A double-differences approach to inference in regression discontinuity designs with a discrete running variable

We present a new approach to nonparametric inference in regression discontinuity designs where the running variable is discrete and may have few support points. The approach is based on an intuitive and comparatively weak identification assumption: In the absence of treatment, the difference in population mean outcomes at the cutoff is not the largest nor the smallest among differences at any two adjacent support points. Test statistic and confidence interval are constructed from a set of t-statistics based on double differences of the outcome variable, eliminating the need to estimate a conditional expectation function. Simulation evidence suggests that statistical power compares favorably to existing methods. We illustrate our approach by revisiting two studies from the economics of education literature.