A Fundamental Measure of Treatment Effect Heterogeneity
In this paper we offer an asymptotically efficient, non-parametric way to assess treatment effect variability via the conditional average treatment effect (CATE) which is a function of the measured confounders or strata, giving the average treatment effect for a given stratum. We can ask the two main questions of the CATE function: What are its mean and variance? The mean gives the more easily estimable and well-studied average treatment effect whereas CATE variance measures reliability of treatment or the extent of effect modification. With the knowledge of CATE variance and hence, CATE standard deviation, a doctor or policy analyst can give a precise statement as to what an individual patient can expect, which we distinguish as clinical effect heterogeneity. We can also assess how much precision in treatment can be gained in assigning treatments based on patient covariates. Through simulations we will verify some of the theoretical properties of our proposed estimator and we will also point out some of the challenges in estimating CATE variance, which lacks double robustness. We will provide a demonstration, featuring software in the targeted learning framework as well as instructions for reproducing all the results here-in.
READ FULL TEXT