Air pollution and cognitive function: the potential protective effect of physical activity

Abstract

Physical activity (PA) may mitigate pollution‐related cognitive decline while concurrently increasing individuals’ exposure to harmful pollutants. Data were obtained from the 2018 China Health and Retirement Longitudinal Study (CHARLS), comprising 17,734 participants aged 45 years or older. Information regarding particulate matter (PM₁, PM_2.5, and PM₁₀), nitrogen dioxide (NO₂), sulfur dioxide (SO₂) and ozone (O₃) was sourced from the China High Air Pollutants (CHAP) database. The assessment of cognitive function was carried out following the approach utilized in the Health and Retirement Study (HRS). The impact of air pollutants on cognitive function was estimated using the two-stage least squares method, with the ventilation coefficient serving as an instrumental variable. The results indicated that all air pollutants were significantly associated with cognitive function. An interquartile range (IQR) increase in PM₁, PM_2.5, PM₁₀, NO₂, SO₂, and O₃ corresponded to decreases in cognitive function of −0.45 (95% CI: −0.76, −0.13), −0.43 (95% CI: −0.74, −0.13), −0.53 (95% CI: −0.90, −0.16), −0.57 (95% CI: −0.96, −0.17), −0.47 (95% CI: −0.80, −0.14), and −1.06 (95% CI: −1.81, −0.31), respectively. Further stratified analyses revealed that higher levels of PA significantly moderated the association between air pollution and cognitive function, suggesting a potential protective effect. The level of PA was found to modify this association of pollution and cognitive function, with higher PA levels seemingly alleviating the adverse cognitive effects of air pollution. These findings underscore the importance of policies that simultaneously target pollution reduction and promote PA to safeguard cognitive health.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-28614-1.

Introduction

Dementia has become a significant global health challenge, with its increasing disease burden posing substantial socio-economic and public health concerns. By 2050, the number of individuals affected by dementia worldwide is projected to reach 152 million¹. China, the country with the highest dementia prevalence, faces an intensified burden due to its accelerated aging process². Compared to the global average, China has a higher proportion of elderly individuals, a faster rate of aging, and a greater socio-economic impact from dementia^2,3. The demographic transition has further exacerbated this challenge. By 2050, China’s population aged 65 and above is expected to reach 365 million⁴, with dementia prevalence potentially tripling compared to 2015⁵. Given the absence of an effective cure for dementia, primary prevention remains a critical strategy to mitigate incidence rates and slow disease progression. Early interventions—including lifestyle modifications, cardiovascular risk management, and cognitive training—may reduce dementia risk and alleviate the anticipated public health burden⁶.

Cognitive function is a fundamental aspect of mental health and quality of life, increasingly recognized as being influenced by environmental and behavioral factors. Among these, air pollution has emerged as a significant global health threat, contributing not only to respiratory and cardiovascular diseases and depression but also to neurodegenerative disorders such as dementia⁷. Growing evidence suggests that exposure to air pollutants may impair cognitive performance through mechanisms including neuroinflammation, oxidative stress, and structural brain alterations^8–11, Elevated air pollution concentrations are associated with a greater accumulation of amyloid proteins in the brain, potentially increasing the risk of Alzheimer’s disease¹².

Studies using data from the China Health and Retirement Longitudinal Study (CHARLS) have established the adverse effects of elevated concentrations of particulate matter (PM₁, PM_2.5, PM₁₀) on cognitive function¹³, with this evidence corroborated by numerous other observational studies, causal inference studies, and meta-analyses^14–19. However, the relationships between nitrogen dioxide (NO₂), sulfur dioxide (SO₂), ozone (O₃), and cognitive function remain poorly understood. For example, a study examining populations in Northwest China found no association between exposure to NO₂ and SO₂ and cognitive function, while a nationwide survey reported no association between exposure to O₃ and NO₂ and cognitive function¹². In contrast, studies conducted in the United States and South Korea demonstrated that O₃ exposure either exhibits a purported “beneficial effect” or has no significant association with cognitive function, whereas SO₂ exposure showed a positive association in certain studies^20,21. These inconsistencies likely arise from inherent endogeneity issues prevalent in observational studies, where unobserved factors—such as local economic conditions, occupational activities, or subtle differences in health behaviors—simultaneously correlate with both cognitive function and air pollution exposure. Omitting these variables introduces potential bias in estimation, resulting in results that are difficult to replicate. Thus, more robust causal inference methods are needed to clarify the precise relationship between air pollution and cognitive function.

Current causal analyses of air pollution’s effects are significantly limited. Most observational studies are affected by endogeneity issues arising from unobserved factors that influence both individual neurological and psychological health as well as air pollution exposure^22,23. Therefore, more rigorous methodological strategies are necessary to establish a causal link between air pollution and cognitive function, minimizing potential biases from confounding factors and ensuring greater validity of the findings. As of now, no studies employing causal frameworks have evaluated the association between air pollution and cognitive function.

Physical activity (PA) is a modifiable lifestyle factor that supports cognitive health by enhancing neural plasticity and reducing inflammation, thereby mitigating cognitive impairment^24,25. However, engaging in high-level exercise in polluted environments increases exposure to air pollutants, raising concerns about whether its neuroprotective effects can counteract the adverse impact of pollution on cognitive function. This interaction between PA and air pollution is complex, as PA in polluted conditions exerts dual effects: it may alleviate pollution-related cognitive decline while simultaneously increasing individuals’ inhalation of harmful pollutants. The conflicting physiological mechanisms underlying these effects have been widely debated in research on physical health outcomes, yet their influence on cognitive function remains poorly understood^26,27. Addressing this uncertainty is crucial for formulating public health guidelines that effectively balance the cognitive benefits of PA with environmental health risks.

China faces the dual challenge of deteriorating air quality and a rapidly aging population, providing a crucial foundation for this study^28,29. While stringent air pollution control measures have improved conditions in some regions, global pollutant concentrations continue to rise, posing long-term risks to the cognitive health of middle-aged and elderly individuals^28,30. Using a nationally representative sample, this study employs exogenous meteorological variations as instrumental variables to systematically assess the causal impact of air pollution on cognitive function. Notably, it is the first to conduct an in-depth analysis of PA level as a moderating factor in this relationship, addressing a key gap in existing research. These findings not only illuminate the intrinsic link between environmental factors and cognitive health but also offer a scientific basis for targeted public health interventions, ensuring cognitive preservation in aging populations under environmental stress.

Methods

Study population

The study population was derived from the China Health and Retirement Longitudinal Study (CHARLS), encompassing 450 villages or communities across 28 of China’s 31 provinces. Employing a multi-stage random sampling method with unequal probabilities, the study began in 2011–2012 and enrolled 17,708 participants from 10,257 households. Follow-up waves of data collection occurred in 2013, 2015, and 2018. Ethical approval was obtained from Peking University’s Biomedical Ethics Review Committee (IRB00001052–11015), and informed consent was secured from all participants. Data from the 2018 wave was selected for this analysis due to its comprehensive information on PA across participants. Of the initial 19,816 participants surveyed, exclusions were made for 362 individuals aged below 45 years, 1,700 missing cognition assessments, 22 lacking PA data, and 398 missing other covariates. Ultimately, the study included 17,734 participants aged 45 years and older (Fig. 1).

Fig. 1.

Flowchart of the study population.

Air pollution assessment

Ambient air pollution was evaluated using data on particulate matter (PM) with diameters ≤ 1 µm (PM1), ≤ 2.5 µm (PM2.5), ≤ 10 µm (PM10), as well as ozone (O₃), sulfur dioxide (SO₂), and nitrogen dioxide (NO₂). These measurements were sourced from the China High Air Pollutants (CHAP) dataset (https://weijing-rs.github.io/product.html) and estimated using space–time extremely randomized trees models. The models utilized inputs such as weather conditions, land use patterns, emission inventories, population distribution, and other spatial–temporal indicators. Predicted daily concentrations for PM1, PM2.5, PM10, and O₃ were mapped at a 1 km × 1 km resolution, while estimates for NO₂ and SO₂ were produced at a 10 km × 10 km scale. Detailed descriptions of the modeling approaches are available in earlier publications^31–36. Model performance was strong, with cross-validation root mean square errors between 0.80 and 0.92.

To preserve confidentiality, participant locations in the CHARLS database were anonymized by geocoding them at the city level. Air pollution exposure was therefore averaged at the city scale to match the survey data. Consistent with established methodologies^37,38, the primary exposure metric was the one-year average pollutant concentration preceding the survey year. Sensitivity analyses further examined the robustness of findings by considering average pollutant concentrations over the three years prior to the survey year.

Cognitive function

Cognitive function was evaluated based on the Health and Retirement Study (HRS) methodology³⁹. Participants underwent in-person assessments encompassing four cognitive areas: orientation, memory, drawing, and calculation, Both orientation and calculation were measured by utilizing the Telephone Interview of Cognitive Status (TICS). For orientation, participants identified the year, month, date, day, and season, with a maximum score of 5. Calculation involved subtracting 7 from 100 serially, up to five times, with each correct response earning 1 point, and the total score of the orientation was 5 points. Memory was tested through immediate and delayed recall of 10 randomly selected words. Immediate recall was scored by the number of words remembered instantly, and the delayed recall was assessed after completing the depression scale, drawing, and calculation test. Participants were awarded 1 point for each accurately remembered word, with a total possible score of 20 for the memory assessment. Participants were instructed to replicate a figure consisting of two overlapping pentagons, earning 1 point for an accurate representation. The overall cognitive score included components for orientation (5 points), calculation (5 points), memory (20 points), and drawing (1 point), resulting in a maximum possible score of 31⁴⁰.

Physical activity

Data on PA were gathered using a standardized questionnaire adapted from the International Physical Activity Questionnaire (IPAQ)⁴¹. Respondent reported the frequency (days per week), duration, and intensity (vigorous, moderate, or light) of activities lasting at least 10 min during a typical week. Examples were provided to clarify intensity levels: light (e.g., walking), moderate (e.g., mopping), and vigorous (e.g., hoeing). Daily PA duration was classified into five categories: < 10 min, < 30 min, < 120 min, < 240 min, and ≥ 240 min. Due to the lack of exact PA durations in some observations, Durations under 10 min were recorded as 0, and those ≥ 240 min were capped at 240 min, following established methods^42,43. Weekly PA time was calculated by multiplying the reported number of active days per week by the median duration of activity per day. PA intensity was quantified using metabolic equivalents (MET)⁴⁴, with values assigned as follows: 8.0 for vigorous, 4.0 for moderate, and 3.3 for light activities, based on previous studies^37,45. Total PA volume was calculated by summing the products of MET values and corresponding weekly durations, using the formula: PA volume = (8.0 × vigorous PA time) + (4.0 × moderate PA time) + (3.3 × light PA time). Eventually, PA levels were categorized as either low or high according to the median MET value, in line with established guidelines^46,47.

Covariates

Our selection of covariates was guided by the principle of confounding adjustment based on prior subject-matter knowledge and established literature. We specifically aimed to include variables that are known or plausibly associated with both air pollution exposure and cognitive function, the omission of which would introduce confounding bias^48–55. Below, we outline the primary rationale for including each group of covariates:

1. Sociodemographic Factors (age, gender, living area, current married, education level, regional category, family residence): These are fundamental determinants of socioeconomic status (SES) and cognitive reserve. SES influences an individual’s place of residence and, consequently, their exposure to air pollution, while also being strongly correlated with cognitive performance and the risk of decline.

2. Lifestyle habits and health (smoking status, drink status, cooking fuel use, number of chronic diseases): Smoking and alcohol consumption are associated with systemic inflammation and oxidative stress, which are also potential pathways for pollution-induced neurotoxicity. Cooking fuel type is a direct source of indoor air pollution and could confound the effect of ambient air pollution. Overall health burden can directly impact cognitive function. Furthermore, health status may influence physical mobility and time spent outdoors, thereby modifying personal exposure to ambient pollution.

3. Environmental Confounders (average temperature, relative humidity, Normalized Difference Vegetation Index (NDVI)): These meteorological and geographical factors directly influence the concentration and dispersion of air pollutants. They may also correlate with regional economic development and have direct or indirect effects on cognitive health.

Education level was divided into three groups: primary school or below, high school, and college or above. Chronic diseases were identified based on self-reported physician diagnoses of any of the following conditions: hypertension, dyslipidemia, diabetes, cancer, lung disease, liver disease, heart disease, stroke, kidney disease, digestive disorders, psychiatric conditions, memory-related disorders, arthritis, or asthma, and the number of chronic diseases was devided into three groups: none, one chronic disease, more than one chronic diseases. Smoking status was categorized into three groups based on self-reported past and current smoking behavior: never smoking, currently smoking, and ever smoking. Drinking status was divided into three categories based on self-reported drinking frequency: never drinking, drinking less than once a month, drinking more than once a month. According to the geographical location of each province, the regional category is divided into three categories: East, Midland, and West. At the urban scale, the annual average temperature and relative humidity were summarized and analyzed using data sourced from the Fifth Generation European Reanalysis (ERA5) gridded meteorological dataset⁵⁶. The NDVI dataset is derived from the Google Earth Engine (GEE) remote sensing cloud computing platform. The dataset is accessible at http://www.nesdc.org.cn.

Statistical methods

Initially, we employed an ordinary least squares (OLS) model to examine the relationship between air pollution and cognitive function, formulated as follows:

1

Among these variables, represents cognitive function, while denotes the intercept term. captures the effect size corresponding to air pollution concentration. X is a covariate vector, and represents the associated coefficient matrix. represents the error term.

Although Eq. (1) incorporates a comprehensive set of control variables, the potential issue of endogeneity may introduce bias into the OLS estimation results^22,23. Failure to account for these factors may result in estimation bias. To mitigate endogeneity concerns and obtain a more robust estimate of the causal effect of air pollution, the instrumental variable (IV) method was employed. This approach helps control for unmeasured confounding variables and alleviates potential estimation bias. The subsequent stratification by PA levels and the inclusion of an interaction term within this IV framework allow us to assess effect modification with greater internal validity than conventional approaches. The ventilation coefficient, defined as the product of wind speed at 10 m and the boundary layer height, serves as an indicator of the atmosphere’s ability to disperse pollutants⁵⁷. It was employed as an IV for air pollution in the analysis^58,59. IV estimation relies on two primary assumptions⁶⁰: (1) Relevance – The instrumental variable must be strongly correlated with air pollution. The F-test for weak instruments assesses whether the correlation between instrumental and endogenous variables is statistically significant. If the F-statistic exceeds 10, the instrumental variable is generally considered valid. A strong correlation exists between the instrumental variable (e.g., ventilation coefficient) and the endogenous variable (e.g., air pollution concentration), eliminating concerns about weak instruments. (2) Exogeneity – The IV must be uncorrelated with the error term in the regression model and should affect cognitive function only through its influence on air pollution, not through any other pathway. The Durbin-Wu-Hausman test examines the exogeneity of the IV. If the p-value is below the significance threshold, the null hypothesis is rejected, indicating that the instrumental variable is not exogenous⁶¹.

First, a higher ventilation coefficient enhances the dispersion of air pollutants, resulting in a negative correlation between the ventilation coefficient and pollution levels. This supports the relevance assumption of the IV approach⁶². Besides, since the ventilation coefficient is determined by large-scale atmospheric conditions and is not influenced by human economic activity^58,63, its direct effect on individual cognitive function is assumed to be minimal, aside from its indirect effect through air pollution. This satisfies the exogeneity assumption.

Information on 10-m wind speed and boundary layer height was obtained from the ERA5 gridded meteorological dataset. The city-level annual average ventilation coefficient for 2017 was calculated and matched to the corresponding survey data. Instrumental variable estimation was conducted using a two-stage least squares (2SLS) model, specified as follows:

2

3

4

Among these variables, represents the ventilation coefficient, and serving as the instrumental variable. denotes the air pollution value fitted according to Eq. (2). In the first stage of the two-stage least squares (2SLS) estimation, air pollution levels are regressed on the ventilation coefficient (serving as the instrumental variable) along with other control variables. This stage produces a fitted value for air pollution, isolating the variation explained by the instrument and covariates. The residual term from this regression captures the unexplained variation, which may reflect unmeasured confounding factors not included in the model.

Subsequently, the predicted air pollution values and the residuals from the first stage are incorporated into the second stage, where the same confounding factors as in the first stage are adjusted to assess the causal effects of air pollution on cognitive function. The air pollution coefficient is standardized using the interquartile range (IQR) to facilitate comparisons of the effects of different air pollution types on cognitive function. To further evaluate the moderating effect of PA, the analysis is stratified by PA type, and an interaction term capturing the product of air pollution and PA is introduced into the model.

To assess the robustness of the results, a series of sensitivity analyses was conducted. First, additional covariates, including nighttime sleep duration and social activity, were incorporated for correction. Second, individuals who changed their place of residence since 2015 (n = 1858) were excluded from the analysis. Third, air pollution exposure values were replaced with the averaged values from 2015, 2016, and 2017. Statistical analyses were performed using R, with a P-value of less than 0. 05 considered indicative of statistical significance.

Result

Descriptive statistics

Table 1 presents the descriptive statistics of the study population. Categorical variables are shown as counts and percentages, while continuous variables are reported as means with standard deviations (SD) or medians with IQR. Among the 17,332 participants, the average age was 61.22 ± 9.63 years; 47.66% were male, 13.79% were single, and 60.40% lived in rural areas. Additionally, 22.76% of participants were illiterate, 65.48% reported no alcohol consumption, and 57.39% were non-smokers. Chronic conditions were common, with 79.88% of individuals reporting at least one chronic disease. Regionally, 33.61% of participants were from eastern China, and 68.97% of households used clean fuels for cooking or heating. The mean cognitive function score was 12.70 ± 7.04. Median annual concentrations (with IQR) of key air pollutants were: PM₁ at 23.25 (10.60) µg/m³, PM_2.5 at 40.94 (21.11) µg/m³, PM₁₀ at 76.19 (47.28) µg/m³, NO₂ at 26.87 (14.77) µg/m³, SO₂ at 15.46 (8.66) µg/m³, and O₃ at 94.48 (19.35) µg/m³. An inverse association was identified between the ventilation coefficient and air pollution levels, with lower ventilation coefficients linked to higher concentrations of pollutants (Fig. 2).

Table 1.

Summary statistics of the study population, N = 17, 334.

Variables	Summary statistics
Age, mean (SD), years	61. 22 (9. 63)
Cognitive function, median (SD)	12. 70 (7. 04)
Gender (%)
Female	9072 (52. 34)
Male	8262 (47. 66)
Education level (%)
Illiterate	3946 (22. 76)
Primary school	7989 (46. 09)
Middle school	3400 (19. 61)
High school and above	1999 (11. 53)
Marital status (%)
Currently married	14,944 (86. 21)
Currently unmarried	2390 (13. 79)
Regional category (%)
East	5826 (33. 61)
Midland	6366 (36. 73)
West	5142 (29. 66)
Family residence (%)
Rural	10,469 (60. 40)
Urban	6865 (39. 60)
Smoking status (%)
Never smoke	9948 (57. 39)
Current smoke	4721 (27. 24)
Ever smoke	2665 (15. 37)
Drinking status (%)
Never drink	11,351 (65. 48)
< 1 /month	1327 ( 7. 66)
≥ 1 /month	4656 (26. 86)
Physical activity level (%)
Low (≤ 69. 3 METs-h/week)	8356 (48. 21)
High (> 69. 3 METs-h/week)	8978 (51. 79)
Cooking fuel use (%)
Clean fuel	11,955 (68. 97)
Solid fuel	5379 (31. 03)
Number of chronic diseases (%)
0	3487 (20. 12)
1	4081 (23. 54)
≥ 2	9766 (56. 34)
PM1, median (IQR), μg/m3	23. 25 (10. 60)
PM2.5, median (IQR), μg/m3	40. 94 (21. 11)
PM10, median (IQR), μg/m3	76. 19 (47. 28)
NO2, median (IQR), μg/m3	26. 87 (14. 77)
SO2, median (IQR), μg/m3	15. 46 (8. 66)
O3, median (IQR), μg/m3	94. 48 (19. 35)
NDVI, median (IQR)	0.33 (0.15)
Ventilation coefficient, median (IQR), m2/s	1114. 78 (477. 06)
Average temperature, mean(SD), ℃	15. 62 (5. 67)
Relative humidity, mean (SD), %	72. 02 (17. 12)

Categorical variables are presented as frequencies and percentages, while continuous variables are expressed as means with SD or medians with IQR.

SD: standard deviation; IQR: interquartile range; PM₁: particulate matter with aerodynamic diameter ≤ 1 µm; PM_2.5: particulate matter with aerodynamic diameter ≤ 2.5 µm; PM₁₀: particulate matter with aerodynamic diameter ≤ 10 µm; NO₂: nitrogen dioxide; SO₂: sulfur dioxide; O₃: ozone.

Fig. 2.

Annual average concentrations of air pollutants and ventilation coefficient (VC) in 2017.

Associations of air pollution and PA levels with cognitive function

In the univariate OLS model (Table 2), we find a statistically significant positive impact of all air pollution on cognitive function. After adjusting all covariates in model 2 using the OLS model, the positive effect on cognitive function was still statistically significant in PM₁, PM_2.5, PM₁₀, NO₂, SO₂; however, we didn’t find any association between O3 and cognitive function.

Table 2.

Associations of air pollution with cognitive function.

Variables	Model 1 (Crude)	Model 2 (Adjusted)
	β coefficient (95% CI)	β coefficient (95% CI)
PM1 (IQR = 10. 60 μg/m3)	0. 32 (0. 23, 0. 41)***	0. 37 (0. 25, 0. 50)***
PM2.5 (IQR = 21. 11 μg/m3)	0. 33 (0. 23, 0. 43)***	0. 34 (0. 19, 0. 48)***
PM10 (IQR = 47. 28 μg/m3)	0. 37 (0. 26, 0. 48)***	0. 45 (0. 27, 0. 64)***
NO2 (IQR = 14. 77 μg/m3)	0. 38 (0. 27, 0. 49)***	0. 21 (0. 03, 0. 39)*
SO2 (IQR = 8. 66 μg/m3)	0. 18 (0. 11, 0. 26)***	0. 18 (0. 06, 0. 30)**
O3 (IQR = 19. 35 μg/m3)	0. 54 (0. 41, 0. 66)***	0. 17 (− 0. 03, 0. 37)

IQR: interquartile range; PM₁: particulate matter with aerodynamic diameter ≤ 1 µm; PM_2.5: particulate matter with aerodynamic diameter ≤ 2.5 µm; PM₁₀: particulate matter with aerodynamic diameter ≤ 10 µm; NO₂: nitrogen dioxide; SO₂: sulfur dioxide; O₃: ozone.

***p < 0. 001,

**p < 0. 01,

*p < 0. 05. 95% confidence interval in parentheses based on robust standard error. Model 1: Crude model; Model 2: Multivariate model adjusted for sociodemographic characteristics (age, gender, living area, marital status, education level, regional category, family residence) , lifestyle habits and health (smoking status, drink status, cooking fuel use, number of chronic diseases), and meteorological conditions (average temperature, relative humidity).

Causal effect of air pollution on cognitive function

The Durbin-Wu-Hausman test⁶¹ was conducted to evaluate the endogeneity of air pollution exposure. All p-values were below 0.001 (Table S1), leading to rejection of the null hypothesis and indicating that air pollution is indeed endogenous. To address this, instrumental variable (IV) estimation was employed.Results from the first-stage IV analysis (Table 3) showed a significant inverse relationship between the ventilation coefficient and pollutant levels. The first-stage F-statistics exceeded the threshold of 10, confirming the strength of the instrument and mitigating concerns about weak instrument bias⁶⁴. In the second-stage IV estimation (Table 4), air pollution exposure was found to significantly impair cognitive performance. An increase of one IQR in pollutant concentration corresponded to the following reductions in cognitive scores: PM₁: − 0.45 (95% CI: − 0.76, − 0.13), PM_2.5: − 0.43 (95% CI: − 0.74, − 0.13), PM₁₀: − 0.53 (95% CI: − 0.90, − 0.16), NO₂: − 0.57 (95% CI: − 0.96, − 0.17), SO₂: − 0.47 (95% CI: − 0.80, − 0.14), O₃: − 1.06 (95% CI: − 1.81, − 0.31).

Table 3.

First-stage of IV estimation: Effects of ventilation coefficient on air pollution (per IQR increase).

	PM1	PM2.5	PM10	NO2	SO2	O3
Ventilation coefficient	− 0. 37*** (− 0. 38, − 0. 36)	− 0. 38*** (− 0. 39, − 0. 37)	− 0. 31*** (− 0. 32, − 0. 30)	− 0. 29*** (− 0. 30, − 0. 28)	− 0. 35*** (− 0. 36, − 0. 34)	− 0. 16*** (− 0. 16, − 0. 15)
F value of first stage	662. 09	820. 49	1219. 29	1161. 53	1114. 6	1105. 05

IQR: interquartile range; PM₁: particulate matter with aerodynamic diameter ≤ 1 µm; PM_2.5: particulate matter with aerodynamic diameter ≤ 2.5 µm; PM₁₀: particulate matter with aerodynamic diameter ≤ 10 µm; NO₂: nitrogen dioxide; SO₂: sulfur dioxide; O₃: ozone.

***p < 0. 001,

**p < 0. 01,

*p < 0. 05. 95% confidence interval in parentheses based on robust standard error. All modes adjusted for sociodemographic characteristics (age, gender, living area, marital status, education level, regional category, family residence) , lifestyle habits and health (smoking status, drink status, cooking fuel use, number of chronic diseases), and meteorological conditions (average temperature, relative humidity).

Table 4.

Second-stage of IV estimation: Effects of air pollutants on cognitive function.

Variables	β coefficient	95% CI
PM1 (IQR = 10. 60 μg/m3)	− 0. 45	(− 0. 76, − 0. 13)**
PM2.5 (IQR = 21. 11 μg/m3)	− 0. 43	(− 0. 74, − 0. 13)**
PM10 (IQR = 47. 28 μg/m3)	− 0. 53	(− 0. 90, − 0. 16)**
NO2 (IQR = 14. 77 μg/m3)	− 0. 57	(− 0. 96, − 0. 17)**
SO2 (IQR = 8. 66 μg/m3)	− 0. 47	(− 0. 80, − 0. 14)**
O3 (IQR = 19. 35 μg/m3)	− 1. 06	(− 1. 81, − 0. 31)**

IQR: interquartile range; PM₁: particulate matter with aerodynamic diameter ≤ 1 µm; PM_2.5: particulate matter with aerodynamic diameter ≤ 2.5 µm; PM₁₀: particulate matter with aerodynamic diameter ≤ 10 µm; NO₂: nitrogen dioxide; SO₂: sulfur dioxide; O₃: ozone.

***p < 0. 001,

**p < 0. 01,

*p < 0. 05. 95% confidence interval in parentheses based on robust standard error. All modes adjusted for sociodemographic characteristics (age, gender, living area, marital status, education level, regional category, family residence), lifestyle habits and health (smoking status, drink status, cooking fuel use, number of chronic diseases), and meteorological conditions (average temperature, relative humidity).

The modifying role of PA

Stratified analysis by PA levels was conducted to explore the correlation between air pollution and cognitive function (see Fig. 3). Results indicated that PA could moderate the correlation between them (p for interaction: PM₁ = 0.006, PM_2.5 = 0.006, PM₁₀ = 0.005, NO₂ = 0.021, SO₂ = 0.004, and O₃ = 0.003). Specifically, individuals with low PA levels experienced adverse cognitive effects from air pollution, whereas those with higher PA levels showed no significant negative impact on cognitive function from air pollution.

Fig. 3.

Effect of air pollution (per IQR) on cognitive function in low and high PA level subgroups. ***p < 0. 001, **p < 0. 01, *p < 0. 05. 95% confidence interval in parentheses based on robust standard error. All modes adjusted for sociodemographic characteristics (age, gender, living area, marital status, education level, regional category, family residence), lifestyle habits and health (smoking status, drink status, cooking fuel use, number of chronic diseases), and meteorological conditions (average temperature, relative humidity). PA: physical activity. IQR: interquartile range; PM₁: particulate matter with aerodynamic diameter ≤ 1 µm; PM_2.5: particulate matter with aerodynamic diameter ≤ 2.5 µm; PM₁₀: particulate matter with aerodynamic diameter ≤ 10 µm; NO₂: nitrogen dioxide; SO₂: sulfur dioxide; O₃: ozone.

Sensitivity analysis

The robustness of the association between air pollution and depressive symptoms was confirmed through several sensitivity analyses. The moderating effect of PA remained stable across all models. Additional adjustments accounted for self-reported participation in social activities and sleep duration. (Table S2) produced results consistent with the primary findings. Further sensitivity tests included: Excluding participants who had changed their residential address since the previous wave (Table S3), and replacing the exposure metric with the average air pollution levels from 2015, 2016, and 2017 (Table S4). All results remained consistent, reinforcing the validity of the main findings.

Discussion

This study examines data from the China Health and Retirement Longitudinal Study (CHARLS) to investigate the relationship between environmental air pollution exposure and cognitive function among middle-aged and older adults in China. The findings demonstrate that PA significantly moderates this relationship. Higher levels of air pollution exposure are associated with poorer cognitive function, particularly among individuals with lower levels of PA. Conversely, this adverse effect is substantially diminished in those who engage in higher levels of PA. The findings remained consistent across various sensitivity analyses and after controlling for potential confounding factors, demonstrating the robustness of the results.

While observational studies have identified an association between air pollution and cognitive function, methodological inconsistencies have led to varying results. For instance, a study on populations in northwest China found no correlation between exposure to atmospheric NO₂ and SO₂ and cognitive function, whereas a nationwide survey in China reported no association between high concentrations of O₃ and NO₂ and cognitive function¹². Similarly, studies in the United States²⁰ and South Korea²¹ observed either no significant association or even a beneficial effect of O₃ on cognitive function, while another study in South Korea found a positive effect of SO₂ exposure²¹. These contradictions may stem from traditional research practices that regress absolute pollution values into equations, overlooking unobserved factors such as socioeconomic status and lifestyle habits. Additionally, reverse causality and inaccuracies in measuring air pollution exposure contribute to endogeneity, potentially distorting the results⁷. The ventilation coefficient, determined by wind speed and mixing layer height, influences atmospheric pollutant dispersion. Our data indicate that higher ventilation coefficients correspond to lower pollutant concentrations and are relatively independent of confounding variables in this study. Thus, Employing the ventilation coefficient as an instrumental variable allows for the isolation of exogenous pollution variations, enabling more reliable causal inference and accurate evaluation of pollution’s impact on cognitive function. However, the application of causal frameworks to assess air pollution’s effects on cognition remains underexplored^18,19.

The impact of air pollution on cognitive function may be attributed to several mechanisms. First, exposure to air pollution can impair brain health through oxidative stress, neuroinflammation, and blood–brain barrier disruption. Pollutants such as PM_2.5 activate microglia, trigger the release of pro-inflammatory cytokines (e.g., IL-1β, TNF-α), and generate reactive oxygen species (ROS), leading to neuronal oxidative damage. This process contributes to beta-amyloid (Aβ) deposition and Tau protein hyperphosphorylation, both hallmarks of Alzheimer’s disease (AD)^8,9, Additionally, air pollutants can cross the blood–brain barrier via the olfactory nerve or bloodstream, degrading tight junction proteins (e.g., occludin and claudin-5). This increases barrier permeability, allowing neurotoxic substances to infiltrate the brain and compromise neurological health^10,11. Second, air pollution exacerbates the risk of stroke, heart failure, asthma, chronic obstructive pulmonary disease (COPD), and depression^53,54. These conditions are strongly linked to cognitive decline⁵⁵. Furthermore, elevated air pollution levels correlate with reduced outdoor activity, physical exercise, and social engagement, indirectly impairing neurological health^65,66.

Among air pollutants, O₃ demonstrated a particularly strong negative impact on cognitive function in our instrumental variable analysis, a finding that aligns with several previous reports. For instance, a study in northwest China indicated that the risk of cognitive impairment associated with O₃ was more than twice that of PM_2.5 and PM₁₀ per 10 μg/m³ increment⁵². The pronounced effect observed in our 2SLS model, in contrast to the null association in the OLS analysis, suggests that conventional regression might fail to capture O₃’s true neurotoxicity due to unaddressed confounding, while the IV approach provides a less biased estimate. The potent neurotoxicity of O₃ may be attributed to its distinctive physicochemical properties and mechanisms of action. As a powerful gaseous oxidant, O₃ primarily induces injury through robust oxidative stress. Upon inhalation, it reacts rapidly with pulmonary surfactants to generate reactive oxygen and nitrogen species (RONS) and a cascade of lipid peroxidation products, which can enter the systemic circulation^67,68. These secondary oxidative molecules are capable of crossing the blood–brain barrier (BBB), triggering neuroinflammation, disrupting BBB integrity, and directly causing oxidative damage to neuronal lipids, proteins, and DNA^68,69. In contrast, PM_2.5 and PM₁₀ largely exert effects indirectly through the adsorption and translocation of harmful components (e.g., heavy metals, polycyclic aromatic hydrocarbons) that subsequently induce systemic inflammation and oxidative stress. The direct and pervasive oxidative assault mediated by O₃ and its derivatives may underlie its particularly strong association with cognitive decline⁷⁰. We also acknowledge the reviewer’s pertinent observation regarding differing spatial resolutions of pollutant data. O₃ was modeled at a 1 km resolution, while NO₂ and SO₂ were estimated at 10 km. This methodological difference could theoretically influence effect estimates. Finer-resolution exposure assessment generally reduces exposure misclassification, potentially allowing for more precise and accurate effect estimation. The stronger effect seen for O₃ might therefore be partly attributable to its more precise spatial modeling compared to NO₂ and SO₂. However, it is noteworthy that PM₁, PM_2.5, and PM₁₀, which were also modeled at 1 km resolution, did not exhibit effect estimates of comparable magnitude to O₃. This suggests that while differential exposure measurement error might play a role, the unique physicochemical and mechanistic properties of O₃ likely constitute a primary explanation for its distinctive effect size.

Traditional perspectives posit that high-level outdoor PA may increase exposure to air pollutants, potentially causing adverse health effects^71,72. This underscores a possible trade-off between PA and air pollution exposure⁷². Nevertheless, extensive research indicates that the benefits of PA outweigh its associated risks in polluted environments^7,65. Our findings reveal that elevated PA levels not only fail to intensify air pollution’s detrimental effects on cognitive function but may counteract them entirely. This suggests that the potential neuroprotective effects of vigorous PA may mitigate air pollution’s harmful impact on cognition^7,65. Moreover, a year-long investigation demonstrated that PA alleviates cognitive decline resulting from short-term PM_2.5 exposure⁷³. Despite these insights, studies examining PA’s moderating role in the relationship between air pollution and cognitive health remain scarce.

The mechanism linking air pollution, PA, and cognitive function may involve several factors. First, air pollution triggers neuroinflammation, which can damage nerve cells. Exercise, however, promotes the release of anti-inflammatory factors, inhibits inflammatory pathways, and mitigates neuronal damage^9,24, Second, high-level PA stimulates the production of antioxidant enzymes that neutralize free radicals, enhance the brain’s antioxidant capacity, and reduce oxidative stress-induced neuronal damage from air pollutants²⁵. Third, high-level exercise significantly increases brain-derived neurotrophic factor (BDNF), a critical neurotrophic factor involved in neuronal growth, differentiation, and maintenance. BDNF supports dopamine neuron function, enhances synaptic plasticity, and facilitates neuronal survival and regeneration, promoting cognitive resilience against air pollution-related damage^9,25,74.

However, a cautious interpretation of this moderating effect is imperative. The observed attenuation of air pollution’s cognitive impact among individuals with high PA levels may not solely be attributed to biological protection. It is plausible that our findings are confounded by underlying differences in exposure patterns. For instance, individuals who report high PA might predominantly engage in indoor activities (e.g., in gyms or sports centers), thereby reducing their personal exposure to ambient pollutants during exercise. Conversely, those with lower PA levels might spend more time outdoors for non-exercise purposes (e.g., commuting or socializing) in polluted environments, leading to higher cumulative exposure. This behavioral difference could create a spurious association where the ‘protective’ effect is partly, or entirely, driven by reduced exposure rather than enhanced physiological resilience. Furthermore, urban–rural disparities could play a significant role. Urban areas, while often having higher pollution levels, also provide greater access to indoor exercise facilities. Rural residents, despite potentially cleaner ambient air on average, may engage in more outdoor agricultural or domestic labor, increasing their direct exposure. Our models, while adjusted for urban–rural residence, cannot fully capture these nuanced behavioral and environmental contexts. Therefore, the relationship between PA, pollution exposure, and cognitive function is likely influenced by a complex interplay of biological, behavioral, and geographical factors.

This study has several limitations. First, while our primary analysis focused on one-year average exposure preceding the follow-up assessment to align with the longitudinal design, cognitive decline is a slow process influenced by cumulative exposure over decades. Second, participants’ locations were identified only at the city level rather than at more granular district or county levels, and air pollution exposure was estimated using citywide average concentrations. Third, our outcome variable, the overall cognitive function score, did not account for the effects of air pollution on specific cognitive domains, and the potential moderating role of PA remains unknown. Future studies should employ more detailed questionnaires to assess cognitive domains comprehensively. Fourth, recall bias may affect the evaluation and collection of PA data, as specific exercise durations were not recorded. Fifth, Although the CHARLS is a longitudinal study, detailed MET data on PA were significantly missing in waves prior to 2018—whereas the 2018 survey wave provides the most complete and standardized MET data. Therefore, to ensure the accuracy and consistency of assessments for the exposure variable PA and outcome variable (cognitive function), we adopted a cross-sectional analysis framework using 2018 CHARLS data. Sixth, as noted previously, we lacked data on several important behavioral modifiers of exposure, including the proportion of time spent indoors, the use of air filtration systems in homes, and mask-wearing habits. These factors could significantly alter personal exposure profiles and introduce non-differential misclassification, likely biasing our results toward the null. The use of a single exposure window may not fully capture the long-term neurotoxic effects of air pollution and could underestimate the true effect size. Future studies with longer follow-up periods and capacity to model life-course exposure are needed.

Conclusions

These findings underscore the potential importance of policies that simultaneously target pollution reduction and promote PA to safeguard cognitive health. Nonetheless, the precise mechanism behind PA’s apparent mitigating role—whether primarily through biological adaptation or behavioral exposure reduction—remains to be fully elucidated. Thus, our results should be considered hypothesis-generating, highlighting the need for future studies equipped with detailed data on activity location (indoor/outdoor), time-activity patterns, and personal exposure monitoring to confirm a causal protective effect.

Environmental implication

The extent to which PA moderates the relationship between air pollution and cognitive function remains uncertain. This study identified a negative correlation between air pollution exposure and cognitive performance. Notably, engagement in PA plays a crucial role in shaping this association. Among individuals with lower levels of PA, air pollution is linked to poorer cognitive function. However, among those who engage in high-level PA, this adverse effect diminishes. This finding is especially important because it offers concrete evidence of air pollution’s detrimental effects on cognitive function, while also indicating that the cognitive advantages of PA may help counterbalance these impacts.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

O₃

Ozone

CHAP

China High Air Pollutants

IQR

Interquartile range

CHARLS

China Health and Retirement Longitudinal Study

PM1

Particulate matter with diameters ≤ 1 µm

PM2.5

Particulate matter with diameters ≤ 2.5 µm

PM10

Particulate matter with diameters ≤ 10 µm

HRS

Health and Retirement Study

TICS

Interview of Cognitive Status

IPAQ

International Physical Activity Questionnaire

MET

Metabolic equivalents

NDVI

Normalized difference vegetation Index

ERA5

Fifth Generation European Reanalysis

GEE

Google Earth Engine

2SLS

Two-stage least squares

OLS

Ordinary least squares

SD

Standard deviations

ROS

Reactive oxygen species

Aβ

Beta-amyloid

AD

Alzheimer’s disease

COPD

Chronic obstructive pulmonary disease

RONS

Reactive oxygen and nitrogen species

BDNF

Brain-derived neurotrophic factor

Author contributions

Lin Zhu: Conceptualization, Writing (original draft), Visualization, Data analysis. Mingjun Zou: Methodology, Data cleaning, Data analysis, Validation, Supervision, Writing (review & editing).

Funding

No fund was received.

Data availability

The datasets referenced in this study are accessible through the China Health and Retirement Longitudinal Study (CHARLS) repository, which you can find at their official website: http://charls.pku.edu.cn.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

The Ethics Review Committee of Peking University reviewed and approved this study involving human participants. Written informed consent was obtained from all participants before their involvement.