Spatiotemporal Changes in Fine Particulate Matter Pollution and the Associated Mortality Burden in China between 2015 and 2016

In recent years, research on the spatiotemporal distribution and health effects of fine particulate matter (PM2.5) has been conducted in China. However, the limitations of different research scopes and methods have led to low comparability between regions regarding the mortality burden of PM2.5. A kriging model was used to simulate the distribution of PM2.5 in 2015 and 2016. Relative risk (RR) at a specified PM2.5 exposure concentration was estimated with an integrated exposure–response (IER) model for different causes of mortality: lung cancer (LC), ischaemic heart disease (IHD), cerebrovascular disease (stroke) and chronic obstructive pulmonary disease (COPD). The population attributable fraction (PAF) was adopted to estimate deaths attributed to PM2.5. 72.02% of cities experienced decreases in PM2.5 from 2015 to 2016. Due to the overall decrease in the PM2.5 concentration, the total number of deaths decreased by approximately 10,658 per million in 336 cities, including a decrease of 1400, 1836, 6312 and 1110 caused by LC, IHD, stroke and COPD, respectively. Our results suggest that the overall PM2.5 concentration and PM2.5-related deaths exhibited decreasing trends in China, although air quality in local areas has deteriorated. To improve air pollution control strategies, regional PM2.5 concentrations and trends should be fully considered.


Introduction
Fine particulate matter (PM 2.5 ) is well known for its negative impacts on human health. Notably, it has contributed to 4.24 million deaths, according to the Global Burden of Disease study 2015 (GBD 2015) [1]. Exposure to PM 2.5 could lead to cardiovascular [2][3][4] and respiratory diseases [5,6], which have been extensively investigated in a number of epidemiological cohort studies. Although the mechanisms are not fully understood, exposure to PM 2.5 has been found to be linked to certain pregnancy outcomes [7,8] and disorders of the nervous system. Among the diseases associated with PM 2.5 concentrations, stroke, ischaemic heart disease (IHD), lung cancer (LC) and chronic obstructive

Ground Monitoring PM 2.5 Data
The hourly PM 2.5 concentrations from January 2015 to December 2016 in 336 cities were obtained from the China National Environmental Monitoring Center (CNEMC), and spatial coverage of the data is illustrated in Figure 1. The micro oscillating balance method and the beta absorption method were used to measure PM 2. 5 . The instruments which measured PM 2.5 concentration in each site were tested by using at least 3 samplers based on HJ 618-2011, according to the regulations published by the Ministry of Environmental Protection of the People's Republic of China [24]. The CNEMC website [25] is a public website that provides air quality data free of charge, and the data are from national air quality automatic monitoring sites and were initially reviewed by the release system. In this paper, daily, monthly and yearly PM 2.5 concentrations were all calculated based on hourly data. Although the source and acquisition process were credible, further filtering and correction were necessary because of some routine maintenance activities, communication failures and power outages at monitoring sites could lead to the absence of data. The data were transformed into z scores (standard scores), and data were removed when the following conditions were met: (1) less than 12 h of valid data were available for a day; (2) the absolute z score was larger than 4; or (3) the increase from the previous value was larger than 6 [26]. After the screening and deletion processes, more than 85% of the hourly data could be used for further calculation. The hourly PM2.5 concentrations from January 2015 to December 2016 in 336 cities were obtained from the China National Environmental Monitoring Center (CNEMC), and spatial coverage of the data is illustrated in Figure 1. The micro oscillating balance method and the beta absorption method were used to measure PM2. 5. The instruments which measured PM2.5 concentration in each site were tested by using at least 3 samplers based on HJ 618-2011, according to the regulations published by the Ministry of Environmental Protection of the People's Republic of China [24]. The CNEMC website [25] is a public website that provides air quality data free of charge, and the data are from national air quality automatic monitoring sites and were initially reviewed by the release system. In this paper, daily, monthly and yearly PM2.5 concentrations were all calculated based on hourly data. Although the source and acquisition process were credible, further filtering and correction were necessary because of some routine maintenance activities, communication failures and power outages at monitoring sites could lead to the absence of data. The data were transformed into z scores (standard scores), and data were removed when the following conditions were met: (1) less than 12 hours of valid data were available for a day; (2) the absolute z score was larger than 4; or (3) the increase from the previous value was larger than 6 [26]. After the screening and deletion processes, more than 85% of the hourly data could be used for further calculation.

Population and Mortality Data
The population and deaths reported at each monitoring site were collected, summarized and combined in the China Death Monitoring Data Set. The China Death Monitoring Data Set was edited by two departments, the National Health and Family Planning Commission of the People's Republic of China and the Chinese Center for Disease Control and prevention. The data quality at each monitoring site was evaluated before data collection, and sites with serious missing reports which might affect the overall results were excluded, as described in the dataset instructions. The latest data of the number and cause of deaths were provided by the China Death Monitoring Data Set 2014. This data set included the deaths from LC (C33-C34), IHD (I20-I25), stroke (I60-I69), and COPD (J40-J44), as well as the total number of deaths in the eastern region (including Beijing, Tianjin, Hebei, Liaoning, Shanghai, Jiangsu, Zhejiang, Fujian, Shandong, Guangdong and Hainan),

Population and Mortality Data
The population and deaths reported at each monitoring site were collected, summarized and combined in the China Death Monitoring Data Set. The China Death Monitoring Data Set was edited by two departments, the National Health and Family Planning Commission of the People's Republic of China and the Chinese Center for Disease Control and prevention. The data quality at each monitoring site was evaluated before data collection, and sites with serious missing reports which might affect the overall results were excluded, as described in the dataset instructions. The latest data of the number and cause of deaths were provided by the China Death Monitoring Data Set 2014. This data set included the deaths from LC (C33-C34), IHD (I20-I25), stroke (I60-I69), and COPD (J40-J44), as well as the total number of deaths in the eastern region (including Beijing, Tianjin, Hebei, Liaoning, Shanghai, Jiangsu, Zhejiang, Fujian, Shandong, Guangdong and Hainan), central region (including Shanxi, Jilin, Heilongjiang, Anhui, Jiangxi, and Henan) and western region (including Inner Mongolia, Guangxi, Chongqing, Sichuan, Guizhou, Yunnan, Tibet, Shaanxi, Gansu, Qinghai, Ningxia and Xinjiang) of China. Provincial population and mortality information was obtained from the provincial Statistical Bulletin of National Economic and Social Development.

Kriging Model
To visualize the regional profiles of PM 2.5 across China, a kriging model was applied to generate a nationally continuous surface based on the PM 2.5 concentrations in 336 cities. Kriging is a high-level statistical process used to generate an estimated surface based on a set of scatter points with z values. As a famous interpolation method, it was first used in the mineral industry and later used in a wide variety of disciplines, such as air pollution mapping [27]. Among the various versions of kriging, we employed ordinary kriging, a commonly used variant of the kriging algorithm [28,29], to construct an unbiased estimator without the requirement of a stationary mean of observed values in advance. The predictive formula of the kriging model is shown in Equation (1): where Z(S 0 ) is the predictive value, Z(S i ) is the measurement at position i, N is the number of measurements, and λ i is the unknown weight of the predictive value at position i. Here, weight relies on not only distance and the predictive positions of measurement points, but also on the spatial arrangement of measurement points. Therefore, the estimation of a semivariogram γ(h) at a distance of h is necessary to provide spatial autocorrelation information for the data set before creating predictive surfaces. To ensure that kriging predictions have positive kriging variances, it is necessary to fit a model-that is, a continuous function or curve-to the empirical semivariogram. Linear, circular, spherical, exponential and Gaussian are some models used to fit different types of phenomena [30,31].
For example, spherical model shows a progressive decrease of spatial autocorrelation until some distance, beyond which the autocorrelation is zero. The formula is shown in Equation (2): where C 0 , C 1 , C 0 + C 1 and a represent nugget, partial sill, sill and range respectively.

Integrated Exposure-Response (IER) Model
An increase in the PM 2.5 concentration could lead to higher rates of LC, IHD, stroke and COPD. To estimate the number of premature deaths caused by PM 2.5 in 2015 and 2016, we applied the IER model proposed by Burnett et al. The IER model is an effective predictor of relative risk. RR is the ratio of the probability of an event occurring in an exposed group to the probability of the event occurring in a comparable, non-exposed group. Notably, the IER model combines the RRs of ambient air pollution, second-hand tobacco smoke, household solid cooking fuel and active smoking, and yields reasonable predictions over a range of concentrations that prevail in China and other highly polluted areas. RR can be calculated using Equation (3): where z is the exposure to PM 2.5 and z cf is the PM 2.5 concentration below which no additional risk exists. Here, we define z cf as a uniform random variable between 5.8 µg/m 3 -the minimum concentration observed in the American Cancer Society Cancer Prevention II cohort [32]-and 8.8 µg/m 3 -the 5th percentile value. For very large z values, RR IER approximates 1 + α. Parameter δ is a power of PM 2.5 to predict risk over a very large range of concentrations. Further, RR IER (z cf + 1) approximates 1 + αγ. Thus, γ = [RR IER (z cf + 1) − 1]/[RR IER (∞) − 1] can be interpreted as the ratio of the RR at low-to-high exposures. α, γ, and δ are unknown parameters, estimated using nonlinear regression methods according to the RR information and the variance estimates of the logarithms of RRs at different PM 2.5 concentration from the available literature [33]. The detailed information about the estimates of (α, γ, δ), and the confidence interval for RR IER can be found in the supplemental material. Two assumptions used in this study should be noted before further analysis: (1) that the proportion of deaths caused by four diseases (LC, IHD, stroke and COPD) compared to total deaths was the same in 2015 and 2014 (this assumption was made because the 2015 China Death Monitoring Data Set has not been published as of the writing of this article), and (2) that city-level death rates are the same as the death rate of the associated province (this assumption was made because the death rates are unavailable for many cities).
The deaths impacts (DI) due to PM 2.5 exposure in each city can be estimated using Equation (4): where DI i,j is the number of deaths due to exposure to PM 2.5 for disease j in city i. Population i is the exposed population in city i. Mortality i,j represents the mortality of disease j in city i. PAF i,j is the population attributable fraction (PAF) of the cause-specific mortality of disease j in city i. Specifically, PAF can be determined using Equation (5): where RR Ci,j is the RR of the specific mortality of disease j at exposure level C i estimated from the IER model. Non-exposure mortality in each city can be computed based on the exposure mortality and PAF in 2015. Thus, cause-specific mortality in 2016 can easily be determined using Equation (6): Then, with the cause-specific mortality and PAF in 2015 and 2016, the mortality trends associated with changes in the PM 2.5 concentration per million people in each city can be obtained.  [36]. However, emissions caused by the increase in the number of motor vehicles and boilers, coupled with the influence of meteorological factors (increased regional sand weather, and rise of temperature and decrease of precipitation in the winter of 2015/2016) contributed to the aggravated PM 2.5 pollution in the west of Xinjiang Province [37]. Overall, 72.02% of cities observed a decrease in the PM2.5 concentration from 2015 to 2016, among which Hohhot and Xiaogan experienced the most significant decreases (of more than 20 μg/m 3 ). The PM2.5 concentration in 24.40% of cities increased by less than 10 μg/m 3 . In addition, 7 and 5 cities exhibited an increase in PM2.5 concentrations 10-20 μg/m 3 and over 20 μg/m 3 , respectively. At the provincial level, 23 out of 31 provinces exhibited decreased PM2.5 concentrations, and significant decreases were observed in Jilin and Hubei (Table S1). By contrast, Xinjiang exhibited the most obvious increase in its PM2.5 concentration among the remaining 8 provinces. All cities in Jilin, Hubei and Shandong exhibited declines in their PM2.5 concentrations, and PM2.5 in more than 90% of the cities in Heilongjiang, Inner Mongolia, Liaoning, Hunan and Zhejiang decreased. By contrast, the PM2.5 concentration in approximately 90.91% of the cities in Shanxi increased to some extent.

Spatial Distribution of PM 2.5
According to the WHO air quality guidelines (AQGs), an annual average concentration of 10 μg/m 3 is set as the long-term guideline value for PM2.5, over which significant effects on survival are observed. Besides this guideline, three interim targets are also defined and may benefit countries in gauging progress over time in the process of reducing PM2.5 exposures. An annual mean PM2.5 concentration of 35 μg/m 3 is defined as the IT-1 level, which is associated with about a 15% higher long-term mortality risk. The IT-2 level is set at 25 μg/m 3 , and this level lowers the risk of premature mortality by about 6%, relative to the IT-1 level. The IT-3 level is 15 μg/m 3 , reducing the mortality risk by approximately 6% relative to IT-2 level. The proportions of the hourly PM2.5 concentrations exceeding the WHO annual threshold values in 2015 and 2016 were not equally distributed at a national scale. Figure S1 shows the distribution of the hourly PM2.5 concentrations of all 336 cities in the ranges of ≤10 μg/m 3 , 10-15 μg/m 3 , 15-25 μg/m 3 , 25-35 μg/m 3 and >35 μg/m 3 . An obvious geographical difference in the distribution of the PM2.5 concentration can be observed in Figure S1. The concentration of PM2.5 has declined in recent years, but it is still far from the WHO standard. A previous study indicated that the areal proportion of China with a PM2.5 concentration of less than 35 μg/m 3 decreased from 1999 to 2011 [38]. Additionally, more than 50% of cities were in a stage of "non-attainment" in September, although September was the month when the largest number of cities exhibited minimum PM2.5 concentrations among 190 Chinese cities in 2014 [34]. Most monitoring sites exceeded the WHO standard, and only 0.38%, 0.41% and 12.93% of stations met the WHO IT-3, IT-2 and IT-1 thresholds, respectively, in China in 2015. Only 1% of susceptible people (of age 61 and over, or 13 and under) lived in areas with relatively safe levels of PM2.5 (less than 10 μg/m 3 ) in 2010 [39], and approximately 14% (181.08 million) of people were exposed to PM2.5 concentrations lower than the WHO IT-1 threshold in 2015 [40].  (Table S1). By contrast, Xinjiang exhibited the most obvious increase in its PM 2.5 concentration among the remaining 8 provinces. All cities in Jilin, Hubei and Shandong exhibited declines in their PM 2.5 concentrations, and PM 2.5 in more than 90% of the cities in Heilongjiang, Inner Mongolia, Liaoning, Hunan and Zhejiang decreased. By contrast, the PM 2.5 concentration in approximately 90.91% of the cities in Shanxi increased to some extent.
According to the WHO air quality guidelines (AQGs), an annual average concentration of 10 µg/m 3 is set as the long-term guideline value for PM 2.5 , over which significant effects on survival are observed. Besides this guideline, three interim targets are also defined and may benefit countries in gauging progress over time in the process of reducing PM 2.5 exposures. An annual mean PM 2.5 concentration of 35 µg/m 3 is defined as the IT-1 level, which is associated with about a 15% higher long-term mortality risk. The IT-2 level is set at 25 µg/m 3 , and this level lowers the risk of premature mortality by about 6%, relative to the IT-1 level. The IT-3 level is 15 µg/m 3 , reducing the mortality risk by approximately 6% relative to IT-2 level. The proportions of the hourly PM 2.5 concentrations exceeding the WHO annual threshold values in 2015 and 2016 were not equally distributed at a national scale. Figure S1 shows the distribution of the hourly PM 2.5 concentrations of all 336 cities in the ranges of ≤10 µg/m 3 , 10-15 µg/m 3 , 15-25 µg/m 3 , 25-35 µg/m 3 and >35 µg/m 3 . An obvious geographical difference in the distribution of the PM 2.5 concentration can be observed in Figure S1. The concentration of PM 2.5 has declined in recent years, but it is still far from the WHO standard. A previous study indicated that the areal proportion of China with a PM 2.5 concentration of less than 35 µg/m 3 decreased from 1999 to 2011 [38]. Additionally, more than 50% of cities were in a stage of "non-attainment" in September, although September was the month when the largest number of cities exhibited minimum PM 2.5 concentrations among 190 Chinese cities in 2014 [34]. Most monitoring sites exceeded the WHO standard, and only 0.38%, 0.41% and 12.93% of stations met the WHO IT-3, IT-2 and IT-1 thresholds, respectively, in China in 2015. Only 1% of susceptible people (of age 61 and over, or 13 and under) lived in areas with relatively safe levels of PM 2.5 (less than 10 µg/m 3 ) in 2010 [39], and approximately 14% (181.08 million) of people were exposed to PM 2.5 concentrations lower than the WHO IT-1 threshold in 2015 [40].

Temporal Trends of PM 2.5
The PM 2.5 concentration in 2016 was 2.27 µg/m 3 lower than that of 2015, which is consistent with the decreasing trend in PM 2.5 from 2014 to 2016 reported in a previous study [41]. Winter, spring, and summer experienced decreases in the PM 2.  Figure 3. The seasonal variations in air pollution were consistent between years [42]. Severe air pollution occurred in winter months because of high emissions from heating and unfavorable meteorological conditions for pollution dispersion (stagnant weather and temperature inversion) [43]. Crop residue burning can also induce an evident PM 2.5 increase in winter and autumn [31]. In contrast, sufficient precipitation and active atmospheric circulation led to low pollution on summer days [34,44,45]. The two curves illustrating the change in the PM 2.5 concentration over the two years differ and display different peak and valley trends and fluctuation ranges.

Temporal Trends of PM2.5
The PM2.5 concentration in 2016 was 2.27 μg/m 3 lower than that of 2015, which is consistent with the decreasing trend in PM2.5 from 2014 to 2016 reported in a previous study [41]. Winter, spring, and summer experienced decreases in the PM2.  Figure 3. The seasonal variations in air pollution were consistent between years [42]. Severe air pollution occurred in winter months because of high emissions from heating and unfavorable meteorological conditions for pollution dispersion (stagnant weather and temperature inversion) [43]. Crop residue burning can also induce an evident PM2.5 increase in winter and autumn [31]. In contrast, sufficient precipitation and active atmospheric circulation led to low pollution on summer days [34,44,45]. The two curves illustrating the change in the PM2.5 concentration over the two years differ and display different peak and valley trends and fluctuation ranges.  [18,40,41]. Enhanced anthropogenic activity during rush hour contributed to the morning peak in the PM2.5 concentration, which was studied through the comparison between PM2.5 diurnal changes in weekdays and weekends [46], as well as the contract between PM2.5 concentration diurnal trends in urban and rural areas [47]. Following the morning peak, the decrease in traffic pollution combined with the increase in convective movement led to the decline in the PM2.5 concentration. In the evening, traffic pollution and cooking emissions led to a gradual elevation of PM2.5 concentration and around midnight, electricity generation caused industrial pollution, increasing the PM2.5 concentration. Thus, seasonal shifts in peak and valley times can potentially be attributed to seasonal changes of residents' behaviors. However, not all cities exhibited diurnal variations that are consistent with the average condition of the 336 cities, which may be explained by the variety of urbanization level, sources of pollution and people's lifestyle. For example, PM2.5 concentration peaks appeared at 8:00 and 20:00 and valleys occurred at 1:00 and 15:00 in Guilin, Guangxi Province. In addition, Changde in Hunan exhibited four small peaks during a day but displayed a narrow  [18,40,41]. Enhanced anthropogenic activity during rush hour contributed to the morning peak in the PM 2.5 concentration, which was studied through the comparison between PM 2.5 diurnal changes in weekdays and weekends [46], as well as the contract between PM 2.5 concentration diurnal trends in urban and rural areas [47]. Following the morning peak, the decrease in traffic pollution combined with the increase in convective movement led to the decline in the PM 2.5 concentration. In the evening, traffic pollution and cooking emissions led to a gradual elevation of PM 2.5 concentration and around midnight, electricity generation caused industrial pollution, increasing the PM 2.5 concentration. Thus, seasonal shifts in peak and valley times can potentially be attributed to seasonal changes of residents' behaviors.
However, not all cities exhibited diurnal variations that are consistent with the average condition of the 336 cities, which may be explained by the variety of urbanization level, sources of pollution and people's lifestyle. For example, PM 2.5 concentration peaks appeared at 8:00 and 20:00 and valleys occurred at 1:00 and 15:00 in Guilin, Guangxi Province. In addition, Changde in Hunan exhibited four small peaks during a day but displayed a narrow range of fluctuation [42].

Disease Burden of PM 2.5
The PM 2.5 concentration, RR and standard error were obtained from the aggregated information in the supplemental material of Burnett et al.'s paper [33] for the burdens of diseases (including LC, IHD, stroke and COPD), based on published sources. The relationship between the PM 2.5 concentration and RR was calculated using the IER model, as shown in Figure 4. The curves of LC and COPD are almost straight lines, which suggest that RR increased uniformly as the PM 2.5 concentration increased. For the curves corresponding to stroke and IHD, RR increased rapidly when the PM 2.5 concentration was below 50 µg/m 3 and then increased slowly. Deaths attributed to PM 2.5 based on the actual populations in 31 provinces and deaths per million residents in 336 cities were calculated based on the PAF. The PM2.5 concentration, RR and standard error were obtained from the aggregated information in the supplemental material of Burnett et al.'s paper [33] for the burdens of diseases (including LC, IHD, stroke and COPD), based on published sources. The relationship between the PM2.5 concentration and RR was calculated using the IER model, as shown in Figure 4. The curves of LC and COPD are almost straight lines, which suggest that RR increased uniformly as the PM2.5 concentration increased. For the curves corresponding to stroke and IHD, RR increased rapidly when the PM2.5 concentration was below 50 μg/m 3 and then increased slowly. Deaths attributed to PM2.5 based on the actual populations in 31 provinces and deaths per million residents in 336 cities were calculated based on the PAF.  Table 1. The total number of deaths due to the four diseases was largest in Henan, at more than 130,000, while the lowest total was observed in Tibet, at less than 2000. This finding reflects an obvious difference between provinces. At the provincial level, 21 provinces experienced different declines in deaths due to changes in PM2.5. Heilongjiang, Hunan and Hubei were the top three provinces with the most obvious decreases of 6391, 5304 and 4705 deaths, respectively. In the remaining 10 provinces, Sichuan, Shanxi and Shaanxi exhibited the most significant increases from 2015 to 2016, with 2769, 2697 and 2553 more deaths, respectively. The decreases or increases in deaths per million residents related to PM2.5 in 336 Chinese cities were calculated, and the results are presented in Table S2 and Figure 5. At the city level, approximately 80% of cities exhibited fewer deaths due to the four diseases. For LC, IHD and COPD, the distributions of changes in deaths were similar, with approximately 70% of the cities exhibiting a  Table 1. The total number of deaths due to the four diseases was largest in Henan, at more than 130,000, while the lowest total was observed in Tibet, at less than 2000. This finding reflects an obvious difference between provinces. At the provincial level, 21 provinces experienced different declines in deaths due to changes in PM 2.5 . Heilongjiang, Hunan and Hubei were the top three provinces with the most obvious decreases of 6391, 5304 and 4705 deaths, respectively. In the remaining 10 provinces, Sichuan, Shanxi and Shaanxi exhibited the most significant increases from 2015 to 2016, with 2769, 2697 and 2553 more deaths, respectively. The decreases or increases in deaths per million residents related to PM 2.5 in 336 Chinese cities were calculated, and the results are presented in Table S2 and Figure 5. At the city level, approximately 80% of cities exhibited fewer deaths due to the four diseases. For LC, IHD and COPD, the distributions of changes in deaths were similar, with approximately 70% of the cities exhibiting a decrease in deaths-ranging from 20% to 50%-and 17% of cities exhibiting an increase in deaths-ranging from 0% to 20%. The proportions of cities in different ranges were similar for stroke. In terms of the change in the total number of deaths due to PM 2.5 from 2015 to 2016, Baoshan, Yichun and Hohhot exhibited the most obvious declines among the 231 cities with fewer deaths, while 23 cities remained unchanged. The remaining 82 cities exhibited more deaths, with Qiannan, Chizhou and Xianyang experiencing the three highest increases.
Our study indicated that approximately 1,126,000 deaths were caused by PM 2.5 in 2015 across China. This total is slightly larger than the value estimated by GBD 2015 (1,108,000) [48]. Moreover, our estimate decreased to approximately 1,092,000 in 2016 due to the overall decline in the PM 2.5 concentration. Notably, the numbers of deaths due to LC, IHD, stroke and COPD caused by PM 2.5 totalled 130,000, 284,000, 592,000 and 120,000, respectively, in 2015 and 124,000, 278,000, 573,000 and 117,000, respectively, in 2016. Our estimates are comparable with those of previous studies. These studies found that premature deaths in China caused by PM 2.5 totaled 807,000 in 2004, 1,250,000 in 2012 [49], 1,367,000 in 2013 [50], and 1,600,000 in 2014 [51]. In 2015, the number of deaths per million people attributed to PM 2.5 was 813 in China. This total was more than four times that in the United States and Japan. Additionally, China accounted for approximately a quarter of global deaths (4,241,000), which suggests that air pollution remains a serious problem in China. However, according to our research, air pollution in China improved significantly from 2015 to 2016, based on the fact that 72.02% of the cities exhibited varying decreases in their PM 2.5 concentrations, and 68.75% of the cities exhibited decreases in mortality attributed to PM 2.5 . Improvements in air quality in China indicate that the policies and measures taken by the government have achieved initial success. To ensure the continuity and effectiveness of air pollution control, it is necessary to reduce heating, cooking and agricultural emissions [52] and adjust regional policies according to changes in PM 2.5 concentration.
deaths-ranging from 0% to 20%. The proportions of cities in different ranges were similar for stroke. In terms of the change in the total number of deaths due to PM2.5 from 2015 to 2016, Baoshan, Yichun and Hohhot exhibited the most obvious declines among the 231 cities with fewer deaths, while 23 cities remained unchanged. The remaining 82 cities exhibited more deaths, with Qiannan, Chizhou and Xianyang experiencing the three highest increases.

Strengths and Limitations
Few studies have explored the PM 2.5 -related mortality burden using local health data (including population and mortality data) in China, and most studies have focused on key cities or regions instead of the entire country. Due to the adoption of different algorithms, low comparability exists between these studies. Our study used hourly, ground-level PM 2.5 monitoring data, population data and mortality data to estimate the deaths in China caused by PM 2.5 using the IER model and PAF. This approach overcomes the limitations of small-scale research. Moreover, these findings can be used to inform people regarding the health impacts of PM 2.5 pollution and provide guidance for regional policy design.
However, some limitations and uncertainties exist in our study. The evaluation was based on city-level PM 2.5 concentrations calculated by averaging the data at all sites, which is the common method used to report daily air quality to the public. However, the uneven distribution of monitoring stations, including more in urban areas and less in suburban and rural areas, makes a simple averaging method less accurate than more complex methods [40]. In addition, the assumptions proposed before further calculations were sources of uncertainties as well, but the assumptions were inevitable due to the unavailability of certain data. Furthermore, the IER model was adopted to estimate the relationships between PM 2.5 and the RRs of four diseases in our study. However, a highly accurate exposure-response model based completely on Chinese cohort studies should be developed in the future, even though the IER model provides reasonable predictions in China and other heavily polluted areas [49].

Conclusions
The spatial distributions, temporal trends and health burdens of PM 2.  19,193 stroke deaths and 3168 COPD deaths) were avoided due to the overall decrease in the PM 2.5 concentration in China from 2015 to 2016. New findings in this paper could enhance public awareness regarding the health risks caused by PM 2.5 . We urge governments in Chinese provinces and cities to take effective measures to curb air pollution. In addition, to avoid a rebound in PM 2.5 pollution, sustainable strategies and joint actions among cities should be considered.