Robust Subgroup Analysis in Clinical Trials and Precision Medicine

摘要：Subgroup analyses have been an important part of the analysis of clinical trial. However, they are commonly over-interpreted and can lead to further misguided research or worse to suboptimal patient care. Today Big Data including molecular and cellular markers only make the problem worse, which is one of the big challenges in converting the vision of precision medicine into reality. In this talk, I will first present our findings on subgroup identification with omics data using novel statistical learning and functional models. Then I will propose a robust method via a semiparametric model to formally test if sub-groups with differential treatment effects in clinical trials exist and if so to classify subjects into subgroups, a critical issue in precision medicine. In more general terms, the method applies to testing if there is a group of subjects with differential outcome. The model is formulated as a geometrical mean of a parametric and a nonparametric component with themixture proportion as a parameter to be estimated from the data. The profile likelihood approach is used for model estimation. The profile likelihood ratio and score statistics are used to test the existence of subgroups. If existence of subgroups is confirmed, we use Neyman-Pearson rule to classify each subjects to one of the subgroups, so that the misclassification error for the treatment favored group is controlled by pre-specified criterion and is minimized for the other subgroup. The large sample properties of the procedure are derived and finite sample performance is studied by simulation. The work is in collaboration with Drs. A. Yuan, and H Fang.