Online Program

Return to main conference page
Wednesday, September 25
Wed, Sep 25, 10:45 AM - 12:00 PM
Currently Used Statistical Models in Phase II Basket Trial Designs

A Robust Bayesian Approach for Basket Trial Design with Rare Tumor Types in Oncology (300975)

*Satrajit Roychoudhury, Pfizer Inc. 

Keywords: Basket trial, Borrowing, Bayesian methods, Bayesian

Prescription Drug User Fee Act (PDUFA) VI encourages the use of complex and innovative trial design that helps efficient and informative decision making in drug development. Basket trial designs in Oncology allows development of multiple tumor types in the same protocol. A Basket trial includes patients with certain disease characteristics (e.g., genetic mutation) in common regardless of the site or origin of cancer in the body. Basket trials are typically small sized, non-randomized and used at the early stage of the drug development. It helps to evaluate potential indications for further development. Although the patients enrolled in a basket trial have the same disease characteristics, that does guarantee homogeneous response across all tumor types. Tumor type often has profound effects on the treatment effect, and some tumor types are more effective. A naïve complete pool analysis may cause a large bias and inflation of type-I error. On the other hand, stratified analysis for each tumor type produces less efficient estimate and often lacks power due to limited sample size. In recent years, several Frequentist and Bayesian methods have been proposed in literature for efficient Basket trial design. All these approaches provide ways to borrow information across different tumor types and increase efficient treatment effect estimate. We proposed a Bayesian Hierarchical model to borrow information across tumor types in multi-arm Phase 2 Oncology trial with rare indications. Due to robust nature, the proposed model allows dynamic borrowing of information between groups. This implies more borrowing when the groups are consistent and less borrowing when the groups differ. In this way, the model is a compromise between the two alternate extremes of either a completely pooled analysis or a separate analysis in each group and provide reasonable strata specific and overall estimate. Design characteristics will be illustrated by data scenarios and simulation.