Generalized error rates for subgroup analyses
*Frank Bretz, Novartis  Willi Maurer, Novartis  Xiaolei Xun, Novartis 

Keywords: multiple testing, decision analysis, subgroup analysis

We consider the problem of comparing the treatment effect of a new drug against a comparator for two non-overlapping subgroups of patients defined by predictive biomarkers, demographic factors or any other classifier. A decision is to be made if and for which of the two subgroups the respective null hypotheses can be rejected and an advantage of the new drug over the comparator be claimed. We argue that in this situation traditional methods to control the Type I error rate are too restrictive and that the standard familywise error rate (FWER) is not appropriate. Instead, we propose decision procedures that allow us to control the FWER, but for which also upper bounds of expected values for more general loss functions can be derived.