A Model-Free Measure of Variable Importance for Sources of Treatment Effect Heterogeneity

• In Event: Causal Inference Meets Machine Learning

Brief Overview

We propose a leave-one-covariate-out approach with honest inference to rank the importance of variables as a source of heterogeneous treatment effects. This approach assuages biases toward continuous variables or variables with multiple cuts.

Authors

• Zenobia Chan, Princeton University

• Marc Thomas Ratkovic, Princeton University