Machine Learning as a Data Driven Approach to Automate Multivariate Matching Methods

Abstract: This paper introduces machine learning to automate the coarsening choices in coarsened exact matching (CEM) as a monotonic imbalance bounding matching class. I suggest to replace the otherwise arbitrary multivariate stratification process with a binary classification tree. This way, I can minimise potential bias caused by subjective preferences. By using the LaLonde (1986) dataset, I systematically compare this novel approach with competing matching specifications, in particular arbitrary CEM and propensity score matching (PSM). While the automated CEM returns more accurate results than PSM, coarsening arbitrarily performs best in terms of reducing imbalance as well as in the post-matching causal estimation.