Identification and Inference of Partial Effects in Sharp Regression Kink Designs

Abstract: The partial effect refers to the impact of a change in a target variable D on the distribution of an outcome variable Y . This study examines the identification and inference of a wide range of partial effects at the threshold in the sharp regression kink (RK) design under general policy interventions. We establish a unifying framework for conducting inference on the effect of an infinitesimal change in D on smooth functionals of the distribution of Y, particularly when D is endogenous and instrumental variables are unavailable. This framework yields a general formula that clarifies the causal interpretation of numerous existing sharp RK estimands in the literature. We develop the relevant asymptotic theory, introduce a multiplier bootstrap procedure for inference, and provide practical implementation guidelines. Applying our method to the effect of unemployment insurance (UI) benefits on unemployment duration, we find that while higher benefits lead to longer durations, they also tend to reduce their dispersion. Furthermore, our results show that the magnitude of the partial effect can change substantially depending on the specific form of the policy intervention.