Abstract:
|
Graphical testing (Bretz 2009) is a multiple testing approach that has been widely adopted in confirmatory clinical trials, where multiple endpoints are usually evaluated simultaneously to demonstrate the efficacy of the treatment. It is a closed testing procedure that achieves strong control of type I error, via passing and recycling the alpha among endpoints given a specified weighting scheme. In this study, we consider the problem of choosing an optimal weighting scheme. Each endpoint is considered to have a corresponding business value. The goal of optimizing the weight scheme is to maximize the Bayesian posterior expectation of the overall business value, while control for its variation (i.e., mitigating the overall risk). To simulate empirically realistic multivariate clinical trial data, we developed a Bayesian bootstrap-based approach, where the assumptions on drug effect size are drawn from a multivariate distribution that approximates our belief on the uncertainty of the assumptions. We also compared the method with another scenario-based approach in terms of its performance in maximizing the overall value versus controlling for the overall risk.
|