Comparative Analysis of Matching Methods and a Multivariate Recursive Probit Model for Estimating the Treatment Effects on a Binary Outcome: A Simulation Study

In this research paper, a comparative analysis of the quality of estimates of the average treatment effects of matching methods and a multivariate recursive probit model in various situations of violation of the unconfoundedness assumption put forward by matching and the exclusion restriction required to be performed in systems of binary equations is carried out. On simulated data generated with different correlations between error terms in the selection and outcome equations, as well as between error terms and control variables, using matching based on the Mahalanobis distance, full matching and a multivariate probit model, the average treatment effects are estimated, provided that different sets of control variables are used in the evaluation process. The key conclusion of the work is that binary equation systems can be considered as a worthy alternative to matching methods with a high correlation between the error terms of the selection and outcome equations, whereas matching can be easily used instead of systems if the violation of the unconfoundedness assumption occurs within a low correlation of error terms of the selection and outcome equations. To demonstrate how the methods work on real data, an estimation of the average effect of higher education on the probability of obesity among Russian women is carried out. Based on the results obtained from simulated data, recommendations are formulated for the use of matching methods and a multivariate recursive probit model in applied econometric research. It is assumed that the results of the work can become the basis for the development of a formal test to verify the feasibility of the unconfoundedness assumption on real data, which further research can be devoted to.