Regression discontinuity: what a sharp cutoff reveals

Regression discontinuity design compares people who fall just above and just below a sharp, arbitrary cutoff, a legal age, a test score, an income threshold, on the theory that people right next to the line are otherwise alike. Whatever difference in outcomes shows up exactly at the cutoff is attributed to whatever changes there, since nothing else about a person changes the instant they cross an age or a score.

The clearest example is the effect of the United States' minimum legal drinking age of 21. Nothing about a person changes biologically the day they turn 21, except their legal access to alcohol, so any jump in outcomes exactly at that birthday can be attributed to drinking rather than to age itself. A 2007 study built around the age-21 cutoff found that legal access to alcohol produced a 21% increase in the number of days people drink, and that this jump in drinking was followed by a discrete 9% increase in the overall mortality rate right at age 21, driven by a 14% increase in motor vehicle deaths, a 30% increase in alcohol-related deaths and overdoses, and a 15% increase in suicides.

Two traps follow. First, the design only identifies an effect exactly at the cutoff, for people close to that age or score; it says nothing directly about people far from the line, who may respond differently. Second, a regression discontinuity result depends on nothing else jumping at the same cutoff. The drinking-age study held up in part because researchers could rule out other 21st-birthday changes competing for the explanation, and any such result becomes suspect if a second policy, a different eligibility rule, or a change in how the outcome is measured happens to kick in at the identical threshold.

So when a headline reports an effect tied to a specific age, score, or income cutoff, check whether the comparison is really built around people just on either side of that line, and whether anything else changes at that same threshold. A sharp, isolated cutoff with nothing else moving at the same point is what makes the design credible. Pair a regression discontinuity claim with a look at how narrow the window around the cutoff was and whether the outcome jumps cleanly, or just drifts, right at the line.