Covariate Selection in Physical Activity Epidemiology: Too Little, Too Much, or Just Right?

Abstract

Many observational studies are interested in the total effect of physical activity on any given health outcome. These studies rely heavily on covariate adjustment, yet decisions about which variables to include vary widely. Such decisions have meaningful consequences, including altering estimated effect sizes and their interpretation, and ultimately, modifying our understanding of how physical activity influences health. For example, inclusion of biological intermediates is likely to constitute overadjustment, while the role of co-occurring health behaviors is more conceptually complex. Despite longstanding recognition of overadjustment bias, there remains little consensus regarding which variables should routinely be included in models that use observational data to explore the effect of physical activity on health. In this paper, we discuss key conceptual definitions and frameworks relevant to covariate selection, such as confounding, mediation, and overadjustment, alongside the shift from statistical toward conceptual approaches including causal inference frameworks (eg, directed acyclic graphs). We outline common biological and behavioral pathways linking physical activity to downstream health outcomes, highlighting the challenges of adjusting for biological intermediates, health-status indicators, and co-occurring health behaviors. We draw on empirical illustrations using data from cohort studies in Brazil, the United Kingdom, and the United States to explore how common adjustment strategies influence estimates of the association between physical activity and each of hypertension, type 2 diabetes, and all-cause mortality. The use of directed acyclic graphs highlights where adjustment decisions may be conceptually simple and where they require careful consideration of the underlying causal assumptions. Our findings demonstrate that adjustment decisions meaningfully influence both the magnitude and interpretation of associations in physical activity epidemiology. We conclude by proposing practical recommendations to improve transparency, interpretation, and covariate selection in future observational epidemiological analyses examining how physical activity influences health outcomes.