A Generalized Bayesian Hierarchical Model in Basket Trials.

Area of Science:

• Oncology
• Biostatistics
• Clinical Trial Design

Background:

• Basket trials in oncology evaluate single treatments across multiple cancer types with shared genomic alterations.
• Standard Bayesian hierarchical models (BHM) borrow information across cancer types but struggle with adaptive borrowing strength and Type I error inflation when treatment effects vary.
• Existing BHM variations attempt to account for heterogeneity but may require complex modifications or hyperparameter tuning.

Purpose of the Study:

• To propose a generalized Bayesian hierarchical model (GBHM) that relaxes variance parameter assumptions in standard BHMs.
• To evaluate the performance of the GBHM against existing BHM variants using extensive simulations based on established study designs.
• To assess the robustness of the GBHM to treatment effect heterogeneity across different cancer histologies.

Main Methods:

• Developed a generalized Bayesian hierarchical model (GBHM) by relaxing the variance parameter assumption of the standard BHM.
• Conducted four simulation studies mirroring setups from existing BHM variant publications.
• Investigated the GBHM with Inverse-Gamma (IG) and Cauchy priors, exploring various hyperparameters.

Main Results:

• The GBHM demonstrated robustness to treatment effect heterogeneity with specific prior choices.
• GBHM with IG(0.01,0.01) prior allows for liberal information borrowing, suitable when Type I error is less critical.
• GBHM with Cauchy(25) prior provides conservative borrowing, recommended for situations prioritizing Type I error control.

Conclusions:

• The proposed GBHM offers a flexible and robust alternative to existing models for oncology basket trials.
• GBHM simplifies implementation compared to other methods, avoiding complex prior specifications.
• The choice between IG and Cauchy priors allows tailoring information borrowing strength based on study objectives.