Regression discontinuity design
Compare units just above and just below a sharp cutoff for treatment.
What it means
Regression discontinuity exploits a rule that assigns treatment based on whether a continuous variable crosses a threshold, comparing units narrowly on either side of the cutoff. Near the boundary, units are assumed to be similar in all respects except treatment status, so a jump in the outcome at the threshold identifies the causal effect — a kind of local randomization created by the rule. Designs are 'sharp' when the cutoff perfectly determines treatment and 'fuzzy' when it only shifts the probability. The estimate is local to the threshold, which sharpens internal validity at some cost to generalizability away from the cutoff.
Examples
Students scoring just above a scholarship cutoff are compared with those just below to estimate the award's effect on graduation.
Researchers compare people just before and just after their twenty-first birthday to estimate what legal drinking does to death rates, since a day either side changes little else.
Every pupil scoring below forty on the maths test is put in summer school; comparing those who scored thirty-nine with those who scored forty-one isolates what the extra teaching did.
First described in Thistlethwaite & Campbell (1960); revived by Hahn, Todd & van der Klaauw (2001).