Bayesian methods incorporating external data are attractive for clinical trials with limited patient availability, but controlling the type I error rate remains challenging. Psioda and Ibrahim (2019) proposed a simulation-based framework using external data, which evaluates the Bayesian type I error and power as rejection probabilities averaged over sampling priors. However, when the external data show a large treatment effect, the null sampling prior can degenerate and the Bayesian type I error rate approaches the frequentist rate. Moreover, the robust mixture prior used as the analysis prior has a second component as informative as the external data, potentially biasing estimates, and the framework does not accommodate hybrid-controlled trials. We propose an alternative sampling prior specification, assess estimation bias under the analysis prior, and extend the framework to hybrid-controlled trials. In simulations, the existing specification required sample sizes comparable to or larger than those of frequentist designs, whereas the proposed specification reduced the required sample size but allowed the frequentist type I error rate to exceed the nominal level. A weakly informative second component also reduced estimation bias under prior-data conflict. A pediatric cutaneous lupus erythematosus case study illustrates the method.
翻译:暂无翻译