Distinguishing the Confounding Factors: Policy Evaluation, High-Dimension and Variable Selection

Variable selection is an important question for policy evaluation when identification of the treatment effect relies on a conditional-on-observables strategy. Recent advances in variable selection methods, such as the Lasso, have been deemed useful for the econometrics of policy evaluation. The Lasso approach focuses on the computational feasibility of exhaustive model selection borrowing from procedures developed in a high-dimensional context. However, it has been seldom applied in policy evaluation works because it raises other di culties such as the choice of a parameter that sets the trade-o between t and sparsity. Two Lasso-based treatment e ect estimators are reviewed and compared on an empirical application, on which they perform well. This paper also illustrates the pitfalls of variable selection in a policy evaluation context.