Myunghyun Song

Research designs exploiting quasi-random assignment of cases to judges and other examiners are increasingly popular, often using examiner leniency as instruments. Interest is also growing in identifying causal effects with multiple unordered treatments, beyond the binary treatment framework. Identification can benefit from instruments that shift cases toward a single alternative without inducing selection into any other treatment. Examiner designs, however, do not directly provide such instruments. We construct them by developing a discrete choice model in which examiners choose the treatment they assess most favorably. Slutsky symmetry of treatment propensities helps identify the model's coefficients without exclusion restrictions, and we develop a semiparametrically efficient estimator of these coefficients. Predicted assessments provide instruments for nonparametrically identifying mean potential outcomes and pairwise treatment effects at margins between two treatments. We develop estimation and inference procedures for the treatment effects and establish their asymptotic properties. Revisiting a recent study of Virginia courts, we obtain different findings on how noncarceral and carceral convictions affect reoffending.