A Bayesian Basket Trial Design Using Local Power Prior

Abstract

In recent years, basket trials, which allow the evaluation of an experimental therapy across multiple tumor types within a single protocol, have gained prominence in early-phase oncology development. Unlike traditional trials, which evaluate each tumor type separately and often face challenges with limited sample sizes, basket trials offer the advantage of borrowing information across various tumor types to enhance statistical power. However, a key challenge in designing basket trials is determining the appropriate extent of information borrowing while maintaining an acceptable type I error rate control. In this paper, we propose a novel three-component local power prior (local-PP) framework that introduces a dynamic and flexible approach to information borrowing. The framework consists of three components: global borrowing control, pairwise similarity assessments, and a borrowing threshold, allowing for tailored and interpretable borrowing across heterogeneous tumor types. Unlike many existing Bayesian methods that rely on computationally intensive Markov chain Monte Carlo (MCMC) sampling, the proposed approach provides a closed-form solution, significantly reducing computation time in large-scale simulations for evaluating operating characteristics. Extensive simulations demonstrate that the proposed local-PP framework performs comparably to more complex methods while significantly shortening computation time.