Three essays on policy evaluation

Abstract

This thesis comprises three chapters on policy evaluation. The first chapter discusses how to test IV validity in the marginal treatment effects model. The second chapter performs the counterfactual decompositions using standardization techniques. The third chapter estimates the variance estimator of the matching estimator under spatial dependence. These three chapters discuss causality, specification, and efficiency in the framework of policy evaluation, which are three important topics in econometrics. The first chapter develops a specification test for IV validity assumptions in marginal treatment effects models. The IV validity assumptions are intractable directly, but they have the strongest testable implication involving two shape restrictions on the conditional joint density function of the outcome and treatment on the propensity score. Our test is based on transforming the shape restrictions into equality restrictions using the LCM operator. Here, the statistics’ null asymptotic distribution is approximated by a newly proposed easyto- implement bootstrap procedure. We propose a Monte-Carlo experiment that examines the finite sample performance. The second chapter proposes methods to perform counterfactual decompositions using standardization techniques based on a partition. We provide a data-driven algorithm to partition data into classes that receives the same predictions in the propensity score or the conditional distribution function. The distinguishing feature of our approach is that it could adapt to the various variable types and many covariates. We apply the method to analyze the gender gap of the Spanish labor market during the period 2004-2017. Results suggest that the occupation categories’ share plays an important role in decreasing the gender gap in employment. The third chapter discusses how to evaluate regional policy under spatial dependence. Spatial dependence among local units leads to size distortions in regional policy evaluation. Our paper proposes a consistent spatial heteroskedasticity and autocorrelation (SHAC) variance estimator for the matching estimator. We also consider two valid bootstrap procedures for the matching estimator adjusted for spatial dependence. The finite sample performance of these approaches is studied by Monte Carlo experiments. Our methods are applied to revisit one immigration policy on German local labor markets’ unemployment rate.