Study design

A time‑stratified case‑crossover design was employed to investigate the association between air pollutants and both nonaccidental and cause‑specific deaths. In this design, each person was considered a case on the date of death and served as their own matched control before and / or after the date of death within the same month (ie, a time stratum).¹⁶ Specifically, each case was treated as a case day on the date of death and matched with 3 or 4 control days that fell on the same weekday as the case day within the time stratum.¹⁷ This approach helps control for the effects of weekly, seasonal, and long‑term time trends.¹⁸ For instance, deaths occurring on Saturday, August 11, 2018, were defined as the time stratum, with that date designated as the case day, while all other Saturdays in August 2018 (August 4, 18, and 25) were identified as the corresponding control days. This design allows for adequate control of individual factors (eg, sex, age, race, or marital status) that are time‑invariant or unlikely to change significantly across time strata.¹⁸

Data collection

As one of the representative regions in the Yangtze river delta, the Qingpu District is designated as a key demonstration zone for the Yangtze river delta integration strategy, reflecting its role in the regional economic and ecological cooperation. It covers an area of 668.5 km² and has a resident population of approximately 1.266 million. Administratively, the Qingpu District consists of 3 subdistricts (streets)—typically urbanized areas with dense populations—and 8 towns that include both urban and rural communities. All 11 subdivisions are directly governed by the district, with no hierarchical dependency between streets and towns.¹⁹ This study primarily focuses on the permanent residents of these 3 streets and 8 towns in the region.

Information on the deaths (both at home and health care facilities) was collected from the cause‑of‑death registration system of the Centre for Disease Control and Prevention in the Qingpu District, Shanghai. These data spanned from March 21, 2013 to June 30, 2019, and included details such as sex, date of birth, date of death, age at death, and the code for the underlying cause of death. The International Classification of Diseases, Tenth Revision, was employed to categorize reported fatalities into nonaccidental (A00–R99), cardiovascular (I00–I99), and respiratory (J00–J99) categories, after excluding accidental deaths. For each of the 22 149 individuals who died from nonaccidental causes in Qingpu, Shanghai, China, between 2013 and 2019, we collected data on the cause of death, date of death, usual residence address, and sociodemographic characteristics. Among these, 6712 individuals died of diseases of the cardiovascular system, while 3275 individuals died specifically of diseases of the respiratory system.

Air pollutants and meteorological data

Data on air pollution were provided by the Shanghai Municipal Bureau of Ecology and Environment, and we focused on mean daily concentrations of PM_2.5 (μg/m³), SO₂ (μg/m³), NO₂ (μg/m³), O₃-8h (μg/m³), and CO (mg/m³) from 3 air monitoring stations in the Qingpu District (including Xiayang, YingPu, and Zhujiajiao subsites) located in the center of the study area. Detailed descriptions of these instruments and methodologies were given in a previous study.²⁰

The meteorological data for this study were mainly extracted from the data of automatic meteorological observation stations in Qingpu District from 2013 to 2019, including the average daily temperature (°C) and relative humidity, and the values of each pollutant were taken as the average value of each station.

Statistical analysis

In this study, demographic characteristics, ambient air pollutants, and meteorological conditions for continuous variables were reported as mean (SD), while categorical variables were expressed as number and percentage. The relationships between PM_2.5, O₃-8h, NO₂, SO₂, CO, and meteorological parameters were assessed using correlation coefficients. Additionally, a conditional logistic regression model was employed to evaluate the impact of a brief exposure to air pollution on the likelihood of overall and cause‑specific death.²¹ We also presented composition‑death relationships for single days (lag 0 to lag 7 days) and moving average lags (lag 01 to lag 07 days) to account for the delayed effects of ambient air pollution. Lag‑01 was defined as the 2‑day moving average of the day of death and the day prior, and so forth. The date of death was recorded as lag‑0, while the day before was designated as lag‑1. We also assigned single‑day and moving average delays to each control day using the same methodology. In this study, we utilized a case‑crossover design and conditional logistic regression analysis to adjust for individual characteristics, including economic factors. A natural spline function with 3 degrees of freedom was employed in all models to account for relative humidity and temperature.²² To investigate the key patterns in the relationship between air pollution and death, we generated exposure‑response curves using air pollutant exposure modeled as a 3‑degree of freedom natural spline function during the periods when each air pollutant was most significantly associated with death. If a nonlinear correlation was observed, a 2‑piece‑wise linear regression model was employed to calculate the threshold effect.

Utilizing weighted quantile sum (WQS) regression analysis, this study further investigated the collective and individual impact of 5 air contaminants on nonaccidental deaths as well as system‑specific death.²³ We selected exposure windows during which the majority of air contaminants exhibited a strong association with death. This technique employs weighted indices as individual exposure variables within a linear model, assigning weights based on air pollution quartiles. The data were divided into training (40%) and validation (60%) subgroups. For each air pollutant in the training subset, we calculated the WQS mixed exposure index by taking a weighted average of over 10 000 bootstrap resamples.²³ In the validation subset, we assessed the association between the likelihood of death and the WQS index. The odds ratio (OR) with 95% CI was used to demonstrate the association between death and each one‑quarter increase in all compounds within the mixture.

By applying conditional logistic regression modeling, we further assessed the modifying effect of temperature on the relationship between air pollution and death. First, the impact of air pollutants on death varies continuously with temperature, as demonstrated by the discrete simulation of the atmospheric pollutant‑temperature response surface represented by temperature stratification.²⁴ The relationship between air pollution and temperature was then analyzed using the cross‑product of temperature categories (low [<⁠25th percentile], medium [25th–75th percentile], and high [>75th percentile]) with atmospheric contaminants.²⁴ Finally, by incorporating a multiplicative interaction term between ambient temperature and air pollutants, we evaluated the interaction effects.

Subsequently, a variety of sensitivity tests were conducted to assess the reliability of our findings. First, a stratified subgroup analysis was conducted based on age (<⁠65 years and ≥65 years) and sex (men and women). The study period was divided into warm and cold seasons according to the case date, allowing for stratified analyses to identify the seasonal patterns in the relationship between constituents and death. We defined the months of April through September as the warm season and the months of October through March as the cold season. A 2‑sample z test was employed to compare the differences in effects across the various stratifications.²⁵ Secondly, we employed a temperature stratification model based on the findings from the bivariate modeling in order to determine whether there is variability in the effects of air pollutants on health outcomes across different temperature levels.

All analyses were performed using R software (version 4.3.2; R Foundation for Statistical Computing, Vienna, Austria). Two‑sided P values below 0.05 were considered significant.

Ethics

Due to the design of the study, no approval of an ethics committee and patient consent were necessary.

Baseline characteristics of the study participants

The basic characteristics of nonaccidental deaths from 2013 to 2019 are presented in Table 1. The analysis included 22 149 nonaccidental deaths, of which 3275 (14.8%) were attributable to respiratory disorders and 6712 (30.3%) to cardiovascular conditions. The average age of the deceased was 78 years, with approximately 85% being over the age of 65. Men accounted for about 52.7% of the deceased, and 51.8% of the cases occurred during the warmer months, from April to September.

Summary statistics for air pollutants and meteorological conditions on case days are presented in Supplementary material, Table S1 and Figure S1. The mean case‑day exposures to PM_2.5, O₃-8h, NO₂, SO₂, and CO were 55.7, 107, 40.6, and 16 μg/m³, and 0.75 mg/m³, respectively. The mean case‑day air temperature was 17.4 °C, and the relative humidity was 75.2%. The Spearman correlation coefficients (r_s) between the daily atmospheric pollutants and meteorological factors are shown in Supplementary material, Figure S2. The majority of air pollutants were significantly and positively correlated (excluding O₃-8h), with medium‑to‑high r_s ranging from 0.47 to 0.67.

Associations of air pollutants with the probability of death

Significant associations were identified between exposure to PM_2.5 (lag 3, lag 03, lag 04, lag 05), O₃-8h (lag 1, lag 2, lag 02, lag 03, lag 04), NO₂ (lag 0, lag 1, lag 3, lag 01, lag 02, lag 03, lag 04, lag 05, lag 06, lag 07), and SO₂ (lag 3, lag 4, lag 02, lag 03, lag 04, lag 05, lag 06, lag 07) and nonaccidental death (Figure 1; Supplementary material, Table S2). Notably, the estimated ORs (95% CI) for PM_2.5 at lag 04, O₃-8h at lag 03, SO₂ at lag 07, and NO₂ at lag 06 concerning nonaccidental deaths reached their maximum values of 1.009 (1.002–1.017), 1.017 (1.008–1.027), 1.051 (1.015–1.089), and 1.027 (1.010–1.045), respectively. The nonlinear relationships between atmospheric pollutants and nonaccidental deaths were significant (P value for nonlinearity <⁠0.001 for PM_2.5, O₃-8h, and NO₂, P value for nonlinearity = 0.003 for SO₂), as illustrated in Supplementary material, Figure S3. After evaluating the associations using 2‑piece‑wise linear regression models, the cutoff points for atmospheric pollutants (including PM_2.5, O₃-8h, SO₂, and NO₂) related to nonaccidental deaths were determined to be 38.4, 100.5, 21.9, and 36.6 μg/m³, respectively.

Supplementary material Figures S4 and S5 display the OR estimates (with 95% CI) for cardiovascular and respiratory deaths per interquartile range increase in ambient air pollutants. Similarly, exposure to air pollutants significantly elevated the risk of respiratory death due to cardiovascular disease at a lag of 07 days, with an OR (95% CI) of 1.029 (1.005–1.053) for PM_2.5, 1.137 (1.054–1.226) for SO₂, 1.068 (1.022–1.116) for NO₂, and 1.041 (1.009–1.073) for CO (Supplementary material, Table S3). For cardiovascular death, significant associations were identified only for O₃-8h (lag 1, lag 2, lag 02, lag 03) and SO₂ (lag 2), while no evident associations were observed for the remaining air pollutants (PM_2.5, NO₂, and CO; Supplementary material, Table S4). To assess whether ambient air pollutants and reproductive lifespan had a nonlinear dose‑response relationship, smooth curve fits based on generalized additive models were also prepared. After controlling for all factors, the results demonstrated a linear relationship between the majority of ambient air pollutants (excluding O₃-8h) and cause‑specific death (Supplementary material, Figures S6 and S7). The nonlinear relationship between O₃-8h and cardiovascular death was significant (P value for nonlinearity <⁠0.001), with the cutoff point determined to be 99.3 μg/m³ (Supplementary material, Figure S6).

Association of a combination of air pollutants with death

We selected a lag of 03 days as the study period to examine the short‑term associations of the exposure to mixed air pollutions on nonaccidental deaths, as this is when the majority of air pollutants were significantly associated with nonaccidental fatalities. In the single‑pollutant analysis, pollutants other than CO strongly correlated with nonaccidental fatalities at lag 03 days. Exposure to a combination of PM_2.5, O₃-8h, NO₂, and SO₂ was significantly associated with an increased rate of nonaccidental fatalities (OR, 1.03; 95% CI, 1.01–1.05). The WQS model, illustrated in Figure 2, was employed to quantify the contribution of the assessed air pollutants to nonaccidental fatalities. The WQS model also identified PM_2.5 (weight = 0.557) and NO₂ (weight = 0.316) as significant contributors to the overall association (Figure 3; Supplementary material, Table S5).

In the single‑pollutant analysis of cardiovascular death, the only significant association was observed between O₃-8h and SO₂ levels at different lag days. Consequently, we further assessed the impact of the air pollutant mixture on respiratory death. We selected a lag of 05 days as the period for analysis, as this was when most air pollutants showed significant associations with respiratory death. The results indicated that the air pollutant combination was significantly and positively associated with respiratory death (OR, 1.1; 95% CI, 1.05–1.15), with SO₂ and PM_2.5 identified as the primary contributors (Figure 3; Supplementary material Table S5).

Interactive effects of temperature and air pollutants on death

Across various temperature levels, Figure 4 and Supplementary material Figures S8 and S9 illustrate the lag patterns of air pollutants on both overall and cause‑specific death. In general, there was a stronger association between an increase in O₃-8h level and death at high than low air temperature. For example, at low, medium, and high air temperature, a percentage increase in cardiovascular death of −0.06% (95% CI, −0.29% to 0.18%), 0.02% (95% CI, −0.10% to 0.14%), and 0.13% (95% CI, 0.02%–0.25%) was associated with every 10 μg/m³ in O₃-8h at lag 04 days (P = 0.001; Supplementary material, Table S6).

Sensitivity analyses

Table 2 presents the subgroupspecific excess risks (ERs; 95% CI) for cardiovascular, respiratory, and nonaccidental deaths associated with the exposure to ambient air pollution, categorized by age and sex. The sex‑specific findings indicate that women are at a significantly higher risk than men for nonaccidental death due to exposure to PM_2.5 and NO_2. For instance, the ER for PM_2.5 was 0.96% (95% CI, 0.21%–1.72%) for women and 0.16% (95% CI, –0.54% to 0.87%) for men (P = 0.02). A similar trend was observed in the relationship between the atmospheric pollutants and respiratory death, with PM_2.5 and SO₂ showing a more pronounced association with respiratory death among women. Additionally, stronger positive associations between the air pollutants (including SO₂, O₃-8h, and CO) and respiratory death were identified in younger individuals than those aged 65 years and older. However, the associations of all air pollutants with death did not vary significantly by season.

Moreover, to investigate the interaction between temperature and air pollution, we employed response surfaces as an experimental tool. Our findings showed that the effects of PM_2.5, NO₂, and CO at extreme temperature levels (both low and high) were more pronounced than those observed in the moderate temperature range for all 3 health outcomes (Supplementary material, Figures S10–S12). When comparing the low‑temperature layer with the other temperature layers, we found that air pollutants significantly influenced both death of specific cause and all‑cause death for O₃-8h and SO₂. Additionally, there were regions where the fitted response surface approached zero. The bivariate response suggested that temperature may affect the relationship between PM_2.5 and death.

Strengths and limitations

Our study has several strengths. To more effectively account for specific confounders that remained constant over time, we employed a case‑crossover design. This research, based on long‑term surveillance data, provides new evidence that exposure to various air pollutants increases likelihood of death in China. This finding could significantly influence regional environmental and public health regulations. Furthermore, we utilized the WQS model to assess the impact of exposure to a combination of multiple air pollutants on death, potentially offering valuable insights into the complex mechanisms underlying the relationship between pollutant mixtures and death. Lastly, our study provides first‑hand epidemiologic evidence of the association between multiple air pollutants and meteorological conditions on death in the general population.

The study also has some limitations. First, exposure to air pollutants was classified based on monitoring sites rather than the residential addresses of each fatal case, which may result in some degree of unavoidable exposure misclassification. This nondifferential measurement error could skew our results toward zero, leading to an underestimation of the true relationship. Second, to validate our findings regarding the robustness of the data, additional large cohort studies are necessary. Our data were derived solely from typical district‑level statistics in Shanghai, China. Third, we lacked sufficient data on the impact of individual behavioral risk factors (such as drinking and smoking habits) on participant death.²² Although employing a case‑crossover approach suggests that certain fixed‑term confounders are fully controlled, it is not feasible to assess other potential individual risk factors without documentation of these specifics. Moreover, the statistical WQS model does not consider the actual levels of exposure or toxic properties of each air pollutant. Finally, due to data unavailability, our analyses could not account for the confounding influence of seasonal influenza.

Conclusions

The study found that both individual and combined ambient air pollutants are associated with respiratory and nonaccidental death. This underscores the necessity for integrated air pollution control, prioritizing PM_2.5 and SO_2. Additionally, we found that high temperatures may amplify the associations of O₃-8h with daily death. Our findings reinforce the existing body of evidence linking mortality to air pollution, and serve as a guide for the development of intervention and mortality prevention programs.