Abstract: During the course of a fixed sample size clinical trial, differences in the size of the treatment effect between strata may become pronounced. If rejecting the null hypothesis of no treatment effect is of paramount concern, it may make sense to increase representation of the more responsive stratum to increase power of the trial. Additionally, strata in which treatment is harmful may need to be dropped for ethical reasons. This paper provides conditions under which such adaptation does not affect the type I error rate. The change in power resulting from adaptation under various strategies is investigated. Frequentist and Bayesian approaches to decision making are explored and a simulation is used to provide guidelines as to whether stratum proportions should be altered. A clinical trial with an early and substantial between stratum difference is reanalyzed under adaptation.
Key words and phrases: Bayesian statistics, decision theory, interim monitoring, interaction, strata.