Exact statistical tests using integer programming: Leveraging an overlooked approach for maximizing power for differences between binomial proportions

Abstract: Traditional hypothesis testing methods for differences in binomial proportions can either be too liberal (Wald test) or overly conservative (Fisher's exact test), especially in small samples. Regulators favour conservative approaches for robust type I error control, though excessive conservatism may significantly reduce statistical power. We offer fundamental theoretical contributions that extend an approach proposed in 1969, resulting in the derivation of a family of exact tests designed to maximize a specific type of power. We establish theoretical guarantees for controlling type I error despite the discretization of the null parameter space. This theoretical advancement is supported by a comprehensive series of experiments to empirically quantify the power advantages compared to traditional hypothesis tests. The approach determines the rejection region through a binary decision for each outcome dataset and uses integer programming to find an optimal decision boundary that maximizes power subject to type I error constraints. Our analysis provides new theoretical properties and insights into this approach's comparative advantages. When optimized for average power over all possible parameter configurations under the alternative, the method exhibits remarkable robustness, performing optimally or near-optimally across specific alternatives while maintaining exact type I error control. The method can be further customized for particular prior beliefs by using a weighted average. The findings highlight both the method's practical utility and how techniques from combinatorial optimization can enhance statistical methodology.