(118g) Metaheuristic Optimization of Graphical Testing Procedures for Clinical Trials

When evaluating the safety and efficacy of compounds during a clinical trial, there is always the risk that a dataset will produce a statistically significant indication despite being the result of random noise, termed a “type-1 error.” Furthermore, when multiple hypotheses are tested in the same study, the likelihood of producing any spurious indications grows unless appropriate steps are taken to control the family-wise type-1 error rate. A commonly used approach is to commit to a graphical hypothesis testing procedure before analyzing the trial results. This involves allocating a “budget” of statistical significance to each clinical effect and committing to procedures for “sharing” significance in a way that strongly controls the family-wise type-1 error rate. Currently, there are a variety of strategies for manually building a graphical hypothesis test, but there are no algorithmic approaches for rigorously optimizing the design of a graphical hypothesis test subject to the inherent uncertainty in clinical outcomes. In this work, we translate the problem of designing graphical hypothesis tests to suitably defined mathematical optimization models. We then demonstrate approaches for systematically optimizing the resulting models. Due to the presence of parametric uncertainty and the highly non-convex nature of the graphical hypothesis testing procedures, we employ metaheuristic optimization algorithms and provide a thorough characterization of their algorithmic performance. This work streamlines the process of identifying clinical trial designs that are both powerful and averse to risk.