All Times EDT
Keywords: Bayesian Augmented Control Design, Hierarchical Modeling, prior distribution, Inverse Gamma Distribution
Bayesian Augmented Control Design (BAC) is an important method to improve the statistical power and lower false positive rate in Proof-of-Concept (POC) trial for time-to-event end points by leveraging historical control data. One challenge in BAC design, however, is the negative impact of the deviation of borrowed historical control data from the current control. One potential solution is the Bayesian Hierarchical Modeling (BHM) that can dynamically borrow different amount of information from historical data based on the concordance of historical and current control. Meanwhile, when there are multiple historical controls with different median estimates, the inter-trial variance can be estimated by BHM in order to fine-tune the historical information borrowed dynamically. In BHM, the key hyper-parameter that controls the extent of information borrowing is Tau^2, the inter-trial variance whose prior distribution can be conveniently modeled by the inverse-gamma distribution (IVG). Currently, the impact of parameterization of IVG on the adaptability of historical borrowing in dynamic BAC is not well understood.
In this research, we study the information sharing behavior of dynamic BAC under different settings of IVG parameters for inter-trial variance Tau^2, in the context of a randomized phase 2 trials for Glioblastoma Multiforme (GBM) with 2:1 randomization ratio between treatment and control. This study provides useful guidance for the practitioners of BAC trial to optimize their prior setting for hyper-variance parameter in order to achieve effective borrowing and reduce the bias introduced by the discordance between the historical and current control cohort. Furthermore, the impact of BHM on the mitigation of the compromised power and inflated type I error resulted from this discordance is also studied.