Extending the Search Space for Genetic Matching: ExpoMatch

• In Event: Political Methodology Virtual Poster Session

Abstract

Matching methods for causal inference have progressed tremendously in the last 45 years, but new approaches still keep appearing in the literature. Among them is Genetic Matching (Diamond & Sekhon, 2012) which improves covariate balance by trying different sets of confounder weights and comparing the balance of the resulting treatment and control groups. This method achieved robust results on many different datasets. We expand on GenMatch by introducing ExpoMatch, an extensive genetic algorithm that searches an even larger solution space by adding an exponent weight to each potential confounder (matching variable). This allows ExpoMatch to find all the possible matched sets that GenMatch can, as well as an extensive set of previously impossible matches, dominating GenMatch in theory (though more computationally expensive in practice). We prove these claims and illustrate their practical importance using simulated datasets with various properties, as well as on the canonical Lalonde (1986) data set. ExpoMatch is computationally extremely expensive, and so we also discuss ways of diagnosing scenarios in which the benefits of ExpoMatch (vs. GenMatch and other related methods) are the most compelling.

Authors

• Idriss Bennis

• Mikulas Plesak, Minerva Schools at KGI