Subgroup Mixable Inference in Personalized Medicine, with an Application to Time-to-Event Outcomes

Abstract: Measuring treatment efficacy in mixture of subgroups from a randomized clinical trial is a fundamental problem in personalized medicine development, in deciding whether to treat the entire patient population or to target a subgroup. We show that some commonly used efficacy measures are not suitable for a mixture population. We also show that, while it is important to adjust for imbalance in the data using least squares means (LSmeans) (not marginal means) estimation, the current practice of applying LSmeans to directly estimate the efficacy in a mixture population for any type of outcome is inappropriate. Proposing a new principle called {\em subgroup mixable estimation}, we establish the logical relationship among parameters that represent efficacy and develop a general inference procedure to confidently infer efficacy in subgroups and their mixtures. Using oncology studies with time-to-event outcomes as an example, we show that Hazard Ratio is not suitable for measuring efficacy in a mixture population, and provide alternative efficacy measures with a valid inference procedure.