There has been a recent increase on research focusing on partial identification of average treatment effects in the program evaluation literature. In contrast with traditional point identification, partial identification approaches derive bounds on parameters of interest based on relatively weak assumptions. Thus, they deliver more credible results in empirical applications. This dissertation extends Instrumental Variable (IV) methods in the program evaluation literature by partially identifying treatment effects of interest when evaluating a program or intervention.;An influential approach for studying causality within the IV framework was developed by Imbens and Angrist (1994) and Angrist, Imbens and Rubin (1996). They show that, when allowing for heterogeneous effects, IV estimators point identify the local average treatment effect (LATE) for compliers, whose treatment status is affected by the instrument. This dissertation advances the current IV literature in two important ways. First, inspired by the common criticism that LATE lacks external validity, this dissertation derives sharp nonparametric bounds for population average treatment effects (ATE) within the LATE framework. Second, the dissertation extends the LATE framework to bound treatment effects in the presence of both sample selection and noncompliance. Even when employing randomized experiments to evaluate programs—as is now common in economics and other social science fields—assessing the impact of the treatment on outcomes of interest is often made difficult by those two critical identification problems. The sample selection issue arises when outcomes of interest are only observed for a selected group. The noncompliance problem appears because some treatment group individuals do not receive the treatment while some control individuals do. The dissertation addresses both of these identification problems simultaneously and derives nonparametric bounds for average treatment effects within a principal stratification framework. More generally, these bounds can be employed in settings where two identification problems are present and there is a valid instrument to address one of them. The bounds derived in this dissertation are based on two sets of relatively weak assumptions: monotonicity assumptions on potential outcomes within specified subpopulations, and mean dominance assumptions across subpopulations.;The dissertation employs the derived bounds to evaluate the effectiveness of the Job Corps (JC) program, which is the largest federally-funded job training program for disadvantaged youth in the United States, with the focus on labor market outcomes and welfare dependence. The dissertation uses data from an experimental evaluation of JC. Individuals were randomly assigned to a treatment group (whose members were allowed to enroll in JC) or to a control group (whose members were denied access to JC for three years). However, there was noncompliance: some individuals who were assigned to participate in JC did not enroll, while some individuals assigned to the control group did. The dissertation addresses this noncompliance issue using random assignment as an IV for enrollment into JC. Concentrating on the population ATE, JC enrollment increases weekly earnings by at least ;The dissertation closes by pointing out that a similar analytic strategy to the one used in this dissertation can be used to address other problems, for example, to bound the ATE when the instrument does not satisfy the exclusion restriction, and to derive bounds on the part of the effect of a treatment on an outcome that works through a given mechanism (i.e., direct or net effects) in the presence of one identification issue (e.g., noncompliance).
展开▼
