|
Abstract:
|
In this paper, we develop a Bayesian adaptive design methodology for basket trials with binary response rate endpoint using model averaging. Several recent design methods have been proposed that attempt to borrow information across baskets in order to increase the efficiency of these trials. Existing methods seek to borrow information based on the degree homogeneity of estimated response rates across all baskets. In reality, an investigational product may only demonstrate efficacy for a subset of baskets, and the degree of efficacy may vary across that subset. Accordingly, an ideal methodology would acknowledge these possibilities and account for them in the information borrowing procedure. We use Bayesian model averaging to borrow information across baskets by averaging over the complete model space for the treatment effects which can include thousands of models. We present results that demonstrate that this computationally feasible Bayesian approach outperforms existing state-of-the-art frequentist approaches without the need to incorporate subjective prior information into the design, even when the Bayesian approach is held to stringent type I error control requirements.
|
