Although it is a part of the community-based collaborative innovation program against the hepatitis B virus, the cohort was established to investigate the prevalence of, as well as the underlying predictors for, several major infectious diseases, such as hepatitis B virus, and major chronic diseases, such as respiratory and cardiovascular outcomes [17]. Due to the availability of data, this study recruited 654,115 participants from 35 randomly selected communities in Guangzhou between 2009 and 2015 [18,19]. Inclusion criteria included participants being permanent residents, capable of undergoing a physical examination, and willing to participate with a signed informed consent form. Exclusion criteria included an inability to complete the questionnaire or physical examination, or the presence of significant mental or cognitive abnormalities evaluated through a combination of past medical history and subjective verbal reports. The flowchart of the study sample selection can be found in Figure S1 in Multimedia Appendix 1. Participants younger than 18 years (n=72,330) or with unknown cause of death (n=1028) were excluded, resulting in a total of 580,757 participants who followed up till 2020.

The information was collected by the interviewers using a computer-based standardized structured questionnaire, which included general demographic characteristics, lifestyle behaviors, and other variables. Physical activity was assessed by participants’ self-reported frequency of exercise in the past week. BMI (kg/m²) was calculated by dividing weight by the square of height, with height and weight being measured according to standard protocols.

The data in our study has been de-identified. This study, which did not involve invasive measures on the participants, was thoroughly reviewed and approved by the institutional review board committee at Sun Yat-sen University (L2017030), ensuring all procedures followed the relevant ethical tenets of the Declaration of Helsinki.

We determined the vital status of each participant during the period from enrollment to December 2020 and confirmed the cause of death for those who died by linking the cohort data to the Death Registry Systems of the Guangzhou Centers for Disease Control and Prevention via the national ID. The main outcome of this study was mortality from respiratory diseases (International Classification of Diseases, 10th Revision [ICD-10]: J00-J99).

Ambient PM data at a 1-km spatial resolution over mainland China were extracted from the ChinaHighAirPollutants database, which has been used in prior studies [20,21]. The concentrations of PM₁, PM_2.5, and PM₁₀, which are more likely to be objective, were obtained from space-time extremely randomized tree models based on meteorological, land use information, and other factors [22-25]. Initially, we converted the residential addresses of participants into standardized longitude and latitude coordinates, and assigned the annual average levels of PMs based on those coordinates using a nearest-neighbor matching approach.

The normalized difference vegetation index (NDVI) is an objective indicator of vegetation cover provided by the Terra Moderate Resolution Imaging Spectroradiometer (MODIS) Vegetation Indices (MOD13Q1). The average NDVI value within a 500-meter buffer around each person’s residential address was assigned based on latitude and longitude coordinates.

The marginal structural Cox model was used to estimate the associations of PM exposure with respiratory mortality, accounting for time-varying variables in the model. This approach is a causal inference method for observational data, which have been increasingly used in recent years [26]. First, we calculated the generalized propensity scores by regressing the exposure variable against the observed confounders. Stabilized inverse probability weights (IPWs) were then developed by using the inverse of the generalized propensity scores to weigh each participant (ie, creating a pseudopopulation) [27]. Then, a time-varying Cox model was developed to estimate the associations of PM exposure with respiratory mortality, accounting for time-varying variables in the model. A total of 3 different methods (ie, linear model, generalized estimating equation, and machine learning) were used to create the IPW, with the generalized estimating equation weighting method being selected due to its superior performance in covariates balancing as diagnosed by the minimal average absolute correlation values (Figure S2 in Multimedia Appendix 1). The generalized variance-inflation factor was used to assess collinearity among the covariates, with values less than 2 indicating no issues of collinearity. Additional methods’ details can be found elsewhere [28].

Table S1 in Multimedia Appendix 1 shows the proportion of missing data for all variables, which were imputed using the chained equation [29]. The complete data set after imputation was used in this study. Finally, we developed 3 models.

First, model 0 was an unadjusted model within the time-varying Cox model framework. Second, model 1 was model 0 plus confounders, including sex, age, demographic and behavioral factors, and NDVI, which were selected based on prior studies. Finally, model 2 was this model refitting model 1 using the marginal structural Cox model and was the final model. Subsequently, the results from the ultimate models were stratified. To investigate the potential influence of exposure to low concentrations, the analysis was further restricted to individuals having an annual PM₁₀ (the largest particles containing PM_2.5 and PM₁) concentration below 70 μg/m³ (World Health Organization [WHO] interim target 1) [30].

Several sensitivity analyses were carried out. First, we compared the main results using 3 different IPW methods. Second, the analysis was repeated using nonimputed data to assess whether the results were influenced by the imputation of missing data. Third, we used different sizes of buffers for the NDVI (ie, 250 m and 1000 m), and assessed the potential impact of unmeasured confounding factors in observational studies using the E-value [31].

In a Chinese community-based cohort study, our results showed that each 1-μg/m³ increase in the concentration of PM₁, PM_2.5, and PM₁₀ was linked to increases in the risk of respiratory mortality by 6.6%, 4.2%, and 4.0%, respectively. Moreover, this risk was reported to be higher among older participants, never-smokers, participants with high exercise frequency, and those residing in low-NDVI areas than among their counterparts. Participants in the low-exposure group tended to be at a 7.6% and 2.7% greater risk of respiratory mortality following PM₁ and PM₁₀ exposure, respectively, compared to the entire cohort.

Our findings showed significant associations of PM exposure with respiratory mortality. Former research on the PM_2.5-respiratory mortality association yielded inconclusive results; a 2013 meta-analysis found a nonsignificant association [32]. However, another updated meta-analysis of 17 studies indicated an elevated risk of respiratory mortality attributable to PM_2.5 and PM₁₀ [3]. In recent years, an increasing number of studies have focused on the health impact of PM₁. A time-series study in China examined the effects of short-term exposure to PM on mortality, showing that a 10-μg/m³ increase in PM₁ concentration was associated with a 0.35% increase in the excess risk of daily mortality [33]. Another 2 case-crossover studies in China found the same increase was linked to a 9% and 56% increase in the hospitalization risk of respiratory diseases and the incidence of doctor-diagnosed asthma, respectively [34,35]. Our research provides further insights into the establishment of national standards concerning PM₁ and developing policies to reduce the health risks associated with ambient air pollution. In the future, it is essential to explore the biological mechanisms involved, including how PM penetrates respiratory systems and the biochemical interactions at the cellular level.

Stratified analysis showed that the older participants and never-smokers were more susceptible to the respiratory impact of PMs than their counterparts. A systematic review of 20 studies consistently indicated greater vulnerability among older adults [36]. One possible explanation is that aging reduces the physiological capacity in the lungs for gas exchange, respiratory mechanics, muscle strength, and ventilatory control, causing prolonged exposure to toxins and resulting in greater lung damage [37]. In regard to the modification effect of smoking status, a cross-sectional study across 6 countries consistently found that the PM_2.5-stroke effect among never-smokers was greater than that among ever-smokers [38]. However, the exact mechanisms need to be further investigated. The mechanisms underlying the interaction between smoking and PMs remain unclear. One possible explanation is that both smoking and exposure to ambient PMs have negative effects on health through pathways involving oxidative stress and inflammation [39]. It also suggests that the negative effects of smoking are dominant in smokers due to their limited kidney function reserve [40], so additional exposure to ambient PMs may not further enhance the effects. However, further research on mechanisms is needed to clarify this hypothesis.

Additionally, we observed that participants with high exercise frequency and lower green space exposure showed higher vulnerability to PMs. The interaction effects between physical activity and PM exposure are controversial in prior studies. In a meta-analysis, exposure to exercise-related air pollution may reduce the positive effects of physical activity on cerebral health [41]. A recent cohort study conducted in China found that when the PM_2.5 concentrations exceeded 54 μg/m³, participants with farming activity levels greater than 15.74 metabolic equivalent of task (MET)–hours per day had a 12% and 18% elevated risk of cerebrovascular disease compared to those with activity levels between 3.43‐8.05 and 8.06‐15.73 MET-hours per day, respectively [42]. Another cohort study across 6 countries revealed that high levels of physical activity could increase the association between PM and stroke [43]. To further clarify this hypothesis, more research is required. Regarding the interactions between PM and green space exposure, in line with our findings, a cohort study among Chinese populations indicated a synergistic effect on mortality when PM decreased while greenness increased [44]. Possible reasons for the health benefits of green space exposure include that green plants absorb harmful air pollutants and help alleviate mental stress, reducing symptoms of anxiety and depression, and promoting psychological relaxation [45]. Therefore, both existing evidence and our findings suggest that greenness can serve as a protective measure to mitigate the adverse effects of air pollution.

Another significant finding was that the low-exposure group showed a stronger association between respiratory mortality and PM exposure compared to the overall population. Our results align with other research that has also found a greater association between PM and mortality among those exposed to lower air pollution concentrations [46,47]. A cohort study among the Canadians showed a 10.4% greater effect estimate over the 0‐5 μg/m³ range compared to the 5‐12 μg/m³ range (23.7% vs 13.3%) [48]. A possible reason is that participants in highly polluted areas may take measures to protect themselves from the harmful impacts of ambient air pollutants on health, such as turning on air purifiers or wearing masks during severe air pollution days. These results suggest that the same concentration reduction in air pollution at relatively lower levels may result in greater health improvements.

We provided the first cohort evidence of the impact of long-term ambient PM exposure across different particle sizes (especially PM₁) on respiratory mortality among the general Chinese population. Currently, there are an insufficient number of epidemiological studies on the effects of PM₁, and no established monitoring standards exist for long-term PM₁ exposure. Therefore, our study provides important quantitative evidence for the establishment of standards and guidelines for PM₁. In addition, our study used causal inference models and considered the temporal changes in PM features by using time-varying PM data, which ensured more robust effect estimates.

However, some limitations need to be discussed. First, the environmental exposure data for the study participants were obtained by matching the pollution simulation with each participant’s residential address. Other individual factors, such as personal indoor air pollution and outdoor physical activity time, were not considered, which may induce individual exposure measurement errors. However, this type of exposure misclassification is usually considered nondifferential and not likely to result in significant bias [49]. Second, owing to the absence of specific sources or chemical components of air pollution, our study could not identify the specific chemical components contributing to the higher risk of mortality from respiratory disease. Third, although we adjusted for demographic characteristics, socioeconomic status, lifestyle factors, and other potential confounding variables, residual confounding could not be ruled out due to the presence of unobserved confounding factors, such as occupational exposure to air pollutants or indoor air quality. However, the sensitivity analysis of E-values indicated that the effect estimates were robust against unobserved confounding factors.

This study suggests potential causal links between PM exposure and respiratory mortality. In addition, older participants, nonsmokers, participants with high exercise frequency, and low-NDVI residents, as well as those in the low-exposure group, appear to exhibit a higher vulnerability to the impacts of PMs, highlighting the necessity for developing strategies focused on reducing PM concentrations and protecting susceptible populations.