2.1 Modeled O₃ and PM_2.5 surface concentration

Global numerical modeling experiments initiated by TF HTAP2, the regional experiments by the Air Quality Model Evaluation International Initiative (AQMEII) over EUR and NAM, and the Model Intercomparison Study-Asia (MICS-Asia) were coordinated to perform consistent emission perturbation modeling experiments across the global, hemispheric, and continental/regional scales (Galmarini et al., 2017). Simulation periods, meteorology, emission inventories, boundary conditions, and model output are also consistent. The Joint Research Centre's (JRC) EDGAR (Emission Data Base for Global Research) team in collaboration with regional emission experts from the U.S. Environmental Protection Agency (US EPA), European Monitoring and Evaluation Programme (EMEP), Centre on Emission Inventories and Projections (CEIP), Netherlands Organization for Applied Scientific Research (TNO), and the MICS-Asia Scientific Community and Regional Emission Activity Asia (REAS) provide a global emission inventory at 0.1^∘ × 0.1^∘ resolution for TF HTAP2 modeling experiments (Janssens-Maenhout et al., 2015). The emission dataset was constructed for SO₂, NO_x, CO, non-methane volatile organic compounds, NH₃, PM₁₀, PM_2.5, black carbon (BC), and organic carbon (OC) and seven emission sectors (shipping, aircraft, land transportation, agriculture, residential, industry, and energy) for the year 2010 (Fig. S1 in the Supplement).

This study uses outputs from 14 global models/model versions (Table S1) participating in TF HTAP2. Overall, TF HTAP2 model resolutions are finer than in TF HTAP1. In TF HTAP2, each model performed a baseline simulation and sensitivity simulations in which the anthropogenic emissions in a defined source region or sector were perturbed (reduced by 20 % in most cases). Based on the number of models that simulated different experiments, we choose to focus on emission reductions from six source regions, three emission sectors, and one global domain. More specifically, all anthropogenic emissions are reduced by 20 % in the NAM, EUR, SAS, EAS, RBU, and MDE continental regions, in the PIN, TRN, and RES emission sectors globally, and in GLO (Fig. S2). Unlike TF HTAP1 (Dentener et al., 2010), which defined rectangular regions that included ocean or some sparsely inhabited regions, TF HTAP2 regions are defined by geopolitical boundaries.

We selected output from the models that provided temporally resolved volume mixing ratios of O₃ and mass mixing ratios of PM_2.5 (“mmrpm2p5”) for the baseline and at least one regional or sectoral emission reduction scenario. Among the 14 models, 11 models reported O₃ and eight reported PM_2.5 for regional emission perturbation scenarios. Four models reported O₃ and four reported PM_2.5 for sectoral emission perturbation scenarios, and 10 models reported O₃ and eight reported PM_2.5for the global emission perturbation. All models used prescribed meteorology for the year 2010, although this meteorology was derived from different (re)analysis products and was not uniform across models. Modeled concentrations are processed by calculating metrics consistent with the underlying epidemiological studies to estimate premature mortality. For O₃, we calculate the average of daily 1 h maximum O₃ concentration for the 6 consecutive months with the highest concentrations in each grid cell (Jerrett et al., 2009), for the baseline and each 20 % emission reduction scenario. While some models reported hourly O₃ metrics, others only reported daily or monthly O₃. We include these models by first calculating the ratio of the 6-month average of daily 1 h maximum O₃ to the annual average of O₃ in individual grid cells, for models reporting hourly O₃, and then applying that ratio to the annual average of ozone for those models that only report daily or monthly O₃, following Silva et al. (2013, 2016b). For PM_2.5, we calculate the annual average PM_2.5 concentration in each cell using the monthly total PM_2.5 concentrations reported by each model (“mmrpm2p5”). Model results for these two metrics are then regridded from each model's native grid resolution (varying from 0.5^∘ × 0.5^∘ to 2.8^∘ × 2.8^∘) to a consistent 0.5^∘ × 0.5^∘ resolution used in mortality estimation. We estimate regional and sectoral multi-model averages for each 20 % emission reduction scenario in the year 2010, but for each perturbation case, we only include models that report both the baseline and perturbation cases.

2.2 Model evaluation

Measurements from multiple observation networks are employed in this study to evaluate the model performance around the world. We evaluate model performance for the 2010 baseline simulation for 11 TF HTAP2 models for O₃ and eight models for PM_2.5 (Table S1). For O₃, we use ground level measurements from 2010 at 4655 sites globally, collected by the Tropospheric Ozone Assessment Report (TOAR) (Schultz et al., 2017; Young et al., 2018). The TOAR dataset identifies stations as urban, rural, and unclassified sites (Schultz et al., 2017). Model performance is evaluated for the average of daily 1 h maximum O₃ concentrations for the 3 consecutive months (3m1hmaxO₃) with the highest concentrations in each grid cell, including models that only report daily or monthly O₃ as described above. This metric for O₃ differs slightly from the 6-month average of daily 1 h maximum metric used for health impact assessment and is chosen because TOAR reports the 3-month metric but not the 6-month metric. For PM_2.5, we compare the annual average PM_2.5, using PM_2.5 observations from 2010 at 3157 sites globally selected for analysis by the Global Burden of Disease 2013 (GBD2013) (Forouzanfar et al., 2016; Brauer et al., 2016). Statistical parameters including the normalized mean bias (NMB), normalized mean error (NME), and correlation coefficient (R) are selected to evaluate model performance.

Tables S2 and S3 present statistical parameters of model evaluation for O₃ and PM_2.5, and Figs. S3–S10 show the spatial O₃ and PM_2.5 evaluation as NMB around the world and in NAM, EUR, and EAS. For 3m1hmaxO₃, the model ensemble mean shows good agreement with measurements globally with a NMB of 7.3 % and NME of 13.2 % but moderate correlation with R of 0.53 (Table S2). For individual models, eight models (CAM-Chem, CHASER_T42, CHASER_T106, EMEPrv48, GEOS-Chem Adjoint, GEOS-Chem, GFDL_AM3, and HadGEM2-ES) overestimate 3m1hmaxO₃ with NMB of 9.2 to 23 % while three models (C-IFS, OsloCTM3.v2, and RAQMS) underestimate 3m1hmaxO₃ by −10.8 to −19.4 % globally (Fig. S3). In the six perturbation regions, the model ensemble mean is also in good agreement with the measurements, with −11.2 to 25.3 % for NMB, 9.8 to 25.3 % for NME, and −0.09 to 0.98 for R. The ranges of NMB for individual models are −18.1 to 32.3 %, −24.1 to 21.3 %, −24.5 to 45.0 %, −26.4 to 24.5 %, −30.5 to 20.3 %, and −35.3 to 5.4 % in NAM, EUR, SAS, EAS, MDE, and RBU, respectively (Figs. S4–S6). Note that some regions (SAS, MDE, and RBU) have very few observations for model evaluation, making the comparison less robust. The underestimated O₃ in the western US and overestimated O₃ in the eastern US in most models are very close to the model performance results of Huang et al. (2017), who compare eight TF HTAP2 models with CASTNET observations (Fig. S4) , as well as earlier studies under HTAP1 (Fiore et al., 2009). Similarly, Dong et al. (2018) find that O₃ is overestimated in EUR and EAS by six TF HTAP2 models, consistent with our ensemble mean result in these two regions (Figs. S5–S6).

For PM_2.5, the model ensemble mean agrees well with measurements globally, with a NMB of −23.1 %, NME of 35.4 %, and R of 0.77 (Table S3). For individual models, only one model (GEOS-Chem Adjoint) overpredicts PM_2.5 by 20.3 %, while the other seven models underpredict PM_2.5 by −60.9 to −7.4 % around the world (Fig. S7). In six perturbation regions, the model ensemble mean is also in good agreement with measurements, with ranges of NMB of −49.7 to 19.4 %, 21.2 to 49.7 % for NME, and 0.50 to 1.00 for R. The ranges of NMB for individual models are −46.6 to 13.9 %, −76.0 to 31.9 %, −35.0 to 49.7 %, −50.4 to 29.5 %, −52.6 to 31.5 %, and −74.1 to −19.8 % in NAM, EUR, SAS, EAS, MDE, and RBU, respectively (Figs. S8–S10). Dong et al. (2018) show that PM_2.5 is underestimated in EUR and EAS by six TF HTAP2 models, consistent with our ensemble mean result in these two regions (Figs. S9–S10). Note that many observations used are located in urban areas, and models with coarse resolution may not be expected to have good model performance. Several models also neglect some PM_2.5 species, which may explain the tendency of models to underestimate.

2.3 Health impact assessment

We use output from the TF THAP2 model ensemble to estimate annual global O₃- and PM_2.5-related cause-specific premature mortality and avoided mortality from the 20 % regional and sectoral emission reductions, following the same methods used by Silva et al. (2016a, b). The annual O₃- and PM_2.5-related premature mortality is calculated using a health impact function based on epidemiological relationships between ambient air pollution concentration and mortality in each grid cell: $\mathrm{\Delta }M=y_{\mathrm{0}}\times \mathrm{AF}\times \mathrm{Pop}$, where ΔM is premature mortality, y₀ is the baseline mortality rate (for the exposed population), AF = $\mathrm{1}-\mathrm{1}/\mathrm{RR}$ is the attributable fraction, where RR is relative risk of death attributable to the change in air pollutant concentration (RR = 1 when there is no increased risk of death associated with a change in pollutant concentration), and Pop is the exposed population (adults aged 25 and older).

For O₃ mortality, we use a log-linear model for chronic respiratory mortality (RESP) from an American Cancer Society (ACS) study (Jerrett et al. 2009), following recent studies including the GBD (Cohen et al., 2017), but Turner et al. (2016) recently published new results for chronic ozone mortality, and adoption of these results would lead to more ozone-related deaths overall (Malley et al., 2017). RR is calculated as

where β is the concentration response factor, and Δx corresponds to the change in pollutant concentrations among simulations with perturbed emissions and the baseline simulation. For O₃, RR = 1.040 (95 % CI: 1.013–1.067 ) for a 10 ppb increase in O₃ concentrations (Jerrett et al., 2009), which from Eq. (1) gives values for β of 0.00392 (0.00129–0.00649). We estimate O₃-related premature deaths due to respiratory disease (RESP) based on decreases or increases in O₃ concentration (i.e., Δx) due to 20 % regional and sectoral emission reduction scenarios relative to the baseline. For regional and sectoral reductions, we do not assume a low-concentration threshold below which changes in O₃ have no mortality effects as there is no clear evidence for such a threshold, following Anenberg et al. (2009, 2010) and Silva et al. (2013, 2016a, b). However, we evaluate global O₃ premature mortality for the baseline 2010 simulation, relative to a counterfactual concentration of 37.6 ppb (Lim et al. 2012), for consistency with GBD estimates (Cohen et al., 2017).

For PM_2.5 mortality, we apply the IER model, which is intended to better represent the risk of exposure to PM_2.5 at locations with high ambient concentrations (Burnett et al., 2014). RR is calculated as

where z is the PM_2.5 concentration in micrograms per cubic meter and z_cf is the counterfactual concentration below which no additional risk is assumed, and the parameters α, γ, and δ are used to fit the function for cause-specific RR (Burnett et al., 2014). The overall PM_2.5-related cause-specific premature deaths related to ischemic heart disease (IHD), cerebrovascular disease (STROKE), chronic obstructive pulmonary disease (COPD), and lung cancer (LC) are estimated using RRs per age group for IHD and STROKE and RRs for all ages for COPD and LC. A uniform distribution from 5.8 to 8.8 µg m⁻³ is used for z_cf as suggested by Burnett et al. (2014), which does not vary in space nor time. For uncertainty analysis, we use results from 1000 Monte Carlo simulations of Burnett et al. (2014) to calculate RR in each grid cell using Eq. (2) or (3). We estimate avoided premature mortality in 20 % emission perturbation experiments by taking the difference in premature mortality estimates with the 2010 baseline. However, in the IER model, the concentration response function flattens off at higher PM_2.5 concentrations, yielding different estimates of avoided premature mortality for identical changes in air pollutant concentrations from less-polluted vs. highly polluted regions. That is, one unit reduction of air pollution may have a stronger effect on avoided mortality in regions where pollution levels are lower (e.g., EUR, NAM) compared with highly polluted regions (e.g., EAS, India), which would not be the case for a log-linear function (Jerrett et al., 2009; Krewski et al., 2009). Therefore, using the IER model in this study may result in smaller changes in avoided mortality in highly polluted areas than using the linear model.

For the exposed population, we use the Oak Ridge National Laboratory's LandScan 2011 Global Population Database at approximately 1 km resolution (30${}^{\prime \prime }$ × 30${}^{\prime \prime }$) (Bright et al., 2012). For the population of adults aged 25 and older, we use ArcGIS 10.2 geoprocessing tools to estimate the population per 5-year age group in each cell by multiplying the country level percentage in each age group by the population in each cell. We obtained cause-specific baseline mortality rates for 187 countries from the GBD 2010 mortality dataset (IHME, 2013). The population and baseline mortality per age group were regridded to the 0.5^∘ × 0.5^∘ grid (Table S4 and Fig. S11). Cause-specific baseline mortality rates vary geographically, e.g., RESP and COPD are relatively more dominant in SAS, IHD in EUR, STROKE in Russia, and LC in NAM.

Finally, we conduct 1000 Monte Carlo simulations to propagate uncertainty from baseline mortality rates, modeled air pollutant concentrations, and the RRs in health impact functions. We use the reported 95 % CIs for cause-specific baseline mortality rates, assuming lognormal distributions. For modeled O₃ and PM_2.5 concentrations we use the absolute value of the coefficient of variation among models in each grid cell, for each 20 % emission perturbation case minus the baseline, assuming a normal distribution. For O₃ RRs, we use the reported 95 % confidence intervals (CIs), assuming a normal distribution. For PM_2.5 RRs, we use the parameter values (i.e., α, γ, δ, and z_cf) of Burnett et al. (2014) for 1000 simulations. One should acknowledge that the range of modeled air pollution concentrations in an ensemble is not a true reflection of the uncertainty in emissions or concentrations. The mean health outcome of the 1000 Monte Carlo simulations (the empirical mean) may differ from the mean when using the mean RR.

We also quantify the uncertainties in mortality due to the spread of air pollutant concentrations across models, RRs, and baseline mortality rates, as contributors to the overall uncertainty, expressed as a coefficient, of variation and compare the result with the Monte Carlo analysis estimate. To do so, we hold two variables at their mean values and change the variable of interest within its uncertainty range; for example, using mean RRs and baseline mortality rates, we analyze the spread of the model ensemble to calculate the coefficient of variation caused by model uncertainty. Given that our 0.5^∘ × 0.5^∘ grid cell resolution can capture most of the population well in a given region, uncertainty associated with population was assumed to be negligible. We estimate the impacts of extra-regional emission reductions on mortality by using the response to extra-regional emission reduction (RERER) metric defined by TF HTAP (Galmarini et al., 2017):

where for a given region i, R_global is the change in mortality in the GLO 20 % reduction simulation relative to the base simulation, and R_region, i is the change in mortality in response to the 20 % emission reduction from that same region i. A RERER value near 1 indicates a strong relative influence of foreign emissions on mortality within a region, while a value near 0 indicates a weak foreign influence. We also estimate the total avoided extra-regional mortality from a source perspective as the sum of avoided deaths outside of each of the six source regions, and from a receptor perspective by summing $R_{\mathrm{global}}-R_{\mathrm{region},i}$ for all six regions.

3.1 Response of O₃ and PM_2.5 concentrations to 20 % regional and sectoral emission reductions

Previous TF HTAP studies reported area-averaged concentrations to quantify source–receptor relationships, averaging concentrations over a region (Doherty et al., 2013; Fiore et al., 2009; Fry et al., 2012; Huang et al., 2017; Stjern et al., 2016; Yu et al., 2013). Here, we present the population-weighted concentration over a region, which is more relevant for health. Among six receptor regions, the population-weighted multi-model mean O₃ concentrations range from 48.38 ± 8.05 ppb in EUR to 65.72 ± 10.08 ppb in SAS with a global average of 53.74 ± 8.03 ppb, while the annual population-weighted multi-model mean PM_2.5 concentrations range from 9.36 ± 2.62 µg m⁻³ in NAM to 39.27 ± 13.50 µg m⁻³ in EAS with a global average of 25.98 ± 5.05 µg m⁻³ (Tables 1 and S5–S6 and Figs. S12–S13).

Table 1Population-weighted multi-model mean O₃ (ppb) and PM_2.5 concentration (µg m⁻³) for the 2010 baseline, for the 6-month O₃ season average of 1 h daily maximum O₃ and annual average PM_2.5, shown with the standard deviation among models.

Download Print Version | Download XLSX

For 20 % perturbation scenarios, in general the impact on the multi-model mean change in surface O₃ and PM_2.5 concentration is greater within the source region (i.e., domestic region) than outside of it (i.e., foreign region) (Figs. 1–2). This is also true for individual model results (Figs. S14–S16). Among six source regions, the emission reduction from SAS has the greatest impact on global population-weighted O₃ concentration (Tables 2 and S5), while that from EAS has the greatest impact on PM_2.5 (Tables 3 and S6). The source–receptor pairs with the greatest changes in O₃ and PM_2.5 concentration reflect the geographical proximity among regions and the magnitude of emissions (Tables 2–3) – e.g., EUR → MDE (0.34 ± 0.08 ppb), EUR → RBU (0.34 ppb ± 0.09), EAS → NAM (0.29 ± 0.14 ppb), EAS → RBU (0.27 ± 0.12 ppb), and NAM → EUR (0.26 ± 0.55 ppb) for O₃, and EUR → RBU (0.26 ± 0.19 µg m⁻³), EUR → MDE (0.18 ± 0.08 µg m⁻³), MDE → SAS (0.12 ± 0.06 µg m⁻³), SAS → EAS (0.08 ± 0.08 µg m⁻³), and EAS → SAS (0.08 ± 0.07 µg m⁻³) for PM_2.5. Our ensemble shows ozone responses in the western US to emission reductions from EAS (Fig. 1c) similar to those modeled by Lin et al. (2012, 2017), who show that a model can capture the measured western US ozone increases due to rising Asian emissions.

Figure 1Global difference in multi-model mean O₃ concentrations (ppb) in 20 % emission reduction scenarios relative to the baseline for the year 2010 in (a) North America (NAM), (b) Europe (EUR), (c) East Asia (EAS), (d) South Asia (SAS), (e) the Middle East (MDE), and (f) Russia–Belarus–Ukraine (RBU), the (g) power and industry (PIN), (h) transportation (TRN), and (i) residential (RES) sectors, and (j) globally (GLO), shown for the 6-month O₃ season average of 1 h daily maximum health-relevant metric.

Download

Figure 2Global difference in multi-model annual mean PM_2.5 concentrations (µg m⁻³) in 20 % emission reduction scenarios relative to the baseline for the year 2010 in (a) North America (NAM), (b) Europe (EUR), (c) East Asia (EAS), (d) South Asia (SAS), (e) the Middle East (MDE), and (f) Russia–Belarus–Ukraine (RBU), the (g) power and industry (PIN), (h) transportation (TRN), and (i) residential (RES) sectors, and (j) globally (GLO).

Download

Table 2Population-weighted multi-model mean change in O₃ (ppb) in receptor regions due to 20 % regional (NAM, EUR, SAS, MDE, and RBU), sectoral (PIN, TRN, and RES), and global (GLO) anthropogenic emission reductions, for the 6-month O₃ season average of 1 h daily maximum. The diagonal, showing the effect of each region on itself, is shown in italics. All numbers are rounded to the nearest hundredth, and are shown with standard deviations among models.

Download Print Version | Download XLSX

Table 3Population-weighted multi-model annual average change in PM_2.5 concentrations (µg m⁻³) in receptor regions due to 20 % regional (NAM, EUR, SAS, MDE, and RBU), sectoral (PIN, TRN, and RES), and global (GLO) anthropogenic emission reductions. The diagonal, showing the effect of each region on itself, is shown in italics. All numbers are rounded to the nearest hundredth, and are shown with standard deviations among models.

Download Print Version | Download XLSX

For each receptor region, reducing foreign anthropogenic emissions by 20 % (estimated by global minus within-region reductions) can decrease population-weighted O₃ concentrations by 29–74 % of the change in O₃ concentration and 8–41 % of the change in PM_2.5 concentration (Tables 2–3). In some cases, regional emission reductions cause small O₃ concentration increases within the source region or in foreign receptors, reflecting O₃ nonlinear responses (Fig. S14). For instance, C-IFS_v2 predicts O₃ concentration increases in EUR by 0.04 ppb from domestic emission reductions, which is in agreement with results from TF HTAP1 (Anenberg et al., 2009). Similarly, CAM-Chem shows more local O₃ increases, particularly in SAS, than other models (Fig. S14). The change in O₃ concentration in foreign receptors is broader than for PM_2.5, reflecting that O₃ has a longer atmospheric lifetime than PM_2.5.

For sectors, TRN emission reductions cause the greatest decrease in global population-weighted O₃ by 1.13 ± 0.19 ppb, while PIN emission reductions cause the greatest decrease in surface PM_2.5 by 1.46 ± 0.56 µg m⁻³ globally (Tables 2–3 and Figs. 1–2). The 20 % emission reductions from individual sectors also have different effects in different regions. Of the three sectors, emission reductions from TRN have the greatest effect on population-weighted O₃ in NAM, EUR, SAS, MDE, and MDE (40–50 % of the global emission reduction) while PIN emission reductions dominate in EAS (57 %). Emission reductions from PIN have the greatest effect on population-weighted PM_2.5 in NAM, EUR, EAS, MDE, and MDE (41–84 %) while RES emission reductions dominate in SAS (43 %). The response of PM_2.5 concentration to sectoral emission reductions differs significantly across models, which reflects in part the PM_2.5 species simulated by each model (Table S1 and Figs. S15–S17). For instance, we found that models that simulate PM_2.5 nitrate (i.e., CHASER_t42 and GEOS-Chem Adjoint) predict a greater impact on PM_2.5 concentration from TRN emission reduction than those without nitrate (i.e., GOCARTv5 and SPRINTARS) (Fig. S17).

3.2 Global mortality burden associated with anthropogenic air pollution

Table 4 shows the annual multi-model mean O₃- and PM_2.5-related premature deaths in six regions and globally for the year 2010 baseline with 95 % CIs based on Monte Carlo sampling. Tables S7–S8 show estimates of premature deaths due to anthropogenic O₃ and PM_2.5 from individual models. For the ensemble model mean, we estimate 290 000 (30 000, 600 000) premature O₃-related deaths globally using a 37.6 ppb counterfactual concentration, and 2.8 million (0.5 million, 4.6 million) PM_2.5-related premature deaths using a uniform distribution of counterfactual concentration from 5.8 to 8.8 µg m⁻³. Highly populated areas of India and EAS have the highest number of O₃- and PM_2.5-related deaths, and those regions together account for 82 and 66 % of the global total O₃- and PM_2.5-related deaths. Compared with the GBD 2015 (Cohen et al., 2017), our global burden estimates are greater than the 254 000 (97 000, 422 000) premature deaths per year for O₃ from GBD, while there are less than 4.2 million (3.7 million, 4.8 million) premature deaths for PM_2.5. Lelieveld et al. (2015) estimate 142 000 (CI: 90 000, 208 000) O₃-related deaths and 3.2 million (1.5 million, 4.6 million) PM_2.5-related premature deaths for 2015. These differences can be explained mainly by exposure estimates. Here we used a multi-model ensemble, whereas Lelieveld et al. (2015) used a single model, and Cohen et al. (2017) used a single model for O₃ and a single model combined with surface and satellite observations for PM_2.5. In addition, Cohen et al. (2017) use RRs for particulate matter for IHD and stroke mortality that are modified from those used by Burnett et al. (2014) and applied age modification to the RRs, fitting the IER model for each age group separately. The updated IER with estimated higher relative risks, together with greater global pollution and baseline mortality rates in the low-income and middle-income countries in EAS and SAS, leads to the higher absolute numbers of attributable deaths and disability-adjusted life-years (DALYs) in GBD 2015 than estimated in GBD 2013 (Forouzanfar et al., 2016). Also, GBD 2015 includes lower child respiratory infection estimates whereas we do not. Our wider range of uncertainty for the global mortality reflects the uncertainty in baseline rates, RRs, and spread of air pollutant concentration across models whereas Cohen et al. (2017) consider national-level population-weighted mean concentrations and uncertainty of IER function predictions at each concentration, and Lelieveld et al. (2015) only account for the statistical uncertainty of the parameters used in the IER functions.

Table 4Annual multi-model empirical mean O₃- and PM_2.5-related premature deaths with 95 % CI from Monte Carlo simulations in parentheses (including uncertainty in baseline mortality rates, RRs, and air pollutant concentration across models) in the year 2010 baseline. All numbers are rounded to three significant figures or the nearest 100 deaths. Empirical mean is the mean of 1000 Monte Carlo simulations.

Download Print Version | Download XLSX

3.3 Effect of regional reductions on mortality

Reducing global anthropogenic emissions of air pollutant by 20 % avoids 47 400 (11 300, 99 000) O₃-related deaths and 290 000 (67 100, 405 000) PM_2.5-related premature deaths (Tables 5–6 and S9–S10). Most avoided air-pollution-related deaths were found within or close to the source region (Figs. 3–6). Reducing anthropogenic emissions by 20 % from NAM, EUR, SAS, EAS, MDE, and RBU can avoid 54, 54, 95, 85, 21, and 22 % of the global change in O₃-related deaths within the source region (the number of avoided deaths within the source region is divided by the number of avoided deaths globally), and 93, 81, 93, 94, 32, and 82 % of the global change in PM_2.5-related deaths, respectively (Tables 5–6). Whereas the most O₃-related premature deaths can be avoided by reducing SAS emissions (20 000 (3600, 42 200) deaths per year), reducing EAS emissions avoids more O₃-related premature deaths (1700 (−1300, 5400)) outside of the source region than for any other region (500 (180, 870) deaths per year to 1300 (−1200, 4400) deaths per year (Table 5). Similarly, while reducing EAS emissions avoids the most PM_2.5-related premature deaths (96 600 (3500, 136 000) deaths per year), reducing EUR emissions avoids more PM_2.5-related premature deaths (7400 (930, 9500) deaths per year) outside of the source region than for any other region (1400 (−320, 2300) deaths per year to 5500 (3 000, 7800) deaths per year) (Table 6). While emission reductions from one region generally lead to more avoided deaths within the source region than outside of it, 20 % anthropogenic emission reductions from MDE (i.e., 79 and 68 % of global avoided deaths outside the source region for O₃ and PM_2.5, respectively) and RBU (78 % for O₃) can avoid more premature deaths outside of the source region than within (Tables 5–6). This result for RBU is in agreement with West et al. (2009b). However, the results for NAM and EUR do not agree with previous studies that found that emission reductions in these regions cause more O₃-related avoided premature deaths outside of the source region than within (Anenberg et al., 2009; Duncan et al., 2008; West et al., 2009b). For PM_2.5, our results are comparable with Anenberg et al. (2014) and Crippa et al. (2017), who found that for most regions, PM_2.5-related avoided premature deaths are higher within the source region than outside. The above difference in results with TF HTAP1 may be in part because of the definition of regions. Whereas the TF HTAP2 regions are defined by geopolitical boundaries, the TF HTAP1 regions are defined by square domains which are larger and include more ocean areas (Anenberg et al., 2009). In addition, updated atmospheric models and emission inputs, as well as different atmospheric dynamics in the single years chosen in TF HTAP1 vs. TF HTAP2 may contribute to the differences.

Figure 3Annual avoided O₃-related premature deaths in 2010 per 1000 km² due to 20 % emission reduction scenarios relative to the base case in (a) North America (NAM), (b) Europe (EUR), (c) East Asia (EAS), (d) South Asia (SAS), (e) the Middle East (MDE), and (f) Russia–Belarus–Ukraine (RBU), the (g) power and industry (PIN), (h) transportation (TRN), and (i) residential (RES) sectors, and (j) globally (GLO).

Download

Table 5Annual avoided multi-model empirical mean O₃-related premature respiratory deaths with 95 % CI from Monte Carlo simulations in parentheses due to 20 % regional (NAM, EUR, SAS, MDE, and RBU), sectoral (PIN, TRN, and RES), and global (GLO) anthropogenic emission reductions in each region and worldwide. The diagonal, showing the effect of each region on itself, is shown in italics. For regional reductions, we also use the RERER (Eq. 4) as the percent of total avoided deaths in each receptor region that result from foreign emission reductions, as well as the percent of global avoided deaths from emission reductions in each source region. All numbers are rounded to three significant figures or the nearest 10 deaths.

Download Print Version | Download XLSX

Table 6Annual avoided multi-model empirical mean PM_2.5-related premature deaths (IHD + STROKE + COPD + LC) with 95 % CI from Monte Carlo simulations in parentheses due to 20 % regional (NAM, EUR, SAS, MDE, and RBU), sectoral (PIN, TRN, and RES), and global (GLO) anthropogenic emission reductions in each region and worldwide. The diagonal, showing the effect of each region on itself, is shown in italics. For regional reductions, we also use the RERER (Eq. 4) as the percent of total avoided deaths in each receptor region that result from foreign emission reductions, as well as the percent of global avoided deaths from emission reductions in each source region. All numbers are rounded to three significant figures or the nearest 10 deaths.

Download Print Version | Download XLSX

Using individual models, different conclusions may result for the relative importance of interregional transport. For example, for O₃, eight models predict that NAM emission reductions cause more O₃-related premature deaths within NAM (i.e CAM-Chem, CHASER_T42, CHASER_T106, C-IFS, GEOS-Chem Adjoint, GEOS-Chem, GFDL_AM3, and HadGEM2-ES), whereas two models predict more deaths outside NAM (i.e., EMEPrv48 and OsloCTM3.v2). Five models suggest that EUR emission reductions cause more O₃-related premature deaths within EUR (i.e., CAM-Chem, CHASER_T42, CHASER_T106, GFDL_AM3, and HadGEM2-ES), whereas four show more deaths outside (i.e., C-IFS, GEOS-Chem Adjoint, EMEPrv48, and OsloCTM3.v2). Each individual model shows that emission reductions from SAS and EAS avoid more O₃-related premature deaths within than outside, and that those from MDE and RBU avoid more O₃-related premature deaths outside than within (Fig. S18). For PM_2.5, each individual model shows that emission reductions from NAM, EUR, SAS, EAS, and RBU avoid more PM_2.5-related premature deaths within than outside, while for emission reductions from MDE, three models (EMEPrv48, GEOS-Chem Adjoint, and SPRINTARS) show more PM_2.5-related premature deaths within, while three (CHASER_T42, GEOS5, and GOCART) show more PM_2.5-related premature deaths outside (Fig. S19). The variation in health effect reflects the differences in processing of natural emissions, atmospheric physical and chemical mechanisms, numerics, etc., across models.

For each receptor region, reducing domestic anthropogenic emissions by 20 % contributes about 66, 39, 84, 72, 45, and 25 % of the total O₃-related avoided premature mortality (from the global reduction) and 90, 78, 87, 87, 58, and 66 % of the total PM_2.5-related avoided premature mortality (from the global reduction) in NAM, EUR, SAS, EAS, MDE, and RBU, respectively (Tables 5–6). Therefore, reducing emissions from foreign regions avoids more O₃ premature deaths in EUR (foreign emissions account for 61 % of total avoided deaths from the global reduction), MDE (55 %), and RBU (75 %) than reducing domestic emissions (Tables 5–6), in agreement with the results for EUR from Anenberg et al. (2009). Whereas EAS has the greatest number of avoided O₃-related premature deaths due to foreign emission reduction (3800 (3600, 3900) deaths per year), RBU has the greatest fraction of O₃ mortality from foreign emission reductions (75 %) (Table 5). Similarly, for PM_2.5, while EAS has the greatest number of avoided PM_2.5-related premature deaths due to foreign emission reductions (13 600 (3500, 18 800) deaths per year), MDE has the greatest fraction of PM_2.5 mortality from foreign emission reduction (42 %) (Table 6).

Overall, adding results from all six regional reductions, interregional transport of air pollution from extra-regional contributions is estimated to lead to more avoided deaths through changes in PM_2.5 (25 100 (8200, 35 800) deaths per year) than in O₃ (6000 (−3400, 15 500) deaths per year), consistent with Anenberg et al. (2009, 2014). This result is due to the greater influence of PM_2.5 on mortality, despite the shorter atmospheric lifetime of PM_2.5 relative to O₃.

The contributions of different factors to the overall uncertainties in mortality are shown in Tables S11–S12, considering uncertainties due to the spread of air pollutant concentrations across models, RRs, and baseline mortality rates, expressed as coefficients of variation. For both O₃ and PM_2.5 mortality, the spread of model results generally contributes most to the overall uncertainty, followed by uncertainty in RRs and in baseline mortality rates, for most source–receptor pairs. The spread of model results is generally wider for PM_2.5 (14 to 3974 % among source–receptor pairs) than for O₃ (13 to 1065 %). The uncertainty in RRs for O₃ mortality has a constant value (33 to 34 %) due to the fixed uncertainty range of RRs from Jerrett et al. (2009), whereas PM_2.5 mortality leads to a wider range of uncertainty (1 to 247 %) in RRs because the uncertainty differs at different PM_2.5 concentrations (Burnett et al., 2014). Low uncertainty in baseline mortality rate was found for most source–receptor pairs (< 20 %) except for the response of PM_2.5 mortality in SAS to 20 % reduction from RBU (66 %).

3.4 Effect of sectoral reductions on mortality

Reducing global anthropogenic emissions by 20 % in three sectors (i.e., PIN, TRN, and RES) together avoids 48 500 (7100, 108 000) O₃-related premature deaths and 243 000 (66 800, 357 000) PM_2.5-related premature deaths globally (Tables 5–6), with the greatest number of avoided air-pollution-related premature deaths in highly populated areas (e.g., NAM, EUR, India, China) (Figs. 3–6). For instance, reducing anthropogenic emissions by 20 % in three sectors together avoids the highest number of O₃-related deaths in SAS (24 000 (6000, 49 600) deaths per year) and PM_2.5-related deaths in EAS (83 400 (29 400, 135 000) deaths per year). We compare our estimates of O₃- and PM_2.5-related premature deaths attributable to PIN, TRN, and RES emissions with previous studies by multiplying our results for 20 % emission reductions by 5 and by combining their sectors to nearly match each of the three sectors in this study (Table 7). Compared with Silva et al. (2016a), our estimate of O₃- and PM_2.5-related premature deaths attributable to PIN and TRN are very comparable, but that attributable to RES is lower here. In comparison with Lelieveld et al. (2015), we estimate greater O₃- and PM_2.5-related premature deaths attributable to PIN and TRN, but fewer for RES.

Figure 4Annual avoided O₃-related premature deaths in 2010 per million people due to 20 % emission reduction scenarios relative to the base case in (a) North America (NAM), (b) Europe (EUR), (c) East Asia (EAS), (d) South Asia (SAS), (e) the Middle East (MDE), and (f) Russia–Belarus–Ukraine (RBU), the (g) power and industry (PIN), (h) transportation (TRN), and (i) residential (RES) sectors, and (j) globally (GLO).

Download

Figure 5Annual avoided PM_2.5-related premature deaths in 2010 per 1000 km² due to 20 % emission reduction scenarios relative to the base case in (a) North America (NAM), (b) Europe (EUR), (c) East Asia (EAS), (d) South Asia (SAS), (e) the Middle East (MDE), and (f) Russia–Belarus–Ukraine (RBU), the (g) power and industry (PIN), (h) transportation (TRN), and (i) residential (RES) sectors, and (j) globally (GLO).

Download

Figure 6Annual avoided PM_2.5-related premature deaths in 2010 per million people due to 20 % emission reduction scenarios relative to the base case in (a) North America (NAM), (b) Europe (EUR), (c) East Asia (EAS), (d) South Asia (SAS), (e) the Middle East (MDE), and (f) Russia–Belarus–Ukraine (RBU), the (g) power and industry (PIN), (h) transportation (TRN), and (i) residential (RES) sectors, and (j) globally (GLO).

Download

Table 7Comparison of O₃- and PM_2.5-related premature deaths attributable to PIN, TRN, and RES emissions with previous studies. Results from this study (for 20 % reductions) are multiplied by 5. For Silva et al. (2016), we combine results for “energy” and “industry” to represent PIN, and use “land transportation” to represent TRN and “residential & commercial” to represent RES. For Lelieveld et al. (2015), we combine the “power generation” and “industry” sectors to represent PIN, and use “land traffic” to represent TRN and “residential energy” to represent RES.

Download Print Version | Download XLSX

Like Silva et al. (2016a) and Lelieveld et al. (2015), different locations show relatively different mortality responses to changes in sectoral emissions. Whereas PIN emission reductions cause the greatest number of avoided O₃-related premature deaths globally (19 300 (1400, 45 000) deaths per year), TRN emission reductions cause the greatest fraction of avoided deaths in most of the six regions (26–53 % of the global emission reduction), except for EAS (58 %) and RBU (38 %), where the effect of reducing PIN emissions dominates. In comparison with other studies (Table 7), our conclusion that PIN emissions cause the most O₃-related deaths and TRN emissions cause the greatest fraction of avoided deaths in most regions agrees well with Silva et al. (2016a). For PM_2.5, reducing PIN emissions avoids the most PM_2.5-related premature deaths globally (128 000 (41 600, 179 000) deaths per year) and in most regions (38–78 % of the global emission reduction), except for SAS (45 %), where the RES emissions dominate. Although these findings differ from those of Lelieveld et al. (2015) and Silva et al. (2016), who find that RES emissions have the greatest impact on PM_2.5 mortality globally and in most regions, all studies agree that PIN emissions have the greatest impact in NAM. Our result is also comparable with Crippa et al. (2017), who find that PIN emissions have the greatest health impact in most countries. Although comparable emission inventories are used (i.e., Lelieveld et al. (2015) and this study use EDGAR emissions while Silva et al. (2016) use RCP8.5 emissions), our lower mortality estimate for RES emissions may be explained by our 20 % reductions relative to the zero-out method and the different years simulated.

Considering results from individual models, we found that mortality from TRN emission reductions shows greater relative uncertainty than from PIN or RES (Tables 5–6 and S9–S10), reflecting a greater spread of results across models. Regional impacts from individual models also differ from the ensemble mean result – e.g., for O₃, GEOS-Chem Adjoint and OsloCTM3.v2 show that reducing PIN emissions causes the greatest fraction of avoided O₃-related deaths in EUR, while GEOS-Chem Adjoint, HadGM2-ES, and OsloCTM3.v2 show that TRN emissions have the greatest fraction of avoided O₃-related deaths in RBU (Fig. S20). For PM_2.5, CHASER_t42 and GEOS-Chem Adjoint show that reducing PIN emissions causes the greatest fraction of avoided PM_2.5-related deaths in SAS (Fig. S21).