Potential over-adjustment bias in cohort studies of air pollution and health: A methodological study

Confounder adjustment is essential in causal inference studies; however, over-adjustment-excessive statistical adjustment-can bias causal effect estimation. This issue is prevalent in environmental health impact studies but has not received sufficient attention. This study aimed to investigate current practices regarding over-adjustment in studies examining long-term air pollution exposure and health outcomes.

We searched PubMed from January 2021 to October 2023 for cohort studies published in high-impact journals that investigated long-term ambient air pollution exposure (including ambient fine particulate, carbon monoxide, compounds, oxynitride, ozone, sulfur dioxide) and health outcomes. Two reviewers independently screened studies and extracted relevant information. Potential over-adjustment was identified based on the Modified Disjunctive Cause Criterion. Descriptive analyses were performed on all data.

A total of 175 studies were included. Only 26 (14.9%) employed directed acyclic graphs (DAGs) for variable selection. More than half of the studies (122, 69.7%) exhibited potential over-adjustment; of these, 120 (68.6%) adjusted for mediators and 2 (1.1%) adjusted for both mediators and colliders. The most commonly adjusted mediators were Body Mass Index (98, 56.0%) and hypertension (25, 14.3%). Studies that applied DAGs demonstrated a significantly lower proportion of over-adjustment (42.3% vs 74.5%).

A considerable proportion of cohort studies on air pollution and health exhibit potential over-adjustment, with a few employing DAGs for confounder selection to mitigate this issue. We emphasize the importance of avoiding potential over-adjustment and advocate for the proper use of DAGs in accordance with established methodological guidelines to reduce the risk of over-adjustment and improve the validity of future research.