The effect of dichotomization of skewed adjustment covariates in the analysis of clinical trials

Abstract

Baseline imbalance in covariates associated with the primary outcome in clinical trials leads to bias in the reporting of results. Standard practice is to mitigate that bias by stratifying by those covariates in the randomization. Additionally, for continuously valued outcome variables, precision of estimates can be (and should be) improved by controlling for those covariates in analysis. Continuously valued covariates are commonly thresholded for the purpose of performing stratified randomization, with participants being allocated to arms such that balance between arms is achieved within each stratum. Often the thresholding consists of a simple dichotomization. For simplicity, it is also common practice to dichotomize the covariate when controlling for it at the analysis stage. This latter dichotomization is unnecessary, and has been shown in the literature to result in a loss of precision when compared with controlling for the covariate in its raw, continuous form. Analytic approaches to quantifying the magnitude of the loss of precision are generally confined to the most convenient case of a normally distributed covariate. This work generalises earlier findings, examining the effect on treatment effect estimation of dichotomizing skew-normal covariates, which are characteristic of a far wider range of real-world scenarios than their normal equivalents.