SER
This article provides defines simple structural pattern that underlies all epidemiologic selection biases, which involves conditioning in some way on a variable that is a “collider”, which distinguishes this form of bias from “confounding”. Not all authors embrace this distinction so categorically (cf. Modern Epidemiology 3rd Edition, pages 137 and 194)
Hernán MA, Hernández-Díaz S, Robins JM. A structural approach to selection bias. Epidemiology 2004 Sep;15(5):615-25. PMID: 15308962
Although recognized theoretically, collider stratification bias was assumed to be a relatively weak phenomena (Greenland Epidemiology 2003 May;14(3):300-6). This paper proposed that it was responsible for a reversal in the sign of an association, a paradox that had baffled perinatal epidemiologists for decades.
Hernández-Díaz S, Schisterman EF, Hernán MA. The birth weight \”paradox\” uncovered? Am J Epidemiol 2006 Dec 1;164(11):1115-20. PMID: 16931543
These authors demonstrate how there can be a harmful exposure that never prevents death in any individual, and yet after some time observing a cohort with deaths due to the exposure, the association measure between exposure and mortality can appear protective.
Flanders WD, Klein M. Properties of 2 counterfactual effect definitions of a point exposure. Epidemiology 2007 Jul;18(4):453-60. Erratum in: Epidemiology. 2008 Jan;19(1):168. PMID: 17473709
Smokers are observed to have a lower incidence of cognitive decline, but this occurs because smokers have a higher mortality rate, and dead people don’t experience cognitive changes. These authors use inverse probability of censoring weights to correct for this bias, such that although smokers are observed to have less cognitive decline, they actually have more decline than non-smokers in adjusted analyses.
Another longstanding epidemiologic paradox is the observation that although obesity is generally associated with higher mortality, it appears protective in those who are diagnosed with a chronic disease. These authors propose collider stratification bias as an important mechanism contributing to this observation.
In yet another real-world application of this phenomenon, these authors demonstrate powerful selection bias in a genetic association study, demonstrating the necessity of choosing covariates based on a causal model, rather than through any kind of automated or statistical criterion.
Day FR, Loh PR, Scott RA, Ong KK, Perry JR. A Robust Example of Collider Bias in a Genetic Association Study. Am J Hum Genet. 2016 Feb 4;98(2):392-3. PMID: 26849114