Abstract

Large power losses were observed in two well-powered placebo-controlled clinical trials as the primary efficacy endpoint excessively declined due to rapid disease progression. Despite the common practice of applying a parametric model on rank transformed data to address the issue of extreme data due to the robustness of ranks less affected by extreme values, there is no clear consensus among practitioners or in published work regarding the conditions under which a rank transformed analysis should be performed. The question remains: in practice, at what point do data become so extreme that parametric tests stop being effective and nonparametric tests become necessary. This paper aims to raise awareness of non-normal data with extreme values in drug development. We evaluate the impact of non-normality and extreme values on statistical performance through the approach of a latent variable model and provide a framework to identify and handle efficacy data with extreme values through the evaluation of the Jarque Bera test and Kurtosis.

Keywords

Rapid disease progression, extreme values, kurtosis, latent variable model, power loss