Improving inference in air pollution epidemiology: the case for rethinking multi-pollutant adjustment

Air quality regulations and programs are vital for protecting the public from harms caused by air pollution. To support these actions, numerous epidemiological studies have sought to identify the pollutants most responsible for adverse outcomes. These studies often used statistical adjustments for co-pollutants in outcome regression models, a practice also commonly applied to assess interactions between co-pollutants. Here, we highlight possible pitfalls of multi-pollutant analyses. Indiscriminate co-pollutant adjustment can induce noncausal associations through collider adjustment, distorting effect estimates for individual air pollutants. We describe the underlying mechanisms and provide empirical evidence on how such bias may realistically influence the relationships between air pollution and health outcomes from a well-characterized Canadian national cohort alongside a simulation study. Additionally, we discuss strategies to mitigate the impact of this bias. Given the widespread interest in multi-pollutant approaches among the scientific and policy communities, greater caution is needed when conducting and interpreting research on multiple pollutants.