Unblinded Adaptive Statistical Information Design Based on Clinical Endpoint or Biomarker

Abstract

The most frequently seen adaptation of a clinical trial design in regulatory submission is adaptation of statistical information, for example, sample size or number of events. Such adaptation can be based solely on the clinical endpoint of interest or early biomarker. In this article, we articulate the technical merits and discuss challenges when statistical information based solely on the clinical endpoint is used as the design aspect for adaptation. We present the interplay between the weighted and unweighted adaptive Z-statistics with, versus without, additional criteria. We contrast Fisher's p-value product test and the modified version to the adaptive weighted Z-test to elucidate a way to minimize the potential heterogeneity of the observed treatment effects between stages in a two-stage adaptive design setting. Another framework pertains when one is using shorter-term biomarker data for adaptation of statistical information, where the final analysis is to test the null hypothesis of no treatment effect based on the ultimate clinical outcome. It has been argued that under such a framework, no additional Type I error rate control is needed for the final analysis since the clinical endpoint is not used for adaptation. However, we show analytically that, for such an adaptation, the maximum Type I error probability can be far greater than the conventional Type I error level. We conclude by providing a few recommendations.