3/24/2017

Post date: Mar 24, 2017 7:49:01 PM

Title: An Empirical Bayesian Approach to Subgroup Analysis and Precision Medicine.

Speaker: Judy Li, FDA

(Joint work with Wei-Chen Chen and John Scott)

How subgroups of patients react heterogeneously to treatment plays an important role in precision medicine. Interpreting the subgroup for the development of individualized treatment rules has been recommended and yet challenging. The risk of overlooking an important subgroup and making a decision based on a false discovery becomes very crucial. At the meantime, the limited sample size, multiplicity, lack of power, and information borrowing make the solutions not so easy. The Bayesian framework provides the ability to incorporate priors with a degree of skepticism, a natural framework for forming models with shrinkage and the possibility to form realistically complex models allowing synthesis of information from a variety of sources. However, under prior-data conflict, optimistic borrowing could be inappropriate. In this talk, we are going to introduce a Bayesian empirical meta-analytic-predictive (eMAP) prior approach which allows robust borrowing under the prior-data conflict situation. We will firstly introduce how the eMAP prior approach could be applied to borrow information from historical studies for the analysis of the new trial data. Then we will illustrate how the eMAP prior approach could be applied to borrow information among subgroups in the context of personalized medicine, particularly, on how to address the problems of subgroups with extreme high or low treatment effect. One of the challenges is that the eMAP prior approach requires a good estimation of the empirical parameter which is numerically infeasible in the multivariate setup. To address this challenge, we derived an analytical approximation for the multivariate marginal distribution under the normal mixture prior assumption. Both simulated data and real data are applied to illustrate our proposed approach.