Key concepts in clinical epidemiology: collider-conditioning bias

Abstract

Collider-conditioning bias is a common concern in clinical epidemiology and can be difficult to recognize. Collider-conditioning bias is defined as the alteration in the association between an exposure and an outcome under study when conditioning on (via restriction, stratification or regression adjustment) a third factor influenced by both exposure and outcome. Collider-conditioning bias may also be induced by conditioning on a third factor influenced by causes of both the exposure and outcome of primary interest. The shared consequence of the exposure and outcome (or their causes) is referred to as a collider. Collider bias can be induced by in-selection (e.g., recruitment) or out-selection (e.g., loss-to-follow-up, mortality) processes, as well as missing data or analytic choices about which variables are included as covariates. Collider-conditioning bias is especially relevant in clinical epidemiology because it is plausible whenever examining determinants of disease progression or prognosis if the exposure under study is also a cause of the initial occurrence of the disease. Collider-conditioning bias can lead to incorrect inferences about the study sample, the source population, and external populations. Drawing directed acyclic graphs (DAGs) can aid in identification of colliders and quantitative bias analysis may be useful to estimate the magnitude of collider-conditioning bias. Many notorious examples of selection bias are attributable to collider-conditioning. Identifying the potential for collider-conditioning can help researchers avoid or correct for the bias.