Abstract: We present a new approach to non-parametric inference in regression discontinuity designs where the running variable is discrete and may have few support points. We make the key identifying assumption that in the absence of treatment, the difference in outcome (population) means at the cut-off is not the largest, nor the smallest, of the differences at any two adjacent support points. Test statistic and confidence interval are based on a set of t-ratios, making for easy implementation. Complementing existing methods, our approach is valid even when local randomization fails, and does not require assumptions about the smoothness of the conditional expectation function.