Confounding mechanisms and adjustment strategies in air pollution epidemiology: a case study assessment with the UK Biobank cohort.

Abstract

Background: Cohort studies are instrumental in examining long-term risks associated with environmental exposures but require appropriate control for various confounding effects. In this contribution, we assessed this issue by investigating the relationship between fine particulate matter (PM2.5) exposure and mortality in a UK-based cohort.

Methods: We analysed data from half a million adults in the UK Biobank linked with time-varying individual-level exposure data and followed up during the period 2006-21. The assessment focused on confounding related to spatial and temporal patterns as well as due to measurable variables, including both contextual and individual-level factors. We performed an evaluation consisting of descriptive analyses, specification and interpretation of direct acyclic graphs (DAGs), and comparison of results from survival models.

Results: We found correlations between PM2.5 exposure and mortality rates across time, geographical areas, and categories of measurable variables. The DAG indicated complex causal pathways and the need to adjust for a wide set of potential confounders. The regression analysis confirmed these patterns: the fully adjusted model estimated a hazard ratio (HR) of 1.25 (95% CI: 1.06-1.49) per 10 μg/m3 increments in PM2.5, but the association reversed to 0.82 (0.76-0.87) when excluding control for recruitment centre, suggesting strong spatial confounding. Calendar time showed stronger confounding effects compared to age. Area-level socio-economic indicators were more important than individual-level counterparts, while lack of control for lifestyle factors led to a noticeable overestimation.

Conclusions: This case-study illustration elucidates various confounding mechanisms in cohort studies on environmental risks and offers a critical evaluation of alternative adjustment strategies.