Subgroup identification for precision medicine: A comparative review of 13 methods

Natural heterogeneity in patient populations can make it very hard to develop treatments that benefit all patients. As a result, an important goal of precision medicine is identification of patient subgroups that respond to treatment at a much higher (or lower) rate than the population average. Despite there being many subgroup identification methods, there is no comprehensive comparative study of their statistical properties. We review 13 methods and use real‐world and simulated data to compare the performance of their publicly available software using seven criteria: (a) bias in selection of subgroup variables, (b) probability of false discovery, (c) probability of identifying correct predictive variables, (d) bias in estimates of subgroup treatment effects, (e) expected subgroup size, (f) expected true treatment effect of subgroups, and (g) subgroup stability. The results show that many methods fare poorly on at least one criterion. This article is categorized under: Technologies > Machine Learning Algorithmic Development > Hierarchies and Trees Algorithmic Development > Statistics Application Areas > Health Care Subgroup (in green) for breast cancer data; sample sizes and estimated treatment effects (log relative risks) beside and below nodes