Regression discontinuity design (RDD)

Regression discontinuity design exploits a sharp, known rule that assigns treatment based on whether a running variable crosses a cutoff: a credit score above 680 gets an offer, an account with more than 100 monthly active users unlocks a feature, a test score above a threshold earns admission. Units just above and just below the cutoff are assumed to be otherwise similar, so any jump in outcomes right at the threshold is attributed to the treatment.

The design is credible to the extent the cutoff is not gameable: if people can manipulate their score to land just above the threshold, the comparison breaks down, so analysts check for suspicious bunching in the running variable near the cutoff before trusting the estimate. The resulting effect is local, it describes units near the cutoff, and does not necessarily generalize to units far from it.

RD is valuable precisely where a true randomized experiment is impossible but a business or policy rule already creates something close to random assignment near its boundary. It is a narrower tool than difference-in-differences or propensity score matching, but when its assumptions hold it can approach the credibility of a randomized A/B test without ever running one.