Subgroup analysis using exchangeability probabilities.

Subgroup analyses in clinical trials are typically conducted using p-values from a univariate test in a one-arm setting or an interaction model in a two-arm setting. We propose an alternate method using Bayesian posterior weights to estimate the exchangeability probabilities of the groups and establish a cutoff value for "significance". Recommended use case scenarios are given for each trial design and method studied where this method outperforms the standard approach. For example, a two-arm trial with a binary outcome, effect size difference of 0.2, and sample size of 800, the power increased by 0.2904 (LRT power = 0.6705, MEM power = 0.9609) while the type I error rate was essentially unchanged from 0.0532 (LRT) to 0.0530 (MEM). Finally, we apply our method to data from a clinical trial for COVID-19 to test for subgroup effects and evaluate performance in a real-world setting.