Causal Inference
Survivorship as a source of confounding, selection bias and effect modification C Mary Schooling* C Mary Schooling Yang Guoyi
Background: Survivorship, i.e., the state or condition of being a survivor, is a source of confounding, selection bias and effect modification that is often overlooked. Here, we describe how it can occur and when it can be averted.
Methods: Directed acyclic graphs were used to illustrate how survivorship could cause confounding and selection bias. Selection diagrams were used to illustrate how survivorship could cause effect modification by factors such as age or health status.
Results: Survivorship can be a confounder that biases towards favouring the exposure when exposure allocation starts or changes after recruitment. Survivorship can be a source of type 1 selection bias that attenuates or reverses estimates, when the exposure and the outcome, or the exposure and a competing risk of the outcome, affect survivorship before recruitment. Survivorship can be a source of type 2 selection bias when survivorship is an effect modifier. Confounding and selection bias by survivorship can largely be averted by ensuring exposure allocation is determined at recruitment. Given, survivorship is usually most relevant to older or sicker groups, bias from survivorship may be evident as associations which attenuate or reverse with advancing age or poorer health status.
Conclusions: Confounding by survivorship should be considered when exposure allocation starts or changes after recruitment. Type 1 selection bias from survivorship should be considered when the exposure starts before recruitment, and type 2 selection bias when survivorship is an effect modifier. Given, long-term exposures are often of interest, and confounding as well as selection bias are difficult to address, study designs less open to these biases might be preferable.