Evaluation of Subset Matching Methods: Evidence from a Monte Carlo Simulation Study

Abstract

In the absence or infeasibility of experiments, matching methods have increasingly been used in making causal claims using observational data. This paper conducts a Monte Carlo simulation study, based on a household panel survey, to compare the performance of some widely used subset matching methods. The methods include the propensity score caliper matching, Mahalanobis distance matching, and coarsened exact matching. Comparisons were made in terms of the ability to reduce covariate imbalances, as well as effective recovery of the real treatment effect. Numerical results from our simulations provided evidence of coarsened exact matching outperforming the other methods. Our results also showed that, except for the Mahalanobis distance matching method, the efficiency of treatment effect estimates decreases with an increasing proportion of treated units.