Estimating Heterogeneous Treatment Effect on Multivariate Responses Using Random Forests

Estimating the individualized treatment effect has become one of the most popular topics in statistics and machine learning communities in recent years. Most existing methods focus on modeling the heterogeneous treatment effects for univariate outcomes. However, many biomedical studies are interested in studying multiple highly correlated endpoints at the same time. We propose a random forest model that simultaneously estimates individualized treatment effects of multivariate outcomes. We consider a popular study design where covariates and outcomes are measured both before and after the intervention. The proposed model uses oblique splitting rules to partition population space to the neighborhood that experiences distinct treatment effects. An extensive simulation study suggests that the proposed method outperforms existing methods in various nonlinear settings. We further apply the proposed method to two nutrition studies investigating the effects of food consumption on gastrointestinal microbiota composition and clinical biomarkers. The method has been implemented in a freely available R package MOTE.RF at https://github.com/boyiguo1/MOTE.RF.

Automated summary

Estimating individualized treatment effects is a popular research topic in statistics and machine learning communities. We propose a random forest model that simultaneously estimates individualized treatment effects for multiple correlated outcomes. Extensive simulation studies and applications show that our method outperforms existing methods in nonlinear settings.