How Arbitrary Cutoffs Isolate Cause and Effect in Observational Data

For a regression discontinuity design to yield a valid causal estimate, one strict condition must hold: the subjects cannot precisely manipulate their position around the cutoff. If a scholarship requires a test score of 85, and students can retake the exam until they hit exactly 85, the people who score 85.1 are fundamentally different from those who score 84.9.

They possess more resources, more persistence, or more time. But if the score is final and the cutoff is rigid, the student who scores 84.9 and the student who scores 85.1 are statistically indistinguishable. The boundary itself acts as a coin flip, assigning one to the treatment group and the other to the control group.[1][2]

This mechanism relies on an assignment variable—sometimes called a running variable—that determines eligibility for an intervention based on a strict numerical threshold. Because the individuals just barely missing the cutoff are nearly identical to those just barely making it, any sudden jump in outcomes at that exact boundary can be attributed entirely to the intervention.[2][5]

The National Bureau of Economic Research (NBER) published a definitive guide to the practice in 2007, cementing its role in modern econometrics. "The defining characteristic of the RD design is that the probability of receiving treatment changes discontinuously as a function of one or more underlying variables," the NBER authors write. This discontinuous jump allows researchers to isolate cause and effect in observational data, bypassing the need for a prospective trial.[1]

The British Medical Journal (BMJ) characterizes this approach as "randomization without controlled trials." In clinical settings, running a randomized controlled trial (RCT) is often expensive, time-consuming, or ethically impossible. If a new drug is believed to be highly effective, withholding it from a control group poses severe ethical dilemmas. Regression discontinuity solves this by analyzing the rules already in place, such as a clinical guideline that mandates treatment only for patients with a specific biomarker level.[4]

A classic application involves the 1,500-gram birth weight threshold. In many hospitals, infants born weighing less than 1,500 grams are automatically sent to the neonatal intensive care unit (NICU), while those weighing 1,501 grams receive standard care. By comparing the mortality rates of infants weighing 1,499 grams to those weighing 1,501 grams, researchers can measure the exact causal impact of NICU admission.[6]

Another widespread example is the Medicare eligibility age in the United States. At exactly 65 years and zero months, a massive portion of the population suddenly gains access to federal health insurance. Researchers use this sharp discontinuity to measure how insurance coverage affects everything from cancer screening rates to emergency room utilization, knowing that a person who is 64 years and 11 months old is biologically identical to someone who is 65.[3][4]

Methodologists divide these designs into two categories: sharp and fuzzy. In a sharp design, the probability of receiving treatment jumps instantaneously from 0 to 1 at the threshold. The Medicare age cutoff is a sharp design. In a fuzzy design, the threshold increases the probability of treatment, but compliance is not perfect. For instance, a test score might guarantee admission to a specialized school, but not every admitted student chooses to attend.[2][5]

The mathematical challenge in any threshold analysis lies in bandwidth selection. Researchers must decide how far away from the cutoff they should look. A narrow bandwidth—comparing only the 84.9 scores to the 85.1 scores—reduces bias because the subjects are incredibly similar. However, it also increases variance because it discards the vast majority of the dataset, leaving a tiny sample size.[1][6]

To prove that the binding constraint holds, econometricians rely on the McCrary density test. This statistical check examines the distribution of the assignment variable to see if there is an unnatural spike just above the cutoff. If a poverty alleviation program requires an income below $20,000, and the data shows a massive cluster of people reporting exactly $19,999, the density test fails. It proves subjects are manipulating their reported income to qualify.[2][6]

The second mandatory check is covariate continuity. Researchers must prove that other baseline characteristics—such as age, race, or prior education—do not suddenly jump at the threshold. If the students scoring 85.1 are suddenly much wealthier on average than those scoring 84.9, the threshold is likely aligned with a different structural advantage, invalidating the natural experiment.[1][5]

Despite the mathematical elegance of the method, its practical execution often falls short in emerging fields. A systematic review published in Epidemiology screened 859 studies applying regression discontinuity in health research. Of the 61 studies that met the strict inclusion criteria, the review found significant methodological gaps in how the core assumptions were tested and reported.[3]

The review revealed that 43% of the included clinical studies failed to report baseline covariate continuity checks, and many omitted the density test entirely. Without these verifications, the claim that the cutoff mimics a randomized trial remains an unproven assertion rather than a mathematical certainty.[3][7]

The MDRC, a prominent social policy research organization, issued practical guidelines in 2012 to standardize these checks. They emphasized that graphical analysis is just as important as the regression models themselves. Plotting the raw data points around the cutoff provides an immediate, intuitive visual check of whether the discontinuity is real or an artifact of the chosen polynomial function.[2]

As machine learning and massive administrative datasets become more accessible, the use of regression discontinuity is accelerating. Algorithms can now scan millions of medical records or tax filings to automatically detect arbitrary thresholds that researchers never knew existed. The limiting factor is no longer finding the data, but rigorously proving that the boundary cannot be gamed.[5][6]