Impact of Continuously Improved Air Quality on Mental Health: A Quasi-experimental Study Based on Longitudinal Observations for Chinese Adults

Abstract

Long-term exposure to fine particulate matter (PM_2.5) has been linked with adverse mental health outcomes. However, questions remain regarding the nature of lagged effects over time and by extension potential benefits over time of continued reduction in pollution. Here, we aim to estimate the long-term association between exposure to PM_2.5 and depressive symptoms in China utilizing longitudinal models for prolonged exposures as well as a quasi-experimental design utilizing data from 23151 participants over 4 longitudinal waves that occurred in 124 cities in China between 2011 to 2018. Mixed-effects models as well as distributed lag nonlinear mixed models were fitted to assess the relationship between PM_2.5 and depressive symptoms. We also assessed the effect of the Clean Air Policy (CAP) based on a quasi-experimental difference-in-differences (DID) design. The overall average PM_2.5 concentrations generally declined with time from 59.40 to 39.35 μg/m³. A 10 μg/m³ increase in PM_2.5 concentration was associated with a 0.86% increase (95% confidence interval [CI]: 0.1, 1.64%) in depression score based on the first three waves of data. However, the associations were sensitive to secular trends. Flexible exposure–lag–response analysis indicated a potentially influential window for lag-years 0–6. Reduction in PM_2.5 led to 19.51% ([CI]: 11.57%, 26.73%) and 28.18%, ([CI]: 5.87%, 45.2%) lower depressive scores in waves 3 and 4, respectively, compared to no reduction or increase in exposures. Our analysis suggests an association between PM_2.5 and depressive symptoms with potential long-term effects of air pollution as well as potential for continued benefit of air pollution reduction over time.

1. Introduction

Mental health disorders are now recognized as a leading contributor to the global burden of disease, accounting for 21–32% of the years lived with disability globally. In addition to morbidity, mental disorders can also cause increases in mortality (e.g., suicide). , While conventional factors such as perinatal depression, aging, lifestyle changes, and hormonal fluctuations are well-known triggers, recent research has increasingly focused on environmental influences, particularly in the context of climate change and rapid urbanization. − Among them, it has been suggested that chronic exposure to high levels of air pollution affects mental health. ,− However, the current research has mainly focused on developed countries and is limited in low- and middle-income developing countries such as China where severe air pollution levels under the status of rapid economic development exist.

To fully understand the relationship between mental health risks and air pollution in China, it is important to consider both large spatial and extensive temporal scales, given the spatiotemporal variations in pollutant concentrations. For instance, a recent study found exposure to higher concentrations of fine particulate matter (PM_2.5) could significantly increase the risk of depression among Chinese middle-aged and older adults. Based on a quasi-experimental analysis, they also assessed the benefit of the clean air policy and observed that reduction of PM_2.5 exposure was associated with a decrease of self-reported feelings associated with depression. Indeed, the high levels of PM_2.5 in China have continuously and rapidly been attenuated beginning in 2013 when more stringent air pollution control policies were put in effect. − However, these studies only focused on the impact on the overall elderly population and lacked information about subgroups or people living in different regions. Additionally, it is not clear what the benefits of continued reduction in pollution over time are in light of changes in other factors like socioeconomic and physiological status? In particular, evidence of longitudinal patterns of exposure–response that incorporate effects of the long-term historical levels of PM_2.5 exposure on mental health is still limited. These research gaps highlight the need for studies that assess this relationship in order to better understand the complexities between air pollution and mental health over time. A better understanding of these issues can help provide insights into the potential implications of air pollution control policies and their effectiveness in reducing the burden of disease over a long-term time frame.

This study aims to explore how changing air pollution levels over time in China affects mental health. We leveraged repeated measures from the China Health and Retirement Longitudinal Study (CHARLS) in order to assess associations between PM_2.5 exposure and self-reported depression-related emotions at long-time scale. We built on previous evidence, , based on shorter time frames of observations, by using additional data extending the investigation period to 2018 and leveraging both a longitudinal analysis including a distributed-lag model approach to assess long-term exposure effects, as well as a quasi-experimental framework with a difference-in-difference (DID) analysis. The overall aim is to quantify the effect of continuously decreasing PM_2.5 pollution due to control policies on mental health.

2. Materials and Methods

2.1. Study Population

This study focused on Chinese adults using a data set from the project of the China Health and Retirement Longitudinal Study (CHARLS) approved by the Ethics Review Committee of Peking University. These data are an open accessed data set and are publicly available from http://opendata.pku.edu.cn/. CHARLS is a national longitudinal survey mainly focused on middle-aged and older individuals. CHARLS employed a four-stage PPS (Probability Proportional to Size) sampling method, selecting 150 counties (stratified by region, urban/rural status, and GDP) and 450 villages/communities based on 2009 population data, with rigorous verification via statistical authorities to ensure unbiased national representation (https://charls.pku.edu.cn/gy/ybjs.htm). It recorded the status of human health including mental and physical risk factors and socioeconomic indicators using a well-established sample design. This project conducted 4 longitudinal survey waves in 2011 (wave 1), 2013 (wave 2), 2015 (wave 3), and 2018 (wave 4) involving 25,430 individuals in China. Details of this data set can be found elsewhere. The age ranges in the study align with common transitions in work, health, and retirement in a Chinese context. The transition to retirement may lead to lifestyle changes, shifts in mental health, and alterations in social support networks, all of which may make them more vulnerable to both air pollution exposure and chronic health challenges. In this study, we estimate the effects of air pollution on mental health using data from all 4 waves. Samples only in urban areas or nonurban areas within one city were also excluded. A total of 62,352 observations, 23,151 participants (Attrition: ∼9%) across 447 communities in 124 cities including both urban and nonurban areas remained after excluding participants with missing covariate information (Figure ).

1.

Map of study cities in China.

2.2. Exposure Assessment

This study used the concentration of annual PM_2.5 as representative of the level of air pollution exposure because it is the primary air pollutant during the survey waves especially before 2013 (wave 2) when the national Clean Air Policy (CAP) had not been formally implemented. We used raster exposure data estimated from satellite remote-sensing measurements due to the lack of nationwide monitoring data on PM_2.5 concentrations before 2013. In addition, we used the full coverage data to distinguish urban and nonurban areas based on urban boundaries to improve the variation and accuracy in the assessment of exposure levels within cities which were not considered in previous studies. ,

The database of PM_2.5 with a resolution of 0.01° used in this study was obtained from a global annual raster data, version V4.GL.03, modeled by the Atmospheric Composition Analysis Group at Washington University in St. Louis, USA (https://sites.wustl.edu/acag/datasets/surface-PM2-5/). This data set was produced by integrating aerosol optical depth (AOD) data from advanced satellite products including the NASA MODIS C6.1, MISR v23, MAIAC C6, and SeaWiFS satellite products, which offer finer spatial resolution, enhanced global coverage, and improved long-term stability. By relating them to surface PM_2.5 concentrations using geophysical relationships derived from GEOS-Chem chemical transport model simulations with updated algorithms, the annual mean geophysical PM_2.5 estimates showed strong agreement with ground-based monitors globally (R ² = 0.81; slope = 0.90). To further reduce residual bias, geographically weighted regression (GWR) was also applied to the geophysical estimates, significantly improving cross–validated performance (R ² = 0.90–0.92; slope = 0.90–0.97) compared to earlier global PM_2.5 assessments. The urban boundaries used in this study were generated using 30 m global artificial impervious area (GAIA) data and had a multitemporal data set in seven representative years (i.e., 1990,1995, 2000, 2005, 2010, 2015, and 2018) (https://iopscience.iop.org/article/10.1088/1748-9326/ab9be3). We applied the boundaries in 2018 for all of the waves to estimate the level of PM_2.5 and air temperature in urban and nonurban areas for each city, respectively.

2.3. Measurement of Mental Health and Covariates

Information on depressive symptoms obtained during the surveys at each wave was used to assess the status of mental health. The depressive symptoms in CHARLS were measured based on the 10-item Center for Epidemiologic Studies Depression scale (CESD10) which were included in the standard questionnaires of the project. Each question in CESD10 reflected the frequency of a specific type of negative mood using a 4-level score: 0 refers to rarely or none, 1 refers to some days, 2 refers to occasionally, and 3 refers to most of the time. As a result, the depression score from CESD10 ranges from 0 to 30 with higher scores indicating a higher frequency of depressive symptoms.

Additional information on covariates that can be risk factors for the outcome was also available. We obtained raster data to estimate air temperature with a resolution of 1 km × 1 km that can be access at National Earth System Science Data Center, National Science & Technology Infrastructure of China (http://www.geodata.cn). This data set was generated based on the global 0.5° climate data set released by CRU (Climatic Research Unit gridded) and the global high-resolution climate data set produced by WorldClim. It was downscaled in China through the Delta spatial downscaling scheme and verified through 496 independent meteorological observation points. We also calculated the mean temperature in urban and nonurban areas separately using the same boundaries as those above to increase the variations within cities. Information on socioeconomic indicators which were recorded along with CESD10 in CHARLS including age, sex, education, marital status, smoking, drinking, indoor temperature maintenance, cooking energy type, building type, rent payment, household tidiness, in-house telephone, and type of residence was also available. Details on the assessment of these covariates can be found elsewhere. , All analyses in this study adhered to the data usage guidelines.

2.4. Statistical Analysis

We incorporated multiple methods to comprehensively answer the research questions in this study. We conducted longitudinal analysis accounting for repeated measures with a distributed lag nonlinear model (DLNM) to assess the exposure–lag–response relationship between long-term exposure to PM_2.5 and depression. Then, we applied a two-way fixed effects difference-in-difference (DID) model to further quantify the effect of clean air policy on depression more accurately and directly and compare the results from different models. We did separate models for different waves to test the continuous effects of exposure and the clean air policy on mental health during a long-term period. These methods complement each other and collectively strengthen the study’s conclusions.

2.4.1. Longitudinal Analysis

In order to quantify the relationship between PM_2.5 exposure and the CESD10 score, we first used a mixed-effects model utilizing repeated observations per participant. The outcome was log-transformed to account for a skewed distribution. Random effect terms were used to account for two clustering effects on the level of community and the individual participant. Models adjusted for age, ambient temperature, sex, education, marital status, smoking, drinking, indoor temperature maintenance, cooking energy type, building type, rent payment, household tidiness, in-house telephone, and urban/nonurban residence. The observations without information on exposure, outcome or covariates were excluded. We incorporated a spline term with 3 degrees of freedom in the regression to model the nonlinear effects of ambient temperature. Separate models were fitted considering all 4 waves and only the first 3 waves, as shown in eq .

$$\log (\mathrm{CESD}10)={\beta }_0+{\beta }_1{\mathrm{PM}}_{2.5,i,t}+{\beta }_2{\mathrm{age}}_{i,t}+\mathit{\phi }{\mathit{\gamma }}_{i,t}+\delta ({\mathrm{community}}_i/i)$$ 1

where i is the subject index, t is the wave index, β₀ represents the intercept term, β₁ is the parameter coefficient for the annual mean PM_2.5, and β₂ is the corresponding parameter coefficient for age while γ is a vector for other adjusted variables described above and φ is a vector for the corresponding coefficients. δ represents the random effects of the records from the same community and participant. To examine whether the associations were sensitive to a potential interaction of PM_2.5 and age, we also applied models with the product term between the two as sensitivity analysis.

To further estimate the potential long-term relationship between exposure and mental health, we also introduced a distributed lag nonlinear model (DLNM) into the longitudinal analysis based on mix-effects model in eq by replacing annual mean PM_2.5 concentration with a crossbasis term. This approach was chosen to simultaneously account for potential nonlinear exposure–response relationships and delayed effects over time, critical features that traditional models fail to capture. We used a natural cubic spline with 3 degrees of freedom to model the exposure–response and examined two functions for the lag–response utilizing a natural cubic spline and strata for different lag-year groups, respectively. The DLNM’s flexibility in modeling both exposure and lag dimensions (via splines and stratified lags) provides a robust basis for identifying critical exposure windows and quantifying the shape of risk associations, offering insights into the long-term mental health burden of air pollution. We only used observations from the first 3 waves in this approach as the basic longitudinal model using annual means of exposure indicated potential for bias when including all 4 waves.

2.4.2. A Two-Way Fixed Effects Difference-in-Difference (DID) Model

To reduce the potential of unmeasured individual level confounding, we further applied a quasi-experimental difference-in-differences (DID) design. The DID design can effectively address both measured and unmeasured (time-invariant) confounding under a key “parallel-trends” assumption. We used a two-way fixed effects model to evaluate how mental health scores differentially changed among participants whose PM_2.5 decreased or did not decrease based on average wave 1 exposures (preperiod). We used information from waves 3 and 4 (both occurring after CAP was implemented) as the “postperiod” in separate analysis. We categorized participants as being in the treatment group if their PM_2.5 exposure decreased in wave 3 (or 4) compared to their exposure in wave 1 and in the control group if their exposure increased or remained the same. The DID estimator for the effect of reduction in PM_2.5 is estimated by the difference between the change in mental health score from the pre- to postperiod (difference 1) and the difference between the treatment and control group (difference 2). There were potentially time-varying covariates including air temperature, marital status, smoking, drinking, indoor temperature maintenance, cooking energy type, building type, rent payment, household tidiness, and in-house telephone across waves. Most of these covariates were balanced within treatment and control group based on wave 3, several of them were not balanced based on wave 4 according to the results from the balance test (Table S1). Therefore, we also controlled all these time-varying covariates in the DID models. A fixed effect for wave and individual participant ID was included in the model as well. In summary, we fitted a two-way fixed effect model as follows:

$$\begin{array}{l}\log (\mathrm{CESD}10)={\beta }_0+{\beta }_1{\mathrm{PMgroup}}_{,i,t}+{\beta }_{2}t+{\beta }_3({\mathrm{PMgroup}}_{,i,t}\times t)+ai+\mathit{\phi }{\mathit{\gamma }}_{i,t}\\ \mathrm{PMgroup}=0\mathrm{for}{C}_{i,t}\ge C_{i,1}\mathrm{and}1\mathrm{for}{C}_{i,t}<C_{i,1}\end{array}$$ 2

where PMgroup denotes the treatment variable as defined by the change in concentration of PM_2.5 (C _{i ,t} ) between later waves and wave 1, t is an indicator for wave (or post- and preperiod), γ are time-varying variables and φ are their corresponding coefficients, and a are the coefficients of the participant index. The coefficient β₃ for the treatment and wave product term corresponds to the DID estimator described above. We also estimated the impact of PM_2.5 on mental health among subgroups to indicate the variation among them using this model. We focused on sex and the type of residence, considering the different exposure levels they might have. Also, as sensitivity analysis, we varied the definition of the treatment group to include only those that experienced at least 1, 3, and 5 μg/m³ decrease in PM_2.5 as the treatment group, compared to those who saw no change or increase in exposure. Observations from wave 2 were assessed in the same year as the implementation of CAP, so we tried the parallel trend tests using wave 1 + 2 as “preperiod” according to Riveros-Gavilanes (2023) and the results supporting the parallel trends assumption. Exposure concentrations did appear to decrease compared to wave 1 beginning with wave 2 however, so we repeated analysis using wave 2 data as the “postperiod” for the sake of comparison.

All statistical analyses were performed in R studio (R version 4.2.3) using packages “lme4” and “dlnm”.

3. Results

3.1. Descriptive Summary

This study involved 62,352 observations from 23,151 participants with complete exposure outcome and covariate data across four waves. The mean age across wave 1 was 58.47 years (standard deviation [SD] 10.17 years; Table ). In wave 1, wave 2, wave 3 and wave 4, the mean CESD10 score was 8.41 (6.33), 7.84 (5.78), 7.82 (6.33) and 8.44 (6.45) and the corresponding concentration of PM_2.5 was 59.40 (22.16), 55.18 (23.09), 47.29 (16.83) and 39.35 (14.78) μg/m³, respectively. The distributions of CESD10 and PM_2.5 are shown in Figure . PM_2.5 concentrations generally declined steadily with time, while the value of CESD10 did not have a monotonic trend over time. Attrition analysis compared PM_2.5 exposure and CESD10 scores between participants retained across all four waves (complete cases) and those lost to follow-up. As shown in Figure S1, no significant differences in PM_2.5 exposure were observed at any wave (all p > 0.10). For CESD10 scores, variation tests indicated no differences in waves 1–3 (all p > 0.1), but a statistically significant difference was detected at wave 4 (mean Δ = 0.63, p = 0.046, Cohen’s d = 0.15), suggesting slightly elevated depressive symptoms among participants retained.

1. Longitudinal Variables of the Subjects Studied.

Variables	Subgroup	Wave 1	Wave 2	Wave 3	Wave 4
		No. of participants (percentage of the total)
Age group	<30	17 (0.10%)	16 (0.09%)	19 (0.09%)	15 (0.08%)
	30–40	93 (0.53%)	66 (0.36%)	105 (0.50%)	38 (0.19%)
	40–50	3941 (22.62%)	3503 (19.04%)	4370 (20.95%)	2424 (12.35%)
	50–60	6148 (35.29%)	6351 (34.53%)	6600 (31.64%)	6424 (32.74%)
	60–70	4556 (26.16%)	5279 (28.70%)	6137 (29.42%)	6455 (32.90%)
	70–80	2110 (12.11%)	2451 (13.32%)	2789 (13.37%)	3163 (16.12%)
	80–90	517 (2.97%)	671 (3.65%)	769 (3.69%)	1007 (5.13%)
	90–110	37 (0.21%)	58 (0.32%)	72 (0.35%)	94 (0.48%)
Building type	One	10831 (60.53%)	6277 (59.49%)	6277 (59.49%)	9547 (60.33%)
	Two	7063 (39.47%)	4275 (40.51%)	4275 (40.51%)	6277 (39.67%)
Cooking energy type	Clean	5710 (31.91%)	4199 (39.83%)	4199 (39.83%)	10523 (50.09%)
	Unclean	12184 (68.09%)	6342 (60.17%)	6342 (60.17%)	10486 (49.91%)
Drinking	Frequent	826 (5.08%)	1169 (6.41%)	1580 (7.63%)	1410 (7.19%)
	Never	11718 (72.12%)	11982 (65.71%)	13400 (64.70%)	13023 (66.40%)
	Often	1720 (10.59%)	2659 (14.58%)	2663 (12.86%)	2474 (12.61%)
	Rare	1984 (12.21%)	2424 (13.29%)	3067 (14.81%)	2705 (13.79%)
Indoor temperature maintenance	Very cold	14912 (83.48%)	8918 (87.27%)	8918 (87.27%)	16439 (84.56%)
	Cold	650 (3.64%)	390 (3.82%)	390 (3.82%)	466 (2.40%)
	Bearable	1836 (10.28%)	837 (8.19%)	837 (8.19%)	2324 (11.95%)
	Hot	108 (0.60%)	20 (0.20%)	20 (0.20%)	18 (0.09%)
	Very hot	356 (1.99%)	54 (0.53%)	54 (0.53%)	193 (0.99%)
Married	No	2236 (12.72%)	2387 (12.89%)	2702 (12.89%)	2906 (14.76%)
	Yes	15346 (87.28%)	16127 (87.11%)	18257 (87.11%)	16780 (85.24%)
Education	Less than lower secondary	21885 (86.21%)	21885 (86.21%)	21885 (86.21%)	21885 (86.21%)
	Upper secondary & Vocational training	716 (2.82%)	716 (2.82%)	716 (2.82%)	716 (2.82%)
	Tertiary	2785 (10.97%)	2785 (10.97%)	2785 (10.97%)	2785 (10.97%)
Sex	Man	12259 (48.22%)	12259 (48.22%)	12259 (48.22%)	12259 (48.22%)
	Woman	13163 (51.78%)	13163 (51.78%)	13163 (51.78%)	13163 (51.78%)
Rent payment for residence	No	17297 (97.17%)	9684 (96.43%)	9684 (96.43%)	9684 (96.43%)
	Yes	504 (2.83%)	358 (3.57%)	358 (3.57%)	358 (3.57%)
Smoking	No	10552 (60.49%)	10552 (57.62%)	11814 (56.81%)	11357 (57.93%)
	Yes	6893 (39.51%)	7760 (42.38%)	8981 (43.19%)	8248 (42.07%)
In-house telephone	No	9187 (51.33%)	6431 (60.96%)	6431 (60.96%)	17017 (80.47%)
	Yes	8710 (48.67%)	4119 (39.04%)	4119 (39.04%)	4129 (19.53%)
Tidiness	Poor	6921 (38.74%)	3352 (32.77%)	3352 (32.77%)	6135 (31.51%)
	Fair	1600 (8.96%)	915 (8.94%)	915 (8.94%)	2517 (12.93%)
	Clear	4628 (25.90%)	2902 (28.37%)	2902 (28.37%)	5136 (26.38%)
	Very clear	1156 (6.74%)	758 (7.41%)	758 (7.41%)	1304 (6.70%)
	Excellent	3562 (19.94%)	2303 (22.51%)	2303 (22.51%)	4377 (22.48%)
Place of residence	Rural	14123 (55.54%)	14123 (55.54%)	14123 (55.54%)	14123 (55.54%)
	Urban	11307 (44.46%)	11307 (44.46%)	11307 (44.46%)	11307 (44.46%)
PM2.5 (μg/m3)	Mean	59.40	55.18	47.29	39.35
	Standard Deviation	22.16	23.09	16.83	14.78
Age (years)	Mean	58.47	59.38	59.08	61.43
	Standard Deviation	10.17	10.28	10.75	10.41
Temperature (°C)	Mean	13.75	14.20	14.45	14.58
	Standard Deviation	5.02	5.24	5.02	5.16
CESD10 (Depression score)	Mean	8.41	7.84	7.82	8.44
	Standard Deviation	6.33	5.78	6.33	6.45

2.

Boxplots of the distribution of depression score and PM_2.5 by CHARLS waves.

3.2. Longitudinal Association between PM_2.5 and Depression Score

Based on 46125 observations from the first 3 waves, a significantly positive association between PM_2.5 and the CESD10 score was observed (Figure ). A 10 μg/m³ increase in PM_2.5 concentrations was associated with a 0.86% increase (95% confidence interval [CI]: 0.10%, 1.64%) in CESD10. Results from a model including a product term between PM_2.5 and age are summarized in Table S1. The product term was statistically significant and negative in direction, indicating that the association between PM_2.5 and mental health score decreased with increasing age. Effect estimates from the model with all 4waves samples were much weaker and negative compared to the model from the 3-wave model with CIs including the null (−0.11%, [CI]: −0.76%, 0.54%). Additionally, effect estimates from all longitudinal models indicated that the depression risk increased with age (Table S2).

3.

Estimated associations between the depression score and PM_2.5 concentration in longitudinal models. The dots indicate the estimated percentage change in CESD10, and the lines indicate the confidence intervals.

3.3. Lag–Response Effect of Exposure

The exposure–lag–response relationships between PM_2.5 exposure and depression score were nonlinear for both exposure–response and lag–response dimensions illustrated in the comprehensive three-dimensional (3D) images (Figure S2). Generally, there was a positive relationship between PM_2.5 exposure and the CESD10 score during lag-years 1–4, then the relationship became negative for lag-years 6–9 with the relationship appearing strongest at lag-year 3 based on results with a natural spline function for the lag–response (Figure a). In addition, the cumulative effects of PM_2.5 were at their maximum for a five-year window (lag0–4) with 0.30% (0.09%, 0.51%), 21.49% (6.35%, 38.80%) and 65.73% (18.48%, 130.10%) increase in the CESD10 score for exposures of 6 μg/m³, 16 μg/m³ and 36 μg/m³, respectively, compared to the lowest observed exposure of 5.59 μg/m³ over the 5-year lag period (Figure S3a). For the highest concentration of 76 μg/m³, the cumulative effect was the highest for a six-year window (lag0–5) with a 103.76% (33.69%, 210.55%) increase in the CESD10 score. Cumulative effects for larger windows decreased given the negative association for lag-years 6–9. The relationship between PM_2.5 exposure and the CESD10 score and changes in the cumulative effects based on results with strata function showed similar trends (Figure b, Figure S3b). Exposure–response trends indicated a more monotonic relationship but a plateauing and even diminishing effect on higher exposures at lag-years 3 and 4 (Figure ).

4.

Lag–response effects of PM_2.5 concentration based on lag 0 to lag 9 years on the depression score (a, natural splines-based models; b, strata-based models) and exposure–response effects at specific lag years (c, natural splines-based models; d, strata-based models).

3.4. Effect of Clean Air Policy on Depression

The DID model based on wave 1 and wave 3 or wave 4 suggested that the reduction in PM_2.5 concentration after CAP (by wave 3) was associated with decreased risk of depression. Figure shows that the decrease in PM_2.5 resulted in a relative reduction in the score of 19.51% ([CI]: 11.57%, 26.73%) and 28.18% ([CI]: 5.87%, 45.2%) in wave 3 and wave 4, respectively, compared to no decrease or increase in exposure in the same period. The reduction effect (28.18%, [CI]: 5.87%, 45.2%) became stronger when the concentration of PM_2.5 was further decreased which was indicated in the result from changes between wave1 and wave4, though CIs were much wider owing to the smaller number of participants in wave 4 (Figure ). Overall changes in the CESD10 score over time between treatment and control groups indicated a decrease for both groups between waves 1 and 3, but a more pronounced increase for the control group compared to the treatment group between waves 1 and 4 (Figure S4). Results from analysis using wave 2 as the post period indicated no change in the outcome associated with reduction in exposure (Figure ). All models showed strong explanatory power (R ² > 0.79) and statistical significance (p < 0.001). Adjusted R ² values (0.405–0.440) reflect robust performance after accounting for model complexity. Total samples ranged from 19,119 to 22,233, with treatment groups comprising 83–99% of observations. While control groups were smaller (especially in Wave1–4), our DID design’s reliance on within-group changes mitigates potential bias from this imbalance. The details of the models were indicated in Table S3.

5.

Estimated excess risk of increased PM_2.5 concentration on the depression score in the DID models. The dots indicate the estimated percentage change in CESD10; the lines indicate the confidence intervals.

In subgroup analyses using the same DID models for waves 1 and 3, results suggested that the effect of exposure reduction was stronger in males compared with females and weaker for rural areas compared with urban areas (Figure S5). In sensitivity analysis with alternative definitions of the treatment groups, we found consistent and comparable levels of reduction in risk (Figure S6).

4. Discussion

In this study, we examined the relationship between longer-term exposure to PM_2.5 and depressive symptoms, as well as the effect of continued reduction in exposure using a quasi-experimental design based on the CHARLS cohort. Our findings suggest an association between air pollution concentrations and an increase in the risk of depressive symptoms. The relationship, however, was not consistent when examining data from three waves of observations compared to four waves in a longitudinal repeated measures analysis. Our results also indicated that the lag–response for the potential effects of exposures may involve a multiyear influential window of exposure. Additionally, results from a DID analysis suggested a reduced risk of depression associated with decreasing PM_2.5 concentrations after CAP, providing further evidence of the efficiency of air pollution controls in improving health in China. Results from this analysis remained consistent when looking at a longer time frame of observations compared to previous studies. ,

Both animal studies and human epidemiologic data suggest that PM_2.5 can enter brain tissue when exposure to high levels of air pollutants. , Containing a variety of toxic elements, PM can affect the function of brain through systemic inflammation and neuroinflammation, , oxidative stress and neuronal damage, neuroendocrine and neurotransmitter dysregulation and epigenetic modifications. − Furthermore, previous studies also showed that low air quality may limit or reduce physical activities, since heavier air pollution could make a neighborhood environment less appealing for outdoor recreation. Research indicated that less outdoor physical activities might be directly negative to mental health status because of lack of chance to body relaxing and pressure releasing or decrease the protectiveness against to inflammation, oxidative stress and so on.

Previous studies have examined the associations between air pollution and mental health among people living in China; however, they were mostly based on cross-sectional studies and in individual cities, thus neglecting heterogeneity in the population. − Fewer studies have focused on the general population at a national level and suggested an epidemiological link between environmental changes and human mental health. , For example, based on 3 waves of surveys (2011, 2013 and 2015) in CHARLS, a recent study conducted longitudinal analysis and reported a 3.63% (95% CI: 2.00%, 5.27%) excess risk of depression with a 10 μg/m³ increase in PM_2.5 concentration. This result is comparable with our analysis when accounting for exposure and age interaction (Table S2). However, we additionally found that the estimate was sensitive to the addition of more observations from a later fourth wave in CHARLS, indicating a lack of robustness and potential for bias in a traditional regression setting. The overall concentration of PM_2.5 continued to significantly decrease in wave 4 as the national action plan of air pollution control was carried out from 2013 to 2017 for the retained participants. At the same time the possibility of a marginally elevated depression score among retained participants may also introduce minor bias in our 4 wave longitudinal estimates. These trends in the outcome are likely due to other time-varying factors including potential changes in natural and artificial (urban) environment like green space which have been shown to be associated with mental health and may be impacted by climate change and urbanization. , The reduction or fragmentation of green spaces may lead to a decrease in outdoor activities for people, which can, in turn, impact their mental health.

In this study, we explored the potential effects of exposure over longer time periods by identifying potentially relevant windows for long-term PM_2.5 exposure on mental health occurring in the first 2–5 years with largest positive effects at lag-year 3 via DLNMs. These results potentially suggest a longer-term framework for the potential effects of air pollution may provide new insights for policy makers in terms of actions for air quality control aiming to benefit human health. However, associations with exposure were negative after lag-year 6, which might be related to a compensatory mechanism or behavioral adaptations. This compensatory mechanism has been found in a few studies explained for atherogenic risk associated with ambient air pollution. − One of these previous studies suggested that adults with middle atherogenic risk exposed chronically to ambient air pollution might show transitory cardiovascular clinical improvement. Biological studies found activation of the Nrf2/HO-1 pathway by curcumin could inhibit oxidative stress in human nasal fibroblasts and neuroinflammation exposed to urban particulate matter. − This protection mechanism could be activated during a longer exposure. The similar patterns found in mental health in our study could be the reflection of these hypothesis. In addition, these negative effects may reflect population-level behavioral adaptations (e.g., reduced outdoor activity, air purifier adoption) after long-term and high-level exposure.

A few previous studies have examined the lag–response effects of air pollutants on mental disorders. ,, However, these studies mainly focused on short-term exposure in local regions in China. For example, Duan et al. (2018), Gao et al. (2017), and Ji et al. (2021) all found lag effects of daily PM on schizophrenia related outcomes for both inland and coastal Chinese cities. However, the lag–response relationships varied by locations, which may be driven by differences in susceptibility among population or by differences in the levels of pollution among cities. Our study further explored the lagged association of PM_2.5 with mental health at an annual level and indicated that air pollution effects may take place over up to 5–6-year periods, though the shape of lag–response estimated in our models is heavily dependent on a-priori assumptions. Additionally, there was some evidence that nonlinearity in exposure–response with steeper risk increases was observed at lower exposure ranges and plateauing of effects were seen at higher exposures though data were scarcer at higher levels of exposure.

We further leveraged a DID approach with a two-way fixed effects model to explore the effect of decreased pollution on the depression score. A significant reduction in risk of depression score was observed for wave 3 and persisted in wave 4, consistent with a continuing effect of gradually decreasing exposure concentrations. In fact, the association comparing treatment and control groups between waves 4 and 1 was even greater than comparing waves 3 and wave 1. This is consistent with the fact that the CAP measures were not concluded in their entirety until 2017. − This result is also consistent with continued benefits from improved air quality which was been found previously. The two-way fixed effects DID model can control unmeasured time-invariant confounders compared to the mixed-effects model, thus lending more confidence in the assessment of excess risk associated with exposure. Unlike the longitudinal repeated measures design, the two-way fixed effects model can also account for time trends in the outcome if they are similar between treatment and control groups had the exposure been absent from both. Assuming this holds, the DID approach is less prone to potential bias by secular treads that is likely driving the inconsistency in findings from the longitudinal repeated measures analysis between models including three and four waves. In general, the depression score increased a little in wave 4 compared to the first three waves, which could be caused by other time-varying factors, as mentioned above. If those factors are differential across treatment and control groups, however, the DID approach for wave 4 may also suffer from bias. We also conducted DID models for samples including only the first 2 waves but did not find significant relationship between PM_2.5 and depression score. This may be because PM_2.5 did not sufficiently decrease between waves 1 and 2. CAP was implemented until the same year as wave 2 observations, and even though annual mean concentrations between wave 1 (2011) and wave 2 (2013) show a decreasing trend, the difference was not as big as that with later waves. This could also be the result of long-term exposure effects of PM_2.5 which would require a longer time frame to observe an effect.

Subgroup DID model analyses suggested that males exhibited a higher reduction in depressive symptoms associated with reduced air pollution than females, which was consistent with other studies focused on mental disorders. ,,, It is plausible that males could spend more time outdoors experiencing higher exposure which could make them more sensitive to the change in air quality. Animal studies also suggest that differences in neurotoxicity between males and females exist due to differences in the expression of enzymes, which may lead to the varied effects of PM_2.5 exposure. Furthermore, we found that people living in urban areas were more sensitive to the reduction in PM_2.5 concentrations than those living in rural areas, although in our study the reduction of PM_2.5 concentrations appeared higher in rural areas (Figure S7). This could be attributed to multiple factors: (1) urban areas typically experience higher baseline PM_2.5 exposure, meaning marginal improvements yield more pronounced health benefits; (2) urban PM_2.5 often contains more toxic components (e.g., traffic-related metals) versus rural PM_2.5 (e.g., natural dust) which could explain why urban populations may experience stronger health benefits from PM_2.5 reductions; , (3) urban populations may have greater awareness and access to air quality data which may amplify perceived and actual health responses. These findings highlight that PM_2.5 mitigation policies could generate disproportionate health gains causing inequality within cities from the perspective of benefits from air pollution control, which need to pay more attention to.

Our study has several strengths including extended use of longitudinal data to understand the relationship between PM_2.5 and mental health over time, as well as use of a flexible distributed-lag-approach to further assess the longer-term effects of PM_2.5 exposure on mental health. We also leveraged a quasi-experimental design to evaluate the effectiveness of air policy controls and examined robustness of models used for epidemiologic studies of air pollution and of mental health.

However, our study also has some important limitations. First, the raster exposure data used may be less accurate than those from monitoring at the level of concentration. We descaled the data to urban vs nonurban level to increase the exposure accuracy, and this data set was assessed having a strong linear relationship with that from monitoring stations, , though it may still be less accurate than that at the individual level. This type of exposure measurement error is typically expected to be nondifferential with respect to the outcome and lead to bias toward the null. Also, the study relied on 2018 urban boundary data to distinguish between urban and nonurban areas which may introduce some inaccuracies in the definition of urban and nonurban areas given the ongoing urbanization process during the long-term analysis. Analogously, the status of mental health was evaluated using self-reported questionnaire data, which may also lead to measurement error. If this error is also nondifferential, it would have likely led to inflated standard errors and, by extension, less power. Besides, our approach to measuring depressive symptoms may not fully capture clinical-level mental health outcomes and might blur the lines between different mental health conditions, making it difficult to draw precise conclusions about mental health or specific mental disease, which emphasizes the need for nuanced research definitions to better target specific aspects of mental health. Furthermore, in the DID model used in our study, while controlling for unmeasured time invariant confounding by design, unmeasured confounding that changes over time may nevertheless lead to bias, such as rapid urbanization, socioeconomic stressors, and regional disparities in healthcare access. , Though we did adjust for some time-varying confounders, this may still be an issue. Finally, in the DLNM analysis we used a priori defined functions in the crossbasis definition based on a repeated measures design that could be less robust to overfitting and model misspecification compared to penalized DLNMs which is data-driven approach.

5. Conclusions

In summary, based on a national level study of Chinese residents, we examined the potential long-term effects of PM_2.5 exposure on depressive symptoms by characterizing the exposure–lag–response, as well as assessing the potential effect of air pollution control on the outcome via a quasi-experimental design. Our results supported the public health interventions for better mental health in China and provide insights on the potential benefit of continued air pollution exposure reductions with respect to mental health.

Data will be made available on request.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/envhealth.5c00033.

• Figures S1–S7 and Tables S1–S3 (PDF)

Xiuling Zhao: Writing– review and editing, Writing– original draft, Visualization, Methodology, Formal analysis, Data curation, Conceptualization. Weiqi Zhou: Writing– review and editing, Supervision. Andreas M. Neophytou: Conceptualization, Writing– review and editing, Validation, Supervision.

Thanks to the China Center for Economic Research, National School of Development, Peking University, for providing the CHARLS data. This study was supported by the National Natural Science Foundation of China (Grant No. 42225104), National Institute of Environmental Health Sciences (NIEHS) of the United States (Grant No. R00ES027511), National Natural Science Foundation of China (Grant No. 42201303) and National Natural Science Foundation of China (Grant No. U21A2010). Thanks to support from the Foundamental Research Youth Team Project of the Chinese Academy of Science.

The authors declare no competing financial interest.