Regression discontinuity design is a method of causal inference that uses a threshold-based assignment rule to estimate the effect of an intervention. When treatment assignment or the probability of receiving treatment changes abruptly at a known cutoff in an observed variable, outcomes immediately on either side of that cutoff can be compared. Under appropriate assumptions, a discontinuity in outcomes identifies a causal effect rather than merely an association. The method is widely used in econometrics, education research, political science, and public policy evaluation.
Basic principle
Many institutions allocate benefits, obligations, or opportunities according to numerical thresholds. A scholarship may require an examination score above a specified level; a program may restrict eligibility to households below an income ceiling; or a regulation may apply only to firms exceeding a size threshold.
The variable governing assignment is called the running variable, assignment variable, forcing variable, or score. The threshold is the cutoff. Treatment denotes the intervention or exposure of interest, not necessarily a medical treatment.
The central idea is that units just above and just below a cutoff may be similar in their underlying characteristics, even though the assignment rule treats them differently. For example, students scoring 79.9 and 80.1 may differ little in academic preparation, but a scholarship awarded at 80 creates a sharp difference in eligibility.
This does not mean that all eligible and ineligible units are comparable. High-scoring students may differ substantially from low-scoring students. Regression discontinuity instead isolates the comparison at the boundary, separating the jump associated with assignment from the smooth relationship between the score and outcomes.
The design is often classified as a quasi-experimental method or a natural experiment. Unlike a randomized controlled trial, however, threshold assignment generally is not randomized. Its credibility depends on explicit assumptions about behavior and outcomes near the cutoff.
Formal framework
Let be the running variable for unit , the cutoff, treatment status, and the observed outcome. Define threshold eligibility as
where is an indicator function. The convention that treatment increases above the cutoff is arbitrary; some programs assign treatment below it.
Using the potential outcomes framework, let and denote outcomes under treatment and no treatment. The observed outcome is
Only one potential outcome is observed for each unit. Regression discontinuity addresses this missing-counterfactual problem by using behavior immediately across the cutoff to infer the otherwise unobserved outcome at the boundary.
Sharp design
In a sharp regression discontinuity design, treatment is determined completely by the threshold:
Suppose the conditional mean functions
are continuous at . Here denotes conditional expectation. With observations available arbitrarily close to the cutoff from both sides, the treatment effect at the boundary is identified by
Under the continuity assumptions, this equals
For a continuously distributed score, the expression at is understood through the continuous extensions of the conditional mean functions, rather than through a positive-probability group exactly at the cutoff.
The crucial requirement is continuity of the potential-outcome means, not continuity of observed outcomes. A discontinuity in observed outcomes is precisely what the design seeks to measure.
Fuzzy design
In a fuzzy regression discontinuity design, crossing the cutoff changes the probability of treatment without determining treatment perfectly. Some eligible units may decline participation, while some ineligible units may obtain treatment through another route.
Define the outcome jump and treatment-probability jump as
If , the fuzzy RD estimand is
This is a local instrumental variables estimand, with cutoff eligibility serving as the instrument. Under suitable continuity conditions, an exclusion restriction, and monotonicity of treatment response to eligibility, it identifies the average treatment effect for compliers at the cutoff: units whose treatment status is changed by eligibility.
The exclusion restriction requires threshold eligibility to affect the outcome only through the treatment being studied. Monotonicity rules out units who would receive treatment when ineligible but not when eligible, under an assignment rule that increases treatment uptake.
The outcome jump also has a distinct interpretation as the local effect of eligibility or assignment. It should not be confused with the effect of actually receiving treatment.
Identification assumptions
The assumptions needed for statistical identification cannot be established solely by fitting a regression.
Continuity of counterfactual outcomes
Without the treatment change, average outcomes must evolve continuously through the cutoff. Smooth relationships between the score and other determinants of the outcome are compatible with the design; an unrelated abrupt change at the same threshold is not.
For example, if scholarship eligibility and admission to a special teaching program begin at the same score, the outcome discontinuity generally cannot identify the scholarship effect alone. It may instead identify the effect of the combined intervention.
Sorting and manipulation
Precise control over the running variable can undermine comparability. If applicants can alter a reported score to place themselves just above the threshold, units on the two sides may differ discontinuously in motivation, resources, or other determinants of outcomes.
Not every opportunity to influence a score invalidates the design. The concern is whether sorting produces discontinuities in relevant counterfactual outcomes. Conversely, an inability to manipulate the score is useful evidence but does not by itself establish every identifying assumption.
Support near the cutoff
The continuity-based argument requires information sufficiently close to the threshold on both sides. If the score takes only a small number of distinct values, or there is a gap around the cutoff, the boundary effect cannot generally be recovered by arbitrarily local comparisons. Additional restrictions on functional form or smoothness are then needed.
This issue is especially important with rounded scores, coarse age categories, and small numbers of elections or administrative units.
Estimation and statistical inference
Local polynomial estimation
A common approach fits separate local linear regressions on either side of the cutoff. With , a pooled representation is
The bandwidth determines which observations enter the local comparison. Separate slopes allow the outcome-score relationship to differ across the cutoff. In a sharp design, the fitted intercept difference estimates the boundary treatment effect.
Observations are often weighted by distance from the cutoff, commonly using triangular weights. Local quadratic or other low-order polynomial fits are also used. For fuzzy designs, the outcome and treatment jumps are estimated locally and combined as a ratio.
A single linear regression over the entire score range can be sensitive to functional-form assumptions. High-degree global polynomials may introduce oscillations, unstable boundary behavior, and misleading uncertainty estimates. Local low-order methods reduce dependence on distant observations.
Bandwidth selection
Bandwidth choice involves a bias–variance tradeoff:
- A narrow bandwidth makes the comparison more local but uses fewer observations.
- A wide bandwidth improves precision but requires the fitted curve to approximate the outcome relationship over a larger region.
Data-driven selectors often target mean squared error or properties of statistical coverage. Results are commonly examined across a range of reasonable bandwidths, rather than selected according to which estimate is statistically significant.
Bias correction and uncertainty
At bandwidths chosen for point-estimation accuracy, smoothing bias may be large enough to impair conventional confidence interval coverage. Robust bias-corrected inference estimates the leading bias, adjusts the point estimate, and accounts for the additional uncertainty from estimating that bias.
Inference must also reflect the dependence structure of the data. Repeated observations on the same units or assignment at an administrative-group level may require appropriate clustering. In fuzzy designs, a small treatment-probability jump creates a weak first stage and can make treatment-effect estimates highly imprecise.
Diagnostic evidence
Several checks assess the plausibility and stability of a design.
Graphical analysis. Plots typically show binned outcome averages against the running variable, with separate fitted curves on either side. They reveal the outcome pattern, data support, and possible discontinuity. Visual impressions alone are not estimates with valid uncertainty measures.
Predetermined covariate balance. Characteristics determined before treatment should generally not jump at the cutoff. Discontinuities may indicate sorting or a competing institutional change. Balance checks provide supporting evidence, not proof of comparability.
Density checks. Tests for a discontinuity in the running variable’s density can reveal unusual bunching near the threshold. Interpretation must account for rounding, discrete scores, administrative recording practices, and the institutional assignment process.
Placebo cutoffs and outcomes. Apparent effects at thresholds where assignment does not change, or in outcomes that treatment should not affect, may expose specification problems or background discontinuities.
Sensitivity analysis. Researchers examine changes in bandwidth, polynomial order, covariate adjustment, or sample definition. A “donut” specification excludes observations immediately around the threshold, sometimes because of heaping or suspected manipulation. This changes the information used to estimate the boundary effect and requires additional extrapolation; it does not automatically repair an invalid design.
Failure to reject a diagnostic null hypothesis does not establish validity, particularly when the checks have low statistical power.
Continuity and local randomization
Two related inferential frameworks are used.
The continuity framework identifies effects through limits of conditional mean functions as the running variable approaches the cutoff. It does not assume literal random assignment within a fixed interval.
The local randomization framework posits that, within a sufficiently narrow window, assignment can be treated as if generated by a specified random mechanism. This permits randomization-based inference and finite-window comparisons, but requires stronger assumptions within the chosen window.
The frameworks may target different populations: a limit-defined effect at the boundary versus an effect for units inside a specified window. Describing nearby units as “almost randomized” is therefore an intuition, not a substitute for specifying the inferential framework.
Historical development and applications
Donald Thistlethwaite and Donald Campbell introduced regression discontinuity design in 1960 in a study of scholarship recognition and subsequent educational outcomes. Later work developed its identification foundations, connections to instrumental variables, and methods for local estimation and inference.
Applications include scholarship and school-admission thresholds, income-based program eligibility, age-based policy rules, and regulations triggered by firm size. Close-election studies use narrow victory margins to study the consequences of winning office, although their validity depends on the electoral process and whether candidates or parties can sort precisely around victory.
Unlike difference-in-differences, which relies on assumptions about untreated outcome trends across groups and time, regression discontinuity relies on behavior around an assignment boundary. A threshold rule alone is insufficient: the rule must create the relevant treatment change without introducing inseparable alternative explanations for the outcome jump.
Scope and limitations
Regression discontinuity usually identifies a local effect, not an average effect for everyone exposed to a policy. A scholarship effect for students near an eligibility score may differ from the effect for much stronger or weaker students. Fuzzy designs are more local still, identifying effects for compliers at that boundary.
Important limitations include sparse data near the cutoff, measurement error in the assignment score, changing eligibility rules, noncompliance, and interference between units. If treatment of one unit affects another unit’s outcome, the usual interpretation based solely on individual treatment status may be inadequate.
Extensions include multiple-cutoff designs, geographic boundary designs, multidimensional assignment rules, and regression kink designs. A regression kink design exploits a change in the slope of a treatment rule rather than a jump in its level, requiring additional derivative-based smoothness assumptions. Geographic designs require particular care because neighborhoods on opposite sides of a boundary may differ in many ways besides the policy under study.
Estimating effects away from a cutoff, pooling effects across thresholds, or predicting a policy expansion requires assumptions beyond those that identify the original local discontinuity.