Monotonicity conditions for avoiding counterintuitive decisions in basket trials

Area of Science:

• Biostatistics
• Clinical Trial Design
• Oncology

Background:

• Basket trials test new treatments across patient subgroups (baskets) with common biomarkers but different tumor sites.
• Bayesian borrowing methods enhance statistical power in basket trials by pooling data from similar baskets.
• Existing methods can yield non-monotonic posterior probabilities, leading to paradoxical trial outcomes.

Purpose of the Study:

• To identify and address counterintuitive decision-making arising from Bayesian borrowing in oncology basket trials.
• To propose monotonicity conditions for robust inference in basket trial designs.
• To evaluate the impact of basket pruning on trial characteristics.

Main Methods:

• Analysis of Bayesian borrowing methods in single-stage basket trials with equal sample sizes.
• Investigation of monotonicity conditions proposed for basket trial inference.
• Simulation studies assessing the effect of basket pruning on type I error rate and statistical power.

Main Results:

• Bayesian borrowing can lead to non-monotonic posterior probabilities, causing counterintuitive results within and between trials.
• Monotonicity conditions are often violated as the number of baskets increases, depending on borrowing strength.
• Pruning baskets can help satisfy monotonicity conditions, with an associated impact on type I error and power.

Conclusions:

• The application of Bayesian borrowing in basket trials requires careful consideration to avoid paradoxical results.
• Proposed monotonicity conditions offer a framework for ensuring reliable inference.
• Basket pruning is a viable strategy to maintain trial integrity, though its effects on statistical efficiency must be evaluated.