Key Factors in Epidemiological Exposure and Insights for Environmental Management: Evidence from Meta-analysis

Abstract

In recent years, the precision of exposure assessment methods has been rapidly improved and more widely adopted in epidemiological studies. However, such methodological advancement has introduced additional heterogeneity among studies. The precision of exposure assessment has become a potential confounding factors in meta-analyses, whose impacts on effect calculation remain unclear. To explore, we conducted a meta-analysis to integrate the long- and short-term exposure effects of PM_2.5, NO₂, and O₃ on all-cause, cardiovascular, and respiratory mortality in the Chinese population. Literature was identified through Web of Science, PubMed, Scopus, and China National Knowledge Infrastructure before August 28, 2023. Sub-group analyses were performed to quantify the impact of exposure assessment precisions and pollution levels on the estimated risk. Studies achieving merely city-level resolution and population exposure are classified as using traditional assessment methods, while those achieving sub-kilometer simulations and individual exposure are considered finer assessment methods. Using finer assessment methods, the RR (under 10 μg/m³ increment, with 95% confidence intervals) for long-term NO₂ exposure to all-cause mortality was 1.13 (1.05-1.23), significantly higher (p-value=0.01) than the traditional assessment result of 1.02 (1.00-1.03). Similar trends were observed for long-term PM_2.5 and short-term NO₂ exposure. A decrease in short-term PM_2.5 levels led to an increase in the RR for all-cause and cardiovascular mortality, from 1.0035 (1.0016-1.0053) and 1.0051 (1.0021-1.0081) to 1.0055 (1.0035-1.0075) and 1.0086 (1.0061-1.0111), with weak between-group significance (p-value=0.13 and 0.09), respectively. Based on the quantitative analysis and literature information, we summarized four key factors influencing exposure assessment precision under a conceptualized framework: pollution simulation resolution, subject granularity, micro-environment classification, and pollution levels. Our meta-analysis highlighted the urgency to improve pollution simulation resolution, and we provide insights for researchers, policy-makers and the public. By integrating the most up-to-date epidemiological research, our study has the potential to provide systematic evidence and motivation for environmental management.

1. Introduction

Air pollution has become the fourth leading cause of premature death worldwide (WHO, 2022). In 2019, the total air pollution resulted in approximately 6.67 million premature mortality globally, with over 1.85 million in China (Stapleton et al., 2008). The six principal air pollutants have received widespread attention in China, among which fine particulate matter (PM_2.5), ozone (O₃) and nitrogen dioxide (NO₂) have been identified as the most critical air pollutants. PM_2.5 has significant and severe health impacts (Chen and Hoek, 2020; Li et al., 2019b; Orellano et al., 2020). O₃ emerging as the primary pollutant in summer (Xiao et al., 2021; Yang et al., 2018b). The NO₂ pollution is not as severe as PM_2.5 and O₃, but was identified as the core air pollution precursor of O₃ and nitrate component of PM_2.5 (He et al., 2020). In comparison, the health effects of the other three air pollutants, coarse particulate matter (PM₁₀), sulfur dioxide (SO₂) and carbon monoxide (CO), are less important. Exposure effects of CO and SO₂ remained unclear (WHO, 2021a), and the most considerable effect particle size of PM₁₀ is PM_2.5 (Lin et al., 2016b), so we do not prioritize their attention. Over the past decade, China implemented robust policies to reduce air pollution and significantly improved air quality (Luo et al., 2024; Luo et al., 2022; Shi et al., 2022). The impact of drastic changes in air pollution levels on exposure effects have seen evidence in other world-wide nations (Burnett et al., 2014), however, have not yet undergone systematic analysis on the Chinese population.

Recently, WHO updated its AQG in 2021 based on up-to-date health evidence(WHO, 2021b)(WHO, 2021b). In 2022, key cities in China achieved a 62.8% compliance rate for environmental air quality compared to the Chinese air quality standards. The annual average concentrations of PM_2.5 and O₃ at 29 μg/m³ and 145 μg/m³, respectively. Approximately 1/4 of the cities exceeded the standard level throughout the year (Yang et al., 2018b). Compared with WHO AQG 2021, China has not yet sufficiently mitigated the health risks posed by air pollutants exposure, necessitating further tightening of air quality standards. The United States Environmental Protection Agency (U.S. EPA) has established a comprehensive revision process for air quality standards, requiring consideration of updates based on air quality criteria (Wang et al., 2022). However, the air quality criteria in China are still in their early stages. WHO AQG 2021 may be a potential reference, but its representativeness is questionable due to the inclusion of limited evidence from the Chinese population (6 from mainland China, 3 from Taiwan, 1 from Hong Kong) (WHO, 2021b)(WHO, 2021b). Other developed integrated models may face similar challenges (Zou et al., 2019). Most existing meta-analyses focus on one pollutant, health outcomes, or exposure period. Few studies covered major pollutants and health outcomes; however, they often aggregated worldwide population effect relationships (Orellano et al., 2020). There has yet to be a comprehensive integration of the exposure effects of major air pollutants in China to support the formation of health-based air quality criteria. Systematic evidence synthesis could lead to new insights into population exposure risks, further motivating the policy revisions and air pollution reduction.

Research on the short-term effects of air pollution exposure in China can be traced back to the beginning of this century (Kan and Chen, 2003). Many solely relied on exposure level assessments derived from monitoring station observations (Dong et al., 2012; Zhang et al., 2011). However, variations in pollution levels exist among city regions, and solely adopting city-level averages would indeed introduce biases for analysis (Sun et al., 2023). Since the Chinese government has taken action to reduce air pollution, background levels have gradually decreased. However, near-source pollution effects (where pollution levels increase near the emission source) have become more prominent at the street scale, typically within or around tens of meters (Lv et al., 2022; Wang et al., 2023b). Such localized differences in pollution levels are challenging for traditional medium-scale resolution models or observational data, which typically cover several to tens of kilometers, to accurately represent and account for. In recent years, various small-scale models have been developed to estimate pollution distribution at the street scale (Wang et al., 2023b), and some epidemiological studies have already introduced these finer models into exposure assessments (Wong et al., 2015). Exposure level precision significantly impacts the resultant effects, highlighting the need for clear articulation of these distinctions, which is a limitation often seen in current reviews and meta-analyses (Luo et al., 2023; Yang et al., 2019). These finer methods can help identify exposure hotspots and focus on high-exposure activities (Al-sareji et al., 2022), such as pinpointing highly polluted sections of street environments (Munir et al., 2022). It would allow policy-makers to concentrate their focus from overall exposure to a certain chain, enabling more targeted regulation (Roy et al., 2024). Furthermore, the government will be able to issue specific travel recommendations to the public, enhancing the temporal and spatial flexibility of activity patterns to reduce individual and population exposure risks (Song et al., 2021).

In this study, we investigated whether variations in exposure assessment precision and pollution levels would lead to differences in the calculation of exposure effects. We conducted a meta-analysis to examine the associations between exposure to PM_2.5, NO₂, and O₃ and all-cause, cardiovascular, and respiratory mortality in the Chinese population. This analysis included sub-group evaluations based on exposure assessment precision and pollution levels. From the quantitative analysis and the included literature, we qualitatively identified key factors in exposure assessment relevant to effect calculation. Ultimately, we highlighted the shortcomings of current epidemiological exposure assessments and provided insights for improving environmental management.

2. Methods

2.1. Literature identification strategies

This report was conducted using the preferred reporting items for systematic reviews and meta-analyses (PRISMA) guidelines. The databases used for identification encompassed Web of Science, PubMed, Scopus, and China National Knowledge Infrastructure (CNKI). The conclusive date for literature identification was August 28, 2023. Relevant literature published before this date was considered the potential candidates, with no specific starting point set. The targeted primary air pollutants were PM_2.5, NO₂, and O₃, and the health outcomes were all-cause, cardiovascular, and respiratory mortality. The study exclusively focused on health evidence from the Chinese population. All literature included in our analysis was conducted on human subjects. We developed a literature identification strategy for each database, combining free text and medical subject headings (MeSH) terms, e.g., target pollutants, health outcomes, study types, and study region, referring to the published meta-analyses with similar topics (Orellano et al., 2020; Yang et al., 2019). The identification logic remained consistent across different databases, listed in Appendix 1. A secondary refinement was conducted for the cited literature in the identification results, serving as an additional relevant source.

Our study was not registered on Cochrane nor PROSPERO. To ensure novelty, we also an identification strategy for existing systematic reviews and meta-analyses, primarily focusing on literature conducted in China. A summary of our findings can be found in Appendix 1. The results have demonstrated that the motivation behind this study is appropriate and necessary. Compared to previous studies, we have made further progress in assessing the impact of 1) employing finer pollution simulation methods and 2) utilizing individual exposure levels instead of group levels on the calculation of the exposure effects.

2.2. Eligibility criteria

Our study focuses on both long- and short-term exposure effects of PM_2.5, NO₂, and O₃. Only cohort studies were included for long-term effects, while time-series and case-crossover studies were included for short-term effects. Literature involving the general public (ordinary people) was included to ensure the targeted population’s representativeness. In contrast, literature on occupational air pollution exposure (e.g. firefighters), patients, or populations with particular sensitivity (e.g. children and pregnant women) or high-risk conditions (e.g. smokers) was excluded, because these populations have unique exposure characteristics, including high-intensity and long-duration exposure or sensitive physiological conditions, which fail to represent the general population and thus do not align with the scopes of this study (Chen and Hoek, 2020; Luo et al., 2023; Orellano et al., 2020). Literature that passed the title/abstract screening and met our target were all included in the systematic review, irrespective of whether they were assessed to have a high risk of bias (Lu et al., 2015a). The three major categories of health outcomes of interest were classified using the 10^th edition of the International Classification of Diseases (ICD-10), including all-cause non-accidental mortality (ICD-10: A00-R99), cardiovascular disease mortality (ICD-10: I01-I99), and respiratory mortality (ICD-10: J00-J99). Literature that did not provide numeric exposure effects or 95% confidence intervals was excluded. The overall eligibility criteria were summarized in terms of population, intervention/exposure, comparison, outcomes, and type of study (PICOT), shown in Table 1.

Table 1. Table of population, intervention/exposure, comparison, outcomes, and type of study (PICOT).

Terms	Targets
Population	Ordinary Chinese adult population (e.g., not patients, children, or other specific or sensitive populations)
Intervention/Exposure	Ambient (not indoor) PM2.5, NO2, and O3 pollution; exposure derived from all kinds of air pollution observations or simulation models and exposure assessment methods
Comparison	Population exposed under 10.0 μg/m3 lower level, or other increments of the mass concentration of air pollutants, or a certain percentage of the data distribution (e.g., IQR)
Outcomes	All-cause, cardiovascular, and respiratory mortality
Type of Study	Cohort studies for long-term exposure; time-series studies and case-crossover studies for short-term exposure

For long-term effects, if there were multiple literature analyzing the same cohort, the one with a larger sample size was included to avoid duplicative sampling and ensure the analysis was based on a more reliable statistical analyze (Ghosh et al., 2021). For short-term effects, considering representativeness at national, provincial, municipal, and county scales, priority was given to studies with broader spatiotemporal coverage (Luo et al., 2023). In all cases, multi-city studies were always preferred over single-city studies; for two or more literature on the same study period, study region, and population, with the same pollutants, health outcomes, and data sources, the most recently published one was included (Lu et al., 2015b). On this basis, we further focused on the diversity of exposure assessment methods. Even when considering the same spatiotemporal scope and target population, we included literature that employed different exposure assessment methods to evaluate the potential influence on the estimated exposure effects attributed to its precision. After selecting the more informative literature, the rest were all excluded from the meta-analysis. Additionally, the literature eligible for inclusion in the meta-analysis required a systematic assessment of bias risk. Literature with a high risk of bias were excluded, with detailed descriptions provided in the following part of the Methods section. Following title/abstract screening, bias risk assessment, and consideration of spatiotemporal representativeness, and those passed were finally included in the meta-analysis.

2.3. Data extraction

The data extraction process for the chosen literature involved independent extraction by two reviewers, then cross-checking and consolidating the results. Microsoft Excel^® was utilized for efficient data extraction. Essential details were meticulously recorded, including the geographical scope, time duration, data source, subject population, sampling size, exposure assessment method, exposure level, health outcome, and more. Association measures such as hazard ratio (HR), relative risk (RR), odds ratio (OR), and excess rate (ER) of the mortality were extracted. Any literature that did not provide detailed effects or had missing effect data was excluded.

For long-term effects, there was often the case that several adjusted models with different confounders were provided in one cohort study. To eliminate the differences in confounding factor adjustment and choose the most suitable model for integration, we compiled the maximum set of confounding factors from the literature included in our research. Then, we made a directed acyclic graph for the associations between air pollution exposures and mortality (shown in Appendix 2). A minimal collection of confounders was developed based on this directed acyclic graph. The selected adjustment model must have the maximum intersection with this minimal set and should preferably not include unnecessary adjustment factors to eliminate the interference of these confounding factors on the study results. For short-term effects, if only a single lag effect for a specific time was provided, that lag effect was selected (Atkinson et al., 2012; Orellano et al., 2020). If multiple lag effects were presented, the selection followed the framework by previous studies (Yang et al., 2019): i) priority was given to lag effects of particular concern to the authors or the most statistically significant effects; ii) non-cumulative lag effects were prioritized, with cumulative lag effects chosen in their absence. Such a framework was devised to choose lag estimates for inclusion in the review without introducing bias (Atkinson et al., 2014).

All included literature should provide at least one effect estimate for each exposure-outcome pair. There was literature assessed multiple air pollutants within the scope of our research and provide evidence for two or more exposure-outcome pairs. These effects from different exposure-outcome pairs were included in corresponding meta-analyses separately (Orellano et al., 2020; Yang et al., 2019).

2.4. Risk of bias (ROB) assessment

A full-text review was conducted for all literature identified through title and abstract screening, and relevant information was extracted while simultaneously performing a ROB analysis to assess potential biases. The ROBINS-E risk assessment framework is a state-of-the-art tool for evaluating bias risk in non-randomized exposure studies, and the detailed criteria can be found elsewhere (Higgins et al., 2024). It encompasses seven bias domains: bias due to confounding, bias arising from measurement of the exposure, bias in selection of participants into the study, bias due to post-exposure interventions, bias due to missing data, bias in measurement of the outcome, and bias in selection of the reported result. Each domain is addressed through a series of signaling questions to gather crucial information about the study and the analysis under assessment. Ultimately, the ROBINS-E framework yields three judgments: the existence of bias risk, the potential direction of bias, and whether the bias risk is sufficiently high to threaten the accuracy of the results. The ROBINS-E method was applied to all the literature included in the systematic review. All literature with a high risk of bias was excluded from the meta-analysis.

2.5. Data analysis

The exposure-effect relationship may be characterized using HR, RR, and OR in the cohorts, time-series, and case-crossover studies included in this research. Although there exist differences in definitions, these three indicators may not differ significantly in numerical value within the scope of this study. The existing meta-analyses did not further distinguish between them (Southerland et al., 2022). Therefore, we uniformly used the ER to represent effect observations. The HR, RR, and OR were generally taken as RR, and the conversion relationship between ER and RR was as follows (Southerland et al., 2022):

$$\text{ER}=\frac{\text{RR}-1}{\text{RR}}$$ (Eq. 1)

The standardized increment in pollutant levels associated with the effect is commonly set at a 10 μg/m³ change in pollutant levels. RR values based on this specific increment were considered standardized. However, some literature utilized alternative increments in pollutant levels, such as the interquartile range (IQR). For them, the original RR reported in the study is converted to a standardized RR using the following formula (Li et al., 2017; Orellano et al., 2020):

$${\text{RR}}_{\text{std }}=\text{exp}\left(\frac{10\times \ln \left({\text{RR}}_{\text{org }}\right)}{{\text{Inc}}_{\text{org }}}\right)$$ (Eq. 2)

where: RR_std refers to the standardized RR (with a pollutant level increment of 10 μg/m³); RR_org refers to the RR reported in the literature with a pollutant level increment different from 10 μg/m³; Inc_org refers to the non-standardized pollutant level increment provided in the literature. Specifically, for O₃, the concentrations and increments provided in non-mass units (e.g., ppb) in some literature will first be converted to μg/m³ before proceeding with further analysis. The p-value was checked for the integrated effects to ensure certainty of significance. A p-value smaller than 0.05 means that the adverse impact of exposure on specific outcomes is significant.

Sub-group analyses were conducted based on gender, age (take 65 as the dividing point), exposure assessment method precisions, and pollution levels:

• (1)Age and gender: For literature that provides exposure effects for age and gender sub-groups, we extracted the corresponding sub-group effects for further analysis and comparison. Physiological gender sub-groups include male and female. Age sub-groups were divided into young (<65) and elderly (>=65). It should be noted that some literature had different age divisions, such as using 75 as a dividing point or having more sub-group divisions (e.g., <45, 45-64, >65), which will be manually matched accordingly.
• (2)Exposure assessment method precisions: Due to the inclusion of information on population exposure assessment methods (pollutant simulation resolution, subject granularity, etc.), we also conducted sub-group analyses based on the precisions of exposure assessment methods. The population exposure assessments were commonly confined to the urban level. Therefore, literature employing monitoring station observations or pollutant datasets at a comparable scale (e.g., 10 km or coarser resolution grid pollution data) were categorized as lower resolution. Meanwhile, literature utilizing more sophisticated pollution datasets, simulation models were classified as higher resolution (such as adopting land use regression, machine learning algorithms, multi-source fusing data, etc.). As for subject granularity, studies that assess individual exposure levels, such as those using residential addresses for exposure allocation, are considered to have a higher subject granularity in their exposure assessment methodology. Conversely, studies that rely on city-wide mortality data and average pollution levels are deemed to have lower subject granularity. It is important to note that due to the limited number of studies available, we were only able to conduct a subgroup analysis based on subject granularity for the long-term effects of PM_2.5 on all-cause mortality. The number of studies for the long-term effects of NO₂ and O₃ was insufficient. Additionally, almost all studies on short-term effects did not reached individual level, which further restricted the possibility of subgroup analysis.
• (3)Pollution levels: Implementing the Action Plan on Air Pollution Prevention and Control after 2013 has significantly contributed to improving air pollution levels. Therefore, taking 2013 as the dividing point, a sub-group analysis was conducted on the included short-term effects to explore the differences in the short-term exposure effects under different pollutant levels. Due to a lack of epidemiological studies, it cannot perform the same analyses for long-term effects. This sub-group analysis did not include the literature that spanned pre- and post-2013 periods. The significance of differences between sub-groups was checked using the interaction analysis and given as the p-value.

After this phase, we employed an integrated conceptualization process based on meta-analysis and subgroup analysis, combining qualitative analysis, prior knowledge, and systematic review together, to identify the key factors influencing the accuracy of exposure assessment methods. The specific steps included: i) Extract information closely related to the research objectives, primarily focusing on the indicators associated with the accuracy of exposure assessment methods and pollution levels. ii) Evaluate the stratification capability among studies of certain exposure-outcome pairs, and determine the comprehensiveness of the included studies. iii) Conduct subgroup analyses for factors that could be stratified. For factors with significant variability (e.g., single layer or too many layers), evaluating their potential impact through literature review and expert judgment, instead of using an over-stratified subgroup analysis. iv) Systematically organize and structure the identified factors based on prior knowledge and reference literature, a process somewhat akin to constructing a directed acyclic graph based on prior knowledge (though not entirely the same). It is important to note that relying solely on qualitative methods may not fully address the heterogeneity among studies and the complexity of detail distribution, potentially leading to bias or the omission of critical factors in statistical analysis. Therefore, we adopted a hybrid approach that combines quantitative analysis with expert judgment based on prior knowledge to enhance the accurate understanding and comprehensive grasp of complex research.

We performed tests to validate heterogeneity among literature to select the most appropriate model. The Cochran’s Q test based on I² and Tau² was taken to examine the heterogeneity of the exposure effects in the included literature and between the sub-groups (Yang et al., 2019). All integrations in this study displayed high heterogeneity (I²>40% and p<0.05). Thus, the random-effects model was employed for integration instead of the fixed-effect model.

It is vital to assess the publication bias to check for potential influence during the manuscript formation and publication. Funnel plots were used to evaluate the publication bias in the included meta-literature. Additional algorithms, such as Begg’s and Egger’s regression, were applied to examine the asymmetry of funnel plots quantitatively (Begg and Mazumdar, 1994; Egger et al., 1997). Begg’s test statistic is based on rank correlation algorithms to explore the distribution differences among literature, with the test statistic following a standard normal distribution. Egger’s test statistic is based on weighted linear regression of effect sizes against standard errors, with the test statistic data following a t-distribution. A p-value>0.05 in those regression models would indicate that statistically significant publication bias might exist in the literature. Since prior research recommended conducting asymmetry tests only when the number of included literature in the meta-analysis is greater than or equal to 10 (Sterne et al., 2011), bias tests for publication were performed only for the short-term effects with a sufficient number of included literature.

The widely adopted meta-regression analysis was employed to investigate the sources of heterogeneity. In this study, essential sources of heterogeneity included study region (nation-wide, regional, county-or city-level), duration of study period, study population (the whole adult population, middle-aged or elderly population, elderly population), sampling size (mortality or total population size), exposure assessment method, pollution simulation resolution, exposure level, and study design. Each variable was individually subjected to meta-regression. We assessed the degree of influence of various factors on heterogeneity based on core indicators of meta-regression: the R² statistic (representing the proportion of between-study variance explained by covariates) and ΔI² (a measure of the remaining percentage of variance attributable to between-study heterogeneity after adjustment for predictor variables).

All data analysis and graphical visualization were performed using the statistical software R version 4.3.0 (https://www.r-project.org/) with the “meta” and “metafor” packages. The meta-analysis was performed using the ‘metagen’ function. The symmetry assessment of the funnel plot was performed using the ‘metabias’ function with Begg’s and Egger’s options, respectively. The meta-regression analysis was performed using the ‘metareg’ function.

Visualization was made with the Origin lab and Microsoft PowerPoint^®. The directed acyclic graph for the associations between air pollution exposures and mortality was created with the help of DAGitty.net (www.dagitty.net, accessed on March 27, 2024).

2.6. Sensitivity analysis

In terms of sensitivity analysis, this study employed the leave-one-out method to assess the robustness of the meta-integrated effects. The leave-one-out method involves cyclically omitting one study to conduct a new meta-analysis and comparing the results obtained with including all literature. A p-value>0.05 indicates that omitting that study significantly affects the integrated effects.

Given the differences in study details, particularly in the selection and combination of confounding factors within the effect models, it is crucial to assess the impact of these variations on the final results through sensitivity analysis. While some meta-analyses rely on effect estimates that authors of epidemiological studies tend to report preferentially (Lu et al., 2015a), our study used a minimal set of confounding factors to ensure consistency. We performed separate screenings and meta-analyses using both approaches and compared the results.

Due to the existence of the literature with the study period overlapping with the COVID-19 pandemic (from 2020), a sub-group analysis was performed additionally in the sensitivity analysis to see the potential influence of the inclusion of COVID-19 pandemic period on the mortality risk among the Chinese population.

The aforementioned meta-analysis was solely based on single-pollutant models from the literature. However, in some studies, after adjusting for multiple pollutants, there have been instances where the RR values shift or the significance of the associations changes. Therefore, we performed a sensitivity analysis to observe the impact of considering the multi-pollutant interactions on the final results. The principle for including multi-pollutant models is as follows: For studies that provided multi-pollutant models, we replaced the relevant RR values originally included in the meta-analysis. During the selection process, the lag model or the confounder-adjusted model was kept consistent with the primary analysis to ensure methodological comparability within the meta-analysis. In cases where multiple multi-pollutant models were available, we prioritized: i) the model adjusting for the highest number of pollutants (e.g., a four-pollutant model is prioritized over a two-pollutant model) and ii) the most significant pollutant-adjusted model.

2.7. Certainty of evidence

To ascertain the certainty of evidence (CoE) for each combination of exposure and health outcomes, we employed the Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) approach (Orellano et al., 2020). The CoE assessment encompasses eight domains: five for downgrading (limitations in studies, indirectness, inconsistency, imprecision, and publication bias) and three for upgrading (large effect size, confounding, and concentration-response gradient). The overall assessment procedure was as follows: 1) for a specific exposure-health outcome pairing, evidence from multiple articles was initially deemed to have a moderate CoE, and subsequent evaluations occur in both 2) downgrading and 3) upgrading domains. Following the initial analysis, assessments were made for the remaining three domains, allowing for the potential upgrading of CoE. The final results were categorized into four potential overall certainties: high (further research was unlikely to alter confidence in the integrated effects), moderate (additional studies might significantly impact confidence in the integrated effects), low (further research was likely to have a significant effect on confidence in the integrated effects), or very low (high level of uncertainty in estimating the integrated effects). Detailed criteria for evidence certainty assessment are listed in Appendix 3.

3. Results

3.1. Meta-analysis for long- and short-term exposure effects

Among the identified 37104 records from the database, there were 2353 records remained to be screened. Totally 231 records were included in the qualitative analysis, and only 95 studies were included in the final meta-analysis: 21 articles for the long-term effects (Cao et al., 2011; Chen et al., 2021b; Giani et al., 2020; Guo et al., 2022; Huang et al., 2023b; Ji et al., 2022; Liang et al., 2022; Liu et al., 2022a; Liu et al., 2022b; Niu et al., 2022; Wang et al., 2023a; Wang et al., 2023c; Wang et al., 2023d; Xia et al., 2023; Yang et al., 2020; Yang et al., 2018a; Yin et al., 2017a; Yuan et al., 2023; Zhang et al., 2022b; Zhang et al., 2023a; Zhang et al., 2023b) and 74 articles for the short-term effects (Cao et al., 2012; Chen et al., 2023a; Chen et al., 2022a; Chen et al., 2019; Chen et al., 2022b; Chen et al., 2021a; Chen et al., 2022c; Chen et al., 2012; Chen et al., 2017; Chen et al., 2018; Chen et al., 2021c; Chen et al., 2022d; Dai et al., 2015; Deng et al., 2021; Dong et al., 2020; Duan et al., 2019; Gao et al., 2023; Gong et al., 2019; Guan et al., 2022; He, 2022; Hu et al., 2021; Ji et al., 2020; Kan et al., 2010; Kan et al., 2007; Li et al., 2015; Li et al., 2013; Li et al., 2018a; Li et al., 2019a; Lin et al., 2022; Lin et al., 2016a; Liu et al., 2018; Liu et al., 2020b; Liu et al., 2023; Liu et al., 2022c; Liu et al., 2021a; Liu et al., 2021b; Liu et al., 2019b; Lu et al., 2015c; Lu et al., 2023; Ma et al., 2023; Ma et al., 2011; Mo et al., 2023; Peng, 2023; Pu et al., 2023; Qian et al., 2015; Qian et al., 2010; Qiu et al., 2018; Qu et al., 2018; Shao et al., 2021; Song et al., 2022; Sun et al., 2022; Tao et al., 2012; Tian et al., 2020; Wong et al., 2001; Wong et al., 2002; Wu et al., 2022; Wu et al., 2019; Wu et al., 2018; Xu et al., 2020a; Xu et al., 2020b; Xu et al., 2022; Yan et al., 2021; Yang et al., 2012; Yang et al., 2013; Yin et al., 2017b; Yu et al., 2019; Zhai et al., 2023; Zhang et al., 2016; Zhang et al., 2017; Zhang, 2019; Zhang, 2023; Zhang et al., 2020; Zhong et al., 2018; Zhou et al., 2022). Reasons of crossing out identified literature during the whole process were listed in Appendix 4. Lists of the included literature were detailed in Appendix 5 and 6. The study selection process is presented in detail in Figure 1. To ensure the representativeness and coverage of the pollution windows, we summarized the pollution levels among literature for each exposure-outcome pairs in Appendix 7. The substantial variability in pollutant ranges confirms the validity of the integrated effects and underscores the rationale and significance of conducting this work.

Figure 1. PRISMA flowchart: systematic literature identification and the in-and-exclusion criteria for epidemiological studies in this study.

For short-term exposure, each 10 μg/m³ increase in PM_2.5 exposure corresponds to RR values (95% CI) of 1.0050 (1.0037-1.0064), 1.0079 (1.0061-1.0097), and 1.0078 (1.0059-1.0097) for all-cause, cardiovascular, and respiratory mortality, respectively. Correspondingly, with the same increment in other pollutants, the RR values of mortality associated with NO₂ exposure are 1.0149 (1.0108-1.0191), 1.0208 (1.0172-1.0244), and 1.0162 (1.0119-1.0205), and for O₃, the RR values are 1.0058 (1.0029-1.0087), 1.0064 (1.0036-1.0092), and 1.0066 (1.0038-1.0095). For long-term exposure, each 10 μg/m³ increase in PM_2.5 exposure corresponds to RR values (95% CI) of 1.12 (1.06-1.19), 1.16 (1.09-1.25), and 1.17 (1.01-1.37) for all-cause, cardiovascular, and respiratory mortality, respectively. For the same increment in other pollutants, the increase in NO₂ exposure is associated with an RR value of 1.11 (1.03-1.20) for all-cause mortality. The increase in O₃ exposure is associated with an RR value of 1.10 (1.01-1.19) and 1.15 (1.03-1.28) for all-cause and cardiovascular mortality, respectively. The forest maps of the meta-analysis result are shown in Appendix 8.

In our results, we found there was no evidence of long-term O₃ exposure effects on respiratory mortality. We attempt to explain it from the following three aspects. First, the impact of O₃ on respiratory diseases remains uncertain. Even at the international level, as reflected in the WHO AQG 2021 (WHO, 2021a), which compiles the most advanced global health evidence, the association between O₃ and respiratory disease mortality is still classified as “low”. The lack of related evidence in Chinese studies may be possibly because no statistically meaningful results have been obtained. Second, the availability of cohort databases suitable for O₃ analysis in China is still limited. According to our research findings, only the CLHLS and CHARLS databases have been utilized in studies on all-cause mortality; and the CHERRY and CCDRFS databases on cardiovascular disease mortality. The limited availability and richness of information provided by these databases may restrict the scope of related research. Third, in existing studies on the effects of O₃ exposure, the research periods span from 2005 to 2018. Despite the use of some advanced pollution simulation methods, various factors may have contributed to significant inaccuracies in the simulation results. These factors include lower early pollution concentrations (Zhang et al., 2022a), a lack of observational data (since China did not incorporate O₃ into ordinary monitoring until 2013), inaccurate characterization of precursors (Chen et al., 2023b), the complexity of formation mechanisms (Xiong et al., 2023), and significant contributions from cross-regional transport (Liu et al., 2020a). These issues may contribute to the lack of significant associations observed in the results.

The meta-analysis differences in exposure effect among age and gender sub-groups are shown in Appendix 9. Briefly, there exist exposure effect differences among age sub-groups between elderly (≥65) and young (<65), but no statistically significant differences were found between males and females. The results of the ROB assessment (including the ROBINS-E and the funnel symmetry test), the meta-regression analysis, the sensitivity analysis (including: i) the leave-one out validation; ii) confounder difference comparison; iii) the impact analysis of COVID-19 pandemic; iv) the impact of taking multi-pollutant models into consideration), and the certainty analysis can be found in Appendix 10-14. All results showed good robustness and reliability to our results.

3.2. The impact of exposure assessment precisions and pollution levels

For various exposure assessment methods, sub-group analyses reveal a potential influence of the exposure assessment method precisions on the quantitative estimation of exposure effects, shown in Figure 3. Only the long-term effects of PM_2.5 on all-cause mortality supported subgroup analysis by subject granularity. Increasing subject granularity to individual exposure level showed a slightly significant impact on effect estimation (with a p-value approaching the 0.1 level). The short-term effects of PM_2.5, the long- and short-term effects of NO₂, and the short-term effects of O₃ supported subgroup analyses based on the resolution of pollution simulations. The refinement of pollution simulation resolution shows a statistically significant difference at the 0.05 level for the long-term effect of NO₂ on all-cause mortality, and also approaching 0.1 level for short-term NO2 exposure. All these mentioned sub-group analyses for NO₂ and PM_2.5 suggest that applying more precise exposure assessment methods tends to yield higher exposure effect values, though with slight significance.

Figure 3.

Exposure effects of PM_2.5, NO₂, and O₃ under different method precisions (including subject granularity and pollution simulation resolution) in population exposure assessment on all-cause, cardiovascular, and respiratory mortality

Since the implementation of China’s Action Plan on Air Pollution Prevention and Control in 2013, there have been significant changes in pollution levels. Implementing the Action Plan on Air Pollution Prevention and Control has led to a substantial decrease in PM_2.5 and NO₂, while an increase in O₃. Among the short-term effects of PM_2.5, there was no significant difference in the effects on respiratory mortality at different pollution levels (p-value=0.63). However, cardiovascular mortality showed a significant difference at the 0.1 level (p-value=0.09) and as similar for all-cause mortality (p-value=0.13). No significant differences were found in short-term exposure effects of NO₂ and O₃ at different pollution levels. Despite the insignificance, the short-term exposure effects on all mortality showed a potential tendency to increase as the pollution levels decreased.

3.3. Key factors in exposure assessment influencing effect calculation

Sub-group meta-analyses have indicated that the exposure assessment precisions can potentially influence effect calculation. However, the sources of limitations in exposure assessment precisions are multifaceted, necessitating identification of the key factors. Here, we identified four major key factors: pollution simulation resolution, subject granularity, micro-environments classification, and pollution levels.

Pollution simulation resolution is a core factor influencing the exposure assessment precisions, and fundamentally determining the subject granularity, the ability to differentiate micro-environments, and the evaluation of pollution levels. For example, assessments using in-situ monitoring data or pollution datasets with kilometer-scale resolution can only evaluate population exposure at the city or town level, unable to differentiate of micro-environments, and with relatively low accuracy in pollution representativeness. Notably, such limitations depend on the distribution nature of the pollutants. The more a pollutant tends to have the near-source characteristics, the greater the impact of pollution simulation accuracy on exposure assessment precisions (Lal et al., 2020). This is supported by the sub-group analysis. NO₂, a highly traffic-related pollutant, shows the most significant response to changes in pollution simulation accuracy for both short- and long-term exposures (Xiang et al., 2022). In contrast, PM_2.5, due to its complex sources and weaker near-source characteristics compared to NO₂, shows a less significant response to changes in pollution simulation accuracy (Karner et al., 2010). Nevertheless, with the improved capability to identify local hotspots, there is a consistent trend where finer precision leads to higher effect.

Subject granularity (population vs. individual) is another crucial factor. Earlier studies, lacking detailed individual-level data, could only use entire populations as the evaluation target. However, due to the physiological and occupational diversity among individuals, exposure characteristics vary significantly among the population (Gerharz et al., 2009; Shimada and Matsuoka, 2011). It is highlighted by our meta-analyses of long-term PM_2.5 exposure. Even with limited literature, the use of more precise pollution inputs allowed the sub-group analysis to further quantify the effect differences between individual and population research. The difference in long-term PM_2.5 exposure effects indicates increasing subject granularity typically enhances the significance of exposure effects (Guo et al., 2020; Shafran-Nathan et al., 2018). Individual studies can more accurately distinguish exposure differences and adjust for potential confounding factors, leading to less biased quantitative results, while the population-level assessment overlooks the exposure differences brought about by the characteristics divergence.

When the subject granularity is refined to the individual level, the importance of differentiating micro-environments (e.g., streets and indoors) becomes apparent. Most of the time, people are exposed in indoor environments, making the adoption of ambient levels as an exposure proxy problematic (Hu et al., 2022; Koistinen et al., 2001; Morawska et al., 2024). Factors like pollutant penetration and filtration (Adams et al., 2001), gaseous processes (Yang et al., 2023), as well as the diversity in source distribution and intensity (Cai et al., 2021; Miao and Ding, 2020; Shimada and Matsuoka, 2011) all contribute to the significant exposure differences among the micro-environments, which is unlikely to be depicted by using ambient levels alone. Among the literature included in the meta-analysis, only one considered the differences in individual NO₂ exposure across indoor and outdoor micro-environments, deriving a relatively higher quantitative effect (Hu et al., 2021). Such differentiation would be crucial for understanding individual cumulative exposure doses and their health impacts.

The sub-group analyses have shown that lower ambient pollution levels tend to yield higher risk values. However, it is noticeable that pollution simulation precision, subject granularity, and micro-environment differentiation ultimately influence the assessed pollution levels. The close interplay among these influencing factors can result in varying trends in the final results. Further research is needed to disentangle the extent of impact these key factors have on the exposure assessment.

3.4. Insights for environmental management

From our study, we identified some limitations in the localization evidence of the main air pollutant exposure to the Chinese population. Firstly, the majority of literature included in the meta-analysis focuses on the long-term effects of PM_2.5, while NO₂ and O₃ remain rare. The cohort studies on the exposure effects of air pollutants in China have only been reported in recent years, and the literature included in this study is still limited, especially for NO₂ and O₃. In particular, about half of the cohort studies focus on elderly or middle-aged populations, and there is still a lack of evidence on the long-term effects of air pollution on young people. It could introduce a considerable bias in estimating integrated effect values. Moreover, there is no evidence of long-term O₃ exposure effects on respiratory mortality in the Chinese population, and the significance of the impact of long-term NO₂ exposure on cardiovascular and respiratory mortality remained with high divergence. Besides, this study found some publication bias, consistent with the results of other analyses, including meta-analyses conducted within China (Luo et al., 2023) and internationally (Orellano et al., 2020). This bias primarily exists on the right side of the funnel, where some effects are concentrated in the high-value area. Therefore, it cannot be ruled out that RR values may be overestimated in existing studies (Orellano et al., 2020). Also, during the literature identification and exclusion process, we found a lack of research on the air pollution exposure risks of specific occupations in China. Conducting such studies could provide evidence under high-exposure levels and offer potential research opportunities to further explore pollutant exposure mechanisms. Last but not the least, for short-term studies, the exposure is mostly unique for the population, while some has refined to individuals for long-terms studies. Our results have shown the difference and highlight the necessity of the refinement.

To further reduce the risk of population exposure and achieve the grand vision of a “Healthy China” and a “Beautiful China,” based on the limitations of existing studies and key factors affecting the accuracy of exposure assessments, we offer the following insights for environmental researchers, policy-makers and for the public. Researchers form scientific basis for management, policy-makers make direct decisions, and the public serves as the largest audience and practitioners of these efforts.

For researchers: i) continue enriching studies on the exposure effects of air pollutants, especially the long-term exposure effects of NO₂ and O₃; 2) identify the independent effect of crucial air pollutant components and explore their interactive effects to lead a deeper understanding of exposure mechanism; iii) further refine the exposure assessment methods, shifting focus from populations to individuals, and distinguishing important exposure micro-environment; iv) make mechanistic improvements to exposure proxies (such as considering activity levels and physiological differences) to reflect exposure differences, e.g., transitioning from concentration assessment to internal exposure assessment. Furthermore, we additionally suggest recommendations for pollutant simulation methods. For NO₂, due to its regional sources and primary emission from traffic, its temporal and spatial variability is highly significant and pronounced. Therefore, it is recommended to use medium-to small-scale coupled models for detailed simulations at the neighborhood scale (Lv et al., 2022). For O₃, the choice of simulation model needs to be flexible (Li et al., 2023a), as the formation of O₃ is closely related to the complex emissions of volatile organic compounds (VOCs) as well as NO₂. Considered the balance between computational efficiency and representativeness of the reactions (Ren et al., 2022), models that simulate NO₂ distribution using a basic photochemical reaction mechanism (NO₂-NO-O₃) are the most commonly employed. For PM_2.5 with quite complex sources, detailed modeling of local sources’ impact is needed. A convenient approach is to use the output of medium-scale models, applying machine learning methods for ultra-high-resolution assimilation (Wang et al., 2023b), or to use medium-to small-scale coupled models similar to those used for NO₂ simulation (Kim et al., 2022).

For policy-makers: i) advocate for the widespread adoption of green technologies and clean energy, particularly in sectors like coal combustion, smelting, and construction, and eventually transit into atmospheric pollution to ultra-low levels to reduce public health risks; ii) imply stricter vehicle emission standards and control non-tailpipe emissions from vehicles, reducing high exposure caused by high emissions in street environments, and similar for other specific micro-environments; iii) enhance atmospheric pollution forecasting and inversion capabilities by integrating multi-source data, machine learning algorithms, and small-scale modeling for the identification of emission and exposure hotspots and ensuring the effective implementation of regulatory measures.

And for the public: i) increase awareness of air pollution and its health impacts, and take appropriate exposure interventions to reduce health risks if necessary; ii) advocate for green lifestyles and reduce individual pollution emissions, especially during periods of heavy pollution; iii) involve in air pollution supervision and made feedback to ensure policy implementation, e.g., report pollution hotspots under no supervision; etc.

4. Discussions

Through meta-analysis, we comprehensively synthesized epidemiological evidence regarding the impact of major air pollutants on long- and short-term mortality in the Chinese population. Sufficient evidence from certain exposure-outcome pairs supported sub-group analyses based on exposure assessment precisions and pollution levels. We found that more refined assessment methods and lower pollution levels potentially yield higher exposure effects. This variability drove us to investigate the sources of heterogeneity, ultimately identifying four key factors influencing the impact of exposure assessment precisions on effect calculation, with pollution simulation precisions being the core factor. Based on these analyses, some forward-looking insights were proposed for environmental management in China. We believe our study offers some advantages over single epidemiological studies: i) we integrated a broader range of the most updated evidence, enhancing the accuracy of exposure effects and reducing biases; ii) by reanalyzing a large volume of samples from the literature, we systematically addressed areas that biased studies may not be able to reach; and iii) meta-analysis allows for a systematic evaluation of the complexity and precision of methodologies and their impacts, something that cannot be achieved through individual studies alone.

We exclusively focused on exposure effect evidence within the native Chinese population, revealing potential differences from those in developed countries. There are many debates and discrepancies regarding the effects of exposure in China and globally. The RR (and its 95% CI) for long-term exposure to all-cause mortality in WHO AQG 2021 are reported as 1.08 (1.06-1.09) for PM_2.5, 1.02 (1.01-1.04) for NO₂, and 1.01 (1.00-1.02) for O₃ (WHO, 2021b)(WHO, 2021b). In contrast, the integrated RR in this study is 1.12 (1.06-1.19) for PM_2.5, 1.11 (1.03-1.20) for NO₂, and 1.12 (1.01-1.24) for O₃, significantly higher than the effect from WHO AQG 2021. Similar trends are also reflected in the effects of long-term PM_2.5 exposure on cardiovascular and respiratory mortality: 1.11 (1.09-1.14) and 1.10 (1.03-1.18) in WHO AQG vs. 1.16 (1.09-1.25) and 1.17 (1.01-1.37) in this study, respectively. The reasons for the higher effects on the Chinese population are complex. Some studies owe it partly to the relatively higher levels of air pollution exposure, the presence of potential differences in sources, and chemical compositions of air pollution in China with unclear interactions among different pollutants and components that may contribute to the differences in exposure effects (Huang et al., 2023a). However, existing explanations are not yet sufficient. Some studies suggest that exposure effects among the Chinese population may be lower than in developed countries, attributing this to potential adaptive responses at higher exposure levels leading to lower impacts (Liu et al., 2019a). The inconsistency in these results requires further in-depth investigation in the future. In comparison, the RR of short-term exposure in this study is comparable to previous research, e.g., 1.0092 (95% CI: 1.0061–1.0123) for cardiovascular mortality, and 1.0073 (95% CI: 1.0029–1.0116) for non-malignant respiratory mortality reported by the WHO AQG 2021 (WHO, 2021b)(WHO, 2021b).

It shows a tendency that the short-term effects of NO₂ on mortality are greater than those of PM_2.5 and O₃, whereas the opposite is observed in the long term. This may reflect potential differences in the exposure mechanisms of different pollutants. NO₂, as a strong oxidant, may react with water to produce nitrate in the short term, potentially causing more severe oxidative reactions and epithelial cell damage, leading to fatal acute diseases (Huang and Tucker, 2020). Moreover, the long-term effects of NO₂ exposure on cardiovascular and respiratory mortality in the Chinese population were insignificant. This contrasts with some recent evidence from other regions, e.g., from the ESCAPE project (Beelen et al., 2014; Dimakopoulou et al., 2014). However, in a European cohort study from the ELAPSE project (Brunekreef et al., 2021), long-term NO₂ exposure showed significant associations with cardiovascular disease (1.089, 1.060-1.120) and respiratory disease (1.101, 1.038-1.168) mortality risks. A study from New Zealand (Hales et al., 2021) presented similar findings. The various scenarios presented in the literature demonstrate inconsistencies in the strength and significance of effects across studies worldwide, not limited to studies conducted in the Chinese region. The literature we included in our study had exposure levels ranging from 40-100 μg/m³, while studies from Europe and the United States had lower NO₂ levels, around 20 μg/m³ or even lower. We speculate that one of the reasons for this lack of significance may be the presence of sublinear exposure effects under high pollution levels (Zhang et al., 2021), although this is inconclusive. Additionally, studies have suggested the possibility of exposure misclassification, which may pose a minor threat to the validity of the conclusions (Yorifuji and Kashima, 2020). Moreover, there may be potential uncontrolled confounders that result in substantial changes in correlation between models after a series of adjustments (Yorifuji and Kashima, 2020). Further research is needed to enrich the relevant studies.

The impact of the exposure assessment method precision on calculation of the exposure effect has been proved by the state-of-art epidemiological studies. In a cohort study involving 60 million people in the United States (Di et al., 2017), the exposure effects derived from multisource integrated pollutant data were significantly higher than those from using pollution levels within a 50 km radius area from monitoring stations, regardless of PM_2.5 or O₃. This indicates that characterizing exposure at the city level using only the average values from sparse monitoring stations is challenging and may not accurately capture the real differences among populations. Furthermore, sub-group analyses of the long-term exposure effects of PM_2.5 suggest that refining the assessment from the urban population to individuals can also have an impact. Studies have compared the exposure assessed at the individual level using stochastic exposure simulation methods with those from the traditional average level and found the exposure effect estimated from individual exposure was higher (Chang et al., 2012). This indicates a negative bias in exposure-response response when ambient levels are used as proxies for exposure. Such negative bias may also be introduced by exposure misclassification (Kloog et al., 2013). Most studies allocate exposure levels based on residential addresses, while actual exposure may occur in any corner of the city. Moreover, crude exposure assessment methods may not account for various factors influencing individual exposure and the health effects associated with environmental pollutants, such as higher pollution levels near local emission sources, indoor and outdoor exposure differences, population mobility patterns (Özkaynak et al., 2013), etc. The uncertainty introduced by the methodology may reduce comparability between studies and should not be overlooked (Yang et al., 2019).

The phenomenon that a decrease in PM_2.5 levels can lead to an increase in RR has been supported by many published studies. As early as the years when the integrated exposure-response curve was developed, it was proposed that the concentration-response function exhibits a sub-linear relationship at higher concentrations (Burnett et al., 2014). Although this argument was revised with the adoption of the global exposure mortality model, it is still evident that at higher concentrations, the slope of the curve shows a slight decline after the cutting point near ~40 μg/m³ or even higher (Burnett et al., 2018). Within a wide range of PM_2.5 levels, a breakpoint has been observed at higher exposure levels. A study involving 1.9 million adults aged 35 to 75 (Li et al., 2023b) found a breakpoint, and other research has identified it near 60.9 μg/m3 in the all-cause mortality curve (Li et al., 2018b), above which the curve became sub-linear and gradually flattens. A few studies have explored the reasons. For example, research published in 2011 attempted to explain it through the adaptability of residents living in areas with severe air pollution (Ma et al., 2011). This argument has been supported by subsequent studies (Cao et al., 2012; Xu et al., 2020b). Additionally, some research suggests that this phenomenon might be due to survivor bias, as susceptible individuals may have passed away before the pollution levels reached such severe levels, leading to a certain degree of bias in the sample (Wong et al., 2008). Finally, the component characteristics of high PM_2.5 pollution could be significantly different from those of low PM_2.5 pollution (Zhu et al., 2018), as emission policy scenarios are likely to be totally different, which accounts for the observed differences (Cao et al., 2012). More research is needed to better explain the relationship between pollution levels and RR values and to accumulate additional evidence.

We consider the existing studies included in our research to be the most scientifically rigorous and advanced in terms of methodology, strictly adhering to relevant protocols in the research process. However, regarding the risk of bias, we still observe a high degree of heterogeneity among existing studies, with varying levels of bias or tendency. We raise this for discussion and concern because such biases may not necessarily reflect the actual situation of the research. We observed variability in research details within literature, which may be the potential contributors of the heterogeneity. Factors such as population sampling and structure, as well as inconsistencies in confounding factors, are sometimes limited by the available databases and the specific research regions, leading to systematic biases due to methodological or data structure inconsistencies. However, based on multiple sensitivity analyses, we believe that these differences do not compromise the robustness of our findings. Given the China’s large population and vast geographical coverage, along with the complex array of factors influencing exposure effects, these systemic biases may be objectively present and difficult to avoid. Instead, there might be the possibility that with further research, such biases might potentially serve as representatives of the diversity of exposure effects.

Comparison of pollution control strategies between different countries can highlight the importance of improving the accuracy of monitoring pollution levels. According to the U.S. EPA Air Quality System, approximately 5,000 air quality monitoring sites are operational across the United States. The European Union member states have collectively established around 4,000 air quality monitoring stations. In China, the National Urban Air Quality Monitoring Network comprises nearly 1,800 stations covering 367 key cities. Comparatively, China’s urban air quality monitoring network is on par with those in Western countries, though it is primarily concentrated in central and eastern regions, with relatively lower coverage in suburban areas and cities with lower population density. Many countries have deployed low-cost sensors widely to provide high-resolution real-time data (Feinberg et al., 2019; Jiao et al., 2016), aiding in the identification of pollution sources and hotspot areas. China has also begun deploying low-cost air pollution sensors in major cities and industrial areas (Chao et al., 2021), but there is still room for improvement in the coverage and density of technology and data applications. In addition, some countries have imposed restrictions on population exposure levels. The European Union released the revised Environmental Air Quality Directive in 2022, setting targets to reduce population average exposure levels by 2030. In contrast, China is still focusing on the ambient pollution levels, leaving constraints on population exposure levels in the future plan. The aforementioned differences highlight urgency for more accurate exposure assessment methods and assessment indices in China.

5. Limitations

Still, this study has some limitations. The current study was not registered, but we still followed the systematic review procedures rigorously. For long-term effects, due to the limited number of included studies, there may exist estimation deviations to the true effect values, and we could not assess the publication bias of the included studies and exclude the probability of potential publication bias. Although this is acceptable (Yang et al., 2019), further assessment is needed as the evidence evolves. Moreover, some studies have suggested that pollutant exposure may interact with other factors, such as high temperatures, climate, and greenness, and that the effects of various pollutant exposures are not independent. Last but not least, although the independent effects of PM_2.5 components have received attention, the lack of extensive observational methods and the inherent biases in pollution data introduce considerable uncertainty into the independent effects of various components (Yang et al., 2019). Limited by the existing evidence, this study did not further analyze and discuss these independent effects.

6. Conclusions

Our present meta-analysis reveals distinct associations between long- and short-term exposure to three primary air pollutants, PM_2.5, NO₂, and O₃, and an elevated risk of all-cause, cardiovascular, and respiratory mortality in the Chinese population. Finer exposure assessment methods potentially increase the quantified outcomes of exposure effects. There is a tendency for the short-term exposure risk change of pollutants to become larger with the decrease in pollutant levels. Four key factors influencing effect calculation in exposure assessment were identified: pollution simulation resolution, subject granularity, micro-environment classification, and pollution levels. The literature identification and the systematic review highlight the urgent need for more research, improvement in exposure assessment precisions, and pinpoint evaluation for risk management of individual exposure in China. And finally, based on the up-to-date exposure evidence, we provide relevant recommendations for researchers, policy-makers, and the public to further advance exposure assessment methods, promote environmental pollution control, and enhance awareness of environmental health.

Highlights.

• The meta-analysis integrated local air pollutant exposure evidence for China.

• Finer exposure assessment and lower level tend to derive higher exposure effect.

• Exposure effects of NO₂ response significant to exposure assessment precisions.

• We identified four key factors in exposure assessment.

• Refined exposure assessment is urgent for precise environmental management.

Figure 2. Short- and long-term exposure effects of PM_2.5, NO₂, and O₃ on all-cause, cardiovascular, and respiratory mortality.

Figure 4. Exposure effects of PM_2.5, NO₂, and O₃ under different pollution levels before and after 2013 on all-cause, cardiovascular, and respiratory mortality.

Data sharing statement

As this study is a systematic review and meta-analysis of published studies, all data are publicly available in the specified references.

This work is also supported by the Pathways to Equitable Healthy Cities grant from the Wellcome Trust (No. 209376/Z/17/Z). For the purpose of Open Access, the author has applied a CC BY public copyright license to any Author Accepted Manuscript version arising from this submission.