Full text of DOI:10.1080/10473289.2000.10464185

Journal of the Air & Waste Management Association
ISSN: 1047-3289 (Print) (Online) Journal homepage: http://www.tandfonline.com/loi/uawm20
Daily Mortality in the Philadelphia Metropolitan
Area and Size-Classified Particulate Matter
Frederick W. Lipfert , Samuel C. Morris & Ronald E. Wyzga
To cite this article: Frederick W. Lipfert , Samuel C. Morris & Ronald E. Wyzga (2000)
Daily Mortality in the Philadelphia Metropolitan Area and Size-Classified Particulate
Matter, Journal of the Air & Waste Management Association, 50:8, 1501-1513, DOI:
10.1080/10473289.2000.10464185
To link to this article: http://dx.doi.org/10.1080/10473289.2000.10464185
Published online: 27 Dec 2011.
Submit your article to this journal
Article views: 76
View related articles
Citing articles: 20 View citing articles
Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=uawm20
Download by: [109.163.234.2]
Date: 12 June 2016, At: 05:58

TECHNICAL PAPER
ISSN 1047-3289 LJ. iApirfe&rWt,aMsteoMrarniasg,e.aAnssdoc.W50y:1z5g01a-1513
Copyright 2000 Air & Waste Management Association
Daily Mortality in the Philadelphia Metropolitan Area and
Size-Classified Particulate Matter
Frederick W. Lipfert and Samuel C. Morris
Environmental Consultants, Northport, New York
Ronald E. Wyzga
EPRI, Palo Alto, California
ABSTRACT
gaseous pollutants and mutually exclusive components
of PM (PM2.5 and coarse particles, SO42– and non-SO42– por-
tions of total suspended particulate [TSP] and PM10), mea-
sured on the day of death and the previous day.
Time-series of daily mortality data from May 1992 to Sep-
tember 1995 for various portions of the seven-county
Philadelphia, PA, metropolitan area were analyzed in re-
lation to weather and a variety of ambient air quality pa-
rameters. The air quality data included measurements of
size-classified PM, SO42–, and H+ that had been collected
by the Harvard School of Public Health, as well as routine
air pollution monitoring data. Because the various pol-
lutants of interest were measured at different locations
within the metropolitan area, it was necessary to test for
spatial sensitivity by comparing results for different com-
binations of locations. Estimates are presented for single
pollutants and for multiple-pollutant models, including
We concluded that associations between air quality and
mortality were not limited to data collected in the same
part of the metropolitan area; that is, mortality for one
part may be associated with air quality data from another,
not necessarily neighboring, part. Significant associations
were found for a wide variety of gaseous and particulate
pollutants, especially for peak O3. Using joint regressions
on peak O3 with various other pollutants, we found that
the combined responses were insensitive to the specific
other pollutant selected. We saw no systematic differences
according to particle size or chemistry. In general, the as-
sociations between daily mortality and air pollution de-
pended on the pollutant or the PM metric, the type of
collection filter used, and the location of sampling. Al-
though peak O3 seemed to exhibit the most consistent
mortality responses, this finding should be confirmed by
analyzing separate seasons and other time periods.
IMPLICATIONS
The most immediate implications of this study are the
apparent variability of the PM results and the relative ro-
bustness of peak O3 as predictors of daily mortality, not
just in Philadelphia per se but in surrounding counties as
well. Still, it is likely that other pollutants may also be im-
portant, and those species that drive the response of the
nephelometer in Camden, NJ, may be among them. This
study does not resolve the ongoing debate about which
pollutants are the best predictors of daily health events in
Philadelphia or whether any of the associations may be
causal. However, it does show that a wide range of can-
didate relationships between daily mortality and ambient
air quality is possible. This finding strongly cautions against
accepting relationships based on only a single pollutant,
and it demonstrates the difficulty in selecting the “cor-
rect” relationship(s) from alternative multiple-pollutant re-
gressions. Further, the available air quality measurements
should be regarded primarily as indices of the overall ur-
ban pollutant mix, and it must be recognized that they
constitute a limited sample of all of the various constitu-
ents that might be harmful. A possible source of further
ambiguity in previous studies may be the coarse resolu-
tion (10 ppb) of the gaseous measurements that were ini-
tially taken in Philadelphia.
INTRODUCTION
Previous Studies of Philadelphia Air Quality
and Health
Philadelphia, PA, has been the site of many epidemio-
logic and air quality studies, beginning with the doctoral
thesis of Wyzga,1,2 which examined time-series mortality
relationships from 1957 to 1966 using a distributed lag
analysis by season. He found a strong association with
coefficient of haze (COH) that varied by year and was
slightly stronger in summer.1 He also found winter rela-
tionships with other pollutants (one at a time)2 and con-
cluded that most of the response was “immediate,” with
the largest mortality responses estimated for COH, NO,
total hydrocarbons, and total suspended particulate (TSP),
in decreasing order of importance. Most other subsequent
Volume 50 August 2000
Journal of the Air & Waste Management Association 1501

Lipfert, Morris, and Wyzga
studies considered particulate (PM) pollutants singly and
jointly in various combinations.3-24
DATA AND METHODS
This project was intended to explore geographical as well
as temporal variability, covering the city of Philadelphia
(PHIL), three nearby suburban Pennsylvania counties (PA),
and three nearby New Jersey (NJ) counties, for 1991–1995.
These seven counties vary greatly in terms of overall popu-
lation and in percentages of elderly, minorities, and the
poor. The daily data used in this study include death
counts, weather, and air quality data from the same ar-
eas. The inclusion of suburban areas is consistent with
the emphasis on secondary pollutants such as PM2.5 and
O3, thought to be relatively uniform across large regions.26
Ambient air quality has also been studied extensively
in Philadelphia. Suh et al. examined the spatial variabil-
ity of fine and coarse PM fractions, SO42–, and H+, and
concluded that local population density was an impor-
tant factor for coarse particles and acidity levels.25 Suh et
al. then found that daily PM2.5 in Philadelphia was highly
correlated with daily PM2.5 in Washington, DC (some 200
km to the south), on the same day (with no allowance for
air mass travel time).26 Liu examined the spatial distribu-
tion of O3 in the metropolitan area and concluded that
the more outlying areas received higher O3 concentra-
tions.27 Olmez et al. measured elemental contributions to
fine and coarse PM fractions for major point source types
in Philadelphia and found substantial primary SO42– emis-
sions from oil-fired combustion.28
Mortality Data
NJ death counts were supplied directly by the State Bu-
reau of Vital Statistics. PA mortality data were extracted
from about 172,000 individual records for PA and PHIL
to provide the necessary daily counts. Deaths due to acci-
dents, suicides, and homicides were removed from all
mortality variables. The average daily death counts are
presented in Table 1.
Objectives and Organization of the Paper
Given the extensive background of air quality and epide-
miologic studies in Philadelphia, we sought to extend the
mortality analysis to include more detailed PM specifica-
tions and data from nearby suburbs by integrating data
of different types from several agencies. The geographic
extension was required to make use of specific air quality
measurements that are only available for some locations.
The ultimate objective of the analysis is to estimate the
associations between mortality and air pollution for the
entire metropolitan area and to examine the results for
consistency and plausibility. A special emphasis is given
to PM collected in different ways, representing various
particle size distributions and chemistry. Previous studies
in Philadelphia have examined a few air quality variables
using different models and methods. In this paper, we
use a single regression methodology that is simple, well
known, and transparent, to examine a large number of
pollutants and mortality variables. In doing so, we seek
to avoid singling out specific combinations that might be
of regulatory importance in favor of examining the en-
semble of candidate relationships as a whole.
The paper begins with a description of the various
data sets and the simple regression techniques that were
used. Phase I of the regression analysis is concerned with
the applicability of air quality data taken in one part of
the region to the prediction of daily mortality in other
parts. We then select the best predictors and use them to
study relationships with daily mortality summed over
the entire seven-county region (Phase II). Finally, we con-
sider numerous alternative PM metrics as predictors of
regional cardiovascular mortality (Phase III). The con-
clusions are intended to provide guidance for the inter-
pretation of daily time-series studies and to identify needs
for further research.
Weather and Other Data
Daily weather data were obtained from Philadelphia In-
ternational Airport and from selected air quality moni-
toring stations for temperature, dew point, pressure, and
wind speed. Daily maximum temperature ranged from
–22 to 36.3 °C for this period. Day-of-week, trend, and
“hot” and “cold” dummy variables were also included in
the dataset.
Air Quality Data
Air quality data were obtained from four different sources:
the Harvard School of Public Health (PM2.5, coarse par-
ticles [CP] = PM10 – PM2.5, PM10, SO42–, and H+ for PHIL),
the city of Philadelphia (gases, PM, and metals), the state
of Pennsylvania (gases and PM), and the state of New Jer-
sey (gases, PM including nephelometry [light scattering],
and COH, for Camden, NJ). All of the PM data except the
Harvard data and data from one station in PHIL were
sampled on a 6th-day schedule. The PA data were avail-
able for 2–4 stations, which were averaged for each day
with at least two stations reporting data. The PA stations
included Bristol (just northeast of the city along the Dela-
ware River), Chester (just southwest of the city along the
river), and Norristown (about 20 km northwest of the city).
Camden is about 4 km southeast of the city.
The PHIL data for gases were only available in units
of 10 ppb; because of this lack of precision, these data
were not used in the analysis. The PHIL PM data were
limited to two stations: the AFS station at Aramingo Av-
enue and Belgrade Street (east side of the city), sampled
every day, and the LAB station at 1501 East Lycoming
1502 Journal of the Air & Waste Management Association
Volume 50 August 2000

Lipfert, Morris, and Wyzga
Table 1. Average daily death counts.
differential variables include “super-coarse”
particles (TSP–PM10) and non-SO42– particles
(various PM measures less SO42– as (NH4)2SO4).
SO42– ion data from the Harvard sites were strati-
fied into the two main cations, H+ and NH4+,
using the charge balance equation. Table 2 pre-
sents the salient statistics of the main air qual-
ity variables used.
Philadelphia
49.63
All Pennsylvania (including Philadelphia)
age <65
23.25
Suburban PA Counties
Bucks
age 65+
70.93
8.55
11.31
14.38
respiratory (ICD 460-519)
cardiac (ICD 390-448)
other
Delaware
41.63
42.88
Montgomery 18.85
Correlations among selected pollutants
Total Pennsylvania
94.17
Camden Subsets
and sampling locations were considered on the
basis of the filtered data to remove the effects
of common seasonal trends. For the gaseous
pollutants, the PA and Camden data were highly
correlated, with coefficients ranging from 0.88
for O3 to 0.78 for CO. Among the different SO42–
measures considered, the best correlation was
between PA and Camden SO42– data (R = 0.72).
SO42– data from the city of Philadelphia were
poorly correlated with other stations (R < 0.4),
age <65
2.92
8.69
1.04
4.86
3.15
2.56
age 65+
respiratory (ICD 460-519)
cardiac (ICD 390-448)
cancer
NJ
Camden
11.61
other counties 14.39
other
Regional Total
120.2
(7 counties)
Street, which had data on PM10 and a few size-classified
samples from 1991, and was sampled every 6 days.
The Harvard PM2.5 sampling began in May 1992 at 59th
and Greenwood Streets (west side of the city). In March
1993, the site was moved to 61st and Elwood, 0.8 km away.
CP concentrations were obtained by subtracting PM2.5 from
PM10, which led to a few negative values for CP; these val-
ues were not used in the regression analyses.
and the Harvard SO42– data were only moderately corre-
lated with PA and Camden data (R = 0.55). Among fine
and inhalable particle measures, COH, nephelometry, and
PM2.5 were moderately correlated (R = 0.5–0.7), as were
intercorrelations among PM10 and TSP sites. The levels of
these correlations suggest the presence of local sources of
various types of particles, including SO42– 28
additional filter artifacts.29
,
or perhaps
The Harvard SO42– and H+ data were obtained from up
to 8 stations across the city that sampled at schedules rang-
ing from hourly to every 6 days; these data were simply
averaged by date. Because we obtained these supplemen-
tary data late in the project, they were only used in Phase
III of the analysis. Other SO42– data were available from
Camden (Teflon PM10 filters), from PA sites (glass-fiber TSP
filters), and from both PHIL sites (glass-fiber TSP filters).
Twenty-nine air pollutants were considered in Phase
I, including all criteria pollutants (except Pb), SO42– (3 sta-
tions), COH and light scattering at Camden, PM2.5, CP =
Regression Methods
Filters were applied to all continuous variables, using the
19-day weighted average Shumway filter for daily data.30
(Filters of the same shape but with lengths of 9, 39, and
79 days were investigated as alternatives. Although the
longer filters induced less serial correlation and fit the
data slightly better for some pollutants (TSP), the fit was
worse for others (O3), with only about 3% change in the
total response. However, longer filters resulted in the loss
of more observations due to missing data, because of the
need to delete more data at the edges of the data gaps. We
thus judged the 19-day filter to be appropriate for this
data set.) For PM data sampled on a 6-day schedule, a
similar weighted filter was devised that included 5 con-
secutive sampling periods (24 days). The weights were
0.075, 0.25, 0.35, 0.25, and 0.075. Filtering has the ad-
vantage of eliminating long-term cycles and trends, with
the disadvantage of inducing small amounts of serial cor-
relation (autocorrelation coefficients around –0.1). How-
ever, it has been shown that such serial correlation has
no effect on the response estimates, whereas incomplete
control of season can have major effects.31
PM10 – PM2.5, Fe, Mn, and Zn (but not the Harvard SO42–
/
H+ data). These metals were selected to represent transi-
tion metals that occur in ambient air at concentrations
high enough to expect reasonable precision. Concentra-
tions of each pollutant were averaged for both the day of
mortality and the previous day. Such averaging was not
possible for the PM species that were collected on a 6-day
schedule; in these cases, regressions were run separately
for lags 0 and 1 and the coefficients were averaged. Hence,
with 6-day data, lags 0 and 1 represent different data sets.
The best predictors were selected from this set for use
in the Phase II regressions. In Phase III, PM variables were
examined more closely, in part by defining “differential”
PM variables by subtracting relevant species, just as CP is
obtained as the difference between PM10 and PM2.5. These
Appropriate deletions in the filtered series were made
at the beginnings and ends of long periods (>7 days) of
missing data; no missing data were imputed. Appropriate
Volume 50 August 2000
Journal of the Air & Waste Management Association 1503

Lipfert, Morris, and Wyzga
Table 2. Statistics of the Phase I and Phase II air quality variables.
Species
Location
Time (hr)
Units
Mean
Std. Dev.
Minimum
Maximum
n
CO
Camden
Camden
Camden
Camden
Camden
Camden
Camden
Camden
Camden
Camden
PA
24
24
24
24
24
24
24
24
24
24
24
1
ppb
µg/m3
ppb
0.75
23.36
20.63
23.96
21.62
0.28
0.40
11.28
31.80
8.91
0.10
4
3.8
77
1787
1050
1801
1799
1798
1770
1795
1041
302
Light Scatter
NO
1.00
3.00
1.00
0.10
1.00
2
273
NO2
O3
COH
SO2
Temperature
ppb
78.0
74.0
1.36
51.0
91
ppb
13.57
0.18
COHs
ppb
8.64
6.67
°F
55.58
5.57
17.67
4.46
2–
SO4
PM10
O3
µg/m3
µg/m3
ppb
0.86
8.0
30.6
72.0
73.0
130.7
80.7
301.3
500.3
132.7
38.0
77.0
33.3
94.5
96.0
7.8
26.97
23.44
44.76
20.42
42.30
108.05
38.49
8.20
12.89
13.86
25.68
8.16
302
1.0
1807
1822
1814
1809
1823
1823
1793
1823
299
O3
PA
ppb
1.7
NO2
NOx
NOx
NO2
SO2
SO22–
SO4
PM10
TSP
CO
PA
24
24
1
ppb
4.3
PA
ppb
36.33
88.30
12.12
5.24
7.0
PA
ppb
9.3
PA
1
ppb
7.3
PA
24
24
24
24
24
1
ppb
1.0
PA
ppb
18.26
10.37
27.00
41.73
1.44
10.43
4.39
2.3
PA
µg/m3
µg/m3
µg/m3
ppm
3.5
PA
13.29
16.59
1.04
6.5
298
PA
16.0
0.0
271
PA
1729
1727
1720
1720
1036
1630
1720
287
CO
PA
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
ppm
0.63
0.40
0.1
3.3
Fe
Phil AFS
Phil AFS
Phil AFS
Phil AFS
Phil AFS
Phil Lab
Phil Lab
Phil Lab
Phil (HSPH)
Phil (HSPH)
Phil (HSPH)
Phil (HSPH-8)
Phil (HSPH-8)
Phil (HSPH-8)
Phil (various)
0.1µg/m3
ng/m3
µg/m3
µg/m3
ng/m3
µ/m3
9.19
4.43
0.0
47.0
196
Mn
21.86
17.55
58.70
26.53
16.72
48.52
32.20
17.28
24.14
6.80
13.95
5.22
2.0
2–
SO4
TSP
Zn
2.2
40.8
184
23.73
32.24
7.49
10.0
0.0
398
2–
SO4
0.0
53.7
116
TSP
PM10
PM2.5
PM10
CP
µ/m3
µ/m3
21.59
16.30
9.62
13.0
7.0
267
95.0
72.6
82.4
28.3
105.5
225.2
421.3
1154.28
273
µg/m3
µg/m3
µg/m3
nmol/m3
nmol/m3
neq/m3
µg/m3*nmol/m3
–0.6
–1.6
–20.0
–6.9
4.9
1183
1196
1181
310
11.59
4.77
H+
8.04
13.80
36.56
63.98
132.37
2–
SO4+
53.07
97.98
68.46
310
NH4
9.7
310
Fe × H+
–55.1
295
deletions of the filtered data were made for missing samples.
All statistical computations were made using SPSS.32
The weather model was developed by running
stepwise regressions for each mortality variable, in which
significant variables were selected from variables for fil-
tered temperature (individual lags up to 5, average of lags
0–1 and of lags 0–5); dummy variables for “hot” and “cold”
days, lagged up to 3 days based on both wet and dry bulb
temperatures; change in barometric pressure, lagged up
to 3 days; and dummy variables for days of the week. The
cold dummy variable was taken at the 5th percentile of
temperature (–5.6 °C), but only the top 0.5% of days
exhibited excess heat mortality. Defining “hot” at >33 °C
(91 °F) resulted in significance, while using the 95th per-
centile (28.9 °C) did not.
The final Phase I model selected for use with each
mortality variable included all of the day-of-week vari-
ables and each weather variable that was significant for
any mortality variable in the stepwise models (9 weather
variables and 6 day-of-week dummies, in all). This resulted
in the inclusion of many non-significant variables in each
model, but the regression performance statistics (adjusted
R2 and standard error of estimate) were degraded by only a
few percent from the optimum models. Because filtering
1504 Journal of the Air & Waste Management Association
Volume 50 August 2000

Lipfert, Morris, and Wyzga
had removed the main causes of mortality variation, R2s
were typically quite low, of the order of 0.01. Phase II and
III models were more parsimonious, including only the
weather variables significant for the mortality variable
being considered.
These selections were intended to cover a range of ages
and causes of death, socioeconomic statuses, and locations
with respect to air quality monitoring stations. The working
hypothesis was that statistically significant responses for the
elderly, and for causes of death for which some causal mecha-
nisms had been suggested, would occur when pollution and
mortality are measured in the same area. The combination
of 12 mortality variables and 29 pollutants yields 348 re-
gression results, of which 17 or 18 would be expected to be
significant due to chance, assuming independence.
These results are presented in Table 3 as a matrix of
single-pollutant risk estimates based on mean pollutant
concentrations, to facilitate comparisons with previous
studies.5,6 We also present average risk estimates based on
(mean-background) concentrations in the right-hand col-
umn. The table is structured to facilitate the assessment of
geographic concordance between pollutant and mortality
source regions. There are 80 significant cells overall, or about
23% of the total, but they are not concentrated among the
“diagonal” cells of geographic concordance (mortality and
air quality data from the same sub-region, enclosed in
double lines in the table). No significant combinations of
NJ mortality and NJ pollution resulted, and only 3 of 9
possibilities were significant for Philadelphia (the lack of
suitable gaseous pollutant data for the city may have been
partly responsible for this outcome).
There were 27 significant combinations among the 96
diagonal cells for the 4 PA counties, for a total of 30 (23%)
of the 129 diagonal cells in Table 3. However, 20% of the
remaining, or “off-diagonal,” cells were also significant, so
that there appeared to be no significant advantage (chi-
square = 0.5) for measuring air quality and counting deaths
in the same area. This conclusion held for both secondary
(O3 and PM2.5) and primary (CO) pollutants.
Table 3 also allows convenient averaging over rows
and columns, as a means of further analysis of the en-
semble of results. For example, the pollutant with the high-
est mortality response based on (mean–background)
concentrations, averaged across all endpoints, was peak
O3 from the PA suburban stations. The next two pollut-
ants were Philadelphia Fe and TSP, which is surprising
given the putative local nature of these presumably large-
particle pollutants. The light-scattering data at Camden
also ranked high, but Philadelphia PM2.5 was well down
in the ranking. As judged by the column averages of Table
3, total mortality in Delaware County had the highest
mean air pollution responses by far (for no known rea-
son); respiratory mortality for the 4 PA counties was low-
est (negative), contrary to expectations.
Measures of Risk
Regression results are reported in terms of the fractional
mortality risk attributed to the mean pollutant concen-
tration (i.e., mean effect or elasticity) with identification
of those variables that met the nominal 0.05 significance
criterion. This measure of attributable risk is free from
subjectivity that might derive from the use of arbitrary
pollution increments in expressing relative risks.33 Because
of the wide variations in sample size and in the distribu-
tional properties of the pollutants considered here, sig-
nificance levels, that is, p or t values, were not used in
Phases I or II as primary indicators of relative importance,33
but t values are reported for the Phase III results. In addi-
tion to the mean pollutant levels, mean concentration
less “background,” where background was taken as the
4th percentile of each frequency distribution, was used as
a basis for comparing risks. The 4th percentile level was
selected because the regressions were based on 2-day pol-
lution averages, which are less variable than single days,
and because the resulting background estimates appeared
to be reasonable.
RESULTS
Phase I: Geographic Concordance
In Phase I, the following (filtered) mortality variables were
regressed against a systematic array of pollutants, taken
one at a time:
(1) Camden County (NJ), total (less accidents, etc.);
(2) Camden County (NJ), ages 65+;
(3) Other suburban Philadelphia counties in NJ, ages
65+;
(4) Metropolitan Philadelphia counties in PA, ages
<65;
(5) Metropolitan Philadelphia counties in PA, ages
65+;
(6) Metropolitan Philadelphia counties in PA, respi-
ratory deaths;
(7) Metropolitan Philadelphia counties in PA, car-
diac deaths;
(8) Metropolitan Philadelphia counties in PA, other
deaths (balance);
(9) Bucks County (PA), total (less accidents, etc.);
(10) Delaware County (PA), total (less accidents, etc.);
(11) Montgomery County (PA), total (less accidents,
etc.); and
Phase II: Regional Mortality Associations
The general conclusion from Table 3 was one of a high
degree of randomness, which may occur with mortality
(12) Philadelphia County (and city), total (less acci-
dents, etc.).
Volume 50 August 2000
Journal of the Air & Waste Management Association 1505

Lipfert, Morris, and Wyzga
p
N
1506 Journal of the Air & Waste Management Association
Volume 50 August 2000

Lipfert, Morris, and Wyzga
variables of low average daily count (for which Poisson
variability dominates), from unreliable or locally biased air
quality measurements, or from differences in the real un-
derlying relationships. Accordingly, we confined the Phase
II regressions to aggregated mortality variables (higher
counts) and to the most significant pollutant variables.
These pollutants were PM2.5, CP, and PM10 from the
Harvard site in Philadelphia (these data were the main
motivation for the study); TSP, Fe, and SO42– from the AMS
site in the city (because of the daily observations); COH
and light scattering from Camden (unique measurements);
and O3, CO, SO2, NO2, and NOx from the suburban PA net-
work (averages over several sites). Mortality data comprised
total daily deaths (less accidents, homicides, and suicides)
by geographic area; deaths due to cardiovascular, respira-
tory, and other causes, and for ages less than or greater
than 65 were summed over the seven-county metropoli-
tan region (defined as “regional” mortality). Pollutants were
regressed first individually and then in combination with
peak O3. The results are given in Tables 4 and 5.
Table 4 presents single-pollutant regression results for
successively larger (nested) geographic areas on the left.
We would expect the responses to local Philadelphia pol-
lutants to diminish from left to right, and the responses
to suburban pollutants to increase. However, in general,
these expectations are not realized; Philadelphia TSP, for
example, shows increasing responses with the larger ar-
eas, as does Philadelphia Fe. PM2.5 and CP (not significant)
show roughly constant responses for all areas, as do many
of the other pollutants. Peak O3 (suburban data) shows
the largest (mean–background) responses in all four col-
umns. Regressions were also performed for the data set
restricted by the availability of SO42- data; results were not
greatly different from those for the full data set, so that
the poor performance of this SO42- variable cannot be at-
tributed to the smaller data set, per se.
There are no significant responses for the regional res-
piratory mortality variable; the largest (nonsignificant) re-
sponse is for CP. For cardiovascular and elderly (65+)
mortality, all pollutants are significant except CP and (Phila-
delphia) SO42–. The largest responses are for peak O3, at about
twice the response to PM2.5. The responses for “other” mor-
tality (all non-external causes except respiratory and car-
diovascular) are surprisingly strong and similar to those
for cardiovascular causes. The responses for under-65 deaths
are stronger in many cases than those for the elderly group,
which is another unexpected finding.
Because of the strong showing of peak O3, joint regres-
sions were run with each of the pollutants in Table 4, as
shown in Table 5. For total mortality by geographic area
(top), the combined pollutant responses are quite similar
throughout the entire table (significance was not formally
evaluated for the sums but may be inferred from the sig-
nificance of the components). The largest combined re-
sponses were from light scattering + O3, NO2 + O3, and CP
+ O3 (for the larger geographic areas). The relatively minor
differences among the combined responses of various pol-
lutants with peak O3 apparently resulted from collinearity
Table 4. Single-pollutant results for aggregated groups. Attributable risks for single pollutants (avg. lag, 0,1) based on mean concentrations less background.
Note: Bold italic= p < 0.055.
Volume 50 August 2000
Journal of the Air & Waste Management Association 1507

Lipfert, Morris, and Wyzga
N
1508 Journal of the Air & Waste Management Association
Volume 50 August 2000

Lipfert, Morris, and Wyzga
considerations. For example, for the seven-county region,
the sum of the separate responses of CP and O3 from Table
4 was 0.0429, which increased to 0.0500 in the joint re-
gression. For PM2.5, the sum of separate responses was
0.0432, which remained about the same in the joint re-
gression, making CP + O3 appear to be the superior com-
bination. The implication is that the separate result for
PM2.5 includes some contribution from O3, while the sepa-
rate result for CP does not (there may also be some differ-
ences due to the reduced data set for PM).
The three column averages increased slightly by geo-
graphic area. The average relative shares of O3 were about the
same in the three geographic areas: 74, 71, and 73%. The
bottom section of Table 5 shows comparable joint regression
results by age and cause (respiratory causes were omitted be-
cause of the poor showing in Table 4). The strongest com-
bined responses were for cardiovascular causes; the largest
combined responses were for CP (93% due to O3), PM10, and
light scattering in Camden, followed closely by PM2.5, Fe, and
TSP in Philadelphia. The next highest mortality category was
for those under age 65, for which CP, PM10, PM2.5, NO2, and
light scattering had the strongest responses. Other causes of
death were next, for which NO2 had the strongest combined
responses, followed by CP. SO2, Fe, SO42–, COH, light scatter-
ing, PM2.5, and PM10 constituted a very homogeneous lower-
response group. The lowest mortality group was for ages 65+,
for which CP, NO2, Fe, TSP, light scattering, PM10, COH, and
PM2.5 had the largest responses. Averages by column for the
bottom section of Table 5 show that O3 contributes 79, 62,
68, and 73% (from left to right) of the combined (mean–
background) responses, on average.
Table 6 lists the code names, definitions, mean val-
ues, and other details of these PM variables; all variables
were filtered and lagged as described above. All together,
75 PM variables were screened for relationships with re-
gional mortality by examining bivariate correlations with
the residuals from mortality regressions that included no
pollutants. Then the 50 PM variables with the highest
correlations with the mortality residuals were entered into
regressions for regional cardiovascular mortality one at a
time, and the responses were compared on the basis of
(mean–background) concentration (Table 7). Because
sample sizes (n) varied by an order of magnitude in this
data set, we defined an empirical goodness-of-fit param-
eter as the product of the t value from the regression and
the square root of the inverse ratio of the sample size to
the maximum sample size (1795): t*(1795/n)0.5. This pa-
rameter allows an approximate adjustment to be made
for sample size, to be used when comparing the perfor-
mance of alternative predictors.
The 10 PM variables with the largest mortality responses
were non-SO42– light scattering (lag 1), Harvard SO42– (aver-
age of lags 0 and 1), Philadelphia non-SO42– TSP (average of
lags 0 and 1), Harvard NH4+ (average of lags 0 and 1), PA TSP
(lag 1), PA super-coarse particles (lag 1), light scattering (av-
erage of lags 0 and 1), non-COH light scattering (lag 1), Phila-
delphia TSP (average of lags 0 and 1), and Philadelphia Fe
(average of lags 0 and 1). However, all but two of these esti-
mates are lower than the estimated response for peak O3
(Table 4). Figure 1 was prepared to compare effect size with
(adjusted) significance level as the measure of a variable’s
importance. The figure shows that, for a given significance
(i.e., adjusted t value), coarser particles tend to have larger
mean responses. This results from typical differences in the
frequency distributions of the various PM measures.33 Con-
siderations of particle size were taken a step further in Figure
2, which compares the ranges of estimated responses (based
on mean concentration less background) by estimated maxi-
mum particle size for each PM metric. The figure shows very
large ranges of responses for both the largest and the small-
est particles. (It is interesting that the Harvard SO42– data
occupy the top and the bottom of the range for 1 µm, for
different lag assumptions.) The most compact grouping is
for PM10, which includes data from three different agencies
and locations which are only moderately intercorrelated.
Two conclusions emerge from Figure 2: first, particle
size per se does not appear to be an important param-
eter; second, a wide range of response estimates is pos-
sible for any particle size range. The large range in the
SO42– estimates in Figure 2 deserves special consideration.
Phase III: Detailed Investigation
of PM Responses
The acquisition of the Harvard SO42– and H+ data, together
with the unexpected behavior of the PM variables in Phases
I and II, prompted a more detailed investigation of the con-
stituents of the more aggregated PM measures, such as TSP
and PM10. In addition, the temporal overlap of the H+ and
Fe data facilitated a consideration of the transition metal
hypothesis, in which dissolved Fe may cause inflamma-
tion and oxidative stress in the lung;35 an interaction vari-
able defined as Fe × H+ was used for this purpose. Differential
PM measures were formed through subtraction. For ex-
ample, “super-coarse” particles are defined as the differ-
ence between TSP and PM10, a large part of which may be
filter artifacts. “Non-SO42–” fractions are estimated by sub-
tracting SO42–, converted from ion to mass concentration
by assuming a composition of (NH4)2SO4 on the filter. Part
of the rationale for subtracting SO42– was the assumption
2–
We have SO4 data from four different sources in this
that most of the filter artifact would be reflected as SO42–
,
study; their mean concentrations vary by about a factor
of 3, and they are poorly intercorrelated. This conflicts
with the conventional wisdom that such measurements
and by subtracting this quantity, the remaining mass would
better reflect particles actually breathed.
Volume 50 August 2000
Journal of the Air & Waste Management Association 1509

Lipfert, Morris, and Wyzga
Table 6. PM variables used.
Code Name
Size Filter Chemistry
Formula
Location Agency
# Obs
Overall Mean Bckgrnd.
Units
are usually reliable, but effects of local sources may be a
possible explanation.28
(1) Daily death counts from any part of the met-
ropolitan area may be associated with air
quality in any other part (Table 3). This find-
ing may have two alternative interpretations:
(1) such associations are essentially random oc-
currences, not to be interpreted causally; or (2)
all pollutants have sufficient regionality that
they may serve as surrogates for the true regional
levels (those parameters that do not succeed in
this regard may be measured less reliably than
the others). However, the errors involved in ex-
tending local measurements to a regional con-
text may be small in relation to the differences
between outdoor ambient and personal expo-
sure. Interpretation 1(1) cautions against con-
sidering any of these associations to be causal.
Interpretation 1(2) says that there may be some
Finally, we created the interaction variable Fe × H+ to
test the hypothesis that acid-coated transition metal par-
ticles might be especially toxic.35 However, although Fe
per se was significantly associated with mortality, Fe × H+
had a much lower mean response (0.01). The very spiky
frequency distribution of this parameter resulted in a lower
response than would have been expected from its high t
value.33
CONCLUDING DISCUSSION
We derive four main conclusions from these extended
analyses. Although the statistically based conclusions
(stated in bold) flow directly from the tables and figures
cited, alternative interpretations are possible, as discussed
for each topic.
1510 Journal of the Air & Waste Management Association
Volume 50 August 2000

Lipfert, Morris, and Wyzga
Table 7. Regional cardiovascular mortality risks associated with PM.
.
Attributable Risks
Code Name
Est. Size, µm # Obs.
Mean
Bckgrnd.
df Regress. Coeff. Mean
Mean–Back
t -value Adjusted t
causal relationship(s), but the uncertainties in
exposure estimates preclude a reliable estimate
of which specific pollutants may be to blame.
Moreover, we have no assurance that the most
important environmental factors have been in-
cluded in the analysis.
(2) The estimated responses to fine and/or acidic
particles are no larger than those for other types
of particles (Table 7, Figure 2). This conclusion is
reinforced by considering the differences in esti-
mated exposure reliability between fine and coarse
particles. Fine particles are thought to penetrate
Volume 50 August 2000
Journal of the Air & Waste Management Association 1511

Lipfert, Morris, and Wyzga
Est. Particle
Size, µm
t-value adjusted to max. sample size
Figure 1. Attributable risk vs significance level for regional cardiovascular mortality, by particle size.
structures more efficiently and to remain airborne
longer once indoors. Coarse particles are known
to settle out of the atmosphere more rapidly, giv-
ing rise to local hot spots and spatial non-
uniformity, with attendant large discrepancies
with respect to regional exposure. These consid-
erations also lead to two interpretations. Interpre-
tation 2(1) would say that neither fine nor coarse
particles may be ruled out, but that both may be
surrogates for some other agent(s) that may or may
not be outdoor air pollution. Interpretation 2(2)
would hypothesize that there may be local sources
of fine particles such that the data from the single
site used here are unrepresentative of the region.
If this is the case, such sources should be identi-
fied and characterized as prime candidates for con-
trol. Interpretation 2(3) combines (1) and (2).
(3) When regressed jointly with peak O3, the com-
bined responses are quite similar for many dif-
ferent pollutants and PM metrics (Table 5). This
0.05
SO4, H+, NH4
TSP, NSTSP, Fe
0.04
0.03
0.02
0.01
0
PM2.5, Neph
PM10, CP
COH
–0.01
0.1
1
10
100
Figure 2. Regional cardiovascular mortality risks (based on mean concentration less background) by estimated particle size.
1512 Journal of the Air & Waste Management Association
Volume 50 August 2000

Lipfert, Morris, and Wyzga
3. Schwartz, J.; Dockery, D.W. Amer. Rev. Respir. Dis. 1992, 145(3), 600-604.
4. Schwartz, J. Environ. Res. 1994, 64(1), 26-35.
has been seen with several mortality variables, but
only one of the several alternative O3 variables was
tested in this way (the original objectives of the
study were not especially concerned with O3). In
addition, this finding has not been confirmed with
season-specific analyses. Interpretation 3(1) would
say that since indoor O3 is substantially attenu-
ated with respect to outdoor levels, the resulting
exposure error has shifted some of the association
to the collinear pollutants in the joint regression.
The alternative interpretation 3(2) would turn this
logic around, stating that the measurement errors
in the other pollutants have shifted some of the
association to O3.34 This dichotomy cannot be re-
solved without more detailed knowledge of per-
sonal exposures to various pollutants than now
exists, especially with regard to the most likely
victims of air pollution.
5. Wyzga, R.E.; Lipfert, F.W. Temperature-Pollution Interactions with
Daily Mortality in Philadelphia. In Particulate Matter: Health and Regu-
latory Issues, VIP 49; A&WMA: Pittsburgh, PA, 1995; pp 3-42.
6. Wyzga, R.E.; Lipfert, F.W. Ozone and Daily Mortality: The Ramifica-
tions of Uncertainties and Interactions and Some Initial Regression
Results. In Tropospheric Ozone: Critical Issues in the Regulatory Process;
Vostal, J.J., Ed.; VIP54; Air & Waste Management Association: Pitts-
burgh, PA, 1994; pp 453-463.
7. Cifuentes, L.; Lave, L.B. Association of Daily Mortality and Air Pollu-
tion in Philadelphia, 1983–88; unpublished manuscript, 1995.
8. Li, Y.; Roth, H.D. Inhal. Toxicol. 1995, 7, 45-48.
9. Samet, J.M.; Zeger, S.L.; Berhane, K. Particulate Air Pollution and Daily
Mortality: Replication and Validation of Selected Studies; Health Effects
Institute: Cambridge, MA, 1995.
10. Kelsall, J.E.; Samet, J.M.; Zeger, S.L.; Xu, J. Amer. J. Epidemiol. 1997,
146(9), 750-762.
11. Samet, J.M.; Zeger, S.; Kelsall, J.; Xu, J.; Kalkstein, L. Particulate Air
Pollution and Daily Mortality: Analyses of the Effects of Weather and Mul-
tiple Air Pollutants; Health Effects Institute: Cambridge, MA, 1997.
12. Samet, J.; Zeger, S.; Kelsall, J.; Xu, J.; Kalkstein, L. Environ. Res. 1998,
77, 9-19.
13. Moolgavkar, S.H.; Luebeck, E.G.; Hall, T.A.; Anderson, E.L. Epidemiol.
1995, 6(5), 476-484.
14. Samet, J.M. Epidemiol. 1995, 6(5), 471-473.
15. Dockery, D.W.; Schwartz, J. Epidemiol. 1995, 6(6), 629-632.
16. Dockery, D.W.; Schwartz, J. Epidemiol. 1996, 7(2), 212-214.
17. Zeger, S.L.; Dominici, F.; Samet, J. Epidemiol. 1999, 10(2), 171-175.
18. Neas, L.M.; Schwartz, J.; Dockery, D. Environ. Health Perspect. 1999,
107(8), 629-631.
(4) The associations between daily mortality and
various PM metrics are highly dependent upon
the specific metric and its frequency distribu-
tion, the type of sampling and analysis equip-
ment used, and the location sampled (Table 7,
Figure 2). This finding may reflect either differ-
ences in local subpopulations or, more likely, dif-
ferences in the degrees to which each set of
ambient measurements may serve as a surrogate
for the personal exposures to the agents of harm.
As with all observational epidemiologic studies, these con-
clusions are subject to the limitations inherent in the use
of ambient concentrations as surrogates for actual expo-
sures. Most of these pollutant metrics should be regarded
as indices of the overall urban mixture. Results should
not be extrapolated beyond the ranges of data investi-
gated, including the time period. It is also possible that
these relationships may differ by season.
19. Morris, R.D.; Naumova, E.N.; Munasinghe, R.L. Amer. J. Public Health
1995, 85(10), 1361-1365.
20. Neas, L.M.; Dockery, D.W.; Koutrakis, P.; Speizer, F.E. Epidemiol. 1999,
10(5), 550-553.
21. Dayal, H.H.; Khuder, S.; Sharrar, R.; Trieff, N. Environ. Res. 1994, 65(2),
161-171.
22. Lang, D.M.; Polansky, M. New England J. Med. 1994, 331(23), 1542-1546.
23. Dayal, H.H.; Chiu, C.Y.; Sharrar, R.; Mangan, J.; Rosenwaike, I.;
Shapiro, S.; Henley, A.J.; Goldberg-Alberts, R.; Kinman, J. J. Natl. Can-
cer Inst. 1984, 73(3), 565-574.
24. Silber, J.H. Forecasting Asthmatic Wheezing Using Temperature Ve-
locity; Pediatr. Emerg. Care 1987, 3, 13-17.
25. Suh, H.H.; Allen, G.A.; Koutrakis, P.; Burton, R.M. J. Air & Waste Man-
age. Assoc. 1995, 45, 442-452.
26. Suh, H.H.; Nishioka, Y.; Allen, G.A.; Koutrakis, P.; Burton, R.M. Environ.
Health Perspect. 1997, 105(8), 826-834.
27. Liu, F. J. Air & Waste Manage Assoc. 1996, 46(3), 207-215.
28. Olmez, I.; Sheffield, A.E.; Gordon, G.E.; Houck, J.E.; Pritchett, L.C.;
Cooper, J.A.; Dzubay, T.G.; Bennett, R.L. J. Air Pollut. Control Assoc.
1988, 38(11), 1392-1402.
29. Lipfert, F.W. Filter Artifacts Associated with Particulate Measurements:
Recent Evidence and Effects on Statistical Relationships; Atmos.
Environ. 1994, 28, 3233-3249.
30. Kinney, P.L.; Ozkaynak, H. Associations of Daily Mortality and Air
Pollution in Los Angeles County; Environ. Res. 1991, 54, 99-120.
31. Lipfert, F.W.; Morris, S.C. A Time-Series Study of Daily Hospitaliza-
tion, Mortality, and Air Pollution in Toronto. Presented at PM2000:
Particulate Matter and Health; A&WMA, Charleston, SC, January
2000; Paper No. PM21.
32. Norusis, M.J. SPSS/PC+; SPSS, Inc.: Chicago, IL, 1990.
33. Lipfert, F.W.; Wyzga, R.E. Statistical Considerations in Determining
the Health Significance of Constituents of Airborne Particulate Mat-
ter; J. Air & Waste Manage. Assoc. 1999, 49, PM182-191.
34. Marshall, J.R.; Hastrup, J.L. Mismeasurement and the Resonance of
Strong Confounders: Uncorrelated Errors; Amer. J. Epidemiol. 1996,
143, 1069-1078.
35. Ghio, A.J.; Stonehuerner, J.; Dailey, L.A.; Carter, J.D. Metals Associ-
ated with Both the Water-Soluble and Insoluble Fractions of an Am-
bient Air Pollution Partial Catalyze an Oxidative Stress; Inhalation
Toxicol. 1999, 11, 37-49.
ACKNOWLEDGMENTS
This project was sponsored by EPRI, under WOs 3253 and
6424. We thank the various city and state agencies that
supplied data for the project, and especially Drs. George
Allen and Helen Suh of the Harvard School of Public
Health. The fine particle data were collected with funds
from U.S. Environmental Protection Agency Cooperative
Agreement #CR-822050 with Harvard; Dr. William Wil-
son helped arrange the data transfer. The views expressed
in this paper are our own and should not be attributed to
any of these institutions or agencies. We thank four anony-
mous reviewers for many helpful comments.
About the Authors
Fred Lipfert (corresponding author) and Sam Morris are in-
dependent consultants to industry and government, special-
izing in air quality- and risk analysis-related topics. Ronald
Wyzga manages air quality, health, and risk research at the
Electric Power Research Institute in Palo Alto, CA. Corre-
spondence should be directed to Dr. F.W. Lipfert, 23 Carll
Court, Northport, NY 11768 (email: flipfert@suffolk.lib.ny.us).
REFERENCES
1. Wyzga, R.E. J. Amer. Stat. Assoc. 1978, 73, 463-472.
2. Wyzga, R.E. Urban Air Pollution and Mortality: Ten Years of Philadelphia
Data. In Proceedings of the ASA Social Statistics Section; 1977; pp 660-664.
Volume 50 August 2000
Journal of the Air & Waste Management Association 1513