Layered policy analysis in program evaluation using the marginal treatment effect

Abstract

This paper proposes a unified approach to derive sharp bounds on conventional policy parameters when the instrumental variables (IVs) are potentially invalid. Using a vine copula approach, we propose a novel characterization of the identified sets for the marginal treatment effect (MTE) and the policy-relevant treatment effect (PRTE) parameters. Our method has various advantages: First, it explicitly demonstrates how imposing different IV-related assumptions with different credibility levels affects the MTE and PRTE’s identified set. Second, it provides a basis for testing model specifications and hypotheses about various imperfect IV-related assumptions. Third, it provides a tractable way to inform policy choices in the presence of uncertainty of the validity of identifying assumptions. Our approach enlarges the MTE framework’s scope by showing how it can be used to inform policy decisions even when valid instruments are not available.