Online Program

Understanding and Estimating Treatment Effect Heterogeneity Using Adaptive Ensemble Methods (302967)

Keywords: heterogeneous treatment effect, ensemble method, adaptive method, interaction, effect modification, subgroup analysis

Health outcomes studies typically provide credible estimates of the average effect of a treatment on a population of interest. But patients, health care providers, and decisionmakers are often interested in how effects vary across sub-populations of patients—heterogeneous treatment effects. Estimation of treatment effect heterogeneity is challenging, even in the era of Big Data. To identify systematic variation in treatment response and separate it from simple random variation, a combination of assumptions about the data-generating process and information in the data itself is required. New methods have been introduced, but it is difficult to know in advance whether a given method will perform well for a particular data set. We will show how ensemble methods, based on weighted averages of estimates from individual models, can be applied to accurately estimate heterogeneous effects. We will examine the basis for the superior performance of these methods, using both simulated and real data from a recent health outcomes study to demonstrate how this approach can be used in exploratory analysis of heterogeneous treatment effects.