This cross-sectional study obtained data from the CHARLS dataset, which is a national longitudinal cohort study that enrolled middle-aged and elderly adults in China, currently consisting of individuals aged 45 years and above, as well as their spouses or children when possible. The baseline survey of CHARLS consecutively employed a multistaged, stratified probability-proportionate-to-size sampling method to enroll respondents, which covered 450 villages in 28 provinces, 150 counties, and districts across mainland China. The baseline survey recorded sociodemographics, lifestyles, and health status through computer-assisted in-person interviews and subsequently assessed participants’ anthropometric measurements and biosamples, including blood and urine. Details about the recruitment method and designs of the CHARLS have been published previously (16).

The current study included the third follow-up of CHARLS participants in 2015, including a total of 13,337 respondents who had complete blood biochemistry data. Exclusion criteria included i) participants whose ages were under 45 (n = 1,150) and ii) participants with incomplete information on indoor fuel or missing anthropometric measurements (n = 421); the study finally included 11,766 individuals in the entire population dataset.

In the current study, BMI, waist circumference (WC), and visceral adiposity index (VAI) were all adopted to define obesity. VAI, an estimate of visceral adiposity that combines lipid parameters and anthropometric measurements, has been noted to be superior to general anthropometric parameters when predicting the occurrence of cardiovascular disease and metabolic syndrome (17, 18). Anthropometric measurements (height, weight) of each participant were recorded in accordance with WHO guidelines (19). Briefly, a stadiometer was used to measure height, and participants stood upright with their back against the vertical backboard of the stadiometer and kept their head in the Frankfort horizontal plane position. A scale was used to record weight, and participants stood with their shoes off (20). WC was measured with a soft measuring tape at the level of the navel. After obtaining anthropometric measurements, the participants’ BMI was computed using the standard formula of weight in kilograms divided by the square of height in meters. According to the guidelines for the prevention and control of obesity in the Chinese population (21), the obese state was defined by BMI cutoff value (obesity, BMI ≥ 28, non-obesity, BMI< 28). The modified WC cutoffs for Chinese populations were used to define central obesity (22), where WC ≥ 90 cm for men and ≥ 80 cm for women. VAI was calculated using a sex-specific formula according to a previous study and divided into quartiles when performing analysis (23), where triglyceride (TG) and high-density lipoprotein cholesterol (HDLc) levels were expressed in mmol/L:

We collected annual concentrations of sulfur dioxide (SO₂), nitrogen dioxide (NO₂), ozone (O₃), and particulate matter with aerodynamic diameters of ≤ 1 μm (PM₁), PM_2.5, and PM₁₀ levels at 10 km spatial resolution from the ChinaHighAirPollutants (CHAP) dataset. The method was described previously (24, 25). Briefly, the space-time extremely randomized tree (STET) model was utilized to estimate ambient AP concentrations in this study, which incorporates satellite remote sensing with machine learning techniques that account for the spatiotemporal variation in AP. The estimations were highly in accordance with surface measurements, as evidenced by average cross-validation coefficient of determination values (CV-R²) of 0.84, 0.93, 0.87, 0.77, 0.92, and 0.90 and root-mean-square error (RMSE) values of 10.07, 4.89, 17.10, 14.6, 10.76, and 21.12 μg/m³ for the daily concentrations of SO₂, NO₂, O₃, PM₁, PM_2.5, and PM₁₀, respectively.

Individual annual exposures to SO₂, NO₂, O₃, PM₁, PM_2.5, and PM₁₀ were geographically matched based on county-level longitude and latitude data. To minimize the disparities in air pollution exposure resulting from individuals’ outdoor work or commuting activities, exposures of AP assigned to each individual were determined as the average air pollution concentrations within a 15 km radius around their corresponding latitude and longitude coordinates. Then, 1-year average exposures for each participant in CHARLS wave 3 were employed in the current study.

Several potential confounders were controlled in the current study, including sociodemographics, lifestyles, health status and meteorological factors. Meteorological factors, such as temperature and specific humidity, were collected from the China Meteorological Administration’s meteorological monitoring stations (available at http://data.cma.cn/). Other confounding factors were collected via standardized questionnaires by experienced interviewers. In this study, sociodemographic covariates were sex (male and female), age level, which was categorized as middle-aged (45 to 64 years) and elderly (65 years and above), residence, which was grouped as rural and urban, and education attainment, grouped as “middle school or below” and “high school or above” (Supplementary Materials). Lifestyle-related covariates included smoking status, categorized as non-smoker, smoker, and second-hand smoker. In relation to smoking status, the term “smoker” referred to a person who had ever smoked before, while “second-hand smoker” was defined as someone who had never smoked himself but had family members who did; alcohol consumption, which was categorized into 3 groups: non-drinker, drink but less than once a month and drink more than once a month; indoor fuel use, which was classified as either “clean fuel use” or “solid fuel use”. Clean fuel use included solar energy, natural gas, biogas, liquefied petroleum gas, and electricity, while solid fuel use included coal, crop residue, and wood burning. Participants who never cooked and those who used other fuels were excluded according to previous study (26). Health status covariates included a self-reported history of hypertension and diabetes mellitus.

The baseline characteristics of the participants were described based on their obesity status. Continuous variables were examined by Kruskal−Wallis tests and presented as the median and the interquartile range (IQR). Categorical variables were presented as counts (percentages) and examined by Chi-squared tests. This study utilized a multilayered analysis strategy. Logistic regression models were performed to estimate the odds ratio (ORs) and 95% confidence interval (CI) of obesity in relation to outdoor AP exposure (per 10 μg/m³ increase in each air pollutant level). The crude estimates were present in Model 0. Based on Model 0, Model 1 was further adjusted for age at baseline visit and sex. Based on Model 1, Model 2 was additionally adjusted for sociodemographic factors (education attainment), lifestyle-related risk factors (alcohol consumption and smoking status), and indoor fuel use.

To investigate potential effect modification by sociodemographic or lifestyle factors, we conducted stratified analyses of the fully adjusted model (Model 2) by age, sex, education attainment, alcohol consumption, and smoking status. In each stratification, except for variables used for stratification, all other covariates were controlled in the models.

To evaluate the reliability of our results, we also conducted several sensitivity analyses via the following methods: a. including only participants who had resided at their present residential location for over 3 years, b. including meteorological factors (temperature and specific humidity) as covariates, c. including history of hypertension and diabetes mellitus as covariates, and iv) excluding participants with a personal history of cardiac events or stroke. A logistic regression-based restricted cubic spline model was implemented to assess the potential nonlinear associations between 6 air pollutants and obesity measured on a continuous scale. To strike a balance between model fit and overfitting of the primary spline for obesity, the knot value corresponding to the minimum Akaike information criterion (AIC) within the knot range from 3 to 5 was selected. Subsequently, variance analysis was conducted to test the statistical significance of the nonlinear relationships. The maximum ORs of obesity on the spline curves correspond to the air pollutant concentrations with the highest risk.

R version 4.2.1 was used for all statistical analyses in this study. Cases with missing covariate values were imputed by the nonparameter method missForest algorithm (27). Statistical significance in this study was defined as a p value less than 0.05 (two-tailed).

In the current study, we analyzed a total of 11,766 middle-aged and elderly individuals (see Figure 1). These respondents were enrolled from 445 communities located in 125 counties across 28 Chinese provinces. Figure 2 displays the geographical distribution of the study participants and the prevalence of obesity. The overall BMI-defined obesity prevalence was 13.3% (1,561/11,766). The prevalence of BMI-defined obesity in all provinces ranges from 5.88% to 26.87%. Tianjin, Xinjiang and Shandong had the top three prevalence rates at 26.87%, 25.93% and 21.92%, respectively (Supplementary Table 1). Regarding spatial distribution patterns, the prevalence of BMI-defined obesity in eastern China was generally higher than that in other regions, particularly in the northeastern regions, and lower in the southern and central regions. Xinjiang, a western province, also maintained a higher prevalence than other regions.

Table 1 shows the characteristics of individuals categorized by BMI-defined obesity. The respondents had a median age of 60 years old, 52.8% were female, and 61.94% lived in rural areas. Participants diagnosed with obesity tended to be slightly younger (57 vs. 60 years, P<0.001) and had a higher probability of being female (63.55%, P<0.001) and second-hand smokers (39.65%, P<0.001). In the current analysis, the median SO₂, NO₂, O₃, PM₁, PM_2.5, and PM₁₀ concentrations were 23.95 μg/m³ (range 16.37-36.22), 30.22 μg/m³ (range 21.84-38.12), 81.43 μg/m³ (range 74.94-92.30), 35.38 μg/m³ (range 31.53-41.08), 52.20 μg/m³ (range 39.62-64.34), and 86.69 μg/m³ (range 62.68-106.49), respectively. A similar distribution of the basic characteristics of participants grouped by WC-defined central obesity was observed and presented in Supplementary Table 2.

The association between AP exposure and VAI quartiles is summarized in Supplementary Table 3. The quartile ranges were ≤ 0.96, 0.96-1.50, 1.50-2.48, and ≥ 2.48. According to the trending analysis, with the increase in VAI quartiles, BMI, systolic blood pressure, diastolic blood pressure, fasting blood glucose, TGs, and the prevalence of hypertension and diabetes increased. Moreover, we also observed a significant increase in the average AP exposure with a step increase by VAI quartiles (all P for trend< 0.001).

Figure 3 presents the relationship between the prevalence of obesity and exposure to AP. The results indicate that obesity and AP exposure had positive associations in all three models (crude, Model 1, and Model 2) with varying effect sizes. In Model 2 of obesity defined by BMI, after fully adjusting for age level, sex, educational attainment, alcohol consumption, smoking status, and indoor fuel use, the ORs and 95% CI of obesity were 1.21 (1.17, 1.26), 1.33 (1.26, 1.40), 1.15 (1.10, 1.21), 1.38 (1.29, 1.48), 1.19 (1.15, 1.22), and 1.11 (1.09, 1.13) per 10 μg/m³ increases in SO₂, NO₂, O₃, PM₁, PM_2.5, and PM₁₀, respectively, suggesting that all 6 ambient AP exposures were significantly associated with the increased prevalence of obesity.

The results of Model 2 in obesity defined by BMI were similar to those defined by WC. When central obesity defined by WC was used as the outcome variable, the relationship between AP and obesity did not change significantly; however, AP exposure seemed to result in slightly higher ORs for central obesity defined by WC. The numerical results corresponding to the associations between 6 air pollutants and obesity presented in Figure 3 can be found in Supplementary Table 4.

Table 2 demonstrates the results of multivariate logistic regression modeling analysis between AP exposure and VAI quartiles. In Model 1, when comparing the highest quartile of VAI to the lowest quartile, the ORs for AP exposure were 1.13 (1.09, 1.18), 1.33 (1.26, 1.40), 1.13 (1.08, 1.18), 1.18 (1.10, 1.26), 1.10 (1.07, 1.13) and 1.07 (1.05, 1.09) per 10 μg/m³ increase in SO₂, NO₂, O₃, PM₁, PM_2.5, and PM₁₀, respectively. After additional adjustments for lifestyle risk factors, education level, and indoor fuel use, the ORs of AP exposure slightly decreased but did not change the association.

The results of several stratified analyses conducted to assess the effects of AP exposure in participants with different characteristics are presented in Table 3, as classified by BMI. Overall, the positive association between AP exposure and the prevalence of BMI-defined obesity was not substantially modified by sex, educational attainment, smoking status, or alcohol consumption. Notably, the observed positive association of NO_2, PM₁, PM_2.5, and PM₁₀ exposures was more pronounced among elderly participants (ORs: 1.50 (1.35, 1.67), 1.60 (1.37, 1.87), 1.26 (1.19, 1.33) and 1.15 (1.12, 1.19) per 10 μg/m³ increases in air pollutant above), while the relationship remained significant between O₃ and both female and second-hand smokers (ORs: 1.21 (1.14, 1.28) for female and 1.28 (1.19, 1.38) for the second-hand smoker), demonstrating a strong interaction between outdoor O₃ exposure and sex and smoking status on BMI-defined obesity. In addition, we observed that participants with lower education levels and those who were second-hand smokers or who drink more than once a month were more likely to develop BMI-defined obesity when exposed to ambient AP, despite there being no statistically significant interactions for education attainment and lifestyle risk factors.

Similar effects were also observed in WC-defined central obesity (Supplementary Table 5). However, there were no significant differences in the interactions between AP exposure and central obesity among participants stratified by different characteristics.

Table 4 displays the estimated ORs of obesity for all models. Sensitivity analysis 1 indicated that the estimated OR of obesity with AP exposure remained similar when only participants who had resided at their present residential location for over 3 years were included. Sensitivity analysis 2 found that when further including temperature and specific humidity, the ORs for obesity due to PM₁ exposure slightly increased. Specifically, the OR was 1.59 (1.42, 1.77) vs. 1.48 (1.36, 1.61) for BMI-defined obesity and 1.57 (1.46, 1.69) vs. 1.53 (1.44, 1.62) for WC-defined central obesity. In sensitivity analysis 3, there was no significant difference in all models after adjusting for either hypertension or diabetes history. In sensitivity analysis 4, the results were consistent with the primary model when a sample with a personal history of cardiac events or stroke was restricted. Overall, sensitivity analysis suggested that the main results were robust concerning the participants.

Supplementary Figure 1 illustrates the association between obesity and AP exposure on a continuous scale using the RCS model. The spline curves indicate a significant nonlinear relationship between obesity and SO₂, PM_2.5, and PM₁₀ exposures according to the RCS models. However, the nonlinear results of NO₂ and PM₁ differ between BMI-defined obesity and central obesity. Regarding the O₃ curves, linearity of the exposure-response association was observed in both BMI-defined obesity and central obesity. As depicted in Supplementary Figure 1A, the risk of obesity increased with the level of SO₂ when the SO₂ concentration was less than 42.9 μg/m³ in BMI-defined obesity and 46.2 μg/m³ in central obesity but decreased when the SO₂ concentration exceeded these values. With regard to the other 5 air pollutants, irrespective of the significance of the linearity effect within the observable range of the corresponding air pollution concentration, the ORs of obesity increased with the AP concentration. The results of the significance tests corresponding to Supplementary Figure 1 are shown in Supplementary Table 6.