Data integration for the assessment of population ... - ACS Publications

2 from 0.64 to 0.91,. 10 reduced out-of-sample, global population-weighted Root .... Ground monitoring (GM) data consisting of annual average PM10 and...
2 downloads 0 Views 2MB Size
Subscriber access provided by University of Sussex Library

Characterization of Natural and Affected Environments

Data integration for the assessment of population exposure to ambient air pollution for global burden of disease assessment. Gavin Shaddick, Matthew Thomas, Heresh Amini, David M Broday, Aaron Cohen, Joseph Frostad, Amelia Green, Sophie Gumy, Yang Liu, Randall V Martin, Annette Prüss-Üstün, Daniel Simpson, Aaron van Donkelaar, and Michael Brauer Environ. Sci. Technol., Just Accepted Manuscript • DOI: 10.1021/acs.est.8b02864 • Publication Date (Web): 29 Jun 2018 Downloaded from http://pubs.acs.org on June 30, 2018

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 36

Environmental Science & Technology

Data integration for the assessment of population exposure to ambient air pollution for global burden of disease assessment Gavin Shaddick1,2, Matthew L. Thomas2,Heresh Amini3,4,5,David Broday6, Aaron Cohen,7,8, Joseph Frostad8, Amelia Green2, Sophie Gumy9, Yang Liu10, Randall V. Martin11,12, Annette Pruss-Ustun9, Daniel Simpson13, Aaron van Donkelaar11, Michael Brauer8,14,* 1

Department of Mathematics, University of Exeter, Exeter, EX4 4QF, UK

2

Department of Mathematical Sciences University of Bath, Bath, BA2 7AY, UK

3

Department of Epidemiology and Public Health, Swiss Tropical and Public Health Institute, Basel, 4051, Switzerland 4

University of Basel, Basel, 4051, Switzerland

5

Department of Environmental Health, Harvard T.H. Chan School of Public Health, Boston, MA 02215, USA 6

Faculty of Civil and Environmental Engineering, The Technion, Haifa, 32000, Israel

7

Health Effects Institute, Boston, MA 02110, USA

8

Institute for Health Metrics and Evaluation, Seattle, WA 98121, USA

9

World Health Organization, Geneva, 1202, Switzerland

10

Department of Environmental Health, Emory University, Rollins School of Public Health, Atlanta, GA 30322, USA 11

Department of Physics and Atmospheric Science, Dalhousie University, Halifax, Nova Scotia, B3H 4R2, Canada 11 13

Harvard-Smithsonian Center for Astrophysics, Cambridge, MA 02138, USA

Department of Statistical Sciences University of Toronto, Toronto, Ontario, M5S 3G3, Canada

14

School of Population and Public Health, The University of British Columbia, Vancouver, BC, V6T 1Z3, Canada *Corresponding author: Michael Brauer, School of Population and Public Health, The University of British Columbia, 3rd Floor – 2206 East Mall, Vancouver BC V6T1Z3 Canada [email protected] Abstract word count: 200 Manuscript word count (excluding tables and figures: 5306): 6806 (7000 max)

ACS Paragon Plus Environment

Environmental Science & Technology

Submitted to: Environmental Science and Technology

ACS Paragon Plus Environment

Page 2 of 36

Page 3 of 36

Environmental Science & Technology

TOC Art

ACS Paragon Plus Environment

Environmental Science & Technology

Page 4 of 36

1

Abstract:

2

Air pollution is a leading global disease risk factor. Tracking progress (e.g. for Sustainable

3

Development Goals) requires accurate, spatially resolved, routinely updated exposure estimates.

4

A Bayesian Hierarchical Model was developed to estimate annual average fine particle (PM2.5)

5

concentrations at 0.1° × 0.1° spatial resolution globally for 2010-2016. The model incorporated

6

spatially-varying relationships between 6003 ground measurements from 117 countries, satellite-

7

based estimates and other predictors. Model coefficients indicated larger contributions from

8

satellite-based estimates in countries with low monitor density. Within and out-of-sample cross-

9

validation indicated improved predictions of ground measurements compared to previous

10

(Global Burden of Disease 2013) estimates (increased within-sample R2 from 0.64 to 0.91,

11

reduced out-of-sample, global population-weighted Root Mean Squared Error from 23 µg/m3 to

12

12 µg/m3). In 2016, 95% of the world’s population lived in areas where ambient PM2.5 levels

13

exceeded the World Health Organization 10 µg/m3 (annual average) Guideline; 58% resided in

14

areas above the 35 µg/m3 Interim Target-1. Global population-weighted PM2.5 concentrations

15

were 18% higher in 2016 (51.1 µg/m3) than in 2010 (43.2 µg/m3), reflecting in particular

16

increases in populous South Asian countries and from Saharan dust transported to West Africa.

17

Concentrations in China were high (2016 population-weighted mean: 56.4 µg/m3) but stable

18

during this period.

1

ACS Paragon Plus Environment

Page 5 of 36

Environmental Science & Technology

19

Introduction

20

Air pollution is a major risk factor for global health. The Global Burden of Disease (GBD)

21

estimated that in 2016 ambient fine particulate matter (PM2.5) was the 6th highest ranking risk

22

factor for global mortality1. It was estimated that 4.1 million deaths annually can be attributed to

23

exposure to ambient PM2.5, with an additional 234,000 deaths attributable to exposure to ambient

24

ozone. Given this impact, air pollution has been identified as a global health priority in the

25

sustainable development agenda, with metrics related to air pollution exposure and health burden

26

now tracked as components of several Sustainable Development Goals (SDGs)2,3. Assessment

27

of the global risks to health and progress in reducing these risks requires accurate estimates of

28

population exposures to PM2.5 for all countries of the world, and increasingly also for sub-

29

national regions, for example those incorporated into the recent analyses of the GBD4.

30 31

Traditionally, the information used in analyses of health risks associated with air pollution has

32

come from ground monitoring (GM) networks, however there are many regions in which

33

monitoring is sparse5, including many low and middle income countries where air pollution

34

levels and disease burdens are high and increasing. Therefore, GM needs to be supplemented

35

with information from other sources to create a comprehensive set of population exposures6.

36

Even in areas where there are relatively dense GM networks, the use of information from

37

satellite retrievals of aerosol optical depth and land use has been shown to be profitable in filling

38

in spatial and temporal gaps in GM networks7,8. In order to assess progress towards meeting the

39

SDGs and to facilitate comparisons between countries, it is important that the methods used to

40

estimate concentrations, population-exposures and SDG metrics be globally consistent and allow 2

ACS Paragon Plus Environment

Environmental Science & Technology

41

for assessment of uncertainties. The latter helps prioritize regions where additional information

42

on exposure is required and facilitates comparison of health burden between air pollution and

43

other risk factors.

Page 6 of 36

44 45

Estimates of exposures used in previous disease burden assessments were based on a

46

combination of satellite-based estimates and chemical transport model simulations9,10.

47

Information from these two sources were calibrated to data from GM using a globally

48

standardized calibration function based on linear regression. This enabled the production of

49

estimates at high spatial resolution with global coverage, and the use of a consistent

50

methodology over both space and time allowed temporal trends to be examined. However, there

51

were a number of limitations with this approach, the most prominent being the use of a single

52

global function to calibrate satellite-based and chemical transport model estimates to GM which

53

didn’t account for spatial variation in the relationships between GM and information from the

54

other two sources.

55 56

While requiring minimal modelling assumptions, the use of a single calibration function tended

57

to underestimate GM in some locations, for example those with high winter-time and/or night-

58

time emissions where satellite retrievals were restricted either at night or seasonally due to more

59

frequent cloud or snow cover. The spatial coherence of biases related to geophysical fields such

60

as PM2.5 composition (van Donkelaar et al.11), together with increases in the number of GM

61

locations worldwide, means that spatially-varying calibration functions are now feasible. A

62

recent example was described by van Donkelaar et al.11 who used geographically weighted 3

ACS Paragon Plus Environment

Page 7 of 36

Environmental Science & Technology

63

regression with predictors including simulated PM2.5 composition to allow variation in

64

calibration between GM and estimates PM2.5 from satellite remote sensing. This analysis showed

65

improvements in the proportion of explained variance (of the GM measurements) compared to

66

uncalibrated satellite-based estimates. Under the auspice of the Global Platform on Air Quality

67

and Health led by the World Health Organization (WHO)12, a global collaboration between

68

research institutions, international, regional and national agencies was initiated with the aim of

69

addressing the above mentioned issues and reducing population exposure to air pollution from

70

particulate matter.

71 72

In this paper, the methodology used to estimate population level exposures to air pollution for

73

recent global disease burden assessments of the GBD (GBD 201513,14 and 20161) and those

74

produced by WHO15 is presented. This methodology allows spatially-varying calibration

75

functions to be fit within a Bayesian hierarchical framework and builds on the developments

76

presented in van Donkelaar et al.11 by incorporating geophysical predictors. Comparisons with

77

the methods used in previous releases (GBD 2013) of the GBD9 are given, with formal

78

evaluation based on the ability to accurately predict GM measurements. A description of

79

estimated concentrations and population exposures for 2016 is given together with an

80

examination of recent trends, from 2010 to 2016, at the country level.

81 82

4

ACS Paragon Plus Environment

Environmental Science & Technology

Page 8 of 36

83

Materials and Methods

84

Ground monitoring (GM) data consisting of annual average PM10 and PM2.5 concentrations from

85

6,003 monitors (3,275 PM2.5 and 2,728 PM10 that were used to estimate PM2.5) were used to

86

calculate the city-level averages reported in the WHO ‘Air Pollution in Cities’ database16.(Figure

87

S1). The majority of measurements were recorded in 2013 and 2014 as there is a lag in reporting

88

measurements and, at the time the database was created, limited data were available for 2015.

89

Where data were not available for 2014 (2760 monitors), data were used from 2015 (18

90

monitors), 2013 (2155), 2012 (564), 2011 (60), 2010 (375), 2009 (49), 2008 (21) and 2006 (1).

91

For locations measuring only PM10, a locally derived conversion factor was used to estimate

92

PM2.5 as described in Brauer et al.9. Monitors with > 75% temporal coverage within a single year

93

were included. A variety of monitoring techniques and operating conditions were used but such

94

information was not available in the database and could not be explicitly included in the

95

modeling.

96

Geophysical satellite-based estimates of annual average surface PM2.5 developed by van

97

Donkelaar et al.11, that combine aerosol optical depth retrievals from multiple satellites with

98

information from the GEOS-Chem17 chemical transport model, were used at 0.1௢ ×

99

0.1௢ resolution (ca. 11 x 11 km resolution at the equator). These were obtained from van

100

Donkelaar et al.18 Estimates of the sum of sulfate, nitrate, ammonium and organic carbon

101

concentrations and the concentrations of mineral dust in PM2.5 simulated using the GEOS-Chem

102

chemical transport model, and a measure combining elevation and the distance to the nearest

103

urban land surface were also available at the same resolution11. 5

ACS Paragon Plus Environment

Page 9 of 36

Environmental Science & Technology

104 105

A comprehensive set of population data on a high-resolution grid was obtained from the Gridded

106

Population of the World (GPW v4) database.19 These data were originally provided at

107

0.0417௢ × 0.0417௢ resolution and then aggregated to 0.1௢ × 0.1௢ resolution to match the grid

108

cells of the satellite-based estimates, as described in the supporting information of Forouzanfar et

109

al, 201614.

110 111

PM2.5

112

In GBD 201010 and GBD 20139 exposure estimates were obtained using a single global function

113

to calibrate available ground measurements to a ‘fused’ estimate of PM2.5 that was calculated for

114

each 0.1௢ × 0.1௢ grid cell. This fused estimate comprised the mean of estimates of annual

115

average PM2.5 from remote sensing satellites and from the TM5-FASST chemical transport

116

model20. This allowed information from both sources to be utilized, was computationally

117

efficient, reduced the amount of missing data (if any) and provided a simple method of

118

smoothing extreme values from either of the sources of data. However, the two estimation inputs

119

were different quantities, with different error structures, and it was not possible to interpret the

120

individual contributions that the variables made to the model, or to separate their contributions in

121

areas where one might be more accurate than the other. In addition, the fused exposure estimates

122

underestimated ground measurements in specific locations (see discussion in Brauer et al.,

123

20169). This underestimation was largely due to the use of a single, global, calibration function,

124

where in reality the relationship between ground measurements and other variables varied

125

spatially. Figure S2 shows differences between the global calibration function used in GBD 2013 6

ACS Paragon Plus Environment

Environmental Science & Technology

126

Page 10 of 36

with those that would be used if the calibration function was fit on a region-by-region basis.

127 128

Van Donkelaar et al.11 used geographically weighted regression (GWR) to allow for spatially-

129

varying calibration functions in this setting. GWR extends the linear regression model by

130

allowing coefficients to vary locally over space and is often used as an exploratory technique to

131

examine spatial variation and non-stationarity. It comprises fitting a series of regression models,

132

one for each of the GM locations, using data from a set of surrounding points21. The definition of

133

‘surrounding points’ may be determined on a measure of distance, e.g. points within a certain

134

distance, or by estimating an optimal bandwidth according to pre-specified criteria. Using each

135

set of data, weighted regression is performed with the weights determined by the distance of the

136

locations from the central point. However, it may be difficult to interpret location-specific

137

coefficients and the construction of confidence intervals for coefficients and prediction intervals

138

for the response variable may be problematic22.

139 140

The recently developed Data Integration Model for Air Quality (DIMAQ), the basis for the

141

production of the estimated concentrations described in this manuscript, is implemented within a

142

Bayesian Hierarchical modelling (BHM) framework23. BHMs provide a flexible framework in

143

which to model complex relationships and dependencies in data. Uncertainty can also be

144

propagated through the model allowing uncertainty arising from different components, both data

145

sources and models, to be incorporated within estimates of uncertainty associated with the final

146

estimates. All modeling is based upon annual average concentrations.

147 7

ACS Paragon Plus Environment

Page 11 of 36

Environmental Science & Technology

148

Within DIMAQ, coefficients in the calibration model relating satellite-based estimates with GM

149

were estimated for each country. Where data were insufficient within a country, information was

150

‘borrowed’ from a higher aggregation (region), and if enough information was still not available,

151

from an even higher level (super-region). This was achieved using a hierarchy of random-effects,

152

with individual country level estimates based on a combination of information from the country,

153

its region and super-region23. The model is designed to produce the most accurate measures of

154

concentrations, and further to that population-exposures, together with valid measures of

155

uncertainty on a country-by-country basis. As such, country was selected as the unit of analysis

156

as national monitoring protocols may vary (although it is acknowledged that there may also be

157

within-country variation in the way that monitoring is performed) and because impact

158

assessment and comparative analyses are all conducted at the country level. However, the nature

159

of the hierarchy is such that boundary effects may occur between countries (or regions/super-

160

regions) where there are substantial differences in the relationships between ground monitors and

161

satellite estimates in neighboring countries which may require post-processing in order to

162

produce a smooth map.

163 164

In DIMAQ, modelling was performed on the log-scale with the unit of measurement being a grid

165

cell (with ~1.4 million in total, globally). Estimates of exposures were obtained by transforming

166

predictions from the model back to the original scale of ground measurements. The following

167

components were included in the model (for further information, see Shaddick et al, 201723):

168 169

1. Continuous explanatory variables: a. Geophysical estimate of PM2.5 (ߤ݃/݉ଷ ) from satellite remote sensing, on the log8

ACS Paragon Plus Environment

Environmental Science & Technology

170 171 172 173 174 175

Page 12 of 36

scale11. b. Estimate of the sum of sulfate, nitrate, ammonium and organic carbon concentrations. simulated using the GEOS-Chem chemical transport model11. c. Estimate of concentrations of mineral dust, simulated using the GEOS-Chem chemical transport model11 d. A measurement of the difference between the elevation at the ground measurement

176

location and the mean elevation within the GEOS-Chem simulation grid cell

177

multiplied by the inverse distance to the nearest urban land surface, on the log scale11

178 179

e. Estimate of population on the log-scale. 2. Discrete explanatory variables:

180

a. Binary variable indicting whether the location of the ground measurement is

181

specifically geocoded or whether a city-level centroid geocode was used.

182

b. Binary variable indicting whether ground measurement monitoring site type, i.e.

183

industrial, residential etc, is known. Site type characterization may, however, differ by

184

country.

185 186 187 188 189 190

c. Binary variable indicting whether measurement is directly measured PM2.5 or converted from PM10 3. Random Effects: a. Grid cell random effects on the intercept, to allow for multiple ground monitors in a grid cell. b. Country, region and super-region hierarchical random effects for the intercept 9

ACS Paragon Plus Environment

Page 13 of 36

191 192 193

Environmental Science & Technology

c. Country, region and super-region hierarchical random effects for the satellite remote sensing term. d. Country, region and super-region hierarchical random effects for the population. The

194

country level random effect allows borrowing of information from neighboring

195

countries.

196

4. Interactions:

197

a. Interactions between the binary variables and the effects of satellite remote sensing

198

b. Interactions between the binary variables and the effects of population

199 200

While both the linear regression approach (as used in GBD 20139) and the GWR approach11 only

201

provide an informal analysis of the uncertainty associated with the resulting exposure estimates,

202

DIMAQ produces a posterior distribution for each prediction location, allowing a full assessment

203

of uncertainty.

204 205

Due to both the complexity of the models and the size of the data, notably the number of spatial

206

predictions that are required, recently developed techniques that perform ‘approximate’ Bayesian

207

inference based on integrated nested Laplace approximations (INLA) were used24. Computation

208

was performed using the R interface to INLA (R-INLA25). The code is provided in the

209

Supporting Information (File S1).

210

10

ACS Paragon Plus Environment

Environmental Science & Technology

Page 14 of 36

211

The primary output of the DIMAQ is a comprehensive, global, set of exposures and country-

212

level population-weighted average exposures for a single year. The model was developed in

213

order to provide estimates for 2014, but for subsequent burden of disease calculations, e.g. GBD

214

2015 and GBD 2016, and to examine trends, estimates for other years were required. The

215

following is a description of how DIMAQ was used with the fixed set of ground measurements

216

together with values of the input variables for each year to produce annual averages for 2010-

217

2016.

218 219

Satellite estimates, populations and quantities estimated using the GEOS Chem model were

220

available for each year from 2010-2015. Population estimates for 2000, 2005, 2010, 2015 and

221

2020 were available from GPW version 4. Population values for each grid cell for 2011, 2012,

222

2012, 2013 and 2014 were obtained by interpolation (and application to 2016) using natural

223

splines with knots placed at 2000, 2005, 2010, 2015 and 2020. To provide PM2.5 concentration

224

estimates for 1990, 1995, 2000, 2005 and 2010 we applied the DIMAQ coefficients to historical

225

satellite-based PM2.5 estimates for 2000-2010, together with GEOS Chem simulations for 1990

226

and 1995.

227 228

These yearly sets of data were used as inputs for DIMAQ, enabling estimates of exposures to be

229

obtained for 2010, 2011, 2012, 2013, 2014 and 2015. Trends presented previously for GBD

230

20139 were based on comparisons of estimated concentrations in 2010 and 2013, the latter being

231

based on an extrapolation of trends in satellite-based estimates from 2010 to 2011. Using this

232

approach, observed differences between estimates for 2010 and 2013 may have been be overly 11

ACS Paragon Plus Environment

Page 15 of 36

Environmental Science & Technology

233

influenced by short-term variability in meteorology rather than longer-term trends in emissions

234

and resulting concentrations. For 2016, estimates of exposures were obtained from predictions

235

from locally-varying regression models. For each cell, a LOWESS model26 was fit to the values

236

within that grid cell over time, with a constraint placed on the rate of change between 2015 and

237

2016 to avoid unrealistic and/or unjustified extrapolation of trends. Measures of uncertainty were

238

obtained by repeating the procedure for the limits of the 95% credible intervals, again on a cell-

239

by-cell basis.

240 241

For GBD 2016, population-weighted mean concentrations, together with 95% uncertainty

242

intervals were calculated for 195 countries and territories, and for sub-national areas in Brazil,

243

China, India, Indonesia, Japan, Kenya, Saudi Arabia, South Africa, Sweden, United Kingdom

244

and the United States (File S2, Supporting Information). There are a number of ways in which

245

uncertainty intervals can be obtained and here we present the results of two approaches: the first

246

to provide consistency with the GBD; and the second utilizing the information that is available

247

within the posterior distributions that are available when using DIMAQ. Consistent with

248

previous GBD assessments9,10, in GBD 2016 the population-weighted means were calculated by

249

combining the mean value in each cell with a corresponding estimate of population followed by

250

aggregation to the national or sub-national level. Uncertainty intervals (95%) around these values

251

were obtained by combining the population estimates with concentrations arising from draws

252

from normal distributions constructed to represent the concentrations in each grid cell, where the

253

parameters (means and variances) for each cell were estimated based on summary measures of

254

the outputs of DIMAQ. This sampling was repeated 1000 times for each grid cell, with the final 12

ACS Paragon Plus Environment

Environmental Science & Technology

Page 16 of 36

255

(population-weighted) mean and 95% CI being calculated from the resulting set of 1000

256

estimates and then aggregated to sub-national or national-level population-weighted means. In

257

the second approach, uncertainty intervals were obtained by taking samples from the DIMAQ

258

posterior distributions in each grid cell and combining these to produce overall uncertainty

259

intervals for each country that include both within- and between- grid cell uncertainty.

260

Population-weighting is applied to the 2.5% and 97.5% quantiles of the posterior distributions

261

for each grid cell (with the same also applied to the means and medians of the posteriors). As the

262

second approach is based directly on the posterior distributions it more closely reflects the

263

uncertainties present in the outputs of the modelling and therefore produces wider uncertainty

264

intervals for national-level population–weighted mean concentrations.

265 266

Ozone

267

As in GBD 20139, a running 3-month average (of daily 1 h maximum values) was calculated for

268

each grid cell over a full year and the maximum of these values was selected . This metric was

269

chosen to align with epidemiologic studies of chronic exposure which typically employ a

270

seasonal (summer) average, and to account for global variation in the timing of the ozone

271

(summer) season. These estimates were simulated with the TM5-FASST27 reduced-form

272

chemical transport model at 0.1° × 0.1° for 1990, 2000, and 2010 using the same emissions data

273

sets and meteorological inputs as described previously 9. Estimates for 1995, 2005, 2015 and

274

2016 were generated by fitting a natural cubic spline. Uncertainty intervals for each country were

275

estimated by assuming a normal distribution with mean equal to the population-weighted mean

276

concentrations and a standard deviation of ±6% of the estimated concentration. As in GBD 2013, 13

ACS Paragon Plus Environment

Page 17 of 36

Environmental Science & Technology

277

given the scarcity of surface ozone measurements throughout the world and the challenges in

278

accessing hourly data from available monitoring sites to develop the desired metric, surface

279

ozone measurements were not utilized when developing the global estimates. Future iterations

280

will incorporate the recently compiled measurement database of the Tropospheric Ozone

281

Assessment Report28

282 283

Results and Discussion

284

Ground measurements were available for 6003 locations from 117 countries, a substantial

285

improvement compared to previous (GBD 2013) estimates that were based on 4037

286

measurements from 79 countries9. Of the 3275 direct PM2.5 measurements, 1046 were from

287

China, 870 from The United States and Canada, 774 from high income countries in Europe, 163

288

from Central, Eastern Europe and Central Asia, 133 from Latin America and the Caribbean, 78

289

from Australasia, 67 from high income Asia Pacific countries, 49 from Sub-Saharan Africa, 31

290

from North Africa and the Middle East, 25 from India, 21 from elsewhere in South Asia and 18

291

from elsewhere in Southeast Asia. The highest measured annual average PM2.5 concentration in

292

the GM database was 288 µg/m3 in New Delhi, India while the lowest was 2 µg/m3, in

293

Anchorage, USA. The highest PM2.5 concentration estimated from PM10 measurements was 202

294

µg/m3 in Johannesburg, South Africa with the lowest being 1 µg/m3, in Muonio, Finland and in

295

Kiruna, Sweden.

296 297

Table 1 and Figure 1 show comparisons between the linear regression approach used in GBD

14

ACS Paragon Plus Environment

Environmental Science & Technology

Page 18 of 36

298

20139 and the DIMAQ hierarchical approach23 for 2014. Average concentrations from ground

299

monitors in each region are shown, together with the within-sample root mean squared error

300

(RMSE). In addition, summaries of out-of-sample validation are shown, in the form of the results

301

from performing cross-validation using 25 combinations of training (80%) and validation (20%)

302

datasets. Validation sets were obtained by taking a stratified random sample, using sampling

303

probabilities based on the cross-tabulation of PM2.5 categories (0-24.9, 25-49.9, 50-74.9, 75-99.9,

304

100+ µg/m3) and GBD super-regions10, resulting in them having the same distribution of PM2.5

305

concentrations and super-regions as the overall set of sites. The following metrics were

306

calculated for each training/evaluation set combination: for model fit - R2 and deviance

307

information criteria (DIC, a measure of model fit for Bayesian models29); for predictive accuracy

308

- RMSE and population weighted root mean squared error (PwRMSE).

309

15

ACS Paragon Plus Environment

Page 19 of 36

Environmental Science & Technology

310 Population weighted

Ground monitor

RMSE (µg/m3)

RMSE (µg/m3)

Average (µg/m3) Region

GBD

GBD

2013

DIMAQ

2013

DIMAQ

23.2

10.9

3.7

6.4

2.7

12.0

5.6

1.7

9.7

6.0

20.4

12.8

5.2

13.9

7.1

42.9

30.6

8.2

20.1

10.8

East

62.4

37.6

16.8

23.6

14.3

Sub-Saharan Africa

50.7

19.6

6.8

38.8

32.3

South Asia

41.2

35.8

17.6

44.8

22.0

Global

27.5

16.9

6.6

23.1

12.1

High income Central Europe, Eastern Europe and Central Asia Latin America and Caribbean Southeast Asia, East Asia and Oceania North Africa / Middle

Table 1: Average concentrations from ground monitors (µg/m3), together with summary measures of predictive ability, globally and by super-region. Results are the (within-sample) root mean squared error (concentration), together with the median values of population weighted 16

ACS Paragon Plus Environment

Environmental Science & Technology

Page 20 of 36

root mean squared error from 25 validation sets (see text for details). A list of countries within each region is provided in the Supporting Information (File S1).

311

Table 1 shows the average of ground measurements together with the RMSE (from within-

312

sample validation). Compared to the model used in GBD 2013, DIMAQ resulted in an

313

improvement in the within-sample fit, an increase in R2 from 0.64 to 0.91 and marked

314

improvements in all regions. Globally the RMSE was reduced by 10.3 ߤg/݉ଷ , set against the

315

context of the global ground monitor average of 27.5 ߤg/݉ଷ . The same pattern was seen in out-

316

of-sample predictive ability where, compared to the model used in GBD 2013, DIMAQ showed

317

improved predictions of ground measurements in all super regions as shown by the median

318

values of population weighted root mean squared error (ߤg/݉ଷ ) from 25 validation sets.

319

Globally, population-weighted RMSE was 12.1 µg/m3 compared to 23.1 µg/m3 when using the

320

GBD 2013 model. Figure 1 shows the same pattern in the correlations between predicted and

321

measured values at ground monitoring locations.

322

17

ACS Paragon Plus Environment

Page 21 of 36

Environmental Science & Technology

323 Figure 1: Comparison of observed and predicted measurements, by super-region, using: (i) the GBD 2013 model, (ii) DIMAQ, for the year 2014. The red line has a slope of one and an intercept of zero. Super-regions are 1: High income (3051 monitors), 2: Southeast Asia, East Asia and Oceania (1312 monitors), 3: Central Europe, Eastern Europe and Central Asia (518 monitors), 4: South Asia (464 monitors), 5: Latin America and Caribbean (308 monitors), 6: North Africa / Middle East (273 monitors) and 7: Sub-Saharan Africa (77 monitors). 324

18

ACS Paragon Plus Environment

Environmental Science & Technology

Page 22 of 36

Figure 2: Coefficients from DIMAQ based on 2014 data, by country: the effect of (log of) the satellite-based estimates. 325 326

Figure 2 shows the DIMAQ coefficients (each a combination of country-region and super-region

327

fixed and random effects) for the satellite-based estimates for each country using data from 2014,

328

the year for which DIMAQ was developed. The size of the coefficients will reflect a number of

329

different factors with larger coefficients reflecting differences between GMs and estimates from

330

satellites, and thus a greater impact of the satellite-based estimates on the overall estimates.

331

Lower coefficients may be observed where there is little variation (in concentrations) beyond the

332

mean (represented by the intercept) and/or in areas where the magnitude of the estimates from

333

the satellites is large. Countries with higher monitor density, e.g. North America, Western

334

Europe, China, Chile, Argentina, tended to have lower coefficients. In some countries, for

335

example Mongolia and countries in sub-Saharan Africa, relatively higher coefficients for 19

ACS Paragon Plus Environment

Page 23 of 36

Environmental Science & Technology

336

population density indicate importance of this factor in the overall estimates (Supporting

337

Information, Figures S3-4). In areas where there are countries that have no monitoring data that

338

are adjacent to those where measurements are out-of-line with satellite estimates, then calibration

339

functions for neighboring countries may differ more than would be expected. As previously

340

mentioned, due to the nature of the random-effects structure in DIMAQ, this can result in the

341

coefficients for neighboring countries being substantially different, resulting in apparent

342

discontinuities in maps of estimated exposures. In order to produce a smoothed map, the

343

coefficients of the model were smoothed (over continuous space) using a thin plate spline30. The

344

resulting set of coefficients were then used to produce estimated exposures as before (Figure 3).

345

Figure 3: (Smoothed) map of estimated annual average PM2.5 concentrations in 2016 (µg/m3). Map created with QGIS Geographic Information System. QGIS Development Team (v2.18, 2016)31

20

ACS Paragon Plus Environment

Environmental Science & Technology

Page 24 of 36

346 347

Based on matching the yearly outputs from DIMAQ to estimates of population on a cell-by-cell

348

basis, it was estimated that in 2016, 95% of the world’s population lived in areas that exceeded

349

the WHO guideline of an annual average of 10 µg/m3

350

population resided in areas with PM2.5 concentrations above the WHO Interim Target 1 (IT-1 of

351

35 µg/m3); 69% lived in areas exceeding IT-2 (25 µg/m3); and 85% lived in areas exceeding IT-3

352

(15 µg/m3).

32

. Fifty eight percent of the global

353 354

Population-weighted mean concentrations in 2016 as incorporated into the GBD 2016 estimates

355

were highest in countries in north (e.g. Niger: 204 µg/m3, Egypt: 126 µg/m3, Mauritania: 124

356

µg/m3) and west (e.g. Cameroon: 140 µg/m3, Nigeria: 122 µg/m3, Burkina Faso, 111 µg/m3)

357

Africa and the Middle East (e.g. Saudi Arabia: 188 µg/m3, Qatar: 148 µg/m3, Kuwait: 111

358

µg/m3), due mainly to windblown mineral dust11. The next highest concentrations appeared in

359

South Asia due to combustion emissions from multiple sources, including household use of solid

360

fuels, coal-fired power plants, agricultural and other open burning, and industrial and

361

transportation-related sources. The annual average population-weighted PM2.5 concentrations

362

were 101 µg/m3 in Bangladesh, 78 µg/m3 in Nepal, and 76 µg/m3 in India and Pakistan. The

363

population-weighted annual average concentration in China was 56 µg/m3. Estimates of the

364

annual average population-weighted PM2.5 concentrations were lowest (≤ 8 µg/m3) in Australia,

365

Brunei, Canada, Estonia, Finland, Iceland, New Zealand, Norway, Sweden and several Pacific

366

island nations. Population-weighted mean concentrations and their 95% uncertainty intervals as

367

used in GBD 20161 for all years and countries as well as sub-national areas in Brazil, China, 21

ACS Paragon Plus Environment

Page 25 of 36

Environmental Science & Technology

368

India, Indonesia, Japan, Kenya, Saudi Arabia, South Africa, Sweden, United Kingdom and the

369

United States are presented in the Supporting Information (File S2). Within countries with sub-

370

national estimates, population-weighted annual average concentrations varied as much as 3-5-

371

fold, for example in India (27.7 µg/m3 in urban Kerala to 147.5 µg/m3 in urban Delhi), China

372

(16.4 µg/m3 in Tibet to 77.0 µg/m3 in Tianjin), Brazil (5.6 µg/m3 in Sergipe to 18.6 µg/m3 in Sao

373

Paulo), Indonesia (9.5 µg/m3 in Maluku to 27.8 µg/m3 in Riau) and Saudi Arabia (82.2 µg/m3 in

374

Tabuk to 238.8 µg/m3 in Eastern Province). Elsewhere within-country variability was lower, for

375

example in the United Kingdom (9.5 µg/m3 in Northern Ireland to 16.0 µg/m3 in Lewisham),

376

Japan (9.5 µg/m3 in Okinawa to 15.0 µg/m3 in Osaka), Kenya (9.7 µg/m3 in Taita-Taveta to 22.9

377

µg/m3 in Busia), Mexico (12.8 µg/m3 in Zacatecas to 28.0 µg/m3 in Tabasco), the United States

378

(5.4 µg/m3 in Alaska to 11.0 µg/m3 in California), and South Africa (21.8 µg/m3 in Eastern Cape

379

to 51.5 µg/m3 in Gauteng).

380 381

As shown by the evaluation, this new (DIMAQ) methodology and its spatially varying

382

calibration resulted in estimates with lower error than other approaches when estimates were

383

compared to GM data. Figure S5 shows, (for 2013 concentrations), the impact of these

384

methodologic changes using exposures previously estimated in GBD 20139. Overall, with

385

DIMAQ increases were seen in population-weighted exposures for most countries, with large

386

increases in countries heavily impacted by windblown mineral dust, e.g. Saudi Arabia, Qatar and

387

Egypt. Only small differences were evident for China, the USA and Western Europe. The global

388

population-weighted mean increased by 39% (from 31.8 µg/m3 using the GBD 2013

389

methodology to 44.2 µg/m3 when using DIMAQ with updated data), mainly driven by large 22

ACS Paragon Plus Environment

Environmental Science & Technology

390

increases in the estimate population-weighted mean concentrations in the highly-populated

391

countries of South Asia (Bangladesh, Pakistan, India).

Page 26 of 36

392 393

Using the DIMAQ methodology for 2010-2016 allowed for a consistent examination of recent

394

trends. Global population-weighted PM2.5 concentrations were 18% higher in 2016 (51.1 µg/m3)

395

than in 2010 (43.2 µg/m3). This reflects increases in air pollution levels in some of the most

396

populous countries of the world, many of them the result of continuing long term trends but also

397

of note are marked increases in West Africa, as identified by satellite data. Increased levels of

398

particulate matter during late 2015 and early 2016 driven by an episode of wind-blown mineral

399

dust from the Sahara impacted heavily populated countries in the region, such as Nigeria33.

400

Accordingly, a proportion of this global increase may prove to be transient, should similar

401

episodes not affect these heavily populated areas in the future. Figure S6 shows trends in

402

population-weighted PM2.5 concentrations for the 10 most highly populated countries in the

403

world, together with the European Union, from 2010 to 2016.

404 405

Changes between 2010 and 2016 for both PM2.5 and ozone are shown in Table 2. India, Pakistan

406

and Bangladesh, have experienced both high levels and sharply increasing trends in PM2.5

407

exposure, while the decrease seen in Nigeria between 2011 and 2014 is reversed in 2015 due to

408

the extensive dust storm at the end of that year33. Concentrations in the other highly populated

409

countries and the European Union experienced little change over this period, with high levels

410

persisting in China, although with a noteworthy stabilization of concentrations in China over this

411

period, compared to clear increases observed since 19909. Ozone trends for the same period are 23

ACS Paragon Plus Environment

Page 27 of 36

Environmental Science & Technology

412

provided for context although these should be interpreted with caution as they are based upon

413

chemical transport model simulations for 1990, 2000 and 2010; 2016 values and trends merely

414

reflect the pre-2010 trends in emissions.

Location

2010 PM2.5 2016 PM2.5

% change 2010 ozone 2016 ozone % change

GLOBAL

43.2

51.1

18.2

59.6

61.6

3.3

Bangladesh

82.4

101.0

22.6

70.0

74.6

6.6

Brazil

11.0

12.6

14.5

49.6

53.6

8.1

China

58.2

56.3

-1.4

62.8

65.7

4.6

European Union

15.2

14.9

-2.0

59.3

58.6

-1.2

India

64.6

75.8

17.3

72.0

76.6

6.4

Indonesia

14.5

16.7

15.6

42.1

44.0

4.5

Japan

12.3

13.1

6.6

59.9

61.7

3.0

Nigeria

51.5

122.5

137.7

67.5

67.3

-0.3

Pakistan

61.4

75.7

23.3

67.2

69.8

3.9

Russia

16.6

15.5

-6.6

48.2

48.3

0.2

8.6

9.2

6.7

67.6

66.3

-1.9

United States 415

Table 2: Changes in population-weighted concentration of PM2.5 and ozone for the 10 most

416

highly populated countries in the world, together with the European Union between 2010 and

417

2016.

418 419

Annual population-weighed exposures to PM2.5, and ozone for each country, together with 24

ACS Paragon Plus Environment

Environmental Science & Technology

420

associated measures of uncertainty, can be found in the Supporting Information. Results are

421

given by country and by year (2010-2016).

Page 28 of 36

422 423

Significant improvements were made in the estimation of global long-term average exposure to

424

PM2.5 when compared to approaches implemented previously. These improvements were

425

facilitated in part by a global collaboration led by the WHO12 and were based on both updated

426

data and methodological advances; more than 6000 ground measurements were used with a

427

model that allowed spatially-varying calibration with satellite-based estimates at 0.1 × 0.1°

428

resolution. The estimates of concentrations using DIMAQ, were used to produce population

429

exposures for the WHO and Global Burden of Disease 2015 and 2016 estimates of disease

430

burden attributable to ambient particulate matter. Compared to earlier exposure estimates, errors

431

in out-of-sample prediction of ground measurements were reduced and the extent of variability

432

explained in ground measurements was increased in all regions of the world. However, as there

433

is still substantial heterogeneity in the magnitude of errors and uncertainties between regions,

434

additional ground monitoring data34, especially in areas currently with limited monitoring, and

435

ongoing improvements to satellite-based estimates35 will be required to reduce errors in large

436

parts of the world. It is important to acknowledge, however, that ground measurement data are

437

included as reported by local agencies and assessed with varying levels of quality assurance. As

438

the exposure estimation methodology for PM2.5 no longer includes direct use of the TM5 or

439

TM5-FASST chemical transport model estimates, the use of satellite-based estimates (which are

440

produced using outputs from the GEOS-Chem model) that are regularly updated along with

441

ground measurements should facilitate regular updating of these exposure estimates as part of 25

ACS Paragon Plus Environment

Page 29 of 36

442

Environmental Science & Technology

SDG reporting.

443 444

Although these estimates advance those developed previously, they incorporate known

445

limitations. For example, the spatially varying calibration model was built from a specific year of

446

satellite-based estimates and included measurements spanning multiple years. The parameters

447

from this model were then applied to the satellite-based estimates for other years. As it is likely

448

that the relationship between satellite-based estimates and ground monitors would vary over time

449

future iterations will more directly model the relationships between temporally aligned satellite-

450

based estimates and ground monitors.

451 452

Future developments of DIMAQ will also incorporate a continuously varying spatial term, in

453

addition to the country structure, embedded within the model. This will facilitate two lines of

454

improvement, one related to eliminating post-processing smoothing currently performed to

455

produce smoothed maps and secondly allowing within-country variation in the calibration

456

equations. Also, additional information on land-use and topography will be included within the

457

model which should reduce errors and uncertainties. Dependent on computational considerations,

458

the fundamental approach used in DIMAQ is scalable and higher resolution (0.01 x 0.01°) global

459

exposure estimates should be feasible using finer resolution satellite-based estimates11,35. Such

460

estimates would be expected to further reduce errors in relation to ground measurements, reduce

461

spatial misalignment with population density data and improve estimation of disease burden

462

attributable to air pollution at a finer spatial scale. In addition to developments to the spatial

463

component of the model, a temporal component will be added, allowing trends to be modelled 26

ACS Paragon Plus Environment

Environmental Science & Technology

Page 30 of 36

464

directly within the model framework, and allowing dependence between years, rather than being

465

performed on time periods.

466 467

Future developments of DIMAQ will also allow the contributions from individual inputs to be

468

identified. The growing availability of information on source contributions to PM2.5 , for example

469

from simulations36, could allow an assessment of the contribution of different sources to the

470

final estimated concentrations37,38. This would require methodological advances in terms of how

471

multiple, potentially highly correlated, inputs are treated in a coherent manner within DIMAQ.

472

An important aspect of the Bayesian hierarchical modelling approach that is the basis of DIMAQ

473

is the ability to propagate uncertainty from the inputs, through the modelling process, to the final

474

estimates. The method used here, which that is based directly on the posterior distributions that

475

are an output of the DIMAQ Bayesian hierarchical model, includes more of the uncertainties

476

present in the outputs of the modelling than previous approaches and therefore produces wider

477

uncertainty intervals for national -level population–weighted mean concentrations. Future

478

developments will allow a breakdown of the overall uncertainty associated with the final

479

estimates to their individual components, including individual input variables. In addition, such

480

global estimates will benefit from ongoing improvements in understanding of AOD – surface

481

PM2.5 conversion factors, such as those derived from the Surface Particulate Matter Network

482

(SPARTAN) of collocated PM2.5 monitors and sun photometers39.

483 484

Analyses clearly show that the current disease burden attributable to ambient air pollution is

485

dominated by effects of exposure to PM2.5 in comparison to the impacts related to ozone 27

ACS Paragon Plus Environment

Page 31 of 36

Environmental Science & Technology

486

exposure1, and as such less attention has been directed towards evaluation and improvement of

487

ozone exposure estimates. However, as global ozone exposure is increasing and projected to

488

continue to increase as a result of climate change40 there is a need to enhance the estimation of

489

ozone exposures, for example by chemical transport model assimilation and fusion with the

490

newly available global ground measurement database (TOAR)28. Further, at present uncertainty

491

in ozone exposures is assumed to be a uniform standard deviation of ±6% of the estimated

492

concentration given the estimation via a single chemical transport model, without direct

493

calibration to ground measurements. Future estimation of ozone using ground measurements and

494

multiple simulations can also identify areas of greater uncertainty in estimates which may be

495

useful for prioritization of new monitoring. In addition, accumulating research suggests

496

additional health impacts related to nitrogen dioxide (NO2) exposure that is distinct from those

497

currently attributable to PM2.5 and ozone41,42. Global high resolution NO2 exposure estimates43,

498

similar to those already developed for North America44,45, western Europe46 and Australia47

499

which combine satellite retrievals, chemical transport models and land use information have

500

recently become available and should facilitate the potential inclusion of nitrogen dioxide in in

501

future global disease burden assessments.

502 503

Estimates of outdoor air pollution exposure and trends developed with globally consistent

504

methodology have facilitated policy analyses6,41,42 which should ultimately inform strategies to

505

mitigate the health impacts of air pollution exposure and reduce emissions of climate forcing

506

agents. As with prior iterations, we encourage others to use these estimates for relevant impact

507

assessment or epidemiologic analyses and we have provided the DIMAQ code (File S2) along 28

ACS Paragon Plus Environment

Environmental Science & Technology

508

with associated files of national and sub-national population-weighted (File S3) and gridded

509

estimates (linked to population data) (File S4) in the Supporting Information.

Page 32 of 36

510 511

Acknowledgements

512

Annette Pruss-Ustun and Sophie Gumy are staff members of WHO. The authors alone are

513

responsible for the views expressed in this publication and they do not necessarily represent the

514

decisions or stated policy of the World Health Organization. We thank Rita van Dingenen and

515

Frank Dentener of the European Commission Joint Research Centre for the ozone estimates from

516

TM5-FASST.

517 518

Supporting Information

519

Additional figures (Figures S1-S6), list of countries by region and super-region (File S1),

520

DIMAQ code (File S2), and a full listing of population-weighted mean and 95% uncertainty

521

interval PM2.5 and ozone concentration estimates for all countries and selected sub-national

522

entities for 2010-2016 (File S3). Gridded estimates (linked to population data, File S4), and

523

associated metadata (File S5).

29

ACS Paragon Plus Environment

Page 33 of 36

Environmental Science & Technology

524

References

525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564

(1)

(2) (3)

(4)

(5) (6) (7)

(8)

(9)

(10)

GBD 2016 Risk Factors Collaborators, A. A.; Abate, K. H.; Abbafati, C.; Abbas, K. M.; Abd-Allah, F.; Abdulle, A. M.; Abera, S. F.; Aboyans, V.; Abu-Raddad, L. J.; AbuRmeileh, N. M. E.; et al. Global, Regional, and National Comparative Risk Assessment of 84 Behavioural, Environmental and Occupational, and Metabolic Risks or Clusters of Risks, 1990-2016: A Systematic Analysis for the Global Burden of Disease Study 2016. Lancet (London, England) 2017, 390 (10100), 1345–1422. United Nations. Goal 3. Sustainable Development Knowledge Platform https://sustainabledevelopment.un.org/sdg3 (accessed May 28, 2018). GBD 2016 SDG Collaborators, A. A.; Abate, K. H.; Abbafati, C.; Abbas, K. M.; AbdAllah, F.; Abdulkader, R. S.; Abdulle, A. M.; Abera, S. F.; Aboyans, V.; Abu-Raddad, L. J.; et al. Measuring Progress and Projecting Attainment on the Basis of Past Trends of the Health-Related Sustainable Development Goals in 188 Countries: An Analysis from the Global Burden of Disease Study 2016. Lancet (London, England) 2017, 390 (10100), 1423–1459. GBD 2016 DALYs and HALE Collaborators, A. A.; Abate, K. H.; Abbafati, C.; Abbas, K. M.; Abd-Allah, F.; Abdulkader, R. S.; Abdulle, A. M.; Abebo, T. A.; Abera, S. F.; Aboyans, V.; et al. Global, Regional, and National Disability-Adjusted Life-Years (DALYs) for 333 Diseases and Injuries and Healthy Life Expectancy (HALE) for 195 Countries and Territories, 1990-2016: A Systematic Analysis for the Global Burden of Disease Study 2016. Lancet (London, England) 2017, 390 (10100), 1260–1344. Carvalho, H.; Morikawa, T.; Ito, A. The Air We Breathe: Differentials in Global Air Quality Monitoring. Lancet Respir. Med. 2016, 4 (8), 603–605. The World Bank. The Cost of Air Pollution : Strengthening the Economic Case for Action; Washington DC, 2016. Kloog, I.; Nordio, F.; Coull, B. A.; Schwartz, J. Incorporating Local Land Use Regression And Satellite Aerosol Optical Depth In A Hybrid Model Of Spatiotemporal PM 2.5 Exposures In The Mid-Atlantic States. Environ. Sci. Technol. 2012, 46 (21), 11913– 11921. Kloog, I.; Chudnovsky, A. A.; Just, A. C.; Nordio, F.; Koutrakis, P.; Coull, B. A.; Lyapustin, A.; Wang, Y.; Schwartz, J. A New Hybrid Spatio-Temporal Model for Estimating Daily Multi-Year PM2.5 Concentrations across Northeastern USA Using High Resolution Aerosol Optical Depth Data. Atmos. Environ. 2014, 95, 581–590. Brauer, M.; Freedman, G.; Frostad, J.; van Donkelaar, A.; Martin, R. V.; Dentener, F.; Dingenen, R. van; Estep, K.; Amini, H.; Apte, J. S.; et al. Ambient Air Pollution Exposure Estimation for the Global Burden of Disease 2013. Environ. Sci. Technol. 2016, 50 (1), 79–88. Brauer, M.; Amann, M.; Burnett, R. T.; Cohen, A.; Dentener, F.; Ezzati, M.; Henderson, S. B.; Krzyzanowski, M.; Martin, R. V; Van Dingenen, R.; et al. Exposure Assessment for Estimation of the Global Burden of Disease Attributable to Outdoor Air Pollution. Environ. Sci. Technol. 2012, 46 (2), 652–660. 30

ACS Paragon Plus Environment

Environmental Science & Technology

565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605

(11)

(12) (13)

(14)

(15)

(16)

(17)

(18)

(19) (20)

(21)

(22)

Page 34 of 36

van Donkelaar, A.; Martin, R. V; Brauer, M.; Hsu, N. C.; Kahn, R. A.; Levy, R. C.; Lyapustin, A.; Sayer, A. M.; Winker, D. M. Global Estimates of Fine Particulate Matter Using a Combined Geophysical-Statistical Method with Information from Satellites, Models, and Monitors. Environ. Sci. Technol. 2016, 50 (7), 3762–3772. World Health Organization. Global platform on air quality and health http://www.who.int/airpollution/global-platform/en/ (accessed May 28, 2018). Cohen, A. J.; Brauer, M.; Burnett, R.; Anderson, H. R.; Frostad, J.; Estep, K.; Balakrishnan, K.; Brunekreef, B.; Dandona, L.; Dandona, R.; et al. Estimates and 25-Year Trends of the Global Burden of Disease Attributable to Ambient Air Pollution: An Analysis of Data from the Global Burden of Diseases Study 2015. Lancet 2017, 389 (10082), 1907–1918. Forouzanfar, M. H.; Afshin, A.; Alexander, L. T.; Anderson, H. R.; Bhutta, Z. A.; Biryukov, S.; Brauer, M.; Burnett, R.; Cercy, K.; Charlson, F. J.; et al. Global, Regional, and National Comparative Risk Assessment of 79 Behavioural, Environmental and Occupational, and Metabolic Risks or Clusters of Risks, 1990–2015: A Systematic Analysis for the Global Burden of Disease Study 2015. Lancet 2016, 388 (10053), 1659– 1724. World Health Organization, 2016. Ambient air pollution: A global assessment of exposure and burden of disease http://www.who.int/phe/publications/air-pollution-globalassessment/en/ (accessed May 28, 2018). World Health Organization, 2016. WHO Global Urban Ambient Air Pollution Database (update 2016) http://www.who.int/phe/health_topics/outdoorair/databases/cities/en/ (accessed May 28, 2018). Bey, I.; Jacob, D. J.; Yantosca, R. M.; Logan, J. A.; Field, B. D.; Fiore, A. M.; Li, Q. B.; Liu, H. Y. G. Y.; Mickley, L. J.; Schultz, M. G. Global Modeling of Tropospheric Chemistry with Assimilated Meteorology: Model Description and Evaluation. J. Geophys. Res. 2001, 106 (D19), 23073–23095. Atmospheric Composition Analysis Group - Dalhousie University. Surface PM2.5 (V4.GL.02.NoGWR). http://fizz.phys.dal.ca/~atmos/martin/?page_id=140 (accessed May 28, 2018). Socioeconomic Data and Applications Center. Gridded Population of the World (GPW v4). http://sedac.ciesin.columbia.edu/data/collection/gpw-v4 (accessed May 28, 2018). Huijnen, V.; Williams, J.; van Weele, M.; van Noije, T.; Krol, M.; Dentener, F.; Segers, A.; Houweling, S.; Peters, W.; de Laat, J.; et al. The Global Chemistry Transport Model TM5: Description and Evaluation of the Tropospheric Chemistry Version 3.0. Geosci. Model Dev. 2010, 3 (2), 445–473. Fotheringham, A. S.; Charlton, M. E.; Brunsdon, C. Geographically Weighted Regression: A Natural Evolution of the Expansion Method for Spatial Data Analysis. Environ. Plan. A 1998, 30 (11), 1905–1927. Wheeler, D. C.; Waller, L. A. Comparing Spatially Varying Coefficient Models: A Case Study Examining Violent Crime Rates and Their Relationships to Alcohol Outlets and 31

ACS Paragon Plus Environment

Page 35 of 36

606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645

Environmental Science & Technology

(23)

(24)

(25) (26) (27)

(28)

(29)

(30) (31) (32) (33) (34)

(35)

(36)

(37)

Illegal Drug Arrests. J. Geogr. Syst. 2009, 11 (1), 1–22. Shaddick, G.; Thomas, M. L.; Green, A.; Brauer, M.; van Donkelaar, A.; Burnett, R.; Chang, H. H.; Cohen, A.; Dingenen, R. Van; Dora, C.; et al. Data Integration Model for Air Quality: A Hierarchical Approach to the Global Estimation of Exposures to Ambient Air Pollution. J. R. Stat. Soc. Ser. C (Applied Stat. 2017. Rue, H.; Martino, S.; Chopin, N. Approximate Bayesian Inference for Latent Gaussian Models by Using Integrated Nested Laplace Approximations. J. R. Stat. Soc. Ser. B (Statistical Methodol. 2009, 71 (2), 319–392. The R-INLA project http://www.r-inla.org/ (accessed May 28, 2018) . Cleveland, W. S. LOWESS: A Program for Smoothing Scatterplots by Robust Locally Weighted Regression. Am. Stat. 1981, 35 (1), 54. Van Dingenen, Rita; Leitao, Joana; Dentener, Frank. A Multi-Metric Global SourceReceptor Model for Integrated Impact Assessment of Climate and Air Quality Policy Scenarios. EGU Gen. Assem. 2014 2014. Chang, K.-L.; Petropavlovskikh, I.; Copper, O. R.; Schultz, M. G.; Wang, T. Regional Trend Analysis of Surface Ozone Observations from Monitoring Networks in Eastern North America, Europe and East Asia. Elem Sci Anth 2017, 5 (0), 50. Spiegelhalter, D. J.; Best, N. G.; Carlin, B. P.; van der Linde, A. Bayesian Measures of Model Complexity and Fit. J. R. Stat. Soc. Ser. B (Statistical Methodol. 2002, 64 (4), 583– 639. Wood, S. N. Generalized Additive Models : An Introduction with R; Chapman & Hall/CRC, 2006. QGIS Development Team. QGIS Geographic Information System. Open Source Geospatial Foundation Project. 2016. http://qgis.osgeo.org World Health Organization. Air Quality Guidelines. Global Update 2005. Particulate Matter, Ozone, Nitrogen Dioxide and Sulfur Dioxide; World Health Organization, 2006. Sand and Dust Storm Warning Advisory and Assessment System. WMO Airborne Dust Bulletin; 2017. Snider, G.; Weagle, C. L.; Martin, R. V.; van Donkelaar, A.; Conrad, K.; Cunningham, D.; Gordon, C.; Zwicker, M.; Akoshile, C.; Artaxo, P.; et al. SPARTAN: A Global Network to Evaluate and Enhance Satellite-Based Estimates of Ground-Level Particulate Matter for Global Health Applications. Atmos. Meas. Tech. 2015, 8 (1), 505–521. van Donkelaar, A.; Martin, R. V; Spurr, R. J. D.; Burnett, R. T. High-Resolution SatelliteDerived PM2.5 from Optimal Estimation and Geographically Weighted Regression over North America. Environ. Sci. Technol. 2015, 49 (17), 10482–10491. Lelieveld, J.; Evans, J. S.; Fnais, M.; Giannadaki, D.; Pozzer, A. The Contribution of Outdoor Air Pollution Sources to Premature Mortality on a Global Scale. Nature 2015, 525 (7569), 367–371. Health Effects Institute. 2016. Burden of Disease Attributable to Coal-Burning and Other Air Pollution Sources in China. https://www.healtheffects.org/publication/burden-disease32

ACS Paragon Plus Environment

Environmental Science & Technology

646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683

(38)

(39)

(40)

(41) (42)

(44)

(45)

(46)

(47)

(48) (49)

Page 36 of 36

attributable-coal-burning-and-other-air-pollution-sources-china (accessed May 28, 2018. Health Effects Institute. 2018. Burden of Disease Attributable to Major Air Pollution Sources in India. https://www.healtheffects.org/publication/gbd-air-pollution-india (accessed May 28, 2018). Snider, G.; Weagle, C. L.; Martin, R. V.; Van Donkelaar, A.; Conrad, K.; Cunningham, D.; Gordon, C.; Zwicker, M.; Akoshile, C.; Artaxo, P.; et al. SPARTAN: A Global Network to Evaluate and Enhance Satellite-Based Estimates of Ground-Level Particulate Matter for Global Health Applications. Atmos. Meas. Tech. 2015, 8 (1). Silva, R. A.; West, J. J.; Lamarque, J.-F.; Shindell, D. T.; Collins, W. J.; Dalsoren, S.; Faluvegi, G.; Folberth, G.; Horowitz, L. W.; Nagashima, T.; et al. The Effect of Future Ambient Air Pollution on Human Premature Mortality to 2100 Using Output from the ACCMIP Model Ensemble. Atmos. Chem. Phys. 2016, 16 (15), 9847–9862. Faustini, A.; Rapp, R.; Forastiere, F. Nitrogen Dioxide and Mortality: Review and MetaAnalysis of Long-Term Studies. Eur. Respir. J. 2014, 44 (3). World Health Organization, Regional Office for Europe. Review of evidence on health aspects of air pollution – REVIHAAP project: final technical report; World Health Organization, 2013. Larkin, A.; Geddes, J. A.; Martin, R. V.; Xiao, Q.; Liu, Y.; Marshall, J. D.; Brauer, M.; Hystad, P. Global Land Use Regression Model for Nitrogen Dioxide Air Pollution. Environ. Sci. Technol. 2017, 51 (12), 6957–6964. Hystad, P.; Setton, E.; Cervantes, A.; Poplawski, K.; Deschenes, S.; Brauer, M.; van Donkelaar, A.; Lamsal, L.; Martin, R.; Jerrett, M.; et al. Creating National Air Pollution Models for Population Exposure Assessment in Canada. Environ. Health Perspect. 2011, 119 (8), 1123–1129. Novotny, E. V; Bechle, M. J.; Millet, D. B.; Marshall, J. D. National Satellite-Based Land-Use Regression: NO2 in the United States. Environ. Sci. Technol. 2011, 45 (10), 4407–4414. Vienneau, D.; de Hoogh, K.; Bechle, M. J.; Beelen, R.; van Donkelaar, A.; Martin, R. V.; Millet, D. B.; Hoek, G.; Marshall, J. D. Western European Land Use Regression Incorporating Satellite- and Ground-Based Measurements of NO 2 and PM 10. Environ. Sci. Technol. 2013, 47 (23), 13555–13564. Knibbs, L. D.; Hewson, M. G.; Bechle, M. J.; Marshall, J. D.; Barnett, A. G. A National Satellite-Based Land-Use Regression Model for Air Pollution Exposure Assessment in Australia. Environ. Res. 2014, 135, 204–211. OECD. OECD Environmental Outlook to 2050; OECD Environmental Outlook; OECD Publishing, 2012. Zhang, Q.; Jiang, X.; Tong, D.; Davis, S. J.; Zhao, H.; Geng, G.; Feng, T.; Zheng, B.; Lu, Z.; Streets, D. G.; et al. Transboundary Health Impacts of Transported Global Air Pollution and International Trade. Nature 2017, 543 (7647), 705–709.

684

33

ACS Paragon Plus Environment