Matching as Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal Inference

Abstract

The fast growing statistical literatures on matching methods in several disciplines offer the promise of causal inference without resort to the difficult-to-justify functional form assumptions inherent in commonly used parametric methods. However, these literatures also suffer from many diverse and conflicting approaches to estimation, uncertainty, theoretical analysis, and practical advice. In this paper, we propose a unified perspective on matching as a method of nonparametric preprocessing for improving parametric methods. This approach makes it possible for researchers to preprocess their data (such as with the easy-to-use matching software we offer with this paper) and then to apply whatever familiar statistical techniques they would have used anyway. Under our approach, instead of using matching to replace existing methods, we use it to make existing methods work better, such as by giving more accurate and considerably less model-dependent causal inferences.