Title: Case studies in calibrating informative hierarchical model priors
Speaker: Kert Viele (Director and Senior Statistical Scientist, Berry Consultants)
Hierarchical models are one of the workhorses of Bayesian statistics. As with any Bayesian method, the prior distribution can affect the performance of the method, particularly when the effective sample size of the data is small. In this talk we will review two case studies of calibrating a hierarchical model prior to achieve desirable trial operating characteristics. In the first example, a sponsor proposes a 16 arm study in a 4x2x2 structure. An additive model obtains much higher power than modeling all arms separately when the truth is in fact additive, but produces large biases when the truth is not additive. We generalize the additive model to include an additive base and arm specific deviation terms, whose prior is a hierarchical model. We will discuss how to calibrate this prior so that posterior tends toward additivity when the truth is additive, and avoids lack of fit when the truth is nonadditive. The second example is a basket trial with multiple groups. The original analysis pooled all the groups, but an outlying group was found and a hierarchical model was utilized as a sensitivity analysis. We will discuss calibrating this prior to include a desirable range of shrinkage possibilities.
SKYPE LINK: https://meet.lilly.com/natanegara_fanni/6T31HJQ8
Join by phone:
+1 (317) 277-1498 (United States) English (United States)
+1 (855) 545-5910 (United States) English (United States)