Exploring sources of uncertainty in premature mortality estimates from fine particulate matter: the case of China

Atmospheric pollution from fine particulate matter (PM2.5) is one of the major concerns in China because of its widespread and harmful impacts on human health. In recent years, multiple studies have sought to estimate the premature mortality burden from exposure to PM2.5 to inform policy decisions. However, different modeling choices have led to a wide array of results, with significant discrepancies both in the total mortality burden and in the confidence intervals. Here, we present a new comprehensive assessment of PM2.5-related mortality for China, which includes quantification of the main sources of variability, as well as of age and province-specific premature mortality trends during 2015–2018. Our approach integrates PM2.5 observations from more than 1600 monitoring stations with the output of a high-resolution (8 km) regional simulation, to accurately estimate PM2.5 fields along with their uncertainty, which is generally neglected. We discuss the sensitivity of mortality estimates to the choice of the exposure-response functions (ERFs), by comparing the widely used integrated exposure response functions (IERs) to the recently developed Global Exposure Mortality Models (GEMMs). By propagating the uncertainty in baseline mortalities, PM2.5 and ERFs under a Monte Carlo framework, we show that the 95% confidence intervals of mortality estimates are considerably wider than previously reported. We thus highlight the need for more epidemiological studies to constrain ERFs and we argue that uncertainty related to PM2.5 estimate should be also incorporated in health impact assessment studies. Although the overall mortality burden remains vast in China (∼1.6 million premature deaths, according to GEMMs), our results suggest that 200 000 premature deaths were avoided and 195 billion US dollars were saved in 2018 compared to 2015, bolstering the mounting evidence about the effectiveness of China’s air quality policies.


Introduction
Exposure to high concentrations of air pollutants, including fine particulate matter (i.e. PM 2.5 ), leads to premature mortality related to cardiovascular and pulmonary diseases, as highlighted by a number of epidemiological studies (Kan et al 2012, Apte et al 2015. As a result of rapid economic growth, industrialization and urbanization, China is experiencing severely degraded air quality conditions (Li et al 2016a) and ranks first for premature mortality due to air pollution with more than one million deaths per year (Lelieveld et al 2015, Gu et al 2018, followed by India with~650 000 deaths (Lelieveld et al 2015). In response to this situation, since the year 2010, the Chinese government has implemented clean air policies to reduce air pollution levels (State-Council 2013) by imposing more stringent emission standards and by promoting transition to cleaner fuels and/or non-fossil energy sources for the industrial, power, and residential sectors (Zheng et al 2018). These actions will contribute to mitigate not only pollution impacts on air quality (Yang and Zhang 2018), but also those on climate change . The recent decreasing trend in pollution concentrations (during 2015-2017) has been linked to the effective implementation of emissions reduction strategies by purely observation-based studies (Silver et al 2018) as well as by model simulations (Ding et al 2019).
Multiple studies have sought to estimate the nationwide mortality burden associated with PM 2.5 in China, by employing different techniques. Most of these analyses rely on regional chemical transport model (CTMs) simulations, which, however, are subject to significant uncertainties that are often not quantified (Lelieveld et al 2015, Silva et al 2016, Butt et al 2017, Hu et al 2017, Zheng et al 2017b, Aunan et al 2018, Gao et al 2018. Depending on the simulation setup and the chosen exposure-response functions (ERFs), significantly different estimates of the total premature mortality burden were presented, ranging from 916 000 premature deaths in 2014 (Archer-Nicholls et al 2016) to 1331 000 in 2013 (Gao et al 2018). Much larger figures (up to~2.47 million) are obtained when using the latest exposure response functions from Burnett et al (2018). Furthermore, the 95% confidence intervals (CI) reported in such studies can be as narrow as (1 090 000-1 190 000) (Aunan et al 2018) or as wide as (168 000-1 796 000) (Gu et al 2018). Besides model simulations and ERFs, other sources of uncertainty include baseline mortalities and exposed population estimates (Kodros et al 2018).
Mortality estimates have also been proposed based on a combination of ground and/or satellite observations (Zhou et al 2015, Jerrett et al 2017, despite being mostly focused on geographically limited areas. A few attempts to combine observations-either from monitoring stations or satellites-and CTM results were made to overcome limitations of prior studies, both worldwide (Ford and Heald 2016, Van Donkelaar et al 2016, Shaddick et al 2018 and in China , Liang et al 2017, Bai et al 2019, Zou et al 2019. For instance, multiple studies used different geostatistical techniques to integrate PM 2.5 ground measurements, CTMs output, satellite observations of AOD and other land use and meteorological factors to estimate PM 2.5 exposure at different spatial scales, ranging from the city-level (Li et al 2016b, Tao et al 2020 to regional and national scales (Liang et al 2017, Bai et al 2019, Wei et al 2019. Although these works improved representation of PM 2.5 spatial patterns, the uncertainty in PM 2.5 concentrations-as well as its propagation to the mortality estimate-was not quantified. Using data assimilation techniques, Liu et al (2016) quantified the PM 2.5 uncertainty from a CTM and observations, although the resolution of that simulation was rather coarse (45 km).
Building from previous attempts to combine observations and modeling results, here we propose a novel approach to quantify the PM 2.5 -related mortality burden and its uncertainty in China. Our methodology integrates ground-based observations from the recently established Chinese monitoring network (>1500 sites) with a high-resolution CTM simulation (8 km). This work aims to: (a) Provide a multi-year assessment of the mortality burden related to PM 2.5 (2015-2018) by nudging model results to the observations. (b) Discuss the effect of using recently revised ERFs (Burnett et al 2018) in the recent mortality trends. (c) Overcome key limitations of previous studies on the uncertainty of such estimates.
Specifically, our methodology allows estimating the uncertainty in PM 2.5 concentrations and its propagation to the final mortality results. We show that confidence intervals of mortality estimates previously reported are likely narrower than the true ones, as they do not take into account PM 2.5 uncertainty (e.g. Reddington et al 2019, Zou et al 2019. We discuss nationwide and province-level mortality trends during 2015-2018 as well as the advantages of using our modeling framework compared to other approaches.

PM 2.5 observations
This investigation is based on hourly PM 2.5 concentrations for Mainland China and Hong Kong. Data are retrieved from online repositories of the Ministry of Ecology and Environment for Mainland China (http://beijingair.sinaapp.com/) and of the Hong Kong Environmental Protection Department (www.epd.gov.hk/epd) for Hong Kong. The monitoring network comprises 1601 and 16 sites for Mainland China and Hong Kong, respectively. PM 2.5 measurements are acquired and analyzed for four consecutive years, from 2015 to 2018, at hourly resolution. Data are pre-processed for use in the mortality assessment as described in the Supplementary Information.

WRF-Chem simulation
Hourly PM 2.5 concentrations are simulated by the Weather Research and Forecasting model with Chemistry (WRF-Chem, version 3.9) (Grell et al 2005, Skamarock andKlemp 2008) at 8 km resolution. This resolution allows to capture important regional details in complex mountainous areas and resolve atmospheric processes at urban and regional scales including formation/transformation/removal of air pollutants, thus facilitating the direct comparison between model results and station measurements.  (Li et al 2017), which incorporates data from the Multiresolution Emission Inventory for China (MEIC, www.meicmodel.org). We select EDGAR-HTAPv2.2 for consistency with the simulation resolution (8 km). Other emission inventories, like ECLIPSE (Klimont et al 2017), provide more up-to-date emissions but at a lower resolution (~50 km), which would cause a mismatch with the resolution applied in our model simulation. The entire 2015 is simulated, with a spin up of 1 month (1-31 December 2014). Details on the simulation setup are provided in the Supplementary Information (table S2).

Integrating observations and WRF-Chem results
To account for the uncertainty of WRF-Chem results, we introduce a statistical methodology which integrates measured data in the analysis. The novelty of the proposed methodology is three-fold: (i) nudging the CTM results towards the observations, (ii) providing a better quantification of the mortality CI by accounting for PM 2.5 uncertainty and (iii) estimating a multiyear premature mortality trend using only one year of computationally-expensive WRF-Chem simulations.
The PM 2.5 concentration field is reconstructed by applying a linear regression model followed by kriging of the residuals (ε) generated by the model: where z is the corrected annual mean concentration of PM 2.5 in each grid cell, x is the logarithmically transformed annual mean concentration simulated by WRF-Chem, β 0 and β 1 are parameters to be estimated. In other words, the observed data are first detrended using WRF-Chem results as the only covariate, and then ordinary kriging is performed on the detrended data to include the residual spatial dependence in the model. A logarithmic transformation is used in equation (1) to work under the Gaussianity assumption, as the original data are slightly skewed (figure S1 (stacks.iop.org/ERL/15/064027/mmedia)). Moreover, the logarithmic transformation avoids unphysically negative PM 2.5 concentrations that could arise from the regression. Since x and ε are Gaussian and the model is additive, we can calculate the variance of the predictions (i.e. the uncertainty in the PM 2.5 field) by adding the variance related to the linear regression part and the variance arising from the kriging predictions. Details on the calculation of prediction variances for linear regression and ordinary kriging are reported in the Supplementary Information for completeness.
This procedure is applied for each year in the 2015-2018 period, with time-varying coefficients (i.e. β 0 , β 1 and the parameters of the variogram are estimated on an annual basis). Linear regression parameters are estimated by maximizing the likelihood of the data, which is equivalent to ordinary least squares for linear models, and the variogram is estimated using a spherical covariance function, which proved to be the most numerically stable covariance function across the years. The values of the estimated parameters can be found in the Supplementary Information (table S1). For comparison with other models, hereinafter we will denote model (1) as 'Regression + Kriging' , whereas the same model without kriging on the residuals ε will be denoted as 'Regressiononly' (i.e. model (1) where we do not model the spatial structure in ε and we assume that the residuals are instead normally distributed with zero mean and variance σ 2 ).

Population exposure and mortality estimates
Long-term premature mortality associated with PM 2.5 is calculated following a well-established approach (e.g. Burnett et al 2014, Apte et al 2015): where M is the estimated premature mortality burden, P is the population count, B is the baseline mortality rate and RR is the relative risk due to exposure to annual mean concentration of PM 2.5 . Subscripts a, d and k refer to the specific age interval, disease and grid cell, respectively. As in previous studies (Gu et al 2018), we consider premature mortality associated to lung cancer (LC), ischemic heart disease (IHD), stroke and chronic obstructive pulmonary diseases (COPD) for individuals above 25 years with 5-years groupings ( where θ, α, µ, ν are age and disease-specific parameters, andz k is the maximum between zero and the difference between the PM 2.5 concentration in grid cell k and the counterfactual concentration (z cf = 2.4 µg m −3 ), i.e. the minimum concentration below which no impacts are assumed to occur. The parameters were estimated by Burnett et al (2018) using cohort studies that cover the global exposure range, including a Chinese male cohort. For comparison purposes, we also calculate RRs with the traditional IERs (Burnett et al 2014), which however include risk information from indoor air pollution and secondhand and active smoking: where α, γ and δ are age-and disease-specific parameters estimated in the GBD study (Dicker et al 2018) along with their uncertainty. Equation (4) is valid only if z k is greater than z cf , otherwise RR is set to 1. The counterfactual concentration for the IERs is uniformly distributed between 2.4 and 5.9 µg m −3 . Total mortality is finally computed as the sum for each health endpoint, age group and grid cell: We also compute the monetary savings as the difference in mortality between two instances in time multiplied by the value of statistical life (VSL), which is assumed to be equal to 978 219 US dollars (in 2011 values), following previous literature (Zou et al 2019). A thorough discussion of the monetary evaluation uncertainty exceeds the scope of this work, but could be considered in future investigations.

Uncertainty quantification
We propagate the uncertainty in the input (i.e. PM 2.5 fields, baseline mortality and ERFs parameters) to the mortality estimates with a numerical approach. PM 2.5 fields, baseline mortality and ERFs are assumed to be Gaussian distributions, with parameters (i.e. mean and standard deviations) specified in the Supplementary Information. For each Monte Carlo simulation, we independently sample from such distributions and we apply equations (2) to (5) to calculate the mortality value for that specific simulation. To ensure robustness of our results (i.e. a coefficient of variation of the estimated total mean mortality less than 0.05%), we perform 108 000 simulations (i.e. 4500 simulations per processor distributed on 24 processors) which resulted in 108 000 total mortality values. The mortality central estimate is therefore the mean of the 108 000 mortality estimates, whereas the 95% confidence interval is calculated as the 0.025-and 0.975-quantiles of the empirical distribution obtained with the 108 000 simulations.
To assess the influence of each input (i.e. PM 2.5 , ERFs parameters and baseline mortality) on the total uncertainty, we run four different cases by marginally assuming no uncertainty in some of the input features (i.e. treating that input feature as an exact value rather than a Gaussian random variable): (i) only ERFs parameters are uncertain (ERFsonly), as usually done in previous literature, (ii) only ERFs and baseline mortalities are uncertain (ERFs + BM), (iii) only ERFs and PM 2.5 are uncertain (ERFs + PM 2.5 ) and (iv) all three input features are uncertain (ERFs + BM + PM 2.5 ).

Model evaluation
Annual PM 2.5 concentrations are generally well reproduced by the high-resolution WRF-Chem simulation, with 93.5% of the modeled values being within a factor of 2 of the measured concentrations ( figure S2). On average, the model overestimates observations, with a mean bias of 10.16 µg m −3 (figure S2(b)), which is likely related to the mismatch between the simulated year and the emission inventory year. Such overestimation mainly occurs during the summer, when the overall difference between the model and the observations is significant ( figure  S2(b)). Conversely, the model is negatively biased at most sites measuring extremely high annual average concentrations (i.e. above 100 µg m −3 ). As clear from figure S3, the most misrepresented areas are the Sichuan province (e.g. Chengdu, which is highly overestimated) and the westernmost cities of China (e.g. Kashgar, which is underestimated). Despite the relatively fine spatial resolution, limitations remain in the model's ability to capture local scale PM 2.5 concentrations, particularly over megacities where urban measurements are highly impacted by local emissions and sub-grid scale dispersion processes not resolved by a CTM.
A more accurate PM 2.5 field can be obtained by integrating observations with the model results, as described in section 2.3. According to the calibration/validation procedure described in the Supplementary Information, the 'Regression + Kriging' statistical model is the most suitable to reproduce the actual observations ( figure 1 and table S3). Major improvements to the WRF-Chem results are thus achieved when integrated with observations, even with a simple linear regression only (the mean bias decreases from 10.2 µg m −3 to −2.50 µg m −3 and Root Mean Square Error (RMSE, defined in the Supplementary Information) from 24.87 µg m −3 to 15.62 µg m −3 ). However, by also relaxing the assumption of iid residuals (ε in equation 1) and modeling the spatial dependence via a kriging procedure, we can accomplish almost zero bias with a threefold decrease in RMSE compared to the WRF-Chem model results. More complex statistical models were also investigated but did not provide significantly more accurate results despite the additional complexity (Supplementary Information).

Premature mortality estimates
The long-term premature mortality burden during 2015-2018 is reported in table 1. Age-and diseasespecific premature mortality estimates indicate that IHD and stroke are responsible for most premature deaths attributable to PM 2.5 (36.3% and 32.0% of total PM 2.5 related deaths, respectively), especially among the young population (figure 2). More than 50% premature deaths due to LC, IHD, COPD and stroke are attributable to PM 2.5 exposure (i.e. RR ⩾ 2, using GEMMs) for adults less than 40 years old. This percentage decreases with age, as other medical conditions leading to those diseases may dominate for the elderly population (figure 2). Given the rather short timeframe considered in this work, the premature mortality distribution by age groups remains almost constant in time (figure S5), i.e. both the baseline mortality and the total population distributions do not vary considerably during the years of study.
Since on average our WRF-Chem simulation overestimates PM 2.5 concentrations, likely because of the mismatch in the emission inventory year, among other reasons, the use of model output only-as done in prior studies with non-zero biases (e.g. Reddington et al 2019)-would overestimate the total annual premature mortality burden by almost 140 000 deaths in 2015 in our work (table 1), compared to the biascorrection methodology (Regression + Kriging).
According to the Regression + Kriging methodology, we estimate that in 2018 around 200 000 premature deaths were avoided compared to 2015. Several factors have contributed to this trend. During 2015-2018, population growth (+1.4% during 2015-2018) and ageing of the Chinese population (+1 year of median age) have slightly offset the reduction in premature deaths due to better air quality conditions,   whereas the decrease in baseline mortality rates positively reinforced such trend. Without any variations in the population age distribution and in baseline mortality, 20 000 additional deaths would have been avoided in 2018 compared to 2015, i.e. changes in population structure outweighed the benefits deriving from improved health conditions. The major improvements in terms of avoided premature deaths are concentrated in major urban areas (e.g. Beijing, Shanghai, Changchun, figure 3).    Nonetheless, some other areas have experienced a slight worsening of their mortality burden in the last few years. The most notable examples are the urban area of Xi'An and the westernmost cities of China, coherently with the findings (in terms of air quality) reported in Silver et al (2018). The spatial distribution of the difference in terms of total mortality values can be found in the Supplementary Information  (figure S6). The provincial-level partitioning of the total mortality burden indicates that the most populated provinces of Shandong and Henan are characterized by the highest premature mortality due to PM 2.5 (figure 4), although Shandong province achieved the largest improvements between 2015 and 2018 (25 000 avoided premature deaths, −18%). When normalizing by mortality, the highest reduction was around −28%, accomplished in the Jilin province in only four years. The leading coal producing regions of Shanxi and Shaanxi instead rank last for health improvements, with almost a zero net change during the study period, as also reported by Zou et al (2019). On the other hand, major steel-related industrial provinces (e.g. Hebei and Jiangsu, Reddington et al 2019) achieved important reductions of 13% and 10%, respectively. The northern, western and central provinces where the residential sector is a major contributor to poor air quality (Tao et al 2020) also showed a considerable decrease (e.g. Bejing, 21%, Sichuan, 11% and Hubei, 16%).

Uncertainty quantification
All the results presented above are related to the estimated mean of the mortality burden. However, such estimates are affected by three layers of uncertainty, associated with the PM 2.5 concentration fields, the RR values and the baseline mortality. Here, we identify the full probability distribution of the mortality estimates for the four different cases presented in section 2.5. Although all the numerical probability distributions are approximately Gaussian (even if some skewness is present, figure 5), the standard deviations are remarkably different. This implies that the traditional 95% confidence interval, computed only with the ERFs uncertainty, is narrower of about 700 000 premature deaths with respect to the case where the uncertainty in PM 2.5 fields, baseline mortality and ERFs is considered (table 2 and figure  S8). Because of a combination of the large sensitivity of GEMMs at high PM 2.5 concentrations and the log-normality of PM 2.5 concentrations, both the lower bound (2.5% percentile) and the upper bound (97.5% percentile) are significantly sensitive to the addition of the PM 2.5 layer of uncertainty. On the other hand, the baseline mortality does not play a major role in the premature mortality confidence intervals as it is known with higher accuracy (the average coefficient of variation of age-and disease-specific baseline mortality rates is approximately 3%).
Finally, figure 6 shows the comparison between the mortality estimates using GEMMs and the traditional IERs. The discrepancy between the two methods is significant, with the upper bound of IERs mortality estimates being approximately the mean of the GEMMs mortality distribution. Because the GEMMs relationship between PM 2.5 and mortality is much stronger at high PM 2.5 concentrations, the difference is more pronounced in the regions where PM 2.5 concentrations are particularly large (e.g. urban areas).

Discussion and conclusions
In this study, we quantify the long-term health burden associated with PM 2.5 exposure in China during 2015-2018. To our knowledge, this is the one of the first studies that integrates ground-based observations and high-resolution model output to accurately estimate the trend of nationwide premature mortality due to PM 2.5 for China. Our approach improves fidelity of premature mortality estimates by accounting and correcting for potential model biases in representing PM 2.5 concentrations. In addition, we present a novel methodology both to account for the uncertainty in CTM results (Solazzo and Galmarini 2016) and to properly quantify the confidence intervals of the premature mortality estimates. Despite many attempts to validate and improve CTM results (Vautard et al 2007), inaccuracies in CTM output systematically occur due to improper spatial resolution (Crippa et al 2017), lack of accurate input data (e.g. emissions estimates (Zheng et al 2017a)), uncertain initial and boundary conditions (Georgiou et al 2018), and incomplete understanding of some physical-chemical processes (e.g. formation of secondary organic aerosol (Giani et al 2019)). These issues can lead to potentially large biases in mortality estimates and thus we stress the importance of data assimilation techniques in calculating the mortality burden.
The statistical methodology presented in this work also allows for an improved characterization of the confidence intervals of mortality estimates, by considering and propagating multiple sources of uncertainty. We show that most of the uncertainty in mortality estimates is associated with the uncertainty in PM 2.5 fields and ERFs parameters, whereas baseline mortality is known with higher accuracy and therefore entails a minor impact on mortality confidence intervals. Hence, we argue that (i) uncertainty related to PM 2.5 estimate should be considered in mortality assessments, (ii) improving CTM results can considerably narrow the confidence intervals in mortality estimates and (iii) that some of the variability in mortality estimates presented in previous literature is likely coming from the neglected PM 2.5 uncertainty.
Results from our analyses suggest that 1 656 500 people prematurely died in 2015 due to PM 2.5 exposure, although a steady decline is estimated during 2015-2018 (~200 000 avoided deaths in 2018 with respect to 2015 over the entire China, which corresponds to an estimated saved monetary value of 195 billion US dollars). Using GEMMs instead of IERs leads to a significantly larger premature mortality burden compared to previous estimates (which usually range between 900 000 and 1 300 000, Reddington et al (2019)), as also highlighted by other studies, e.g. Burnett et al (2018) and Xue et al (2019b). The uncertainty related to the total premature deaths attributable to PM 2.5 remains vast, especially when taking into account different exposure response functions as well as the uncertainty related to PM 2.5 levels. If we consider both the results obtained with IERs and GEMMs together (i.e. assigning an uninformative prior probability of 50% to each of the two models), the range of our estimate (95% confidence interval) would be 574 000-2078 700 in 2018, which is much wider than previously reported. Although GEMMs are likely to be more representative of the health response to ambient PM 2.5 concentrations than IERs, as the former only include cohort studies of outdoor air pollution, there is no clear evidence of GEMMs providing more reliable results than IERs yet (Xue et al 2019b). More research and epidemiological studies are thus needed to constrain and evaluate the different exposure response functions as well as to narrow the current confidence intervals. Nonetheless, the trend of avoided deaths in recent years (~200 000 avoided deaths in 2018 compared to 2015, or 50 000 avoided deaths per year) is coherent with recent findings (42 800 avoided deaths per year, between 2013 and 2017, Zou et al (2019)), where estimates were derived from the use of a corrected version of the IERs to take into account exposure-related human activity patterns. Despite a considerably different total mortality burden, a similar trend in recent years was also found also in Ding et al (2019) (57 400 avoided deaths per year, between 2013 and 2017), which however relied on CTM simulations only, without any observations. The similarity of these three results, obtained with three independent methodologies, bolsters the robustness of the estimated trends.
Multi-year assessments of premature mortality usually require (i) large computational power to run expensive multi-year simulations and (ii) accurate and up-to-date emission inventories at fine resolution. Whilst computational power has been increasingly more accessible in recent years, CTMready official emission inventories are often available for discontinuous years and released belatedly. The main advantage of our approach is that it partly compensates the need for large computational power (thus allowing for very high resolution in space) and it does not require yearly varying emission inventories which generally require considerable processing before utilization as input to CTMs. Moreover, the presented methodology allows the calculation of the uncertainty in PM 2.5 spatio-temporal fields and its propagation to the mortality estimates, which overcomes the limitations of deterministic CTM output. However, we acknowledge that our methodology should be used for limited and consecutive years (e.g. 4 years, in our work), to ensure that the base case simulation provides reasonable results for different meteorological and emission conditions. This procedure can thus be intended to fill the gaps for those years where official emission inventories are not available, provided that observations of PM 2.5 are accessible. Future investigations should be devoted to the refinement of the presented methodology, for example using non-parametric approaches to estimate the PM 2.5 field and uncertainty (e.g. with machine learning techniques (Bai et al 2019, Xue et al 2019a). Population uncertainty, estimated with different global products (e.g. the Gridded Population of the World, v4), can also be included in the Monte Carlo assessment, even though we anticipate that its impact would be of the same order of the baseline mortality. Furthermore, this methodology could prove useful to quantify mortality trends in highly populated regions where the health impacts of air quality have been only partially quantified.
Although we find substantial reductions in the premature mortality rates associated with PM 2.5 during 2015-2018, and this trend has been reported since 2013 (Zheng et al 2017b), the health burden is still vast and multiple challenges await China in the near future. Climate change is expected to lead to more intense and frequent stagnation events and heat waves, which can exacerbate the PM 2.5 related premature mortality (Hong et al 2019). Even though emission reduction policies might be strengthened in the near future (Peng et al 2017), the premature mortality burden could still be negatively affected by the predicted population growth and ageing (Madaniyazi et al 2015, Cohen et al 2017, i.e. a slight reduction in PM 2.5 concentration could be more than compensated by an increase in the exposed population. Furthermore, socioeconomic factors like economic growth, industrialization and urbanization will strongly influence PM 2.5 concentrations (Li et al 2016a). As the economic growth is a vital goal in China's 13th five-year plan (Aglietta and Bai 2016), the government is urged to find a way to decouple the effects of economic growth on PM 2.5 pollution.
In light of our findings and other recent studies (Silver et al 2018), we conclude that China is headed in the right direction towards the decoupling of these two vital aspects, even though more stringent emission reduction policies will be needed to cope with all the challenges highlighted above.