Human mortality attributable to outdoor air pollution in China during the period 2016–2020

In this study, a latest reanalysis dataset of atmospheric composition, the Global Exposure Mortality Model and a log-linear exposure-response function were employed to estimate the national deaths attributable to fine particulate matter (PM2.5) and ozone (O3) pollution in China for the period 2016–2020, including the lockdown due to COVID-19 pandemic in 2020. The national mortality attributable to long-term PM2.5 exposure decreased year by year from 2.18 million (95% confidence interval (1.83, 2.51), the same hereinafter) in 2016 to 1.99 million (1.66, 2.30) in 2020. In particular, the number in 2020 was 133.16 thousand less than 2019 owing to the reduced emissions during the pandemic, and the mortality attributable to short-term PM2.5 exposure dropped from 46.86 thousand in 2019 to 36.56 thousand in 2020. However, because O3 concentrations have kept increasing during the period, the national mortality attributable to long-term O3 exposure increased from 132.79 thousand (128.58, 137.00) in 2016 to 197.00 thousand (190.98, 203.03) in 2020. In addition, compared to before the pandemic, the national mortality attributable to short-term O3 exposure showed an increase in February, April and May of 2020, and the sharpest year-on-year increase of 162% occurred in April. The different trends of mortality after anthropogenic emissions were reduced pose a challenge for policy-makers and researchers.


Introduction
Ambient air pollution is an important environmental risk to human life. Among the pollutants harmful to human health, fine particles with a diameter of 2.5 µm or less (PM 2.5 ) and ozone (O 3 ) play an important role in inducing deaths, especially in China where PM 2.5 and O 3 are two key pollutants (Gao et al 2015, Wang et al 2020a, Kou et al 2021, Liu et al 2021. To mitigate air pollution, Chinese government started to implement the Clean Air Action Plan in 2013 aimed at reducing anthropogenic emissions (State Council of the People's Republic of China 2013). In general, the effects of reduced emissions since 2013 are credited to the decreased PM 2.5 concentrations thereafter (e.g. Zhang et al 2019), but the effects are uncertain for O 3 owing to the complicated relationships between O 3 and its precursors (Ding et al 2019), and some researchers have even found O 3 pollution has been showing an increasing trend (e.g. Li et al 2019a, Chen et al 2021, Liu et al 2021. Furthermore, anthropogenic emissions were greatly reduced in China in 2020 because of the unprecedented large-scale lockdown imposed in response to the COVID-19 pandemic. Therefore, in this context, it is meaningful to investigate the impact of both long-term and short-term exposure to outdoor air pollution on human mortality under the influence of the implementation of the policies and regulations, while doing so requires simulating the air pollutant concentrations accurately and then accurate simulation can enable the estimation of human mortality induced by air pollution. Moreover, against the background of the pandemic causing a direct loss of human life in China in 2020, it is also meaningful to investigate the health impacts of the reduced air pollution. This is an indirect impact of the pandemic on human health, and research in this respect will benefit policy-makings in the future in terms of the control of air pollutant emissions. So far studies on the impacts of outdoor air pollution on death have not yet been very many and ample on the effect of the COVID-19 lockdown, because the pandemic just began in 2020. Hao et al (2021) studied the long-term health impacts of PM 2.5 in 2020, but they used predicted monthly PM 2.5 concentrations for 2020 due to data unavailability and only focused on selected Chinese cities. It would be meaningful and interesting to investigate the long-term and short-term health impacts of outdoor air pollution related to the pandemic in 2020 across the whole China based on the latest available data of atmospheric composition.
Quantifying PM 2.5 and O 3 concentrations and their distributions as precisely as possible is a necessary condition for estimating the number of deaths attributable to outdoor air pollution. Employing time series of observed concentrations is the most straightforward method for estimating the related human mortality. For example, Hao et al (2021) used the time series of observed PM 2.5 concentrations for the period 2015-2019 and predicted PM 2.5 concentrations for 2020 to study the long-term health impacts of PM 2.5 in Chinese cities. However, because observational sites are spatially sparse and their distribution is usually inhomogeneous, and observational data are often temporally interrupted, observational datasets are not fit for the type of estimation. Instead, results from numerical simulations are employed. For example, Zhang et al (2019) used a chemical transport model and an emissions inventory to study the effect of reduced emissions on the PM 2.5 concentration and its health impacts during 2013-2017 in China. Guttikunda and Goel (2013) used a numerical dispersion model and local meteorological data to estimate the health impacts of particulate pollution in 2010 in Delhi, India and its satellite cities. In their research, however, the model errors had not been reduced as much as possible, since either meteorological conditions or emissions were constant and unchanged.
Reanalysis datasets of atmospheric composition incorporate the advantages of numerical simulations and observations. As they are homogeneous and continuous spatially and temporally, they have been employed to estimate pollution-related human mortality in recent years. For Gao et al (2015), Gao et al (2017) used reanalysis data to study the health impacts of PM 2.5 pollution during the winter haze in 2013 in Beijing, China as well as in Eastern China. Moreover, an atmospheric environmental monitoring network was built in China in 2013 to carry out chemical observations. The observational datasets have enabled the improvement of air quality forecasting via data assimilation and thus benefitted reanalysis datasets of atmospheric composition. Liu et al (2021) for the first time employed long-term reanalysis data for the period 2013-2018 to estimate the human mortality attributable to outdoor air pollution in China. However, source emissions were constant and not changing with time in their study.
In recent years, a joint data assimilation system that produced a reanalysis dataset of atmospheric composition was developed, which optimized source emissions as well as concentrations (Kou et al 2021). Moreover, meteorological fields were also assimilated with inputted meteorological observations. In this study, we used a 5 year high-resolution atmospheric composition reanalysis dataset for the period 2016-2020 to estimate the human mortality attributable to outdoor PM 2.5 and O 3 exposure in China as well as investigate the health impacts of the lockdown imposed due to the COVID-19 pandemic in 2020. The dataset, which was produced by a joint data assimilation system that assimilated meteorological and chemical fields as well as source emissions simultaneously, can provide concentration fields of conventional air pollutants and an updated emissions inventory. Moreover, the effects of reduced emissions in recent years can be studied against the background of both changing meteorological conditions and changing emissions. Since the dataset is the latest reanalysis product of atmospheric composition available in China and it includes the assimilation of meteorological and chemical observations as well as source emissions, it is employed in this study to investigate the effects of the policies, which can benefit policy-makings in the future.

Methodology
The methodology used in this study to estimate the human mortality attributable to outdoor PM 2.5 and O 3 pollution is briefly described as follows: 2.1.1. Deaths attributable to long-term exposure to PM 2.5 and O 3 The number of deaths attributable to long-term exposure to PM 2.5 and O 3 was estimated using equation (1), where M is the human mortality attributable to air pollution exposure of a specific year, P i, j is the population of grid cell i in age group j, y i, j, k is the baseline mortality rate of grid cell i in age group j for cause of death k, and RR i, j, k is the relative risk function of grid cell i in age group j for cause of death k. In this study, the 20 age groups include 0-4, 5-9, 10-14, 15-19, 20-24, 25-29, 30-34, 35-39, 40-44, 45-49, 50-54, 55-59, 60-64, 65-69, 70-74, 75-79, 80-84, 85-89, 90-94 and over 95 years, while the three causes of death include noncommunicable diseases and lower respiratory infections (NCD+LRI), chronic obstructive pulmonary disease (COPD) and all-cause. The relative risk function of long-term PM 2.5 exposure for NCD+LRI used by the Global Exposure Mortality Model (GEMM) is thought to be better than the widely used integrated exposure-response model in providing estimates for highly polluted areas such as China (Burnett et al 2018, Chen et al 2020. Thus, GEMM is used in this study and is expressed as where z = max (0, c i − c 0 ), in which c i is the annual average PM 2.5 concentration in grid cell i and c 0 is the minimal concentration of exposure, which in this study was 2.4 µg m −3 , below which no association between PM 2.5 and mortality was assumed. θ j,k , α j,k , µ j,k and ν j,k are parameters that describe the shape of the relative risk function and are provided by Burnett et al (2018) with inclusion of the Chinese cohort.
The relative risk function of long-term O 3 exposure for COPD developed by GBD 2017 Risk Factor Collaborators (2018) is expressed as where in which c i is the summer (1 April to 30 September) average maximum 8 h moving-average O 3 concentration in grid cell i and c 0 is the minimal concentration of exposure, which in this study was 29.1-35.7 ppb with a uniform distribution, below which no association between O 3 and mortality was assumed. rr was set to be 1.06 (CI 95: 1.02-1.10) (Wang et al 2020a).

Deaths attributable to short-term exposure to PM 2.5 and O 3
The daily number of deaths attributable to short-term exposure to PM 2.5 and O 3 for all-cause death was estimated with equation (4), where M is the human mortality within a day, P i, j is the population of grid cell i in age group j, y i, j, k is the baseline mortality rate of grid cell i in age group j for cause of death k, N is the number of days of a year, and RR i,d is the relative risk function of grid cell i and the day d of a year, which is expressed as Here, for PM 2.5 , c i is the daily average PM 2.5 concentration in grid cell i, β is set to be 0.0004 (CI 95: 0.00022-0.00059), and c 0 is set to be 35 µg m −3 (Lu et al 2015); for O 3 , c i is the daily maximum 8 h moving-average O 3 concentration in grid cell i, β is set to be 0.0004 (CI 95: 0.00028-0.00052), and c 0 is set to be 100 µg m −3 (Ye et al 2020).
The 95% confidence intervals of the number of mortalities are estimated using the maximum likelihood method with a Gaussian distribution from multiple simulations with varying parameters in the relative risk functions, and thus the 95% confidence intervals only reflect the uncertainty of the parameters.
In this study, PM 2.5 pollution episode represents the polluted day on which the daily average PM 2.5 concentration was equal to or greater than 75 µg m −3 . Similarly, O 3 pollution episode represents the polluted day on which the daily maximum 8 h moving-average O 3 concentration exceeded 160 µg m −3 .

Air pollution exposure
The PM 2.5 and O 3 concentration data for the period 2016-2020 were obtained from a Chinese reanalysis dataset of atmospheric composition with a grid resolution of 45 km × 45 km and an hourly temporal resolution (Kou et al 2021). This dataset was produced by assimilating surface chemical observations of PM 10 , PM 2.5 , SO 2 , NO 2 and O 3 under the framework of simultaneously optimizing chemical concentrations and source emissions. Moreover, meteorological fields were also assimilated simultaneously. Compared to previous datasets produced in China, the advantage of this one is that the uncertainties of the chemical initial conditions, source emissions and meteorological simulations are all reduced through the use of chemical and meteorological observations. Details of the joint data assimilation system can be found in Peng et al (2017), Peng et al (2018), (2020) and Kou et al (2021).
To facilitate the estimation of human mortality, the concentrations of PM 2.5 and O 3 at the lowest grid layer (nearest to the ground) in the reanalysis dataset of atmospheric composition were linearly interpolated to the grid of the population distribution data.

Population
The population distribution with a horizontal resolution of 0.1 • × 0.1 • was adapted from the LandScan data (LandScan 2019, https://landscan.ornl.gov/) (Rose et al 2020) for years 2016-2019. Although the population distribution for 2020 was not available yet, it is reasonable to assume that the difference in the population distribution between 2019 and 2020 should not be significantly large. Thus, the health impacts of air pollution in 2020 was estimated in this study by using the population distribution for 2019 still. Since the baseline mortality rates were only available at the scale of the whole China, the baseline mortality rates at each grid of the population distribution data are calculated according to the following equation, which is based on the method by Apte et al (2015),

Baseline mortality
where I i is the baseline mortality rate of grid cell i for a certain cause of death, I is the national baseline mortality rate for the cause of death, RR i is the relative risk function of grid cell i and is calculated by equations (2) or (3), depending on the cause of death, and the national relative risk function PWRR is calculated as follows: where P i is the population of grid cell i and P is the national population. In addition, as the data for 2020 were not available, the baseline mortality rates for 2019 were used to estimate the human mortality for 2020. In this study, like in Wang et al (2020a) and Liu et al (2021), the baseline mortality rate of NCD+LRI was used for calculation of deaths attributable to long-term PM 2.5 exposure and the baseline mortality rate of COPD was used for deaths attributable to long-term O 3 exposure, while the all-cause baseline mortality rate was used for deaths attributable to short-term PM 2.5 and O 3 exposure. In addition, the national average baseline mortality rate for cause of death k within a year was calculated as ∑ where P i, j is the population of grid cell i in age group j, and y i, j, k the baseline mortality rate of grid cell i in age group j for cause of death k.

Ambient air pollutant concentrations in China from reanalysis dataset
Since 2013, the PM 2.5 pollution in China has substantially abated year by year, which is attributed to reduced emissions because the Clean Air Action Plan began to be implemented in 2013 (e.g. Zhang et al 2019).
According to the evaluation of the reanalysis PM 2.5 concentrations in Kou et al (2021), the study reproduced the concentration changes well (figure S1), as the reanalysis data were generally in good agreement with the observations and thus a reasonable representation of the atmosphere. The annual average PM 2.5 concentration constantly decreased during the study period, with the national average concentration being 23. 58, 22.74, 20.94, 20.55 and 17.27 µg m −3 in 2016, 2017, 2018, 2019 and 2020, respectively. The major high-concentration centers were in central, eastern and southwestern China, including the North China Plain, Yangtze River Delta (YRD) and Sichuan Basin regions, which are home to major urban agglomerations and industrial areas, while significant reductions in the PM 2.5 concentration also occurred in these regions. Similarly, the occurrence of pollution episodes became less frequent during these years, as shown in figure  S2. The number of days of PM 2.5 pollution episodes decreased during the study period, with the national average being 24.54 (i.e. 6.7% in the year), 21.75 (5.96%), 18.66 (5.11%), 20.11 (5.51%) and 15.02 (4.10%) in 2016, 2017, 2018, 2019 and 2020, respectively, indicating that the air quality in terms of PM 2.5 pollution has overall improved in recent years. In particular, significant improvement has happened in the major high-concentration centers in central, eastern and southwestern China, where there are large and dense populations ( figure S3).
In terms of monthly ambient PM 2.5 concentration, the peak occurred in winter while the trough occurred in summer in both 2019 and 2020 (figure S4). Compared with the same period in 2019, the monthly average concentration decreased significantly in 2020, and the decrease was overall uniform and mostly between 20 and 30%. The sharpest decrease occurred in February of 2020 when PM 2.5 concentration dropped by 7.83 µg m −3 . As shown in figure S5, although the national daily average PM 2.5 concentration was mostly less than 35 µg m −3 , the PM 2.5 concentration in urban agglomerations with relatively concentrated population was still high, especially in the urban agglomerations in the YRD and the Beijing-Tianjin-Hebei (BTH) regions. In 2020, the daily average PM 2.5 concentration in the whole China and several urban agglomerations decreased to varying degrees in February and March compared with the same months in 2019.
However, the O 3 concentrations have not changed with the same trend as the PM 2.5 concentrations have, which was indicated in our study as well as in other researches (e.g. Chen et al 2021, Liu et al 2021). Figure

Mortality burden of long-term exposure to air pollution
In China in recent years, the baseline mortality rates for most diseases have continued to decrease for a single age group because of medical progress. For example, according to the Global Burden of Disease 2019 study (GBD 2019 Risk Factors Collaborators 2020), the death rates of COPD, LRI, NCDs and stroke have decreased during the past decades in China in nearly all age groups. However, the national baseline mortalities (e.g. the NCD+LRI baseline mortality in figure S10) have shown an increasing trend (e.g. Sun et al 2018, Zhou et al 2019. This is most likely because the population has been aging in China in recent years. To discern the effect of the air pollutant concentration on the excess mortality, the net increment of mortality in each year after 2016 was calculated, which was the mortality estimated with the concentration field in the year minus the mortality estimated with the concentration field in 2016, as shown in table 1. The changes of the national mortality attributable to outdoor PM 2.5 exposure were in consistence with the PM 2.5 concentration changes. Figure 1 shows the number of mortalities attributable to long-term  exposure to outdoor PM 2.5 . In general, the national number of deaths attributable to long-term PM 2.5 exposure decreased year by year from 2.18 million (95% confidence interval (1.83, 2.51), the same hereinafter) in 2016 to 1.99 million (1.66, 2.30) in 2020. In addition, the net increment of mortality attributable to the change in PM 2.5 concentration kept being negative while its absolute value increased steadily, indicating that the pollution reduction and emission control strategies implemented in recent years have delivered noticeable achievements in protecting public health in China.
Our study also reveals that in terms of spatial distribution, the mortalities attributable to PM 2.5 exposure in different regions of China have not always displayed a decreasing trend as the national mortality has. Figure 2 shows the trend of the number of deaths attributable to long-term PM 2.5 exposure during 2016-2020. The nationwide mortality decreased during the study period. However, during 2016-2019 the mortality increased in urban areas even though the annual average concentration of PM 2.5 decreased in those same areas according to figure S1, and in the areas the mortality decreased significantly in 2020 due to PM 2.5 concentration decreasing significantly. The increase was mainly caused by the increase of the baseline mortality in those areas (figure S10), even though the PM 2.5 concentration had been decreasing. This is also supported by the findings of Hao et al (2021), which indicated reduction of emissions has not led to decrease of mortality attributable to long-term exposure in Chinese cities.
Contrary to PM 2.5 , the deaths attributable to long-term exposure to O 3 pollution increased from 132.79 thousand (128.58, 137.00) in 2016 to 197.00 thousand (190.98, 203.03) in 2020. Such increases in recent years have been steady and continuous, indicating the O 3 pollution and related human mortality have kept deteriorating in China, despite great efforts in mitigating air pollutant emissions since the implementation of the Clean Air Action Plan began in 2013. It reflects the complicated physicochemical mechanism of near-surface O 3 formation , and the mitigation of O 3 -related human health costs keeps a challenging task ahead. For O 3 exposure, in terms of spatial distribution, the mortalities in different regions of China have shown the same trend as the national mortality has, indicating the O 3 pollution and induced mortality have been deteriorating nationwide. Figure 3 shows the trend of the number of deaths attributable to long-term O 3 exposure during 2016-2020. The nationwide mortality kept increasing and the trend was consistent with the increase of the seasonal average maximum 8 h O 3 concentration in figure S6, except in some regions with sparse populations. The results reveal that the mortality increased with the seasonal average O 3 concentration and number of O 3 polluted days.

Impact of COVID on mortality burden of air pollution
In 2020, anthropogenic emissions were greatly reduced in China because of the unprecedented large-scale lockdown imposed in response to the COVID-19 pandemic . As shown in table 1, the net decrement of the mortality related to long-term PM 2.5 exposure increased from 212.16 thousand in 2019 to 345.33 thousand in 2020 and the increase was the largest year-by-year change, suggesting a significant impact of the pandemic. Similarly, the national number of deaths attributable to short-term PM 2.5 exposure decreased from 46.86 thousand (25.91, 68.74) in 2019 to 36.56 thousand (20.20, 53.66) in 2020. Compared with 2019, the mortality attribute to short-term PM 2.5 exposure decreased significantly during late January to March of 2020 (figure 4), with the sharpest decrease of 3.91 thousand occurring in February.
Contrary to PM 2.5 , the national number of deaths attributable to short-term O 3 exposure increased from 128.62 thousand (90.52, 166.30) in 2019 to 135.17 thousand (95.34, 174.41) in 2020, indicating the O 3 pollution and induced human mortality have kept deteriorating and the trend was not even interrupted by the pandemic in 2020. The number of deaths in April and May of 2020 exploded, which was much more than the number of the deaths attributable to short-term PM 2.5 exposure, with the sharpest year-on-year increase of 162% occurring in April (figure 5).
Why has O 3 been increasing while PM 2.5 has been decreasing in recent years in China? According to some researches (Lu et al 2018, Li et al 2019b, Wang et al 2020b, besides hotter meteorological conditions favorable to the formation of O 3 in recent years, the increase of O 3 was related to the decrease of PM 2.5 , because PM 2.5 scavenged hydroperoxy (HO 2 ) and nitrogen oxides (NO x ) radicals that would produce O 3 by photochemical reactions, although the scavenging could also make O 3 less sensitive to NO x but more sensitive to volatile organic compounds (VOCs) that had not decreased in China. In other words, while   PM 2.5 concentration is high, VOCs can produce O 3 ; while PM 2.5 concentration is low, both NO x and VOCs can produce O 3 , leading to higher O 3 concentration. In addition, the decrease of PM 2.5 increased the atmospheric transparency and thus strengthened the solar radiation reaching the ground, enhancing photochemical reactions that induced higher O 3 concentration too.
To better analyze the impact of the lockdown, the daily data for the province of Hubei was plotted. As shown in figure 6, during the tightest coronavirus lockdown period (from Jan 24 to Mar 24), during which transportation had come to a standstill and industrial production except in some important sectors (such as power plants) had almost stopped as most of people stayed home, the reduction of daily average PM 2.5 concentration led to a decrease of 437 deaths, but the increase of O 3 concentration reduced this health benefit. The effect of PM 2.5 concentration on the mortality attributable to short-term exposure was greater than that of O 3 . However, in April, the surge in daily average O 3 concentration led to an increase of 999 deaths, which was more than the decrease attributable to the reduction of PM 2.5 concentration during the same period.

Limitations and uncertainties
This study has the following major limitations: (1) firstly, the parameters in the relative risk functions and the functions themselves have uncertainties; (2) the mortality for 2020 was not estimated with the population data and baseline mortality rates for 2020, which were not available for this study; (3) the national baseline mortality rates were processed for each grid of the population distribution data to estimate the mortalities for each grid, since the baseline mortality rates were only available at the scale of the whole China; (4) both the national baseline mortality rates and the population distribution data employed to estimate the deaths attributable to short-term exposure are made public yearly, but in fact they change all the time. These limitations inevitably brought about uncertainties for the results of this study, which should be improved in next works.
Our estimates of nationwide mortality attributable to long-term and short-term O 3 exposure are larger than those of Liu et al (2021) during 2016-2018. For example, in 2018 the nationwide mortality attributable to long-term and short-term O 3 exposure in our study is approximately 26.7 thousand and 62.0 thousand larger than their results, respectively. However, the year-by-year trend of increasing O 3 concentrations and induced deaths is the same as theirs. The reason for the larger nationwide mortality attributable to O 3 exposure in the present study is that the reanalysis dataset of atmospheric composition used here has larger O 3 concentrations, which needs to be further investigated in future work.

Conclusions
The number of deaths induced by long-term and short-term PM 2.5 exposure decreased year by year, and the impacts of pollution control strategies and pandemic on the deaths attributable to long-term exposure are more significant than on the deaths attributable to short-term exposure. The positive impact of China's Clean Air Action Plan implemented in recent years on human health is evident, and reducing PM 2.5 pollution is more beneficial to mitigating loss of human life in the long term than in the short term. Also, it underscores that more works are needed to mitigate the impacts of long-term PM 2.5 exposure on human mortality by further reducing the annual average concentration of PM 2.5 , especially as the population has kept aging in China in recent years, since the baseline mortality rates are always higher in aged population than in young population. Although the monthly or yearly average concentration of PM 2.5 is lower than 35 µg m −3 , considering the excess daily average concentration higher than 35 µg m −3 that occurred on 119 of 1827 d during 2016-2020 (i.e. the occurring probability was 6.5%), short-term exposure to PM 2.5 pollution still causes deaths, which calls attention to the short-term health impact of sudden PM 2.5 pollution events.
The case of O 3 is on the contrary. The deaths induced by long-term and short-term O 3 exposure have increased year by year. Moreover, the year-on-year variations induced by long-term exposure were overall larger than those induced by short-term exposure, thus human health and welfare would have been more benefited in the long term than in the short term had O 3 pollution been reduced during the study period. Reducing O 3 pollution needs to be focused on in the coming years, since it poses an urgent challenge for policy-makers and researchers.
The lockdown in 2020 can be taken as an example for simulation of the effect of substantial emission reduction. However, compared to the preceding years, the excess mortality associated with outdoor O 3 exposure was still increased although the excess mortality attributable to PM 2.5 exposure was further reduced. The result reflects the difference of the formation mechanism between PM 2.5 and O 3 , since the decrease of PM 2.5 could contribute to the increase of O 3 , because less PM 2.5 attenuated the scavenging of HO 2 and NO x that could produce O 3 by photochemical reactions that could also be enhanced by stronger solar radiation through more transparent atmosphere induced by less PM 2.5 . While the simultaneous reduction of NO x and VOCs may be an effective approach for decreasing PM 2.5 and O 3 both, more researches on the simultaneous reduction of PM 2.5 and O 3 are required in China.

Data availability statement
The data cannot be made publicly available upon publication because no suitable repository exists for hosting data in this field of study. The data that support the findings of this study are available upon reasonable request from the authors.