Keywords: propensity score matching, matching, methods, causal
Background: Propensity score matching (PSM) is often employed to address selection bias. While PSM leads to similar 1st and 2nd moments for matched variables, the underlying distribution can remain different between groups. We report a novel PSM variation, which addresses the non-normal distribution of health expenditures. Methods: Typically, case and control observations are only considered for matching if they are in within a specified caliper distance of a PS derived from logistic regression. To obtain the matched population, we utilize a weighted bipartite graph; potential matches are assigned an edge weight equal to the absolute value of the difference in cost. The maximum weight matching, obtained via the Hungarian algorithm, yields a cohort with optimally matched cost distributions. Results: Using a naïve greedy nearest neighbor algorithm, underlying cost distributions for the matched cohorts differed (median: $2,947 vs $2,342) vs after applying our method (median: $2,947 vs $2,836). Conclusion: Cohorts well-matched on prior medical cost is critical in assessment of economic outcomes. Given the distribution of costs, standard matching techniques may not be sufficient.