On Extrapolation of Treatment Effects in Multiple-Cutoff Regression Discontinuity Designs

We investigate how to learn treatment effects away from the cutoff in multiple-cutoff regression discontinuity designs. Using a microeconomic model, we demonstrate that the parallel-trend type assumption proposed in the literature is justified when cutoff positions are assigned as if randomly and the running variable is non-manipulable (e.g., parental income). However, when the running variable is partially manipulable (e.g., test scores), extrapolations based on that assumption can be biased. As a complementary strategy, we propose a novel partial identification approach based on empirically motivated assumptions. We also develop a uniform inference procedure and provide two empirical illustrations.