On study designs and hypotheses for clinical trials with predictive biomarkers

Recent interest in conducting clinical trials with predictive biomarkers has generated research in comparing relative efficiency of different trial designs. We find these comparisons of efficiency mostly misleading since they are based on different hypotheses. In this paper, we discuss several commonly used trial designs and consider the hypotheses that each design is capable to address. We first consider the ideal situation of no classification errors, then the more realistic situation where marker assay's sensitivity, specificity and the rule of classification are imperfect. We pay special attention to the differences between treatment utility versus absolute treatment effect, and marker by treatment interaction versus marker utility.