Research on Matching Methods for Causal Inference in Experimental and Observational Studies
Overview
The estimation of causal effects is a central goal of social science research. In this project, we develop matching methods that can help empirical researchers conduct reliable and efficient causal inference in both experimental and observational studies.First, we clarify the misunderstandings commonly held by applied researchers about matching and propensity score methods. We introduce a general framework where matching methods can be considered as a preprocessing procedure that improves the robustness of parametric regression models. In this view, matching methods should not be thought as an alternative to regression techniques. We also address the other methodlogical issues such as balance checking and standard error calculation. An easy-to-use package is developed in the free software language R that implements various matching methods.Second, we show that matching methods can be useful for experimetal studies. In particular, matched-pair designs can recoup the loss of efficiency that is common in cluster randomized experiments. The matched-pair cluster randomized design provides a robust and efficient method for estimating treatment effects in the presence of interference among units. We apply our proposed methodology to the randomized evaluation of the Mexican universal health insurance program. This program evaluation is one of the largest such health policy experiments in the history.Third, we consider the relationships between matching methods and linear fixed effects estimators. Our analysis shows that fixed effects estimators, which are the primary workforce of applied social scientists for panel data and other analyses, can be shown to be equivalent to particular matching estimators. This analysis in turn provides us insights about the limitations of fixed effects estimators. Finally, we address these limitations by developing matching methods for the time-series cross-sectional data.
