Keywords: Subgroup analysis, clustering, prognostic modeling
Motivated by a growing interest in personalized medicine, over the past 10 years the biostatistics literature has grown rapidly on the topic of understanding treatment effect heterogeneity in clinical trials. Topics have included: data mining and machine learning techniques to identify promising subgroups, general tools to better understand and characterize heterogeneity and approaches focused on setting treatment decision rules. Unfortunately, this literature has had less impact in practice and remains somewhat theoretical. One reason for this is that clinical trials are typically designed to detect population average treatment effects. In addition, primary outcomes can be crude general measures of response composed from many components.
With this background in mind, in this talk we discuss a different approach, where data from multiple clinical research is broken down into its most basic components allowing a more sophisticated grouping of patients or modeling of multiple outcomes over time. The presentation will begin with a review of previous research and then an example of a typically primary outcome measure that is likely to make the study of treatment effect heterogeneity challenging. Next we will show how breaking the primary end-point down into basic components can improve an understanding of treatment effect. Finally, we will look at formal modeling techniques that can help in the understanding of treatment effect heterogeneity