Proper metrics for clinical trials: transformations and other procedures to remove non‐normality effects

Abstract

A simulation study was performed to study the effects of non-normality and to examine procedures to ameliorate possible loss of power when data are incorrectly assumed to be normally distributed. It was found that only distributions with high asymmetry or heavy tails seriously affect the t-test. The Box-Cox likelihood ratio test appears to have some advantages over the others, but this must be offset by the greater complexity in making the results understandable to non-statisticians. The variability in outcomes with the different procedures demonstrates the importance of specifying such procedures a priori.