Collider scope: when selection bias can substantially influence observed associations

Summary

This methodological paper argues that selection bias in epidemiological studies can induce collider bias—a form of bias occurring when conditioning on a variable that is influenced by two independent exposures—leading to substantially distorted estimates of associations. Through simulation, the authors demonstrate that even modest influences on study selection or attrition can generate misleading results for both phenotypic and genetic associations, challenging the common assumption that selection bias primarily affects prevalence estimates rather than association estimates. The paper emphasises the importance of characterising factors influencing selection and attrition, and suggests sensitivity analyses using available data (such as DNA on birth cohort participants) to quantify potential bias.

UK applicability

The findings are directly applicable to UK birth cohort studies (such as the Avon Longitudinal Study of Parents and Children) and other large longitudinal studies conducted in the United Kingdom, where understanding selection and attrition mechanisms is critical for valid inference. The recommendations for sensitivity analyses and adjustment strategies would be particularly relevant to UK research governance and quality assurance in longitudinal studies.

Key measures

Bias in association estimates (phenotypic and genotypic); impact of selection-related collider conditioning; polygenic score prediction of study participation

Outcomes reported

The study examined how selection bias arising from non-random recruitment or attrition in large-scale cohort and cross-sectional studies can induce collider bias and substantially distort estimates of associations between exposures and outcomes. Simulations demonstrated that even modest selection effects related to phenotypes can generate biased estimates of both phenotypic and genotypic associations.

Topic tags