Toward a More Nuanced Understanding of the Statistical Properties of a Median Split

Abstract

Some behavioral researchers occasionally wish to conduct a median split on a continuous variable and use the result in subsequent modeling to facilitate analytic ease and communication clarity. Traditionally, this practice of dichotomization has been criticized for the resulting loss of information and reduction in power. More recently, this practice has been criticized for sometimes producing Type I errors for effects regarding other terms in a model, resulting in a recommendation of the unconditional avoidance of median splits. In this paper, we use simulation studies to demonstrate more thoroughly than has been shown in the literature to date when median splits should not be used, and conversely, to provide nuance and balance to the extant literature regarding when median splits may be used with complete analytical integrity. For the scenario we explicate, the use of a median split is as good as a continuous variable. Accordingly, there is no reason to outright reject median splits, and oftentimes the median split may be preferred as more parsimonious.

Highlights • We show that the categorical rejection of median splits is incorrect. • We show when median splits are appropriate, and when they are not. • When median splits are appropriate, they may be preferred as more parsimonious.

Keywords: Median split, Median-split, Dichotomization, Categorization