Assessing the impact of endpoint shopping on power in confirmatory clinical trials

When planning a confirmatory study, one of the important aspects is choosing a primary endpoint and evaluating power to show a difference. Data from preliminary studies are generally used for such planning. It is natural to want to use the endpoint from the preliminary study that best differentiates between test drug and control as the primary endpoint in the confirmatory study. However this leads to the possibility of bias in estimation of the effect size in the preliminary study, and, hence, lower than anticipated power in the confirmatory study. In this paper we quantify the impact of such endpoint shopping on the power of confirmatory studies. We find the upper bound on bias and show that for low to moderate correlation it is not very conservative. We derive the asymptotic distribution of the treatment effect and propose a test forequal treatment effects for the data from the preliminary study when endpoints are correlated. We study properties of this test. We propose a strategy to use the data from the preliminary study to plan confirmatory studies with unbiased or conservative estimates of power.