Regression discontinuity is a research method that estimates the causal effect of a treatment or policy by exploiting a cutoff rule. Whenever eligibility for something is determined by a score, an age, or a threshold on some measurable scale, researchers can compare outcomes for people who just barely qualified to those who just barely missed out. Because these two groups are nearly identical in every other way, the difference in their outcomes reveals the true effect of the treatment itself.
The method is considered one of the most credible ways to measure cause and effect outside of a randomized experiment. It shows up across economics, public health, education, and political science, and understanding how it works helps you evaluate the quality of research claims you encounter.
How the Cutoff Creates a Natural Experiment
Every regression discontinuity design has two essential ingredients: a running variable and a threshold. The running variable is the continuous score or measurement that determines who receives a treatment. The threshold (or cutoff) is the specific value on that score where assignment changes. People above the cutoff get the treatment; people below it don’t, or vice versa.
Think of a scholarship program that awards funding to every student who scores 80 or above on a standardized test. A student who scored 80.1 gets the scholarship. A student who scored 79.9 doesn’t. These two students are essentially the same in terms of ability, preparation, and background. The tiny difference in their scores is largely a matter of luck. So if the scholarship recipients go on to have better outcomes (higher college graduation rates, for example), you can attribute that difference to the scholarship rather than to pre-existing differences between the groups.
The causal effect is measured as the “jump” or discontinuity in outcomes right at the cutoff. Researchers estimate what the average outcome looks like just above the threshold and just below it, then calculate the gap. If outcomes change smoothly along the running variable but suddenly jump at the cutoff, that jump is the treatment effect.
Sharp vs. Fuzzy Designs
Regression discontinuity comes in two varieties, depending on how strictly the cutoff rule is followed.
In a sharp design, the cutoff is absolute. Everyone on one side receives the treatment, and no one on the other side does. There are no exceptions. The scholarship example above is a sharp design: score 80 or above, you get the money; score below 80, you don’t. Because the rule is followed perfectly, the analysis is straightforward.
In a fuzzy design, the cutoff changes the probability of receiving treatment but doesn’t guarantee it. Some people above the cutoff don’t take up the treatment, and some below it manage to receive it anyway. This happens frequently in real-world policy. A government program might target households below a certain income level, but not every eligible household enrolls, and some ineligible households find alternative routes in. In fuzzy designs, researchers use the cutoff as a kind of nudge that shifts the likelihood of treatment, then adjust their estimates to account for imperfect compliance. The analysis is more complex, but the core logic remains the same.
Why the Effect Is “Local”
One of the most important things to understand about regression discontinuity is that its findings apply only to people near the cutoff. The estimated effect is called a local average treatment effect, and it describes what happens for individuals whose scores are close to the threshold, not for everyone in the population.
This matters for interpretation. If a study finds that a tutoring program boosted test scores using a regression discontinuity at a cutoff of 50 points, that result tells you about students who scored around 50. It doesn’t necessarily tell you what the program would do for students who scored 20 or 90. Those students might respond very differently to the same intervention.
Researchers face a tradeoff when deciding how close to the cutoff they look. A narrow window around the threshold gives you the most credible comparison (the people on either side are nearly identical), but you end up with fewer observations and less statistical power. A wider window gives you more data to work with, but increases the risk that people far from the cutoff differ in meaningful ways. Even with wider windows, though, the result is still interpreted as a local effect concentrated around the threshold. This tradeoff means regression discontinuity studies generally require large sample sizes to produce reliable results.
What Could Go Wrong
The key assumption behind regression discontinuity is continuity: if the treatment didn’t exist, outcomes would change smoothly across the cutoff with no sudden jump. This assumption breaks down when people can manipulate the running variable to land on their preferred side of the threshold.
Imagine a program that provides benefits to anyone with a blood pressure reading above a certain value. If patients or doctors know about the cutoff, they might time measurements, change behavior, or otherwise game the system to push readings above or below the threshold. When this happens, the people just above and just below the cutoff are no longer comparable, and the entire design falls apart.
Researchers have a standard check for this problem, known as a density test. The idea is intuitive: if people are gaming the cutoff, you’d expect to see a suspicious pile-up of observations on the desirable side. The density test looks for exactly that, checking whether the number of people just above the cutoff is suspiciously different from the number just below. If the distribution of the running variable is smooth through the cutoff, it suggests no one is manipulating their way across it.
Another routine check involves looking at whether pre-existing characteristics (age, gender, income, health status) change smoothly across the cutoff. If those background traits show a sudden jump at the threshold, something other than the treatment may be driving the results.
A Real-World Example: Medicare at Age 65
One of the clearest applications of regression discontinuity in health research uses the fact that Americans become eligible for Medicare at exactly age 65. A person who is 64 years and 11 months old is essentially identical to someone who just turned 65, but one has Medicare and the other likely doesn’t. That sharp eligibility rule creates a natural experiment.
A study published in The BMJ used this approach to examine how Medicare eligibility affected care for trauma patients. At age 65, the share of patients with Medicare coverage jumped by 9.6 percentage points, confirming the cutoff was real and meaningful. The researchers then looked at what changed in how these patients were treated and how they fared.
The findings were nuanced. Hospital stays shortened by about a third of a day (roughly 5%) after the Medicare cutoff, while discharges to nursing homes increased by 1.56 percentage points and discharges to home decreased by about 2 percentage points. This suggests Medicare shifted where patients went after the hospital, not necessarily how they were treated inside it. Emergency treatments, blood transfusions, ICU admissions, ventilator use, and mortality showed no meaningful change at the age-65 threshold. The actual medical care patients received in the hospital was largely the same on both sides of the cutoff.
This example illustrates both the power and the limitations of the method. It can isolate the effect of insurance coverage from the effect of aging, something that would be impossible with a simple comparison of insured and uninsured patients. But the findings only speak to the experience of 65-year-olds, not to what Medicare might do for a 45-year-old or an 80-year-old.
Why Researchers Value This Method
Regression discontinuity sits in a sweet spot between randomized experiments and ordinary observational studies. It doesn’t require researchers to randomly assign people to groups, which is often impractical or unethical. But unlike a simple comparison of treated and untreated people, it has a built-in mechanism for ensuring the groups being compared are genuinely similar. The cutoff creates that similarity automatically, as long as people can’t precisely control which side they land on.
The tradeoff is generalizability. Because the effect is local to people near the threshold, you can’t automatically extend the findings to a broader population. And because adequate power requires large datasets, the method works best in settings where administrative records or large surveys provide enough observations near the cutoff. When those conditions are met, regression discontinuity produces some of the most believable causal evidence available outside of a controlled trial.