The Same but Different?: A Systematic Review of the Impact of Selection and Collider Bias on Internal Validity

Recent work conceptually unifying selection and collider-restriction bias as threats to internal validity implies that their impact on observed associations should similarly align. We reviewed epidemiologic literature to summarize existing knowledge about the impact of selection and collider bias.

We systematically searched for peer-reviewed, methodologic articles and general epidemiology textbooks published in English from 1 January 2000 to 12 July 2024. We included sources that focused on internal validity and discussed the magnitude or direction of selection or collider bias. We abstracted conclusions about the likely magnitude and direction of bias, which stratum or strata are affected when restricting analyses to a subset, and the conditions under which the consequences of bias were evaluated.

We retained 33 of 5508 identified articles and 12 of 205 textbooks for data abstraction. Overall, we found that collider bias articles conveyed its impact as minimal while selection bias sources described variable effects. We also observed that most collider bias sources evaluated bias under the sharp null (assuming no relationship between the exposure and outcome) and found differences between how selection and collider bias sources discussed the role of interaction and the strata affected.

Although collider-restriction and selection bias affecting internal validity are considered theoretically equivalent, conclusions differ about their consequences for study results. Investigating collider bias not under the sharp null and considering the role of both multiplicative and additive interaction between the causes of a collider may improve our ability to predict and quantify its impact on internal validity.