Weighted sum and order statistics methods for dynamic information borrowing in basket trials.

In basket trials, the same investigational therapy is studied on multiple sub-populations simultaneously under a single protocol. The goal of basket trials is to identify the sub-populations in which the therapy is effective. Basket trials have become a popular and generally accepted study design in disease areas including but not limited to oncology and rare diseases, for their advantages in operation and ethical considerations. Extensive research work on information borrowing has been conducted to explore the statistical efficiency in basket trials. In this paper, two novel frequentist methods for basket trials are proposed. The first method borrows information to minimize the mean squared errors in the treatment effect estimation. The second method uses information across all baskets to optimize the multiple testing task in detecting the treatment effects in each basket. Extensive simulation studies show that the proposed methods substantially improved statistical efficiency in basket trials while limiting family-wise error rate inflation. Both methods can be implemented with common statistical models with or without adjustment for covariates.