|
Abstract:
|
Oncology research has shifted from evaluating a single treatment in patients with a specific type of cancer to the development of targeted therapies designed to treat tumors with specific genetic mutations regardless of the tumor histology or site, otherwise known as ‘basket trials’. The rarity of a genetic mutation may make it infeasible to enroll enough patients to sufficiently power a study for individual tumor subtypes. Independently testing indications may lead to bias from sampling variability. On the other hand, the assumption of homogeneous responses may also lead to large biases and inflated type I error rates depending on the heterogeneity of subtypes. To overcome these assumptions, we utilize a Bayesian hierarchical model based on a non-parametric prior. Compared to other models that have been proposed for borrowing information across subtypes, our model does not depend of a pre-specified exchangeability parameter and does not force potentially limiting distributional assumptions on the prior parameter. We show that indication effect size estimation using the proposed Bayesian semi-parametric models produces less bias, leading to more robust go/no-go decision outcomes.
|
