Air quality improvements and health benefits from China's clean air action since 2013

The Weather Research and Forecasting (WRF) model version v3.5.1 [19] and the Models-3 Community Multiscale Air Quality (CMAQ) model version 5.1 [20] were used to simulate daily variation of PM_2.5 concentrations in China with a horizontal resolution of 36 km × 36 km. The WRF model is driven by the National Centers for Environmental Prediction Final Analysis (NCEP-FNL) reanalysis data [21] as initial and boundary conditions. Meteorological parameters simulated by the WRF model together with emission inventories were applied to drive the CMAQ model. Anthropogenic emissions for mainland China during 2013–2015 were derived from the Multi-resolution Emission Inventory of China (MEIC) model [22], while anthropogenic emissions for other Asian countries and regions were obtained from the MIX Asian emission inventory [23]. Details of model configurations are documented in Zheng et al [24] as well as in the supplementary information ((SI) text S1 available at stacks.iop.org/ERL/12/114020/mmedia).

In this study, the WRF–CMAQ modeling system was utilized in two parts. First, continuous PM_2.5 simulations from the WRF–CMAQ model were used as a major dataset to retrieve our optimal PM_2.5 concentrations in the three-stage data fusion model. Second, a set of sensitivity simulations based on the WRF–CMAQ model and the continuous PM_2.5 simulations were used together to quantify the impacts of meteorological variations on air quality as sensitivity analysis.

In this study, we applied a three-stage data fusion model proposed in our previous study [17] to predict PM_2.5 concentrations in China (excluding Taiwan, map of China is illustrated in figure S1) from 2013–2015, whose spatial resolution is 0.1° × 0.1°. Our first-stage model utilized a modified linear mixed-effects (LMEs) model [25] to predict surface PM_2.5 by calibrating satellite-based AOD and PM_2.5 simulations from the WRF–CMAQ model using routine measurements. The second-stage model fused information from satellite-based PM_2.5 and calibrated-CMAQ PM_2.5 estimates from the first-stage model based on the maximum likelihood estimator (MLE). In the third-stage model, the MLE residuals were interpolated over the whole study domain by the spatiotemporal Kriging [26] approach. The optimal PM_2.5 estimates were then derived from the sum of the fused PM_2.5 values obtained in the second stage and the residuals calculated in the third stage. The PM_2.5 concentrations for each county (i.e. county average) were then derived by zonal averaging over our optimal PM_2.5 surfaces, and the derived county level PM_2.5 maps were utilized as the base datasets for further analysis. We calculated national and provincial population-weighted PM_2.5 concentrations based on the county level PM_2.5 maps and county level population data from the 2010 China census. All aftermentioned national and provincial PM_2.5 concentrations present population-weighted values. Details on the PM_2.5 predictions are documented in Xue et al (2017) [17].

We applied a strict cross-validation (CV) procedure to evaluate our optimal PM_2.5 maps against surface measurements (SI: text S2). The CV on the daily estimates during the three years implies a good performance of our fusion model with correlation coefficient (R) value and normalized mean bias (NMB) of 0.84 and −7.5%, respectively, which are comparable to previous studies [27–30]. We also evaluated our daily estimates by years. According to the CV results, mean PM_2.5 observations for the three years were 70.8, 56.3, and 50.2 μg m⁻³, respectively, while mean PM_2.5 estimates were 63.9, 51.4, and 46.9 μg m⁻³, respectively. The CV results indicates that our model performs well in all three years with R values and NMB ranging from 0.83 to 0.88 and from −9.6% to −6.6%, respectively. The CV results also shows that mean PM_2.5 observations and mean PM_2.5 estimates reduced by 20.5% and 19.6% during 2013–2014, respectively, and reduced by 10.8% and 8.8% during 2014–2015, respectively. The consistent reduction ratios of PM_2.5 concentrations from observation and our estimates during the three years indicate the credibility of annual PM_2.5 variations estimated in this study (SI: table S3). The estimated PM_2.5 levels in this study are also comparable to that predicted in previous studies. For example, the population-weighted PM_2.5 concentration in 2013 predicted in this study was 60.5 μg m⁻³, while Brauer et al [31] reported a value of 55 μg m⁻³, and Zhang et al [32] reported a value of 61 μg m⁻³. We thus utilized the derived optimal PM_2.5 estimates as the base dataset for further analysis.

To identify the significance of the PM_2.5 abatements during the analysis period, we applied the Mann–Kendall trend analysis [33] to daily estimates of PM_2.5 concentrations at the county level. Mann–Kendall trend analysis is a non-parametric trend test which utilizes the relative magnitude of the data rather than the absolute value to evaluate significance of trends [34]. This estimator has been widely applied to investigate trends in environmental time series datasets (e.g. Du et al [35], Shen et al [36]).

To evaluate the premature mortality attributable to PM_2.5 exposure, the relationships between chronic exposure and attributable mortalities, which are called C–R relationships, must be determined. The IER functions developed for the Global Burden of Diseases Study include cohort studies at high levels of exposure concentrations (i.e. active and second-hand tobacco smoke) and provide C–R relationships for a full range of ambient PM_2.5 distributions [18], which are suitable for relevant studies in China. We utilized the IER relationships to calculate the premature mortality attributable to chronic PM_2.5 exposure in China for adults (age ≥ 25). Four leading endpoints were considered, including lung cancer (LC), chronic obstructive pulmonary disease (COPD), ischemic heart disease (IHD), and stroke. The relative risk (RR) for each endpoint can be obtained by equation (1) [18]:

$$\begin{equation} RR(C)=\left\{ \begin{array}{l@{}l@{}l@{}} 1+\alpha(1-{\rm e}^{-\gamma(C-C_0)^\delta}), & \, {\rm if} \ C < C_0\\ 1, & \, {\rm if} \ C \le C_0 \end{array}\right., \end{equation} \tag{ 1 }$$

where C is the ambient PM_2.5 exposure concentration; C₀ is the endpoint-specific theoretical minimum-risk concentration of PM_2.5; and α, γ, and δ are parameters that determine the shape of C–R curves. A distribution of 1000 point estimates of C₀, α, γ, and δ parameters provided by the IER were utilized to calculate mean RR and its 95% confidence intervals (CIs). We further employed the following equation to calculate the endpoint-specific (subscript: e) premature mortalities attributable to PM_2.5 exposures for the population subgroup p (population by age, sex, and urban/rural residence) in county c:

$$\begin{equation} {{{M}}_{p,e,c}}\left( {{{{C}}_c}} \right) = {{{P}}_{p,c}} \cdot \,{{{B}}_{p,e}} \cdot \frac{{{{R}}{{{R}}_{p,e}}\left( {{{{C}}_c}} \right) - 1}}{{{{R}}{{{R}}_{p,e}}\left( {{{{C}}_c}} \right)}}, \end{equation} \tag{ 2 }$$

where P_p,c represents county-level population amount of a population subgroup, B_p,e represents the national-level cause-specific mortality incidence for a population subgroup and RR_p,e(C_c) represents the relative risk of endpoint e among a population subgroup p during exposure to a specific PM_2.5 level (i.e. C_c). Detailed county-level demographic information was obtained from the 2010 China census, which is the most recent and accurate national census. To exclude the impacts of population growth and mortality incidence changes on the estimated mortality variations, we applied mortality incidence of 2010 for mortality estimation during 2013–2015. The cause-specific national mortality incidence statistics for all population subgroups in 2010 were obtained from the Ministry of Health of China [37] (SI: table S4). Because IER provides age-specific C–R functions for stroke and IHD, the attributable mortality of these two endpoints are calculated by age [18].

Zoom In Zoom Out Reset image size

Figure 1 illustrates the optimal PM_2.5 estimates in China from 2013 to 2015. The names of each Chinese province are depicted in figure S1. In 2013, high levels of PM_2.5 concentrations were observed in northern, eastern and central China, especially in the BTH region and the surrounding provinces, including Shandong, Shanxi, and Henan. Hebei was the most polluted province in 2013 and presented population-weighted PM_2.5 concentrations of 102 μg m⁻³. All the most polluted counties with annual average PM_2.5 concentrations greater than 130 μg m⁻³ in 2013 are located in Shijiazhuang City and Xingtai City, which are both prefecture-level cities in Hebei Province.

The nationwide PM_2.5 concentrations decreased considerably, especially in the most polluted regions (i.e. BTH and surrounding provinces) as shown in figures 1(a)–(c). During the analysis period, the national population-weighted PM_2.5 concentration declined from 60.5 μg m⁻³ in 2013 to 47.5 μg m⁻³ in 2015, representing a reduction of 21.5% in two years. Although the derived PM_2.5 trend is directly comparable to the trend from monitoring data, the monitoring data confirmed the decrease of PM_2.5 pollution levels; for the 74 cities with full coverage of surface PM_2.5 concentration data, the annual mean PM_2.5 concentrations decreased by 23.6% during the same period (from 72 μg m⁻³ to 55 μg m⁻³) [38].

Figure 2 further confirms the marked reduction of population-weighted PM_2.5 concentrations at the provincial level. The population-weighted PM_2.5 at the provincial level showed an overall reduction (figure 2(a)). From 2013 to 2015, the population-weighted PM_2.5 concentrations declined in all 31 provinces. The greatest reductions, at more than 20 μg m⁻³, were observed in Hebei Province, Tianjin Municipality, and Henan Province, and reductions that exceeded 15 μg m⁻³ were observed in Shandong, Hubei, Anhui, Jiangsu, Shaanxi, and Shanxi provinces.

Zoom In Zoom Out Reset image size

Moreover, reductions of more than 20% were observed in 17 provinces, and reductions of more than 15% were observed in the remaining eight provinces. These findings indicate a nationwide alleviation of PM_2.5 pollution. Figure 2(c) and (d) illustrate the distribution of population exposed to various PM_2.5 levels. The population exposure level clearly shifted towards a lower value during 2013–2015, with population exposed to annual PM_2.5 levels exceeding the World Health Organization Air Quality Interim Target-1 (WHO-IT1) standard (i.e. 35 μg m⁻³) declining significantly from 80% to 66%.

Figure 1(d) shows the reduction of annual PM_2.5 concentrations during 2013–2015 for counties where significant decreasing trends of PM_2.5 were identified by the Mann–Kendall test. The PM_2.5 concentrations declined significantly for 87% of the total counties (91% of the total population), and the reduction of PM_2.5 concentration at the county level reached up to 63.5 μg m⁻³. All of the counties with a PM_2.5 reduction greater than 50 μg m⁻³ were located in Shijiazhuang and Xingtai, the two most polluted cities in the country, which indicates the effectiveness of aggressive air quality improvement measures in the most polluted regions in China. The Mann–Kendall test was also applied to the daily average national population-weighted PM_2.5 concentrations and the result confirmed the significant reduction of national-level PM_2.5 exposure in China during 2013–2015.

Our study estimated that 1.22 million (95% CI: 1.05, 1.37) premature adult deaths were attributable to PM_2.5 pollution over China in 2013, in which stroke, IHD, COPD, and LC contributed 51%, 25%, 14%, and 10%, respectively (figure 3). The estimated mortality in 2013 is close to the results of recent studies that estimated PM_2.5 attributable mortality in China using the IER functions with values from 1.1–1.37 million [5, 10, 39–41]. The similar distribution patterns of PM_2.5 concentrations and population (see SI: figure S5) aggravate the public health burden. In 2013, four provinces (Shandong, Henan, Jiangsu, and Sichuan) with annual PM_2.5-attributable mortality greater than 90 000 were all populous and highly polluted, with populations greater than 70 million and provincial PM_2.5 levels above 60 μg m⁻³. In 2013, Shandong, which had a population of 96 million (70 million adults) and a PM_2.5 level of 87.4 μg m⁻³, had the highest annual PM_2.5-induced mortality, with a value of 113 800 (95% CI: 97.8, 128.1).

Zoom In Zoom Out Reset image size

Improved air quality has alleviated the disease burden attributable to PM_2.5 exposure in China. During 2013–2015, declined PM_2.5 concentrations might have led to a 110 800 (95% CI: 94.8, 127.4) (9.1%) reduction of the PM_2.5-attributable mortality. Among all endpoints, stroke accounted for 46% of the total mortality reduction and COPD contributed to approximately 21% of the total reduction (figure S6). The different shapes of the C–R functions for each endpoint led to variations in the reduction ratio of each endpoint-induced mortality. The mortality caused by LC and COPD decreased by over 14%, whereas the mortality caused by IHD and stroke decreased by only 6% and 8%, respectively.

Driven by the decline of ambient PM_2.5 levels, all provinces showed reduced PM_2.5-attributable mortality during 2013–2015 (figure 3). In all these provinces, Guangdong achieved the highest reduction in mortality, with a value of 9942 (95% CI: 7688, 12050), and Henan, Jiangsu, Shandong, Sichuan, Hunan, Zhejiang, Hebei, Fujian, Jiangxi, Hubei, and Anhui provinces each achieved reductions in mortality of over 5000 for the analysis period, which accounted for 70% of the national mortality reduction. The considerable reduction of mortality in these provinces can be attributed to their high base mortality in 2013 and the significant reduction of PM_2.5 concentrations. Most of these 12 provinces are populous, with more than 50 million residents (except for Fujian and Jiangxi), and population-weighted PM_2.5 concentrations in all these provinces decreased by more than 18% between 2013 and 2015. In addition to these 12 provinces, 12 other provinces had mortality reductions of more than 1000. In contrast, reduced mortality in comparatively clean regions including Ningxia, Qinghai, and Tibet are all less than 400 owing to comparatively small populations and low base PM_2.5 concentrations. Specifically, air quality improvements during 2013–2015 only led to 45 (95% CI: 25, 73) abated attributable deaths in Tibet, which can be attributed to its low population (about three million) and extremely low PM_2.5 concentrations (around 8 μg m⁻³ during the three years).

Because the shapes of the C–R functions utilized in this study are supralinear, the reduction of PM_2.5 concentrations may not have produced consistent mortality reduction ratios. As depicted in figure 4, the highest PM_2.5 reduction ratio was 30%, in Jiangxi Province; however, the mortality reduction ratio in Jiangxi was only 16%. Tibet achieved the greatest PM_2.5-attributable mortality reduction ratio of 24%, and its population-weighted PM_2.5 concentrations declined by 8%. Tibet and Yunnan are the only two provinces that obtain higher reduction ratio in mortality than in PM_2.5 level because of the steep C–R relationship at lower concentrations [40]. During 2013–2015, PM_2.5 reductions in relatively clean provinces yielded higher marginal health benefits (i.e. mortality abatement induced by unit reduction of PM_2.5 level) than reductions in polluted provinces. For example, although Guizhou and Shandong both achieved PM_2.5 reductions of 20%, the mortality reduction was much higher in Guizhou (12%) than in Shandong (6%), which was primarily attributed to the lower PM_2.5 level in Guizhou Province.

Zoom In Zoom Out Reset image size