Berkson's Paradox: A Common Statistical Error in Genomics Analyses

Abstract

Berkson's paradox describes a scenario where a spurious correlation arises between two independent variables within a subset of samples selected based on a third variable which the first two variables cause. The conditions for Berkson's paradox are perfect in genomics due to many variables we study being causally intertwined, common selection criteria being biased either intentionally or not by the variables of interest, and our propensity to look towards correlations for windows into underlying mechanism. Here, we provide examples of cases where spurious correlations are likely to arise in common genomics analyses, and suggest analysis approaches to avoid misinterpreting spurious correlations.