Machine learning methods for heterogenous area-specific treatment effect estimation

Abstract

Small area estimation (SAE) methodology is popular for estimating parameters at the level of subpopulations with small sample sizes. It is often of interest for statistical offices to quantify the effect of an intervention at the small area level, for example, the effect of a new policy targeting unemployment across counties in a region. We develop new procedures for the estimation of heterogeneous area-specific average treatment effects at the small area level in observational studies. In particular, we use nonparametric machine learning (ML) methods which are ubiquitous in causal inference as they excel in terms of modelling flexibility and predictive abilities. We compare the empirical performance of our new estimators with the parametric alternatives. In addition, we provide an easy to follow set of instructions which indicate how different causal estimators should be used in SAE practical applications.