducted in the sample location
itself. However, if no such study is available, a concentration-response function derived from
a study that was performed in a different location (or a pooled analysis concentration-response
function derived from several such studies) is used. The precision of these predictions
therefore rests in part on the "transferability" of the concentration-response relationship from
one location to another. That is, it rests on the assumption that the relationship between
ambient PM (either PM-10 or PM-2.5) and a given health endpoint is the same in the two
locations.
The relationship between average ambient PM concentration and the incidence of a
given health endpoint, the concentration-response relationship, depends on (1) the relationship
between individual exposure and average ambient PM concentration and (2) the relationship
between the health response and individual exposure (as described formally in Appendix 1).
One or both of these relationships may depend on the exposed population (for example, the
extent of susceptible subgroups) and/or the composition of the PM and other air quality
indicators to which the population is exposed. Both the population and the composition of PM
and other air quality indicators could vary significantly from one location to another. In this
case, the concentration-response relationship could vary significantly from one location to
another as well. There are various reasons why one or both of the relationships upon which
the concentration-response relationship depends might vary from one location to another.
The relationship between individual exposure and average ambient PM concentration °
might differ among locations if people's behavior patterns differ significantly from one
location to another. Suppose, for example, people in the study city spend a lot more time
outdoors than people in the sample location. Suppose also that ambient (outdoor)
concentrations of PM are greater in both locations than indoor concentrations. Then a given
ambient PM concentration will be associated with higher individual exposures among people in
the study location than among people in the sample location.
Suppose, alternatively, that coarse particles are less likely to infiltrate indoor air. In
this case, PM-10 with a high proportion of coarse particles would result in lower indoor
exposure than PM-10 with a high proportion of fine particles. If the percent of time spent
indoors is the same in different locations, then the location with the coarser PM-10 would have
lower individual exposure to PM-10 than the location with the finer PM-10.
Abt Associates, Inc. p. 30 July 3, 1996, Revised
-------
The relationship between the health response and individual exposure might differ
among locations if, for example, the population in one location has a higher proportion of a
susceptible subgroup than another location. Closely matching populations used in studies to
the populations of the sample locations is not possible for many characteristics (for example,
smoking status, workplace exposure, socioeconomic status, and the prevalence of highly
susceptible subgroups).
Alternatively, the PM-10 in one location may be largely fine particles (PM-2.5),
whereas in another location it may be made up predominantly of coarse particles. If PM-2.5 is
more potent than coarse particles in causing the health effect, then there will be a greater
incidence of the health effect corresponding to a given level of individual exposure to PM-10
in the first location, all else equal, than in the second location (see Appendix 4).
Other pollutants, such as carbon monoxide and ozone, may also play a role in causing
health effects, either independently or in combination with PM. Interlocational differences in
these pollutants could also induce differences in the concentration-response relationship
between one location and another.
In summary, the concentration-response relationship in one location may not be the
same as the concentration-response relationship in another location. Even if the relationship
between the health response and individual exposure is the same in both locations, the
relationship between individual exposure and ambient concentrations may differ between the
two locations. Similarly, even if the relationship between individual exposure and ambient
concentrations is the same in both locations, the relationship between the health response and
individual exposure may differ between the two locations. In either case, the concentration-
response relationship would differ.
Instead of a single concentration-response function that characterizes the relationship
between ambient PM (either PM-10 or PM-2.5) and a given health endpoint everywhere in the
United States, then, a more realistic model may be a distribution of concentration-response
functions, or, equivalently, a distribution of values of the parameter p in the concentration-
response function. If concentration-response functions were available for all health endpoints
for each of the sample locations, the precision of the risk analyses would be improved. The
necessity of applying concentration-response functions estimated in locations other than the
sample locations for which risk is being analyzed introduces uncertainty into the results of the
risk analyses. This is particularly apparent in the case of mortality, for which there are many
estimated concentration-response functions. The degree of this uncertainty is assessed by
Monte Carlo methods, as described in Section 9.
The uncertainty associated with the application of concentration-response functions
from epidemiological studies to the sample locations is nonetheless smaller than the uncertainty
associated with the use of concentration-response functions from animal studies. Such studies
Abt Associates, Inc. p. 31 My 3, 1996, Revised
-------
are often used in risk assessments, especially when epidemiological results are not available,
but are not used in the risk analysis described in this report.
3.1.3. Extrapolation beyond observed air quality levels
Although a concentration-response function describes the theoretical relationship
between ambient PM and a given health endpoint for all possible PM levels (down to zero),
the estimation of a concentration-response function is based on real ambient PM values that are
limited to the range of PM concentrations in the location in which the study was conducted.
The actual shape of the concentration-response function is not known outside the observed air
quality range. Several of the mortality studies discussed in the Criteria Document (EPA,
1996a), including Pope et al.(1992), reported measured PM-10 levels as low as 4 /ig/m3.
Nonetheless, the concentration-response relationship may be less certain towards the lower end
of the concentration range if few days had such low concentrations.
The risk analyses assume that the estimated concentration-response functions
adequately represent the true concentration-response relation down to background levels in the
sample locations, in cases in which this background level is above the lowest concentrations
used to derive the concentration-response functions. For studies in which the lowest
concentrations studied are likely to be above background (e.g., the long-term exposure study
of Pope et al. 1995) estimates of risk are not generated for concentrations below the minimum
concentrations observed in the studies. The estimates of risk for the lowest concentrations
considered are more uncertain than the estimates for concentrations in the middle of the range
of study data.
The concentration-response relationship may also be less certain towards the upper end
of the concentration range being considered in a risk analysis, particularly if the PM
concentrations in the sample location exceed the PM concentrations observed in the study
location. Even though it may be reasonable to model the concentration-response relationship
as exponential over the ranges of PM concentrations typically observed in epidemiological
studies, it is unlikely to be exponential over a very wide range of PM levels.7 Rather, at very
high PM concentrations, the concentration-response function is likely to begin to flatten out.
In a location such as Los Angeles, where pollution levels are generally higher than average in
the United States, and possibly substantially higher, it is possible that some PM concentrations
fall in the range in which the exponential function is no longer appropriate. To the extent that
this is the case, it would contribute to overestimation of PM-related health effects incidence in
such a location.
The standards that EPA has chosen to evaluate in the second phase of the risk analysis
lie in the middle range of pollution levels observed in epidemiological studies. Applying
'Although the concentration-response functions are exponential, they are practically linear. It is still
unlikely, however, that a linear function is appropriate over a very wide range of PM concentrations.
Abi Associates, Inc. p. 32 July 3, 1996, Revised
-------
uniform linear rollbacks to the concentration distributions in the sample locations, however,
will result in some modeled PM concentrations well below these levels. It is possible that
there is a minimum concentration below which PM is not associated with health effects. To
the extent that reducing concentrations below such a concentration is counted as reducing PM-
related health effects, the health benefits attributed hi the risk analyses to reducing PM are
overestimated. The degree of overestimation depends on the frequency of modeled PM
.concentrations that are lower than the lowest level at which effects actually occur, and how
much lower the modeled PM concentrations are. Sensitivity analyses address the sensitivity of
the results of the risk analysis to different assumptions about the minimum concentration at
which health effects occur.
3.2. The air quality data
3.2.1. Appropriateness of the PM indicator
The ozone ah" quality risk analysis modeled people's activity patterns and their
resulting exposure to different ozone micro-environments. This was of interest because
controlled experiments have measured people's reactions to varying ozone concentrations. No
such controlled experiments for PM were used in the PM risk analyses. Instead, the PM risk
analysis used estimates of average ambient air quality in a given location as measured at
monitors. This matches the measure of exposure used in PM epidemiology studies.
PM is measured hi units of mass per unit volume, typically in micrograms per cubic
meter, rather than in density of particles (i.e., in parts per million). The only distinction made
between different kinds of particles in the risk analysis is one of size: PM-2.5 vs. PM-10.
This may not be the only distinction of interest. The chemical composition of PM, for
example, was not considered in any of the risk analyses (as it was not in most of the
epidemiological studies used in these analyses).
3.2.2. Adequacy of PM air quality data
The method of averaging data from monitors across a metropolitan area in the risk
analysis is similar to the methods used to characterize ambient air quality in most of the
epidemiology studies. The important issue, however, is whether any biases in the
measurement of average daily ambient PM concentrations in the study location are matched in
the sample location, as discussed in Section 3.1.1. Ideally, the measurement of average daily
ambient PM concentrations in the study location are unbiased. In this case, unbiased risk
predictions in the sample location depend, in part, on an unbiased measurement of average
daily ambient PM concentrations in the sample location as well. If, however, the
measurement of average daily ambient PM concentrations in the study location are biased,
unbiased risk predictions in the sample location are still possible if the measurement of average
daily ambient PM concentrations in the sample location incorporate the same bias as exists in
Abr Associates, Inc. p. 33 July 3, 1996, Revised
-------
the study location measurements. Because this is not known, however, the adequacy of the
PM measurements in the sample locations is a source of uncertainty in the risk analysis.
PM air quality data are not available for all days of the year chosen for risk analysis in
either of the sample locations. The change in the incidence of a health effect over the course
of the year corresponding to a given change in daily PM levels is calculated based on the
assumption that PM levels on those days with PM data are representative of levels on those
days without PM data (see Section 2 for an explanation of the method of extrapolating changes
in health effects incidence to an entire year). Where available concentration data are evenly
distributed throughout the year, the extrapolation can be performed in a single step. Where
available concentration data are unevenly distributed through the year, results from different
parts of the year are scaled separately, and the results added. This avoids bias due to seasonal
differences hi average PM levels and monitoring frequencies.
Because the PM data in each sample location are limited to a specific year, the results
of the risk analyses are generalizable to the present only to the extent that ambient PM levels
in the available data are similar to current ambient PM levels. A substantial difference
between PM levels in the years used in the risk analyses and current PM levels could imply a
substantial difference in predicted incidences of health effects. This is not expected to be a
large problem, however, because adequate PM-10 and PM-2.5 monitoring data are available
for Philadelphia County for 1992-1993 and for Southeast Los Angeles County for 1995.
3.3. Baseline health effects incidence rates
3.3.1. Quality of incidence data
Local incidence data were obtained for mortality and for hospital admissions for both
Philadelphia and Los Angeles (see Section 6). This is clearly preferable to using nonlocal
data, such as national incidence rates. As with any health statistics, however, misclassification
of disease, errors in coding, and difficulties in correctly assigning residence location are
potential problems. These same potential sources of error are present in most epidemiological
studies. In most cases, the reporting institutions and agencies utilize standard forms and codes
for reporting, and quality control is monitored.
When national rates are used, the estimated rates are generally considered reliable, due
to the large sample size available. As the source population becomes smaller and the event
rarer, the reliability may decrease, due the infrequency of occurrence. Most endpoints
considered in this report are common occurrences and the locations are sufficiently large,
however, that the statistics reported are likely to be adequately representative of the
occurrences, even though the data are limited to one year.
Incidence rates for some health endpoints (in particular, for lower respiratory
symptoms) were obtained from the studies reporting the concentration-response functions.
Abt Associates, Inc. p. 34 July 3, 1996, Revised
-------
There is greater uncertainty in the application of incidence data obtained from specific studies
to locations across the country, because the rates are specific to a certain location, time, and
cohort. Where possible, baseline incidence rates were obtained for the age cohorts matching
the populations studied. In addition, some rates taken from studies (for example, lower
respiratory symptoms from Pope et al. 1991) are age-specific. Since actual rates for some
endpoints (including lower respiratory symptoms) are known to vary with age, rates for some
age cohorts may not be accurately represented. For example, it is likely that lower respiratory
symptom rate? are underestimated for young children.
Regardless of the data source, if actual incidence rates are higher than the incidence
rates used risks will be underestimated. If incidence rates are lower than the incidence rates
used, then risks will be overestimated. For most of the concentration-response functions, the
incidence rates affect the estimation of the changes in the number of cases associated with
changes in PM, but not the estimation of the percentage changes in PM-related cases. The
uncertainties in identifying the correct baseline incidence rates therefore affect only one
portion of the results.
Both morbidity and mortality rates change over time for various reasons. One of the
most important of these is the age distribution of the population. The old and the extremely
young are more susceptible to many health problems than is the population as a whole. The
most recent available data were used in the risk analysis. However, the average age of the
population in many locations will increase as the post-WWII children age. Consequently, the
baseline incidence rates for some endpoints may rise, resulting in an increase in the number of
cases attributable to any given level of PM pollution. Alternatively, areas which experience
rapid in-migration, as is currently occurring in the south and west, may tend to have a
decreasing mean population age and corresponding changes in incidence rates and risk.
Although temporal changes in incidence are relevant to both morbidity and mortality
endpoints, however, the most recent available data were used in all cases, so temporal changes
are not expected to be a large source of uncertainty.
3.3.2. Lack of daily health effects incidence rates
Both ambient PM levels and the daily health effects incidence rates corresponding to
ambient PM levels vary somewhat from day to day. Those risk analyses based on
concentration-response functions estimated by short-term exposure studies calculate daily
changes in incidence and sum them over the days of the year to predict an annual change in
health effect incidence.
Most of the concentration-response functions calculate relative risk, that is, percent
change in health effects, which depends only on the change in PM levels (and not the actual
value of either the initial or final PM concentration). This percent change is multiplied by a
baseline incidence in order to determine the change in health effects incidence. That is,
Abt Associates, Inc. p. 35 July 3, 1996, Revised
-------
A incidence = baseline incidence (PM) * RR(&PM),
where the relative risk (RR) depends only on the change in PM levels (APM), not the actual
values, and the baseline incidence may depend on the actual PM concentrations.
If PM does indeed affect health, then actual incidence rates can be expected to be
somewhat higher than average on days with high PM concentrations, and somewhat lower than
average on days with low PM concentration. However, only annual average incidence rates
are available from vital statistics sources. Daily incidence rates corresponding to "as is" daily
PM levels were therefore approximated by the average daily incidence rates, calculated from
the annual figures:
A incidence - average baseline * RR(&PM).
The annual average baseline incidence rate is expected to be lower than the actual baseline
incidence rate on days with high PM, and so the predicted change in incidence will also be
lower than it should be. Similarly, the annual average baseline incidence is expected to be
higher than the actual baseline incidence rate on days with low PM, and so the predicted
change in incidence will also be higher than it should be. The change in health effects
incidence may therefore be slightly underestimated on days with high PM levels and slightly
overestimated on days with low PM levels. Both effects are small, however, and should
largely cancel one another.
3.4. Further caveats
3.4.1. Highly susceptible subgroups
Highly susceptible subgroups, such as asthmatics and people with cardiovascular or
pulmonary disease, are of particular concern with regard to PM pollution. These groups
comprise, presumably, a substantial portion of hospital admissions, deaths, and morbidity
counts enumerated in this risk analysis. To the extent that a location has a larger (smaller)
proportion of people in these groups than the location in which a concentration-response
function was estimated, however, risk may be underestimated (overestimated). This is one of
the reasons that concentration-response functions may not be transferable, as discussed in
Section 3.1.2. In addition, the health effects experienced by a susceptible subgroup may be
much greater than those experienced by the population at large. Given the lack of data on the
representation of various potentially susceptible subgroups in specific locations, this question
cannot be addressed with certainty given available data.
Abt Associates, Inc. p. 36 July 3, 1996, Revised
-------
3.4.2. Possible omitted health effects
Although this risk analysis method considers both mortality and a variety of morbidity
health effects, it does not include all health effects which may result from PM exposure. Only
a subset of those endpoints that have been the subject of quantitative epidemiological studies
are enumerated. Other possible health effects reported to be associated with short-term
exposures to PM-10 include emergency room visits for asthma (Schwartz et al. 1993),
respiratory hospitalization in children (Pope et al. 1991), school absences (Ransom and Pope
1992), symptoms of cough (Schwartz et al. 1994; Ostro et al. 1991), and asthma medication
usage (Pope et al. 1991). Other possible health effects reported to be associated with short-
term exposures to PM-2.5 include respiratory-related restricted activity days and work loss
days in adults (Ostro et al. 1987). Health effects that have been associated with long-term
exposures to PM-10, not included in the risk analysis, are chronic bronchitis in adults (Abbey
et al. 1995a) and decreased lung function in children (Raizenne et al. 1996). Other possible
effects of concern include cardiovascular and respiratory episodes and diseases not measured
in the hospitalization studies, and effects in those under 65 for health effects for which only
studies on those over 65 are available. The omission of some endpoints may lead to an
underestimate of total risk to the population.
Abt Associates, Inc. p. 37 July 3, 1996, Revised
-------
Abt Associates, Inc. p. 38 July 3, 1996, Revised
-------
4. Air Quality Assessment: The PM Data
This section describes the PM-10 and PM-2.5 data for Philadelphia County and
Southeast Los Angeles County used in the risk analysis. Average ambient PM-10
concentration in a sample location on a given day is represented by the average of reported
PM-10 levels at the different monitors in that location on that day. The same approach is used
for PM-2.5. This approach is consistent with what has been done in epidemiological studies
estimating PM concentration-response functions. Also, because people are often quite mobile
(e.g., living hi one part of a city and working hi another), a city-wide average PM level may
be a more meaningful measure of ambient PM concentration than PM levels at individual
monitors. Ito et al. 1995 found that averaging PM-10 concentrations reported at monitors hi
different places generally improved the significance of the association between PM-10 and
mortality in Chicago, compared with using individual monitors.
Frequency graphs of average daily PM levels hi each of the sample locations are shown
for both PM-10 and PM-2.5 for Philadelphia County hi Exhibit 4.1, and for PM-10 and PM-
2.5 in Southeast Los Angeles County hi Exhibits 4.2 and 4.3, respectively.
4.1. The Philadelphia County PM data
Air pollution data were collected in Philadelphia County by the Harvard School of
Public Health (HSPH) Exposure Assessment and Engineering Program for the EPA's
Atmospheric Research and Exposure Assessment Laboratory. Each monitor recorded
pollution levels for PM-10 and PM-2.5. No monitor gave information on every day. Exhibit
4.4 lists the monitors used. Exhibits 4.5 and 4.6 show the number of days on which PM-10
and PM-2.5 concentration data were available at each monitor. Concentration data were
available almost every day in Philadelphia County (as compare with the situation in Southeast
Los Angeles County, described below). Exhibits 4.7 through 4.9 summarize the reported PM-
10 and PM-2.5 concentrations at the three monitors used, and for a composite monitor
assumed to report on each day the average of any concentrations reported by the three
monitors.
Because not all monitors report PM concentrations on all days, the effect of estimating
missing concentrations was explored in a previous analysis8 of a data set including monitors
with many more days missing than those used in this analysis. Adding estimated values where
there had been missing values proved to have virtually no effect on the distribution of PM
concentrations. Therefore, no such estimation was attempted in this analysis, and Philadelphia
air quality is represented on each day by the average of any concentrations reported by the
three monitors. The method used to adjust health effect incidence estimates from those
corresponding to the number of non-missing days in the year to estimates corresponding to a
full year is described in Section 2.
8 Proposed Methodology for PM Risk Analyses in Selected Cities. Abt Associates. Febmary 1996.
Abt Associates, Inc. p. 39 July 3, 1996, Revised
-------
Exhibit 4.1
Daily Average PM Concentration Frequencies
Philadelphia County, September 1992 - August 1993
0.3
0.25
2- °-2
£ 0.15
.1
£ 0.1
0.05
0
PM-10
Data Available on 358 Days
|bmwidth=5ugta3 [
I
I
J
II. 1..
0 20 40 60 80 100 120 140 160 180 200
24-hour Average PM-10 Concentration
0.3
0.25
g- 0.2
c
3
tu
il 0.15
or 0.1
0.05
0
PM-2.5
Data Available on 353 Days
(bin width = 5xigAn3 I
i
J
I
t
ltin . .
3 20 40 60 80 100 120 140 160 180 200
24-hour Average PM-2.5 Concentration
Abi Associates, Inc.
p. 40
July 3, 1996, Revised
-------
Exhibit 4:2
Daily Average PM-10 Concentration Frequencies
Southeast Los Angeles County, 1995
£ 0^
H
%
> 2
1st Quarter
DaoAvaibOlaa-ilODa^
b» tndik - S »(«> 1
MILL
lilt
5 40 60 8O 10O 120 14O 160 1«0 20
24J
0 1
2nd Quarter
Data Awaibbb on 27 Day*
Ikumdtb -Suitoil
1
-- 4-
i Ir
J 2O 40 SO » 100 l» 140 WO 1W 2&
2tl>our Avmge W-ICCcrcwwUo'
0
3rd Quarter
Daa Available on S3 Day*
O.J
II l
liT .
0 20 40 60 80 100 120 140 160 180 200
4th Quarter
Data Auaibtik on 89 Day*
_ ll I
ll.lllllllllll lll.l .l.ll I. . ..
0 » 40 60 «O 100 120 14O 160 1*0 200
47
y«/y 3, 1996, Revised
-------
Exhibit 4.3
Daily Average PM-2.5 Concentration Frequencies
Southeast Los Angeles County, 1995
1st Quarter
DaaAuaibMtonttDay*
0.)
i"
i
0.1
0
0
|k-^^-Su,^|
20
.
40 60 «0 100 120 140 160 1SO 20O
OJ
|02
i
i
0.1
0
2nd Quarter
0*aA»ibbleoq27Day*
J
I
|b>. «,4ik -i*fmi\
I
I
1
» 20 40 6O «O 100 120 14O 160 1SO 20O
24-ti«ug« >U.2£Ccfic«fi» 160 180 20'
bwour M«jg« ' V-2S Ca
-------
Exhibit 4.4. Philadelphia County Monitor IDs and Locations
Monitor Code
N/E
PBY
TEM
AIRS ID (if AIRS monitor)
421010037
Location
Philadelphia NE Airport: Grant/ Ashton Roads
59th St. and Greenway Ave.
13th St. and Montgomery Ave.
Exhibit 4.5. Number of Days on which PM-10 Concentration Data are Available, by
Quarter. Philadelphia, September 1992 - August 1993.
N/E
PBY
TEM
Composite
Ql
52
88
47
87
Q2
87
88
83
91
Q3
77
90
70
92
Q4
35
84
31
88
Year Total
251
350
231
358
Exhibit 4.6. Number of Days on which PM-2.5 Concentration Data are Available, by
Quarter. Philadelphia, September 1992 - August 1993.
N/E
PBY
TEM
Composite
Ql
52
81
43
84
Q2
81
84
89
91
Q3
76
87
71
90
Q4
34
86
35
88
Year Total
243
338
238
353
Exhibit 4.7. PM-10 Concentrations by Quarter
Philadelphia, September 1992 - August 1993
N/E
PBY
TEM
Composite
Ql
15.7
18.6
18.6
18.1
Q2
21.3
24.3
25.0
23.7
Q3
26.2
29.0
30.3
27.1
Q4
18.7
19.6
23.9
20.1
Weighted
Year Avg.
20.5
22.9
24.5
22.2
2nd Daily
Max
70.7
72.3
76.8
67.2
All concentrations are in /xg/m3.
Abt Associates, Inc.
p. 43
July 3, 1996, Revised
-------
Exhibit 4.8. PM-2.5 Concentrations, by Quarter
Philadelphia, September 1992 - August 1993
N/E
PBY
TEM
Composite
Ql
12.9
14.3
13.6
13.8
Q2
16.3
17.6
17.3
17.0
Q3
20.0
21.8
22.0
20.9
Q4
13.2
14.2
15.6
14.1
Weighted
Year Avg.
15.6
17.0
17.1
16.5
2nd Daily
Max
65.1
72.6
70.0
69.3
All concentrations are in /ig/m3.
Exhibit 4.9. Percentile Points of Composite Distribution
Philadelphia, September 1992 - August 1993
PM-10
PM-2.5
10th
10.4
7.2
25th
14.2
9.7
50th
19.7
14.0
75th
28.2
21.0
90th
44.2
29.5
Max
72.6
69.8
All concentrations are in /*g/m3.
4.2. The Southeast Los Angeles County PM data
Two California Air Resources Board monitors in Los Angeles County, designated
Central Los Angeles (CELA) and Diamond Bar (DEAR) were selected to represent air quality.
A portion of southeastern Los Angeles County selected for use in the analysis includes the
portion of the county with the highest PM-10 levels. The region included in this analysis
approximates the portion of the county reported to have an annual average PM-10 level above
40/xg/m3 in 1994 (from "Air Quality Standards Compliance Report," South Coast Air Quality
Management District, 1995). The two monitors reported concentrations infrequently during the
first quarter of 1995, somewhat more frequently in the second quarter of 1995, and almost
every day during the second and third quarters. Exhibits 4.10 and 4.11 show the number of
days for which data were available in each quarter. Exhibits 4.12 and 4.13 show the average
concentration reported in each quarter, as well as a weighted year average (weighted by
quarter, as per the 1987 Federal Register notice, Vol. 52, No. 126, p. 24667 (July 1, 1987),
and as used for annual average data reported in AIRS) and the second-highest reported value.
Percentile points of the distributions are not provided, because the unequal distribution of
monitor-days among quarters, even for the composite monitor, provides a skewed picture of
air quality. All of the tables include statistics for the CELA and DEAR monitors, as well as
for a composite monitor, assumed to report on each day the average of whatever
concentrations were reported by the CELA and DEAR monitors.
Abt Associates, Inc.
p. 44
July 3, 1996, Revised
-------
Exhibit 4.10. Number of Days on which PM-10 Concentration Data are Available, by
Quarter. Southeast Los Angeles County, 1995.
CELA
DBAR
Composite
Ql
12
10
13
Q2
26
26
27
Q3
83
82
86
Q4
85*
81
89
Year Total
206
199
215
*One concentration, on October 30, was omitted as an obvious error.
Exhibit 4.11. Number of Days on which PM-2.5 Concentration Data are Available, by
Quarter. Southeast Los Angeles County, 1995.
CELA
DBAR
Composite
Ql
12
9
13
Q2
26
26
27
Q3
80
82
85
Q4
83
84
89
Year Total
201
201
214
Exhibit 4.12. PM-10 Concentrations by Quarter
Southeast Los Angeles County, 1995
CELA
DBAR
Composite
Ql
36.4
30.4
32.6
Q2
45.8
40.0
42.4
Q3
52.0
43.2
47.5
Q4
72.8
72.2
72.4
Weighted
Year
Average
51.7
46.4
48.7
2nd Daily
Max
195.2
170.7
19^.4
All concentrations are in
Exhibit 4.13. PM-2.5 Concentrations by Quarter
Southeast Los Angeles County, 1995
CELA
DBAR
Composite
Ql
21.6
18.5
20.6
Q2
26.0
20.2
23.3
Q3
27.0
23.1
25.1
Q4
45.6
47.6
45.8
Weighted
Year Average
30.1
27.3
28.7
2nd Daily
Max
115.7
129.3
106.2
All concentrations are in /xg/m3.
Abt Associates, Inc.
p. 45
July 3, 1996, Revised
-------
Abt Associates, Inc. p. 46 July 3, 1996, Revised
-------
5. Concentration-Response Functions
5.1. Concentration-response functions taken from the literature
Study selection decisions are among the most important subjective decisions that must
be made in a risk analysis. These judgements include decisions on which endpoints to
consider quantitatively, as well as decisions about which of the available scientific studies
should be used for quantitative risk analysis. The concentration-response functions reported
by epidemiological studies are estimates of the relationships between PM-10 and PM-2.5 and a
variety of health endpoints. The choice of functions for use in the risk analysis was guided by
the PM Criteria Document (EPA 1996a, Tables 13-3 to 13-5).
The studies highlighted by the CD, and those used in this risk analysis to derive
quantitative estimates of risk, used PM-10 or PM-2.5 mass (or other fine particle indicators) as
their indicator of PM. This eliminated, for example, the extensive studies of air pollution and
mortality conducted in Philadelphia (e.g., Health Effects Institute 1995, Moolgavkar et al.
1995a, Schwartz and Dockery 1992). In addition, two studies were omitted because other
studies in the same locations were considered more appropriate for use in this risk analysis.
Styer et al., 1995 estimated a concentration-response function for mortality in Chicago only
for autumn; Ito and Thurston 1996 estimated a concentration-response function for mortality in
Chicago for the whole year, and was therefore preferable. Dockery et al. 1992 (St. Louis and
East Tennessee) was superseded by Schwartz et al. 1996, which considered much larger data
sets in the same locations. Finally, for some studies of respiratory symptoms, the definitions
of cough and lower respiratory symptoms can overlap; thus for risk analyses using these
studies only lower respiratory symptoms were evaluated.
The health endpoints and the epidemiology studies used to estimate concentration-
response functions for these endpoints that have been considered in the risk analysis are
summarized in Exhibit 5.1.
As can be seen in Exhibit 5.1, most of the concentration-response functions were not
estimated in either Philadelphia or Los Angeles. If the concentration-response relationship for
a given combination of health endpoint and PM indicator were the same everywhere, then a
concentration-response function estimated in one location could be applied to another location,
and the only uncertainty would be from the usual sampling error associated with any estimate.
There is no reason to believe, however, that the concentration-response relationship is the
same everywhere. If a concentration-response function has not been estimated for the location
of interest (e.g., Philadelphia County or Southeast Los Angeles County), this presents the
problem of how to best estimate the concentration-response relationship in that location, based
on information from other locations. Even if a concentration-response function has been
estimated in the location of interest, the estimate of that location-specific function may be
improved by incorporating information from other locations into the estimate. The more
sampling error there is around the location-specific
Abt Associates, Inc. p. 47 July 3, 1996, Revised
-------
Exhibit 5.1. Studies of the Health Effects Associated with Participate Matter Pollution
Used in the Risk Analysis
Endpoint
Short-Term Exposure
Mortality:
Long-Term Exposure
Mortality:
Study
Ito and Thurston, 1996
Schwartz et al., 1996a
Pope et al. 1992
Schwartz, 1993
Kinney et al . 1995
Schwartz etal., 1996a
Pope et al. 1995
City /Location
Chicago, IL
Six Cities
Utah Valley, UT
Birmingham, AL
Los Angeles, CA
Six Cities
51 U.S. Cities
PM Indicator
PM-10
PM-10
PM-10
PM-10
PM-10
PM-2.5
PM-2.5
Hospital Admissions:
Respiratory Disease:
COPD:
Pneumonia:
Ischemic Heart Disease:
Congestive Heart Failure:
Respiratory Symptoms:
Schwartz et al. 1995
Schwartz et al. 1995
Schwartz, 1996
Thurston et al. 1994
Schwartz, 1994a
Schwartz, 1994b
Schwartz, 1994c
Schwartz, 1996
Schwartz, 1994a
Schwartz, 1994b
Schwartz, 1994c
Schwartz, 1996
Schwartz & Morris, 1995
Schwartz & Morris, 1995
Dockery et al., 1989
Pope et al. 1991
Ostroetal., 1995
Tacoma, WA
New Haven, CT
Spokane, WA
Ontario, CA
Birmingham, AL
Detroit, MI
Minneapolis, MN
Spokane, WA
Birmingham, AL
Detroit, MI
Minneapolis, MN
Spokane, WA
Detroit, MI
Detroit, MI
Six Cities
Utah Valley, UT
Los Angeles, CA
PM-10
PM-10
PM-10
PM-2.5*
PM-10
PM-10
PM-10
PM-10
PM-10
PM-10
PM-10
PM-10
PM-10
PM-10
PM-10
PM-10
PM-10
*Thurston et al.(1994) reports both a PM-10 and a PM-2.5 coefficient. In a later paper (Thurston and Kinney,
1995), however, the authors interpret their findings as "clear that the FP [fine particle] portion of the mass
(including particle strong acidity, H+) is driving the apparent relationships seen for the PM-10 and TSP metrics.
The risk analysis therefore uses only the PM-2.5 results from Thurston et al. (1994).
Abi Associates, Inc.
p. 48
Julv 3, 1996, Revised
-------
estimate, the greater will be the improvement achieved by incorporating information from
other locations. This is discussed below.
5.2. Estimation of a distribution of P's, estimation of P in any given location,
and characterization of the uncertainty surrounding that estimate
The concentration-response function is an important component of the risk analyses and
a source of substantial uncertainty in those analyses. The exponential concentration-response
function (equation (1), Section 2) commonly assumed in the epidemiological literature on
paniculate matter pollution health effects and used as the basis for the risk analyses implies
that the relationship between a given change in PM concentration, Ax, and the corresponding
change in the health endpoint, Ay,
A '* ij&X 1 T
(see Section 2 and Appendix 3), depends critically on the value of p. The larger the value of
P, the greater the change in the health effect associated with a given change in PM
concentration. For ease of discussion, the health endpoint is taken to be mortality. However,
the discussion below applies to any health endpoint.
It is possible that there is only a single value of p, that is, that the concentration-
response relationship between PM and mortality is the same everywhere. If this is the case,
different estimates of P reported by different epidemiological studies are all estimates of the
same underlying parameter and differ from each other only because of sampling error.
A more general and a more plausible model, however, is that there is not just a single
concentration-response relationship between PM and mortality, but that this relationship varies
from one location to another, depending on such factors as the composition of the PM and the
composition of the exposed population. Even if the form of the concentration-response
function is the same everywhere, the value of p may change from one location to another,
reflecting differences in these factors. For example, it may be the case that in Philadelphia
County,
Ay = y[e ' - 1] ,
whereas in Los Angeles,
Ay = y[e 2 - 1] .
with P, * P2.
Abt Associates, Inc. p. 49 July 3, 1996, Revised
-------
The concentration-response relationship between PM and mortality (for example,
throughout the United States) may be characterized, then, by a distribution of P's. For any
given interval of possible values of p, this distribution describes the probability that p (and
therefore the concentration-response relationship) at any particular location is within that
interval.9-10
If there is an underlying distribution of P's, then differences in reported estimates of p
among studies carried out on a single population in a single location (using identical averaging
times, methodology, etc.) would reflect only sampling error, because all such studies are
estimating the same p. Differences in reported estimates among studies carried out on
different populations in different locations, however, may also reflect differences in the P's
being estimated. There are, then, two potential sources of variability among concentration-
response estimates:
(1) within-study variability, or sampling error (so that even two studies estimating
the same P are likely to report different estimates), and
(2) between-study variability derived from the fact that studies may be estimating
different underlying parameter values,
and associated with these two sources of variability is uncertainty about the correct
concentration-response function for a given location.
If the underlying distribution of P's were known, this distribution could be used to
characterize the uncertainty surrounding an estimate of P applied to a given location (in the
absence of an estimated concentration-response function for that location). Suppose, for
example, that p005 is the 5th percentile of the distribution of P's and P095 is the 95th percentile.
Then, for a randomly selected location (e.g., for a sample location), there is a 90 percent
probability that P in that location falls within the interval (p005, PO.PS)- The distribution of P's
thus allows uncertainty bounds to be associated with any estimate of P applied to a particular
location.
'This is technically referred to as a probability density function, but will be called a distribution here for
simplicity.
'°Since the concentration-response relationship (P) in a given location is uncertain, the mortality predicted
using any given concentration-response relationship is also uncertain. Corresponding to the underlying
distribution of P's, there is a distribution of PM-related mortality, describing the probability that the incidence of
mortality associated with a given PM level falls within a given interval.
Abt Associates, Inc. p. 50 July 3, 1996, Revised
-------
5.2.1. Estimation of the distribution of P's
The distribution of p's is not known, however, and must be estimated. Once the
distribution is estimated, the interval from the 5th percentile to the 95th percentile of the
estimated distribution, referred to as the "90 percent credible interval," is used to characterize
the uncertainty associated with P in any location for which a concentration-response function
has not been estimated (and for which there is therefore no information more specific than the
general distribution of P's).
If only a single study has estimated a concentration-response function for a given
endpoint, then the only available information about the distribution of P's comes from that
study. Within the general case of n studies reporting P's, this is just the special case in which
n= 1. The discussion below refers to the general case of n studies (where n may be greater
than or equal to 1).
If each study were reporting the true concentration-response function for the location
studied, then the set of reported P's would be a sample from the underlying distribution of P's
and could therefore be used to help estimate this distribution. What each study reports,
however, is an estimate of the concentration-response function (or, equivalently, an estimate of
P) for the location studied. The reported P for each study location therefore has some
sampling error associated with it.
Under the assumption that the true P's in the various study locations are all drawn from
the same distribution of p's, an estimate of p for a given study location that uses information
from all the study locations is generally better than an estimate that uses information from only
the given study location (see, for example, Efron and Morris, 1973; Laird and Ware, 1982;
and Laird and Ware, 1984).u That implies that the estimates of P reported for each study
location can be improved upon. Suppose, for example, that the concentration-response
function for PM-10 and mortality has been estimated in locations A, B, C, and D. Let MLEA
denote the maximum likelihood estimate of P in location A, MLEB the maximum likelihood
estimate of p in location B, and so on. Let pooled(A,B,C,D) denote a pooled estimate of the
concentration-response function, pooling the concentration-response functions in locations A
through D. Then a weighted average of MLEA and pooled(A,B,C,D) is a better estimator of P
in location A than MLEA, and similarly for locations B, C, and D.
A good estimate of the distribution of P's, then, relies on first adjusting the P's
estimated in individual study locations by incorporating information from the other study
locations (if there is more than one study location). The estimation of the distribution of P's
for a given combination of health endpoint and PM indicator is a three-step procedure which
"Estimator A is better than estimator B if the mean squared error associated with estimator A is less than
the mean squared error associated with estimator B.
Abt Associates, Inc. p. 51 July 3, 1996, Revised
-------
efficiently uses the available information as follows:
In step 1, a random effects pooled estimate of p is calculated, using the P's reported by
those studies that estimated a concentration-response function for the given combination of
health endpoint and PM indicator. The random effects pooled estimate is based on the
assumption that there is an underlying distribution of P's, with variance tj2. The pooled
estimate is a weighted average of the reported P's (which are assumed to be estimates of P's
drawn from the distribution). The weights are a function of both the sampling error (the
within-study variability) and r\2 (the between-study variability). The calculation of the weights
is described in Appendix 2. The pooled estimates, calculated for all health endpoint PM
indicator combinations for which there is more than one study, are given in Section 5.2.2.
In step 2, P in each study location is re-estimated, using a weighted average of the P
reported for that location and the random effects pooled estimate calculated in step 1. (This
shifts the individual P's towards the pooled estimate.) The standard error of the estimate of P
is similarly recalculated. (This reduces the standard errors associated with the individual P's.)
The uncertainty associated with P in the ith location is, as before the adjustment, described by
a normal distribution. The adjustment in step 1 simply shifts the mean of that distribution
from the P reported by the study to the re-estimated p. (The adjustment also reduces the
standard deviation of this normal distribution by incorporating information from all study
locations into the estimate of P in the ith study location.) The formulas for the adjusted mean
and standard deviations are given in Appendix 5.
In step 3, the underlying distribution of P's is estimated as an (unweighted) average of
the normal distributions derived in step 2. Suppose, for example, that three epidemiology
studies reported estimates of P for PM and mortality, each in a different location. The
available information about what P might be in a randomly selected location (not necessarily
one of the three for which P has been estimated) is contained, then, in three normal
distributions, adjusted in step 2 above. These adjusted distributions, denoted flt f2, and f3, are
centered at three different values, Pl5 P2, and P3 (the adjusted P's). The underlying
distribution of P's is estimated as follows: For any possible value of P, x, the value of the
distribution of P's at x is estimated as
fix) = ~ [/!(*) + f2(x) + /3(x)] ,
normalized so that the area under the distribution function integrates to 1. In general, the
estimate of the distribution of P's based on n studies is
n , = i
Abt Associates, Inc: p. 52 July 3, 1996, Revised
-------
normalized so that the distribution function, f(x), integrates to 1.
If n= 1, that is, if there is only a single study for a given combination of health
endpoint and PM indicator, the estimated distribution of P's is just the distribution for that
study ~ that is, a normal distribution, with mean equal to the reported P in the study and
standard deviation equal to the reported standard error of the mean. (There is no pooled
estimate when n= 1, and therefore no adjustment of the P reported by the single study.)
The 5th and 95th percentiles of this estimated distribution, f(x), are then the 90 percent
credible interval - that is, the estimate of the interval within which the true value of P at a
randomly selected location (e.g., the location of interest) lies with 90 percent probability.
The adjustment of the normal distributions in step 2 above may also be seen in a
Bayesian framework, in which the distributions reported by the studies are considered prior
distributions, and the adjusted distributions are considered posterior distributions,
incorporating the evidence of the random effects analysis. The procedure is referred to as an
"empirical Bayes" estimation procedure because the prior distributions are based on empirical
evidence (see, for example, Efron and Morris, 1971; Efron and Morris, 1972a; Efron and
Morris, 1972b; Efron and Morris, 1973; and Cox and Hinkley, 1974) . Exhibit 5.3 shows the
prior and posterior distributions for the set of 10 functions relating short-term PM-10 exposure
and mortality. The adjustment pulls all mean relative risk estimates towards the random-
effects distribution average relative risk of 1.040, with those starting furthest from the average
being changed the most. In addition, the standard deviations are reduced, since the
combination of several coherent analyses reduces overall uncertainty.
The classic Monte Carlo technique, which consists of generating a large number of
observations from a known distribution, is another technique often used to assess uncertainty.
When there are several sources of uncertainty, or equivalently, several parameters in the
model whose values are uncertain, Monte Carlo methods are often used to generate
observations from several distributions. On each of a large number of iterations, an
observation is randomly drawn from each of the distributions, and the model output based on
the parameter values drawn is calculated. (This is referred to as "propagating uncertainty
through the model.") N iterations thus generate N model output values. As N approaches
infinity, the distribution of output values approaches the distribution of output values
consistent with the distributions of input parameter values. A ninety percent confidence
interval around the model output can then be derived from the 5th and 95th percentiles of this
distribution.
The distribution, f(x),whose derivation is discussed above, can be shown to be the limit
distribution of a Monte Carlo "propagation of uncertainty" procedure in which there is only a
single source of uncertainty (the concentration-response function). That is, this distribution
would have been approached by a Monte Carlo procedure analogous to the procedure used in
the typical "propagation of uncertain" exercise. Because there is only a single source of
Abt Associates, Inc. p. 53 July 3, 1996, Revised
-------
uncertainty, the Law of Large Numbers allows the analytic calculation of the distribution
reached in the limit, that is, as the number of trials in the Monte Carlo procedure approaches
infinity. The analytic limit distribution was used in these risk analyses.
Exhibit 5.3. Empirical Bayes Estimation of Distributions of Relative Risk for Mortality
for a 50 /tg/m3 Increase in PM-10: Prior and Posterior Distributions
Study
Pope et al '92
Schwartz '93
Kinney et al '95
Schwartz et al 1996a
Ito et al '95
Location
Utah
Birmingham
LA
Boston
Knoxville
St. Louis
Steubenville
Portage
Topeka
Chicago
Prior (unadjusted)
mean
1.076
1.054
1.025
1.061
1.046
1.030
1.046
1.035
0.975
1.025
std. dev.
0.017
0.022
0.014
0.013
0.023
0.012
0.020
0.028
0.036
0.006
Posterior (adjusted)
mean
1.056
1.044
1.032
1.052
1.042
1.034
1.042
1.039
1.032
1.027
std. dev.
0.011
0.012
0.010
0.010
0.012
0.009
0.012
0.013
0.013
0.006
As an alternative to the three-step method described above, a standard functional form
for the underlying distribution of P's may be assumed (for example, it may be assumed to be a
normal distribution or a beta distribution). In this case, the reported estimates of P could be
treated as a sample from this distribution and used to estimate the values of the parameters of
the (assumed) distribution. Because the method described above does not impose any
particular standard functional form on the underlying distribution of p's, however, but instead
allows the reported estimates of P and the user's confidence in these estimates to suggest the
form, this method is preferred as a way of providing an estimate of the underlying distribution
that is most consistent with the evidence from the epidemiological studies.
5.2.2. Estimating P in a given location
In the absence of any location-specific information, the most reasonable estimate of P
in a given location (other than the study locations) is the mean of the estimated distribution of
P's. When there is only a single study that has estimated a concentration-response function,
the estimated distribution of P's is just a normal distribution with mean equal to the p reported
by the study. In this case, the single reported P is therefore the best estimate of P in the given
Abt Associates, Inc.
p. 54
July 3, 1996, Revised
-------
location. When there is more than one study, it can be shown that the mean of the estimated
distribution of P's is the random effects pooled estimate derived in step 1 and used in step 2.
This pooled estimate, then, is the best estimate of p in the given location. The uncertainty
bounds around the estimate are just the 90 percent credible interval, described above. (In the
case of a single study, this is the same as the 90 percent confidence interval around the mean.)
5.2.3. Pooled estimates of p
Many studies have attempted to determine the influence of paniculate matter pollution
on human health. Usually this involves estimation of relative risk for a given change in
pollutant concentration. Each study provides an estimate of the relative risk, along with a
measure of the uncertainty of the estimate. Because uncertainty decreases as sample size
increases, combining data sets is expected to yield more reliable estimates of relative risk.
Combining data from several comparable studies in order to analyze them together is often
referred to as meta-analysis.
For a number of reasons, including data confidentiality, it is often impractical or
impossible to combine the original data sets. Combining the results of studies in order to
produce better estimates of relative risk provides a second-best but still valuable way to
synthesize information (DerSimonian and Laird, 1986). This is referred to as "pooled
analysis" in this report. This kind of pooled analysis requires that all estimates of relative risk
be made using the same change in pollutant concentration. The method of pooled analysis
used is described briefly below and in more detail in Appendix 2. Appendix 3 discusses how
relative risks for different pollutant concentration changes can be made comparable.
One method of pooled analysis is simply averaging all reported relative risks. This has
the advantage of simplicity, but the disadvantage of not taking into account the measured
uncertainty of each of the estimates. Estimates with great uncertainty surrounding them are
given the same weight as estimates with very little uncertainty.
It seems reasonable that a "pooled estimate" which combines the estimates from
different studies should give more weight to estimates from studies with little reported
uncertainty than to estimates with a great deal of uncertainty. The exact way in which weights
are assigned to estimates of relative risk from different studies in a pooled analysis depends on
the underlying assumption about how the different estimates are related to each other.
Under the assumption that there is a distribution of P's (referred to as the random
effects model), the different relative risks (or p's) reported by different studies may be
estimates of different underlying relative risks (corresponding to a given change in
concentration), rather than just different estimates of the same relative risk. The random-
effects model is preferred here to the fixed effects model (which assumes that there is only one
Abt Associates, Inc. p. 55 July 3, 1996, Revised
-------
P everywhere), because it does not assume that all studies are estimating the same parameter.12
(Some researchers (e.g. Moolgavkar & Luebeck, 1996; hi press) suggest, however, that the
heterogeneity hi estimates of PM effects makes single estimates of risk difficult to estimate.)
Several pooled analyses, performed on a number of combinations of PM-10 studies
identified in the Criteria Document for PM-10 (EPA, 1996a), are described below. Studies
were aggregated for analysis based upon the health endpoint of concern (only mortality and
hospital admissions are considered) and the period of exposure evaluated (e.g., one day, three
day). In addition, a pooled function for mortality and PM-2.5, derived by Schwartz et
al.!996a from their studies in six cities, is also presented.
5.2.3.1. Pooled analyses of mortality PM-10 concentration-response functions
Pooled analyses were performed on various subsets of the following short-term
exposure mortality studies cited in the Criteria Document for PM-10 (EPA, 1996a):13
Ito and Thurston 1996 (Chicago, IL);
Kinney et al. 1995 (Los Angeles, CA);
Pope et al. 1992 (Utah Valley, UT);
Schwartz 1993 (Birmingham, AL);
Schwartz et al. 1996a (Boston; Knoxville, TN; St. Louis; Steubenville, OH;
Portage, WI; Topeka, KS)
Exhibit 5.4 shows the relative risks reported in the original studies for a change in PM-10
concentration of 50 /xg/m3. It is explained in Appendix 3 how the relative risk corresponding
to one PM concentration change can be adjusted to reflect other concentration changes. Exhibit
5.5 shows the studies included in each pooled analysis.
The "all averaging times" pooled analysis includes studies that used the average PM
concentration on a single day as the pollution indicator as well as studies that used the average
PM concentration over a 2-, 3- or 5-day period. Those studies which use multi-day averages
are in effect using a smoothed data set, comparing each day's mortality to recent average
12 In studies of the effects of PM-10 on mortality, for example, if the composition of PM-10 varies
among study locations the underlying relationship between mortality and PM-10 may be different from one study
location to another. For example, fine particles make up a greater fraction of PM-10 in Philadelphia County than
in Southeast Los Angeles County. If fine particles are disproportionately responsible for mortality relative to
coarse particles, then one would expect the true value of p for PM-10 in Philadelphia County to be greater than
the true value of P for PM-10 in Southeast Los Angeles County. This would violate the assumption of the fixed
effects model.
''Although there are five studies, Schwartz et al. 1996a effectively conducted six separate studies in the
six cities listed. Following Schwartz et al., the pooled analyses treat these six cities as separate studies. There
are therefore ten PM-10 mortality studies on which pooled analyses were based.
Abt Associates, Inc. p. 56 July 3, 1996, Revised
-------
exposure rather than simply exposure on the same day. The averaging tunes applied
Exhibit 5.4. Relative Risks of Mortality Associated With a Change in
PM-10 Concentration of 50
Study
Ito and Thurston, 19%
Kinney et al., 1995
Popeetal., 1992
Schwartz, 1993
Schwartz etal.,19%a
Schwartz etal., 1996a
Schwartz etal., 1996a
Schwartz etal., 1996a
Schwartz etal., 1996a
Schwartz etal., 1996a
Location
Chicago
Los Angeles
Utah
Birmingham
Boston
Knoxville, TN
St. Louis
Steubenville,
OH
Portage, WI
Topeka, KA
Relative
Risk
1.025
1.025
1.076
1.054
1.061
1.046
1.030
1.046
1.035
0.975
Standard Error
0.006
0.014
0.017
0.022
0.013
0.023
0.012
0.020
0.027
0.036
Exhibit 5.5. Studies Included in Each Pooled Analysis for PM-10
Study
Ito et al, '95
Kinney et al. '95
Pope et al. '92
Schwartz '93
Schwartz et al. 1996a
all
averaging
times
/
/
/
/
/
single-day
averaging
only
,
/
2-day
averaging
only
/
/
multi-day
averaging
only
/
/
to the PM data used in the risk analyses were chosen to correspond to the majority of functions
used in a pooled analysis. Because seven of the ten functions included in the "all averaging
times" pooled analysis are based on two-day PM averages, two-day PM averages were used
with this pooled function. Single-day PM concentrations were used with the one single-day
function; two-day PM averages were used with the two two-day functions; and five-day PM
averages were used with the two multi-day functions (one of which is a three-day function and
the other of which is a five-day function). The more nearly linear the concentration-response
function, however, the less difference it makes whether multi-day averaging functions are used
with single-day PM data. (If the functions were perfectly linear, it would make no difference
at all.) The concentration-response functions considered here are nearly linear, as discussed
above. Because all of the studies in the "all averaging times" pooled analysis focus on the
results of short-term exposure to pollution (as opposed to long-term exposure, measured in
years), it is appropriate to consider these studies together.
Abt Associates, Inc.
p. 57
July 3, 1996, Revised
-------
Exhibit 5.6 shows the average relative risk of the mortality studies considered in each
pooled analysis, as well as the relative risk estimates (and standard errors) from pooled
analyses based on the fixed effects model and, where possible, the random effects model. The
random effects model is preferred because it takes into account possible geographic variability,
and pooled functions were based on the random effects model except when this was not
possible due to insufficient difference among the reported studies (see Appendix 2). Relative
risks are expressed for a 50 /ig/m3 increase in PM-10 concentration. For the single-day
averaging time study, it was not possible to calculate a random effects model estimate, and so
only the result based on the fixed effects model is shown. Exhibit 5.7 shows the relative risk
results of the original studies and the inverse variance weighting pooled analyses graphically,
including the 95 percent confidence bounds.
Exhibit 5.6. Pooled Analyses of Mortality for PM-10 Relative Risk Estimates for a 50
/ig/m3 Increase in PM-10
Group
All averaging times
Single-day averaging
only
2-day averaging only
> 2 -day averaging
only
N
10
1
7
2
Arithmetic
Average
Relative Risk
1.037
1.025
1.031
1.065
Fixed Effects
Inverse Variance
Weighting
est. RR
1.035
1.025*
1.032
1.068
s.e.
0.004
0.014*
0.005
0.013
Random Effects
Inverse Variance
Weighting
est. RR
1.040
n/a
1.035
n/a
s.e.
0.007
n/a
0.007
n/a
* Although there is only one single-day averaging study (Kinney et al., 1995), and therefore no weighting is
possible, the figures in this column are presented to include the standard error.
Exhibits 5.6 and 5.7 illustrate that the selection of studies to include in the pooled
analysis is a critical choice. For example, pooling the two studies that use average PM
concentrations over three to five days increases the relative risk estimated for a 50 ^g/m3
increase in PM-10, compared to that estimated by the pooled function based on studies that use
average PM concentrations over only two days.
5.2.3.2. Pooled analyses of mortality PM-2.5 concentration-response functions
Exhibit 5.8 summarizes the available short-term exposure mortality studies that use
PM-2.5 as the paniculate matter indicator. Schwartz et al. 1996a reported relative risk
estimates for six cities, along with an estimate for all six cities combined, based on a method
similar to the popled analysis method described above, Exhibit 5-8 presents Schwartz et al.'s
relative risk estimates for a 25 /xg/m3 increase in PM-2.5 for each of the six cities separately as
well as the relative risk estimate based on the pooled function provided in the study.
Abi Associates, Inc.
p. 58
July 3, 1996, Revised
-------
§
T3
O,
Relative Risk (95% Confidence Intervals)
o
00
CO
tn
Ito and Thuston, 1996
Kinney et al., 1995
Pope etal., 1992
5"
a Schwartz, 1993
<
^- Schwartz et al. , 1996a (Boston)
S.
£ Schwartz et al. , 1 996a (Knoxville, TN)
C
g.
{J| Schwartz et al., 1 996a (St. Louis)
Schwartz et al., 1996a (Steubenville, OH)
Schwartz et al., 1996a (Portage, Wl)
Schwartz et al., 1996a (Topeka, KA)
POOled Function All averaging times
(Random Effects) , J , ,
1 ' 2-day averaging time
POOled Function > 2-day averaging time
(Fixed Effects)
IK+H
il * i
1 1*1
1 1 * 1
1 1 <* 1
1 1*1
It A 1
|l * 1
II A i
1 1
I, A __.,|
'1 * '
i .,_,,A. 1
' 1 ' '
i A _ , ,,j
* 1
1 ^
1 ^
i A. i
1
3
Q.
o
m
8
o ~
55' 01
I ^
c st £ 2?
Q- 3 a: SL
" sf. S ST
2 a)
o o <§
S 3 w
ft)
3
Q.
2.
-------
Exhibit 5.8. Short-term Exposure Mortality Studies Using PM-2.5 as the
Indicator of Particulate Matter, with Pooled Analyses.
Relative Risk Estimates for a 25 ftg/m3 Increase in PM-2.5
Study City
Watertown, MA
Knoxville, TN
St. Louis, MO
Steubenville, OH
Portage, WI
Topeka, KS
Pooled function, using fixed
effects inverse variance
weighting
Relative Risk
(95% Confidence
Interval)
1.056
(1.038, 1.074)
1.035
(1.005, 1.066)
1.028
(1.010, 1.043)
1.025
(0.998, 1.053)
1.030
(0.993, 1.071)
1.020
(0.951, 1.092)
1.038
(1.028, 1.048)
Source
Schwartz et al.
1996a
Schwartz et al.
1996a
Schwartz et al.
1996a
Schwartz et al.
1996a
Schwartz et al.
1996a
Schwartz et al.
1996a
Schwartz et al.
1996a
5.2.3.3. Pooled analyses of morbidity concentration-response functions
Exhibit 5.9 shows the results of pooled analyses performed on studies of various other
health effects of PM-10, as measured by hospital admissions. The four studies for hospital
admissions for pneumonia and COPD come from Schwartz 1994a (Birmingham), Schwartz
1994b (Detroit), Schwartz 1994c (Minneapolis/St. Paul), and Schwartz 1996 (Spokane). The
three figures used for "total respiratory" hospital admissions are from Schwartz 1995 (New
Haven, CT and Tacoma, WA) and Schwartz 1996 (Spokane). A fourth study of respiratory
hospital admissions, using data from Cleveland, was excluded, because it considers 2-day
average PM-10 rather than same-day average PM-10 as the other studies do. (Including the
Cleveland study makes virtually no difference in the results.) In only one case was the
variation in results among the studies great enough to allow the use of a random-effects model.
In the other cases, there is insufficient evidence that the different studies are estimating
different underlying concentration-response functions.
Abt Associates, Inc.
p. 60
July 3, 1996, Revised
-------
Exhibit 5.9. Pooled Analyses of Effects of PM-10 on Hospital Admissions
Relative Risk Estimates for a 50 /tg/m3 Increase in PM-10
Health Effect*
hospital admissions-
"total respiratory"
hospital admissions-
pneumonia
hospital admissions-
COPD
N
3
4
4
Arithmetic
Average
Relative
Risk
1.098
1.071
1.163
Fixed Effects
Inverse Variance
Weighting
est. RR
1.088
1.069
1.137
s.e.
0.022
0.014
0.023
Random Effects
Inverse Variance
Weighting
est. RR
n/a
n/a
1.140
s.e.
n/a
n/a
0.024
*A11 hospital admissions in this exhibit refer to individuals age 65 or older.
5.2.4. Quantitative assessment of uncertainty surrounding P's applied to Philadelphia
and Los Angeles: results
An uncertainty analysis (i.e., estimation of the distribution of p's and calculation of a
90 percent credible interval) was carried out in each case in which a pooled analysis was
performed (see Section 5.2.3) ~ that is, in each case in which there is more than one reported
estimate of the concentration-response function.14 (The one exception is the concentration-
response function for PM-10 and mortality, based on 1-day averaging, for which there is only
one study. This was included within the set of pooled analyses for mortality for
completeness.)
The estimated distribution of P's can be translated into a distribution of avoided health
effects incidences corresponding to a given change in PM concentrations, as described above.
Alternatively, it can be translated into a distribution of relative risks associated with a given
change in PM concentrations, because each value of P corresponds to a particular relative risk
for a given change in PM (see Section 2 and Appendix 3). The results of the uncertainty
analyses, presented below, are in terms of relative risks associated with a 50 /ig/m3 change in
PM-10 or a 25 jig/m3 change in PM-2.5. The mean, the 95 percent credible interval (i.e., the
2.5 and 97.5 percentile points), and the 90 percent credible interval (i.e., the 5 and 95
percentile points) of the estimated distribution of relative risks for each uncertainty analysis are
presented in Exhibit 5.10. The 90 percent credible intervals are reported along with the
uln those cases for which there is only a single study, the best estimates of the 5th and 95th percentiles of
the distribution of P's are the 5th and 95th percentiles of the normal distribution with mean equal to the (J reported
by the study and standard deviation equal to the standard error of the mean reported by the study. There is
therefore no "analysis" necessary.
Abt Associates, Inc.
p. 61
July 3, 1996, Revised
-------
quantitative estimates of risk. This approach is consistent with the 90 percent confidence
intervals used to characterize uncertainty in other Agency risk analyses, such as those
conducted for ozone and lead. In addition, the 95 percent confidence intervals around the
pooled analysis estimate of the mean of each underlying distribution of relative risks is
presented for comparison both with the uncertainties as reported in the original studies and
with the 95 percent credible interval from the estimated distribution of relative risks.
The distinction between a credible interval and the corresponding confidence interval
around the pooled estimate is worth a further note for the sake of clarity. The 90 percent
confidence interval around the pooled estimate is derived from the standard error of the
estimate of the mean, which is a measure of how good an estimate of the mean of the
underlying distribution the pooled analysis estimate is. The 90 percent confidence interval
around the pooled estimate is the interval within which the true mean of the underlying
distribution lies with 90 percent confidence. (For a precise definition of a confidence interval,
see, for example, Mood et al., 1974, p. 375.) The confidence interval around the pooled
estimate, then, comprises uncertainty bounds around the true mean of the distribution.
The 90 percent credible interval, consisting of the 5th and 95th percentiles of the
estimated underlying distribution, comprises uncertainty bounds around the true value of p (or
the true relative risk) in a particular location. In the absence of any further information about
that location, the 90 percent credible interval is an estimate of the interval within which P in
that location will fall with 90 percent probability. This is the appropriate measure of
uncertainty surrounding the estimate of P applied to a specific location.
The 90 percent credible interval will always be at least as wide as, and usually wider
than the 90 percent confidence interval around the pooled estimate of the mean. The greater
the variance of the underlying distribution (i.e., the larger if) the more of a discrepancy there
will be between the two types of uncertainty bounds. As r)2 approaches zero, the 90 percent
credible interval approaches the 90 percent confidence interval around the pooled estimate of
the mean. This is the case, for example, when there is only a single study. In this case, there
is no evidence of differing P's (because P has been estimated in only a single location) and
therefore no evidence that r|2 is positive. In this case, the 90 percent credible interval equals
the 90 percent confidence interval around the estimate of the mean, which is based only on the
within-study sampling error.
To illustrate the comparison between the confidence interval around the pooled estimate
of the mean of the distribution of P's and the corresponding credible interval for a location-
specific p, Exhibits 5.11 and 5.12 show graphically the normal distributions representing the
within-study variability around the P's reported by the ten "all averaging times" mortality PM-
Abt Associates, Inc. p. 62 July 3, 1996, Revised
-------
Exhibit 5.10. Results of Uncertainty Analyses: Means and Ninety-Five Percent Credible Intervals of Estimated Distributions
of Relative Risk
Health Endpoint
Number
of Studies
Mean of
Estimated
Distribution of
Relative Risk
95% Confidence
Interval around
Random Effects
Pooled Analysis
Estimate*
95% Credible
Interval (2.5th and
97.5th percentile
points)
For a 50 j*g/m3 Increase in PM-10:
A. Mortality
all averaging times
2-day averaging time
> 2-day averaging time
1-day averaging time
10
7
2
1
1.040
1.035
1.068
1.025
(1.026, 1.053)
(1.021, 1.049)
(1.043, 1.093)**
(0.998, 1.052)**
(1.014, 1.069)
(1.013, 1.059)
(1.043, 1.093)**
(0.998, 1.052)**
B. Morbidity: Hospital admissions
"Total respiratory"
COPD
pneumonia
3
4
4
1.089
1.140
1.069
(1.045, 1.131)**
(1.093, 1.187)
(1.042, 1.097)**
(1.045, 1.131)**
(1.087, 1.195)
(1.042, 1.097)**
For a 25 j*g/m3 Increase in PM-2.5:
Mortality
6 cities
(in one
study)
1.036
(1.026, 1.047)
(1.019, 1.053)
90% Credible
Interval (5th and
95th percentile
points)
(1.018, 1.064)
(1.017, 1.055)
(1.047, 1.090)**
(1.002, 1.048)**
(1.053, 1.125)**
(1.094, 1.185)
(1.046, 1.093)**
(1.022, 1.051)
* The random effects pooled analysis estimate of central tendency is the same as the mean of the uncertainty analysis distribution based on the same weights.
** A random effects pooled analysis could not be performed. Results are from a fixed effects pooled analysis.
Aht Associates, Inc.
p. 63
July 3, 1996. Revised
-------
10 studies as well as both the estimated distribution of P's and the corresponding pooled
analysis estimate of the mean of the distribution. Exhibit 5.11 shows the ten unadjusted
normal distributions representing the within-study variability around each of the ten reported
P's. Exhibit 5.12 superimposes on the ten adjusted normal distributions (1) the pooled
analysis estimate of the mean of the distribution of P's and the normal distribution around this
estimate, based on the standard error of the estimate, and (2) the estimated distribution of P's
generated from the underlying ten adjusted normal distributions, using the three-step
estimation procedure described in Section 5.2.1.
Note that, even though the distributions around the reported estimates of P's are
normal, the estimated distribution of P's derived from them is not normal. As noted above, it
does not have a standard functional form but instead reflects the evidence from the particular
studies on which it is based. Note also that it is possible, when sampling from the estimated
distribution of P's, to select a negative number. The probability of doing so, however, is
extremely small.
5.2.5. Translating a 90 percent credible interval for P into a 90 percent credible
interval for avoided health effect incidence
As in Section 5.2 above, the health effect will be taken to be mortality for ease of
discussion in this section. The discussion is, however, generalizable to any health effect. For
a given set of PM reductions in a given location, to any value of P there corresponds a
predicted avoided mortality. There is therefore a distribution of avoided mortality
corresponding to the distribution of P's. Ideally, the 5th and 95th percentiles of the estimated
distribution of avoided mortality would compose the 90 percent credible interval around a
point estimate of avoided mortality in the location of interest. If the concentration-response
function is linear, the mortality predicted by the concentration-response function evaluated at
the x percentile P would be the same as the x percentile of the distribution of mortality. That
is, if PX denotes the x percentile value of p from the distribution of P's, and Ayx denotes the x
percentile value of the distribution of avoided mortality from the corresponding distribution,
then
.
by, = y[e '
If, however, the concentration-response function is convex or concave, the equality becomes
an inequality. Because the concentration-response functions are almost linear, however, the
discrepancy will be very small, and the 5th and the 95th percentile P's from the estimated
distribution of P's were applied to the air quality data in a given location (Philadelphia County
or Southeast Los Angeles County) to obtain a close approximation to a 90 percent "credible
interval" in avoided incidences .of the given health effect in that location.
Abt Associates, Inc. p. 64 July 3, 1996, Revised
-------
Exhibit 5.11: Probability Density of Beta (Slope) Coefficients
For PM-10 Mortality Studies in Ten Locations (Before Empirical Bayes Adjustments)
Ito & Thurston, 1996
St. Louis
Steubenville
Kinney et al.,
1995
Pope et al 1992
Schwartz,
1993
0 0.001 Knoxville
Beta (Slope) Coefficient
0.002
Curves labelled with only a city name come from Schwartz et al., 1996a.
Abt Associates, Inc.
p. 65
July 3, 1996, Revised
-------
Exhibit 5.12: Mortality Studies from ten Locations (Adjusted Using Empirical Bayes Method),
Pooled Analysis Central Tendency Estimate, and Uncertainty Analysis Estimated Distribution
Results Based on Random Effects Weighting
Pooled Analysis Central'
Uncertainty Analysis
Estimated Distribution
Tendency Estimate
(with surrounding normal distribution,
using standard error of estimate)
-------
6. Baseline Health Effects Incidence Rates
Incidence rates are required inputs for many, but not all, concentration-response
functions. Many of the epidemiology studies used in tins analysis directly estimate the
percentage change in incidence (i.e., the relative risk), rather than the absolute nupher of
cases for an endpoint. To estimate the number of PM-associated cases using 4hese studies, it
is necessary to know the baseline incidence, that is, the number of cases hi a location, before*
change in PM air quality. «
Incidence rates express the occurrence of a disease or event (e.g., asthma episode,
death, hospital admission) in a specific period of time, 'usually per year. Rates are expressed
either as a value per population group (e.g., the number of cases in Philadelphia Ckninr^joria
value per number of people (e.g., number of cases per 10,000 residents), and may be^6 «nd
sex specific. Incidence rates vary among geographic areas due to differences hi population
characteristics (e.g, age distribution) and factors promoting illness (e.g., smoking, air
pollution levels).15 The sizes of the populations in Philadelphia County and Southeast Los
Angeles County that are relevant to the risk analyses reported here (i.e., the populations to
which the baseline incidences refer) are given hi Exhibit 6.1.
Exhibit 6.1. Relevant Population Sizes for Philadelphia County and Southeast
Los Angeles County
Population
Total
Ages * 65
Children, ages 8-12
Children, ages 10-12
Asthmatic Children,
ages 9-11
Asthmatic African-American
Children, ages 7-12
Philadelphia County
1,586,000
240,800 (15.2%)
102,900 (6.5%)
61,700 (3.9%)
3,900* (0.25%)
Southeast Los Angeles
County
3,636,000
322,100 (8.9%)
282,100 (7.8%)
165,800 (4.6%)
10,700* (0.29%)
1,800* (0.05%)
Incidences for asthmatic children were obtained using the national asthma prevalence among children (6.3%).
The incidence of asthmatic African-American children ages 7-12 in Southeast L.A. County, for example, is
3,636,000 multiplied by {0.0937 (the proponion of the population that is ages 7-12) x 0.085 (the proportion of the
population that is African-American) x 0.063 (the proponion of the national population of children that are
asthmatic)}.
15 Incidence rates also vary within a geographic area due to the same factors; however, statistics
regarding within-city variations are rarely available and are not necessary for this analysis.
Abt Associates, Inc.
p. 67
July 3, 1996, Revised
-------
6.1. Sources
data
"^afes,
(death
. None of the
tfaefedendpainTmynt
^health departments,co
to hospital
planning medical
endpoints in the
for specific communicable
report to the federal
sis are required to be reported
to the required iederal reporting, many state and local
on some additiooal|fe|^oints. 'These most often are *
discharge diagnoses; Which are collected to assist in
may also be cotted&l ^particular studies of health issues Df
& -
^_ attaoomany of flic efidpomts co^ ^
at the n^PV level (national averages);feril.the
ents and hospital
rates.
forhospii
disuse, and ischemic
Philadelphia County and
;nter for Health Statistics
, mortality, and health services data. Baseline
neumonia, COPD, "total respiratory," congestive
Failure) were obtained for Philadelphia County from the
Delaware Valley Hospital Council for 1993-1994 (fiscal year 1994), and for Los Angeles from
California's Office of Statewide Health Planning and Development Data Users Support Group.
Finally, in the absence of other sources of baseline incidence data for respiratory symptoms
and acute bronchitis, baseline rates for these health endpoints were taken from the studies
which estimated the concentration-response functions used for these endpoints in the risk
analysis.
Baseline health effects incidence rates used in the risk analysis are given in Exhibit 6.2.
In all cases, the incidence rates listed correspond to the ages of the populations studied in the
relevant epidemiology studies, e.g., individuals over 65 years of age. The national incidence
rates given in Exhibit 6.2 may differ from those given in the Criteria Document for several
reasons, including differences in the years used and differences in the ICD codes included
within a health effects category.
Abi Associates, Inc.
p. 68
July 3, 1996, Revised
-------
Exhibit 6.2. Baseline Health Effects Incidence Rates
Health Effect
Short-Term Exposure Mortality*
tper 10Q,000 general population/year)
Umg^erto^Exponire Mortality (age 30 and ottlet)
$per 100,000 general population/year)
%Jfjl»»Ki«M+«. - ' '- ''"!, -. . - .
mOnMUiiyi >,?, --«:.-<-' -v-:-
. ' ' ' ,\. '''- ' '"' '
Philadelphia
County
12*°
1*54*
Southeast Los
Angeles County
676 ,*%;..
. v " >,;^a
657* ' ;
National
Average3
830
o^.',- .
" Y .'
A. Hospital AdmbefcM (per 100,000 general population/year) ', ,
"' "'"-'*3^i^'
Total ttiplratory hospital admissions0 (all ages):
ICD codes 466, 480-482, 485, 490-493
Total respiratory hospital admissions (65 and older):
ICD codes 460-519
COPD admissions (65 and older): ICD codes 490-4%
Pneumonia admissions (65 and older): ICD codes 480-487
Ischemic heart failure (65 and older): ICD codes 41O414
Congestive Heart Disease (65 and older): ICD code 428
816
650
202
257
614
487
427 -v5^
428
116
205
307
197
t+. -
' *S:£.:VG.
504
103
229
450
231
B. Respiratory Symptoms (percent of relevant population)
Lower Respiratory Symptoms (LRS) in children, ages 8-12
(number of cases of symptoms per day)
Lower Respiratory Symptoms (LRS) in asthmatic children, ages
9-1 1 (number of days of symptoms)
Shortness of breath (number of days) in asthmatic African-
American children, ages 7-12
(Doctor diagnosed) acute bronchitis in children ages 10-12 per
yr-
0.15%**
16%**
6.5%**
0.15%**
16%**
5.6%**
6.5%**
~
All incidence rates are rounded to the nearest unit.
a. National rates for hospital admissions for patients over 64 years of age were obtained from Vital and Health
Statistics, Detailed Diagnoses and Procedures, National Hospital Discharge Survey, 1990. June, 1992. CDC.
Hyattsville, Md. Each rate is based on the number of discharges divided by the 1990 population of 248,709,873
b. Mortality figures exclude suicide, homicide, and accidental death, which corresponds to the measures used in
the epidemiological studies employed in this analysis.
c. Although a baseline incidence rate is not needed for calculating the.incidence of total respiratory hospital
admissions associated with PM (because the concentration-response function predicts cases rather than percent
change), it is needed for calculating the PM-related percent change in total incidence.
Abi Associates, Inc.
p. 69
July 3, 1996, Revised
-------
* Although county-specific total mortality incidences (over all ages) were available for both Philadelphia and Los
Angeles, age-specific mortality incidences were not available. Baseline mortality incidences among individuals
aged 30 and over in Philadelphia and Southeast Los Angeles Counties were therefore estimated by applying
national age-specific death rates to county-specific age distributions, and adjusting the resulting estimated age-
specific incidences so that the estimated total incidences (including all ages) equaled the actual county-specific
total incidences. For example, if me total of the estimated age-specific incidences obtained in this way was 5%
higher man the actual totalinej
-------
7. Assessment of the Health Risks Associated with "As Is" PM Concentrations Above
Background
7.1. Results and sensitivity analyses
The results of theJfirst phase of the risk analysis, assessing die health risks associated
with "as is" PM concentrations are given in Exhibit 7.1 for Philadelphia County in 1992-1993
and Exhibits 7.2 and 7.3 for Southeast Los Angeles County in 1995: Because Southeast Los
Angeles County'was not in attainment of current PM-10 standards in 1995, the health risks
associated with "is fe^JIfl concentrations in that location was assessed in two ways. First,
thelassessment W£s''earned out using "as is" PM concentrations (Exhibit 7.2). Second, health
risks were assessed using daily PM concentrations adjusted to simulate attainment of current
standards (Exhibit 7.3). The method of adjusting daily PM concentrations to simulate
attainment of current standards is described in Section 2.2.
All estimated incidences were rounded to the nearest 10, except lower respiratory
symptoms, which are reported to the nearest 1000, and shortness of breath among African-
American asthmatics, which is rounded to the nearest 100. All percentages were rounded to
one decimal place. Rounding was done for convenience of presentation and is not intended to
imply a particular level of precision.
There is substantial uncertainty surrounding all estimates of incidence associated with
"as is" PM concentrations, in both Philadelphia County and Southeast Los Angeles County.
The incidence of a health effect predicted to be associated with "as is" PM concentrations in a
given location depends on the concentration-response function in that location. Because the
true concentration-response functions (for the relevant health effects) are not known, they must
be estimated. If concentration-response functions had been estimated for Philadelphia County
and Southeast Los Angeles County specifically, then the only uncertainty associated with using
these estimated concentration-response functions in the risk analyses for Philadelphia County
and Southeast Los Angeles County would be the uncertainty as to how well the estimated
concentration-response functions approximate the true concentration-response functions in
these locations. This uncertainty is typically expressed as a 90 or 95 percent confidence
interval around the estimate.
However, because concentration-response functions have, for most health endpoints,
not been estimated for Philadelphia County or Southeast Los Angeles County specifically,
concentration-response functions estimated in other locations, or a central tendency estimate
derived by pooling these, have been used instead. This adds a second source of uncertainty to
the risk analyses. If there is true geographic variability in the concentration-response function
for a given health effect, then it is uncertain how well the concentration-response function in
one location (or the mean of concentration-response functions in several locations)
approximates that for a different location. (This is discussed in more detail in Section 3 and
Section 5.2.)
Abi Associates, Inc. p. 71 July 3, 1996, Revised
-------
To assess the total uncertainty surrounding the concentration-response function applied
to a given location (e.g., Philadelphia County) in the risk analysis, in this case, the full range
of possibilities of what die function hi that location might be is characterized by estimating the
distribution of possible values of the "slope" parameter, p, in the concentration-response
function. This distribution is estimated based on the limited information from studies
Lacking further itlli( ion, this
__
distribution characterizes the a^l possibifitjcs of what die conotitWresponse function
might be inJE-ftj^aMnly selected io^tionanywi*ere in die United States. If nothing more about
a location & known, then, it is estimated that with 90 percent probability, p in that location
lies between the 5th and the 95th perccntiles of this estimated
-------
Exhibit 7.1
Estimated Annual Health Risks Associated with "As Is" PM Concentrations
in Philadelphia County, September 1992- August 1993 (for base case assumptions)
Health Effects*
Mortality
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
Lower Respiratory
Symptoms
in Children""
(A) Associated with short-term exposure (all ages)
(B) Assoc. with long-term exposure (age 30 and over)
(51 locations)
(C) Total Respiratory
(all ages)
(D) Total respiratory
(>64 years old)
(E) COPD
(>64 years old)
(F) Pneumonia
(>64 years old)
(G) Ischemic Heart Disease "'
(»64 years old)
(H) Congestive Heart Failure **'
(>64 years old)
(I) Lower Respiratory Symptoms (# of cases)
(8-1 2 year olds)
(J) Lower Respiratory Symptoms (# of days)
(9-1 1 year old asthmatics)
(K) Doctor-diagnosed Acute Bronchitis assoc-
iated with long-term exposure (10-12 year olds)
Health Effects Associated with PM-10 Above Background"
Incidence
220
(160 - 290)
:: : ::
250
(150-340)
120
(80-150)
80
(50 - 100)
80
(30-120)
110
(50-160)
< 10000 >
(8000- 11000)
< 16000 >
(6000 - 25000)
<190>
( 20 - 370 J
Percent of Total Incidence
1.1%
(0.8-1.4)
:: : :
2.4%
(1.5-3.3)
3.7%
(2.5-4.7)
1.9%
(1.3-2.6)
0.8%
(0.3-1.3)
1 .4%
(07-2.1)
17.5%
(15.3-19.6)
6.8%
(2.4 - 10.91
0.3%
(0.0 -0.6L
Health Effects Associated with PM-2.5 Above Background"
Incidence
370
(230 - 510)
860
(540-1170)
280
(70-450)
:::
: : :
70'
(30-120)
100
(50-150)
< 11000 >
(6000-15000)
Percent of Total Incidence
1.8%
(1.1-2.5)
4.7%
(2.9-6.4)
2.0%
(0.5 - 3.5)
--
0.7%
(0.3-1.2)
1.3%
(0.6 - 2.0)
20.0%
(10.3-28.2)
_____
* Health effects are associated with short-term exposure to PM, unless otherwise specified.
** Health effects incidence was quantified across the range of PM concentrations observed in each study, when possible, but not
below background level Background PM-10 is assumed to be 8 ug/m3; background PM-2.5 is assumed to be 3.5 ug/m3.
*" PM-2 5 results based on using PM-2.5 mass as PM-10 mass in the PM-10 functions.
""'Angle brackets <> indicate incidence calculated using baseline incidence rates reported in studies, with no adjustment for
location-specific incidence rates This increases the uncertainty in the incidence estimates.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All the numbers in parentheses are interpreted as 90% credible intervals based on uncertainty analysis
that takes into account both statistical uncertainty and possible geographic variability.
See text in Chapter 7 for details.
Sources of Concentration-Response (C-R) Functions:
(A) PM-10 C-R function based on pooled results from
studies in 10 locations; PM-2.5 C-R function based on pool
results from studies in six locations.
(B) Pope etal., 1995
(C)Thurston, etal., 1994
(0) PM-10 C-R based on pooled results from 4 functions
(E) PM-10 C-R based on pooled results from 4 functions
(F) PM-10 C-R based on pooled results from 4 functions
(G) Schwartz & Morris, 1995
(H) Schwartz & Morris. 1995
(I) Schwartz, etal., 1994
(J) Pope etal., 1991
(K) Dockery et al., 1989
Abt Associates, Inc.
p. 73
July 3, 1996, Revised
-------
Exhibit 7.2
Estimated Annual Health Risks Associated with "As Is" PM Concentrations
in Southeast Los Angeles County, 1995* (for base case assumptions)
Health Effects'*
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
Lower Respiratory
Symptoms
in Children "'"
(«j nssociaieo wnn
short-term exposure (all ages)
(B) Associated with short-term exposure
(all ages; study done in Los Angeles)
(C) Associated with long-term exposure
(age 30 and over; 51 locations)
(D) Total Respiratory
(all ages)
(E) Total Respiratory
(>64 years old)
(F) COPD
(=64 years old)
(G) Pneumonia
(>64 years old)
(H) Ischemic Heart Disease""
(>64 years old)
(I) Congestive Heart Failure*"*
(>64 years old)
(J) Lower Respiratory Symptoms (# of cases)
(8-1 2 year olds)
(K) Lower Respiratory Symptoms (# of days)
(9-1 1 year old asthmatics)
(L) Days of shortness of breath (7-12 year old
African American asthmatics in Los Angeles)
(L) Doctor-diagnosed Acute Bronchitis assoc-
iated with long-term exposure (10-12 year olds)
Health Effects Associated with PM-10 Above Background-
Incidence | Percent of Total Incidence
ouu
(570- 1020)
400
(40 - 750)
__ ., __
1,070
(660-1460)
440
(310-560)
420
(290-550)
260
(100-420)
290
(140-430)
< 62000 >
(56000 - 68000)
< 1 15000 >
(43000-175000)
<7200>
(2400- 10900)
< 5090 >
(680 - 7750)
J.OTb
(2.3-4.1)
1.6%
(0.2-3.1)
_
6.9%
(4.2-9.4)
10.3%
(7.3-13.1)
5.6%
(3.9-7.3)
2.3%
(0.9-3.7)
4.1%
(2.0-6.1)
41.4%
(37.2 - 45.2)
18.4%
(6.9 - 28.0)
19.3%
(6.4 - 29.2)
3.1%
(0.4-4.7)
Health Effects Associated with PM-2.5 Above Background"*
Incidence
9UU
(540-1230)
......
2,800
(1800-3800)
1,200
(330-2080)
:::
......
160
(60-260)
180
(90 - 270)
<51000>
(28000 - 68000)
......
~ ** ""
:::
Percent of Total Incidence
9. f 7b
(2.2 - 5.0)
.
11.9%
(7.5-16.0)
7.7%
(2.1 - 13.4)
_ _
__
« _ _
1.4%
(0.6 - 2.3)
2.5%
(1.2-3.8)
34.4%
(19.1-45.7)
_
__ _ ,
* Southeast Los Angeles County was not in attainment of current PM-10 standards (50 ug/m3 annual average
standard and 150 ug/m3 daily standard) in 1995. Figures shown use the actual reported concentrations.
" Health effects are associated with short-term exposure to PM. unless otherwise specified.
*" Health effects incidence was quantified across the range of PM concentrations observed in each study, when possible, but not
below background level. Background PM-10 is assumed to be 6.0 ug/m3 and background PM-2.5 is assumed to be 2.5 ug/m3.
"" PM-2.5 results based on using PM-2.5 mass as PM-10 mass in the PM-10 functions.
""Angle brackets <> indicate incidence calculated using baseline incidence rates reported in studies, with no adjustment for
location-specific incidence rates. This increases the uncertainty in the incidence estimates.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All numbers in parentheses are interpreted as 90% credible intervals based on uncertainty
analysis that takes into account both statistical uncertainty and possible geographic variability.
See text in Chapter 7 for details.
Sources of Concentration-Response (C-R) Functions:
(A) PM-10 C-R function based on pooled results from
studies In 10 locations; PM-2.5 C-R function based on pooled
results from studies In six locations.
(B)Kinneyetal.,1995
(C) Pope el al., 1995
(D) Thurston, et at.. 1994
(E) PM-10 C-R based on pooled results from 4 functions
(F) PM-10 C-R based on pooled results from 4 functions
(6) PM-10 C-R based on pooled results from 4 functions
(H) Schwartz & Morris. 1995
(I) Schwartz & Morris. 199S
(J) Schwartz, et al., 1994
(K) Pope et al., 1991
(L) Dockery et al., 1989
Abt Associates, Inc.
p. 74
July 3, 1996, Revised
-------
Exhibit 7.3
Estimated Annual Health Risks Associated with Attainment of Current Standards
in Southeast Los Angeles County, 1995* (for base case assumptions)
Health Effects"
Mortality
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
.ower Respiratory
Symptoms
in Children *****
(A) Associated with
short-term exposure (all ages)
(B) Associated with short-term exposure
(all ages: study done in Los Angeles)
(C) Associated with long-term exposure
(age 30 and over 51 locations)
(D) Total Respiratory
(all ages)
(E) Total Respiratory
(>64 years old)
(F) COPD
(>64 years old)
(G) Pneumonia
(>64 years old)
(H) Iscrtemic Heart Disease****
(=64 years old)
(I) Congestive Heart Failure"**
(>64 years old)
(J) Lower Respiratory Symptoms (# of cases)
(8-1 2 year olds)
(K) Lower Respiratory Symptoms (# of days)
(9-11 year old asthmatics)
(L) Days of shortness of breath (7-12 year old
African American asthmatics in Los Angeles)
(L) Doctor-diagnosed Acute Bronchitis assoc-
iated with long-term exposure (10-12 year olds)
Health Effects Associated with PM-10 Above Background*"
Incidence
630
(450 - 800)
290
(30 - 550)
840
(520- 1160)
350
(240 - 440)
330
(230 - 430)
200
(80 - 330)
230
(110-340)
< 52000 >
(46000 - 57000)
< 93000 >
(34000-143000)
< 5200 >
(1700-8100)
< 3760 =
(470-6190)
Percent of Total Incidence
Z.6%
(1.8-3.3)
1.2%
(0.1 - 2.2)
5.4%
(3.3 - 7.4)
8,2%
(5.8 - 10.5)
4.4%
(3.1 -5.8)
1.8%
(0.7 - 2.9)
3.2%
(1.5-4.8)
34.8%
(31.0-38.4)
14.9%
(5.5 - 23.0)
14.1%
(4.6-21.8)
2.3%
(0.3 - 3.7)
Health Effects Associated with PM-2.5 Above Background*"
Incidence
(430 - 970)
2.050
(1250-2690)
940
(250-1630)
_ .. u
130
(50-200)
140
(70-210)
< 43000 >
(23000-58000)
Percent of Total Incidence
2.9%
(1.7-3.9)
8.6%
(5.4-11.7)
6.1%
(1.6-10.5)
1.1%
(0.4-1.8)
2.0%
(1.0-3.0)
28.7%
(15.4-39.0)
* Southeast Los Angeles County was not in attainment of current PM-10 standards (50 ug/m3 annual average
standard and 150 ug/m3 daily standard) in 1995. "As is" daily PM-10 concentrations were first rolled
back to simulate attainment of these standards. "As is" daily PM-2.5 concentrations were rolled back
by the same percent as daily PM-10 concentrations. See text in Chapter VI for details.
"* Health effects are associated with short-term exposure to PM. unless otherwise specified.
*" Health effects incidence was quantified across the range of PM concentrations observed in each study, when possible, but not
below background level Background PM-10 is assumed to be 6.0 ug/m3 and background PM-2.5 is assumed to be 2.5 ug/m3.
*" PM-2.5 results based on using PM-2.5 mass as PM-10 mass in the PM-10 functions.
""'Angle brackets <> indicate incidence calculated using baseline incidence rates reported in studies, with no adjustment for
location-specific incidence rates This increases the uncertainty in the incidence estimates.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All numbers in parentheses are interpreted as 90% credible intervals based on uncertainty
analysis that takes into account both statistical uncertainty and possible geographic variability.
See text in Chapter 7 for details.
Sources of Concentration-Response (C-R) Functions:
(A) PM-10 C-R function based on pooled results from
studies in 10 locations; PM-2.5 C-R function based on pooled
results from studies in six locations.
(B)Kinneyetal .1995
(C) Pope et al., 1995
(D) Thurston. et al., 1994
(E) PM-10 C-R based on pooled results from 4 functions
(F) PM-10 C-R based on pooled results from 4 functions
(G) PM-10 C-R based on pooled results from 4 functions
(H) Schwartz & Morris, 1995
(I) Schwartz & Morris, 1995
(J) Schwartz, et al., 1994
(K) Pope et al., 1991
(L) Dockery et al., 1989
Abt Associates, Inc.
p. 75
July 3, 1996, Revised
-------
Exhibit 7.4
Number of Days on Which PM Concentration Reported
in Southeast Los Angeles County, 1995
Exceeds Maximum Observed PM in Studies for Each Endpoint
Health Effects
Mortality (all ages)
Hospital Admissions
Respiratory
Ispilal Admissions
rdiac
Her Respiratory
nptoms
Children
(A) Associated with
short-term exposure
(B) Associated with short-term exposure
(study done in Los Anaeles)
(C) Associated with long-term exposure
(51 locations)
(D) Total Respiratory
(all ages)
(E) Total Respiratory*
(>64 years old)
(H) Ischemic Heart Disease*
(>64 years old)
(F) COPD*
(>64 years old)
(G) Pneumonia*
(>64 years old)
(I) Congestive Heart Failure*
(>64 years old)
(J) Lower Respiratory Symptoms (# of cases)
(8-12 year olds)
(K) Lower Respiratory Symptoms (# of days)
(9-1 1 year old asthmatics)
(L) Days of shortness of breath (7-12 year old
African American asthmatics in Los Anaeles)
(L) Doctor-diagnosed Acute Bronchitis assoc-
iated with lonq-term exposure (10-12 year olds)
Southeast Los Angeles AQ Data, 1995
Composite monitor
Annual average PM-10: 49ug/m3
Annual average PM-2.5: 29 ug/m3
Daily Max PM-10 197ug/m3
Daily Max PM-2.5: 1 20 ua/m3
Maximum Concentration
(in ug/m3)
251 (PM-10)
170 (PM-2.5)"
177 (PM-10)
34 (PM-2 5)
66 (PM-2.5)
83 (PM-10)'
83 (PM-10)*
83 (PM-1 DI-
BS (PM-10)'
83 (PM-10)'
117 (PM-10)
86 (PM-2.5)
195 (PM-10)
101 (PM-10)
59 (PM-10)"'
Number of days witn AQ data on wnicn Reported
PM-10
0
2
n/a
n/a
33
33
33
33
33
16
1
19
falls within range
PM-2.5
0
n/a
falls within range
22
12
12
12
12
12
10
n/a
n/a
n/a
Sources of Concentration-Response (C-R) Functions:
(A) PM-10 C-R function based on pooled results from
studies in 10 locations; PM-2.5 C-R function based on pooled
results from studies in six locations.
(B)Kinneyetal..1995
(C)Popeetal, 1995
* Based on reported 90th percentile of distribution reported in study.
" Based on reported 95th percentile of distribution reported in study
"* Based on reported PM-15 distribution.
(D)Thurston, etal., 1994
(E) PM-10 C-R based on pooled results from 4 functions
(F) PM-10 C-R based on pooled results from 4 functions
(G) PM-10 C-R based on pooled results from 4 functions
(H) Schwartz & Morris, 1995
(I) Schwartz & Morris. 1995
(J) Schwartz, et a!.. 1994
(K) Pope et al.. 1991
(L)Dockeryetal., 1989
Abi Associates, Inc.
p. 76
July 3, 1996, Revised
-------
Another reason that the estimated percentage of lower respiratory symptoms may be so
large is that the original study of lower respiratory symptoms among a general population of 8-
12 year old children (Schwartz et al. 1994) restricted the analysis to a period from April to
August. During these months, the respiratory symptoms incidences from other causes are at a
minimum. Thus, applying the changes observed in the odds ratio associated with PM over
these months across an entire year could result hi an overestimate of the percentage incidence
of new cases of respiratory symptoms overall.
The effect of other pollutants, such as ozone, as confounders or effects modifiers,
would also be higher in the summer months and could possibly lead to some overestimation of
effects. Schwartz et al. 1994 indicates that a general consistency of findings in both summer
studies in the Six Cities and winter studies in Utah Valley suggests that the association with
respiratory symptoms is not limited to photochemically produced aerosols. However, the
consistency between studies is greater for cough than for lower respiratory symptoms, the
endpoint examined in this analysis (Schwartz et at. 1994, Table 6). Thus questions still exist
as to whether factors such as the differing prevalence of respiratory symptoms between whiter
and summer or effects modification by ozone may play some role hi the large odds ratio for
lower respiratory symptoms predicted by the Six Cities study (Schwartz et al. 1994).
As seen in Exhibits 7.2 and 7.3, Pope et al. (1991) also examined lower respiratory
symptoms among asthmatic children (a possible sensitive subgroup) hi the Utah Valley, and
found significantly lower PM-associated incidence. Ostro et al. (1995) examined shortness of
breath among African-American asthmatic children and estimated a PM effect roughly 50
percent stronger than that from the Utah Valley. Shortness of breath might be considered a
less severe effect than "lower respiratory symptoms" (and therefore be exacerbated by lower
levels of PM), and African-Americans may be a sensitive subgroup for PM effects on
asthmatic symptoms. Both of these factors might lead one to expect a somewhat strong PM
effect on shortness of breath in African-American asthmatics, as observed. The results from
Ostro et al. (1995) might therefore be considered consistent with the symptom results
presented in Exhibits 7.2 and 7.3.
Although in most cases concentration-response functions estimated in the sample
locations were not available, for short-term exposure mortality there is a single preliminary
(unpublished) study carried out in Philadelphia for PM-2.5 and a single (published) study
(Kinney et al. 1995) carried out in Los Angeles for PM-10. The results from these functions
are compared with the results using the corresponding pooled analysis functions for short-term
exposure mortality in Exhibit 7.5.
Abt Associates, Inc. p. 77 July 3, 1996, Revised
-------
Exhibit 7.5. Comparison of Predicted Short-Term Exposure Mortality Incidence Using
Pooled Analysis Functions and Location-Specific Functions
Concentration-
Response Function
PM-10 pooled
analysis function
based on 10 studies
Study done in Los
Angeles (Kinney et
al., 1995)
PM-2.5 pooled
analysis function
based on 10 studies
Study done in
Philadelphia
(Dockery et al.,
Abstract, 1996)*
Health Effects Associated with PM-10
Above Background in Southeast Los
Angeles County
Incidence
800
(570-1020)
400
(40 - 750)
Percent of Total
Incidence
3.3%
(2.3-4.1)
1.6%
(0.2-3.1)
Health Effects Associated with PM-2.5
Above Background in Philadelphia County
Incidence
370
(220 - 510)
510
(190 - 840)
Percent of Total
Incidence
1.8%
(1.1-2.5)
2.5%
(0.9-4.1)
*This study is as yet only in the form of an unpublished abstract and is therefore not included among the studies in
the Exhibits of results.
Exhibit 7.6 provides a different way of looking at the "as is" results for Philadelphia
County. The histogram shows the number of days on which PM-2.5 is within a given range in
the Philadelphia County data. The line shows the number of deaths associated with PM-2.5 on
those days. The number of deaths associated with PM-2.5 depends both on the number of
days at a given concentration and on the concentration itself. Therefore the bulk of PM-
related mortality is associated with PM-2.5 concentrations of between 12 and 24 /^g/m3 simply
because most days have concentrations in that range. The small number of days with large
concentrations of PM-2.5, although they contribute more deaths per day, do not contribute as
large a number of deaths overall.
Several sensitivity analyses were performed to assess the sensitivity of the results of
analyses in the first phase of the risk analysis to various assumptions underlying the analyses.
The sensitivity analyses and the exhibits presenting their results are summarized in Exhibit
7.7.
Abt Associates, Inc.
p. 78
July 3, 1996, Revised
-------
Exhibit 7.6
Distributions of Two-Day Mean PM-2.5 Concentration
and of Mortality Associated With
the Excess of Those Concentrations Above Background
Philadelphia, September 1992 - August 1993
120
100
to
o>
to
o
-------
Exhibit 7.7. Summary of Sensitivity Analyses Associated with "As Is" Phase of RiskAnalysis
Sensitivity Analysis of:
the effect of alternative background levels on predicted health effects
associated with "as is" PM-10 and PM-2.5 in Philadelphia County
the effect of alternative background levels on predicted health effects
associated with "as is" PM-10 and PM-2.5 in Los Angeles County
the effect of alternative cutpoint models on predicted health effects
associated with "as is" PM-10, using two different methods of slope
adjustment in Philadelphia County
the effect of alternative cutpoint models on predicted health effects
associated with "as is" PM-2.S, using two different methods of slope
adjustment in Philadelphia County
the effect of alternative cutpoint models on predicted health effects
associated with "as is" PM-10, using two different methods of slope
adjustment in Los Angeles County
the effect of alternative cutpoint models on predicted health effects
associated with "as is" PM-2.5, using two different methods of slope
adjustment in Los Angeles County
the effect of combining different averaging times in pooled short-term
exposure mortality functions on predicted health effects associated
with "as is" PM-10 in Philadelphia County
the effect of using concentration-response functions for short-term
mortality from different individual studies on predicted health effects
associated with "as is" PM-10 and PM-2.5 in Philadelphia County
the effect of copollutants on relative risks for a change of 50 ng/m3
PM-10 or 25 Mg/m3 PM-2.5*
the effect of copollutants on predicted health effects associated with
"as is" PM in Philadelphia County
the effect of copollutants on predicted health effects associated with
PM after meeting current PM-10 standards in Los Angeles County
the effect of differing cutpoints on estimated mortality associated with
long-term exposure to PM-2.5 (no slope adjustment) in Philadelphia
County
the effect of historical previous air quality on estimated mortality
associated with long-term exposure to PM-2.5 (in Philadelphia
County)
Exhibits)
Exhibits 7.9 and 7.
Exhibits 7. 11 and 7
Exhibits 7. 13 and 7
Exhibits 7. 15 and 7
Exhibits 7.17 and 7
Exhibits 7. 19 and 7
10
.12
.14
.16
.18
.20
Exhibit 7.21
Exhibit 7.22
Exhibit 7.23
Exhibit 7.24
Exhibit 7.25
Exhibit 7.26
Exhibit 7.27
This sensitivity analysis is not location-specific. It examines the effect of having copollutants in the models
estimated by epidemiological studies on relative risks associated with specified changes in PM concentration.
The sensitivity analyses of alternative cutpoint models considered the effect of using
alternative concentration-response models. The exponential model estimated by most
Abr Associates, Inc.
p. 80
July 3, 1996, Revised
-------
epidemiological studies (and therefore used in most of the risk analyses (see Section 2,
equation 1)) assumes that there is no PM level at which the relationship between PM and the
health effect fundamentally changes. Using the relative risk version of the model (see
Appendix 3, equation 9), this means that there is a linear relationship between the natural
logarithm of the relative risk, ln(RR), and PM, as shown hi Exhibit 7.18. As an alternative to
this simple linear relationship, a "hockey stick" model was considered. The hockey stick
model is determined by (1) a cutpoint, below which the model is a horizontal line (i.e., the
slope is zero), and (2) the positive slope of the line corresponding to PM concentrations
greater than the cutpoint. Changing the cutpoint and/or the positive slope results hi different
hockey stick models.
Exhibit 7.7a compares the results of using a cutpoint with no slope adjustment and of
using a cutpoint while doubling the slope. In each case, there is no additional risk at or below
the cutpoint; that is, the relative risk at or below the cutpoint is equal to one (so that the
natural logarithm of the relative risk is zero). Appendix E of the Staff Paper (EPA, 1996b)
discusses the choice of the particular cutpoints presented. Philadelphia "as is" results, and
Los Angeles results assuming attainment of current standards are examined. The base case
results (which require no slope adjustment, since no cutpoint is imposed) are provided for
comparison. Doubling the slope roughly doubles the estimate of mortality associated with
short-term exposure to PM-2.5.
Exhibit 7.7a. Comparison of the Effect of Cutpoints with and without Slope Adjustment.
Mortality Associated with Short-Term Exposure to PM-2.5.
Cutpoint
Background
("Base Case")
lO^g/m3
18 Mg/m3
30 fjig/m3
Philadelphia County
no adjustment
1.8%
(1.1,4.4)
1.0%
(0.6, 1.3)
0.4%
(0.2, 0.5)
0.09%
(0.06,0.13)
slope doubled
.
1.9%
(1.2,2.6)
0.7%
(0.4, 1.0)
0.2%
(0.1,0.3)
Southeast Los Angeles County
no adjustment
2.9%
(1.7,3.9)
1.9%
(1.1,2.6)
1.1%
(0.7, 1.5)
0.5%
(0.3, 0.7)
slope doubled
3.7%
(2.2, 5.0)
2.2%
(1.3,3.0)
1.0%
(0.6, 1.4)
Exhibits 7.13 through 7.20 provide another perspective on slope adjustment. Two
different methods of adjusting the positive slope of the "hockey stick" concentration-response
function were used. The methods result in different slope adjustments being applied when
different cutpoints are selected.
Abt Associates, Inc.
81
Jul\ 3, 1996, Revised
-------
The first slope adjustment method preserves the area of the triangle formed by (1) the
x-axis, (2) a vertical line at the maximum observed PM concentration in the study that
estimated the original exponential concentration-response function, and (3) the original
concentration-response function itself (i.e., the linear relationship between ln(RR) and PM).
That is, the slope of the hockey stick is adjusted so that the area under the new line (down to
the x-axis and out to the vertical line at the maximum observed PM level in the study) is the
same as the corresponding area under the original function. The original function spans a
wider range on the x-axis, and a smaller range on the y-axis, than the adjusted function. That
is, to compensate for fewer PM-associated health effects at low concentrations (and no effects
at all below the cutpoint), the adjusted function rises more quickly than does the original
function.
If the actual PM concentrations where the original function was estimated were evenly
distributed along the x-axis, then the area under the original function (the triangle described
above) is an approximation of the total health effects predicted by the original function. In
general, however, distributions of PM concentrations tend to be skewed to lower
concentrations, with relatively long tails. A better approximation would take account of this.
However, such an approximation would predict an even steeper slope than the method used
here, since the days with lower PM concentrations would need to account for the health effects
previously accounted for by days with high PM concentrations.
The second slope adjustment method assumes that the relative risk associated with the
maximum observed concentration remains the same in the hockey stick model as in the
original model that did not assume a cutpoint. Thus the positive-sloped portion of the hockey
stick extends from the cutpoint to the relative risk achieved by the original function at the
maximum observed concentration. This method adjusts the slope by less than the first method.
Exhibit 7.8 illustrates the two slope adjustment methods. It is important to keep in
mind that these adjustment methods are illustrative, rather than definitive. Different choices of
slope adjustments can yield substantially different results. Proper evaluation of the effect of
cutpoints would require re-analysis of original health and air quality data, as noted above.
Exhibits 7.22, 7.23, and 7.24 show the effect of including copollutants in the analysis.
The figures presented in these exhibits derive from models that include PM and one other
pollutant at a time. No models considering three or more pollutants at a time are included.
Abt Associates, Inc. p. 82 July 3, 1996, Revised
-------
Exhibit 7.8
Sensitivity Analysis: Slope Adjustment
PM-10 Pooled Mortality Function
TJ
c
3
g
s
o
CO
CD
o
2
CO
(A
£
0)
.>
*3
n
0)
"SiT
0.20
0.16
0.12-
0.08'
0.04
0
"
^ '
Slope Adjustment ,''
Method 1 -^^^^ ,'' .^
, " ^^^^^
, ''^f''
'^^^^
Original Function <-^^ '
\^^ / "^w
^"''' -^ ^\
^^ ' " ^ ^^
^--' ^"^^ Slope Adjustment
,, -^^ Method 2
^--' ./^^
8 30 125 220 251
Eastern Example Highest
Background Outpoint PM-1 0 l_6VelS ((JO/m3) Observed
PM-10 Level
in Studies
Relative Risks shown are the risks associated with elevated PM-10 levels relative to the
risks associated with the background PM level (8 ug/m3).
Abl Associates, Inc.
p. 83
July 3, 1996, Revised
-------
Exhibit 7.9
Sensitivity Analysis: The Effect of Alternative Background Levels on
Predicted Health Effects Associated With "As-ls" PM-10
Philadelphia County, September 1992 - August 1993
Health Effects*
Mortality (all ages)
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
Lower Respiratory
Symptoms in Children
(A) Associated with short-term exposure
(B) Total Respiratory
(>64 years old)
(E) Ischemic Heart Disease
(>64 years old)
(C) COPD
(>64 years old)
(D) Pneumonia
(>64 years old)
(F) Congestive Heart Failure
(>64 years old)
(G) Lower Respiratory Symptoms (# of cases)
(8- 12 year olds)
(H) Lower Respiratory Symptoms (# of days)
(9-1 1 year old asthmatics)
Percent of Total Incidence Associated with PM-10 Above Background**
BASE CASE
= 8 ug/m3
1.1%
(0.8 -1.4)
2.4%
(1.5 -3.3)
3.7%
(2.5 -4.7)
1.9%
(1.3 -2.6)
0.8%
(0.3 -1.3)
1.4%
(0.7 - 2.1 )
17.5%
(15.3 -19.6)
6.8%
(2.4 - 10.9 )
Background
= 5ug/m3
1.3%
(1.0 -1.7)
2.87%
(1.8 -4.0)
4.4%
(3.1 -5.7)
2.3%
(1.6 -3.1)
1.0%
(0.4 -1.5)
° 1.7%
(0.8 -2.5)
20.8%
(18.2 -23.3)
8.2%
(2.9 - 13.0 )
Background
= 1 1 pg/m3
0.9%
(0.6 - 1.1 )
1.9%
(1.2 -2.7)
3.0%
(2.1 -3.8)
1.6%
(1.1 -2.1)
0.6%
(0.2 -1.0)
1.1%
(0.5 -1.7)
14.2%
(12.4 -15.9)
5.5%
(2.0 - 8.8 )
* Health effects associated with short-term exposure to PM.
** Health effects incidence was quantified across the range of PM concentrations observed in each study,
when possible, but not below background level Background PM-10 is assumed to be 8 ug/m3 .
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All numbers in parentheses are interpreted as 90% credible intervals based on
uncertainty analysis that takes into account both statistical uncertainty and
possible geographic variability. See text in Chapter 7 for details.
Sources of Concentration-Response (C-R) Functions:
(A) PM-10 C-R function based on pooled results from
studies in 10 locations.
(B) PM-10 C-R based on pooled results from 4 functions
(C) PM-10 C-R based on pooled results from 4 functions
(D) PM-10 C-R based on pooled results from 4 functions
(E) Schwartz & Morris, 1995
(F) Schwartz & Morris, 1995
(G) Schwartz, et al., 1994
(H) Pope etal., 1991
Abt Associates, Inc.
p. 84
July 3, 1996, Revised
-------
Exhibit 7.10
Sensitivity Analysis: The Effect of Alternative Background Levels on
Predicted Health Effects Associated With "As-ls" PM-2.5
Philadelphia County, September 1992 - August 1993
Health Effects*
Mortality (all ages)
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac (>64 years
Lower Respiratory
Symptoms in Childrer
(A) Associated with short-term exposure
(B) Total Respiratory
(all aqes)
(C) Ischemic Heart
Disease***
(D) Congestive
Heart Failure***
(E) Lower Respiratory Symptoms
(# of cases) (8-12 years old)
BASE CASE Background
= 3.5 uo/m3
1.8%
(1.1 -2.5)
2.0%
(0.5 - 3.5 )
0.7%
(0.3 -1.2)
1.3%
(0.6 - 2.0 )
20.0%
(10.3 -28.2)
Background
= 2.0 ua/m3
Z.1%
(1.2 -2.8)
2.2%
(0.6 - 3.9 )
0.8%
(0.3 -1.3)
1.5%
(0.7 -2.2)
22.2%
(11.5 -31.1)
ve Background"1
Background
= 5.0 ua/m3
1.6%
(1.0 -2.2)
1.8%
(0.5 - 3.1 )
0.7%
(0.3 -1.1)
1.2%
(0.6-1.8)
17.8%
(9.2 -25.2)
* Health effects associated with short-term exposure to PM.
** Health effects incidence was quantified across the range of PM concentrations observed in each study,
when possible, but not below background level. Background PM-2.5 is assumed to be 3.5 ug/m3.
*** PM-2.5 results based on using PM-2.5 mass as PM-10 mass in the PM-10 functions.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All the numbers in parentheses are interpreted as 90% credible intervals based on uncertainty
analysis that takes into account both statistical uncertainty and possible geographic variability.
See text in Chapter 7 for details.
Sources of Concentration-
Response (C-R) Functions:
(A) PM-2.5 C-R function based on
pooled results from 6
locations.
(B)Thurston, etal.. 1994
(C) Schwartz & Morris, 1995
(0) Schwartz & Morris, 1995
(E) Schwartz, etal., 1994
Abt Associates, Inc.
p. 85
July 3, 1996, Revised
-------
Exhibit 7.11
Sensitivity Analysis: The Effect of Alternative Background Levels on
Predicted Health Effects Associated With PM-10 After Attainment of Current Standards*
Southeast Los Angeles County, 1995
Health Effects**
Mortality (all ages)
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
-Ower Respiratory
Symptoms in Children
(A) Associated with short-term exposure
(A') Associated with short-term exposure
(Study done in Los Anaeles)
(B) Total Respiratory
(>64 years old)
(E) Ischemic Heart Disease
(>64 years old)
(C) COPD
(>64 years old)
(D) Pneumonia
(>64 years old)
(F) Congestive Heart Failure
(>64 years old)
(G) Lower Respiratory Symptoms (# of cases)
(8- 12 year olds)
(H) Lower Respiratory Symptoms (# of days)
(9-1 1 year old asthmatics)
BASE CASE
= 6 ua/m3
2.6%
(1.8 -3.3)
1.2%
(0.1 -2.2)
5.4%
(3.3 -7.4)
8.2%
(5.8 -10.5)
4.4%
(3.1 -5.8)
1.8%
(0.7 -2.9)
3.2%
(1.5 -4.8)
34.8%
(31.0 -38.4)
14.9%
(5.5 -23.0)
Background
= 4 uo/m3
2.7%
(1.9 -3.4)
1.2%
(0.1 -2.2)
5.7%
(3.5 - 7.9 )
8.7%
(6.1 - 11.1 )
4.7%
(3.3 -6.1 )
1.9%
(0.7 -3.1 )
3.4%
(1.6 -5.1 )
36.7%
(32.7 -40.3)
15.7%
(5.8 -24.2)
m BacKaround*"
Background
= 8ua/m3
2.4%
(1.7 -3.1 )
1.2%
(0.1 -2.2)
5.1%
(3.1 -7.0)
7.7%
(5.4 -9.9)
4.2%
(2.9 -5.5)
1.7%
(0.7 - 2.8 )
3.0%
(1.5 -4.5)
33.0%
(29.4 -36.3)
14.0%
(5.2 -21.7)
* Current standards are 50 ug/m3 annual average PM-10. 150 ug/m3 second daily maximum PM-10.
" Health effects associated with short-term exposure to PM.
*** Health effects incidence was quantified across the range of PM concentrations observed in each study,
when possible, but not below background level.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All numbers in parentheses are interpreted as 90% credible intervals based on
uncertainty analysis that takes into account both statistical uncertainty and
possible geographic variability. See text in Chapter 7 for details.
Sources of Concentration-Response (C-R) Functions:
(A) PM-10 C-R function based on pooled results from
studies in 10 locations.
(A) Kinney et al., 1995
(B) PM-10 C-R based on pooled results from 4 functions
(C) PM-10 C-R based on pooled results from 4 functions
(D) PM-10 C-R based on pooled results from 4 functions
(E) Schwartz & Morris, 1995
(F) Schwartz & Morris, 1995
(G) Schwartz, etal., 1994
(H)Popeetal, 1991
Abt Associates, Inc.
p. 86
July 3, 1996, Revised
-------
Exhibit 7.12
Sensitivity Analysis: The Effect of Alternative Background Levels on
Predicted Health Effects Associated With PM-2.5 After Attainment of Current Standards*
Southeast Los Angeles County, 1995
Health Effects"
Mortality (all ages)
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac (>64 years old)
Lower Respiratory
Symptoms in Children
(A) Associated with
short-term exposure
(B) Total Respiratory
(all ages)
(C) Ischemic Heart
Disease*"*
(D) Congestive
Heart Failure*"*
(E) Lower Respiratory Symptoms
(# of cases) (8-1 2 years old)
Percent of Total incidence Associated with PM-2.5 Above Background*"*
BASE CASE
Background = 3.5 pg/m3
2.9%
(1.7 -3.9)
6.1%
(1.6 -10.5)
1.1%
(0.4 -1.8)
2.0%
(1.0 -3.0)
28.7%
(15.4 -39.0)
Background
= 1.0pg/m3
3.1%
(1.9 -4.2)
6.5%
(1.8 -11.2)
1.2%
(0.5 -2.0)
2.1%
(1.0 -3.2)
30.6%
(16.5 -41.4)
Background
- 4.0 M9/m3
2.7%
(1.6 -3.7)
5.6%
(1.5 -9.7)
1.1%
(0.4 -1.7)
1.8%
(0.9 -2.8)
26.7%
(14.4 -36.3)
* Current standards are 50 ug/m3 annual average PM-10, 150 ug/m3 second daily maximum PM
** Health effects associated with short-term exposure to PM.
*** Health effects incidence was quantified across the range of PM concentrations observed in each study,
when possible, but not below background leve
**** PM-2.5 results based on using PM-2.5 mass as PM-10 mass in the PM-10 functions.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All the numbers in parentheses are interpreted as 90% credible intervals based on
uncertainty analysis that takes into account both statistical uncertainty and possible
geographic variability. See text in Chapter 7 for details.
Sources of Concentration-
Response (C-R) functions:
(A) PM-2.5 C-R function
based on pooled results
from 6 locations.
(B)Thurston, etai., 1994
(C) Schwartz & Morris, 199
(D) Schwartz & Morris, 199
(E) Schwartz, etal., 1994
Abt Associates. Inc.
p. 87
July 3, 1996. Revised
-------
Exhibit 7.13
Sensitivity Analysis: The Effect of Alternative Cutpoint Models on
Predicted Health Effects Associated With "As-ls" PM-1
Slope Adjustment Method 1*
Philadelphia County, September 1992 - August 1993
Mortality (all ages)
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
Lower Respiratory
Symptoms in Children
Health Effects"
(A) Associated with short-term exposure
(B) Total Respiratory
(>64 years old)
(C) Ischemic Heart Disease
(>64 years old)
(D) Congestive Heart Failure
(>64 years old)
(E) Lower Respiratory Symptoms (# of cases)
(8- 12 year olds)
Percent of Total Incidence Associated with PM-10 Above Cutpoint
BASE CASE
= 8 ug/m3
1.1%
(0.8 -1.4)
2.4%
(1.5 -3.3)
0.8%
(0.3 -1.3)
1.4%
(0.7 - 2.1 )
17.5%
(15.3 - 19.6 )
Cutpoint
= 20ug/m3
0.4%
(0.3 - 0.6)
1.3%
(0.8-1.7)
0.3%
(0.1 -0.4)
0.5%
(0.2 - 0.2)
9.3%
(5.4 - 12.7)
Cutpoint
= 30ug/m3
0.2%
(0.1-0.2)
0.7%
(0.4 - 0.9)
0.1%
(0.1-0.2
0.2%
(0.1-0.1)
6.3%
(3.9-8.1)
Cutpoint
= 40ug/m3
0.1%
(0.0-0.1)
0.4%
(0.2 0.6)
0.1%
(0.0 - 0.1)
(, 0.1%
(0.1-0.2)
4.7%
(3.4 - 5.5)
* Two methods examine the potential impact of a concentration-response function having a steeper slope (i.e., larger coefficient) above
specified outpoints In both methods the slope below the cutpoint Is set = 0, while the slope above the cutpoint is set to be greater
than the slope in the original study. In Adjustment Method 1, the cutpoint C-R relationship is modeled to intersect with the original
relationship, exceeding the RRs predicted for the original study at higher concentrations. The relationship was modeled to match the reduction in
the range of PM concentrations upon application of the cutpoint with an identical percentage increase in the risk observed
at the highest concentration. Method 2 estimates a smaller increase in the slope. See text in Section 7 for details.
"Health effects associated with short-term exposure to PM
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All numbers in parentheses are interpreted as 90% credible intervals based on
uncertainty analysis that takes into account both statistical uncertainty and
possible geographic variability. See text in Chapter 7 for details.
Sources of Concentration-
Response (C-R) functions:
(A) C-R function based on pooled
results from 10 locations.
(B) C-R function based on pooled
results from 4 locations.
(C) Schwartz & Morris, 1995
(D) Schwartz A Morris, 1995
(E) Schwartz, et al., 1994
Ahi Associates, Inc.
p. 88
July 3, 1996, Revised
-------
Exhibit 7.14
Sensitivity Analysis: The Effect of Alternative Outpoint Models on
Predicted Health Effects Associated With "As-ls" PM-10
Slope Adjustment Method 2*
Philadelphia County, September 1992 - August 1993
Health Effects"
Mortality (all ages)
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
Lower Respiratory
Symptoms in Children
(A) Associated with short-term exposure
(B) Total Respiratory
(>64 years old)
(C) Ischemic Heart Disease
(>64 years old)
(D) Congestive Heart Failure
(>64 years old)
(E) Lower Respiratory Symptoms (# of cases)
(8- 12 year olds)
Percent of Total Incidence Associated with PM-10 Above Cutpolnt
BASE CASE Background
= 8 ug/m3
1.1%
(0.8 -1.4)
2.4%
(1.5 -3.3)
0.8%
(0.3 - 1.3 )
1.4%
(0.7 - 2.1 )
17.5%
(15.3 -19.6)
Outpoint
= 20 pg/m3
0.4%
(0.3 - 0.5)
1.0%
(0.6-1.3)
0.3%
(0.1-0.4)
0.5%
(0.2-0.7)
7.9%
(4.5-11.0)
Outpoint
= 30ug/m3
0.1%
(0.1-0.2)
0.4%
(0.3-0.6)
0.1%
(0.0 - 0.2)
0.2%
(0.1 -0.3)
4.1%
(2.4-5.6)
Outpoint
= 40 ugym3
0.1%
(0.0-0.1)
0.2%
(0.1-0.3)
0.0%
(0.0-0.1)
0.1%
(0.0-0.1)
2.5%
(1.5-3.2)
* Two methods examine the potential impact of a concentration-response function having a steeper slope (i.e.. larger coefficient)
above specified cutpoints. In both methods the slope below the cutpoint is set = 0, while the slope above the outpoint is set to
be greater than the slope in the original study In Adjustment Method 2, the slope is increased so that the new C-R function estimates
the same health risk at the highest observed PM value as the original function. Method 1 estimates a larger increase in the slope.
See text in Section 7 for details.
"Health effects associated with short-term exposure to PM.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All numbers in parentheses are interpreted as 90% credible intervals based on
uncertainty analysis that takes into account both statistical uncertainty and
possible geographic variability. See text in Chapter 7 for details.
Sources of Concentration-
Response (C-R) functions:
(A) C-R function based on pooled
results from 10 locations.
(B) C-R function based on pooled
results from 4 locations.
(C) Schwartz & Morris, 1995
(D) Schwartz & Morris, 1995
(E) Schwartz, et al., 1994
Abt Associates, Inc.
p. 89
July 3, 1996, Revised
-------
Exhibit 7.15
Sensitivity Analysis: The Effect of Alternative Outpoint Models on
Predicted Health Effects Associated With "As Is" PM-2.5
Slope Adjustment Method 1*
Philadelphia County, September 1992 -August 1993
1
Health Effects"
Mortality Tall ages)
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
Lower Respiratory
Symptoms in Childre
(A) Associated with
short-term exposure
(B) Total Respiratory
(all aqes)
(C) Ischemic Heart
(>64 vears old)
(D) Congestive Heart
(>64 vears old)
(E) Lower Respirator
(8- 12 vears old)
BASE CASE:
= 3.5 uo/rn3
1.8%
(1.1 -2.5)
2.0%
(0.5 -3.5)
0.7%
(0.3 -1.2)
1.3%
(0.6 -2.0)
20.0%
(10.3 -28.2)
Cutpoint
= 10ua/m3
1.1%
(0.7-1.5)
1.4%
(0.4 - 2.4)
0.4%
(0.1 - 0.6)
0.7%
(0.3-1.0)
13.2%
(7.2-18.6)
Cutpoint
* 18 ua/m3
6.5%
(0.3 - 0.6)
0.8%
(0.2-1.4)
0.2%
(0.1 -0.3)
0.4%
(0.2 - 0.5)
9.9%
(5.7-13.1)
* Two methods examine the potential impact of a concentration-response function having a steeper slope (i.e., larger coefficien
specified cutpoints. In both methods the slope below the cutpoint is set = 0, while the slope above the cutpoint Is set to be
than the slope in the original study. In Adjustment Method 1, the cutpoint C-R relationship is modeled to intersect with the
relationship, exceeding the RRs predicted for the original study at higher concentrations. The relationship was modeled to
the range of PM concentrations upon application of the cutpoint with an identical percentage increase in the risk observed
at the highest concentration. Method 2 estimates a smaller increase in the slope. See text in Section 7 for details.
** Health effects associated with short-term exposure to PM.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All the numbers in parentheses are interpreted as 90% credible intervals based on uncertainty analysis
that takes into account both statistical uncertainty and possible geographic variability.
See text in Section 7 for details.
Sources of Concentration-
Response (C-R) functions:
(A) C-R function based on pooled
results from six locations.
(B) Tnurston, et al., 1994
(C) Schwartz & Morris, 1995
(D) Schwartz & Morris, 1995
-------
Exhibit 7.16
Sensitivity Analysis: The Effect of Alternative Cutpoint Models on
Predicted Health Effects Associated With "As Is" PM-2.5
Slope Adjustment Method 2*
Philadelphia County, September 1992 August 1993
Health Effects"
Mortality (all ages)
Hospital Admissions
^espiratorv
Hospital Admissions
Cardiac
-ower Respiratory
SvmDtoms
(A) Associated with snort-term exposure
(B) Total Respiratory
fall aaesl
(C) Ischemic Heart
(>64 vears old)
(D) Congestive Heart
(>Q4 vears old)
{E} Lower Respiratory
(8 - 12 vears old)
BASE CASE: Background
* 3.5 u^i/m3
1.8%
(1.1 -2.5)
2.0%
(0.5 -3.5)
07%
(0.3 -1.2)
1.3%
(0.6 -2.0)
20.0%
(10,3 -28.21
inrr.i7;iiaUi',B*»j*rnwr,niair
^^^^^^^^^^^^^^^1
^BnirRiMlH
1.0%
(0.6-1.41
1.2%
(0.3 - 2.2)
0.4%
(0.2-06)
0.7%
(0.3 - 1 0)
12.2%
(6.5 - 17.3)
Cutpoint
= 18uoAn3
0.41*
10 3 - OB)
0.6%
(0.2 . 1.1)
0.2%
(0.1.0.3)
0.3%
(0.2-0.51
7.1%
(3.0-9.8)
Outpoint
(0.1 - 0.2)
0.2%
(0.1-04)
0.1%
(0.0-0.1)
0.1%
(0.0-01)
3.8%
12.4. 4.7)
Two methods examine the potential impact of a concentration-response function having a steeper slope (i.e.. larger coefficient)
above specified cutpoints In both methods the slope below the Cutpoint is set = 0. while the slope above the Cutpoint Is set to
be greater than the slope in the original study In Adjustment Method 2, the slope is increased so that the new C-R function estimates
the same health risk at the highest observed PM value as the original function Method 1 estimates a larger increase in the slope.
See text in Section 7 for details.
"Health effects associated with short-term exposure to PM
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All the numbers in parentheses are interpreted as 90% credible intervals based on uncertainty analysis
that takes into account both statistical uncertainty and possible geographic variability.
See text in Chapter 7 for details.
Sources of Concentration-
Response (C-R) functions:
(A) C-R function based on pooled
results from six locations.
(B)Thurston. !!.. 1994
(C) Schwartz & Morris, 1995
(0) Schwartz & Morris. 199S
(E) Schwartz el al.. 1994
Abt Associates, Inc.
p. 91
July 3, 1996, Revised
-------
Exhibit 7.17
Sensitivity Analysis: The Effect of Alternative Outpoint Models on
Predicted Health Effects Associated With PM-10 After Attainment of Current Standards*
Slope Adjustment Method 1**
Southeast Los Angeles County, 1995
Health Effects'"
Mortality (all ages)
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
Lower Respiratory
Symptoms in Children
(A) Associated with short-term exposure
(A1) Associated with short-term exposure
(Study done in Los Anqeles)
(B) Total Respiratory
(>64 years old)
(C) Ischemic Heart
(>64 years old)
(D) Congestive Heart
(>64 years old)
(E) Lower Respiratory Symptoms (# of cases'
(8-12 year olds)
MJ»cl^jM»i^jl-IjLl!lAjllUI'H
JH^HEE^^REH^^I
2.6%
(1.8 -3.3J
1.2%
(0.1 -2.2)
5.4%
(3.3 -7.41
1.8%
(0.7 -2.9)
3.2%
(1.6 -4.8)
34.8%
(31.0 -38.4)
1.8%
(1.3-2.3)
1.1%
(0.1-2.1)
4.6%
(2.8-6.3)
1.0%
(04-17)
1.8%
(0.9 - 0.9)
26.4%
(16.1 - 34.6)
1.2%
(0.9-1.6)
0.8%
(0.1 -1.5)
4.1%
(2.6 - 5.6)
0.9%
(04-15)
1.7%
(0.8 - 0.8)
27.7%
(18.4-34.1)
Outpoint
40 ua/m3
0.8%
(0.6-1.0)
0.6%
(0.1-1.1)
3.7%
(2.3-4.9)
0.9%
(03-1.4)
1.5%
(0.7 - 2.2)
27.2%
(21.4-30.3)
* Current standards are 50 ug/m3 annual average PM-10. 150 ug/m3 second daily maximum PM-1
" Two methods examine the potential impact of a concentration-response function having a steeper slope (i.e., larger coefficient) above
specified cutpoints. In both methods the slope below the cutpoint is set = 0, while the slope above the outpoint is set to be greater
than the slope in the original study. In Adjustment Method 1. the cutpoint C-R relationship is modeled to intersect with the original
relationship, exceeding the RRs predicted for the original study at higher concentrations. The relationship was modeled to match the reduction in
the range of PM concentrations upon application of the cutpoint with an identical percentage increase in the risk observed
at the highest concentration. Method 2 estimates a smaller increase in the slope. See text for further information.
*" Health effects associated with short-term exposure to PM.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All numbers in parentheses are interpreted as 90% credible intervals based on
uncertainty analysis that takes into account both statistical uncertainty and
possible geographic variability. See text in Chapter 7 for details.
Sources of Concentration-
Responw (C-R) function*:
(A) C-R function based on pooled
results from 10 locations.
-------
Exhibit 7.18
Sensitivity Analysis: The Effect of Alternative Outpoint Models on
Predicted Health Effects Associated With PM-10 After Attainment of Current Standards*
Slope Adjustment Method 2**
Southeast Los Angeles County, 1995
Health Effects'"
Mortality (all ages)
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
Lower Respiratory
Symptoms in Children
(A) Associated with short-term exposure
(A1) Associated with short-term exposure
(Study done in Los Anaeles)
(B) Total Respiratory
(>64 years old)
(C) Ischemic Heart
(>64 years old)
(D) Congestive Heart
(>64 years old)
(E) Lower Respiratory Symptoms (# of cases
(8-12 year olds)
Percent of Total Incidence Associated with PM-10 Abov
BASE CASE Background
8 ughn3
2.6%
(1.8 -3.3)
1.2%
(0.1 -2.2)
5.4%
(3.3 -7.4)
1.8%
(0.7 -2.9)
3.2%
(1.5 -4.8)
34.8%
(31.0 -38.4)
Outpoint
= 20uo/m3
1.6%
(1.2-2.1)
1.0%
(0.1 -2.0)
3.5%
(2.2-4.8)
1.1%
(0.4-1.8)
1.9%
(0.9 - 2.8)
22.9%
(13.7-30.5)
cutooint
Outpoint
= 30 uo/m3
(0.8-1.4)
0.7%
(0.1-1.3)
2.7%
(1.7-3.7)
0.8%
(0.3-1.3)
1.4%
(0.7-2.1)
19.4%
(12.0-25.2)
Cutpdnt
(0.5 - 0.9)
0.5%
(0.0 0.9)
2.0%
(1.3-2.7)
0.6%
(0.2-1.0)
1.1%
(0.5-1.6)
18.7%
(11.3-20.5)
* Current standards are 50 ug/m3 annual average PM-10. 150 ug/m3 second daily maximum PM-1
** Two methods examine the potential impact of a concentration-response function having a steeper slope (i.e., larger coefficient)
above specified cutpoints. In both methods the slope below the outpoint is set = 0, while the slope above the cutpoint is set to
be greater than the slope in the original study. In Adjustment Method 2. the slope is increased so that the new C-R function estimates
the same health risk at the highest observed PM value as the original function. Method 1 estimates a larger increase in the slope.
"Health effects associated with short-term exposure to PM.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All numbers in parentheses are interpreted as 90% credible intervals based on
uncertainty analysis that takes into account both statistical uncertainty and
possible geographic variability. See text in Chapter 7 for details.
Sources of Concentration-
Response (C-R) functions:
(A) C-R function based on pooled
results from 10 locations.
(A'JKinneyetal.. 1995
(B) C-R function based on pooled
results from 4 locations.
(C) Schwartz & Morris, 1995
(D) Schwartz & Morris, 1995
(E) Schwartz, etal., 1994
Abt Associates, Inc.
p. 93
July 3, 1996. Revised
-------
Exhibit 7.19
Sensitivity Analysis: The Effect of Alternative Outpoint Models on
Predicted Health Effects Associated With PM-2.5 After Attainment of Current Standards*
Slope Adjustment Method 1**
Southeast Los Angeles County, 1995
Health Effects'"
Mortality (all ages)
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
_ower Respiratory
Symptoms in Children
(A) Associated with
with short-term exposure
(B) Total Respiratory
(all ages)
(C) Ischemic Heart
Disease (>64 years old)
(D) Congestive Heart Failure
(>64 years old)
(E) Lower Respiratory Symptoms
(8- 12 years old)
Percent of Total Incidence Associated with PM-2.5 Above Cutpomt
BASE CASE:
Backgrount = 3.5 |ig/m3
2.9%
(1.7 -3.9)
6.1%
(1.6 -10.5)
1.1%
(0.4 -1.8)
2.0%
(1.0 -3.0)
28.7%
(15.4 -39.0)
Outpoint
= 10ug/m3
2.1%
(1.3-2.8)
4.9%
(1.3-8.5)
0.7%
(0.3-1.1)
1.2%
(0.6-1.8)
22.0%
(12.5-29.7)
Cutpoint
18 pg/m3
1.3%
(0.8-1.8)
4.0%
(1.1-7.0)
0.6% i
(0.2 - 0.9)
1.0%
(0.5-1.5)
21.0%
(13.2-26.5)
Cutpoint
30 ug/m3
O.f%
(0.4-1.0)
3.4%
(0.9 - 5.9)
0.4%
(0.2 - 0.7)
0.7%
(0.4-1.1)
19.7%
(17.0-20.9)
* Current standards are 50 ug/m3 annual average PM-10,150 ug/m3 second daily maximum PM-1
" Two methods examine the potential impact of a concentration-response function having a steeper slope (i.e., larger coefficient) above
specified cutpoinls. In both methods the slope below the cutpoint is set = 0, while the slope above the cutpoint is set to be
than the slope in the original study. In Adjustment Method 1, the cutpoint C-R relationship is modeled to intersect with the o
relationship, exceeding the RRs predicted for the original study at higher concentrations. The relationship was modeled to match the
reduction in the range of PM concentrations upon application of the cutpoint with an identical percentage increase in the risk observed
at the highest concentration. Method 2 estimates a smaller increase in the slope. See text for further information.
*** Health effects associated with short-term exposure to PM.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All the numbers in parentheses are interpreted as 90% credible intervals based on uncertainty analysis
that takes into account both statistical uncertainty and possible geographic variability.
See text in Section 7 for details.
Sources of Concentration-
Response (C-R) functions:
(A) C-R function based on
pooled results from
six locations
(B)Thurston, etal., 1994
(C) Schwartz & Moms, 1995
(D) Schwartz & Morris, 1995
(E) Schwartz etal., 1994
Abt Associates, Inc.
p. 94
July 3, 1996, Revised
-------
Exhibit 7.20
Sensitivity Analysis: The Effect of Alternative Outpoint Models on
Predicted Health Effects Associated With PM-2.5 After Attainment of Current Standards*
Slope Adjustment Method 2**
Southeast Los Angeles County, 1995
Health Effects'"
Mortality (all ages)
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
Lower Respiratory
Symptoms in Children
(A) Associated with
short-term exposure
(B) Total Respiratory
(all ages)
(C) Ischemic Heart
Disease (>64 years old)
(O) Congestive Heart
Failure (>64 years old)
(E) Lower Respiratory
Symptoms (8-12 years
Percent of Total Incidence Associated with PM-2.5 Above Cutpoint
BASE CASE:
Background = 3.5 ug'mJ
J.9%
(1.7 -3.9)
6.1%
(1.8 -10.5)
1.1%
(0.4 -1.8)
2.0%
(1.0 -3.0)
28.7%
(15.4 -39.0)
Cutpoint
= 10ug/m3
1.9%
(1.2-2.7)
4.4%
(1.2-7.6)
0.7%
(0.3-1.2)
1.3%
(0.6 - 1 .9)
20.4%
(11.5-27.9)
Cutpoint
= 18uoym3
1.2%
(0.7-1.7)
3.1%
(0.8-S.4)
0.5%
(0.2 - 0.8)
0.9%
(0.4-1.3)
15.8%
(9.3 - 20.9)
Cutpoint
« 30uo/m3
(0.4-0.8)
2.0%
(0.5-3.4)
0.3%
(0.1-0.5)
0.5%
(0.2 - 0.8) '
13.1%
(9.1 - 15.5)
Current standards are 50 ug/m3 annual average PM-10.150 ug/m3 second daily maximum PM-
* Two methods examine the potential impact of a concentration-response function having a steeper slope (i.e.. larger coefficie
above specified cutpoints. In both methods the slope below the cutpoint is set = 0. while the slope above the outpoint is s
be greater than the slope in the original study. In Adjustment Method 2, the slope is increased so that the new C-R function estimates
the same health risk at the highest observed PM value as the original function. Method 1 estimates a larger increase in the slope.
"Health effects associated with short-term exposure to PM.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All the numbers in parentheses are interpreted as 90% credible intervals based on uncertainty analysis
that takes into account both statistical uncertainty and possible geographic variability.
See text in Chapter 7 for details.
Sources of Concentration-
Response (C-R) functions:
(A) C-R function based on
pooled results from
six locations.
(B) Thurston. et a!.. 1994
(C) Schwartz & Morris, 199
(D) Schwartz & Morris. 199
(E) Schwartz et al.. 1994
Abt Associates, Inc.
p. 95
My 3, 1996, Revised
-------
Exhibit 7.21
Sensitivity Analysis: Effect of Combining Different Averaging Times
In Pooled Short-Term Exposure Mortality Functions on
Predicted Health Effects Associated With "As-ls" PM-10
Philadelphia County, September 1992 - August 1993
Matching study and data
averaging times
Using 2-day average PM data
Using 1-day average PM data
Using 5-day average PM data
Percent of Total Incidence Associated with PM-10 Above Background*
BASE CASE"
(10 studies)
2-day average PM
1.1%
(0.8 -1.4)
2-day average PM
same
1-day average PM
1.1%
(0.8 -1.4)
5-day average PM
1.1%
(0.8 -1.4)
Studies using 1-day
(1 study)
1 -day average PM
0.4%
(0.0 -0.8)
2-day average PM
0.4%
(0.0 -0.8)
Studies using 2-day
(7 studies)
2-day average PM
1.0%
(0.5 -1.5)
2-day average PM
same
Studies using 3-5 day
(2 studies)
5-day average PM
1.8%
(1.3 -2.4)
2-day average PM
1.9%
(1.3 -2.4)
Health effects incidence was quantified across the range of PM concentrations observed in each study,
when possible, but not below background level. Background PM-10 is assumed to be 8 ug/m3 .
** The base case is a random-effects pooled function used with 2-day average PM data
All other pooled functions are also random effects, except the pooled function derived from
studies using 3-5 day average PM data, for which a fixed effects model was used, since
it is not possible to calculate a random effects model for those two functions.
The numbers in parentheses for pooled functions are NOT standard confidence int
All numbers in parentheses are interpreted as 90% credible intervals based on
uncertainty analysis that takes into account both statistical uncertainty and
possible geographic variability. See text in Chapter 7 for details.
The studies that contribute to the pooled
function are:
1-day: Kinneyetal., 1995 (Los Angeles)
2-day: Ito and Thurston, 1996 (Chicago)
Schwartz et al. 1996 (Boston, MA;
Knoxville, TN; St. Louis. MO;
Steubenville, OH; Portage, Wl;
Topeka, KS)
3-day: Schwartz 1993 (Birmingham, AL)
5-day: Pope et al., 1992 (Utah Valley)
Abt Associates, Inc.
p. 96
July 3, 1996, Revised
-------
Exhibit 7.22
Sensitivity Analysis: The Effect of Considering Different Epidemiology
Studies Relating Mortality and Short-Term Exposure to PM on
Estimated Annual Mortality Risks Associated with "As Is" PM Concentrations
in Philadelphia County, September 1992- August 1993 (for base case assumptions)
Study
h'ooied Tunciion
Ito&Thurston 1996
Kinney et al. 1995
Popeet al. 1992
Schwartz 1993
Schwartz et al., 1996a
Location
Chicago
Los Angeles
Utah Valley
Birmingham, AL
Boston
Knoxville, TN
St. Louis
Steubenville, OH
Madison, Wl
Topeka, KS
Associated wun KIVI-IU
Incidence
uu
(160 - 290)
170
(70 - 270)
80
(10-160)
420
(270 - 560)
300
(100-490)
360
(230 - 480)
280
(490 - 490)
160
(70 - 290)
230
(30 - 430)
140
(-130-410)
-150
(-560 - 250)
Percent of Total
I.V/o
(0.8-1.4)
0.8%
(0.3-1.3)
0.4%
(0.0 - 0.8)
2.1%
(1.4-2.8)
1.5%
(0.5 - 2.4)
1.8%
(1.1-2.4)
1.4%
(2.4 - 2.4)
0.8%
(0.3-1.4)
1.1%
(0.2-2.1)
0.7%
(-0.6 - 2.0)
-0.7%
(-2.7-1.2)
Mssociaieo wun rm-z.a
Incidence
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
Percent of Total
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
Health effects incidence was quantified across the range of PM concentrations observed in each study, when possible, but not
below background level. Background PM-10 is assumed to be 8 ug/m3; background PM-2.5 is assumed to be 3.5 ug/m3.
The presence of negative numbers (for Madison, Wl and Topeka, KS) is due to statistical uncertainty in the estimation
of relative risks, and does not reflect a belief that increased particulate matter pollution may actually be beneficial to health.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All the numbers in parentheses are interpreted as 90% credible intervals based on uncertainly analysis
that takes into account both statistical uncertainty and possible geographic variability.
See text in Section 7 for details.
Abt Associates, Inc.
p. 97
July 3, 1996, Revised
-------
Table 7.23
Sensitivity Analysis: Effect of Copollutants
Relative Risks for Change of 50 ug/m3 PM-10 or 25 ug/m3 PM-2.5
End point
Mortality
Hospital
Admissions
All respiratory
(all ages)
All respiratory
(ages >64)
Pneumonia
(ages >64)
COPD
(ages >64)
Ischemic Heart Disease
Congestive Heart Failure
Study, Pollutant. & Location
Ito & Thurston 1995. PM-10
Chicago
Kmney et al., 1995. PM-10
Los Angeles
Pope 1994. PM-10
Utah Valley, summer only
Thurston et al.. 1994, PM-2 5
Ontario, Canada
Schwartz 1995, PM-10
New Haven
Schwartz 1995, PM-10
Tacoma
Schwartz 1994, PM-10
Minneapolis/St. Paul
Schwartz 1994, PM-10
Detroit
Schwartz 1994, PM-10
Detroit
Schwartz & Morris 1995. PM-10
Detroit
Schwartz & Morris 1995, PM-10
Detroit
Relative risk
no copollutant
1.02
(1.02-1.04)
1.02
(1.00-1.06)
1.11
(0.95-1.31)
0.086*
(0.024 - 0.16 )
1.06
(1.01 - 1.12)
1.10
(1.04-1.16)
1.028
(1.011 -1.047)
1.050
(1.024-1.077)
Relative Risk with
daily average
S02
1.07
(1.02-1.13)
1.11
(1.03-1.19)
1.024"
(1.005-1.043)
Relative Risk with
daily 1-hour maximum
CO
1.02
(0.99-1.04)
1.025
(1.007-1.044)
1.038
(1.011 - 1.064)
Relative Risk with
daily average
O3
1.02
(1.01-1. OS)
1.14
(0.96-1.37)
1.09
(1.01-1.18)
1.12
(0.98-1.26)
1.0*
(1.02-1.14)
1.06
(1.03-1.09)
1.10
(1.06-1.16)
Relative Risk with
daily 1-hour maximum
O3
1.02
(1.00-1.05)
1.19
(1.00-1.43)
0.045*
(-0.028 - 0.12 )
Results presented In bold come from functions used in the base case analysis.
The number of significant digits given for each relative risk is the same as the number reported in the original study.
* Thurston et al. 1994 provides a function relating changes in PM to changes in the number of cases.
The relative risk calculated from this coefficient may vary widely from location to location, depending on baseline incidences.
Therefore, the coefficient, adjusted to a rate per 100,000 people, is reported, instead of a relative risk.
" Based on 1-hour maximum SO2.
Abi Associates, Inc.
p. 98
July 3, 1996, Revised
-------
Exhibit 7.24
Sensitivity Analysis: Effect of Co pollutants
Predicted Health Effects Associated With "As-ls" PM*
Philadelphia County, September 1992 -August 1993
Health Effects
ivionaiiiy
Hospital
Admissions
All respiratory
iaqes >64)
All respiratory
(ages >64)
Pneumonia
(ages >64)
COPD
(aqes >64)
Ischemic Heart Diseas
Congestive Heart Failu
Study & Location
iio & i nursioh iyyo, ^ivi-iu
Chicago
Kinneyetal., 1995. PM-10
Los Anqeles
Pope 1994, PM-10
Utah Valley, summer only
Thurstonetal., 1994, PM-2.
Ontario, Canada
Schwartz 1995, PM-10
New Haven
Schwartz 1995, PM-10
Tacoma
Schwartz 1994, PM-10
Minneapolis/St. Paul
Schwartz 1994, PM-10
Detroit
Schwartz 1994, PM-10
Detroit
Schwartz & Morris 1995, PM
Detroit
Schwartz & Morris 1995. PM
Detroit
percent OT loiai inciaonce associated wun i*wi BDOVG oacKS
with
no cooollutant
U.BYo
(0.3 - 1.3)
0.4%
(0.0 - 0.8)
3.0%
(-1.5-7.2)
2.0%
(0.5 - 3.5)
2.4%
(0.3 - 4.5)
3.2%
(-0.2 - 6.4)
0.8%
(0.3-1.3)
1.4%
(0.7-2.1)
with
daily average
S02
1.9%
(0.6 - 3.4)
2.9%
(1.0-4.7)
0.7%"
(0.1-1.2)
with daily
-hour maximu
CO
0.3%
(-0.0 - 0.7)
0.7%
(0.2-1.2)
1.1%
(0.3-1.8)
with
daily average
O3
U.O70
(0.2 - 0.9)
3.7%
(-1.3-8.3)
2.4%
(0.4 - 4.6)
3.2%
(-0.2 - 6.4)
2.2%
(0.6 - 3.8)
1.6%
(0.7 2.5)
2.8%
(1.5-4.2)
with daily
1 -hour maximu
O3
0.4%
(0.0 - 0.8)
4.8%
(-0.2 - 9.4)
1.1%
(-0.7 - 2.8)
Health effects associated with short-term exposure to PM. Incidence was quantified across the range of PM concentrations observed in each stu
but not below background PM levels, assumed to be 8 ug/m3 for PM-10 and 3.5 ug/m3 for
" Based on 1-hour maximum SO2.
The numbers in parentheses for pooled functions are NOT standard confidence intervals. All numbers in parenthes
as 90% credible intervals based on uncertainty analysis that takes into account both statistical uncertainty and possi
variability. See text for details.
Abt Associates, Inc.
p. 99
July 3, 1996, Revised
-------
Exhibil 7.25
Sensitivity Analysis: Effect of Copollutants on
Predicted Health Effects Associated With PM* After Attainment of Current Standards**
Southeast Los Angeles County, 1995
Health Effects
wionamy
Hospital
Admissions
All respiratory
(all ages)
All respiratory
(ages >64)
Pneumonia
(ages >64)
COPD
(apes >64)
Ischemic Heart Disease
Congestive Heart Failure
Study & Location
no & i nurston iyyt>, KM-IU
Chicago
Kinneyetal.. 1995. PM-10
Los Angeles
Pope 1994. PM-10
Utah Valley, summer only
Thurston el al.. 1994. PM-2.5
Ontario, Canada
Schwartz 1995. PM-10
New Haven
Schwartz 1995. PM-10
Tacoma
Schwartz 1994, PM-10
Minneapolis/St. Paul
Schwartz 1994, PM-10
Detroit
Schwartz 1994. PM-10
Detroit
Schwartz & Morris 1995, PM-10
Detroit
Schwartz 8 Morris 1995. PM-10
Detroit
rercBm OT loiai inciavn
with
no copollutant
l.STfr
(0.7-3.1)
1.2%
'(0.1 - 2.2)
6.7%
(-3.6-15.7)
NA
NA
5.5%
(0.6 10.1)
7.2%
(-0.4 14.0)
1.8%
(0.7 - 2.9)
3.2%
(1.5-4.8)
;a associaura wnn r*m a
with
daily average
SO2
4.4%
(1.3-7.6)
6.6%
(2.2-10.5)
1.5%*~
(0.3 - 2.8)
love oacKgrouno
with dally
1 -hour maximum
CO
1.0%
(-0.1 - 2.0)
1.6%
(0.4 - 2.8)
2.4%
(0.7 - 4.0)
with
daily average
03
(0.5-2.1)
B.3%
(-3.0-17.9)
5.5%
(1.0-10.3)
7.2%
(-0.4-14.0)
5.1%
(1.5-8.4)
3.7%
(1.7-5.7)
6.4%
134 - 9.3)
with dally
1 -hour maximum
O3
1.2%
(0.1 - 2.2)
10.7%
(-0.4-20.1)
NA
NA
* Health effects associated with short-term exposure lo PM Incidence was quantified across the range of PM concentrations observed in each study.
but not below background PMIevels. assumed lo be 6 ug/m3 for PM-10 and 2.5 ug/m3 for PM-2.5.
" Current standards are 50 ug/m3 annual average PM-10, 150 ug/m3 second daily maximum PM-10.
*" Based on 1-hour maximum SO2.
The numbers in parentheses for pooled functions are NOT standard confidence intervals. All numbers in parentheses are interpreted
as 90% credible intervals based on uncertainty analysis that takes into account both statistical uncertainty and possible geographic
variability. See text for details.
Abt Associates, Inc.
p. 100
July 3, 1996, Revised
-------
Exhibit 7.26
Sensitivity Analysis: The Effect of Differing Outpoints on Estimated
Mortality Associated with Long-term Exposure to PM-2.5
Philadelphia County, September 1992 - August 1993
(A) Mortality associated with
long-term exposure
BASE CASE
Lowest
4.7%
(2.9 - 6.4)
Outpoint = 12.5ug/m3
2.5%
(1.6-3.5)
Outpoint = 15ug/m3
1.0%
(0.6-1.3)
Outpoint = 18ug/m3
0.0%
(0.0-0.0)
(A) Pope et at., 1995
Health effects incidence was calculated down to the lowest level observed in the study (9 ug/m3).
No adjustments to the slope were performed.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All the numbers in parentheses are interpreted as 90% credible intervals based on uncertainty analysis
that takes into account both statistical uncertainty and possible geographic variability.
See text in Chapter 7 for details.
Abt Associates, Inc.
p. 101
July 3, 1996, Revised
-------
Exhibit 7.27
Sensitivity Analysis: The Effect of Concentration-Response Function Slope
on Estimated Mortality Associated with Long-term Exposure to PM-2.5
Philadelphia County, September 1992 - August 1993
(f\) Monamy associated witn
long-term exposure
BASt UASC
Assuming AQ as
4./7o
(2.9 - 6.4)
Assuming relevant AQ
J.OVo
(2.2 - 4.91
Assuming relevant AQ
£A7o
(1.5-3.3)
Health effects incidence was calculated down to the lowest level observed in the stud (A) Pope et al., 1995
* Adjusted function from Pope et al., 1995. Had historical air quality been 50% higher, the
relative risk calculated by the study would have been two thirds of that reported. Had historical
air quality been twice as high, the relative risk calculated would have been half that reported.
See text for details.
The numbers in parentheses for pooled functions are NOT standard confidence intervals.
All the numbers in parentheses are interpreted as 90% credible intervals based on uncertainty analysis
that takes into account both statistical uncertainty and possible geographic variability.
See text in Chapter 7 for details.
Abt Associates, Inc.
p, 102
July 3, 1996, Revised
-------
7.2. Uncertainty Analyses
A set of risk analyses considering different sample locations, different PM indicators,
different PM standards, different health endpoints, and different concentration-response
functions for each health endpoint can be expected to produce a set of results with a large
degree of variability. Some of this variability reflects actual variability in the underlying
population parameters. For example, "as is" PM concentrations in one location may be much
greater than in another location. Some of the variability in the outcomes of the risk analyses,
however, reflects uncertainty about the true values of the parameters of the risk analyses. The
substantial variability generated by applying different concentration-response functions to a
given sample location, for example, does not reflect real variability that exists in that location,
but instead reflects uncertainty about the actual concentration-response relationship between
PM and the population response hi that location.
There are several sources of possible uncertainty associated with the estimation of PM-
related incidence in the risk analysis (as discussed in Section 3). These include, for example,
uncertainty about the appropriate concentration-response function for a given location,
uncertainty about the baseline health effect incidence rates in the location, uncertainty about
average daily PM concentrations in the location, and uncertainty about the background PM
concentration in the location. In addition, there are sources of uncertainty inherent in any
empirical investigation, such as the uncertainty about the correct functional form of the
(concentration-response) model.
Some of these analysis input components are likely to involve much greater degrees of
uncertainty than others. The concentration-response function, for example, is considered to be
a source of substantial uncertainty. In contrast, because it was possible to obtain local baseline
incidence rates for many of the health endpoints, the degree of uncertainty associated with
these incidence rates is considered to be quite small. Similarly, although there may be some
uncertainty associated with average daily PM concentrations (especially related to those days
for which no PM monitoring was done), the overall level of uncertainty associated with PM
concentrations is judged to be small relative to that associated with the concentration-response
function. First, the percent of days without monitoring was very small in Philadelphia County
and fairly small in Southeast Los Angeles County. Second, if the uncertainty is associated
largely with random measurement errors, then small errors in one direction on one day are
likely to be counterbalanced by small errors in the other direction on another day. Because
daily incidences are totaled over the year, minor daily discrepancies related to PM
measurement errors are likely to largely cancel each other out in the total.
In addition to differences in the degrees of uncertainty associated with different
components of the analysis, there are differences in the degree to which these uncertainties can
be quantified. A sensitivity analysis (like those presented in Section 7.1) considers how the
end result of an analysis varies as a particular input parameter value is varied. Such an
analysis requires only the possible alternative parameter values but does not require that the
Abi Associates, Inc. p. 103 July 3, 1996, Revised
-------
probabilities of each possible input parameter value be known. In contrast, a quantitative
assessment of the uncertainty associated with an input parameter requires either the
distribution of possible parameter values, or, at a minimum, some information on which to
base an estimate of this distribution.
The source of the largest amount of uncertainty hi the risk analysis is likely to be the
concentration-response function. In addition, although the amount of information about the
distribution of possible values of the parameter in the concentration-response function varies
from one function to another (e.g., for short-term exposure mortality there is substantial
information, whereas for hospital admissions for respiratory illness there is much less), there
is some information for each category of concentration-response function. For these reasons,
the concentration-response function is the primary focus of the uncertainty analysis.
Uncertainty bounds characterizing the uncertainty associated with the concentration-response
function alone were derived for each combination of health endpoint and PM indicator. The
methods used to derive these uncertainty bounds are discussed in Section 5.2. The emphasis
in this discussion is on the general case when there is more than one study reporting a
concentration-response function for a particular combination of health endpoint and PM
indicator (e.g., short-term exposure mortality and PM-10). The results of these uncertainty
analyses are shown in Section 5.2.4.
Other sources of uncertainty were sequentially added to the uncertainty associated with
the concentration-response function, using Monte Carlo propagation of uncertainty methods,
which allow multiple sources of uncertainty to be considered simultaneously. The method and
results of this propagation of uncertainty are presented below.
7.2.1. A Monte Carlo analysis: propagation of uncertainties from several sources
The sensitivity analyses presented in Section 7.1 (and for meeting alternative standards,
in Section 8 below) illustrate the dependence of the results of the risk analyses on certain key
parameters and assumptions. The uncertainty analysis presented in Sections 5.3.1 and 5.3.2
attempts to quantify the uncertainty associated with a single parameter of the risk analysis,
namely, the concentration-response function. While sensitivity analyses may help to identify
those parameters that most influence the results of an analysis, such analyses do not indicate
the likelihood that the true value of the parameter is any of those values considered. In
addition, neither the sensitivity analyses nor the uncertainty analysis presented above consider
more than one source of uncertainty at a time. This section presents a set of analyses intended
to more fully characterize the uncertainty surrounding the risk estimates presented in Sections
7.1 and 8.1. The analyses use a Monte Carlo procedure to include uncertainty from several
sources simultaneously. These integrated uncertainty analyses are limited to consideration of
mortality associated with short-term and long-term exposure to PM-2.5.
Abt Associates, Inc. p. 104 July 3, 1996, Revised
-------
As noted above, a Monte Carlo procedure refers to the drawing of observations from a
known distribution. Uncertainty from several sources may be propagated through the risk
model by simultaneously drawing observations from a set of distributions, one for each source
of uncertainty considered. Suppose, for example, there are three unknown parameters in the
analysis model, each of which has been estimated. Each parameter estimate is the mean of a
distribution of possible values. On each iteration of the Monte Carlo procedure, an
observation is randomly selected from each of the three distributions. Each iteration therefore
produces a triple - three values, one for each unknown parameter. Given these three
particular parameter values, there is a particular value of the endpoint of the analysis (e.g., a
particular value of avoided mortality). On each of many iterations, the Monte Carlo procedure
produces a randomly selected set of parameter values (selected from the distributions for these
parameters) which in turn produces a particular value of the endpoint of the analysis. The
procedure therefore produces a distribution of values of the analysis endpoint (e.g., a
distribution of avoided mortality) corresponding to the set of distributions of parameter values.
This distribution characterizes the uncertainty surrounding the analysis endpoint resulting from
the uncertainty surrounding the input parameters considered.
Monte Carlo propagation of uncertainty analyses were carried out for both the short-
term exposure mortality and the long-term exposure mortality risk analyses. The sources of
uncertainty included in the Monte Carlo analysis are listed in Exhibit 7.28, along with
theircorresponding distributions. Uncertainties were incorporated into the analysis one at a
time in order to demonstrate the effect of each one.
The distributions of parameter values are not known but instead were estimated from
the available information The estimation of the distribution of p in the concentration-response
function, for example, is described in Section 5.2. The distributions for background PM-2.5
were based on estimates of background PM-2.5 concentrations ranging from 2 /xg/m3 to 5
jug/m3 in the Eastern U.S. and from 1 /ig/m3 to 4 /-ig/m3 in the Western U.S. With no further
information about the distribution of background concentrations in the Eastern or Western
U.S., uniform distributions on these intervals seemed reasonable.
Not all uncertainties associated with the model are incorporated into the propagated
uncertainty analysis. Some uncertainties, including, for example, the possible influence of
copollutants, were excluded from this analysis due to a lack of sufficient quantitative
information from which to estimate a distribution. Despite a lack of quantitative information,
cutpoints were included because of their key role in the analysis. Because there is no
information on which to base an estimate of the distribution of values of a possible cutpoint,
three alternative distributions, each consisting of weights for four discrete cutpoint values,
were used to illustrate the impact on the risk estimates of alternative viewpoints about the
likelihood of a threshold existing above background PM-2.5 levels. The weighting schemes,
which are included for illustrative purposes, are shown in Exhibit 7.29. Case I represents a
judgement that concentration-response functions are likely to be valid down to either
background or 10 /xg/m3; case III represents a judgement that concentration-response functions
Abt Associates, Inc. p. 105 July 3, 1996, Revised
-------
are likely to have outpoints at or above 18 /ig/m3; and case n is intermediate between the other
two.
Exhibit 7.28. Summary of Uncertainties Incorporated into Integrated Uncertainty
Analysis
Uncertainty
Coefficient (P) in
concentration-response
function
Cutpoints in concentration-
response function
Background PM-2.5
concentration
Form of PM reductions to
achieve alternative standards
Distribution
Short-term exposure function (for which there were
several epidemiological studies): 200 points representing
the estimated distribution of p for short-term exposure
mortality, derived in Section 5
Long-term exposure function: normal distribution based
on results of the single study (Pope et al. , 1995) used in
the risk analysis
4 cutpoints, 3 discrete weighting schemes,
2 slope adjustment methods
uniform distributions on the intervals [2,5] and [1,4]
(/ig/m3) for Philadelphia County and Los Angeles County,
respectively
(these are the ranges identified in the Criteria Document)
130 points representing distribution of a certain kind of
non-proportionality (described in Section 8)
Exhibit 7.29. Three weighting schemes for cutpoints
in integrated uncertainty analyses:
Mortality associated with short-term
exposure to PM-2.5
Background
10ug/m3
1 8 ng/m3
30 jag/m3
Case I
0.50
0.30
0.15
0.05
Case II
0.20
0.30
0.30
0.20
Case III
0.05
0.15
0.50
0.30
Ideally, to derive a concentration-response function with a cutpoint. the data on which the
original concentration-response function (without a cutpoint) was based would be re-analyzed,
excluding those points falling below the selected PM level. If a threshold existed in the original
Abt Associates, Inc.
p. 106
July 3, 1996, Revised
-------
data, this would presumably result in a different, steeper estimated function. Since the data on
which the concentration-response functions were based are not readily available, two methods of
adjusting the slope of concentration-response functions (described in Section 7.1) were
considered. In the integrated uncertainty analyses, they are given equal weight.
Exhibits 7.30 and 7.31 show the integrated uncertainty analysis for mortality due to
short-term exposure for "as is" conditions in Philadelphia County, and results assuming
attainment of current PM-10 standards in Southeast Los Angeles County, respectively. Each
vertical bar represents an estimate of the health effects that includes a certain set of
uncertainties (identified below the bars). The mean estimate is shown, as well as the 5th,
25th, 75th, and 95th percentiles. Note that in weighting case HI, the 25th percentile becomes
very close to the 5th percentile, and the 75th percentile becomes very close to the mean. This
is a result of the discrete nature of the distribution of outpoints (as well as of the specific
weights assigned to each cutpoint). The lowest thirty percent of values in the distribution of
estimates for Case in come from iterations when the cutpoint is assumed equal to 30 /ng/m3.
Iterations when the cutpoint is assumed to be lower give substantially higher values; therefore
the 25th percentile is very close to the 5th percentile (compared to the range of values). A
jump would be expected to occur around the 30th percentile. Similarly, since the next fifty
percent of values come from iterations when the cutpoint is assumed equal to 18 jig/m3, and
since these values are substantially lower than those predicted when the cutpoint is assumed to
be lower, the 75th percentile value is close to the mean (again, compared to the range of
values). If all the individual values obtained were plotted, the graph would show clusters of
points.
Because the lower cutpoints yield larger ranges of values (as well as larger values in
general; see Exhibits 7.15 and 7.16), Case I , which gives greater weight to the lower
cutpoints, does not show this kind of bunching of the distribution. Case II is intermediate
between Cases I and HI. Similarly, because of Los Angeles County's generally higher PM
concentrations, this phenomenon is not as pronounced in Exhibit 7.31.
Abt Associates, Inc. p. 107 July 3, 1996, Revised
-------
Exhibit 7.30
Uncertainty Analysis: Effect of Uncertainty of Relative Risks,
Background Concentrations, Outpoint and Slope Adjustment Method
Mortality Associated With Short-Term Exposure to PM-2.5
Philadelphia, September 1992 - August 1993
(Population: 1.6 Million)
3.0%
2.5%
2.0%
1.5%
1.0%
0.5%
0.0%
95th V. Me
75th %ile
(|) Mean
25th«/.ile
5th % ile
0
600
500
400
300
200
100
Uncertainty in
Just RR
(Background
= 3.5 mi/m')
Uncertainty in
RR and Background
(cutpoint = background)
Case I
I
Case II
Case III
I
Uncertainty in RR, Background,
Cutpoint, and Slope Adjustment Method
Risk
Associated
with PM-2.5
above
Cutpoint
(Number of
Deaths and
as Percent of
Total
Mortality)
Uncertainty in background concentration enters into these calculations only when
the cutpoint is set equal to background. The other cutpoints are greater than the
highest background concentration considered. When a cutpoint other than background
is chosen, each of the two slope adjustment methods has a fifty percent chance of selection.
Cutpoint Weighting Schemes
Background
lOpg/m3
IS^g/m3
30 ng/m3
Casel
0.5
0.3
0.15
0.05
Case II
0.2
0.3
0.3
0.2
Case III
0.05
0.15
0.5
0.3
Abt Associates, Inc.
p. 108
July 3, 1996, Revised
-------
Exhibit 7.31
Uncertainty Analysis: Effect of Uncertainty of Relative Risks,
Background Concentrations, Outpoint and Slope Adjustment Method
Mortality Associated With Short-Term Exposure to PM-2.5
After Meeting Current PM-10 Standards
Southeast Los Angeles County, 1995
(Population: 3.6 Million)
3.5%
3.0%
2.0%
1.5%
1.0%
0.5%
0.0%
1
-
4
(
i
i
95th % ile '
75th % ile
JMean (
254% ile
5th % ile *
1
I
)
(
1
'
1
(
t
Uncertainty in Uncertain
Just RR RRandBac
(Background (cutpoint = '
= 2.5 |ig/m')
1
)
(
'
1
)
I
ty in Case I Case II Case III
keround I 1
lackground) 1
Uncertainty in RR, Background,
IWU
900
Risk
700 Associated
with PM-2.5
eoo above
Cutpoint
(Number of
400 Deaths and
as Percent of
300 Total
Mortality)
200 J/
100
0
Cutpoint. and Slope Adjustment Method
Uncertainty in background concentration enters into these calculations only when
the cutpoint is set equal to background. The other cutpoints are greater than the
highest background concentration considered. When a cutpoint other than background is chosen,
each of the slope adjustment methods has a fifty percent chance of selection.
Cutpoint Weighting Schemes
Background
10ug/m3
18ug/m3
30 ug/m3
Case I
0.5
0.3
0.15
0.05
Case 11
0.2
0.3
0.3
0.2
Case III
0.05
0.15
0.5
0.3
Abt Associates, Inc.
p.109
July 3, 1996, Revised
-------
A similar uncertainty analysis was carried out for mortality associated with long-term
exposure. Exhibit 7.32 shows the cutpoint weighting schemes for this analysis. Like the
weighting schemes for short-term exposure cutpoints (see Exhibit 7.29), these are used for
illustrative purposes. Exhibit 7.33 shows the results of the uncertainty analysis for mortality
associated with long-term exposure in Philadelphia County.
Exhibit 7.32. Three weighting schemes for cutpoints
in integrated uncertainty analyses:
Mortality associated with long-term
exposure to PM-2.5
9 ug/m3
12.5 \ig/m3
15 ug/m3
18 ug/m3
Case I
0.55
0.20
0.15
0,10
Case II
0.35
0.35
0.20
0.10
Caselll
0.10
0.20
0.40
0.30
Abt Associates, Inc.
p.110
JH/V 3, 1996, Revised
-------
Exhibit 7.33
Uncertainty Analysis: Effect of Uncertainty of Relative Risk
and Cutpoint
Mortality Associated With Long-Term Exposure to PM-2.5
- Philadelphia, September 1992 - August 1993
(Population: 1.6 Million)
6.0%
5.0%
4.0%
3.0%
2.0%
1.0%
0.0%
<
' 95A%ile
I
75th %ile
) Mean
25th % ile
(
' 5th % ile
(
)
(
)
(
.
>
1000
Risk
BOO Associated
with PM-2.5
above
coo Cutpoint
(Number of
Deaths and
400 as Percent of
Total
Mortality)
200
0
Uncertainty in Case I Case II Case III
Just RR | |
(Cutpoint 1
= 9 ug/m') Uncertainty in RR and Cutpoint
The lowest observed level in the long-term exposure mortality study (Pope et al.. 1995) is
9 (ig/m3. Because this is above background PM-2.5 in Philadelphia, background does not
enter into these calculations.
Cutpoint Weighting Schemes
9Hg/m3
12.5 Ug/m3
15Hg/m3
1 8 Hg/nf
Case I
0.55
0.20
0.15
0.10
Case 11
0.35
0.35
0.20
0.10
Case 111
0.10
0.20
0.40
0.30
Abi Associates, Inc.
p.Ill
Jul\ 3, 1996, Revised
-------
8. Assessment of the Health Risk Reductions Associated with Attainment of Alternative
PM Standards
8.1. Results and sensitivity analyses
The results of the second phase of the risk analysis, assessing the health risk reductions
associated with attaining alternative PM-2.5 standards (as opposed to current standards), are
given in Exhibit 8.1 for Philadelphia County in 1992-1993 and Exhibit 8.2 for Southeast Los
Angeles County in 1995. Because Southeast Los Angeles County is not in attainment of
current PM-10 standards (and was not in attainment in 1995), PM-2.5 concentrations were
adjusted prior to the risk analyses, to simulate attainment of current standards. The method
for adjusting daily PM-2.5 concentrations to simulate attainment of alternative PM-2.5
standards and to simulate attainment of current PM-10 standards (in Southeast Los Angeles
County) is described in Section 2.2.
The results of the analyses in this second phase of the risk analysis follow a pattern
similar to that of the first phase. Predicted reductions in health effects incidence and predicted
reductions hi PM-related percent of total incidence associated with attaining alternative PM-2.5
standards in Southeast Los Angeles County are uniformly greater than those predicted in
Philadelphia County. The generally higher pollution levels and greater population size in
Southeast Los Angeles County are, as in the first phase of the risk analysis, the primary
reasons for this.
Because reduction of PM by the linear rollback method removes a given percent of
daily PM over background, the amount of PM removed each day depends on what background
PM concentration is. Analyses were conducted to illustrate the sensitivity of predicted avoided
mortality to changes in assumed background PM concentration. The results are shown in
Exhibit 8.3 for Philadelphia County and Exhibit 8.4 for Southeast Los Angeles County.
8.2 An assessment of the plausibility of linear rollbacks and associated
sensitivity analysis
As described in Section 2.2, the method of adjusting daily PM concentrations to
simulate attainment of alternative standards could significantly influence the results of these
analyses, especially when the alternative standard is a daily standard.
To assess the plausibility of the linear rollback method, analyses were carried out to
evaluate the extent to which historical changes in PM-2.5 air quality can be modeled using
linear rollbacks. The historic changes in PM-2.5 have not been the result of a PM-2.5 control
strategy, however. The PM-2.5 changes likely result from control programs for other
pollutants (especially PM-10, ozone, and sulfates) and from weather variability. The pattern
of changes that have occurred in the past, therefore, may not necessarily accurately estimate
the changes that may occur from future control programs.
Abt Associates, Inc. p. 112 July 3, 1996, Revised
-------
Exhibit 8.1
Estimated Changes in Health Risks Associated with Meeting Alternative PM-2.5 Standards
in Philadelphia County, September 1992 - August 1993 (for base case assumptions)
The Daily Standards Allow One Exceedance at Each Monitor; the Annual Standards Apply to the Annual Average at Each Monitor*
Mortality
Hospital Admissions
Respiratory
Hospital Admissions
Cardiar,
(F) Lower Respiratory Symptoms (
Health Effects*
Percent Reduction in Total Incidence:'"*
(B) Associated with long-term exposure (age 30 and over)
Percent Reduction in PM-Associated Incidence:
Percent Reduction in Total Incidence:
(C) Total
(alt
Percent Reduction in PM-Associated Incidence:
Percent Reduction In Total Incidence:
(D) Ischemic Heart Disease*""
(>64
64 years old)
Range of Percent Reductions in PM-Associated Incidence:
Range of Percent Reductions in Total Incidence:
B-12 yr. okls) '
Percent Reduction in PM-Associated Incidence:
Percent Reduction in Total Incidence:
PM-2.5-Associaled
Incidence
associated with
current standards**
(230 - 510 )
MO
(540-1170)
260
(70 -450)
70
(30 -120)
100
(50 -150)
< 11000>
(6000 -15000)
Reduction in tncl
20 ug/m3 annual
(0-0)
0.0%
0.0%
0
(0 -0)
0.0%
0
0
(0-0)
0.0%
0.0%
0
(0 -0)
0
(0-0)
0.0% - 0.0%
0.0% - 0.0%
<0>
(0 -0)
0.0%
0.0%
ctenc* Associated wtth MMI
20 ug/m3 annual
and 65 ug/m3 dairy
40
(20 -60)
10.8%
0.2%
170
(130-2SO)
20.0%
OB%
30
(10 -50)
11.5%
02%
to
(0-10)
10
(10 -20)
10.0%- 14.3%
0.1% -0.1%
< 1000>
(1000 -2000)
91%
i.a%
Ing AlWnMlm SUntfw*
20 UQ/mJ MnuH
MdMugMdMy
12U
(70 -170)
52.4%
0.6%
500
(920-720)
57.8%
2.7%
M
(20-190)
34.0%
0.7%
20
(10 -40)
90
(20 -90)
2V.e% SO.6%1
0.2% -0.4%
<4000>
(2000 -6000)
96.4%
7.9%
20 ug/m3 nnual
md2Sug/m3
(4000 -11000)
72.7%
14.6%
Healln effects are associated with short-term exposure 1o HM. unless otherwise specified.
" Health effects incidence was quantified across the range of PM concentrations observed in each study, when possible, but nol below background
PM-2.5 level. Background PM-2.5 is assumed to be 3.5 ug/m3 in Philadelphia County.
" The percent reduction in PM-associated incidence achieved by attaining alternative standards as opposed to the current standards is the reduction In
incidence divided by the incidence associated with current standards. For example, the percent reduction in PM-associated incidence of mortality
associated wilh short-term exposure to PM-2.5 achieved by meeting both a 20 ug/m3 annual and a 65 ug/m3 daily standard is 40/370-10.6%.
""* The percent reduction in total incidence achieved by attaining current or atternalive standards is the reduction in incidence achieved by attaining
the standard divided by the total (not only PM-associated) incidence
PM-2 5 results based on using PM-2 5 mass as PM-10 mass in (tie PM-10 functions
Angle brackets <> indicate incidence calculated using baseline incidence rates reported in studies, with no adjustment for location-specific
incidence rates This increases the uncertainty in the incidence estimates.
The numbers in parentheses for pooled functions are NOT standard confidence intervals. All the numbers in parentheses are interpreted 88 90% credible Intervals
based on uncertainty analysis that takes into account both statistical uncertainty and possible geographic variability. See text in Chapter 7 for details.
*The one exceedance form of the daily std. requires that the second highest concentration (the second daily max.) at each monitor (rounded to the nearest ug/m3) meets the »td.
The highest second daily maximum concentration at a monitor in Philadelphia is 72.6 ug/tn3.
The annual standard requires that the annual average at each monitor (rounded to the nearest 0.1 ug/m3) meets the std. The highest annual avg. at a monitor in Philadelphia Is 17.1 ug/m3.
Therefore Ihe 20 ug/m3 annual standard is already met in Philadelphia.
Soureai of ConctilU UofrRMporna (C-R) Functions:
(A) C-R function b«Md on pootod
tMub ftwn ttudin to itn bcrton*.
(B) Pop* MM., 1M5
(QTtujnton. !!.. 1OT4
(0) Sdmutt t Moms. 1M5
(E) SdiwMz « Moms. 1M5
(F)ScflwttU, otal.. 19»4
Abl Associates, Inc.
p.113
July 3, 1996, Revised
-------
Exhibits 1 (cent)
Estimated Changes in Health Risks Associated with Meeting Alternative PM-2.5 Standards
in Philadelphia County, September 1992 - August 1993 (for base case assumptions)
The Daily Standards Allow One Exceedance at Each Monitor; the Annual Standards Apply to the Annual Average at Each Monitor*
Health Effects'
Mortality
Hospital Admissions
Respiratory
Hospital Admissions
Cardiac
(A) Associated with short-term exposure (all ages)
Percent Reduction in PM-Associated Incidence:"*
Percent Reduction in Total Incidence:""
(B) Associated with long-term exposure (age 30 and over)
KM-^.3-ASSOClaleO
Incidence
associated with
current standards**
37U
(230 - 510 )
860
(540-1170)
Percent Reduction in PM-Associated Incidence:
Percent Reduction in Total Incidence:
(C) I otal Respiratory 260
(all ages) (70 - 450 )
Percent Reduction in PM-Associated Incidence:
Percent Reduction in Total Incidence:
(D) Ischemic Heart Disease
(>64 years old)
(E) Congestive Heart Failure
(>64 years old)
70
(30 - 120 )
100
(50 -150)
Range of Percent Reductions in PM-Associated Incidence:
Range of Percent Reductions in Total Incidence:
(F) Lower Respiratory Symptoms (8-12 yr olds)
<11000>
(6000 - 15000 )
Percent Reduction in PM-Associated Incidence:
Percent Reduction in Total Incidence:
Reduction in Incidence Associated with Meeting Alternative Standards
1 5 ug/m3 annual
60
(40 - 80 )
16.2%
0.3%
230
(170-380)
27.4%
13%
40
(10 -70)
15.4%
0.3%
10
(0 - 20 )
20
(10 20 )
14 3% - 20.0%
0.1% - 0.3%
<2000>
(1000 -3000)
18.2%
3.6%
1 5 ug/m3 annual
and 65 ug/m3 daily
BO
(40 -80)
16.2%
03%
230
(170-380)
27.4%
1.3%
40 .
(10 - 70 )
15.4%
0.3%
10
(0 -20)
20
(10-20)
14.3% -20.0%
0/1% - 0.3%
<2000>
(1000 -3000)
18.2%
3.6%
15 ug/m3 annual
and SO uo/ms daly
120
(70 - 170 )
32.4%
0.6%
500
(5320-720)
57.9%
2.7%
80
(20 -150)
34.6%
0.7% '
20
(10-40)
30
(20-50)
28.6% - 30.0%
0.2% - 0.4%
<4000>
(2000 -6000)
364%
7.3%
15 ug/m3 annual
and 25 uo/m3 daily
zoo
(160 - 360 )
70.3%
1.3%
860
(540-1170)
100.0%
4.7%
160
(50 -310)
69.2%
1.4%
50
(20 - 80 )
70
(30 -110)
70.0% -71 .4%
0.5% - 0.9%
<8000>
(4000 -11000)
727%
14.6%
Health effects are associated with shorl-term exposure to PM, unless otherwise specified.
' Health effects incidence was quantified across the range of PM concentrations observed in each study, when possible, but not below background
PM-2 5 level. Background PM-2.5 is assumed to be 3.5 ug/m3 in Philadelphia Count
The percent reduction in PM-associated incidence achieved by attaining alternative standards as opposed to the current standards is the reduction in
incidence divided by the incidence associated with current standards For example, the percent reduction in PM-associated incidence of mortality
associated with short-term exposure to PM-2 5 achieved by meeting both a 15 ug/m3 annual and a 65 ug/m3 daily standard is 607370=16.2%.
* The percent reduction in total incidence achieved by attaining current or alternative standards is the reduction in incidence achieved by attaining
Ihe standard divided by the total (not only PM-associated) incidence
" PM-2 5 results based on using PM-2.5 mass as PM-10 mass in the PM-10 functions.
'"Angle brackets <> indicate incidence calculated using baseline incidence rates reported in studies, with no adjustment for location-specific
incidence rates This increases the uncertainty in the incidence estimates
Sources of ConcantiaUon-Resptrse (C-R) Functions:
(A) C-R function based on pooled
results from studies in six locations
(B)Popaatal.. 199S
(C)Thurston, at*., 1994
(0) Schwartz & Morris. 1995
(E) Schwartz & Morris, 1995
(F) Schwartz, et el., 1994
The numbefs in parentheses for pooled functions are NOT standard confidence intervals. All the numbers in parentheses are interpreted as 90% credible intervals
based on uncertainty analysis that takes into account both statistical uncertainty and possible geographic variability. See text in Chapter 7 tor details.
#The one exceedance form of the daily std. requires that the second highest concentration (the second daily max.) at each monitor (rounded to the nearest ug/m3) meets the std.
The highest second daily maximum concentration at a monitor in Philadelphia is 72.6 ug/m3.
The annual standard requires that the annual average at each monitor (rounded to the nearest 0.1 ug/m3) meets the std. The highest annual avg. at a monitor In Philadelphia is 17.1 ug/m3.
Therefore Ihe 20 ug/m3 annual standard is already met in Philadelphia.
Abt Associates, Inc.
p. 114
July 3, 1996, Revised
-------
Exhibit 82
Estimated Changes in Health Risks Associated with Meeting Alternative PM-2.5 Standards
in Southeast Los Angeles County, 1995* (for base case assumptions)
The Dally Standards Allow One Exceedance at Each Monitor; the Annual Standards Apply to the Annual Average at Each Monitor*
Health Effects
Mortality
Hospital Admissions
Respiratory
Cardiac
(A) Associated with short-term exposure (all ages)
rM-z.b-Kelated
associated with
current standards"
fiu
(430 - 970 )
Percent Reduction in PM-Associated Incidence:*"
Percent Reduction in Total Incidence:"*'
(B) Associated with long-term exposure (age 30 and over)
Percent Reduction in PM-Associated Incidence:
Percent Reduction in Total Incidence:
(C) Total Respiratory
(all ages)
Percent Reduction in PM-Associated Incidence:
Percent Reduction in Total Incidence:
([)) Iscnemic Heart Disease
(>64 years old)
(E) Congestive Heart Failure ""*
(»64 years old)
Range of Percent Reductions in PM-Associated Incidence:
Range of Percent Reductions in Total Incidence:
(F) Lower Respiratory Symptoms (8-12 yr olds)
2050
(1290-2770)
9*H)
(250 -1630)
130
(50 - 200 )
140
(70 -210)
< 43000 >
(23000 - 58000 )
Percent Reduction in PM-Associated Incidence:
Percent Reduction in Total Incidence:
Reduction In Incidence Associated wttll Maeting ArtemaHv* Standards
20 ug/m3 annual
(80 - 190 )
197%
0.6%
550
(340 - 760 )
27.0%
2.3%
160
(50 -310)
19.1%
1.2%
20
(10-40)
30
(10 -40)
15.4% -2 1.4%
02% - 0.4%
< 10000>
(5000 -15000)
23.3%
67%
20 uo/m3 annual
and 65 ug/m3 dally
2/U
(160 - 370 )
38.0%
1.1%
1130
(710-1550)
55.5%
4.8%
350
(90 -600)
37.2%
2.3%
(20-80)
50
(30-80)
35.7% - 38.5%
0.4% - 0.7%
< 19000 >
(9000 -27000)
44.2%
12.7%
20 uo/m3 annual
and 50 uo/m3 daily
3/0
(220 -510)
52.1%
1.5%
1580
(990-2150)
77.3%
6.6%
490
(130 -850)
52.1%
3.2%
/o
(30-110)
70
(40 -110)
50.0% - 53.8%
OS* -1.0%
< 25000 >
(13000 -36000)
58.1%
16.7%
20 uo/m3 annual
and 25 u°/m3 daily
Mu
(330 - 750 )
77.5%
2.2%
2050
(1290 - 2770 )
100.0%
8.6%
730
(200 -1260)
777%
4.7%
luu
(40 -160)
110
(SO -170)
76.9% - 78.6%
0.9% -16%
< 35000 >
(18000 -49000)
61.4%
23.3%
n exposure lo PM, unless otherwise specified
* Los Angeles County was not in attainment of current PM-10 standards in 1995 Figures shown assume actual PM-10 concentrations
are first rolled back to simulate attainment of these standards, and that actual PM-2.5 concentrations are rolled back by the same
percent as PM-10. See text in Chapter VI for details.
" Health effects incidence was quantified across Ihe range of PM concentrations observed in each study, when possible, but not below background
PM-2 5 level. Background PM-2 5 is assumed to be 25 ug/rn3 in Southeast Los Angeles County.
*" The percent reduction in PM-associated incidence achieved by attaining alternative standards as opposed to the current standards is the reduction in
incidence divided by the incidence associated with current standards For example, the percent reduction in PM-associated incidence
of mortality associated with short-term exposure to PM-2 5 achieved by meeting both a 20 uo/m3 annual and a 65 ug/m3
daily standard is 270/710 = 38 0%.
**" The percent reduction in total incidence achieved by attaining current or alternative standards is the reduction in incidence achieved by attaining
the standard divided by the total (not only PM-associated) incidence
***** PM-2 5 results based on using PM-2.5 mass as PM-10 mass in Ihe PM-10 functions
Angle brackets <> indicate incidence calculated using baseline incidence rates reported in studies, with no adjustment for location-specific
incidence rales This increases the uncertainty in the incidence estimates.
The numbers in parentheses for pooled studies are NOT standard confidence intervals. All the numbers in parentheses are interpreted as 90% credible intervals
based on uncertainty analysis that takes into account both statistical uncertainty and possible geographic variability. See text in Chapter 7 for details.
#The one exceedance form of the daily std. requires that the second highest concentration (the second daily max.) at each monitor (rounded to the nearest ug/m3) meets the std.
The highest second daily maximum concentration at a monitor in L.A. is 101.7 ug/m3 (after adjustment to simulate attainment of current stds).
The annual standard requires that the annual average at each monitor (rounded to the nearest 0.1 ug/m3) meets the std.
The highest annual avg. at a monitor in L.A. is 24.1 ug/m3 (after adjustment to simulate attainment of current stds).
Sources of Concentration-Response (C-R) Functions:
(A) C-R function based on pooled results from
studies In 6 locations
(B) Pope at al.. 1995
(C)Thuraton,«tal.. 1994
(D) Schwartz t Morris. 1995
(E) Schwartz & Morril. 1995
(F) Schwartz. Mai.. 1994
Abt Associates, Inc.
p.115
July 3, 1996, Revised
-------
Exhibit 8.2 (cont.)
Estimated Changes in Health Risks Associated with Meeting Alternative PM-2.5 Standards
in Southeast Los Angeles County, 1995* (for base case assumptions)
The Dally Standards Allow One Exceedance at Each Monitor; the Annual Standards Apply to the Annuil Average at Each Monitor*
Respiratory
Hospital Admissions
Cardiac
Health Effects
(A) Ai>uL(dlcJ wilh al mi l-ltii i 1 **p-jiijit (dll dyca)
PM-2 5-Related
associated with
current standards"
(430 -970)
Percent Reduction in PM-Associated Incidence:*"
Percent Reduction in Total Incidence:****
(B) Associated with long-term exposure (age 30 and over)
2050
< 1290-2770)
Percent Reduction in PM-Associated Incidence:
Percent Reduction in Total Incidence:
(C) Total Respiratory
(all ages)
Percent Reduction in PM-Associated Incidence:
Percent Reduction In Total Incidence:
(D) Ischemic Heari Disease
(>64 years old)
(E) Congestive Heart Failure
(>B4 years old)
940
(250 - 1630 )
130
(50 - 200 )
140
(70 -210)
Range of Percent Reductions in PM-Associated Incidence:
Range of Percent Reductions in Total Incidence:
|F) Lower Respiratory Symptoms {8-12 yr olds)
<43WQ>
(23000 - 56000 )
Percent Reduction in PM-Associated Incidence:
Percent Reduction in Total Incidence:
Reduction In Incidence Associated wtth MMting Alfttrmthn Standards
15ug/m3 annual
3UQ
(180 -410)
42.3%
1.2%
1260
(790-1720)
61.6%
5.3%
400
(110 -680)
426%
2.6%
50
(20 - 90 )
60
(30-90)
38.5% - 42.9%
0.4% - 0.6%
< 21000 *
(10000 -30000)
486%
14.0%
15 ug/m3 annual
nd65ug/m30 -410)
42.3%
1.2%
1220
(790-1720)
61.6%
5.3%
400
(110 -AN)
42.8%
2.8%
50
(20-90)
60
(30-90)
38.5% - 42.9%
0.4% -0.8%
< 21000 >
(10000 -30000)
48.8%
14.0%
15 ugvms annual
and 50 ug
(13000 -36000)
56.1%
16.7%
1 5 ug/m3 annual
and 25 ugYm3 daily
(330 750 )
77.5%
2.2%
2050
(1290-2770)
100.0%
6.6%
730
(200 -1260)
77.7%
4.7%
100
(40 -160)
110
(50 -170)
76.9% - 78.6%
0.9% -1.6%
< 35000 >
(18000 -49000)
81.4%
23.3%
affn effects are associated with snort-lerm exposure to PM. unless otherwise specified
' Los Angeles County was not in attainment of current PM-10 standards in 1995 Figures shown assume actual PM-10 concentrations
are first rolled back to simulate attainment of these standards, and that actual PM-2 5 concentrations are rolled back by the same
percent as PM-10. See text in Chapter VI for details
** Health effects incidence was quantified across the range of PM concentrations observed in each study, when possible, but not below background
PM-2 5 level Background PM-2.5 is assumed lo be 2.5 ug/m3 in Southeast Los Angeles County.
*** The percent reduction in PM-associated incidence achieved by attaining alternative standards as opposed to the current standards is the reduction in
incidence divided by the incidence associated with current standards For example, the percent reduction in PM-associated incidence
of mortality associated with short-term exposure to PM-2.5 achieved by meeting both a 15 ug/m3 annual and a 65 ug/m3
daily standard is 300/710 = 42 3%
"" The percent reduction in total incidence achieved by attaining current or alternative standards is the reduction in incidence achieved by attaining
the standard divided by the total (not only PM-associated) incidence.
***** PM-2 5 resulls based on using PM-2 5 mass as PM-10 mass in the PM-10 functions
Angle brackets <> indicate incidence calculated using baseline incidence rates reported in studies, with no adjustment for location-specific
incidence rates. This increases Ihe uncertainty in the incidence estimates.
The numbers in parentheses for pooled studies are NOT standard confidence intervals. All the numbers in parentheses are interpreted as 90% credible intervals
based on uncertainty analysis that takes into account both statistical uncertainty and possible geographic variability. See text in Chapter 7 for details.
#The one exceedance form of the daily std. requires that the second highest concentration (the second daily max.) at each monitor (rounded to the nearest ug/m3) meets the Std.
The highest second daily maximum concentration at a monitor in LA. is 101.7 ug/m3 (after adjustment to simulate attainment of current stds).
The annual standard requires that the annual average at each monitor (rounded to the nearest 0,1 ug/m3) meets the std.
The highest annual avg. at a monitor in LA: is 24.1 ug/m3 (after adjustment to simulate attainment of current stds).
Sources of ConcentratiorvRasponse (C-R) Functions:
(A) C-R function based on pooled results (TORI
studies in 6 locations
(B) Pope el at., 1995
(C)Thurston.etal., 1W4
(0) Schwartz & Morris. 1995
(E) Schwartz & Mom's, 1995
(F) Schwartz. etaJ., 1994
Abt Associates, Inc.
p. 116
July 3, 1996, Revised
-------
Exhibit 8.3
Sensitivity Analysis: The Effect of Alternative Background Levels on
Estimated Changes in Health Risks Associated with Meeting a PM-2.5 Standard of 15 ug/m3 Annual, 50 ug/m3 Daily
in Philadelphia County, September 1992 - August 1993 (for base case assumptions)
Mortality
Health Effects*
(A) Associated with Short-term exposure (all ages)
Percent Reduction In PM-Associated Incidence:"*
Percent Reduction In Total Incidence:"**
(B) Associated with long-term exposure (age 30 and over)
PM-2.5-Associaled
Incidence associated with
current standards"
Background = 3.5 ug/m3
370
(230 - 510 )
860
(540-1170)
Percent Reduction in PM-Assoclated Incidence:
Percent Reduction In Total Incidence:
Reduction In Incidence Associated with Meeting Standard
Background
= 3.5 ug/m3
120
(70 -170)
32.4%
0.6%
550
(500-720)
57.9%
2.7%
Background
« 2.0 uo/m3
130
(60 -190)
35.1%
0.6%
540
(360-780)
63.2%
3.0%
Background
* 5.0 ug/m3
110
70-150)
.'!',]$;';::W.T%
:?'.::': 0.5% .
450
(300450)
52.8*;; :'. ': ";:
2J* ;!''' ;'.
* Health effects are associated with short-term exposure to PM, unless otherwise specified
** Health effects incidence was quantified across the range of PM concentrations observed in each study, when possible, but not below background
PM-2.5 level. Background PM-2.5 is assumed to be 3.5 ug/m3 in Philadelphia County.
*** The percent reduction in PM-associated incidence achieved by attaining alternative standards as opposed to the current standards Is the reduction in
incidence divided by the incidence associated with current standards. For example, the percent reduction in PM-associated incidence of mortality
associated with short-term exposure to PM-2.5 achieved by meeting both a 15 ug/m3 annual and a 50 ug/m3 daily standard, assuming that background
PM-2.5 concentration is 3.5 ug/m3 is 120/370 = 32.4%.
"** The percent reduction in total incidence achieved by attaining current or alternative standards is the reduction in incidence achieved by attaining
the standard divided by the total (not only PM-associated) incidence.
The numbers in parentheses for pooled functions (mortality associated with short-term exposure) are NOT standard
confidence intervals. All the numbers in parentheses are interpreted as 90% confidence intervals based on uncertainty
analysis that takes into account both statistical uncertainty and possible geographic variability. See text in Chapter 7 for details.
Sources of Concentration-
Response (CR) functions:
(A) C-R function based on
pooled results from studies
in 6 locations.
(B) Pope at at.. 1995
Ah/ Associates, Inc.
p.117
My 3, 1996, Revised
-------
Exhibit 8.4
Sensitivity Analysis: The Effect of Alternative Background Levels on
Estimated Changes in Health Risks Associated with Meeting a PM-2.5 Standard of 15 ug/m3 Annual, 50 ug/m3 Daily
Southeast Los Angeles County, 1995
Health Effects-
Mortality
(A) Associated with short-term exposure (all ages)
Percent Reduction In PM-Associated Incidence:***
Percent Reduction In Total Incidence:"*'
{B) Associated with long-term exposure (age 30 and over)
PM-2 5-Associated
Incidence associated wit
current standards**
Background = 2.5 ug/m3
710
(430 -970)
2050
(1290-2770)
Percent Reduction in PM-Associated Incidence:
Percent Reduction In Total Incidence:
Reduction In Incidence Associated with Meeting Standard
Background
= 25 ug/m3
370
(220 -510)
52.1%
1.5%
1580
(990-2150)
77.3%
6.6%
Background
= 1 Oug/m3
390
(240 - 540 )
54.9%
1.6%
1670
(1050-2270)
81.5%
7.0%
Background
«4.0ugftn3
350
(210 -480)
«*.3* . :.;
' US**. <: '.
1480
(930-2030)
72.5%
6.2%
* Health effects are associated with short-term exposure to PM. unless otherwise specified.
" Health effects incidence was quantified across the range of PM concentrations observed in each study, when possible, but not below background
PM-2 5 level Background PM-2.5 is assumed to be 2.5 ug/m3 in Los Angeles County.
*** The percent reduction in PM-associated incidence achieved by attaining alternative standards as opposed to the current standards is the reduction in
incidence divided by the incidence associated with current standards. For example, the percent reduction in PM-associated incidence of mortality
associated with short-term exposure to PM-2.5 achieved by meeting both a 15 ug/m3 annual and a 50 ug/m3 daily standard, assuming that background
PM-2.5 concentration is 2.5 ug/m3 is 370/710 = 52 1%
"** The percent reduction in total incidence achieved by attaining current or alternative standards is the reduction in incidence achieved by attaining
the standard divided by the total (not only PM-associated) incidence.
The numbers in parentheses for pooled functions (short-term exposure) are NOT standard confidence intervals.
All the numbers in parentheses are interpreted as 90% credible intervals based on
uncertainty analysis that takes into account both statistical uncertainty and possible geographic variability. See text in Chapter 7 for details.
Sources of Concentration-
Response (CR) functions:
(A) C-R function based on
pooled results from studies
in 6 locations.
(B) Pope etal., 1995
Abt Associates, Inc.
p. 118
July 3, 1996, Revised
-------
The historic PM-2.5 data come from monitors operated by the California Air Resources
Board (CARD) and from the EPA's National Aerometric Monitoring System (NAMS) and the
Harvard monitoring system. Some monitors reported concentrations on only a few days in a
given year; such data were excluded. Only monitor-years represented by fifty or more days
were included in the analysis. This corresponds to standard one-in-six monitoring, with
allowances for a limited number of missed days. Exhibit 8.5 shows the number of days
available at each monitor for each year (blank entries indicate either no monitoring or
insufficient monitoring). In all, 230 monitor-years from 57 sites are represented.
Air quality data from different years at a single monitor could not be compared
directly, because different years were represented by different numbers of days. Therefore,
the reported concentrations from each monitor-year were grouped into deciles16, and all of the
observations in each decile were averaged to produce a representative concentration for that
decile. The averaging (as opposed to selecting specific percentile values) was meant to
promote stability of results between two monitor-years actually reporting the same air quality
(with sampling error).17 The second-highest reported concentration was also retained for each
year. Each monitor-year was therefore represented by ten decile averages plus the second
highest reported concentration.
To determine the extent to which the historic air quality changes were linear, the
deciles from one monitor-year were regressed against the corresponding deciles from another
monitor-year. In the primary analyses consecutive years from single monitors were compared;
these results were later compared to those comparing the years from a given site with highest
and lowest average reported concentrations.
A regression gives a linear equation of the form
y = Ox + P .
A linear rollback over background, however, is represented by an equation of the form
y = A(x - B) + B ,
where x is the earlier year's PM decile, y is the later year's PM decile, and B is background
concentration, which is subtracted from all PM concentrations, and therefore from each PM
decile, x. If the equations are to agree that is, if the relationship between one year's PM
16 Each decile was one tenth of the observations, rather than spanning one tenth of the concentration
range.
"Analysis of data from several years in Pittsburgh indicated that most of the distribution was in any case
fairly uniformly distributed, with significant variations only at the highest concentrations reported.
Abt Associates, Inc. p.119 July 3, 1996, Revised
-------
Exhibit 8.5
PM-2 5 Monitors with at Least Two Years In Which >50 Days Each Year Have Reported PM Concentrations: Number of Days with Reported Concentrations
TIARB Monitors
AIRS Monitors
State
6
22
23
25
29
34
36
37
41
42
45
48
49
County
19
37
65
107
71
3
9
11
15
25
189
13
39
61
67
67
5
17
29
51
3
101
79
29
141
167
201
355
439
35
Monitor
1000246
1300694
1400699
1500203
1500250
1600715
1700720
1700728
2000002
2600785
2900794
3300144
3400305
3600188
3900252
4300382
5000567
5400568
7000060
7000072
Monitor
5
4002
8001
2002
10
1005
3
1
4002
2
2003
5001
7002
11
4
56
69
77
1016
9
4
1
1001
3001
15
80
21
27
4
7
36
37
1002
24
12
60
3
Fresno. CA
El Centra, CA
Lone Pine, CA
Bakersfield. CA
Taft College. CA
Corcoran. CA
Anderson Springs, CA
Glenbrook. CA
Madera. CA
Mammoth Lakes. CA
Truckee. CA
Rubidoux. CA
Sacramento, CA
Trona. CA
Stockton-Hazetton. CA
San Jose, CA
Modesto, CA
Visalia, CA
Azusa, CA
North Long Beach. CA
Fresno. CA
Long Beach. CA
Rubidoux. CA
Visalia. CA
New Orleans. LA
Presque Isle, ME
Acadia Nafl Park, ME
Augusta. ME
Ware. MA
Boston. MA
Clayton, MO (SL Co. I
Ferguson. MO (SL Co )
St. Ann. MO(SLCo)
Newark. NJ
Elizabeth. NJ
New York City. NY
New York City. NY
New York City. NY
Syracuse, NY
Wmslon-Salem. NC
Canby, OR
Bend. OR
Central Point. OR
Medford. OR
Portland, OR
Portland, OR
Pittsburgh, PA
Pittsburgh, PA
Philadelphia. PA
Columbia. SC
San Antonio. TX
El Paso, TX
Texas City. TX
Harris Co.. TX
Corpus Christi, TX
Fort Worth, TX
Salt Lake Co., LIT
62 83 84 85
51 85
74
87 66
104
59 91 61
56 54
59 58
60 58
65 91 71
60 82 65
86
110
128
86
111
112
112
B6
141
95
188
90
169
54
57
61
57
58
56
170
164
164
93
125
155
132
54
87 88 89
se
56
58
61
62
59
58
52 54
55
53
56
57
60
60 57
60 60
57 54
61 57
111 61 61
111 92
90
68
81
173 63 58
59 60 58
61 59 53
53
58 52
59 55
60 52
55
171 168 167
59 61
179 115
134 81
146 71
160 88
56
90
59
58
58
57
52
59
54
52
62
56
54
58
59
57
56
57
52
53
159
59
91
60
55
S3
X
56
51
55
58
58
53
53
57
57
59
55
164
62
92
SB
se
52
X
5»
59
5fl
66
59
01
58
56
60
S9
60
61
56
57
93
60
61
52
55
57
61
60
60
73
56
60
61
60
61
58
61
59
56
59
52
56
84
J M
61
57
60
62
60
60
60
61
62
60
sa
60
51
55
Abi Associates, Inc.
p. 120
My 3, 1996, Revised
-------
deciles and another year's PM deciles is to be linear then it is necessary that A equal a, and
P
B =
1 - A
Since changes hi ah- quality from year to year tend to be small, however (since A = 1),
the denominator hi the equation above is often close to zero. Therefore, small changes in the
estimate of A lead to large changes hi the estimated background, and regressions for different
years estimate backgrounds differing by an order of magnitude (and in some cases substantially
higher than maximum PM-2.5 concentrations). In order to avoid this, the intercept was
constrained to be zero hi all regressions. When background concentrations were considered,
they were actually subtracted from all concentration measurements, and regressions again
performed with the intercept constrained to be zero.
All estimated slope coefficients were significant at the 0.01 confidence level, and the
regressions explained the vast majority of the variation in almost all year pairs. Exhibit 8.6
describes the distributions of r2 values obtained in four sets of regressions. The first set
includes all consecutive years at single monitors. The second set considers a single pah: of
years from each monitor, the year with the lowest average concentration and the year with the
highest average concentration. These two sets assume no background concentration. The final
two sets consider consecutive year pah's in the eastern and western United States separately,
incorporating an estimate of background concentration in each case. Note that including
reasonable estimates of background concentrations improves the predictive power of the worst
of the regressions (a minimum r2 with estimated backgrounds of 0.71, as opposed to 0.61).
Exhibit 8.6. Distributions of r2 Statistics for Regressions on Different Sets of Year-Pairs
All, Consecutive Years
(no background)
All, High Year vs. Low Year
(no background)
East, Consecutive Years
(background = 3.5 ^g/ni3)
West, Consecutive Years
(background = 2.0 fig/m3)
N
130
57
45
85
Mean
(St. Dev.)
0.95
(0.06)
0.94
(0.07)
0.96
(0.04)
0.95
(0.05)
Min
0.61
0.63
0.77
0.71
5th
%-ile
0.86
0.79
0.88
0.86
95th
%-ile
0.99
0.99
0.99
0.99
The statistics in Exhibit 8.6 show that a linear rollback above some background can
account for the vast majority of the change in PM concentrations between two years.
Abt Associates, Inc.
p. 121
Julv 3, 1996, Revised
-------
However, predicting the proper slope (percent change) is more difficult. The percent change
in the mean, as might be expected, is generally close to the percent change predicted by the
regression. Therefore, rollbacks to meet annual standards present little problem. However,
the percent change necessary to bring the second highest value down to meet a standard is not
particularly well correlated (correlation coefficient ~ 0.66) with the percent change indicated
by the regression. Exhibit 8.7 gives some statistics on the distribution of the ratio of the
regression slope to the ratio of second highest values. When the two percent changes are in
agreement, the ratio is one. When the second high changes more than the distribution as a
whole, the ratio is less than one, and when the second high changes less than the distribution
, as a whole, the ratio is greater than one.
Exhibit 8.7. Statistics on the Distribution of
(Regression Slope)/(Ratio of Second High Values)
All Consecutive Years (129 pairs)
Mean
Standard Deviation
1st percentile
5th percentile
25th percentile
50th percentile
75th percentile
95th percentile
99th percentile
1.02
0.18
0.66
0.72
0.90
1.03
1.12
1.27
1.62
These statistics, along with the correlation coefficient of 0.66, show that the regression
slopes are not well predicted by the ratio of second high values. (Linear regression accounts
for 43% of the variation; exponential and logarithmic forms do worse.)
A sensitivity analysis was carried out for Philadelphia County to examine the effect of
different rollback methods on PM-related health effects. The results are given in Exhibits 8.8
and 8.9. Rollbacks designed to meet annual standards, which remove the same total amount of
PM from the air no matter how the reductions are distributed, result in similar changes in
incidence. They result in exactly the same reductions when the reductions are calculated using
functions relying on only the annual mean. The small differences produced by the different
methods when reductions are measured using functions relying on daily PM concentrations are
due to the slight nonlinearity of the functions. (Linear functions would produce identical
results under the three methods.)
Abt Associates, Inc.
p. 122
July 3, 1996, Revised
-------
Rollbacks designed to simulate the attainment of daily standards, however, produce
notably different results, since they result in the elimination of widely different amounts of PM
from the air. The percent rollback necessary to bring the second highest value down to a
given value remains the same no matter what the rollback method; however, the amount of
PM removed from the lower 90% of the distribution changes. When the higher concentrations
are reduced more than the lower, the result is that the lower concentrations are reduced less
than under a strictly proportional rollback, so less PM is removed from the air and health
effects are reduced less. Conversely, when higher concentrations are reduced less than lower
ones, the result is that the lower concentrations are reduced more than under a strictly
proportional rollback, so more PM is removed from the air and health effects are reduced
more.
The degree to which these deviations from proportional rollbacks might be expected to
be observed in areas attempting to come into attainment with new standards is unclear,
especially the case in which higher concentrations are reduced less than those in the bulk of the
distribution. For any sets of standards in which daily standards were controlling in an area, a
set of control strategies that reduced high concentrations less than the overall distribution
would actually make achieving attainment more difficult, although this case might approximate
an instance in which reducing high peaks was particularly difficult for some reason.
Ab! Associates, Inc. p. 123 July 3, 1996. Revised
-------
Exhibit 8.8
Sensitivity Analysis: Effect of Alternative Rollback Methods on Mortality Estimates
Short-term Exposure (Pooled Function) and Long-term Exposure PM-2.5 Mortality Functions
Philadelphia County, September 1992 -August 1993
Initial Air Quality: 16.S ug/m3 annual average, 69.3 ug/m3 2nd daily maximum
(A) Mortality essotiated with
hart-term exposure
(B) Mortality associated with
long-torm exposure
Alternative
15 ug/m3 annual
50 ug/mS dairy
15 ug/m3 annual
50 uoym3 daily
Percent Change In PH-Aaaedalad Incidence"
All PM conceiiliauons
11.4%
2».T%
20:8%
53.8%
Higher PM
11.4%
21.5%
208%
39.1%
Higher PM
11.4%
390%
20.9%
70.5%
Portion of Proportional Rollback
Higher PM
1004%
72.6%
100.2%
72.6%
Higher PM
100.4%
131.3%
100.5%
131.1%
Health effects incidence was quantified across the range
of PM concentrations observed in each study, but not below
background PM-2.5 level, which is assumed to be 3.5 ug/m3.
(A) C-R function based on studies in 6 cities
(B) Pope etal., 1995
In the alternative rollback cases, the upper 10% of the PM distribution was reduced by more or less than the lower 90%. See text in Section 8.2 for details
Abt Associates, Inc.
p. 124
Jul\ 3, 1996, Revised
-------
Exhibit 8.9
Sensitivity Analysis: Effect of Alternative Rollback Methods on Mortality Estimates
Short-term Exposure (Pooled Function) and Long-term Exposure PM-2.5 Mortality Functions
Philadelphia County, September 1992 - August 1993
Details of Rollbacks for Proportional and One Alternative Rollback
InWal Atr Quality: 16.5 ugrml annual average, 61.3 ugfm3 2nd dally maximum
(A) Mortaely associated Mitti
then mm exposure
(B)Mortaety associated wW>
tonfl>arm exposure
Alterative
15ug/m3annua
50 uaym3 daily
15uo/m3annua
SOucymSdaiy
Entire AC
AORoeback
11.3%
29.4%
11.3%
28.4%
distribution reduced equally
RwuHngAQ
15.0
81.9
12.7
600
15.0
61.9
1Z7
50.0
Pel UK it change
11.4%
29.7%
20.8%
53.8*
Uooer 10% of AQ distribution reduced more
AQ Rollback Require
15.6%
9.8%
29.4%
18.4%
15.6%
9.8%
29.4%
18.4%
Resulting AQ
15.0
59.2
13.7
50.0
15.0
59.2
13.7
50.0
Percent Chang* i
11.4%
21.5%
20.8%
38.1%
Portion of Proportional
100.4%
72.6%
100.2%
72.6%
HeaHh effects incidence was quantified acrou me range
<* PM concentratiora oUerved In eech study, but net batow
background PM-2.5 level, which a assumed to be 3.5 uB/m3.
** The percent of PM-auociated incidence achieved by the alternative rollback
method (i.e., the upper 10% o( the air quaWy diitribution being reduced by
more than the lower 90% of the distribution). For example, mine tecond
row 62.6% - 18.6V29 7% x 100.
(A) C-R function based on studies in 6 cities
(B) Pope et al, 1996
Abi Associates, Inc.
p. 125
July 3, 1996, Revised
-------
8.3. Uncertainty analyses
Section 7.2 describes the integrated uncertainty analysis methods. Similar analyses
were conducted to assess the uncertainty surrounding estimates of avoided health effects
associated with attaining alternative PM-2.5 standards. Results are presented in terms of the
number of cases avoided, rather than the number of cases remaining
Exhibit 8.10 shows the estimated health benefits associated with attaining an annual
standard of 15 /ig/m3 and a daily standard of 50 /ig/m3 in Southeast Los Angeles County,
progressively including more sources of uncertainty from left to right in the diagram. The
first line shows the estimate when only uncertainty in relative risk is included. The next line
shows the estimate when uncertainty in relative risk and background (but not cutpoint) is
included, and the next three lines show the estimates using the three cutpoint weighting cases.
The four lines on the right repeat the last four lines on the left, adding uncertainty in the form
of the rollback. The diagram shows that adding this uncertainty does not significantly change
the uncertainty in the estimates produced by the model.
Exhibit 8.11 compares the benefits associated with meeting an annual standard of 15
/ig/m3 and either no daily standard or a daily standard of 65, 50, or 25 /xg/m3 in Southeast Los
Angeles County. For each standard, estimates are presented assuming that the cutpoint is
equal to background, as well as for cutpoint weighting cases I, II, and III. All the estimates
presented in Exhibit 8.11 include uncertainty in the form of the rollback, as described in
Section 8.2 and Exhibit 8.7.
Abt Associates, Inc. p. 126 July 3, 1996, Revised
-------
Exhibit 8.10
Uncertainty Analysis: Effect of Uncertainty of Relative Risk, Background Concentration, Cutpoint, Slope Adjustment
Method, and Form of Rollback
Reduced Risk Associated with Meeting a PM-2.5 Standard of 15 f4.g/m3 Annual and 50 u,g/m3 Daily
Mortality Associated With Short-term Exposure to PM-2.5
Southeast Los Angeles County, 1995 (Population: 3.6 Million)
Reduced Risk
Associated
with Meeting
PM-2.5
Standards
(Number of
Deaths and as
% of
Total
Mortality)
25%
20%
1.5%
1.0%
0.5%
0.0%
«
«
95th % ile
) Mean (
5th % ile
1
1
3
1
I
1
>
(
1
1
J\
c
»
)
1
«
1
)
/
1
«
)
(
1
1
)
Just RR RR and Case I Case II Case HI RR and Case I Case 1 Case I
(background = background background
25ug/m3)
Cutpoint = RR. background, and Cutpoint = RR, background. »nd
Background Cutpoint Background Outpoint
600
500
400
300
200
100
0
Proportional Rollback
Cutpoint Weighting Schemes
Background
10 ng/in'
ISjig/m'
3 (>Hg/m'
Case I
0.5
0.3
O.I 5
0.05
Case II
0.2
0.3
0.3
0.2
Case III
0.05
O.I 5
0.5
0.3
Non-Proportional Rollback
Mean Reduced Risk as % of Total PM-Associated Risk
I
Proportional
Non-l'roportional
Just RR
52.1%
--
RR and Background
52.3%
52.4%
Case 1
60.5%
60.7%
Case 11
69.5%
70.0%
Case ill
78.8%
79.2%
Abt Associates, Inc.
p.127
July 3, 1996, Revised
-------
Exhibit 8.11
Uncertainty Analysis: Effect of Uncertainty of Relative Risk, Background Concentration, Cutpoint, Slope Adjustment
Method, and Form of Rollback
Reduced Risk Associated with Meeting Alternative PM-2.5 Standards
Mortality Associated With Short-term Exposure to PM-2.5
Southeast Los Angeles County, 1995 (Population: 3.6 Million)
Reduced Risk
Associated
with Meeting
PM-2.5
Standards
(Number of
Deaths and as
%of
Total
Mortality)
3.5%
3.0%
2.5%
2.0%
1 .5%
1 .0%
0.5%
0.0%
«
i
} <:
»
> |
» «
i
)
95th %ile
1
1
> «
1
5 «
P f.
>
5 ,
«
^
)
1
)
^ (
m
Meui (
R
<
1
Sth % He
)
»
)
/
(
»
<
1
\
)
(
>
1
«
1
1
)
/
«
*
1
)
1
CD o o o roooo CD o o O CD O o O
(UttQitt ai(uo)Qi ta to to to m at g> p
o w w « o w tft (LfACpw
*j" fl> re o> j^fcoa * ffi ft $
800
700
600
500
400
300
200
100
0
I I
15 ng/m' annual
no daily
Cutpoint Wcighling Schemes
I I
I I
15 (ig/m1 annual
65 (ig/m1 daily
15 ng/m1 annual
50 tig/m1 daily
1S u,g/m3 annual
25 ng/m3 daily
Mean Reduced Risk as % of Total PM-Associated Risk
Background
10 ng/m'
IS^ig/m'
30 fig/in1
Case 1
0.5
0.3
0. 1 5
0.05
Case II
0.2
0.3
0.3
0.2
Case III
0.05
0.15
0.5
0.3
(Hg/m1)
15 Annual only
15 Annual/65 Daily
15 Annual/50 Daily
15 Annual/25 Daily
KK and Background
42.2%
42.2%
52.4%
76. 9%
(Jase 1
49.7%
49.7%
60.7%
83.6%
(Jase 11
57.9%
57.9%
70.0%
90.0%
(Jase 111
67.3%
67.3%
79.2%
95.4%
Aht Associates, Inc.
p. 128
July 3, 1996. Revised
-------
9. Characterization of Risk Associated with PM Pollution: Interpreting the Results of the
Risk Analysis
This section discusses some issues related to the interpretation of the results of the risk
analyses presented above. Several risk analyses were carried out for each of the sample
locations (1) to assess the risks associated with "as is" PM levels and just attaining current
PM-10 standards, and with combinations of alternative daily and annual proposed standards;
(2) for several different health endpoints; (3) considering PM-10 and PM-2.5; and (4) using
concentration-response functions estimated by different studies as well as pooled analysis
concentration-response functions. The following points are discussed below:
(1) For any given concentration-response function, both the predicted health risk
associated with "as is" PM concentrations and the predicted risk reductions associated with
attaining alternative standards differ substantially between the two sample locations;
(2) At each sample location, both the predicted health risk associated with "as is" PM
concentrations and the predicted risk reductions associated with attaining alternative standards
are surrounded by substantial uncertainty;
(3) The PM-related health risk in one location may appear greater or smaller than the
PM-related health risk in another location, depending upon how PM-related health risk is
measured as a percent of total incidence or in absolute number of cases. (This distinction is
not obvious from the particular two locations examined in this report, but could well occur in
comparisons of PM-related health risks in other locations);
(4) The indicator of paniculate matter (i.e., PM-10 vs. PM-2.5) may be very important
in assessing PM-associated health risks;
(5) The mortality associated with annual average PM-2.5, estimated using the long-
term exposure studies, is notably greater than the mortality associated with daily average PM-
2.5, estimated by the short-term exposure studies; the estimate of mortality associated with
long-term exposure is also more uncertain, given greater uncertainty surrounding past
exposures; and
(6) The health effects incidence estimated by a risk analysis is based on the assumption
that the concentration-response relationship applies for all PM concentrations considered in the
analysis (e.g., down to the lowest PM level observed in the study which estimated the
concentration-response function or down to background level). If the relationship between PM
and a given health effect does not extend as low as the lowest level considered in an analysis,
then the predictions of incidence for that health effect will be overstated. Similarly, if the PM
concentrations considered in a risk analysis far exceed those observed in the study estimating
the concentration-response function, it may be the case that the estimated concentration-
response function is inappropriate for the very high PM concentrations considered in the risk
Abi Associates, Inc. p. 129 July 3, 1996, Revised
-------
analysis. The reader must infer the projected impact of PM on a location's public health based
on his or her judgement of the lowest concentrations for which a relationship between PM and
health can plausibly be drawn and the highest concentrations for which an exponential
concentration-response function is an appropriate model.
Each of the above points is discussed in turn below.
9.1. Variability of predicted health risks
There are substantial differences in the PM-related health risks estimated in
Philadelphia County and in Southeast Los Angeles County, even when the same concentration-
response function is used in both locations. When the measure of risk is the incidence of
health effects associated with "as is" PM concentrations or the incidence of health effects
associated with attaining alternative PM standards, these differences reflect, to a large extent,
the substantial differences in (1) the sizes of the exposed populations in Philadelphia County
(population in 1990 = 1.6 million) and Southeast Los Angeles County (population in 1990 =
3.6 million), (2) the baseline health effects incidence rates in the two locations, and (3)
estimated PM levels in the two locations (as is in Philadelphia, assuming attainment of current
standards in Los Angeles). When the measure of risk is in percentage terms (e.g., the number
of cases avoided due to attaining a standard divided by the total number of cases in the absence
of the standard), the difference in predicted risk based on a given concentration-response
function reflects only the difference in estimated PM levels between the locations (see Section
9.3).
The differences discussed above do not include the differences that would be expected
due to the fact that the concentration-response function for any given health endpoint probably
varies from one location to another. The actual differences in PM-related risks between the
sample locations could therefore be greater than the differences apparent in the risk analyses.
(It is possible, but unlikely, that the concentration-response functions and PM exposure
estimates would vary in such a way that the actual differences are less than the differences
apparent in the risk analyses.) The uncertainty surrounding the differences in risk estimates
between the two locations stems from the same sources of uncertainty surrounding risk
estimations within each location, described below.
9.2. Uncertainty surrounding predicted risks
As noted in Section 2, the risk analysis requires knowledge about relationships between
ambient PM concentrations and health effects, information on ambient air quality, baseline
incidence rates, and population sizes. Uncertainty in estimating each of these four factors
contributes to the uncertainty of predicted risks.
One of the primary quantifiable sources of uncertainty in the risk analyses is the
concentration-response function. The predictions of changes in health risks associated with
Abt Associates, Inc. p. 130 July 3, 1996, Revised
-------
changes in PM concentrations (e.g., to attain an alternative standard) depend crucially on this
function. For example, there are two estimates of short-term exposure mortality associated
with PM-10 above background in Southeast Los Angeles County (Exhibit 7.2), one using a
pooled analysis concentration-response function and the other using a concentration-response
function estimated by a study done in Los Angeles (Kinney et al., 1995). The PM-related
incidence of mortality estimated by the former is twice that estimated by the latter.
There are two sources of uncertainty associated with the concentration-response
function in the risk analyses. First, a concentration-response function appropriate for one
location may not be appropriate for another location. As discussed in Sections 3 and 9, true
underlying values of P in this function are likely to vary from one location to another. There
is therefore uncertainty associated with applying concentration-response functions estimated in
study locations (or functions derived by pooling these functions) to the sample locations.
Second, because the concentration-response functions are empirically estimated functions,
there is uncertainty surrounding these estimates. The Monte Carlo analyses presented in
Section 9, and the credible intervals (presented with all estimated incidences) derived from
these Monte Carlo analyses indicate a substantial degree of both kinds of uncertainty.
One reason for the uncertainty surrounding estimates of concentration-response
functions is the difficulty of accounting for possible confounding factors, including weather
and other pollutants. Both of these are often highly correlated with elevated PM
concentrations, the first because weather conditions can keep PM hi the air longer than usual
(in inversions, for example), and the second because many pollution sources emit more than
one pollutant. As in any regression, it is difficult to determine the separate effects of highly
correlated variables. In addition, a study recently reported by the Health Effects Institute (HEI
1995) found that terms taking into account the interactions of multiple pollutants were
significant predictors of health effects.
A concentration-response function could be biased if the measurement of average
ambient PM concentration is inaccurate in a systematic way. Most epidemiological studies use
the average PM levels reported at some number of PM monitors as the measure of the average
ambient PM concentration. This may or may not yield accurate measurements of the actual
daily average ambient PM concentrations in the study city. Depending on how the monitors
are placed, it could yield systematically inaccurate, or biased measurements. What is
important for the purpose of the risk analysis, in this case, is whether the measurement of
daily average ambient PM concentrations in the sample location is biased in the same way as
in the study city. That is, a systematic bias in the measurement of daily average ambient PM
concentrations in the study city is not a problem for the risk analysis if there is the same
systematic bias in the measurement of daily average ambient PM concentrations in the sample
location. Unfortunately, whether this is the case is unlikely to be known. Uncertainty about
the degree to which this is the case is another uncertainty in the risk analyses.
Abi Associates, Inc. p.131 July 3, 1996, Revised
-------
Finally, the prediction of health effects incidence (e.g., associated with attainment of a
PM standard or with "as is" PM concentrations) depends on the accuracy of the baseline
incidence data used. Obviously, multiplying the baseline incidence of some health effect by
some factor will multiply the predicted PM health effects by the same factor. Because county-
specific rates for mortality and hospital admissions are available from public health agencies
for both Philadelphia County and Los Angeles County, these estimates are considered quite
reliable. In the absence of city-specific rates for incidence of some respiratory symptoms, the
risk analyses used rates from the epidemiological studies. This introduced another component
of uncertainty into the risk analyses.
9.3. Importance of the measure of risk
A PM-related health effect risk may be characterized hi terms of the percent of the total
incidence of the health effect that is associated with PM concentrations above a certain level or
the actual number of cases (i.e., the incidence) of the health effect associated with PM
concentrations above a certain level. Both measures are presented in this report. The measure
of risk used may affect the assessment of the degree of risk in one location relative to another.
Suppose, for example, that location X has much higher PM levels than location Y. If PM-
related mortality risk is measured in terms of the percent of total mortality associated with
PM, location X will appear to have greater PM-related mortality risk than location Y.
However, if location Y's population is much larger than location X's, it is quite possible that a
greater incidence of PM-related mortality will be predicted in location Y than location X.
(This possibility was not apparent hi the results from Philadelphia County and Southeast Los
Angeles County because both the PM levels and the population of the latter exceed that of the
former.)
Unlike the incidence of PM-related mortality, the percent of total mortality that is PM-
related is affected by neither the size of the exposed population nor the baseline health effect
incidence rate. Given a concentration-response function (i.e., a 0), this measure of risk is
affected only by the actual PM concentrations (relative to the alternative being considered).
For an individual considering the PM-related risks to himself in one location versus another,
the percent of the total incidence of the health effect is the appropriate measure of risk. It
yields an ordering of locations by their PM-related risk that is the same as the ordering
achieved by simply measuring the PM concentrations in the different locations. PM levels
were higher in Southeast Los Angeles County than in Philadelphia County, so for any given
individual the PM-related health risks are correspondingly higher in Southeast Los Angeles
County than in Philadelphia County. (Compare, for example, the "percent of total incidence"
columns for Philadelphia County, in Exhibit 7.1, with the corresponding columns for
Southeast Los Angeles County, in Exhibit 7.2.)
To measure risks to society, however, the number of affected individuals is important.
Even if the risk per individual is higher in location A than location B, if many more
individuals are exposed in location B, the total risk to society may be greater in location B.
Abt Associates, Inc. p. 132 July 3, 1996, Revised
-------
The actual number of PM-related cases and the percent of all cases that are PM-related are
both appropriate measures of risk that deal with different questions.
9.4. Importance of the indicator of PM: PM-10 vs. PM-2.5
One important reason that the concentration-response relationship between PM-10 and
a given health endpoint may vary from one location to another is that the composition of PM-
10 varies significantly throughout the United States. In some areas, paniculate matter is
composed mostly of coarse particles; in other areas there is a much larger fine particle
fraction. In Philadelphia County, for example, PM-2.5 comprises about 73 percent of PM-10
. (see Exhibits 4.7 and 4.8), whereas in Southeast Los Angeles County, PM-2.5 comprises only
about 59 percent of PM-10 (Exhibits 4.12 and 4.13). If all particle sizes are equally harmful
in causing a health effect, then this type of variability hi PM-10 composition will not matter.
If different size particles are differentially harmful, however, a given concentration of PM-10
in one location may have a different impact on health than the same concentration of PM-10 in
another location. (The chemical composition of the PM may be important as well. For
example, there has been some investigation of the health impacts of sulfates, a common
chemical component of PM pollution. However, particle size is one important factor and the
one for which the most complete data are available.)
Suppose, for example, that only the fine fraction (PM-2.5) adversely affects human
health and the rest of PM-10 has no adverse effect at all. Suppose also that in location A, only
fifty percent of PM-10 is PM-2.5 whereas in location B 90 percent of PM-10 is PM-2.5. In
this case, 1 jig/m3 of PM-10 in location A translates into 0.5 /ng/m3 of PM-2.5, whereas in
location B 1 /ig/m3 of PM-10 translates into 0.9 ng/m3 of PM-2.5. What appear to be equal
exposures in the two locations when PM-10 is used as the indicator of paniculate matter
pollution exposure, are actually substantially different exposures to the only component that
actually affects health.
There are, then, two important issues for a risk analysis concerning the choice of
indicator of paniculate matter pollution:
(1) different components of PM-10 may affect a health endpoint to different degrees
(i.e., may be differentially "potent"), and
(2) the ratios of these different components within PM-10 may change over time or
from place to place, changing the health effects associated with paniculate matter
pollution.
If all particle sizes are equally potent in causing a given health effect, then the
relationship between PM-10 and that health effect is adequately described by the basic model
introduced in equation (1) in Section 2. If, however, PM-2.5 and the coarse fraction of PM-10
are differentially potent, then that model may not adequately describe the relationship between
Abt Associates, Inc. p. 133 July 3, 1996, Revised
-------
PM-10 and the health effect, because a given concentration of PM-10 may be associated with
different levels of health effect depending on the composition of the PM-10.
A generalization of the basic exponential concentration-response relationship between a
health effect and PM, either PM-10 or PM-2.5 presented in Section 2 (equation (1) is
presented hi Appendix 4. In this generalization, health effect incidence is a function of both
PM-2.5 and the coarse fraction of PM-10. The basic model (equation (1)) is then seen to be
the special case in which PM-2.5 and the coarse fraction are equally potent. It is also shown
that if PM-2.5 and the coarse fraction are not equally potent, then the relationship between
PM-10 and a given health endpoint will depend both on the relative potencies of the two
fractions and on the ratio of PM-2.5 to PM-10. If this ratio varies from one location to
another, then the PM-10 concentration-response function will vary as well, even if the exposed
populations are identical.
In the absence of such information, risk analyses considering PM-10 use only PM-10
concentration-response functions (acknowledging the uncertainty introduced by the possibility
that the composition of the PM-10 in the study location may differ from the composition of the
PM-10 in the sample locations in which the function is applied). Risk analyses considering
only PM-2.5 use PM-2.5 concentration-response functions when they are available. In a few
cases (e.g., for ischemic heart disease and congestive heart failure), PM-2.5 data are used with
PM-10 functions (in the absence of PM-2.5 functions). The health effects incidence
predictions from such analyses are unlikely to overestimate the PM-2.5-related health risks.
9.5. Risk predictions based on concentration-response functions from long-term
exposure studies versus those from short-term exposure studies
The long-term exposure study (Pope et al., 1995) estimates the apparent effect of PM
on mortality to be much greater than do short-term exposure studies. This suggests that the
effects of long-term exposure may not merely be the sum of the effects of short-term exposures
over the course of a year. It is possible that the long-term exposure study is detecting
mortality related to long-term PM exposure, in addition to the mortality precipitated by short-
term PM exposure. It is not unreasonable to suspect that prolonged exposure to elevated PM
levels, as well as exposure to short-term peak PM levels, might cause health problems. In
addition, the effects of the two types of exposure might be related.
While long-term exposure studies use long-term average PM concentration (which can
be approximated by annual average PM concentration) as the PM indicator, these studies are
generally conducted in such a way that they may be detecting effects due to PM exposure over
some longer period. For example, average PM concentrations over the course of five years
might be the appropriate measure. Such possible discrepancies between the actual relevant
exposure period (e.g., five years) and the exposure period considered by a long-term exposure
study (e.g., one year) could have at least two effects.
Abt Associates, Inc. p. 134 July 3, 1996, Revised
-------
It is possible that the full benefits of reducing PM predicted by such studies would not
appear in the first year after reductions to attain a standard, but would be "phased in"
gradually as concentrations during successive years were also reduced. If average PM
concentrations over five years is the appropriate measure, for example, the benefits of a
standard would gradually increase to their full level over the course of the five years after the
new standard had been attained. The risk analysis does not attempt to determine the
appropriate exposure period for the long-term exposure study on mortality. The estimated
annual benefits of reduced long-term exposure are assumed to be completely achieved by the
future year for which attainment of the new standard is being modeled.
It is also possible that the predicted incidence of mortality that is associated with an
annual average PM level could be either an over- or underestimation of risk. This could be
the case if the actual relevant exposure period is substantially different from one year, and the
average PM levels over the relevant exposure period are substantially different from the annual
average PM levels used in a long-term exposure study.
Finally, the prematurity of deaths associated with PM may be of crucial importance,
whether the short-term exposure or the long-term exposure study is used. PM-related deaths
that are days or even weeks earlier than they would be at lower PM levels may cause less
public-health concern than PM-related deaths of individuals who could otherwise expect to live
many more years. Even finding that most of the people who die during PM episodes are
seriously ill would not resolve the question, since without the additional stress of high PM
concentrations, some of those people might fully recover and live significantly longer. This
issue so far has not been resolved in the epidemiological literature.
As indicated in the CD (EPA, 1996a) and the Staff Paper (EPA, 1996b), the public
health burden of ambient PM-mediated mortality depends on both the number of deaths and
the shortening of life that PM exposure causes or promotes (CD, p. 13-44). Risk analysis
estimates of percentage incidence (and incidence counts) of mortality associated with PM could
vary substantially depending on the general prematurity of death involved. For instance, if
prematurity of death associated with short-term exposures to PM generally was only on the
order of days or weeks, then ambient PM concentrations in the two risk analysis locations
would be expected to have less of an impact on annual mortality incidence than indicated by
the base case analysis.18 In this case, PM would be temporally associated with the proportion
of annual mortality events reported in the risk analysis, but PM would be associated with a
lower proportion of the overall mortality rate. This would result because many of the events
that PM would be associated with, if the assumption of little prematurity of death in general is
accurate, would have occurred in the absence of PM involvement days or weeks later. Thus
the absence or reduced concentrations of PM would not affect mortality rates to the same
extent to which PM was temporally associated with mortality events.
18Such a pattern of apparent mortality displacement of only a few days is often seen for some other
environmental effects, such as high temperature (Staff Paper, V-17).
Abt Associates, Inc. p. 135 July 3, 1996, Revised
-------
For this reason, the alternate health measure of Life Years Lost is often employed to
measure the public health burden of environmental factors, rather than simply estimates of
mortality incidence. As the CD indicates, however, confident quantitative determination of
years of life lost to ambient PM exposure is not yet possible (CD p. 13-44). A couple of
studies of mortality from short-term exposures suggest that some portion of PM-induced
mortality may occur among individuals already so ill that they would soon die without PM
exposure (Spix et al. 1993; Ciraentes and Lave 1996), while other studies (Dockery et al.
1993; Pope et al. 1995) report associations between PM and changes in mortality rates,
findings that cannot be solely explained by death-bed effects (i.e., mortality with little
prematurity) (Utell and Frampton 1995). A sustained reduction in particle levels in Utah
Valley over 14 months was also accompanied by a drop in mortality rates, consistent with the
hypothesis that a substantial portion of PM-associated mortality may involve mortality of
sufficient prematurity to affect mortality rates.
Incomplete knowledge of the true excess mortality and prematurity of death associated
with PM exposures complicates this risk analysis and adds uncertainty to the interpretation of
the mortality risk estimates. This difficulty would be expected to most greatly complicate
interpretation of the estimates of mortality associated with short-term exposures compared to
the estimates of mortality associated with long-term exposures.
9.6. Dependence of results on the assumption that the concentration-response
relationship is applicable at low concentrations
The change in health effects incidence predicted by a risk analysis to be associated with
reducing "as is" PM levels either to background or to concentrations that meet alternative
standards is based on the assumption that the concentration-response relationship applies down
to the lowest concentration considered in the analysis. If the relationship between PM and a
given health effect does not extend to the lowest PM levels under consideration, then the
predictions of health effects incidence will be biased. If, for example, there is a level above
the lowest PM level considered below which there are no health effects, then the change in
health effects incidence will be overstated. The degree of overstatement will depend on the
discrepancy between the lowest level considered and the lowest PM level at which there are
PM-related health effects. In the absence of knowledge concerning the lowest level at which
effects occur, the reader must infer the projected impact of paniculate matter on a location's
public health based on his or her judgement of the lowest concentrations for which the given
relationship between PM and health can plausibly be drawn and the plausible shape of the
concentration-response function below that level. This issue has been partially explored in a
series of sensitivity and uncertainty analyses presented in Section 7.
Abt Associates, Inc. p. 136 July 3, 1996, Revised
-------
Appendix 1: The Relationship Between the Ambient Concentration-Response Function
and the Individual Exposure-Response Function
1. Individual exposure versus ambient concentration and the individual exposure-
response relationship versus the ambient concentration-response relationship
There has been persistent concern that epidemiological studies which use ambient PM
concentration data as a surrogate for individual exposure to PM may produce biased estimates
of individual exposure-response functions relating health effects to individual exposure. This
is a valid concern. If such functions were used with individual exposure data from the study
location, the predicted health responses could be biased, as discussed below. If such functions
are used with ambient PM concentration data from the study location, however, there is no
reason to suspect that the predicted health responses would be biased (unless there are other
sources of bias). The relationship estimated in the studies is, after all, between the population
health response and average ambient PM concentration.
To help clarify some potential confusion, two relationships are distinguished. The
relationship between a health response and individual exposure to PM is referred to as an
individual exposure-response relationship. On an individual level, this is the relationship
between the actual exposure to PM (in /xg/m3) experienced by the individual and the
probability that that individual will exhibit the health response. On an aggregate level, it is the
relationship between the average exposure to PM (in /xg/m3) by individuals in the population
and the population response (number of individuals exhibiting the health response).19
The relationship between a health response and ambient PM concentration is referred to
as the ambient concentration-response relationship. It is the relationship between the average
ambient concentration of PM (in ^g/m3) and the population response.20 Both the individual
exposure-response relationship and the ambient concentration-response relationship are of
interest. The individual exposure-response relationship is of clear scientific interest. This is
the relationship that epidemiological studies would presumably estimate if they had data on
individual exposure. Because the NAAQS influence ambient concentrations of PM, it is the
ambient concentration-response relationship that is of interest for this risk analysis, because the
risk analysis examines the risk reduction associated with changing ambient concentrations,
rather than the risk reduction associated with changing individual exposure (which is not
directly controlled by the NAAQS).
"Each individual has an average exposure to PM over time (e.g., a daily average). The average
individual exposure to PM is the average, over all individuals, of these time-averages. Because there are factors
other than average ambient PM concentration that affect individual exposure (as described in the model discussed
below), average individual exposure to PM does not necessarily equal average ambient PM concentration.
:°The average ambient PM concentration is an average over both time and space (e.g., the average over a
given geographic area of 24-hour averages).
Abt Associates, Inc. p. 137 July 3, 1996, Revised
-------
The two relationships (the individual exposure-response relationship and the ambient
concentration-response relationship) are related by the connection between individual exposure
and ambient concentration, as detailed in the model below. Let
Z = average individual exposure to PM (in /ig/m3),
X = average ambient concentration of PM (in /ig/m3),
Y = the health response (e.g., mortality),
Q = a vector of variables, other than ambient concentration, that affect individual
exposure, (e.g., average percent of time spent indoors), and
T = a vector of variables, other than individual exposure to PM, that affect the health
response.
For ease of illustration, certain functional forms are assumed. The relationship
between individual exposure and ambient concentration is assumed to be linear. The
relationship between the health response and individual exposure to PM is assumed to be log-
linear. In addition, it is also assumed, for ease of discussion, that Q and T are each single
variables.
The equations that follow are, as noted above, intended only to provide an example to
illustrate the relationship between average individual exposure, average ambient concentration,
and population health response. While deviations from any of the assumptions stated in this
example may alter the particular functional forms of the equations presented below, they will
not affect the basic ideas discussed in this appendix.
The relationship between average individual exposure to PM (Z) and average ambient
concentration of PM (X) is given by
Z = a, + fix + QoQ (16)
and the relationship between the population health response (Y) and average individual
exposure (Z), the individual exposure-response relationship, is given by
In7 = 6 + AZ + yT . (17)
Substituting equation (1) into equation (2) yields the ambient concentration-response
relationship:
= 6 + X(ao + $X + 6o2) + jT (18)
= a + $X + 60 + jT , (19)
Abt Associates, Inc. p.138 July 3, 1996, Revised
-------
where
a = 6 + Aao , p = AP0 , and, 0 = A.6, . (20)
It is of scientific interest to estimate an individual exposure-response relationship,
equation (2). Lacking data on individual exposure, Z, however, epidemiological studies use
ambient concentration, X, as a surrogate for individual exposure, Z. Such studies estimate
equation (4), then, instead of equation (2). If the estimate of p is an unbiased estimate of p,
then it can also be an unbiased estimate of X, the coefficient of individual exposure in the
individual exposure-response function, if and only if p0 = 1. That is,
P = XP = A, if and only if p = 1 .
Therefore the source of the bias in estimating the coefficient of individual exposure, A.,
does not exist for estimating the coefficient of ambient concentration, p. This is precisely
because it is the ambient concentration-response relationship, rather than the individual
exposure-response relationship that is being estimated in the epidemiological studies. Whereas
this may present a problem for the epidemiological studies that seek to estimate an individual
exposure-response relationship, it should not present a problem for a risk assessment, for
which it is the ambient concentration-response relationship that is of interest.
2. Other possible sources of bias in the estimate of the ambient concentration-response
relationship and mitigating influences
Although epidemiological studies usually include in their models those variables that
are likely to affect the health response of interest (variables in the vector T, such as
temperature and time trends), they do not necessarily include those variables that may affect
individual exposure (variables in the vector Q, such as the average percent of time spent
indoors). While the actual ambient concentration-response relationship is equation (4), the
model estimated is more likely to be
Inr = a + pX + y^ - (22)
That is, the model estimated may have omitted variables. This raises the possibility that
omitted variables could cause bias in the estimates of coefficients in the model. Those
variables that are omitted, however, are likely to be highly correlated with the variables in the
vector T - in particular, with temperature. (For example, the percent of time spent indoors
should be correlated with temperature.) While omission of these variables may cause bias in
the coefficient of temperature, then, it is unlikely to cause bias in the coefficient of X. In fact,
the more highly correlated an omitted variable is with a variable in T, e.g., temperature, the
less of a bias problem there is in estimating the coefficient of ambient concentration, p.
Abt Associates, Inc. p. 139 July 3, 1996, Revised
-------
Including variables such as temperature in the model, then, would tend to mitigate any bias
problem.
3. Transferability of a concentration-response relationship estimated in one location to
another location
It is argued in sections 1 and 2 above that, while using ambient concentration as a
proxy for individual exposure may produce a biased estimate of the individual exposure-
response relationship in the study location, it should produce an unbiased estimate of the
ambient concentration-response relationship in that location. Applying ambient concentration
data from the study location, then, should produce unbiased predictions of health response in
that location (barring any other possible sources of bias).
If ambient concentration data from a different location are applied to the estimated
ambient concentration-response function from the study location, will the predicted health
response in that different location be biased? That depends on whether the ambient
concentration-response relationship in the study location is the same as that in the location to
which it is being applied. This is the issue of transferability.
Recall that the ambient concentration-response relationship estimated is
In7 = a
where P = A,p0. If either A, or p0 differs between one location and another, then p will
differ, and the ambient concentration-response relationship will be different as well. PO might
differ among locations if, for example, coarse particles are less likely to infiltrate indoor air.
In this case, PM-10 with a high proportion of coarse particles would result in lower indoor
exposure than PM-10 with a high proportion of fine particles. If the percent of time spent
indoors is the same, then the location with the coarser PM-10 would have lower individual
exposure to PM-10 than the location with the finer PM-10. A might differ among locations if,
for example, the population in one location has a higher proportion of a susceptible subgroup
than another location. Another reason X might differ among locations is if the composition of
the PM among locations differs and the composition of the PM affects its toxicity (see
Appendix 4).
While applying an ambient concentration-response function estimated in one location to
another location may give biased health response predictions, the direction or magnitude of the
bias is not known, and will depend on the particular pair of locations. This issue is less one of
bias and more one of uncertainty and is addressed in Section 9 of the report.
Abt Associates, Inc. p. 140 July 3, 1996, Revised
-------
Appendix 2: Pooling the Results of Different Studies
Many studies have attempted to determine the influence of paniculate matter pollution
on human health. Usually this involves estimation of a parameter P in a concentration-
response function, which may be linear or non-linear, as discussed above. Each study
provides an estimate of p, along with a measure of the uncertainty of the estimate. Because
uncertainty decreases as sample size increases, combining data sets is expected to yield more
reliable estimates of p. Combining data from several comparable studies in order to analyze
them together is often referred to as meta-analysis.
For a number of reasons, including data confidentiality, it is often impractical or
impossible to combine the original data sets. Combining the results of studies in order to
produce better estimates of P provides a second-best but still valuable way to synthesize
information (DerSimonian and Laird, 1986). This is referred to as pooling results in this
report. Pooling requires that all of the studies contributing estimates of p use the same
functional form for the concentration-response function. That is, the P's must be measuring
the same thing.
One method of pooling study results is simply averaging all reported P's. This has the
advantage of simplicity, but the disadvantage of not taking into account the uncertainty of each
of the estimates. Estimates with great uncertainty surrounding them are given the same weight
as estimates with very little uncertainty. For example, consider the three studies whose
results are presented in Exhibit A2.1.
Exhibit A2.1. Three Sample Studies.
Study
Study 1
Study 2
Study 3
Estimate of p
0.75
1.25
1.00
Standard
Deviation
0.35
0.05
0.10
Variance
0.1225
0.0025
0.0100
The average of the three estimates is 1.0. However, the study 2 estimate has much less
uncertainty associated with it (variance = 0.0025) than either the study 1 or study 3 estimates.
It seems reasonable that a pooled estimate which combines the estimates from all three studies
should therefore give more weight to the estimate from the second study than to the estimates
from the first and third studies. A common method for weighting estimates involves using
their variances. Variance takes into account both the consistency of data and the sample size
used to obtain the estimate, two key factors that influence the reliability of results.
Abt Associates, Inc.
p. 14]
July 3, 1996, Revised
-------
The exact way in which variances are used to weight the estimates from different
studies in a pooled estimate depends on the underlying model assumed. The next Section
discusses the two basic models that might underlie a pooling and the weighting scheme derived
from each.
A2.1 The fixed effects model
The fixed effects model assumes that there is a single true concentration-response
relationship and therefore a single true value for the parameter p. Differences among P's
reported by different studies are therefore simply the result of sampling error. That is, each
reported P is an estimate of the same underlying parameter. The certainty of an estimate is
reflected in its variance (the larger the variance, the less certain the estimate). Pooling that
assumes a fixed effects model therefore weights each estimate under consideration in
proportion to the inverse of its variance.
Suppose there are n studies, with the ith study providing an estimate p, with variance \,
(1 = 1, -.,n). Let
v,
denote the sum of the inverse variances. Then the weight, Wj , given to the ith estimate, PJ , is
1/v
w. =
This means that estimates with small variances (i.e., estimates with relatively little uncertainty
surrounding them) receive large weights, and those with large variances receive small weights.
The estimate produced by pooling based on a fixed effects model, then, is just a
weighted average of the estimates from the studies being considered, with the weights as
defined above. That is,
P/e = E ", * P,
The variance associated with this pooled estimate is the inverse of the sum of the inverse
variances:
Vfe ~
1
Abt Associates, Inc. p. 142 July 3, 1996, Revised
-------
Exhibit A2.2 shows the relevant calculations for this pooling for the three sample studies
summarized in Exhibit A2.1.
Exhibit A2.2. Fixed Effect Model Calculations.
Study
1
2
3
Sum
Pi
-0.75
1.25
1.00
Vi
0.1225
0.0025
0.0100
1/v,
8.16
400
100
£ = 508.16
Wi
0.016
0.787
0.197
£ = i.ooo
Wi*Pi
0.012
0.984
0.197
I = 1.193
The sum of weighted contributions in the last column is the pooled estimate of p based on the
fixed effects model. This estimate (1.193) is considerably closer to the estimate from study 2
(1.25) than is the estimate (1.0) that simply averages the study estimates. This reflects the fact
that the estimate from study 2 has a much smaller variance than the estimates from the other
two studies and is therefore more heavily weighted in the pooling.
The variance of the pooled estimate, vfe, is the inverse of the sum of the variances, or
0.00197. (The sums of the PJ and vs are not shown, since they are of no importance. The sum
of the 1/Vj is S, used to calculate the weights. The sum of the weights, ws , 1 = 1, ,.., n, is 1.0,
as expected.)
A2.2 The random effects model
An alternative to the fixed effects model is the random effects model, which allows the
possibility that the estimates Ps from the different studies may in fact be estimates of different
parameters, rather than just different estimates of a single underlying parameter. In studies of
the effects of PM-10 on mortality, for example, if the composition of PM-10 varies among
study locations the underlying relationship between mortality and PM-10 may be different
from one study location to another. For example, fine particles make up a greater fraction of
PM-10 in Philadelphia than in El Paso. If fine particles are disproportionately responsible for
mortality relative to coarse particles, then one would expect the true value of p in Philadelphia
to be greater than the true value of P in El Paso. This would violate the assumption of the
fixed effects model.
The following procedure can test whether it is appropriate to base the pooling on the
random effects model (vs. the fixed effects model):
A test statistic, Qw , the weighted sum of squared differences of the separate study estimates
from the pooled estimate based on the fixed effects model, is calculated as:
Abt Associates, Inc.
p.143
July 3, 1996, Revised
-------
e. - E, 1
Under the null hypothesis that there is a single underlying parameter, p, of which all the Pi's
are estimates, Qw has a chi-squared distribution with n-1 degrees of freedom. (Recall that n is
the number of studies in the meta-analysis.) If Qw is greater than the critical value
corresponding to the desired confidence level, the null hypothesis is rejected. That is, in this
case the evidence does not support the fixed effects model, and the random effects model is
assumed, allowing the possibility that each study is estimating a different p.
The weights used in a pooling based on the random effects model must take into
account not only the within-study variances (used in a meta-analysis based on the fixed effects
model) but the between-study variance as well. These weights are calculated as follows:
Using Qw , the between-study variance, if, is:
I V^
It can be shown that the denominator is always positive. Therefore, if the numerator is
negative (i.e., if Qw < n-1), then if is a negative number, and it is not possible to calculate a
random effects estimate. In this case, however, the small value of Qw would presumably have
led to accepting the null hypothesis described above, and the meta-analysis would be based on
the fixed effects model. The remaining discussion therefore assumes that if is positive.
Given a value for if , the random effects estimate is calculated in almost the same way
as the fixed effects estimate. However, the weights now incorporate both the within-study
variance (v) and the between-study variance (if). Whereas the weights implied by the fixed
effects model used only vi5 the within-study variance, the weights implied by the random
effects model use v; +T)2.
Let Vj* = Vj +if. Then
Abt Associates, Inc. p. 144 July 3, 1996, Revised
-------
and
1/v
I
w, =
S'
The estimate produced by pooling based on the random effects model, then, is just a weighted
average of the estimates from the studies being considered, with the weights as defined above.
That is,
rand
ft-
The variance associated with this random effects pooled estimate is, as it was for the fixed
effects pooled estimate, the inverse of the sum of the inverse variances:
Vrand
I, 1/V
The weighting scheme used in a pooling based on the random effects model is basically
the same as that used if a fixed effects model is assumed, but the variances used in the
calculations are different. This is because a fixed effects model assumes that the variability
among the estimates from different studies is due only to sampling error (i.e., each study is
thought of as representing just another sample from the same underlying population), while the
random effects model assumes that there is not only sampling error associated with each study,
but that there is also between-study variability each study is estimating a different underlying
P. Therefore, the sum of the within-study variance and the between-study variance yields an
overall variance estimate.
A2.3 An example
This Section demonstrates the relevant calculations for pooling using the example in
Exhibit A2.1 above.
First calculate Qw , as shown in Exhibit A2.3.
Abi Associates, Inc. p. 145 July 3, 1996, Revised
-------
Exhibit A2.3: Calculation of Q«
Study
1
2
3
Pi
0.75
1.25
1.00
l/Vi
8.16
400
100
l/V^Pi-Pfe)2
1.601
1.300
3.725
I = Qw = 6.626
In this example the test statistic Qw = 6.626. The example considers three studies, so Qw is
distributed as a chi-square on two degrees of freedom. The critical value for the 5 percent
level (i.e., corresponding to a 95 percent level of confidence) for a chi-square random variable
on 2 degrees of freedom is 5.99. Because Qw = 6.626 > 5.99, hence the null hypothesis is
rejected. That is, the evidence does not support the fixed effects model. Therefore assume the
random effects model is appropriate.
Then calculate the between-study variance:
2 _ 6.626 - (3 - 1)
508.16 -
170066.65
508.16
= 0.0267
From this and the within-study variances, calculate the pooled estimate based on the random
effects model, as shown in Exhibit A2.4.
Exhibit A2.4. Random Effects Model Calculations.
Study
1
2
3
Sum
Pi
0.75
1.25
1.00
Vi + n2
0.1492
0.0292
0.0367
l/(Vi +T12)
6.70
34.25
27.25
£ = 68.20
Wj*
0.098
0.502
0.400
£ = 1.000
Wj* X P|
0.0735
0.6275
0.400
1 = 1.101
The random effects pooled estimate, Prand , is 1.101. It's variance, vrand , is l/(68.2) = 0.015.
Abt Associates, Inc.
p. 146
July 3. 1996, Revised
-------
Appendix 3: The Concentration-Response Function and Relative Risk
The basic "Poisson regression" concentration-response relationship commonly found in
the epidemiological literature is
y = B e** , (35)
where x is the PM level, y is the incidence of the health endpoint of interest at PM level x, (J
is the coefficient of PM, and B is the incidence at x=0, i.e., when there is no paniculate
matter. (Either incidence or incidence rate may be used as long as y and B are consistent.)
If x denotes the actual ("as is") PM level and y denotes the baseline incidence (rate) of
the health endpoint, i.e., the incidence (rate) corresponding to the "as is" PM level, letting XQ
denote some specified alternative PM level and y0 denote the incidence (rate) associated with
that alternative PM level, then
y = B e*° . (36)
The change hi health effects incidence, Ay = y0 - y, corresponding to a given change
in PM levels, Ax = XQ - x, can be derived from equations (1) and (2)21 as follows:
First, dividing equation (2) by equation (1) yields
o _ 6 p(*0 *) RAX /IT\
e - ev (37)
y e^
Then multiplying through by y yields
y0 =y**** (38)
Subtracting y from both sides gives
Av ~ y^ ~ y ~ ye ~ y = y[£ ~ 1 ] (39)
2lBecause the Poisson regression form of concentration-response function (equation (1)) is by far the most
common form, the discussion that follows assumes that form.
Abi Associates, 'Inc. p.147 July 3, 1996, Revised
-------
or
Ay=y(e**-\}. (40)
Alternatively, the change in health effects incidence can be calculated using relative
risk. Relative risk (RR) is a well known measure of the comparative health effects associated
with a particular exposure comparison. The risk of mortality at PM level XQ relative to the
risk of mortality at PM level x, for example, may be characterized by the ratio of the two
mortality rates: the mortality rate among individuals exposed to PM level XQ, i.e., y0, and the
mortality rate among (otherwise identical) individuals exposed to PM level x, i.e., y. This is
the left-hand side of equation (3). That is,
n
*»=*=£_: =,«*-*> = e^ , (41)
or
= *PA* (42)
Given a concentration-response function (i.e., a particular value for the coefficient, P),
then, and a particular change in PM levels, Ax, the relative risk associated with that change in
PM, denoted as RRAx, can be calculated from equation (8). This is particularly significant,
because it means that the relative risk corresponding to any change in PM levels is easily
calculated. In particular, using equation (8), it is straightforward to convert a relative risk
corresponding to one Ax into a relative risk corresponding to a different Ax. Suppose, for
example, that a relative risk from a study reflects the relative mortality risks associated with
the two PM levels, XQ and Xj. Then from equation (8),
**- -""'"' <«>
Solving for P yields
P =
Now the relative risk corresponding to Ax = (x, - x2) can be calculated, using P and equation
(8) again as
Abt Associates, Inc. p. 148 July 3, 1996, Revised
-------
Substituting equation (8) into equation (6), it becomes clear that the change in health
effects incidence, Ay, corresponding to a given change in PM levels, Ax, can be calculated
based on the relative risk corresponding to Ax as:
Aj> = y(RR^ - 1) . (46)
Equations (6) and (12) are simply alternative ways of writing the relationship between a given
change in PM levels, Ax, and the corresponding change in health effects incidence, Ay.
Note that, in the above, the baseline health effect incidence (rate), y, refers to the
incidence (rate) corresponding to the "as is" PM level. This is because the baseline incidences
used in the calculations for the risk analysis are drawn from available health statistics which
include the effects of exposure to air pollution. Changes in incidence, Ay, correspond to
reductions in PM concentrations. Because Ax is negative (a reduction in PM concentration),
RRAx will be less than 1, and Ay will also be negative that is, the number of cases of the
health effect avoided.
If the general population is not typically exposed to the risk factor of interest, however,
then the baseline incidence (rate) would be the incidence in the absence of exposure to the risk
factor under consideration. In this case, the relative risk associated with exposure to the risk
factor would be positive the increase in cases due to exposure, as opposed to the baseline
incidence, in the absence of exposure. This is a common situation, an example of which is
provided by Samet and Spengler, 1993.
The formula for "population attributable risk" given by Samet and Spengler, 1993
B (RR - 1)
B RR
where RR is the relative risk associated with an increase in the risk factor (i.e., Ax is positive
here), and B is the baseline incidence now, the incidence of the health effect in the absence
of exposure to the risk factor under consideration. That is, the formula computes the increase
in incidence divided by total incidence.
Assume now that the risk factor in both cases (i.e., in the risk analysis reported here
and in Samet and Spengler) is PM. Then a relationship exists between the baseline incidence,
B, defined by Samet and Spengler as the incidence in the absence of exposure to the risk
factor, and the baseline incidence used in the calculations far this risk analysis, y, defined as
Abt Associates, Inc. p. 149 July 3, 1996, Revisec
-------
the incidence associated with the "as is" level of the risk factor (PM). In particular^ = B *
RR. That is, the incidence of the health effect associated with "as is" PM levels is the
incidence of the health effect in the absence of PM (or at background PM level) times the
relative risk associated with the change in PM levels
Substituting B = y/RR from above in the Samet and Spengler formula yields
or, now cancelling y,
1 - (l/RR) ,
which is the formula used in the calculations in the risk analysis. If the relative risk RR is
associated with an increase in pollution, Ax, then a decrease in pollution, -Ax, is associated
with a relative risk 1/RR. Therefore, the formula used in these calculations computes the
percent change in incidence associated with a decrease in pollution from some previously
existing level (in this analysis, from observed concentrations to background or cutpoint
concentrations). The percent is calculated with respect to the incidence at the existing level,
that is, from the full incidence in the population.
As an example, assume that RR = 1.2. Then Samet and Spengler's formula gives
(1.2 - 1)7(1.2) = .167, that is, 16.7% of incidence attributable to the risk factor. Our formula
is 1- (1/1.2) = 1- .833 = .167, the same result given by Samet and Spengler.
Abj Associates, Inc. p. 150 Juh 3, 1996, Revised
-------
Appendix 4: A Generalization of the Basic Concentration-Response Function
This appendix presents a model which generalizes the basic concentration-response
model to explicitly incorporate two important considerations: (1) the fine and coarse fractions
of PM-10 may affect a health endpoint in very different ways, and (2) the ratio of fine to
coarse particles hi PM-10 can change over time or from place to place, changing the health
effects associated with paniculate matter pollution.
Assume that fine and coarse particles cause health effects independently of one another.
Assume further that the Poisson regression model used in most epidemiological studies is an
appropriate model. The model postulates that the effects of a pollutant are multiplicative, that
is, that an increase in pollution implies some percent increase in health effects. Therefore, if
B is the hypothetical "base incidence" when paniculate concentration is zero, then health
effects are estimated using the equation
y =
(50)
where xf = the amount of fine particles (PM-2.5) (in micrograms per cubic meter),
x,. = the amount of coarse particles (PM-10 minus PM-2.5),
Pf = the beta regression coefficient measuring the effect of fine particles,
PC = the beta regression coefficient measuring the effect of coarse particles,
B = the (hypothetical) base incidence rate when no paniculate matter is present, that
is, when xf =xc = 0, and
y = the health effect (mortality is used for purposes of discussion below).
Similarly, let
xio = xf + xc = the total amount of PM-10, and
Pio = the beta regression coefficient measuring the effects of PM-10.
Note that if pf = pc (i.e., if there is no difference in potency between the fine and the coarse
fractions), then the generalized model (1) reduces to the basic model,
y = B * e^ . (51)
Because multiplication is commutative, and because exp[a]*exp[b] = exp[a+b], it does
not matter in what order exposure to additional pollutants is thought of as taking place. (Note
that one could assume a base rate B' at some non-zero pollutant combination, xf° and xc°; the
equation would remain the same under the change of variables xf' =xf - xf° and xc' =xc - xc°.
This would make it unnecessary to model health effects at pollution levels below the range of
Abi Associates, Inc. p. 15] July 3, 1996, Revised
-------
data. None of the following discussion would be changed; the equation as written is retained
for convenience.)
Define
q = rrr= 71 (52)
and
' = TT (53)
The parameter q is just the ratio of PM-2.5 to PM-10. The parameter r may be thought of as
the relative "potency" of coarse and fine particles in causing mortality (or, equivalently, as a
measure of their relative toxicities). If coarse and fine particles are equally potent in causing
mortality (i.e., if particle size doesn't matter), then r=l. If only fine particles matter, in
which case PC = 0, then r=0. Assuming that fine particles cause at least their share of the
mortality associated with PM-10, r will lie somewhere between 0 and 1 .
Equation (2) implies
* *,0 (54)
and therefore also
xc = 0-9> * *,0 (55)
Equation (3) implies also that
Pc = r * p7 . (56)
Substituting equations (4), (5), and (6) back into the model (equation (1)) yields
y = B * eP/-«",o^('-<^,<,> (57)
Abt Associates, Inc. p.152 July 3, 1996, Revised
-------
or, rearranging terms,
= B *
= B *
where
P10 = P/*
The factor [q + r(l - q)] may be thought of as a "scaling" factor that, under the
assumptions stated above, converts the PM-2.5 coefficient to the corresponding PM-10
coefficient.
Even if the relationship between PM-2.5 and mortality (i.e., pf ) is the same
everywhere, then, the relationship between PM-10 and mortality (i.e., (J10) may vary from one
place to another if the ratio of PM-2.5 to PM-10 (q) varies and/or the relative potencies of the
fine and coarse fractions (r) varies.
Suppose, for example, that fif = 0.001 everywhere and that the relative potencies of the
fine to coarse fractions (r) is 0.5 everywhere (i.e., the fine fraction is twice as harmful as the
coarse fraction everywhere). Suppose, however, that in Provo, Utah, only twenty percent of
PM-10 is fine particles whereas in Philadelphia, eighty percent is fine particles. The
coefficient of PM-10 in Provo is then
= 0.001 *[0.20 + 0.5(0.80)] = 0.0006
whereas in Philadelphia it is
= 0.001*[0.80 + 0.5(0.20)] = 0.0009.
Therefore, if the composition of PM-10 and/or the relative potencies of the fine and
coarse fractions of PM-10 varies significantly from one location to another, the PM-10
concentration-response function estimated in one location may not be entirely applicable to a
different location.
The generalized model (equation (1)) allows more specific analysis of alternative policy
options. It also demonstrates that the PM-10 model, which does not distinguish between fine
and coarse particles, may significantly misestimate the health effects associated with a given
concentration of PM-10 if there really is a difference in the potency of PM-2.5 and the coarse
Abt Associates, Inc. p. 153 My 3, 1996, Revised
-------
fraction and if the proportion of PM-10 that is fine particles varies significantly from one
location to another.
Abt Associates, Inc. p. 154 July 3, 1996, Revised
-------
Appendix 5; Adjustment of Means and Standard Deviations of Distributions for
Location-Specific P's
Location-specific estimates of P are adjusted to take into account all the information
about p in all locations for which it has been estimated, assuming the random effects model in
which location-specific P's are regarded as a sample from an underlying distribution of P's. Let
Pi denote the estimate of P in the ith location;
Vj denote the variance of the estimate, ps ;
\\2 denote the variance of the underlying distribution of P's;
Ppooied denote the estimate of the mean of the distribution, derived by pooling the sample
of estimates of the ft's ; and
vpooied denote the variance of the estimate of the mean,
The unadjusted probability distribution describing the probability that the true value of P
in the ith location is within any given interval is a normal distribution with mean equal top( and
variance equal toVj.
The adjusted probability distribution is a normal distribution with mean equal to
+ I/I2)
and variance equal to
[\/(\/v + i/ri2)] + [v .yoiV/n/F + I/iff]
* v I i ' j L. pooiCCf vi'L / i j j
The adjusted mean is a weighted average of the original estimate, PJ, and the pooled
estimate of the mean of the distribution, Ppo0]ed. The larger the variance around the location-
specific estimate, PJ (i.e.. the less certain it is), the less weight it has in the adjusted mean.
The first term in the adjusted variance combines the within study variance (Vj) and the
between study variance, T|2. The second term in the adjusted variance is a correction for the fact
that the mean of the distribution is not known but is only estimated (by Ppo0|ed). This estimate
therefore has some variability associated with it.
Abr Associates, Inc. p. 155 July 3, 1996, Revised
-------
References
AIRS. Aerometric Information Retrieval System [database]. 1995. ResearchTriangle Park,
NC. U.S. Environmental Protection Agency, Office of Air Quality Planning and
Standards.
Abbey, D. E., Lebowitz, M.D., Mills, P.K., Petersen, F.F., Beeson, W.L., Burchette, R.J.
1995. Long-term ambient concentrations of participates and oxidants and development
of chronic disease in a cohort of nonsmoking California residents. In: Phalen, R.F.,
Bates, D.V., eds. Proceedings of the colloquium on paniculate air pollution and
human mortality and morbidity; January 1994; Irvine, CA. Inhalation Toxicol. 7: 19-
34.
American Lung Association. 1995. Dollars and Cents: The economic andhealth benefits of
potenetial paniculate matter reductions in the United States.
Cifuentes, L., Lave, L. 1996. Association of daily mortality and air pollution in
Philadelphia, 1983-1988. /. Air Waste Management Assoc. In press.
Cox, D., and Hinkley, D. 1974. Theoretical Statistics. Chapman and Hall, London.
DerSimonian, R. And N. Laird. 1986. Meta-analysis in clinical trials. Controlled Clinical
Trials 7:177-188.
Dockery, D.W., F.E. Speizer, D.O. Stram, J.H. Ware, J.D. Spengler, and E.G. Ferris, Jr.
1989. Effects of inhalable particles on respiratory health of children. Am. Rev.
Respir. Dis. 139: 587-594.
Dockery, D.W., J. Schwartz, and J.D. Spengler. 1992. Air pollution and daily mortality:
associations with particulates and acid aerosols. Environ Res. 59: 362-373.
Dockery, D.W., et al. 1993. An association between air pollution and mortality in six U.S.
cities. The New England Journal of Medicine. 329 (24): 1753-1759.
Dockery, D.W. et al. 1996. Health effects of acid aerosols on North American children:
respiratory symptoms. Env. Health Persp. Submitted.
Efron, B. and Morris, C. 1971. Limiting the risk of Bayes and empirical Bayes estimators.
Part I: The Bayes case. J Am. Stat. Assoc. 66: 807-815.
Efron, B. and Morris, C. 1972a. Emprical Bayes on vector observations: An extension of
Stein's method. Biometrika 59: 335-347.
Abt Associates, Inc. p. 156 July 3, 1996, Revised
-------
Efron, B. and Morris, C. 1972b. Limiting the risk of Bayes and empirical Bayes estimators.
Part II: The empirical Bayes case. J Am. Stat. Assoc. 67: 130-139.
Efron, B. and Morris, C. 1973. Stein's estimation rule and its ccompetitors - An empirical
Bayes approach. J Am. Stat. Assoc. 68: 117-130.
Empire State Electric Energy Research Corporation (ESEERCO). 1994. New York State
Environmental Externalities Cost Study. Report 2: Methodology. Prepared by:
RCG/Hagler, Bailly, Inc., November.
Graves, E.J. 1994. Detailed Diagnoses and Procedures National Hospital Discharge Survey
1992. NCHS. Vital Health Statistics Series 13, No. 118. August. Figures do not
include newborns.
Health Effects Institute. 1995. Paniculate air pollution and daily mortality: replication and
validation of selected studies. The phase I report of the Particle Epidemiology
Evaluation Project.
Ito, K., Kinney, P., and G.D. Thurston. 1995. Variations hi PMi0 concentrations within two
metropolitan areas and their implications to health effects analyses. Inhalation
Toxicol., 7:735-745.
Ito, K., and Thurston, G. 1996. Daily PM-10/Mortality Associations: An Investigation of
At-Risk Sub-populations. J. Exposure Anal, and Environ. Epidemiol., to appear.
Kinney, P.L., Ito, K. and G.D. Thurston. 1995. A sensitivity analysis of mortality/PM,0
associations in Los Angeles. Inhalation Toxicology 7: 59-69.
Krupnick AJ, Harrington W, Ostro B. 1990. Ambient Ozone and Acute Health Effects:
Evidence from Daily Data. Journal of Environmental Economics and Management
18:1-18.
Mood, A., Graybill, F., and Boes, D. 1974. Introduction to the Theory of Statistics.
McGraw Hill, New York.
Moolgavkar, S.H., and E. G. Luebeck. 1996. Paniculate Air Pollution and Mortality: A
Critical Review of the Evidence. Epidemiology (in press).
Moolgavkar, S.H., Luebeck, E.G., Hall, T.A., Anderson E.L. 1995. Air Pollution and
Daily Mortality in Philadelphia. Epidemiology 6:476-484.
Ostro, B.D. 1987. Air pollution and morbidity revisited: a specification test. /. Eniron.
Econ. Manage. 14: 87-98.
Abt Associates, Inc. p. 157 July 3, 1996, Revised
-------
Ostro, B.D., Lipsett, M.J., Wiener, M.B., and J.C. Seiner. 1991. Asthmatic responses to
airborne acid aerosols. American Journal of Public Health 81: 694-702.
Ostro, B.D., Lipsett M., Mann J., Braxton-Owens H., and White M.. 1995. Air pollution
and Asthma Exacerbations Among African-American Children in Los Angeles. In:
Phalen, R.F., Bates, D. V., eds. Proceedings of the colloquium on paniculate air
pollution and human mortality and morbidity, part II; January, 1994; Irvine, CA.
Inhalation Toxicol. 7:711-722.
Pope, C.A., ffl. 1991. Respiratory hospital admissions associated with PM10 pollution in
Utah, Salt Lake, and Cache Valleys. Arch. Environ. Health 46: 90-97.
Pope, C.A., HI. 1994. Paniculate Pollution and Mortality hi Utah Valley. Prepared for:
Critical Evaluation Workshop on Paniculate Matter - Mortality Epidemiology Studies;
Raleigh, North Carolina, November 9-11, 1994.
Pope, C.A., ffl, D.W. Dockery, J.D. Spengler, and M.E. Raizenne. 1991. Respiratory
health and PMj0 pollution: a daily time series analysis. Am. Rev. Respir. Dis. 144:
668-674.
Pope, C.A., III, Schwartz, J., and M.R. Ransom. 1992. Daily mortality and PM10 in Utah
valley. Arch. Environ. Health 47: 211-217.
Pope, C.A., ffl, M.J. Thun, M.M. Namboodiri, D.W. Dockery, J.S. Evans, F.E. Speizer,
and C.W. Heath, Jr. 1995. Paniculate air pollution as a predictor of mortality in a
prospective study of U.S. adults. Am. J. Respir. Crit. Care Med. 151: 669-674.
Raizenne, M., Neas, L., Damokosh, A., Dockery, D., Spengler, J., Koutrakis, P., Ware, J.,
Speizer, F. 1996. Health effects of acid aerosols on North American children:
pulmonary function. Environ. Health Perspect. Accepted.
Ransom, M. and Pope, C.A. III. 1992. Elementary school absences and PM-10 Pollution in
Utah Valley. Environmental Research 58: 204-219.
Samet, J. And Spengler, J. Prevention of Respiratory Diseases from Indoor and Outdoor Air
Pollution. InHirsch, A. et al., ed. 1993. Prevention of Respiratory Diseases.
Marcel Dekker, Inc. New York.
Science Applications International (ICF Kaiser). 1995. Statistical Support for the PM
NAAQS: Interim Report. SYF APP-95/039.
Schwartz, J. 1991. Paniculate air pollution and daily mortality in Detroit. Environ. Res.
56:204-213.
Abt Associates, Inc. p. 158 July 3, 1996, Revise*
-------
Schwartz, J. 1991/1992. Participate air pollution and daily mortality: a synthesis. Public
Health Rev. 19: 39-60.
Schwartz, J. 1993. Air pollution and daily mortality in Birmingham, Alabama. Am. J.
Epidemiol. 137: 1136-1147.
Schwartz, J. 1994a. Air Pollution and Hospital Admissions in Elderly Patients in
Birmingham, Alabama. American Journal of Epidemiology 139:589-98.
Schwartz, J. 1994b. Air pollution and hospital admissions for the elderly in Detroit,
Michigan. Am. J. Respir. Crit. CareMed. 150:648-55.
Schwartz, J. 1994c. PM,0, Ozone and Hospital Admissions for the Elderly in Minneapolis-St.
Paul, Minnesota. Archives of Environmental Health 49(5): 366-374.
Schwartz, J. 1995. Short term fluctuations in air pollution and hospital admissions of the
elderly for respiratory disease. Thorax 50(5): 531-538.
Schwartz, J. 1996. Air Pollution and Hospital Admissions for respiratory disease.
Epidemiology 7(1): 1-9.
Schwartz, J., D., Dockery, D.W., Neas, L.M., Wypij, D., Ware, J.H., Spengler, J.D.,
Koutrakis, P., Speizer, F.E., Ferris, B.G., Jr. 1994. Acute Effects of Summer Air
Pollution on Respiratory Symptom Reporting in Children. Am. J. Respir. Crit. Care
Med. 150: 648-655.
Schwartz, J., and D.W. Dockery. 1992a. Paniculate air pollution and daily mortality in
Steubenville, Ohio. Am. J. Epidemiol 135: 12-19.
Schwartz, J., and D.W. Dockery. 1992b. Increased mortality in Philadelphia associated with
daily air pollution concentrations. Am. Rev. Respir. Dis. 145: 600-604.
Schwartz, J., and R. Morris. 1995. Air pollution and cardiovascular hospital admissions.
Am. J. Epidemiol., 142: 23-35.
Schwartz, J., Dockery, D., and L. Neas. 1996a. Is Daily Mortality Specifically Associated
With Fine Particles? J. Air & Waste Man. Assoc. (Accepted).
Schwartz, J., Dockery, D., Ware, J., Spengler, J., Wypij, D., Koutrakis, P., Speizer, F., and
B. Ferris., Jr. 1989. Acute effects of acid aerosols on respiratory symptom reporting
in children. USEPA, Boston, MA. For presentation at the 82nd Annual Meeting and
Exhibition. Anaheim, CA. pp. 1-10. June 25-30.
Abt Associates, Inc. p. 159 " July 3, 1996, Revised
-------
Schwartz, J., Dockery, D., Neas, L., Wypij, D., Ware, J., Spengler, J., Koutrakis, P.,
Speizer, F., and B. Ferris., Jr. 1994. Acute effects of summer air pollution on
respiratory symptom reporting in children. Am. J. Respir. Crit. Care Med. 150: 1234-
1242.
Schwartz, J., Slater, D., Larson, T., Pierson, T. and J. Koenig. 1993. Paniculate Air
Pollution and Hospital Emergency Room Visits for Asthma in Seattle. Am Rev. Respir
Dis 147: 826-831.
Spix, C., Heinrich, J., Dockery, D., Schwartz, J., Volksch, G., Schwinkowski, K., Collen,
C., Wichmann, H. 1993. Air pollution and daily mortality in Erfurt, East Germany,
1980-1989. Env. Health Perspect. 101:518-526.
Styer, P., McMillan, N., Gao, F., Davis, J., and J. Sacks. 1994. The effect of airborne
paniculate matter on daily death counts. Environ. Health Perspect. 103: 490-197.
Thurston, G. Ito, K., Hayes, C., Bates, D. and M. Lippmann. 1994. Respiratory Hospital
Admission and Summertime Haze Air Pollution in Toronto, Ontario: Consideration of
the Role of Acid Aerosols. Environmental Research 65: 271-290.
Trijonis, J. C., et al. 1990. Visibility: Existing and Historical Conditions -- Causes and
Effects. NAPAP Report 24.
U.S. Bureau of the Census (USBC). 1990. Statistical Abstract of the United States: 1990.
110th Edition. Washington D. C. U.S. Department of Commerce.
U.S. Bureau of the Census. 1992. Statistical Abstract of the United States (112th edition).
Washington, DC. U.S. Department of Commerce.
U.S. Department of Commerce. 1994. County City Data Book.
U.S. Environmental Protection Agency (USEPA). 1986. Review of the National Ambient Air
Quality Standards for Paniculate Matte: Updated Assessment of Scientific and
Technical Information Addendum to the 1982 OAQPS Staff Paper. Prepared by the
Office of Air Quality Planning and Standards, Research Triangle Park, North Carolina.
EPA 450/05-86-012.
U.S. Environmental Protection Agency (U.S. EPA). 1994. Review of the National Ambient
Air Quality Standards for Sulfur Oxides: Assessment of Scientific and Technical
Information. Supplement to the 1986 OAQPS Staff Paper Addendum. Air Quality
Management Division, Office of Air Quality Planning and Standards. Research
Triangle Park, NC. EPA Report No. EPA-452/R-94-013.
Abi Associates, Inc. p. 160 July 3, 1996, Revised
-------
U.S. Environmental Protection Agency (U.S. EPA). 1996a. Air Quality Criteria for
Paniculate Matter. Office of Health and Environmental Assessment, Office of
Research and Development. Research Triangle Park, NC.
U.S. Environmental Protection Agency (U.S. EPA). 1996b. Review of the National Ambient
Air Quality Standards for Paniculate Matter: Policy Assessment of Scientific and
Technical Information. OAQPS Staff Paper. Office of Air Quality Planning and
Standards. Research Triangle Park, NC.
U.S. Environmental Protection Agency (U.S. EPA). 1996c. Review of National Ambient Air
Quality Standards for Ozone: Assessment of Scientific and Technical Information. Air
Quality Management Division, Office of Air Quality Planning and Standards. Research
Triangle Park, NC.
Utell, M., Frampton, M. 1995. Particles and mortality: a clinical perspective. In: Phalen,
R. F., Bates, D. V., eds. Proceedings of the colloquium on paniculate air pollution
and human mortality and morbidity, part II; January 1994; Irvine, CA. Inhalation
Toxicol. 7: 645-655.
Whitfield, R. G. et al. 1996. A Probabilistic Assessment of Health Risks Associated With
Shon-Term Exposure to Tropospheric Ozone. Report prepared for U.S. EPA Office
of Air Quality Planning and Standards. Research Triangle Park, NC.
Wood, Elaine. 1995. National Center for Health Statistics. Personal communication with
Matthew Wiener, Abt Associates. October 31.
Abt Associates, Inc. p. 161 July 3, 1996, Revised
------- |