Air pollution is a well-recognized environmental threat to human health (Goshua et al., 2022). According to the World Health Organization (WHO), an estimate of 2.4 million deaths annually is attributable to causes related to air pollution (Sierra-Vargas and Teran, 2012). The Global Burden of Disease study identified ambient PM2.5 as the leading risk factor for global mortality and disability-adjusted life years in 2021 (Clark et al., 2025). Moreover, air pollution is an established risk factor for various diseases, such as ischemic heart disease, acute lower respiratory infections, and lung cancer (Chen and Hoek, 2020).

Cohort study is one of the frequently employed study designs for investigating the potential causal relationship between air pollution and health outcomes. Although advances in geospatial satellite-based technology have improved the accuracy of exposure measurement (Hoek et al., 2008), establishing such causal relationships remains challenging due to confounding bias, which significantly threatens the validity of causal inference (Hemkens et al., 2018). To reliably estimate exposure-outcome associations from observational data, appropriate adjustment for relevant confounding variables is essential (VanderWeele, 2019).

Appropriate identification of potential confounders is a crucial first step in causal inference. Misidentification can lead to inappropriate adjustment, particularly over-adjustment, which is further interpreted as excessive adjustment for mediators and colliders, may introduce bias and compromise the validity of the estimated exposure-outcome relationship (Schisterman et al., 2009). Directed acyclic graph (DAG) is a valid approach to visualize assumed causal structure among variables (Feeney et al., 2025), which can systematically organize complex causal relationships and identify potential confounders, thereby helping to avoid over-adjustment (Tennant et al., 2021a). However, the proper application of DAG for confounder adjustment in environmental epidemiological studies remains limited (Gao et al., 2025).

In practice, over-adjustment is often insufficiently acknowledged in studies of air pollution and health, and its prevalence may partly explain why the magnitude and even the existence of the reported causal relationship between air pollution and health outcomes remain controversial (Mortimer et al., 2022; Liang et al., 2014). For example, two studies investigating the association between PM2.5 and cardiovascular diseases (CVD) in the Chinese population reached different conclusions, with one using DAG for variable selection (Liu et al., 2022), whereas the other did not (Liu et al., 2021). The latter inappropriately adjusted for BMI, a mediator, leading to an underestimation of the association (hazard ratio [HR] 1.12, 95% CI 1.06–1.21) compared with the DAG-based estimate (HR 1.29, 95% CI 1.15–1.45).

Therefore, the present study aims to investigate the extent of potential over-adjustments in studies of long-term air pollution exposure and health outcomes, and seeks to provide recommendations for minimizing over-adjustment in future research.