Risk-Based Prioritization among Air Pollution Control Strategies in the Yangtze River Delta, China

Background The Yangtze River Delta (YRD) in China is a densely populated region with recent dramatic increases in energy consumption and atmospheric emissions. Objectives We studied how different emission sectors influence population exposures and the corresponding health risks, to inform air pollution control strategy design. Methods We applied the Community Multiscale Air Quality (CMAQ) Modeling System to model the marginal contribution to baseline concentrations from different sectors. We focused on nitrogen oxide (NOx) control while considering other pollutants that affect fine particulate matter [aerodynamic diameter ≤ 2.5 μm (PM2.5)] and ozone concentrations. We developed concentration–response (C-R) functions for PM2.5 and ozone mortality for China to evaluate the anticipated health benefits. Results In the YRD, health benefits per ton of emission reductions varied significantly across pollutants, with reductions of primary PM2.5 from the industry sector and mobile sources showing the greatest benefits of 0.1 fewer deaths per year per ton of emission reduction. Combining estimates of health benefits per ton with potential emission reductions, the greatest mortality reduction of 12,000 fewer deaths per year [95% confidence interval (CI), 1,200–24,000] was associated with controlling primary PM2.5 emissions from the industry sector and reducing sulfur dioxide (SO2) from the power sector, respectively. Benefits were lower for reducing NOx emissions given lower consequent reductions in the formation of secondary PM2.5 (compared with SO2) and increases in ozone concentrations that would result in the YRD. Conclusions Although uncertainties related to C-R functions are significant, the estimated health benefits of emission reductions in the YRD are substantial, especially for sectors and pollutants with both higher health benefits per unit emission reductions and large potential for emission reductions.


Research
The Yangtze River Delta (YRD), which generally refers to southern Jiangsu Province, eastern and northern Zhejiang Province, and the municipality of Shanghai, is the fast est growing economic development region in China and one of the most densely popu lated regions in the world. Shanghai is one of the world's largest cities, with > 18 million longterm residents and a population den sity of > 40,000 people/km 2 in some districts. Accompanying this economic development has been a dramatic increase in energy con sumption and air pollution emissions. For example, although the Shanghai metropolitan area and the provinces of Jiangsu and Zhejiang constitute only 2% of the area of China, their emissions of sulfur dioxide (SO 2 ), nitrogen oxides (NO x ), and fine particulate matter [aero dynamic diameter ≤ 2.5 μm (PM 2.5 )] accounted for 12%, 15%, and 12%, respectively, of total emissions in China in 2006, which increased by 36%, 55%, and 14%, respectively, from 2001 to 2006 (Zhang et al. 2009). NO x emis sions are of particular concern because they have increased the fastest and are forecasted to increase even more (Chen et al. 2006).
Several studies (Kan et al. 2004;Li et al. 2004;Streets et al. 1999) have evaluated the health benefits of air pollution control in Shanghai, primarily SO 2 and PM 10 , and occasionally sulfate particles. Similar studies have been conducted in other parts of China, such as a recent estimate of annual deaths attributable to air pollution in the Pearl River Delta (Loh et al. 2008). In another study in the Pearl River Delta area, Wang et al. (2005) investigated how the emissions from different sectors influenced the concentrations of gas eous pollutants including ozone. And Wang and Mauzerall (2006) quantified the total health damages from PM due to anthropo genic emissions from Zaozhuang, Shangdong Province. Besides these studies on regional air pollution, both the magnitude of the air pollution problem in China at the national level and the contribution from the power plants have been estimated by several stud ies, including one of the first to quantify the national burden of air pollution (World Bank 1997). In another study, Wang and Smith (1999) focused on the electric power sector in the context of determining the second ary benefits of greenhouse gas reductions. In addition, a largescale study conducted (Ho and Nielsen 2007) assessed the health dam ages of air pollution in China and examined several pollution control policies and how they might affect economic performance.
Researchers have pointed out that emission reductions in different sectors may have differ ent levels of effectiveness on reducing human exposure (Li et al. 2004;Streets et al. 1999) and that the benefits of many pollution control measures likely far exceed the costs; however, the variance by sector and its policy impli cations for future pollution control have not been investigated systematically in China. For example, none of the previous studies included both ozone and PM 2.5 in assessing the health damages, although exposures have been associ ated with increased mortality and a variety of other health outcomes (Bell et al. 2006(Bell et al. , 2007Levy et al. 2005;Pope and Dockery 2006). The exclusion of ozone is partly because of past emphasis on the power sector, where SO 2 and PM receive greater attention, but is also attrib utable to limitations in the atmospheric models used in previous studies (Li et al. 2004). Also, most studies used PM 10 and total suspended particles (TSPs) to estimate population expo sure to PM in China, whereas epidemiologic studies in the United States and worldwide have demonstrated more robust associations with PM 2.5 .
Our study will fill this gap by comparing how emission control strategies across different sectors (e.g., power generation, mobile sources, industry) influence population exposures and health risks related to PM 2.5 and ozone in the YRD area. The sectoral details will help guide development strategies that are economically and environmentally optimal, providing the basis for policy makers to determine how to prioritize future control efforts among the dif ferent sectors in the YRD.

Materials and Methods
In this study, we applied air pollution health impact assessment methods, following approaches articulated elsewhere (National Research Council 2002;World Health Organization 2000). Briefly, this entailed esti mating baseline emissions and the marginal contribution from individual source cate gories, use of a chemistrytransport model to characterize population exposures asso ciated with source category emissions, and application of concentration-response (CR) functions from the epidemiological literature along with characterization of population pat terns to quantify health impacts.

CMAQ modeling and emission inventory.
A stateofthescience Eulerian grid modelthe Community Multiscale Air Quality (CMAQ) modeling system [version 4.6; U.S. Environmental Protection Agency (EPA), Research Triangle Park, NC, USA] was applied to an emission inventory we developed to esti mate the baseline concentrations as well as the marginal concentration change associated with hypothetical control strategies for multiple sectors in the YRD. CMAQ has capabilities to simulate the various chemical and physical processes important for understanding atmo spheric processes and thus allowed us to model population exposure to pollutants such as ozone and PM 2.5 . Although CMAQ is currently used in analyses in the United States (Cohan et al. 2007) and in some air quality assessments in China (Fu et al. 2009), it has rarely been used for risk assessments in China and other devel oping countries. We applied a threeway nested model with 27km (covering all of China), 9km (covering eastern China), and 3km (covering the YRD) grid resolutions, respec tively [see Supplemental Material, Figure S1 (doi:10.1289/ehp.1001991)]. Outputs from the 27km and 9km domains were used as boundary conditions for the 3km domain. To develop meteorological inputs for CMAQ, we used the fifthgeneration mesoscale model (MM5) developed by the U.S. National Center for Atmospheric Research. For more detailed information on CMAQ and MM5 configura tions, see Supplemental Material , Table S1.
The emission inventory for the YRD con sisted of emissions mainly from four sectors: power plants, industrial sources (e.g., metal lurgical, mineral, cement, chemical indus tries, small industries such as coke and brick production), mobile sources, and domestic life (e.g., livestock, residential, and biomass burn ing). Power plants and emissions from pro duction processes of major industrial sources were modeled as point sources. Mobile sources mainly included onroad vehicle emissions in the major cities. Emissions from domestic life and the fugitive emissions from industry were modeled as area sources.
In developing our emission inventory, we used the Asian emission inventory for 2006 that was developed for the INTEXB project of the U.S. National Aeronautics and Space Administration (Zhang et al. 2009). The reso lution of the INTEXB emission inventory is 30 min × 30 min (~ 55 km × ~ 55 km). To capture finer resolution within our emission inventory, we incorporated various sources of input, using collaborations with local agencies, Google Earth (version 5.1.3535.3218; Google Inc., Mountain View, CA, USA) to identify the locations of large point sources based on their addresses, updated road networks and LandScan population data to characterize the distribution of mobile source emissions, and the statistical yearbook of the major cities in the YRD for domestic fuel emissions from the consumption of coal, liquified petroleum gas, coal gas, and natural gas.
After conducting the base-case simulation using CMAQ, we validated it using available monitoring data for ozone and PM 2.5 collected in Shanghai before proceeding with additional model runs. Because PM 2.5 is not currently a criteria air pollutant in China and there are no publicly available ozone monitoring data, the monitoring data we had access to for model validation were all from research monitor ing sites located in Shanghai and operated by Fudan University [see Supplemental Material, Figure S2 (doi:10.1289/ehp.1001991)]. Hourly ozone monitoring data were available from three stations for 4 weeks (2-9 May, 12-19 June, 10-17 July, and 14-21 August 2006). Daily average PM 2.5 monitoring data were available from one station for 1-29 April, 1-26 August, and 1-28 November 2006.
Once the validation of the baseline run was satisfactory, we designed scenarios to model the marginal contribution of selected emission sectors in the YRD to the base line concentrations in the 3km resolution domain. Although some underestimation of the total health benefits would have resulted, given the significant regional transport of PM 2.5 and ozone, the size of the 3km domain provides YRD environmental authorities with the most relevant information for control strategy prioritization and decision making.
For this study, we modeled sectoral impacts in total (e.g., all power plants in the YRD simultaneously). Table 1 lists the emission reduction scenarios in more detail. In design ing the scenarios, we considered several factors. First, we focused on NO x emission control, a stated interest of decision makers and a pollut ant for which emissions have been increasing the fastest in recent years among the pollutants modeled (Zhang et al. 2009). Second, because NO x emissions influence PM 2.5 and ozone concentrations, and given the anticipated health impacts of PM 2.5 and ozone, we also considered other pollutants [e.g., SO 2 , vola tile organic compounds (VOCs), and primary PM 2.5 ] that can affect PM 2.5 and ozone con centrations. Lastly, we tried to cover the cur rent and proposed pollution control measures by the Shanghai government in preparation for World Expo 2010 (Chen CH, personal communication), although these scenarios do not necessarily correspond to specific official control measures. In each of the four major sectors (power, industry, mobile, and domes tic), we included scenarios for reducing NO x alone (scenarios 1, 3, 7, and 10) and scenarios for reducing NO x along with other pollutants to evaluate the relative magnitude of impacts (scenarios 2, 5, and 9). We also considered scenarios for reducing VOC alone (scenarios 4 and 8), allowing for analyses of interactions between NO x and VOC controls, as well as one scenario (scenario 6) to check for non linearity of concentration changes to the mag nitude of emission reductions. The magnitude of the emission reductions corresponds with specific control technologies for the power sec tor and an approximation of technologically and financially plausible emission reductions in other sectors, although the logistical feasibility and costs of controls clearly vary across sectors.
In each case, we estimated the total expo sure and public health impacts but focused on the marginal impacts per ton of emissions, allowing for direct comparisons among sectors. Although CMAQ calculates speciated PM 2.5 concentrations, we report the total PM 2.5 concentrations in calculating the population Power NO x alone (SCR alone) 85% 2 Power NO x + SO 2 (SCR + FGD) 85% for NO x + 90% for SO 2 3 Mobile NO x alone 20% 4 Mobile VOC alone 20% 5 Mobile NO x + VOC + PM 20% 6 Mobile NO x + VOC + PM 50% 7 Industry NO x alone 20% 8 Industry VOC alone 20% 9 Industry NO x + VOC + PM 20% 10 Domestic NO x alone 20% Abbreviations: FGD, fluidized gas desulfurization; SCR, selective catalytic reduction.
volume 118 | number 9 | September 2010 • Environmental Health Perspectives exposure under different scenarios, which includes changes in multiple constituents. To estimate the pollutantspecific benefits, we used the incremental population exposure change between scenarios. For example, the difference in population exposure between scenario 1 and the base-case allowed us to calculate exposure reductions associated with NO x emissions, whereas the difference between scenarios 2 and 1 allowed us to approximate the exposure reductions associated with SO 2 emissions. Population data. To calculate the popula tion exposure and subsequent health impacts, we collected population data from two dif ferent sources. First, we used LandScan 2007 as the primary basis for estimating exposures. LandScan is a worldwide population database compiled on a 30second × 30second latitude/ longitude grid. Census counts (at subnational level) were apportioned to each grid cell based on likelihood coefficients, which are based on proximity to roads, slope, land cover, night time lights, and other information (Dobson et al. 2000;Oak Ridge National Laboratory 2009). Similar to Census data, LandScan pro vides a single population estimate for each location, although these estimates include diurnal movements and collective travel habits, whereas most censuses count people at their nighttime residences. In addition, LandScan includes annual updating of data inputs, as well as global coverage, which potentially allows for easy comparisons with other parts of the world in the future.
However, LandScan does not provide the same level of detail as Census data in terms of population demographics, so we obtained additional information necessary for sub sequent calculation of health benefits (e.g., baseline mortality rates) from China Census 2000 (China Data Center 2003). To calcu late the health benefits, we applied spatially variable baseline mortality rates, to reflect the fact that the YRD area includes both urban and rural areas and that their residents may have different disease patterns, socioeconomic status, and life expectancy. For counties in the YRD, the average baseline mortality rate is 0.6%, with the 5th and 95th percentile rate at 0.34% and 0.84%, respectively. We assumed these estimates based on Census 2000 are applicable to population estimates from LandScan 2007 data.
Population exposure. The population expo sure change under different scenarios was calcu lated by combining population in each location with the corresponding concentration change.
Geographical information system (ArcGIS) software (version 9.3; ESRI, Redlands, CA, USA) was used to convert population data to match the grid structure of CMAQ. Because the emissions of different pollutants under study vary significantly by sector, we focused our analyses on the marginal benefits per ton of emission reductions. We facilitated these com parisons by using the concept of intake fraction (iF), the fraction of a material released from a source that is inhaled or ingested (Bennett et al. 2002). We calculated iF as (Σ C i × P i ) × (BR/Q), where C i is the marginal concentra tion in grid cell i (micrograms per cubic meter) associated with source emission rate Q (micro grams per day, noting that Q can be the same pollutant as C or a precursor to C), P i is the population count in the grid cell, and BR is a nominal breathing rate of 20 m 3 /day. We cal culated iFs for primary PM as well as secondary PM associated with various particle precursors (SO 2 , NO x , VOCs) and ozone (defined as the mass of ozone inhaled per unit mass of NO x or VOC emissions). Because numerous particle constituents are influenced by precursor emis sion changes in CMAQ, we did not focus on iF values for individual constituents but discuss the dominant constituents for all secondary PM iFs. In addition, because emissions of mul tiple pollutants can influence PM 2.5 and ozone concentrations, we estimated the pollutant specific and sectorspecific iFs by comparing the population exposures among scenarios.
Health effects. Although detailed charac terization of health risks is beyond the scope of our investigation, we developed CR functions for PM 2.5 and ozone mortality to allow our CMAQ outputs to be integrated into a com mon metric and to allow for initial evaluation of the magnitude of health benefits, the domi nant pollutants, and the key uncertainties. For both PM 2.5 and ozone, we determined a central estimate, plausible lower bound, and plausible upper bound. These are not meant as formal 95% CIs but were used to construct uncertainty distributions when combining PM 2.5 and ozone health benefit estimates.
As a general point, there are multiple limitations in applying CR functions largely derived from the United States or Europe to China. There are differences in baseline dis ease patterns and age distributions, health care systems, pollutant levels and composition, and exposure modifiers. There are also complexities given the more recent focus on PM 2.5 in U.S. and European studies but the use of PM 10 or TSP in China given available monitoring data. To develop applicable CR functions, we used a combination of evidence from the global literature and the Chinese literature.
First we considered PM: Two studies that developed CR functions applicable to China (Aunan and Pan 2004;Levy and Greco 2007) concluded that the Chinese timeseries mor tality literature yielded estimates on the order of 0.3-0.4% increases in allcause mortality per 10μg/m 3 increase in daily PM 10 con centrations, slightly lower than the global literature. In a recent study that examined three cities in China (Wuhan, Shanghai, and Hong Kong), Wong et al. (2008) found a pooled CR function for timeseries mortal ity of 0.37% (95% CI, 0.21-0.54%), similar to the values reported above. CR functions in China were higher for cardiovascular and respiratory mortality, with patterns similar to those seen in the global literature.
However, health risk assessments for PM generally apply evidence from cohort mor tality studies (National Research Council 2002;World Health Organization 2000), given the strength of available studies and supporting evidence for mortality risks from longterm exposure (e.g., evidence that PM contributes to accelerated atherosclerosis; Floyd et al. 2009). Recent syntheses of the cohort mortality litera ture (Levy et al. 2009) and expert elicitation studies (Industrial Economics 2006) found that a 1% increase in mortality per 1μg/m 3 increase in annual PM 2.5 concentrations was a reasonable cen tral estimate, falling between estimates from the Harvard Six Cities Study (Laden et al. 2006;Schwartz et al. 2008) and the American Cancer Society study (Jerrett et al. 2009;Pope et al. 2002). There is no cohort mortality evi dence available in China, but earlier cross sectional studies yielded CR functions roughly comparable to those from the U.S. cohort studies (Levy and Greco 2007).
Despite the lack of Chinese cohort mor tality evidence, the literature is sufficiently compelling to indicate that mortality risk due to longterm exposure would be expected, and comparisons of the timeseries estimates indicate reasonable concordance between the Chinese and U.S. literature despite the large differences in ambient concentrations and other factors. Thus, for our central estimate, we used a 1% increase in allcause mortality per 1μg/m 3 increase in annual PM 2.5 con centrations. For our bounds, we note that a recent study (Levy et al. 2009) used 0.3% as a lower bound and 2.0% as an upper bound, representing the median values across experts for the 5th and 95th percentiles of the uncer tainty distribution in the recent expert elicita tion study. We maintained this upper bound (which slightly exceeds the central estimate from the Harvard Six Cities Study) but used a lower bound of 0.1%, reflecting a value similar to the timeseries evidence and the uncertainties in determining a cohort mortal ity effect in China without direct evidence. These CR functions were applied identically to all particle constituents, given a lack of sys tematic information to support quantitative differential toxicity, especially with respect to atmospheric conditions in China.
Ozone had not been characterized in pre vious studies developing CR functions for China, in part because of a lack of systematic evidence in the global literature at the time of those investigations. However, three recent metaanalyses and multicity studies (Bell et al. 2005;Ito et al. 2005;Levy et al. 2005) found evidence of an independent ozone effect in the timeseries literature; ozone mortality was recently evaluated in multiple Chinese cities (Wong et al. 2008), and recent evidence from the American Cancer Society cohort study (Jerrett et al. 2009) provides some indica tion of a longterm ozone effect on respiratory mortality. Thus, ozone mortality merits inclu sion in our investigation.
Similar to PM 2.5 , there is no evidence of ozone cohort mortality for China. In the U.S. cohort literature, an ozone effect on mortal ity was significant in one recent publication (Jerrett et al. 2009), but only for respiratory mortality in models including PM 2.5 . The CR function for respiratory mortality cor responded to a 2.7% increase per 10μg/m 3 increase in 8hr maximum ozone following the conversions above. This is significantly greater than the timeseries estimates (albeit for respiratory mortality, only a fraction of allcause mortality), but because of the lack of an impact for allcause mortality or corrobo ration from other studies, we do not use this evidence for our primary CR functions.
Combining this evidence, we considered 0.3% reduction in allcause mortality per 10μg/m 3 reduction in 8hr maximum ozone as a reasonable central estimate (reflecting the Chinese threecity study and two of the metaanalyses), with 0.15% as a lower bound (reflecting the lower confidence limits of the various studies) and 0.45% as an upper bound (reflecting the upper confidence limit of the Chinese threecity study as well as mod est weight on the emerging cohort mortality evidence).
Health benefit estimation. Although we did not develop the upper and lower bound CR functions as specific percentiles of uncer tainty distributions, we wished to estimate net health benefits across PM 2.5 and ozone, neces sitating some combination of distributions. To approximate the overall net mortality change considering the uncertainty in CR functions, we assumed that the CR functions for PM 2.5 and ozone each follows a triangular distribu tion with central estimate as the mode and the lower and upper bounds as the minimum and maximum values for the distribution. We per formed Monte Carlo analysis using SAS 9.1 (SAS Institute Inc., Cary, NC, USA) to com bine the distributions, noting that uncertainty in other risk assessment components was not considered.

Comparison between CMAQ modeling and monitoring data.
To validate the performance of CMAQ, we compared ozone and PM 2.5 modeling results in the base-case scenario with observation data for part of 2006 at four monitoring stations in Shanghai, using U.S. EPA guidance (U.S. EPA 2007). The mean normalized bias (MNB) between model and observational data was relatively low for both daily average PM 2.5 (0.5%) and daily maximum 8hr ozone (-4.3%), indicating a lack of systema tic model bias. For hourly ozone, the model underestimated concentra tions somewhat, with MNB of -25.3% and normalized mean error of 29%. In general, model performance was considered adequate for our application.
Comparison between census and LandScan data. To provide validation of population counts, we compared China Census 2000 and LandScan 2007. For the 3km resolution YRD domain, the mean percentage difference by county was about 9%. Supplemental Material, Figure S3 Table 2, we show the estimated emission rates by sector and pollutant in the base-case scenario. Of note, emissions of many pollutants are high in the industry sec tor, in part due to the active manufacturing industry in the YRD and the lower penetra tion rate for pollution control technologies compared with the power sector.  Figure 1 shows the annual average PM 2.5 concentration in the YRD domain in the base-case scenario. The estimated mean annual PM 2.5 concentration in the YRD domain was 38.4 μg/m 3 , although with significant variation across the study domain (range, 12.7-132.8 μg/m 3 ). Similarly, Figure S4 in Supplemental Material (doi:10.1289/ehp.1001991) shows the annual average 8hr maximum ozone con centration, which ranged from 17 to 54 ppb.

Base-case emissions and ambient concentrations. In
iF variation by pollutant and sector. To estimate the pollutantspecific iFs, we com pared the population exposures among sce narios (Table 3). iFs ranged significantly across pollutants, with more modest differences across sectors. Primary PM 2.5 from the industry and mobile source sectors have the highest iF of 1.4 × 10 -5 , which means that for every metric ton of primary PM 2.5 emitted from either sector, 14 g is eventually inhaled by the total population in the YRD domain. For second ary PM 2.5 , the iF was greatest for SO 2 emis sions from the power sector, with a value of 1.2 × 10 -6 . Among the different species of PM 2.5 modeled by CMAQ, 98% of the PM 2.5 concentration reduction was attributable to sulfate and ammonium particles. For second ary PM 2.5 from NO x emissions, the iFs from power plants, mobile sources, and industry are nearly identical (~ 3.9 × 10 -7 ). In each case, the concentration reduction was driven by nitrate and ammonium reductions with an offsetting increase in sulfate (35% of the magnitude of the nitrate and ammonium reductions for the power sector, 20% for mobile sources, and 19% for industry). In contrast, the domestic emissions sector has a negative iF for secondary PM 2.5 from NO x emissions, potentially attrib utable to two factors. The low NO x emissions within the domestic sector translated into a small reduction in secondary nitrate population exposure. Second, when compared with other scenarios, the domestic sector had the greatest increase in ozone concentrations per unit NO x emission reductions, which contributed to greater oxidizing power of the atmospheric and subsequent increases in other PM 2.5 species (e.g., secondary sulfate and secondary organic aerosols). As a result, there is an increase in overall PM 2.5 concentration corresponding to the NO x emission reductions within the domestic sector.
Across all scenarios, reducing NO x emis sions alone led to an increase in ozone popu lation exposure (as indicated by the negative values in Table 3, "Ozone from NO x "). However, when VOC emissions are reduced alone, ozone population exposures are reduced and the iF is on the order of 1 × 10 -6 (as shown in Table 3, "Ozone from VOC"). This indicates that the YRD model domain is VOC limited in terms of ozone formation, explain able by the high baseline NO x emissions and low biogenic VOCs. Findings were similar using 1hr maximum ozone concentrations, with iFs approximately 10-30% higher. Of note, within scenarios with concurrent NO x and VOC controls, we observed only modest differences from the sum of NO x and VOC controls applied separately.
Health benefit variation by pollutant and sector. Although the health benefits per ton of emission reductions are approximately pro portional to the iF values in Table 3 (when baseline mortality rates are about constant in different parts of the domain), the abso lute benefits of the control scenarios will also depend on the magnitude of emission reduc tions. As indicated in Table 4, the greatest mortality reduction is achieved by controlling primary PM 2.5 emissions from the industry sector and by controlling SO 2 emissions from the power sector, with approximately 12,000 fewer deaths per year (lower bound of 1,200, upper bound of 24,000). This is attributable to the high primary PM 2.5 iF and relatively high magnitude of emission reductions from the industry sector, whereas the SO 2 emis sions from the power sector have an order of magnitude lower iF but an order of magnitude higher emission reduction.
For control scenarios addressing NO x emissions, the health benefits from second ary PM 2.5 reductions (which are themselves reduced by offsetting increases in sulfate con centrations) are blunted by adverse health impacts associated with ozone increases (Table  4). Although the NO x control scenarios still have positive net benefits (excluding the domestic sector), the net benefits are small relative to the aforementioned benefits for primary PM 2.5 and SO 2 controls.

Discussion
To maximize the health benefits of emission reductions among different pollutants from different sectors, there are two major factors to From SO 2 1.2 × 10 -6 2, 1 From NO x 3.9 × 10 -7 1 3.9 × 10 -7 3 3.9 × 10 -7 7 -2.1 × 10 -7 10 From VOC 2.4 × 10 -7 4 1.3 × 10 -7 8 Ozone From NO x -6.8 × 10 -7 1 -6.9 × 10 -7 3 -6.9 × 10 -7 7 -1.5 × 10 -5 10 From VOC 1.7 × 10 -6 4 1.4 × 10 -6 8 Blank cells indicate values not estimated in any scenario runs. iF results reported are unitless. To calculate pollutantspecific iFs, we compared population exposures among scenarios, where each number is the difference between the scenario and the baseline scenario. For example, 1 means the corresponding iF is calculated based on the population exposure difference between the baseline scenario and scenario 1. When multiple scenarios are listed, iF was calculated based on the difference between each scenario listed and the baseline case, as well as the difference among the scenarios listed. All values are provided to two significant figures, and sums may not add due to rounding. A positive value in the last four columns means mortality reduction (or fewer deaths), and a negative value means mortality increase (or more deaths). Values in parentheses represent plausible upper and lower bounds for pollutant-specific mortality changes and 5th and 95th percentile values from a Monte Carlo simulation for net mortality changes. Scenario 6 is not shown here because it is included as a sensitivity test. Calculating pollutant-specific population exposures and mortality changes based on this scenario would require additional modeling scenarios (e.g., two additional scenarios similar to scenarios 3 and 4, but with a reduction percentage of 50%, respectively).
consider-the population exposure reduction per unit emission reduction and the amount of emissions that can be reduced from different sectors. As expected and shown in prior studies in China (Zhou et al. 2003(Zhou et al. , 2006, the highest population exposure reduction per unit emis sion reduction for PM 2.5 is from controlling primary PM 2.5 emissions rather than particle precursors. Because of the relatively high pri mary PM 2.5 emissions in the industry sector and the feasibility of emissions reductions, the potential health benefits are substantial. Considering particle precursors, SO 2 emission reductions yielded greater PM 2.5 exposure reductions than did NO x or VOC emissions reductions. In contrast, previous studies in China (Zhou et al. 2006) found similar iFs for sulfate from SO 2 and nitrate from NO x , both on the order of 10 -6 . In particular, the iFs for nitrate PM 2.5 formed from NO x emissions from power plants in Shanghai and the provinces of Jiangsu and Zhejiang ranged from 2 to 5 × 10 -6 , versus 4 × 10 -7 in the present study. The differences are likely due to two factors: The modeling domain in the previous study covers all of China, and the previous study used an atmo spheric model (CALPUFF) that did not cap ture the offsetting increase in sulfate when NO x emissions are reduced.
For ozonerelated health benefits, our results show that VOC control is more effec tive than NO x control, due to the YRD area being hydrocarbon limited in ozone forma tion, where ozone concentrations increase with increasing hydrocarbons (e.g., VOC) and decrease with increasing NO x (Jacob 1999). One previous study (Wang et al. 2005) found that in the Pearl River Delta area of China, urban areas are VOC limited in ozone forma tion and the nonurban areas are NO x limited, where ozone concentrations increase with increasing NO x and are insensitive to hydro carbons. Our study shows that the YRD as a whole (which contains many dense urban areas) is hydrocarbon limited, although it was beyond the scope of our study to explore the implications of source controls in different regions of the YRD or the potential long range ozone formation that could occur from NO x controls. Despite the hydrocarbon limited ozone formation in the YRD and the relatively low particle formation per unit NO x emissions, NO x emission reductions in power, industry, and traffic sectors are net beneficial, across the range of CR functions simulated for ozone and PM 2.5 .
Limitations. Several limitations could potentially influence the interpretation of our findings. First, in estimating the health benefits per ton of emission reductions, we implicitly assumed a linear relationship between pollutant emission reductions and population exposure reductions. There are many non linear processes in the atmosphere chemistry that could make this assumption faulty. However, our findings suggest that non linearities are limited given the emission changes in our study. For example, the emission reductions in scenario 6 were 2.5 times those of scenario 5 (e.g., 50% vs. 20% reductions of NO x , VOC, and PM from the mobile sector), and the resulting total popula tion exposures in scenario 6 were 2.53 times greater for PM 2.5 and 2.44 times greater for ozone, indicating reasonable linearity.
Second, because PM 2.5 is not currently a criteria air pollutant in China and there are no publicly available ozone monitoring data, the monitoring data we had access to for model validation are all from research monitoring sites located in Shanghai, which limited our ability to validate the model performance in other parts of the YRD modeling domain. Although further validation would have been ideal, vari ous components of the model (e.g., the original emissions inventory input, the application of CMAQ in China) have been previously evalu ated and validated, increasing our confidence in our findings. The lack of available monitor ing data also emphasizes the need for publicly available comprehensive information systems in order to support health risk analyses and other environmental evaluations in China.
Third, although the CR functions lever aged a combination of epidemiological evi dence from the global literature and from China, our health benefit estimates are domi nated by risks from longterm PM 2.5 expo sure, for which there is no evidence within China. Moreover, the CR function derived from the U.S. cohort studies would imply an extremely large mortality gradient across the YRD and between different areas of China, which is challenging to validate and interpret. That said, we did not consider it appropriate to omit cohort mortality entirely, given its biological plausibility, and no evidence exists to quantitatively deviate from the available cohort evidence. More generally, our conclu sions about the relative importance of various source sectors are robust to this assumption. Our core findings therefore remain interpre table despite this large uncertainty, and we recommend that the CR function for PM 2.5 mortality in China be reevaluated as more evidence from China becomes available.
Fourth, our analysis considers mortality only from PM 2.5 and ozone. Incorporation of other impacts, including morbidity outcomes or ecological damage from acid deposition, could be potentially influential if economic valuation of mortality were lower in China relative to economic valuation of morbidity than in the United States.

Conclusion
Despite the limitations, we demonstrated in this study a systematic approach to compare the effectiveness of pollutant control strategies across different sectors in a highly exposed and highly populated region of China. The use of the stateofthescience air quality model CMAQ and a spatially resolved emission inventory allowed us to jointly consider PM 2.5 and ozone exposures for different emission reduction scenarios. Our findings indicate sig nificant variation across pollutants in health benefits per ton of emission reduction. The public health benefits of realistic controls for SO 2 emissions from the power sector and pri mary PM 2.5 emissions from the industry sector are roughly comparable, given higher emission reductions for the former and higher popula tion exposures per ton of the latter, with lower benefits from NO x control strategies. This is attributable in part to the hydrocarbonlimited nature of the YRD, as well as to the lower sec ondary PM 2.5 formation per ton of NO x emis sions relative to other particle precursors. Our findings, in combination with plausible emis sions reduction estimates and their costs, pro vide the basis for prioritizing pollution control strategies in the YRD and provide a template for comparable analyses elsewhere.