How Local Randomization Near a Cutoff Isolates Causal Effects in Observational Data

When a medical guideline dictates that patients with a blood pressure of 140 mm Hg receive medication while those at 139 mm Hg do not, the arbitrary boundary dictates treatment. This sharp division creates a natural laboratory for researchers. By comparing the individuals immediately adjacent to the threshold, statisticians isolate the exact impact of the drug without running a multi-million-dollar randomized controlled trial, because the patients on either side of that one-point divide are biologically indistinguishable.

This mechanism is the foundation of the Regression Discontinuity Design (RDD). According to a foundational 2007 guide published by the National Bureau of Economic Research, this approach leverages the fact that subjects cannot precisely manipulate their position around a strict cutoff. As authors David Lee and Thomas Lemieux note in the paper, "the RD design is often thought to be the most credible non-experimental research design" because it closely mimics a true experiment.[1]

Historically, researchers interpreted RDD through a continuity framework, assuming that potential outcomes change smoothly across the threshold and using regression lines to extrapolate to the boundary. However, a distinct methodological branch has formalized the "local randomization" framework. As detailed in a 2018 review published on arXiv, this approach explicitly assumes that within a very narrow window around the cutoff, the assignment of treatment is as good as random.[4]

The distinction matters fundamentally for how evidence is weighed and processed. In the continuity approach, researchers use data from far away from the cutoff to fit their models. The local randomization approach, outlined by methodologists at Princeton University and Cambridge University Press, discards the distant data entirely. It treats the narrow band—for example, test scores falling strictly between 84.5 and 85.5—as a pure lottery, analyzing it exactly as one would analyze a clinical trial.[3][6]

To execute this, analysts rely on specific computational tools designed to find that perfect window. The `rdlocrand` package, detailed in the Stata Journal in 2016 and maintained as an open-source repository on GitHub, provides the algorithmic infrastructure. It allows researchers to systematically select the bandwidth where baseline characteristics—like age, income, or prior education—are statistically identical on both sides of the divide.[5][7]

Finding that window requires a delicate mathematical balance. If the window is too wide, the subjects on either side are no longer identical; the student scoring 90 is fundamentally different from the student scoring 80. If the window is too narrow, the sample size drops to a level where statistical power vanishes, making it impossible to distinguish a true effect from random noise with a standard p-value of 0.05.

A 2017 analysis published by Emerald Group Publishing explored this exact trade-off when interpreting RDD as a local experiment. The authors demonstrated that the validity of the causal claim rests entirely on the inability of subjects to sort themselves across the boundary. If a student knows the scholarship cutoff is 85 and can retake the test until they hit exactly 85.1, the randomization assumption collapses entirely.[8]

To detect this manipulation, researchers deploy density tests, most notably the McCrary sorting test. If the distribution of the running variable—the score or metric determining treatment—shows a massive spike immediately above the cutoff and a hollowed-out valley immediately below it, it signals that individuals have gamed the system. In such cases, the local randomization framework cannot be applied, and any causal claims are void.

When the assumptions hold, the results are remarkably robust and have reshaped multiple fields. In political science, local randomization is frequently used to study the incumbency advantage by examining elections decided by less than a 1 percent margin. The municipality where a candidate won by 50.1 percent is virtually identical to the one where they lost with 49.9 percent, isolating the pure effect of holding office from the underlying political leanings of the town.

The framework also demands rigorous covariate balance checks before any conclusions are drawn. Before looking at the final outcome, researchers must prove that the treatment and control groups within the window share identical pre-treatment traits. If the algorithm selects a window where the average age is 42 on one side and 48 on the other, the local randomization assumption is violated, and the window must be narrowed.

The evolution of these methods reflects a broader shift in empirical economics and data science toward design-based inference. Rather than relying on complex, opaque models to adjust for hundreds of confounding variables, the local randomization approach forces the researcher to find a transparent, naturally occurring lottery in the data, making the assumptions visible and testable.[2]

The external validity of these findings remains strictly bounded by the mathematics of the design. The treatment effect calculated at the cutoff applies only to the individuals near that cutoff. Whether a medication works for a patient with a blood pressure of 180 mm Hg, or a scholarship helps a student scoring 95, cannot be answered by the discontinuity.

The precision of the local experiment is achieved precisely by sacrificing the ability to generalize globally, leaving researchers to weigh the absolute certainty of a narrow claim against the uncertainty of a broad one.[9]