Estimates of global mortality attributable to particulate air pollution
using satellite imagery
Jessica Evans a,n
, Aaron van Donkelaar b
, Randall V. Martin b,c
, Richard Burnett d
,
Daniel G. Rainham e
, Nicholas J. Birkett f
, Daniel Krewski f
a
McLaughlin Centre for Population Health Risk Assessment, Institute of Population Health, University of Ottawa, 1 Stewart Street, Ottawa, Ontario, Canada K1N 6N5
b
Department of Physics and Atmospheric Science, Dalhousie University, Halifax, Nova Scotia, Canada B3H 3J5
c
Harvard-Smithsonian Center for Astrophysics, Cambridge, MA 02138, USA
d
Environmental and Radiation Health Sciences Directorate, Healthy Environments and Consumer Safety Branch, Health Canada, Environmental Health Center, Tunney’s Pasture,
Ottawa, Ontario, Canada K1A 0K9
e
Environmental Science, Faculty of Science, LSC 827, Dalhousie University, 1355 Oxford Street, Halifax, Nova Scotia, Canada B3H 4R2
f
Department of Epidemiology and Community Medicine, University of Ottawa, Ottawa, Ontario, Canada K1H 8M5
a r t i c l e i n f o
Article history:
Received 14 January 2012
Received in revised form
13 July 2012
Accepted 13 August 2012
Available online 6 September 2012
Keywords:
Health burden
Global
Mortality
Particulate matter
Satellite imaging
a b s t r a c t
Background: Epidemiological studies of the health effects of air pollution have traditionally relied upon
ground-monitoring stations to measure ambient concentrations. Satellite derived air pollution measures
offer the advantage of providing global coverage.
Objective: To undertake a global assessment of mortality associated with long-term exposure to fine
particulate air pollution using remote sensing data.
Methods: Global PM2.5 exposure levels were derived from the MODIS and MISR satellite instruments.
Relative risks and attributable fractions of mortality were modeled using previously developed
concentration–response functions for the association between PM2.5 and mortality.
Results: The global fraction of adult mortality attributable to the anthropogenic component of PM2.5 (95%
CI) was 8.0% (5.3–10.5) for cardiopulmonary disease, 12.8% (5.9–18.5) for lung cancer, and 9.4% (6.6–11.8)
for ischemic heart disease.
Conclusion: This study demonstrates the feasibility of using satellite derived pollution concentrations in
assessing the population health impacts of air pollution at the global scale. This approach leads to global
estimates of mortality attributable to PM2.5 that are greater than those based on fixed site ground-level
measures of urban PM2.5, but more similar to estimates based on global chemical transport model
simulations of anthropogenic PM2.5.
& 2012 Elsevier Inc. All rights reserved.
1. Introduction
Increased mortality due to exposure to fine particular matter
(PM2.5) has been documented in a number of cohort studies,
including the Harvard Six Cities Study (Dockery et al., 1993;
Krewski et al., 2000), the American Cancer Society (ACS) cohort
study (Krewski et al., 2009; Pope et al., 2002), the Veterans Aging
Cohort Study (Lipfert et al., 2006), the Medicare cohort studies
(Zeger et al., 2008), and the Nurses’ Health Study cohort (Puett
et al., 2008). Significant associations between PM2.5 and mortality
from all causes, cardiovascular disease, cardiopulmonary disease,
and lung cancer have been reported (Pope and Dockery, 2006).
Air pollution was estimated to represent 1.4% of the total
mortality attributable to 26 risk factors assessed in the previous
World Health Organization’s (WHO) global burden of disease
project (GBD) (Ezzati et al., 2002). The previous WHO GBD study
estimated that urban PM2.5 exposure was responsible for approximately
712,000 cardiopulmonary and 62,000 lung cancer deaths
in 2000 (Cohen et al., 2004). More recent estimates extending
beyond urban areas are far greater with 3.5 million cardiopulmonary
and 220,000 lung cancer annual deaths being attributed
to the anthropogenic component of PM2.5 (Anenberg et al., 2010).
Comparison of the previous and current GBD studies indicates
much higher exposure assessment due to the inclusion of satellite
remote sensing and chemical transport models to provide global
estimates (Brauer et al., 2011).
Aerosol dispersion varies spatially and temporally, and can display
both local and regional patterns, depending on their source (Chu
et al., 2003). These characteristics of aerosols limit the ability of fixed
site ground-based PM monitors to capture large-scale, regional and
global PM distributions (Gupta et al., 2006). Satellite-mounted sensors
can account for spatial variability and long-range transport of air
Contents lists available at SciVerse ScienceDirect
journal homepage: www.elsevier.com/locate/envres
Environmental Research
0013-9351/$ - see front matter & 2012 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.envres.2012.08.005
n
Corresponding author.
E-mail address: jessica.evans@mail.mcgill.ca (J. Evans).
Environmental Research 120 (2013) 33–42
pollution, and thus provide global coverage of aerosols by measuring
concentration gradients between stationary monitors, as well as in
regions that lack ground monitoring capabilities all together (EngelCox
et al., 2004).
The Moderate Resolution Imaging Spectroradiometer (MODIS)
and the Multi-angle Imaging Spectroradiometer (MISR), two instruments
onboard the National Aeronautics and Space Administration’s
Terra satellite, were designed specifically for the purpose of
retrieving aerosol data (King et al., 1999). Both MISR and MODIS
derived aerosol pollution measures have been found to be well
correlated with ground-level aerosol pollution measures under
certain conditions (Chu et al., 2003; Gupta et al., 2006; Hutchison
et al., 2005; Kahn et al., 2005; Liu et al., 2004; van Donkelaar et al.,
2006). For instance, MODIS aerosol optical depth (AOD) and daily
ground-based PM10 measures were linearly correlated, r¼0.82 at a
site in northern Italy (Chu et al., 2003); MODIS AOD and average
binned ground-based PM2.5 concentrations were linearly correlated,
r¼0.96, across major urban areas (Gupta et al., 2006); and, annual
mean MISR versus surface-level PM2.5 concentrations were linearly
correlated, r¼0.78–0.81, over the US (Liu et al., 2004). In another
study, annual mean PM2.5 concentrations obtained from MODIS and
MISR in conjunction with a global atmospheric chemical transport
model were linearly correlated with surface PM2.5 measurements in
Canada and the United States, r¼0.58 for MISR and r¼0.69 for
MODIS (van Donkelaar et al., 2006). The agreement between
satellite derived and ground based measures of PM2.5 concentration
improves when data from MISR, MODIS, and chemical transport
models are combined (van Donkelaar et al., 2010).
The present study used PM2.5 data generated from both MODIS
and MISR instruments to assess global mortality attributable to
the total as well as the anthropogenic components of PM2.5.
2. Materials and methods
2.1. Study design overview
This cross-sectional study estimated the global population fractions of adult
mortality and expected number of deaths that can be attributed to chronic PM2.5
exposure. Satellite imagery was used to derive PM2.5 concentration estimates, and
previously developed concentration–response (C–R) functions were used to calculate
relative risks of associations between PM2.5 and four causes of mortality. Risk
coefficients for mortality from all causes, cardiopulmonary disease (ICD-9 400-440;
460-519), lung cancer (ICD-9 162), and ischemic heart disease (ICD-9 410-416) in
relation to ambient PM2.5 were based on the ACS cohort study. Ischemic heart disease
(IHD), a subset of cardiopulmonary disease (CPD), was assessed as it is one of the
chronic CPD causes of mortality most strongly associated with PM (U.S. EPA, 2009).
Relative risks and mortality data were used to estimate the population fractions of
mortality attributable to chronic PM2.5 exposure.
Our methodology is based on the WHO’s 2000 GBD assessment for urban air
pollution (Cohen et al., 2004). The satellite-derived PM2.5 measures were compared
with estimates presented in the 2000 GBD and studies using similar
methods.
2.2. Global assessment of mortality attributable to PM2.5
2.2.1. Base case scenario
The following C–R function approximating chronic PM2.5 exposure to risk of
mortality was employed as the base case:
RRXÀX0
¼
expðaþbXÞ
expðaþbX0Þ
¼ exp bðXÀX0Þ
 Ã
ð1Þ
where RRXÀX0
is the relative risk at exposure ‘X’, compared to the reference
exposure ‘X0’; b is the parameter estimate for the association between chronic
PM2.5 exposure and cause of mortality (Krewski et al., 2009); X is the mean PM2.5
ambient concentration; and, X0 is the reference exposure level for PM2.5. While Eq.
(1) is an exponential model, it will be referred to as a ‘linear’ since risk appears to
be increasing linearly with PM2.5 within the lower concentration ranges observed
in this study. This was previously demonstrated in analyses of the ACS cohort,
where the shape of the C–R function was found to be linear within the ranges of
PM2.5 observed (Krewski et al., 2000; Pope et al., 2002).
In 1982, the American Cancer Society (ACS) initiated a prospective cohort
study and enrolled 1.2 million adults 30 years of age and older who were members
of households with at least one individual 45 years of age or older. Participants
completed a confidential questionnaire that included demographic, smoking
history, alcohol use, diet, and education information. Associations between
PM2.5 and mortality have been reported for three follow-up periods—a 7-year
follow-up (Pope et al., 1995), a 16-year follow-up (Pope et al., 2002), and an 18year
follow-up (Krewski et al., 2009). Risk coefficients in the current study were
derived from relative risk estimates published in the most recent analysis of the
ACS cohort study, where the confounding and modifying effects of ecological
covariates on the air pollution–mortality association was examined at various
scales with adjustment for 44 individual-level covariates as well as seven
ecological covariates, and multiple-level spatial autocorrelation was assessed
using a random effects Cox survival model (Krewski et al., 2009). For the base
case scenario, the reference exposure level was set to 5.8 mg/m3
, the lowest
concentration observed in 116 cities in the United States during the 1999–2000
PM2.5 exposure collection period (Krewski et al., 2009). Thus, for the base case, it
was assumed that 5.8 mg/m3
is the theoretical minimum threshold concentration,
below which PM2.5 has no impact on mortality.
2.2.2. Derivation of fractions of mortality attributable to PM2.5
The fraction of mortality attributable to PM2.5 at the country-level (Ostro
2004) was derived as follows:
AF ¼
RRXÀX0
À1
RRXÀX0
ð2Þ
The expected number of deaths attributable to PM2.5 for each country was
calculated by multiplying the number of country-level adult deaths due to a
particular cause by the attributable fraction (AF) for that country. Subregional and
global AFs and expected mortality estimates were derived from country-level
relative risks (based on country-level PM2.5 pollution) as well as country mortality
inputs. Attributable numbers of country-level deaths were aggregated to the
subregional and global levels and were divided by the total number of subregional
and global deaths to obtain subregional and global AFs respectively.
2.2.3. Derivation of subregional relative risks
While country-level relative risk data were used to develop subregional and
global AF and expected mortality estimates, subregional relative risks were
calculated to enable regional comparison of risk associated with PM2.5. Linear
relative risks for outcome ‘i’ in subregion ‘k’ were determined by subtracting the
reference exposure (X0) from country-level mean weighted PM2.5 concentrations
(Xc), and then multiplying each country-level Xc ÀX0 value by the proportion of the
subregion exposed at that level (Pc), where Pc¼(country ‘c’ population within
subregion ‘k’)/(subregion ‘k’ population). Country-level Pcn(Xc ÀX0) values were
then summed at the subregional level, giving a weighted subregional exposure
average, termed Ck; in addition, Ck values were used in the linear risk function to
obtain the subregional relative risk RRk ¼ exp binCk
À Á
.
2.3. Fine particulate matter (PM2.5) concentrations
Detailed information on the satellite-derived PM2.5 concentrations employed
in this work has been published elsewhere (van Donkelaar et al., 2010). PM2.5
concentrations were estimated by combining total-columnar aerosol optical depth
(AOD) measures with the GEOS-Chem chemical transport model (www.geoschem.org;
Bey et al., 2001) to account for factors that affect the relationship
between AOD and PM2.5. Satellite-derived PM2.5 concentrations were standardized
to 35% relative humidity for consistency with ground-based monitors. GEOS-Chem
simulations were used to determine the fraction of the PM2.5 data from natural
sources such as dust, and to a smaller extent, sea salt. Ambient PM2.5 concentrations
were produced in formats that included the dust and sea salt components,
PM2.5_Total, as well as formats that excluded the natural dust components of PM2.5,
PM2.5_No Dust, based upon simulated aerosol speciation. Removal of the dust
component of PM2.5 resulted in concentrations that approximated the anthropogenic
components of PM2.5.
As described in van Donkelaar et al. (2010), a global 0.11 Â 0.11 map of ambient
country-level PM2.5 concentrations was developed based on a composite average
of MODIS and MISR total columnar aerosol measurements taken during 2001–
2006. To have been included, a given pixel must have had a total of at least 50
valid measurements over the 6-year observation period. The data were averaged
within country boundaries and weighted according to population distributions
using a 2.50
 2.50
global map of 2005 population density cells from the Socioeconomic
Data and Applications Center (Gridded Population of the World version
3; http://sedac.ciesin.columbia.edu/). The PM2.5 concentrations derived by these
methods were shown to significantly agree with global ground based measurements,
with spatial correlation coefficients ranging from 0.77 to 0.83 (van Donkelaar et al.,
2010).
J. Evans et al. / Environmental Research 120 (2013) 33–4234
2.4. Mortality data
Country-level mortality estimates were required for the subregional and
global attributable fraction calculations. Detailed country-level mortality data
for adults aged 20 years and older was obtained from the WHO Mortality Database
for 132 countries; data from 2005 or the next closest year of available data was
obtained (Mortality, ICD-7-10; http://www.who.int/whosis/mort/download/en/
index.html). Mortality data from the 2004 WHO GBD update, which is modeled
for 192 countries, was used to calculate the expected numbers of deaths
attributable to PM2.5 (Mortality and morbidity; http://www.who.int/healthinfo/
global_burden_disease/estimates_regional/en/index.html).
2.5. World Health Organization subregions
Data were structured to respect WHO subregional categories. In the WHO’s
GBD analyses, 192 member states are divided into five mortality strata: (A) very
low child mortality, low adult mortality; (B) low child mortality, low adult
mortality; (C) low child mortality, high adult mortality; (D) high child mortality,
high adult mortality; and (E) high child mortality, very high adult mortality (WHO,
2004). These five strata are applied to the six WHO regions: Africa (AFR), Americas
(AMR), South-East Asia (SEAR), Europe (EUR), Eastern Mediterranean (EMR), and
Western Pacific (WPR).
3. Results
3.1. Fine particulate matter concentrations
Fig. 1 shows a global comparison of satellite derived PM2.5_Total
and PM2.5_No Dust concentrations for 231 countries. Country-level
geographic and population coverage of the PM2.5 data from
132 countries used in the study was generally good, with median
coverage of 97% and 99% respectively. Mean weighted PM2.5_Total
country concentrations ranged from 2.8 mg/m3
to 51.5 mg/m3
,
while the range of mean weighted PM2.5_No Dust country concentrations
was 1.0–47.4 mg/m3
. Table 1 reveals the estimated mean
weighted subregional total and anthropogenic PM2.5 concentrations
along with country-level mortality estimates by WHO subregion.
3.2. Global assessment of mortality attributable to PM2.5
Subregional relative risks for total PM2.5 calculated under the
base case scenario are presented in Table 2 for all four causes of
mortality of interest. Countries in the Mediterranean regions
(EMR) had PM2.5 concentrations with large natural dust components.
Subregions WPR-B, EMR-D, and EMR-B respectively had
the highest relative risk estimates for all four causes of death for
PM2.5_Total. However, once the natural dust component of PM2.5
was removed, the relative risks for EMR-B and EMR-D became
comparable to estimates for other subregions, while those of
WPR-B remain large.
The global fraction of mortality attributable to PM2.5_Total (95%
CI) was 7.1% (4.1–9.8) for all causes, 12.1% (8.0–15.8) for cardiopulmonary
disease, 16.8% (7.8–24.1) for lung cancer, and 17.5%
(12.3–21.8) for ischemic heart disease, under the base case. The
global fraction of mortality attributable to PM2.5_No Dust was 4.5%
(2.6–6.2) for all causes, 8.0% (5.3–10.5) for cardiopulmonary
disease, 12.8% (5.9–18.5) for lung cancer, and 9.4% (6.6–11.8) for
ischemic heart disease.
Fig.1. Global satellite-derived PM2.5_Total (top) and PM2.5_No Dust (bottom) measures averaged over 2001–2006. Data was standardized to 35% relative humidity, and
adjusted for sampling to better represent a true annual sample (adapted from van Donkelaar et al., 2010).
J. Evans et al. / Environmental Research 120 (2013) 33–42 35
The fractions of mortality attributable to PM2.5_No Dust were
23–46% lower than those associated with PM2.5_Total. Eastern
Mediterranean (EMR) and Western Pacific (WPR) subregions
had the largest AFs of mortality associated with PM2.5, while the
African (AFR) subregions had the smallest. The mortality attributable
to PM2.5 for the EMR subregions was predominately related
to the non-anthropogenic component of PM2.5, while the AF of
mortality associated with PM2.5 in the WPR subregions was
predominantly associated with the anthropogenic component.
The WPR-B subregion, which includes several countries with
large populations and high PM2.5 pollution levels such as China,
South Korea, and Taiwan, contributed greatly to global estimates
of mortality associated with PM2.5. Removal of this subregion
from the analysis decreased base case global estimates of mortality
by 50% for all cause mortality, 59% for cardiopulmonary
disease mortality, 64% for lung cancer mortality, and 31% for
ischemic heart disease mortality.
The numbers of adult deaths attributable to PM2.5 in 2004 was
over 3.3 million for all causes, 2.5 million for cardiopulmonary
disease, 1.3 million for ischemic heart disease, and 222,000 for
lung cancer, under the base case. Attributable mortality estimates
decreased with removal of the natural dust component, with
2.1 million all cause, 1.6 million cardiopulmonary disease,
677,000 ischemic heart disease, and 170,000 lung cancer deaths
estimated to be attributable to the PM2.5_No Dust.
3.3. Sensitivity analysis
The impact of three analytic choices – the shape of the C–R
function as well as lower and upper threshold effects – was
explored through sensitivity analyses. Table 3 summarizes global
all cause, lung cancer, cardiopulmonary disease and ischemic
heart disease deaths for adults attributable to PM2.5 under all nine
scenarios considered.
3.3.1. Concentration–response function
When extrapolating beyond the highest PM2.5 concentrations
observed in a cohort study, 30 mg/m3
(Dockery et al., 1993;
Krewski et al., 2009), it has been suggested that use of a linear
function may be inappropriate since it could overestimate mortality;
in this regard, the use of a log-linear exposure function has
been recommended such that its slope would flatten at higher
PM2.5 concentrations (Ostro 2004).
Table 1
Satellite derived mean weighted PM2.5 concentrations and all cause (AC), lung cancer (LC), cardiopulmonary disease (CPD) and ischemic heart disease (IHD) mortality data
for 132 countries summarized by WHO subregion.
WHO region WHO
subregion
Number of countries
represented
2005 population
(millions)
PM2.5_Total
(lg/m3
)a
PM2.5_No Dust
(lg/m3
)a
Number of adults deaths (000s)b
AC CPD LC IHD
Africa AFR-D 5 2 7.0 1.3 17 7 c
2
AFR-E 2 59 6.2 5.3 780 202 6 16
Americas AMR-A 3 339 11.3 10.1 2762 1202 182 502
AMR-B 41 462 7.9 5.9 2864 1035 53 280
AMR-D 5 68 9.5 8.2 301 89 3 13
Eastern EMR-B 5 96 24.4 4.7 128 31 1 6
Mediterranean EMR-D 2 232 38.2 11.4 456 184 2 12
Europe EUR-A 26 413 15.5 11.5 4016 1809 209 577
EUR-B 16 223 16.4 8.8 1574 922 53 286
EUR-C 9 235 10.3 8.1 3704 2241 88 1134
South-East SEAR-B 2 85 15.1 12.9 566 99 8 16
Asia SEAR-D 1 0.4 6.6 3.2 1 c c c
Western WPR-A 5 155 14.8 11.4 1264 560 72 112
Pacific WPR-B 10 1,450 47.3 41.0 1519 672 65 119
World 132 3820 26.3 21.3 19,953 9051 742 3075
a
Mean subregional PM2.5 levels were weighted to 2005 country populations.
b
Data from 2005 or next closest available; adults ages Z20 years (http://www.who.int/whosis/mort/download/en/index.html).
c
Estimates o1000.
Table 2
Relative risk of mortality associated with PM2.5_Total under the base case scenario (95% confidence intervals).
WHO region WHO subregion All cause Cardiopulmonary disease Lung cancer Ischemic heart disease
Africa AFR-D 1.02 (1.01–1.04) 1.04 (1.02–1.06) 1.04 (1.02–1.07) 1.08 (1.05–1.12)
AFR-E 1.01 (1.00–1.01) 1.01 (1.01–1.01) 1.01 (1.00–1.01) 1.02 (1.01–1.02)
Americas AMR-A 1.04 (1.02–1.06) 1.07 (1.04–1.10) 1.08 (1.03–1.12) 1.15 (1.09–1.20)
AMR-B 1.02 (1.01–1.03) 1.03 (1.02–1.04) 1.04 (1.01–1.06) 1.07 (1.04–1.09)
AMR-D 1.03 (1.02–1.04) 1.05 (1.03–1.06) 1.05 (1.02–1.08) 1.10 (1.06–1.14)
Eastern EMR-B 1.15 (1.08–1.23) 1.25 (1.15–1.37) 1.28 (1.11–1.48) 1.60 (1.36–1.89)
Mediterranean EMR-D 1.28 (1.15–1.42) 1.48 (1.27–1.72) 1.54 (1.20–1.98) 2.27 (1.70–3.02)
Europe EUR-A 1.08 (1.04–1.11) 1.12 (1.07–1.18) 1.14 (1.06–1.23) 1.28 (1.17–1.39)
EUR-B 1.08 (1.05–1.12) 1.14 (1.08–1.19) 1.15 (1.06–1.25) 1.31 (1.19–1.43)
EUR-C 1.03 (1.02–1.05) 1.06 (1.03–1.08) 1.06 (1.03–1.10) 1.12 (1.08–1.17)
South-East SEAR-B 1.07 (1.04–1.11) 1.12 (1.07–1.17) 1.13 (1.05–1.22) 1.26 (1.16–1.37)
Asia SEAR-D 1.01 (1.00–1.01) 1.01 (1.01–1.01) 1.01 (1.00–1.02) 1.02 (1.01–1.03)
Western WPR-A 1.07 (1.04–1.11) 1.12 (1.07–1.17) 1.13 (1.05–1.22) 1.26 (1.16–1.37)
Pacific WPR-B 1.37 (1.19–1.57) 1.65 (1.36–2.00) 1.73 (1.26–2.39) 2.85 (1.97–4.12)
J. Evans et al. / Environmental Research 120 (2013) 33–4236
The following log-linear C–R function was explored:
RRXÀX0
¼
exp½aþg lnðX þ1ފ
exp½aþg lnðX0 þ1ފ
¼
ðX þ1Þ
ðX0 þ1Þ
!g
ð3Þ
where the parameter g represents the log-linear association between
PM2.5 exposure and mortality and was based on ACS cohort (Krewski
et al., 2009).
Log-linear models, compared to their linear counterparts,
produced higher world attributable fraction estimates. A comparison
of the linear and log-linear concentration–response functions
are shown for all four causes of death in Fig. 2. The log-linear C–R
functions for PM2.5_Total increased the attributable numbers of
deaths across all four causes by 3–18% when the reference value
was the same as the base case, and by 61–107% when the
reference value was set to 2.8 mg/m3
.
3.3.2. Minimum exposure threshold
Since the base case reference value assumes a theoretical minimum
threshold effect and ignores any potential health benefits that
Table 3
Attributable number of global adult all cause (AC), lung cancer (LC), cardiopulmonary disease (CPD) and ischemic heart disease (IHD) deaths (000s) in 2004 due to PM2.5,
under nine scenarios.
Sensitivity
scenario
Fraction
of PM2.5
Shape of
C–R curve
Reference
exposure
(lg/m3
)
Attributable deaths in 2004 (000s)a
(95% confidence interval) {% change}b
AC CPD LC IHD
Base case Total Linearc
5.8 3313 (1943–4593) 2477 (1631–3224) 222 (103–318) 1256 (887–1566)
No dust Linearc
5.8 2093 (1225–2907) {-37} 1645 (1081–2146) {-34} 170 (78–245) {-23} 677 (476–849) {-46}
Choice of C–R
function
Total Log-lineard
5.8 3591 (2278–4867) {8} 2551 (1738–3292) {3} 235 (128–328) {6} 1487 (1087–1825) {18}
No dust Log-lineard
5.8 2303 (1460–3124) {-30} 1694 (1156–2186) {-32} 186 (101–260) {-16} 840 (608-104) {-33}
Minimum
exposure
threshold
Total Linearc
2.8 4220 (2470–5859) {27} 3080 (2024–4019) {24} 264 (123–378) {19} 1669 (1175–2087) {33}
No dust Linearc
1.0 3335 (1943–4652) {1} 2472 (1615–3244) {0} 234 (108–338) {5} 1254 (871–1590) {0}
Total Log-lineard
2.8 6333 (4045–8525) {91} 4319 (2974–5525) {74} 357 (199–490) {61} 2606 (1938–314)6 {107}
No dust Log-lineard
1.0 7390 (4748–9892) {123} 4947 (3433–6282) {100} 427 (242–577) {92} 2896 (2175–3468) {131}
Maximum
exposure
threshold
(30 mg/m3
)
Total Linearc
5.8 2640 (3722–5859) {-20} 1924 (1232–2569) {-22} 176 (78–262) {-21} 1123 (772–1431) {-11}
a
Attributable deaths were calculated using modeled global mortality data from the 2004 WHO GBD update (Mortality and morbidity; http://www.who.int/healthinfo/
global_burden_disease/estimates_regional/en/index.html).
b
Percent change compared to the base case for PM2.5_Total.
c
Linear risk coefficients (95% CI): AC—0.00751 (0.00421À0.01089); CPD—0.01205 (0.00742À0.01672); LC—0.01328 (0.00554À0.02103); IHD—0.02523
(0.01630À0.03415) (Krewski et al., 2009).
d
Log-linear risk coefficients (95% CI): AC—0.12019 (0.07407À0.16776); CPD—0.18878 (0.12343À0.25385); LC—0.21136 (0.10807À0.31562); IHD—0.39392
(0.26995À0.51239) (Krewski et al., 2009).
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1.40
1.60
0
PM2.5_Total Concentration (ug/m3)
RelativeRiskACMortality
Linear Loglinear
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1.40
1.60
1.80
2.00
RelativeRiskCPDMortality
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1.40
1.60
1.80
2.00
RelativeRiskLCMortality
0.00
0.50
1.00
1.50
2.00
2.50
3.00
3.50
RelativeRiskIHDMortality
10 20 30 40 50 60 0
PM2.5_Total Concentration (ug/m3)
10 20 30 40 50 60
0
PM2.5_Total Concentration (ug/m3)
10 20 30 40 50 60 0
PM2.5_Total Concentration (ug/m3)
10 20 30 40 50 60
Linear Loglinear
Linear LoglinearLinear Loglinear
Fig. 2. Comparison of linear and log-linear concentration–response functions for the relative risk of all cause (AC), cardiopulmonary disease (CPD), lung cancer (LC), and
ischemic heart disease (IHD) mortality associated with PM2.5_Total.
J. Evans et al. / Environmental Research 120 (2013) 33–42 37
may arise when reducing PM2.5 concentrations below 5.8 mg/m3
,
several sensitivity scenarios were explored where the reference value
was set to the lowest levels observed for PM2.5_Total and PM2.5_No Dust
concentrations, 2.8 mg/m3
and 1.0 mg/m3
, respectively. This was
explored for the total and anthropogenic components of PM2.5, for
both the linear and log-linear models.
Decreasing the theoretical minimum reference value from
5.8 mg/m3
to 2.8 mg/m3
for PM2.5_Total, and from 5.8 mg/m3
to
1.0 mg/m3
for PM2.5_No Dust resulted in higher attributable fractions.
Decreasing the reference value for PM2.5_Total to 2.8 mg/m3
produced increases of 19–33% in attributable mortality estimates.
3.3.3. Maximum exposure threshold
To assess the impact of extrapolating risk beyond 30 mg/m3
,
the highest ambient PM2.5 exposure level observed in both the
ACS cohort (1979–1983) and the Harvard Six Cities study (1979–
1985), a scenario was run where excess risk was constrained to be
no greater than that associated with a concentration of 30 mg/m3
.
Maximizing concentrations to 30 mg/m3
resulted in an 11–22%
decrease of in the expected number of global deaths due to PM2.5.
However, these decreases were only observed for three subregions,
EMR-B/D and WPR-B, where country PM2.5 concentrations
of above 30 mg/m3
were observed.
Overall, the range of attributable mortality was large across all
nine scenarios. Attributable mortality estimates ranged from
2.1 to 7.4 million for all causes, 1.6–4.9 million for cardiopulmonary
disease, 677,000–2.9 million for ischemic heart disease,
and 170,000–427,000 for lung cancer. Subregions with the highest
PM2.5 concentrations were most affected in the sensitivity
scenarios.
3.4. Uncertainty analysis
Paciorek and Liu (EHP, 2009) discuss the uncertainty in satellitebased
PM2.5 estimates. To quantify additional uncertainty related to
PM2.5 measurements, systematic error estimates were developed by
combining sampling induced and retrieval-inherent uncertainties
for each pixel and developing a population-weighted mean of those
errors for each country (modified Figs. 6 and 7, van Donkelaar et al.
2010). The systematic error values were used to develop countrylevel
upper and lower concentration estimates for the total fraction
of PM2.5. Table 4 shows the propagation of pollutant concentration
ranges into uncertainty estimates associated with global attributable
fractions due to systematic satellite measurement error under the
base case. Uncertainty associated with risk coefficient variability and
satellite measurement error were then imputed by combining upper
and lower limits for risk coefficients with upper and lower values for
PM2.5 concentrations in the linear concentration–response function
(Table 4). The absolute magnitude of uncertainty associated with
global AF mortality estimates increased by approximately 70% for all
causes and cardiopulmonary disease, 40% for lung cancer, and 100%
for ischemic heart disease when comparing uncertainty due to risk
coefficient variability versus uncertainty due to both risk coefficient
variability and systematic satellite measurement error.
3.5. Use of model parameters from the 2000 GBD for urban air
pollution
The attributable fractions of mortality derived in our study were
larger than those derived in the 2000 GBD for urban air pollution.
Table 5 compares global and subregional AFs using the different
input factors employed in this study, relative to the 2000 GBD
analysis. When the risk coefficients as well as the reference and
maximum exposures from the 2000 GBD for urban air pollution are
applied to the PM2.5 concentration estimates employed in the
present analysis, AF estimates decrease by roughly 43% for lung
cancer and 55% for cardiopulmonary disease (Table 5, column 2).
4. Discussion
This work represents one of the first comprehensive attempts to
estimate global mortality attributable to measured PM, and specifically,
that attributable to the anthropogenic component of PM2.5. In
comparison with previous studies aiming to estimate global mortality
attributable to PM2.5, which were limited to assessing the
effects of urban air pollution due to the lack of monitors in rural
regions (Cohen et al., 2004) or which were based on simulated PM2.5
concentrations (Anenberg et al., 2010), the current assessment
employed global PM2.5 exposure measures from two satellite
instruments, representing the most comprehensive observationally
based PM2.5 data currently available at a global scale.
4.1. Total versus anthropogenic components of PM2.5
Total concentrations of PM2.5 were modeled as the base case
and served as the basis for sensitivity analysis for comparison
purposes with other important work in this area, namely the
WHO GBD for urban air pollution, thus allowing for commentary
on the use of satellite imagery in such an analysis. However, the
risk coefficients employed in this study more accurately represent
the association between mortality and the anthropogenic versus
Table 4
Uncertainty ranges for global attributable fractions (%) of mortality associated with fine particulate matter under the base case.a
Cause of
mortality
Uncertainty ranges for global attributable fractions (%)
Global
attributable
fraction (%)
Uncertainty due to variability in
risk coefficient estimatesb
Uncertainty due to systematic
satellite measurement errorc
Uncertainty due to variability in risk coefficients
and systematic satellite measurement error
All cause 7.1 4.1À9.8 5.2À9.1 3.0À12.5
Cardiopulmonary
disease
12.1 8.0À15.8 9.3À15.2 6.1À19.6
Lung cancer 6.8 7.8À24.1 13.5À20.1 6.2À28.5
Ischemic heart
disease
17.5 12.3À21.8 12.4À22.6 8.7À28.0
d
Uncertainty ranges were imputed by combining upper and lower limits for risk coefficients with upper and lower values for PM2.5 concentrations (calculated based on
systematic error) in the linear concentration–response function.
a
AF estimates for the total fraction of PM2.5 using a linear concentration–response function and reference PM2.5 concentration of 5.8 mg/m3
.
b
Risk coefficient variability is represented by the 95% confidence interval associated with the linear risk coefficient estimates (Krewski et al., 2009).
c
Systematic satellite measurement error estimates were developed by combining sampling induced and retrieval-inherent uncertainties for each pixel; estimates
were population weighted at the country-level (modified Figs. 6 and 7, van Donkelaar et al., 2010).
J. Evans et al. / Environmental Research 120 (2013) 33–4238
natural dust components of PM2.5, such that the coefficients were
based on PM2.5 measures from urban areas in the United States
where natural dust comprises only a small fraction of the total
particle mass and is primarily derived from human activities.
Thus, due to composition differences in PM2.5 from anthropogenic
versus natural sources the risk coefficients, the risk coefficients
may not be representative of the risks associated with the natural
dust component of PM2.5. While the natural dust component of
PM2.5 is associated with health effects, pollution control policies
would be intrinsically different for natural versus anthropogenic
components of PM2.5.
4.2. Sensitivity analyses
4.2.1. Concentration–response function
The log-linear model was introduced into the 2000 GBD for
urban air pollution because it was believed to be a reasonable way
to characterize excess risk at higher concentrations, such that the
rate of increase, while still increasing monotonically with PM2.5,
decreases at higher concentrations (Cohen et al., 2004; Pope et al.,
2002), as has been observed for daily mortality in time-series
studies (Schwartz et al., 2002). The log-linear function was proposed
as an alternative to the linear model when modeling health risk
beyond the highest air pollution levels observed in epidemiological
studies in which, up to those levels, risk was observed to increase
approximately linearly with increasing PM2.5 concentrations. In the
2000 GBD for urban air pollution, the burden estimates for subregions
with low exposure levels were 63% higher in log-linear
versus linear scenarios, while burden estimates for subregions with
high exposures remained unchanged or were slightly reduced
(Cohen et al., 2004). Similar results were found in this study, where
AFs of mortality were substantially greater under the log-linear
versus linear scenarios for subregions with relatively low concentrations
of PM2.5. However, global AFs of mortality were only slightly
greater with the log-linear compared to linear models since those
subregions that contributed most to global AF estimates were those
with large populations and high pollution levels that had lower AFs
under the log-linear scenario.
4.2.2. Minimum exposure threshold
Selection of the theoretical minimum reference exposure was a
critical part of this assessment as previous estimates of mortality
attributable to PM2.5 have been sensitive to this variable (Cohen et al.,
2004). As would be expected, decreasing the minimum exposure
threshold from 5.8 mg/m3
, to 2.8 mg/m3
for PM2.5_Total and 1.0 mg/m3
for PM2.5_No Dust resulted in higher attributable fraction estimates.
When modeling risk for concentrations below 5.8 mg/m3
, we assumed
that the relationship between PM2.5 and mortality was either linear or
log-linear. However, given the limited evidence regarding the relationship
between PM2.5 risk of mortality at low PM2.5 concentrations,
it is difficult to comment on the accuracy of these increased AF
estimates with lower reference values. The base case avoided extrapolating
the C–R function below the concentrations observed in the
epidemiological study from which the risk coefficients were derived.
4.2.3. Maximum exposure threshold
Extrapolation of risk beyond the highest PM2.5 exposure levels
observed in the ACS cohort and Harvard Six Cities study, 30 mg/m3
,
involved supposition about the shape of the C-R function. The
base case assumed that risk increases exponentially with ambient
PM2.5 concentrations for the entire range of country-level concentrations
observed in this work. Overall, maximizing risk at
30 mg/m3
decreased AF estimates for three subregions where
country-level concentrations above this level were observed
(EMR-B/D and WPR-B), resulting in moderate decreases in global
AF estimates across all four causes of mortality.
4.3. Comparison with previous estimates of global mortality
attributable to PM2.5
Two key factors leading to differences between the current
study and the 2000 GBD for urban air pollution are related to the
risk coefficients and reference value used. The risk coefficients
employed in our study were based on a more sophisticated model
(a Cox random effects survival model with one level of clustering
and adjustment for seven ecological covariates), which resulted in
larger risk coefficients. The choice of the reference exposure also
Table 5
Attributable fractions (AF) of lung cancer and cardiopulmonary disease mortality associated with PM2.5 under various scenarios.
AF lung cancer mortality AF cardiopulmonary disease mortality
2000 GBD for urban
air pollutiona
Satellite PM2.combined 2000
GBD model parametersb
Current
analysis
2000 GBD for urban
air pollutiona
Satellite PM2.5 combined with 2000
GBD model parametersb
Current
analysis
World 5.0 9.5 16.8 3.0 5.5 12.1
AFR-D 5.0 0.4 0.8 2.0 0.5 1.2
AFR-E 2.0 0.0 1.1 1.0 0.0 1.0
AMR-A 3.0 2.9 7.0 2.0 2.2 6.4
AMR-B 4.0 0.9 2.5 3.0 0.7 2.3
AMR-D 6.0 1.9 5.3 3.0 1.2 4.3
EMR-B 4.0 13.4 23.2 3.0 10.0 21.3
EMR-D 8.0 18.7 31.0 4.0 14.0 28.6
EUR-A 2.0 6.2 12.2 1.0 4.5 11.0
EUR-B 5.0 6.3 12.4 3.0 4.8 11.6
EUR-C 3.0 2.3 6.0 2.0 1.6 5.3
SEAR-B 6.0 7.1 13.6 4.0 3.7 9.1
SEAR-D 4.0 0.0 1.1 3.0 0.0 1.0
WPR-A 4.0 6.2 11.9 3.0 4.5 10.8
WPR-B 10.0 20.3 33.3 6.0 14.1 28.4
Risk coefficients (per 1 mg/m3
) 0.00789a
0.00789a
0.01328c
0.00575a
0.00575a
0.01204c
Reference exposure (mg/m3
) 7.5 7.5 5.8 7.5 7.5 5.8
Maximum exposure (mg/m3
) 50.0 50.0 None 50.0 50.0 None
a
AF estimates and risk coefficients (linear concentration–response functions) from Cohen et al. (2004) (note: confidence intervals not reported in this work).
b
Estimates based on risk coefficients, reference exposure, and maximum exposure threshold from the 2000 GBD while using satellite derived PM2.5_Total
concentrations from the current study.
c
Risk coefficients from Krewski et al. (2009) (linear concentration–response functions).
J. Evans et al. / Environmental Research 120 (2013) 33–42 39
contributed to the differences observed between the studies. In the
base case, the 2000 GBD used the lowest observed concentration in
the 1979–1983 observation period of the ACS, 7.5 mg/m3
, while the
current study used the lowest concentration from the 1999–2000
period, 5.8 mg/m3
. When the risk coefficients and reference value
from the base case assessment of the 2000 GBD for urban air
pollution were applied to the current analysis, the AFs for lung
cancer and cardiopulmonary disease decreased to 5.5% and 9.5% for
cardiopulmonary disease and lung cancer respectively. While these
estimates are greater than those derived in the previous GBD for
urban air pollution, 3% and 5% for cardiopulmonary disease and lung
cancer respectively (Cohen et al., 2004), the remaining differences in
AFs of mortality between the current study and 2000 GBD may have
arisen as a result of including rural ambient PM2.5 pollution in the
burden estimates.
In comparison to more recent estimates of global annual mortality
attributable to anthropogenic PM2.5 (Anenberg et al., 2010), the
estimated numbers of deaths in 2004 attributable to anthropogenic
PM2.5 from our study were roughly 50% (1.6 versus 3.5 million) and
20% (170,000 versus 222,000) lower for cardiopulmonary disease
and lung cancer respectively. Despite using different terms (‘linear’ in
the current study versus ‘log-linear’ in the Anenberg et al. study),
both studies appear to have employed the same exponential C–R
function to derive relative risk estimates. Furthermore, both studies
appear to have employed the same risk coefficients for the association
between PM2.5 and mortality, and both included rural areas in
the assessment. One source of discrepancy between the two studies
relates to assumptions regarding background concentrations. When
Anenberg et al. employed the same background concentration that
was assumed in our base case (5.8 mg/m3
), the attributable numbers
of deaths in our study were approximately 35% lower for cardiopulmonary
disease (1.6 versus 2.5 million), and were negligibly different
for lung cancer (170,000 versus 164,000). The remaining differences
observed for the number of cardiopulmonary deaths attributable to
anthropogenic PM2.5 are potentially due to differences in mean
weighted anthropogenic PM2.5 concentrations in relation to cardiopulmonary
mortality rates at the country-level. The satellite derived
PM2.5 estimates differ considerably and tend to be higher than the
simulated estimates based on the MOZART chemical transport model
employed by Anenberg et al. Discrepancies in PM2.5 estimates based
on satellite measures compared to MOZART simulations could in part
be due to the finer resolution of exposure data used in this study,
whereby the resolution of PM2.5 estimates based on MOZART
simulation is 2.81Â 2.81 versus 0.11Â 0.11 for the satellite-based
estimates. The studies differed as well in the handling of missing
mortality data. Following observations that mortality rates can vary
substantially within a given WHO subregion, the current study
included only countries for which mortality data existed. In comparison,
Anenberg et al. applied subregional estimates when countrylevel
data was not available. Thus, the unique interaction between
the distribution of PM2.5 measures and mortality rates in each study
may explain the differences in attributable mortality estimates
between the studies. Ultimately, the estimates presented by both
studies are associated with a large degree of uncertainty, such that
when statistical uncertainty and the uncertainty associated with
choice of model parameters are considered, the range of estimates
presented by the two studies overlap.
4.4. Uncertainty
Estimating health effects from fine particle air pollution at the
global scale is inherently associated with uncertainty. The choice
of model parameters provided an array of global attributable
fraction estimates that ranged from a half to three quarters below
baseline estimates up to double baseline estimates, depending on
the cause of mortality. Accounting for uncertainty associated with
systematic satellite measurement error as well as variability associated
with risk coefficients increased the absolute uncertainty by
40–100%, as compared to the risk coefficient variability alone.
Additional uncertainty was associated with sampling variation in
the concentration–response estimates from the ACS compared to
other studies, represented by the variation in risk estimates among
cohort studies assessing mortality and PM (U.S. EPA, 2009). Uncertainty
is also inherent within country-level mortality rate estimates;
however, the degree to which this uncertainty influenced our
estimates for PM2.5 cannot be completely quantified.
4.5. Study limitations and future work
The distribution of causes of mortality varies substantially
throughout the world, and the risk coefficients for the association
between mortality and PM2.5 employed in this study were based on
a cohort from the U.S. population. While the risk of all cause
mortality due to PM2.5 was found to be significant in the U.S., this
association may primarily be the result of the association between
cardiopulmonary disease mortality and PM2.5, a large cause of
mortality in the U.S. population. It is recommended that the AFs of
all cause mortality associated with PM2.5 at the global level be
interpreted with caution.
The ACS cohort was a sample of individuals aged 30 years or
older with relatively high socioeconomic status (Krewski et al.,
2000). However, vulnerability to PM may vary by age, ethnicity,
socioeconomic status (SES), risk behaviors such as smoking, and
underlying disease status. For instance, education was shown to
modify the effects of long-term PM exposure on mortality in the ACS
cohort, such that individuals with lower educational attainment had
higher risks of mortality (Krewski et al., 2000, 2009; Pope et al.,
2004). The APHENA study also reported effect modification by SESrelated
variables and age on the association between acute PM
exposure and mortality (Samoli et al., 2008). Although the relationship
between acute PM exposure and mortality is relatively consistent
across various countries and continents (Anderson et al.,
2004; Samoli et al., 2008), if the effects of long-term PM exposure on
mortality are more severe among populations with lower measures
of SES, then these results may underestimate the magnitude of
health burden from air pollution since a large portion of the world’s
population lives below poverty line. We were limited in our ability
to undertake a comprehensive analysis of covariates influencing the
association between PM2.5 and mortality as the previously published
risk coefficients for the association between PM2.5 and mortality, as
well as global mortality data were not available by detailed age,
ethnicity, SES, lifestyle, and underlying disease strata. Prospective
cohort studies on PM exposure and mortality data in developing
countries are lacking, and research is required to better characterize
global burden attributable to long-term PM exposure. Despite these
limitations, the ACS cohort is one of the longest cohort studies on
PM2.5 and mortality, and it has undergone numerous sets of
independent analysis, and is regarded as one of the best available
sources of data on the risk of mortality associated with long term
exposure to PM2.5 (Industrial Economics Inc., 2006; Krewski, 2009).
A key contribution of future work would be the derivation of
more refined PM2.5 estimates. Despite improvements in remote
sensory technology over the past decade, there is still a need for
greater precision and accuracy in satellite measures and modeling
(Hoff and Christopher, 2009). As such, sources for PM2.5 concentration
estimates should include a combination of ground monitors,
satellite-based sensors, and advanced modeling techniques to supplement
areas with poor satellite coverage, such monsoonal regions
with persistent cloud. However, despite more refined satellite
derived exposure estimates, the ecological nature of this work will
likely continue to be a limitation, particularly at the global scale.
Although the use of ecological measures of air pollution exposure
J. Evans et al. / Environmental Research 120 (2013) 33–4240
can be expected to underestimate risk (Mallick et al., 2002), the use
of finer scales of geographic resolution has been shown to increase
estimates of the mortality due to air pollution in Los Angeles (Jerrett
et al., 2005), but not in New York (Krewski et al., 2009). Methods
such as population weighting, while not equivalent to personal
exposure estimates, can help to ensure that concentrations observed
in more population dense regions are given more weight.
Aside from differentiating between natural and anthropogenic
sources of PM2.5 in an attempt to address differing toxicities with
PM2.5 composition, it was essentially assumed that all PM2.5
components pose the same health risks. While it is acknowledged
that this is a limitation of this work, unfortunately the epidemiological
evidence is not sophisticated enough at this stage to look
at risk coefficients specific to anthropogenic and natural dust
components (U.S. NRC, 2009). Developing more refined risk
estimates for the association between specific PM2.5 components
and health outcomes based on good cohort data is an important
area to be addressed in future work.
Future research should also adjust for factors affecting the
association between PM2.5 and mortality that differ globally by
region. This would involve the development of regional mortality
data as well as risk coefficients stratified by socioeconomic
covariates of interest. Furthermore, the associations between
outdoor PM2.5 exposure and mortality should be considered in
conjunction with other related causes of mortality, such as
smoking and indoor air pollution.
4.6. Study strengths
The PM2.5 concentrations employed in this study were obtained
from two different satellite instruments, MISR and MODIS, via
methods that optimized the strengths of each sensor while minimizing
limitations. It is generally acknowledged that combining measures
from different sensors can reduce the uncertainty in column
retrievals and ultimately PM2.5 estimates (Hoff and Christopher,
2009). Global PM2.5 estimates were found to have good correlation
with ground-level measurements in North America and elsewhere
(van Donkelaar et al., 2010)
Spatial autocorrelation due to the complex spatial patterns in the
ACS data has been observed (Krewski et al., 2000). The Cox survival
model assumes independence of individual survival times over
space and time, an assumption that is not met if survival clusters
at the ecological level. If longevity varies spatially, even after
controlling for the appropriate risk factors, risk estimates and their
variances could be biased. Our study addresses this concern to a
certain extent through the use of risk coefficients derived from a
random effects survival model applied to the ACS cohort (Krewski
et al., 2009).
5. Conclusion
Previous studies of the health effects of air pollution have
traditionally relied on data derived from ground-level monitors.
Techniques for deriving PM2.5 from remote sensing data have
been evolving quickly, and satellite imagery technology is now at
a stage where concentrations for several major pollutants, and
specifically PM2.5, can be determined with relative accuracy. Data
from remote sensing technologies are quickly becoming useful in
environmental health research. This study demonstrated the
feasibility of using satellite derived pollution concentrations in
the global assessment of mortality risk associated with exposure
to fine particles. This approach leads to global estimates of
mortality attributable to PM2.5 that are greater than those based
on fixed site measures of urban PM2.5, but more similar to than
estimates based on global simulations of anthropogenic PM2.5,
depending on the cause of mortality assessed. Regardless of the
uncertainty inherent with analyses of global mortality in association
air pollution, fine particulate air pollution is an important
contributor to worldwide variations in rates of mortality.
Competing financial interests
The authors declare they have no actual or potential competing
financial interests.
Acknowledgments
A.v D. was supported by graduate fellowships from the Natural
Sciences and Engineering Research Council of Canada (NSERC)
and the Killam Trust. DK is the NSERC Chair in Risk Science at the
University of Ottawa.
Appendix A. Supporting information
Supplementary data associated with this article can be found in
the online version at http://dx.doi.org/10.1016/j.envres.2012.08.005.
References
Anderson, R., Atkinson, R., Peacock, J., Marston, L., Konstantinou, K., 2004. MetaAnalysis
of Time-Series Studies and Panel Studies of Particulate Matter (PM)
and Ozone (O3). Department of Community Health Sciences, St. George’s
Hospital Medical School, London, United Kingdom. Report of a WHO task
group. Available at: /http://www.euro.who.int/__data/assets/pdf_file/0004/
74731/e82792.pdfS (accessed 10.10.10).
Anenberg, S.C., Horowitz, L.W., Tong, D.Q., West, J.J., 2010. An estimate of the
global burden of anthropogenic ozone and fine particulate on premature
human mortality using atmospheric modelling. Environ. Health Perspect. 118,
1189–1195.
Bey, I., Jacob, D.J., Yantosca, R.M., Logan, J.A., Field, B., Fiore, A.M., et al., 2001.
Global modeling of tropospheric chemistry with assimilated meteorology:
model description and evaluation. J. Geophys. Res. 106, 23,073–23,095.
Brauer, M., Amann, M., Burnett, R., Cohen, A., Dentener, F., Ezzati, M., et al., 2011.
Exposure assessment for estimation of the global burden of disease attributable
to outdoor air pollution. Environ. Sci. Technol. 46, 652–660, http://dx.doi.
org/10.1021/es2025752. (online 8 December 2011).
Chu, D.A., Kaufman, Y.J., Zibordi, G., Chern, J.D., Mao, J., Li, C., et al., 2003. Global
monitoring of air pollution over land from the earth observing system-terra
moderate resolution imaging spectroradiometer (MODIS). J. Geophys. Res. 108,
4661, http://dx.doi.org/10.1029/2002JD003179. (online 5 November 2003).
Cohen, A.J., Anderson, H.R., Ostro, B., Pandey, K.D., Krzyzanowski, M., Kuenzli, N.,
et al., 2004. Urban air pollution. In: Comparative Quantification of Health
Risks: Global and Regional Burden of Disease due to Selected Major Risk
Factors. World Health Organization, Geneva, Switzerland. (accessed 20.06.08).
Dockery, D.W., Pope, C.A., Xu, X., Spengler, J.D., Ware, J.H., Fay, M.E., et al., 1993.
An association between air pollution and mortality in six U.S. cities. N. Engl. J.
Med. 329, 1753–1759.
Engel-Cox, J.A., Hoff, R.M., Haymet, A.D.J., 2004. Recommendations on the use of
satellite remote-sensing data for urban air quality. J. Air Waste Manage. Assoc.
54, 1360–1371.
Ezzati, M., Lopez, A.D., Rodgers, A., Hoorn, S.V., Murray, C.J.L., 2002. Selected major
risk factors and global and regional burden of disease. Lancet 360, 1347–1360.
Gupta, P., Sundar, A.C., Wang, J., Gehrig, R., Lee, Y., Kumar, N., 2006. Satellite
remote sensing of particulate matter and air quality assessment over global
cities. Atmos. Environ. 40, 5880–5892.
Hoff, R.M., Christopher, S.A., 2009. Remote sensing of particulate pollution from
space: have we reached the promised land? J. Air Waste Manage. Assoc. 59,
645–675.
Hutchison, K.D., Smith, S., Faruqui, S.J., 2005. Correlating MODIS aerosol optical
thickness data with ground-based PM2.5 observations across Texas for use in a
real-time air quality prediction system. Atmos. Environ. 39, 7190–7203.
Industrial Economics Inc., 2006. Expanded Expert Judgment Assessment of the
Concentration–Response Relationship Between PM2.5 Exposure and Mortality.
United States Environmental Protection Agency, Office of Air Quality Planning
and Standards, Research Triangle, North Carolina. Available at: /http://www.
epa.gov/ttn/ecas/regdata/Uncertainty/pm_ee_report.pdfS (accessed 15.08.08).
Jerrett, M., Burnett, R.T., Ma, R., Pope, C.A., Krewski, D., Newbold, B., Thurston, G.,
Shi, Y., Finkelstein, N., Calle, E., Thun, M., 2005. Spatial associations between
air pollution and mortality in Los Angeles. Epidemiology 16, 727–736.
J. Evans et al. / Environmental Research 120 (2013) 33–42 41
Kahn, R.A., Gaitley, B.J., Martonchik, J.V., Diner, D.J., Crean, K.A., Holben, B., 2005.
Multiple imaging spectroradiometer (MISR) global aerosol optical depth validation
based on 2 years of coincident aerosol robotic network (AERONET) observations. J.
Geophys. Res. 110, D10S04, http://dx.doi.org/10.1029/2004JD004706.
King, M.D., Kaufman, Y.J., Tanre´, D., Nakajima, T., 1999. Remote sensing of tropospheric
aerosols from space: past, present, and future. Bull. Am. Meteorol. Soc. 80 (11),
2229–2259.
Krewski, D., Burnett, R., Goldberg, M.S., Koover, K., Siemiatycki, J., Abrahamowicz, M.,
et al., 2000. Reanalysis of the Havard Six Cities study and the American Cancer
Society Study of Particulate Air Pollution and Mortality. Health Effects Institute,
Boston, MA, Available at: /http://pubs.healtheffects.org/view.php?id=6S(accessed
15.08.08).
Krewski, D., Jerrett, M., Burnett, R.T., Ma, R., Hughes, E., Shi, Y., et al., 2009.
Extended Analysis of the American Cancer Society Study of Particulate Air
Pollution and Mortality. Health Effects Institute., Boston, MA, Available at:
/http://pubs.healtheffects.org/getfile.php?u=478S.
Krewski, D., 2009. Evaluating the effects of ambient air pollution on life expectancy.
NEJM 360, 413–415.
Lipfert, F.W., Wyzga, R.E., Baty, J.D., Miller, J.P., 2006. Traffic density as a surrogate
measure of environmental exposures in studies of air pollution health effects:
long-term mortality in a cohort of US veterans. Atmos. Environ. 40, 154–169.
Liu, Y., Sarnat, J.A., Coull, B.A., Koutrakis, P., Jacob, D.J., 2004. Validation of
multiangle imaging spectroradiometer (MISR) aerosol optical thickness measurements
using aerosol robotic network (AERONET) observations over the
contiguous United States. J. Geophys. Res. 106, D06205, http://dx.doi.org/
10.1029/2003JD003981. (online March 19 2004).
Mallick, R., Fung, K., Krewski, D., 2002. Adjusting for measurement error in the Cox
proportional hazards regression model. J. Cancer Epidemiol. Prev. 7, 155–164.
Ostro, B., 2004. Outdoor Air Pollution: Assessing Environmental Burden of
Disease at National and Local Levels. World Health Organization, Geneva,
Switzerland, Available at: /http://www.who.int/quantifying_ehimpacts/publi
cations/ebd5.pdfS.
Paciorek, C.J., Liu, Y., 2009. Limitations of remotely-sensed aerosol as a spatial
proxy for fine particulate matter. Environ. Health Perspect. 117, 904–909.
Pope, C.A., Thun, M.J., Namboodiri, M.M., Dockery, D.W., Evans, J.S., Speizer, F.E.,
Heath, C.W.J., 1995. Particulate air pollution as a predictor of mortality in a
prospective study of U.S. adults. Am. J. Respir. Crit. Care Med. 151 (3), 669.
Pope, C.A., Burnett, R.T., Thun, M.J., Calle, E.E., Krewski, D., Ito, K., 2002. Lung
cancer, cardiopulmonary mortality, and long-term exposure to fine particulate
air pollution. J. Am. Med. Assoc. 287, 1132–1141.
Pope, C.A., Burnett, R.T., Thurston, G.D., Thun, M.J., Calle, E.E., Krewski, D., 2004.
Cardiovascular mortality and long-term exposure to particulate air pollution:
epidemiological evidence of general pathophysiological pathways of disease.
Circulation 109 (1), 71–77.
Pope, C.A., Dockery, D.W., 2006. Health effects of fine particulate air pollution:
lines that connect. J. Air Waste Manage. Assoc. 56, 709–742.
Puett, R.C., Schwartz, J., Hart, J.E., Yanosky, J.D., Speizer, F.E., Suh, H., et al., 2008.
Chronic particulate exposure, mortality, and coronary heart disease in the
Nurses’ Health Study. Am. J. Epidemiol. 168, 1161–1168.
Samoli, E., Peng, R., Ramsay, T., Pipikou, M., Touloumi, G., Dominici, R., et al., 2008.
Acute effects of ambient particulate matter on mortality in Europe and North
America: results from the APHENA study. Environ. Health Perspect. 116,
1480–1486.
Schwartz, J., Laden, F., Zanobetti, A., 2002. The concentration–response relation
between PM2.5 and daily deaths. Environ. Health Perspect. 110, 1025–1029.
U.S. EPA (United States Environmental Protection Agency), 2009. Integrated Health
Effects of Long-Term PM Exposure. In: Integrated Science Assessment for Particulate
Matter (final report). U.S. Environmental Protection Agency, Washington, DC.
Available at: oaspub.epa.gov/eims/eimscomm.getfile?p_download_id¼494949
(accessed February 2010).
U.S. NRC (National Research Council), 2009. Global Sources of Local Pollution: An
Assessment of Long-Range Transport of Key Pollutants to and from the United
States. National Academies Press, Washington, DC 248 pp.
van Donkelaar, A., Martin, R.V., Park, R.J., 2006. Estimating ground-level PM2.5
using aerosol optical depth determined from satellite remote sensing.
J. Geophys. Res. 111, D21201, http://dx.doi.org/10.1029/2005JD006996.
van Donkelaar, A., Randall, M., Brauer, M., Kahn, R., Levy, R., Verduzco, C., et al.,
2010. Global estimates of exposure to fine particulate matter concentrations
from satellite-based aerosol optical depth. Environ. Health Perspect. 118,
847–855, http://dx.doi.org/10.1289/ehp.0901623. (online 16 March 2010).
WHO (World Health Organization), 2004. The World Health Report 2004—Changing
History. World Health Organization, Geneva, Switzerland. (accessed 14.09.08).
Zeger, S., Dominici, F., McDermott, A., Samet, J., 2008. Mortality in the medicare
population and chronic exposure to fine particulate air pollution in urban
centers (2000-2005). Environ. Health Perspect. 116, 1614–1619, http://dx.doi.
org/10.1289/ehp.11449. (online 12 August 2008).
J. Evans et al. / Environmental Research 120 (2013) 33–4242