Ostro et al.
145
conducted a daily mortality time series
study of nine California cities using data from 1999
through 2002. They avoided the use of GAM models by
using Poisson regression models that incorporated natural
or penalized splines to control for time, seasonality, tem-
perature, humidity, and day of week. Random-effects
meta-analysis was used to make pooled estimates. Rela-
tively small but statistically significant PM
2.5
-mortality
associations were observed (see Table 1). Several analyses
have been conducted
146,147
using data from 10 U.S. cities
with daily PM
10
monitoring. Statistically significant
PM
10
-mortality associations were consistently observed,
including a reanalysis
148
using more stringent GAM con-
vergence criteria (see Table 1).
A study evaluated daily mortality and air pollution in
14 U.S. cities
149
using the case-crossover study design
rather than daily time series. The exposure of each mor-
tality case was compared with exposure on a nearby day.
Potential confounding factors, such as seasonal patterns
and other slowly varying covariates, were controlled for
by matching (rather than statistical modeling as in the
time series approach). Statistically significant PM
10
-mor
-
tality associations were observed (Table 1). When the data
were also analyzed using daily time series analysis, for
comparison purposes, estimated PM
10
mortality associa
-
tions were similar.
One of the largest and most ambitious multicity daily
time series studies is the National Morbidity, Mortality,
and Air Pollution Study (NMMAPS). This study grew out
of efforts to replicate several early single-city time series
studies
150
and was designed to address concerns about
city selection bias, publication bias, and influence of co-
pollutants. A succession of analyses included as few as 20
U.S. cities
151,152
and as many as 100 cities.
153–155
Although
the PM-mortality effect estimates were somewhat sensi-
tive to various modeling and city selection choices, there
was “consistent evidence that the levels of fine particulate
matter in the air are associated with the risk of death from
all causes and from cardiovascular and respiratory illness-
es.”
151
Excess risk estimates are presented in Table 1. Be-
cause the NMMAPS analysis included many cities with
substantially different levels of copollutants, the influ-
ence of copollutants could be directly evaluated. The PM-
mortality effect was not attributable to any of the copol-
lutants studied (NO
2
, CO, SO
2
,orO
3
).
A parallel research effort, the Air Pollution and
Health: A European Approach (APHEA) project, examined
the short-term PM-mortality effects in multiple European
cities. Initially, this research effort analyzed daily mortal-
ity data from Յ15 European cities, including 5 from Cen-
tral-Eastern Europe, using a common protocol.
156
Daily
mortality was found to be significantly associated with
PM and sulfur oxide concentrations,
157,158
although the
effect estimates were sensitive to approaches to control-
ling for long-term time trends and seasonality.
159,160
A
continuation and extension of the APHEA project, often
referred to a APHEA-2, included analyses of daily mortal-
ity and pollution data for Յ29 European cities.
161,162
APHEA-2 also found that PM air pollution was signifi-
cantly associated with daily mortality counts (see Table
1). Furthermore, the use of GAMs with strict convergent
criteria or parametric smoothing approaches did not sub-
stantially alter the estimated PM-mortality effects.
162
Sub-
sequent analysis of APHEA-2 data found PM-mortality
effects with both cardiovascular and respiratory mortality
(see Table 1).
163
Mortality associations with PM were also observed for
nine French cities
164
and three Australian cities.
165
Two
Asian multicity studies have reported daily mortality as-
sociations with measures of PM (see Table 1). The first was
a study of seven major Korean cities.
166
Measures of PM
10
or PM
2.5
were not available, and PM was measured only as
TSP. Although it was suggested that SO
2
may have func
-
tioned better as a surrogate for PM
2.5
in Korea’s ambient
air than TSP, mortality associations were observed with
TSP, as well as with SO
2
. The second analyzed data from
the 13 largest Japanese cities
167
with mortality data for the
elderly (aged Ն65 years) and suspended PM (special pur-
pose monitoring, approximately PM
7
; i.e., PM with a 50%
cutoff diameter of ϳ7 m). GAM and generalized linear
models were used (estimated using SAS rather than S plus
software).
Summary and Discussion
It seems unlikely that relatively small elevations in expo-
sure to particulate air pollution over short periods of only
1 or a few days could be responsible for very large in-
creases in death. In fact, these studies of mortality and
short-term daily changes in PM are observing small ef-
fects. For example, assume that a short-term elevation of
PM
2.5
of 10 g/m
3
results in an ϳ1% increase in mortality
(based on the effect estimates summarized in Table 1).
Based on the year 2000 average death rate for the United
States (8.54 deaths/1000 per year), a 50-g/m
3
short-term
increase in PM
2.5
would result in an average of only 1.2
deaths per day in a population of 1 million (compared
with an expected rate of ϳ23.5/day). That is, on any given
day, the number of people dying because of PM exposure
in a population is small.
It is remarkable that these studies of mortality and
short-term changes in PM are capable of observing such
small effects. Uncertainties in estimating such small
effects legitimately create some doubts or concerns re-
garding the validity or accuracy of these estimates. Never-
theless, associations between daily changes in PM concen-
trations and daily mortality counts continue to be
observed in many different cities and, more importantly,
in large multicity studies, which have much less oppor-
tunity for selection or publication bias. The estimated size
of these associations is influenced by the methods used to
control for potential confounding by long-term time
trends, seasonality, weather, and other time-dependent
covariates. However, numerous researchers using various
methods, including alternative time series analytic ap-
proaches and case-crossover designs, continue to fairly
consistently observe adverse mortality associations with
short-term elevations in ambient PM.
LONG-TERM EXPOSURE AND MORTALITY
Although daily time series studies of acute exposures con-
tinue to suggest short-term acute PM effects, they provide
little information about the degree of life shortening,
pollution effects on longer-term mortality rates, or the
Pope and Dockery
Volume 56 June 2006 Journal of the Air & Waste Management Association 713
role of pollution in inducing or accelerating the progress
of chronic disease.
168
Several analyses of pollution and
mortality data, as early as 1970, reported that long-term
average concentrations of PM
2.5
or sulfate are associated
with annual mortality rates across U.S. metropolitan ar-
eas.
169–175
These population-based cross-sectional mortal-
ity rate studies were largely discounted by 1997 because of
concern that they could not control for individual risk
factors, such as cigarette smoking, which could poten-
tially confound the air pollution effects. With regard to
the mortality effects of long-term PM exposure, recent
emphasis has been on prospective cohort studies
176
that
can control for individual differences in age, sex, smoking
history, and other risk factors. However, because these
studies require collecting information on large numbers
of people and following them prospectively for long pe-
riods of time, they are costly, time consuming, and, there-
fore, much less common. A brief summary of results from
these studies is presented in Table 2.
Original Harvard Six Cities and ACS Studies
By 1997, two cohort-based mortality studies had reported
evidence of mortality effects of chronic exposure to fine
particulate air pollution. The first study, often referred to
as the Harvard Six Cities Study,
26
reported on a 14- to
16-yr prospective follow-up of Ͼ8000 adults living in six
U.S. cities, representing a wide range of pollution expo-
sure. The second study, referred to as the ACS study,
linked individual risk factor data from the ACS, Cancer
Prevention Study II with national ambient air pollution
data.
27
The analysis included data from Ͼ500,000 adults
who lived in Յ151 metropolitan areas and were followed
prospectively from 1982 through 1989. Both the Harvard
Six Cities and the ACS cohort studies used Cox propor-
tional hazard regression modeling to analyze survival
times and to control for individual differences in age, sex,
cigarette smoking, education levels, body mass index, and
other individual risk factors. In both studies, cardiopul-
monary mortality was significantly and most strongly
associated with sulfate and PM
2.5
concentrations.
Although both the Harvard Six Cities and ACS studies
used similar study designs and methods, these two studies
had different strengths and limitations. The strengths of
the Harvard Six Cities Study were its elegant and relatively
balanced study design, the prospective collection of
study-specific air pollution data, and the ability to present
the core results in a straightforward graphical format. The
primary limitations of the Harvard Six Cities Study were
Table 2. Comparison of percentage increase (and 95% CI) in relative risk of mortality associated with long-term particulate exposure.
Study Primary Sources Exposure Increment
Percent Increases in Relative Risk of Mortality
(95% CI)
All Cause Cardiopulmonary Lung Cancer
Harvard Six Cities, original Dockery et al. 1993
26
10 g/m
3
PM
2.5
13 (4.2, 23) 18 (6.0, 32) 18 (Ϫ11, 57)
Harvard Six Cities, HEI reanalysis Krewski et al. 2000
177
10 g/m
3
PM
2.5
14 (5.4, 23) 19 (6.5, 33) 21 (Ϫ8.4, 60)
Harvard Six Cities, extended analysis Laden et al. 2006
184
10 g/m
3
PM
2.5
16 (7, 26) 28 (13, 44)
a
27 (Ϫ4, 69)
ACS, original Pope et al. 1995
27
10 g/m
3
PM
2.5
6.6 (3.5, 9.8) 12 (6.7,17) 1.2 (Ϫ8.7, 12)
ACS, HEI reanalysis Krewski et al. 2000
177
10 g/m
3
PM
2.5
7.0 (3.9, 10) 12 (7.4, 17) 0.8 (Ϫ8.7, 11)
ACS, extended analysis Pope et al. 2002
179
10 g/m
3
PM
2.5
6.2 (1.6, 11) 9.3 (3.3, 16) 13.5 (4.4, 23)
Pope et al. 2004
180
12 (8, 15)
a
ACS adjusted using various education
weighting schemes
Dockery et al. 1993
26
10 g/m
3
PM
2.5
8–11 12–14 3–24
Pope et al. 2002
179
Krewski et al. 2000
177
ACS intrametro Los Angeles Jerrett et al. 2005
181
10 g/m
3
PM
2.5
17 (5, 30) 12 (Ϫ3, 30) 44 (Ϫ2, 211)
Postneonatal infant mortality, U.S. Woodruff et al. 1997
185
20 g/m
3
PM
10
8.0 (4, 14) – –
Postneonatal infant mortality, CA Woodruff et al. 2006
186
10 g/m
3
PM
2.5
7.0 (Ϫ7, 24) 113 (12, 305)
c
–
AHSMOG
b
Abbey et al. 1999
187
20 g/m
3
PM
10
2.1 (Ϫ4.5, 9.2) 0.6 (Ϫ7.8, 10) 81 (14, 186)
AHSMOG, males only McDonnell et al. 2000
188
10 g/m
3
PM
2.5
8.5 (Ϫ2.3, 21) 23 (Ϫ3, 55) 39 (Ϫ21, 150)
AHSMOG, females only Chen et al. 2005
189
10 g/m
3
PM
2.5
– 42 (6, 90)
a
–
Women’s Health Initiative Miller et al. 2004
190
10 g/m
3
PM
2.5
– 32 (1, 73)
a
VA, preliminary Lipfert et al. 2000, 2003
190,192
10 g/m
3
PM
2.5
0.3 (NS)
d
––
VA, extended Lipfert et al. 2006
193
10 g/m
3
PM
2.5
15 (5, 26)
e
––
11 CA counties, elderly Enstrom 2005
194
10 g/m
3
PM
2.5
1(Ϫ0.6, 2.6) – –
Netherlands Hoek et al. 2002
195
10 g/m
3
BS
17 (Ϫ24, 78) 34 (Ϫ32, 164) –
Netherlands Hoek et al. 2002
195
Near major road 41 (Ϫ6, 112) 95 (9, 251) –
Hamilton, Ontario, Canada Finkelstein et al. 2004
197
Near major road 18 (2, 38) – –
French PAARC Filleul et al. 2005
198
10 g/m
3
BS
7 (3, 10)
f
5(Ϫ2,12)
f
3(Ϫ8,15)
f
Cystic fibrosis Goss et al. 2004
200
10 g/m
3
PM
2.5
32 (Ϫ9, 93) – –
a
Cardiovascular only;
b
Pooled estimates for males and females; pollution associations were observed primarily in males and not females;
c
Respiratory only;
d
Reported to be nonsignificant by author; overall, effect estimates to various measure of particulate air pollution were highly unstable and not robust to selection
of model and time windows;
e
Estimates from the single pollutant model and for 1989 –1996 follow-up; effect estimates are much smaller and statistically
insignificant in an analysis restricted to counties with nitrogen dioxide data and for the 1997–2001 follow-up; furthermore, county-level traffic density is a strong
predictor of survival and stronger than PM
2.5
when included with PM
2.5
in joint regressions;
f
Estimates when six monitors that were heavily influenced by local
traffic sources were excluded; when data from all 24 monitors in all areas were used, no statistically significant associations between mortality and pollution were
observed.
Pope and Dockery
714 Journal of the Air & Waste Management Association Volume 56 June 2006
the small number of subjects from a small number of
study areas (that is exposures) in the Eastern United
States. In contrast, the major strength of the ACS study
was the large number of participants and cities distributed
across the whole United States. The primary limitation of
the ACS was the lack of planned, prospective collection of
study-specific air pollution and health data and the reli-
ance on limited, separately collected subject and pollu-
tion data. However, the ACS study provided a test of the
hypotheses generated from the Harvard Six Cities Study
in an independently collected dataset. These two studies,
therefore, were complementary.
Reanalyses and Extended Analyses of Harvard
Six Cities and ACS Studies
In the mid-1990s, the Harvard Six Cities and the ACS
prospective cohort studies provided compelling evidence
of mortality effects from long-term fine particulate air
pollution. Nevertheless, these two studies were controver-
sial, and the data quality, accessibility, analytic methods,
and validity of these studies came under intense scruti-
ny.
81
There were calls from political leaders, industry rep-
resentatives, interested scientists, and others to make the
data available for further scrutiny and analyses. There
were also serious constraints and concerns regarding the
dissemination of confidential information and the intel-
lectual property rights of the original investigators and
their supporting institutions. In 1997, the investigators of
the two studies agreed to provide the data for a intensive
reanalysis by an independent research team under Health
Effects Institute (HEI) oversight, management, sponsor-
ship, and under conditions that assured the confidential-
ity of the information on individual study participants.
The reanalysis included: (1) a quality assurance audit of
the data, (2) a replication and validation of the originally
reported results, and (3) sensitivity analyses to evaluate
the robustness of the original findings. The reanaly-
sis
177,178
reported that the data were “generally of high
quality” and that the results originally reported could be
reproduced and validated. The data audit and validation
efforts revealed some data and analytic issues that re-
quired some tuning, but the adjusted results did not differ
substantively from the original findings. The reanalysis
demonstrated the robustness of the PM-mortality risk es-
timates to many alternative model specifications. The re-
analysis team also made a number of innovative method-
ological contributions that not only demonstrated the
robustness of the PM-mortality results but substantially
contributed to subsequent analyses. In the reanalysis, per-
sons with higher educational attainment were found to
have lower relative risks of mortality associated with
PM
2.5
in both studies.
Further extended analyses of the ACS cohort
179,180
included more than twice the follow-up time (Ͼ16 years)
and approximately triple the number of deaths. The mor-
tality associations with fine particulate and sulfur oxide
pollution persisted and were robust to control for individ-
ual risk factors including age, sex, race, smoking, educa-
tion, marital status, body mass index, alcohol use, occu-
pational exposures, and diet and the incorporation of
both random effects and nonparametric spatial smooth-
ing components. There was no evidence that the PM-
mortality associations were because of regional or other
spatial differences that were not controlled in the analy-
sis. These analyses also evaluated associations with ex-
panded pollution data, including gaseous copollutant
data and new PM
2.5
data. Elevated mortality risks were
most strongly associated with measures of PM
2.5
and sul
-
fur oxide pollution. Coarse particles and gaseous pollut-
ants, except for sulfur dioxide (SO
2
), were generally not
significantly associated with elevated mortality risk.
Jerret et al.
181
assessed air pollution associations of
the ϳ23,000 subjects in the ACS cohort who lived in the
metropolitan Los Angeles area. PM-mortality associations
were estimated based on PM
2.5
measures from 23 moni
-
toring sites interpolated to 267 residential zip code cen-
troids for the period between 1982 and 2000. Cox pro-
portional hazards regression models controlled for age,
sex, race, smoking, education, marital status, diet, alcohol
use, occupational exposures, and body mass.
179
In addi-
tion, because variations in exposure to air pollution
within a city may correlate with socioeconomic gradients
that influence health and susceptibility to environmental
exposures, zip code-level ecological variables were used to
control for potential “contextual neighborhood con-
founding.”
182,183
The mortality associations with the in-
trametropolitan PM
2.5
concentrations were generally
larger than those observed previously in the ACS cohort
across metropolitan areas.
A recent analysis of the Harvard Six Cities cohort
184
extended the mortality follow-up for 8 more years with
approximately twice the number of deaths. PM
2.5
concen
-
trations for the extended follow-up years were estimated
from PM
10
and visibility measures. PM
2.5
-mortality asso
-
ciations, similar to those found in the original analysis,
were observed for all-cause, cardiovascular, and lung can-
cer mortality. However, PM
2.5
concentrations were sub
-
stantially lower for the extended follow-up period than
they were for the original analysis, especially for two of
the most polluted cities. Reductions in PM
2.5
concentra
-
tions were associated with reduced mortality risk and
were largest in the cities with the largest declines in PM
2.5
concentrations. The authors note that, “these findings
suggest that mortality effects of long-term air pollution
may be at least partially reversible over periods of a de-
cade.”
184
Other Independent Studies
Woodruff et al.
185
reported the results of an analysis of
postneonatal infant mortality (deaths after 2 months fol-
lowing birth determined from the U.S. National Center
for Health Statistics birth and death records) for ϳ4 mil-
lion infants in 86 U.S. metropolitan areas between 1989
and 1991 linked with EPA-collected PM
10
. Postneonatal
infant mortality was compared with levels of PM
10
con
-
centrations during the 2 months after birth controlling
for maternal race, maternal education, marital status,
month of birth, maternal smoking during pregnancy, and
ambient temperatures. Postneonatal infant mortality for
all causes, respiratory causes and sudden infant death
syndrome (SIDS) were associated with particulate air pol-
lution. Woodruff et al.
186
also linked monitored PM
2.5
to
Pope and Dockery
Volume 56 June 2006 Journal of the Air & Waste Management Association 715
infants who were born in California in 1999 and 2000 and
who lived within 5 mi of a monitor, matching 788 post-
neonatal deaths to 3089 survivors. Each 10-g/m
3
in
-
crease in PM
2.5
was associated with a near doubling of the
risk of postneonatal death because of respiratory causes
and a statistically insignificant increase of ϳ7% for death
from all causes (Table 2).
The Adventist Health Study of Smog (AHSMOG) co-
hort study related air pollution to 1977–1992 mortality in
Ͼ6000 nonsmoking adults living in California, predomi-
nantly from San Diego, Los Angeles, and San Francisco.
187
All-cause mortality, nonmalignant respiratory mortality,
and lung cancer mortality were significantly associated
with ambient PM
10
concentrations in males but not in
females. Cardiopulmonary disease mortality was not sig-
nificantly associated with PM
10
in either males or females.
This study did not have direct measures of PM
2.5
but
relied on TSP and PM
10
data. In a follow-up analysis,
188
visibility data were used to estimate PM
2.5
exposures of a
subset of males who lived near an airport. All-cause, lung
cancer, and nonmalignant respiratory disease (either as
the underlying or a contributing cause) were more
strongly associated with PM
2.5
than with PM
10
.Inare
-
cent analysis of the AHSMOG cohort, fatal coronary heart
disease was significantly associated with PM among fe-
males but not among males.
189
The association between long-term PM
2.5
exposure
and cardiovascular events (fatal and nonfatal) were ex-
plored in the Women’s Health Initiative Observational
Study.
190
Based on measurements from the nearest mon-
itor, air pollution exposures were estimated for ϳ66,000
postmenopausal women without prior cardiovascular dis-
ease. After adjusting for age, smoking, and various other
risk factors, an incremental difference of 10 g/m
3
of
PM
2.5
was associated with a 14% (95% confidence interval
[CI], 3–26%) increase in nonfatal cardiovascular events
and with a 32% (95% CI, 1–73%) increase in fatal cardio-
vascular events.
Lipfert et al.
191,192
assessed the association of total
mortality and air pollution in a prospective cohort of
ϳ50,000 middle-aged, hypertensive, male patients from
32 Veterans Administration (VA) clinics followed for ϳ21
years. The cohort had a disproportionately large number
of current or former smokers (81%) and African-Ameri-
cans (35%) relative to the U.S. population or to other
cohorts that have been used to study air pollution. Air
pollution exposures were estimated by averaging air pol-
lution data for participants’ county of residence at the
time of entrance into the cohort. Only analyses of total
mortality were reported. In addition to considering mor-
tality and average exposures over the entire follow-up
period, three sequential mortality periods and four expo-
sure periods were defined and included in various analy-
ses. Lipfert et al.
193
extended the follow-up of the VA
cohort and focused on traffic density as the measure of
environmental exposure. It was suggested that traffic den-
sity was a more “significant and robust predictor of sur-
vival in this cohort” than PM
2.5
. However, of the various
measures of ambient air pollution, PM
2.5
was most
strongly correlated with traffic density (r ϭ 0.50). In single
pollutant models, PM
2.5
was associated with mortality
risk resulting in risk estimates comparable to other co-
horts (see Table 2). Overall in the VA analyses, effect
estimates to various measures of PM were unstable and
not robust to model selection, time windows used, or
various other analytic decisions. It was difficult, based on
the preliminary results presented, to make conclusive sta-
tistical inferences regarding PM-mortality associations.
Enstrom
194
reported an analysis of ϳ36,000 elderly
males and females in 11 California counties followed be-
tween 1973 and 2002. Countywide PM
2.5
concentrations
were estimated from outdoor ambient monitoring for the
time period 1979 –1983. For approximately the first half
of the follow-up period (1973–1983) and for the time
period approximately concurrent with PM
2.5
monitoring,
a small PM
2.5
-mortality association was observed (10
g/m
3
of PM
2.5
was associated with a 4% [95% CI, 1- 7%]
increase risk of mortality). No PM
2.5
-mortality risk asso
-
ciations were observed for the later followup (1983–2002).
For the entire follow-up period, only a small statistically
insignificant association was observed (Table 2).
In a pilot study, Hoek et al.
195
evaluated the associa-
tions between mortality and PM based on a random sam-
ple of 5000 participants in the Netherlands Cohort Study
on Diet and Cancer, originally 55–69 yr of age and fol-
lowed for Ͼ8 yr. Although the effect estimates were not
very precise, the adjusted risk of cardiopulmonary mor-
tality was nearly double for individuals who lived within
100 m of a freeway or within 50 m of a major urban road.
Based on residential location of participants and interpo-
lation of pollution data from the Netherlands’ national
air pollution monitoring network, average background
concentrations of black smoke ([BS] or British smoke mea-
sured by optical densities or light absorbance of filters
used to gather PM from the air
196
) for the first 4 yr of
follow-up were estimated. Background plus local traffic-
related BS exposures were estimated by adding to the
background concentration a quantitative estimate of liv-
ing near a major road. Cardiopulmonary mortality was
associated with estimates of exposure to BS, and the asso-
ciation was nearly doubled when local traffic-related
sources of BS in addition to background concentrations
were modeled.
In an exploration of the relationship between prox-
imity to traffic air pollution and mortality observed in the
Netherlands study, an analysis using a cohort of 5228
persons Ͼ40 yr of age living in Hamilton, Ontario, Can-
ada, was conducted.
197
Somewhat higher mortality risks
were observed for individuals who lived within 100 m of
a highway or within 50 m of a major road.
Filleul et al.
198
reported an analysis of ϳ14,000 adults
who resided in 24 areas from seven French cities as part of
the Air Pollution and Chronic Respiratory Diseases
(PAARC) survey. Participants were enrolled in 1974, and a
25-year mortality follow-up was conducted. Ambient air
pollution monitoring for TSP, BS, nitrogen dioxide, and
NO was conducted for 3 yr in each of the 24 study areas.
When survival analysis was conducted using data from all
24 monitors in all of the areas, no statistically significant
associations between mortality and pollution were ob-
served. However, when the six monitors that were heavily
Pope and Dockery
716 Journal of the Air & Waste Management Association Volume 56 June 2006
influenced by local traffic sources were excluded, nonac-
cidental mortality was significantly associated with all
four measures of pollution, including BS (Table 2). In
addition to PM, mortality was associated with nitrogen
oxides. Nitrogen oxide concentrations were also signifi-
cantly associated with mortality risk in a cohort of Nor-
wegian men,
199
but no measure of PM was available.
Finally, a unique study of the effects of ambient air
pollution was conducted utilizing a cohort of ϳ20,000
patients Ͼ6 yr old who were enrolled in the U.S based
Cystic Fibrosis Foundation National Patient Registry in
1999 and 2000.
200
Annual average air pollution exposures
were estimated by linking fixed-site ambient monitoring
data with resident zip code. A positive, but not statisti-
cally significant, association between PM
2.5
and mortality
was observed. PM
2.5
was associated with statistically sig
-
nificant declines in lung function (FEV
1
) and an increase
in the odds of two or more pulmonary exacerbations.
Summary and Discussion
As can be seen in Table 2, for both the Harvard Six Cities
and the ACS prospective cohort studies, the estimated
effects for all-cause and cardiopulmonary mortality were
relatively stable across different analyses. The Harvard Six
Cities estimates, however, were approximately twice as
large as the ACS estimates. Two main factors may explain
these differences in estimated PM-mortality effects.
First, both the reanalysis and extended analyses have
found that persons with higher educational attainment
had lower relative risk of PM-related mortality. The ACS
cohort overrepresented relatively well-educated individu-
als relative to the Harvard Six Cities study. To provide a
tentative estimate of how this overrepresentation may
have influenced the pooled-effect estimates from the ACS
study, various schemes for adjusting the ACS effect esti-
mates by reweighting the regression coefficients were
tried. A relatively conservative approach was to calculate
a pooled ACS estimate by weighting the effect estimates
by education level from the ACS cohort with the propor-
tions of participants from each education level from the
Harvard Six Cities cohort based on the Krewski et al.
177
reanalysis (Part II, Table 52). A more aggressive approach
was to use the Cox proportional hazard regression coeffi-
cients for the ACS extended analysis
179
that were esti-
mated for each of the three education levels. Pooled,
weighted estimates were then calculated using weights
(proportion of sample within each of the three education
levels from Krewski et al.
177
, Part II, Table 52) for both the
Harvard Six Cities study and the ACS study, and then the
ratio of the pooled, weighted estimates was used to adjust
the originally reported ACS effect estimates. As can be
seen in Table 2, reweighting to account for the overrep-
resentation of relatively well-educated individuals in the
ACS cohort explains part, but not all, of the difference in
effect estimates between the Harvard Six Cities and ACS
studies.
Second, the geographical areas that defined the com-
munities studied in the Harvard Six Cities study were, on
average, substantially smaller than the metropolitan areas
included in the ACS study. Indeed, an analysis of the Los
Angeles metropolitan area ACS participants showed that
interpolated PM
2.5
air pollution concentrations resulted
in effect estimates comparable with estimates from the
Harvard Six Cities Study. Similarly, in the Netherlands
study, when local sources of particulate pollution expo-
sure in addition to community-wide background concen-
trations were modeled, the elevated relative risk estimates
also approximately doubled. These results suggest that
PM-mortality effect estimates based on analysis that only
uses metropolitan-wide average background concentra-
tions may underestimate the true pollution-related health
burden and suggests the importance of analyses with
more focused spatial resolution.
In 1997, Vedal
80
argued that the evidence for sub-
stantive health effects because of chronic or long-term
exposure to particulate air pollution was weak. Since then,
the HEI reanalysis of the Harvard Six Cities and ACS
prospective cohort studies and the subsequent extended
analyses of these cohort studies have strengthened the
evidence of long-term, chronic health effects. Reanalyses
are not as convincing as new, independent cohort studies.
The results from the independent Women’s Health Initia-
tive Study
190
add to the evidence that long-term exposure
increases the risk of cardiovascular disease in women. The
evidence is further bolstered by results from the infant
mortality studies,
185,186
the Netherlands study,
195
and the
Hamilton study
197
but less so by the mixed results from
the AHSMOG studies,
187–189
the French PAARC study,
198
the VA analyses,
191–193
and the 11 California counties
study.
194
With regard to the infant mortality find-
ings,
185,186
although the analyses are based on cross-sec-
tional or long-term differences in air pollution, the time
frame of exposure for the infants was clearly shorter than
for adults (a few months vs. years). The relevant time
scales of exposure for different age groups, levels of sus-
ceptibility, and causes of death need further exploration.
TIME SCALES OF EXPOSURE
The PM-mortality effect estimates from the long-term
prospective cohort studies (Table 2) are substantially
larger than those from the daily time series and case-
crossover studies (Table 1). The much larger PM-mortality
effect estimates from the prospective cohort studies are
inconsistent with the supposition that they are due to
short-term harvesting or mortality displacement. If pollu-
tion-related excess deaths are only because of deaths of
the very frail who have heightened susceptibility and who
would have died within a few days anyway, then the
appropriate time scale of exposure would be only a few
days, and impacts on long-term mortality rates would be
minimal.
Mortality effects of short-term exposure, however,
may not be attributed primarily to harvesting. Long-term
repeated exposures to pollution may have more broad-
based impacts on long-term health and susceptibility.
Much of the difference in PM-mortality associations ob-
served between the daily time series and the prospective
cohort studies may be because of the dramatically differ-
ent time scales of exposure (a few days vs. decades). Ef-
fective dose, in terms of impact on risk of adverse health
effects, is almost certainly dependent on both exposure
concentrations and length of exposure. It is reasonable to
expect that effect estimates could be different for different
Pope and Dockery
Volume 56 June 2006 Journal of the Air & Waste Management Association 717
time scales of exposure, that long-term repeated expo-
sures could have larger, more persistent effects than short-
term transient exposures, and that long-term average ex-
posures could be different from the cumulative effect of
short-term transient exposures.
Neither the daily time series studies nor the prospec-
tive cohort studies were designed to evaluate the alterna-
tive time scales of exposure. These studies were designed
primarily to exploit obvious, observable sources of expo-
sure variability. Short-term temporal variability is exam-
ined in the daily time series studies. In most of these
studies, various approaches are used to focus only on
short-term variability while taking out or controlling for
longer-term temporal variability, such as seasonality and
time trends. Thus, by design, opportunities to evaluate
effects of intermediate or long-term exposure are largely
eliminated. The other important dimension of exposure
variability is spatial (or cross-sectional) variability of long-
term average concentrations. The major prospective co-
hort studies have been designed primarily to exploit this
much longer-term spatial variability. Efforts to estimate
the dynamic exposure-response relationship between
PM
2.5
exposure and human mortality must integrate evi
-
dence from long-term, intermediate, and short-term time
scales.
201
Studies of Intermediate Time Scales of Exposure
Before 1997, there was hardly any reported research that
evaluated intermediate time scales of exposure. One ex-
ception was research related to the operation of a steel
mill in Utah Valley.
20,28,202
During the winter of 1986–
1987, a labor dispute and change in ownership resulted in
a 13-month closure of the largest single source of partic-
ulate air pollution in the valley, a local steel mill. During
the 13-month closure period, average PM
10
concentra
-
tions decreased by 15 g/m
3
, and mortality decreased by
3.2%.
A more recent evaluation of PM-related changes in
mortality using an intermediate time scale was conducted
in Dublin, Ireland.
203
During the 1980s, a dominant
source of Dublin’s ambient PM was coal smoke from do-
mestic fires. In September of 1990, the sale of coal was
banned, resulting in a 36-g/m
3
decrease in average am
-
bient PM as measured by BS. After adjusting in Poisson
regression for temperature, RH, day of week, respiratory
epidemics, and standardized cause-specific death rate in
the rest of Ireland, statistically significant drops in all of
the nontrauma deaths (Ϫ5.7%; 95% CI, Ϫ7.2% to
Ϫ4.1%), cardiovascular deaths (Ϫ10.3%; 95% CI, Ϫ12.6%
to Ϫ8%), and respiratory deaths (Ϫ15.5%; 95% CI,
Ϫ19.1% to Ϫ11.6%) were observed.
As noted above, in the extended analysis of the Har-
vard Six Cities cohort,
184
fine particulate concentrations
were substantially lower for the 8-yr extended follow-up
period than they were for the original analysis, especially
for two of the most polluted cities. These reductions in
PM
2.5
concentrations were associated with reduced mor
-
tality risk, suggesting that the mortality effects were at
least partially reversible within a time scale of just a few
years. Furthermore, the reductions in PM
2.5
in the ex
-
tended follow-up compared with the original study pe-
riod were associated with improved survival, that is, a
relative risk of Ϫ27% (95% CI, Ϫ43% to Ϫ5%) for each
10-g/m
3
reduction in PM
2.5
.
Daily Time Series Studies with Longer Time
Scales or Extended Distributed Lags
Several researchers have developed methods to analyze
daily time series data for time scales of exposure substan-
tially longer than just a few days. A primary motivation of
this effort was to explore the “harvesting,” or mortality
displacement hypothesis. If pollution-related excess
deaths occur only among the very frail, then the excess
deaths during and immediately after days of high pollu-
tion should be followed by a short-term compensatory
reduction in deaths. To explore whether or not this phe-
nomena could be observed, Zeger et al.
204
proposed fre-
quency decompositions of both the mortality counts and
air pollution data. They applied frequency domain log-
linear regression
205
to mortality data from a single city
(Philadelphia, PA) and found larger PM effects on the
relatively longer time scales, a finding inconsistent with
harvesting. This work was extended by Dominici et al.
206
to a two-stage model that allowed for combining evidence
across four U.S. cities with daily PM
10
levels. They found
the PM-mortality associations were larger at longer time
scales (10 days to 2 months) than at time scales of just a
few days. Schwartz
207–209
applied a related approach using
smoothing techniques to decompose the data into differ-
ent time scales in two separate analyses using data from
Chicago, IL, and Boston, MA, and also found that the
PM-mortality associations were much larger for the longer
time scales.
An alternative approach to evaluate longer time
scales is the use of extended distributed lags in time series
analyses. Distributed lag models have long been used in
econometrics
210,211
and have more recently been applied
in air pollution epidemiology.
31,212
Studies using distrib-
uted lag models to evaluate associations from 5 to Յ60
days after exposure have been conducted using data from
10 U.S. cities,
213,214
European cities from the APHEA-2
project,
215,216
and Dublin.
217
In all of these analyses, the
net PM-mortality effect was larger when time scales
longer than a few days were used.
Summary and Discussion
For comparison purposes, Table 3 provides a simple sum-
mary of estimated excess risk of mortality estimates for
different studies with different time scales of exposure.
These results do not provide the complete picture, but
they suggest that the short-term, daily time series air
pollution studies are not observing only harvesting or
mortality displacement. These results also suggest that
daily time series studies capture only a small amount of
the overall health effects of long-term repeated exposure
to particulate air pollution. Because the adverse health
effects of particulate air pollution are likely dependent on
both exposure concentrations and length of exposure, it
is expected that long-term repeated exposures would have
larger, more persistent cumulative effects than short-term
transient exposures. PM-mortality effect estimates for in-
termediate time intervals provide evidence that the dif-
ference in PM-mortality associations observed between
the daily time series and the prospective cohort studies
Pope and Dockery
718 Journal of the Air & Waste Management Association Volume 56 June 2006
are at least partially because of the substantially different
time scales of exposure.
SHAPE OF CONCENTRATION-RESPONSE
FUNCTION
Understanding the shape of the concentration-response
function and the existence of a no-effects threshold level
has played a critical role in efforts to establish and evalu-
ate ambient air quality standards and related public
health policy. This information is also vital in economic
and public policy analyses that require estimating the
marginal health costs of pollution. An early analysis by
Ostro
110
evaluated the shape of the concentration-re-
sponse function and the existence of a no-effects thresh-
old in London mortality and air pollution data for 14
winters (1958–1972). Linear spline functions that allowed
for different response relationships below and above 150
g/m
3
were estimated. Mortality effects were observed
even in winters without historically severe pollution epi-
sodes, and there was no evidence of a threshold. Schwartz
and Marcus
17
plotted the same London data after sorting
the observations in order of increasing pollution levels
and taking the means of adjacent observations. No
threshold was observed; in fact, the slope of the concen-
tration-response function was steeper at lower concentra-
tions than at higher concentrations.
In the early 1990s, various approaches were used to
evaluate the shape of the concentration-response func-
tion. For example, researchers often divided pollution
concentrations into quintiles (or quartiles) and included
indicator variables for different ranges of air pollution in
the time series regression models. This allowed for the
estimated adjusted relative risk of death to be plotted over
various levels of pollution.
19–23
The associations generally
appeared to be near linear with no clear threshold.
218
The
development and use of various parametric and nonpara-
metric smoothing approaches not only allowed for more
flexible handling of long-term time trends, seasonality,
and various weather variables, but they also allowed for
direct exploration of the shape of the concentration-re-
sponse function.
219
Such analyses were conducted in nu-
merous single-city daily time series studies.
24,71,112,220
Generally the shapes of the estimated concentration-re-
sponse function were not significantly different from lin-
ear and were not consistent with well-defined thresh-
olds.
218
However, the lack of statistical power to make
strong statistical inferences regarding function shape, and
the generalizability of single-city estimates of the concen-
tration-response relationships were questioned.
Multicity Daily Time Series Mortality
Since 1997, methods have been developed to explore the
shape of the PM-mortality concentration-response func-
tions in daily time series studies of multiple cities, which
enhance statistical power and generalizability. Schwartz
and Zanobetti
221
estimated a pooled or combined concen-
tration-response function for 10 U.S. cities. The combined
or “meta-smoothed” concentration-response function
was estimated using Poisson regression models fitting
nonparametric smoothed functions for PM
10
and calcu
-
lating the inverse variance weighted average across the 10
cities for each 2-g/m
3
increment of PM
10
. The estimated
combined 10-city concentration-response function was
near linear with no evidence of a threshold (see Figure 1a).
Schwartz et al.
222
applied essentially the same approach
on daily mortality and BS data from eight Spanish cities,
finding a near linear concentration-response function
with no evidence of a threshold (see Figure 1b).
An alternative approach to estimating multicity PM-
mortality combined concentration-response functions
was proposed by Daniels et al.
223
and Dominici et al.
224
They developed flexible modeling strategies for daily
time series analyses that included spline and threshold
Table 3. Comparison of estimated excess risk of mortality estimates for different time scales of exposure.
Study Primary Sources
Time Scale
of Exposure
% Change in Risk of Mortality Associated with an
Increment of 10 g/m
3
PM
2.5
or 20 g/m
3
PM
10
or BS
All Cause
Cardiovascular/
cardiopulmonary Respiratory Lung Cancer
Daily time series Table 1 1–3 days 0.4–1.4 0.6–1.1 0.6–1.4 –
10 U.S. cities, time series, extended
distributed lag
Schwartz 2000
213
1 day 1.3 – – –
2 days 2.1 – – –
5 days 2.6 – – –
10 European cities, time series, extended
distributed lag
Zanobetti et al. 2002
215
2 days 1.4 – – –
40 days 3.3 – – –
10 European cities, time series, extended
distributed lag
Zanobetti et al. 2003
216
2 days – 1.4 1.5 –
20 days – 2.7 3.4 –
30 days – 3.5 5.3 –
40 days – 4.0 8.6 –
Dublin daily time series, extended
distributed lag
Goodman et al. 2004
217
1 day 0.8 0.8 1.8 –
40 days 2.2 2.2 7.2 –
Dublin intervention Clancy et al. 2002
203
months to year 3.2 5.7 8.7 –
Utah Valley, time series and intervention Pope et al. 1992
20
5 days 3.1 3.6 7.5 –
13 months 4.3 – – –
Harvard Six Cities, extended analysis Laden et al. 2006
184
1–8 yr 14 – – –
Prospective cohort studies Dockery et al. 1993
26
10ϩ yr 6–17 9–28 – 14–44
Pope et al. 2002
179
Pope and Dockery
Volume 56 June 2006 Journal of the Air & Waste Management Association 719
concentration-response functions and applied these
methods to data from the 20 largest U.S. cities from the
NMMAPS project. PM-mortality concentration-response
functions were estimated using three different modeling
approaches: (1) models with log-linear functions for PM,
(2) flexible smoothed functions, and (3) models that as-
sumed or allowed for specific PM threshold levels. For
all-cause mortality and for cardiopulmonary mortality,
linear models without thresholds fit the PM-mortality
association better than threshold models or even flexible
cubic spline models (see Figure 1c). The researchers
225,226
extended these analyses to the 88 largest cities in the
United States. Although they found regional differences,
the overall pooled concentration-response function for
the nation was nearly linear (see Figure 1d).
Samoli et al.
227
applied regression spline models to
flexibly estimate the PM-mortality association to data
from 22 European cities participating in the APHEA
project. They observed some heterogeneity in effect esti-
mates across the different cities, but the pooled estimated
PM-mortality association was not significantly different
from linear (see Figure 1e).
Figure 1. Selected concentration-response relationships estimated from various multicity daily time series mortality studies (approximate
adaptations from original publications rescaled for comparison purposes).
Pope and Dockery
720 Journal of the Air & Waste Management Association Volume 56 June 2006
Cross-Sectional and Prospective Cohort
Mortality Studies
Given the small number of cross-sectional and prospec-
tive cohort studies, the shape of the concentration-re-
sponse function with long-term chronic exposure has not
been as carefully explored as with the daily time series
studies. It has long been observed that long-term average
sulfate and/or fine particulate air pollution concentra-
tions are associated with mortality rates across U.S. urban
areas (especially after adjusting for age, sex, and
race).
169–175
Figure 2a presents U.S. metropolitan area
mortality rates for 1980
228
adjusted based on 1980 cen-
sus
229
age-sex-race-specific population counts plotted
over mean PM
2.5
concentrations as compiled and re
-
ported by Krewski et al.
177
Figure 2b presents adjusted
mortality rates or rate ratios for U.S. cities plotted over
corresponding PM
2.5
concentrations based on the ex
-
tended analysis of the Harvard Six Cities Study.
184
The
mortality effects can reasonably be modeled as linear or
log linear.
The extended follow-up analysis of the ACS study
more fully evaluated the shape of the concentration
response function by using a robust locally weighted
regression smoother.
179
The nonparametric smoothed
exposure-response relationships between cause-specific
mortality and long-term exposure to PM
2.5
are also pre
-
sented in Figure 2c. Relative risks for all-cause, cardiopul-
monary, and lung cancer mortality increased across the
gradient of PM
2.5
. Although some inevitable nonlinearity
is observable, goodness-of-fit tests indicated that the as-
sociations were not significantly different from linear (P Ͼ
0.20). The shape of the exposure-response function at
concentrations above the range of pollution observed in
this analysis remains poorly estimated. Because concen-
trations above this range of pollution occur in many other
parts of the world, an attempt to quantify the global
burden of disease attributable to exposure to air pollution
required projected effect estimates at higher concentra-
tions.
230
A log-linear fit of PM
2.5
, where the slope of the
concentration-response function decreases at higher con-
centrations, also fit the data.
230
The concentration-response function for long-term
exposure to particulate air pollution and other health end
points has not been systematically explored. However,
various studies are suggestive. For example, Gauderman et
al.
231
reported results from the Children’s Health Study
that prospectively monitored the growth in lung function
of school children ages 10 –18 yr who lived in 12 Southern
California communities with a relatively wide range of air
pollution. Over the 8-yr period, deficits in lung function
Figure 2. Selected concentration-response relationships estimated from various studies of long-term exposure (approximate adaptations from
original publications rescaled for comparison purposes).
Pope and Dockery
Volume 56 June 2006 Journal of the Air & Waste Management Association 721
growth were associated with PM
2.5
and accompanying
combustion-related air pollutants. As can be seen in Fig-
ure 2d, the concentration-response relationship between
PM
2.5
and the proportion of 18-yr-olds with FEV
1
Ͻ80%
of predicted appears to be near linear, without a discern-
ible threshold.
Summary and Discussion
Recent empirical evidence about the shape of the PM
concentration-response function is not consistent with a
well-defined no-effects threshold. Concentration-re-
sponse functions estimated from various multicity time
series studies are illustrated in Figure 1 and concentration-
response functions for long-term exposure studies are il-
lustrated in Figure 2. These concentration-response func-
tions have been adapted from the original publications
and put on common scales for easy comparison. The best
empirical evidence suggests that, across the range of par-
ticulate pollution observed in most recent studies, the
concentration-response relationship can reasonably be
modeled as linear. From a public policy perspective, at
least with regard to ambient air quality standard setting, a
linear concentration-response function without a well-
defined safe threshold level might be inconvenient. As
argued elsewhere,
218,232
from at least one perspective,
these results are good news, because they suggest that
even at common levels of air pollution, further improve-
ments in air quality are likely to result in corresponding
improvements in public health.
CARDIOVASCULAR DISEASE
Before the mid-1990s there was evidence of cardiovascu-
lar effects of PM air pollution. Deaths associated with the
severe pollution episodes of Meuse Valley, Belgium,
4
Do-
nora, PA,
9
and London
10
were due to both respiratory and
cardiovascular disorders, often in combination.
6,7
Analy-
ses of a less severe episode
38
observed stronger pollution-
related associations with cardiovascular than with respi-
ratory deaths. As noted earlier, many daily time series
mortality studies and the early prospective cohort stud-
ies
26,27
also observed that pollution was associated with
both respiratory and cardiovascular deaths (see Tables 1
and 2). Because it was unclear how these findings were
influenced by diagnostic misclassification or cross-coding
on death certificates, cardiovascular and respiratory
deaths were often pooled together as cardiopulmonary
deaths in the analyses.
26,27
Beginning in the mid-1990s,
several daily time series studies reported pollution-related
associations with hospitalizations for cardiovascular dis-
ease.
233–237
Although there was evidence of cardiovascular health
effects of PM air pollution, early research focused largely
on respiratory disease, including research dealing with
effects on asthma, obstructive pulmonary disease, respi-
ratory symptoms, and lung function.
52
Furthermore, be-
fore 1997, studies of ambient particulate air pollution and
health were rarely published or discussed in cardiovascu-
lar journals. Beginning in the late 1990s, studies dealing
with air pollution and cardiovascular disease were being
published, including in journals of cardiovascular medi-
cine, where they were receiving useful editorial discus-
sion
238–241
and reviews.
138,242–249
In 2004, the American
Heart Association published a Scientific Statement that
concluded that “studies have demonstrated a consistent
increase risk for cardiovascular events in relation to both
short- and long-term exposure to present-day concentra-
tions of ambient particulate matter.”
250
Long-Term Exposure and Cardiovascular Disease
Table 4 provides a brief overview of recent evidence of
cardiovascular and related effects associated with PM air
pollution. Several studies provide evidence that long-term
PM exposure contributes to cardiovascular morbidity and
mortality. As illustrated in Figure 3, initial and extended
analyses of the Harvard Six Cities and ACS cohorts con-
sistently observed PM
2.5
associations with cardiovascular
mortality. An extended analysis of the ACS cohort that
focused on cardiopulmonary mortality found that long-
term PM
2.5
exposures were strongly associated with isch
-
emic heart disease, dysrhythmias, heart failure, and car-
diac arrest mortality.
180
Relatively strong associations
between PM
2.5
and ischemic heart disease mortality were
observed in the metropolitan Los Angeles subcohort.
181
There are three interesting studies that have evalu-
ated the impact of long-term exposure to PM air pollution
and the development and progression of cardiovascular
disease. The first
251
explored associations between air pol-
lution and blood markers of cardiovascular risk, specifi-
cally fibrinogen levels and counts of platelets and white
blood cells. Data from the Third National Health and
Nutrition Examination Survey were linked with air pollu-
tion data. After controlling for age, race, sex, body mass
index, and smoking, elevated fibrinogen levels and plate-
let and white blood cell counts were all associated with
exposure to PM
10
. A second study
252
collected lung tissue
samples during necropsies of individuals who died be-
cause of violent causes and who lived in relatively clean
and polluted areas near Sao Paulo, Brazil. Individuals who
lived in more polluted areas had histopathologic evidence
of subclinical chronic inflammatory lung injury. A third
study used data on 798 participants from two clinical
trials conducted in the Los Angeles metro area.
253
PM
2.5
was associated with increased carotid intima-media thick-
ness (CIMT), a measure of subclinical atherosclerosis. A
cross-sectional contrast in exposure of 10-g/m
3
of PM
2.5
was associated with an ϳ4% increase in CIMT.
Short-Term Exposure and Cardiovascular
Disease
As noted above, there have been many studies that have
reported associations between short-term exposures to
particulate air pollution and cardiovascular mortality (see
Table 1). Studies reporting PM associations with cardio-
vascular hospitalizations have been more recent, but
there are now dozens of such studies. Table 5 presents a
comparison of pooled estimates of percentage increase in
relative risk of hospital admission for cardiovascular dis-
ease estimated across meta-analyses and multicity studies
of short-term changes in PM exposures. In addition, there
have been several recent studies that have reported asso-
ciations between PM exposure and stroke mortality and
hospitalizations. Several of these studies have been
from Asian countries with relatively high stroke mor-
tality.
254–257
However, a recent case-crossover study of
Pope and Dockery
722 Journal of the Air & Waste Management Association Volume 56 June 2006
