A Subgroup Identification Method with Interaction Filtering and Quantitative Criteria

Subgroup identification has always been of great interest among the many functions and applications of statistical learning. In the pharmaceutical area, it is desirable to find a subgroup with enhanced treatment effect so that we can efficiently lower the number of patients required for a trail and improve the success rate of drug development projects. A more familiar name for this application is called personalized medicine, which has drawn great attention recently. A majority of work has been done regarding the personalized medicine with their pros and cons. Some methods focus on the detection of subgroup effects but do not provide any way to select patients. Some methods have a tree regression style and provide a detailed picture of each patient’s performance, they are hence more optimized for prediction rather than subgroup identification. Some methods try to “maximize the effect” in the training dataset but tend to be too greedy. There are also methods trying to build a score system to stratify the patients. In this dissertation, we propose a subgroup identification method with interaction filtering and quantitative criteria. More specifically, the method consists of two steps. Step 1 can select relevant interactions related to the individual treatment benefit without modeling the main effects. Step 2 can select a desired subgroup based on some quantitative criteria without relying on any specific model. The proposed method works for both the continuous and the survival response, and is shown to have a better performance than some popular existing methods.