Testing for and Characterizing Treatment Effect Heterogeneity Under the Neyman-Rubin Potential Outcomes Framework

• Activity Number: 526
• Type: Topic Contributed
• Date/Time: Wednesday, August 7, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Luke Miratrix*+ and Avi Feller and Peng Deng
• Companies: Harvard University and Harvard University and Harvard University
• Keywords: randomized clinical trials ; causal analysis ; permutation tests ; bootstrap tests

Abstract

Most standard analyses of experimental data assume-either implicitly or explicitly-that treatment effects are constant across units. Understanding treatment effect variation, i.e., how a treatment differentially impacts subjects, is a needed extension because it allows for a richer understanding of an intervention's overall causal effect. This is particularly relevant for personalized medicine, where we wish to only treat those who benefit from a drug, or social interventions, where there are finite resources limiting the number of treated units. We design and explore nonparametric tests inspired by resampling methods in conjunction with the Neyman-Rubin potential outcomes framework to test for treatment effect heterogeneity. In particular, we focus on using the Kolmogorov-Smirnoff statistic in conjunction with different methods for estimating the mean effect, which is a nuisance parameter in our context. These tests can then be extended to allow for more general models of treatment effect variation, such as effects that vary across strata but are constant within each stratum.