Regression Discontinuity in Policy Evaluation
Also known as: Policy RD Design, Threshold-Based Policy Evaluation, Cutoff Rule Evaluation, Eligibility-Threshold Design
Regression discontinuity (RD) is a quasi-experimental design for estimating the causal effect of a policy that is assigned by a sharp threshold on some continuous eligibility score — an income line for a benefit, a test score for a scholarship, a vote share for winning office, a population cutoff that triggers a regulation. Units falling just below and just above the cutoff are nearly identical except for their treatment status, so comparing their outcomes isolates the policy's effect at the threshold. First used by Thistlethwaite and Campbell in 1960 and revived as a workhorse of policy evaluation by economists in the 2000s, RD is widely regarded as the quasi-experimental design with the strongest claim to internal validity.
Key highlights
- Transparent, rule-based identification with strong internal validity — the closest a quasi-experiment comes to a randomised experiment near the cutoff.
- The key assumption (smoothness of other factors at the cutoff) is partly testable through density and covariate-continuity checks, unlike untestable selection-on-observables assumptions.
- Exploits administrative eligibility rules that pervade policy, so suitable settings and data are common and often already collected.
- Graphically intuitive: a single plot of outcomes against the running variable communicates the result to non-technical policymakers.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use RD to evaluate a policy whenever eligibility or treatment is determined by a known cutoff on a continuous score and units cannot precisely manipulate their position relative to that cutoff. It is ideal for means-tested benefits, merit thresholds, class-size or financing rules triggered by enrollment counts, electoral close-race designs, and regulatory thresholds tied to firm size or population. It is not appropriate when assignment is not rule-based, when the running variable can be precisely gamed, when there are too few observations near the cutoff, or when the policy question concerns effects far from the threshold. The estimate is local to the cutoff, so it answers what the policy does at the margin of eligibility, not its average effect over the whole population.
Strengths & limitations
- Transparent, rule-based identification with strong internal validity — the closest a quasi-experiment comes to a randomised experiment near the cutoff.
- The key assumption (smoothness of other factors at the cutoff) is partly testable through density and covariate-continuity checks, unlike untestable selection-on-observables assumptions.
- Exploits administrative eligibility rules that pervade policy, so suitable settings and data are common and often already collected.
- Graphically intuitive: a single plot of outcomes against the running variable communicates the result to non-technical policymakers.
- Estimates a local effect only at the cutoff, which may not generalise to units far from the threshold or to a shifted cutoff.
- Requires substantial data density around the cutoff; effective sample sizes can be small, yielding imprecise estimates and wide confidence intervals.
- Validity collapses if agents can precisely manipulate the running variable to sort onto the favourable side of the cutoff.
- Results can be sensitive to bandwidth choice and the order of the polynomial, and high-order global polynomials can produce misleading discontinuities.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
What is the difference between sharp and fuzzy RD in a policy context?
In a sharp design, crossing the cutoff deterministically changes treatment — everyone above the income line loses the benefit, everyone below keeps it. In a fuzzy design, crossing the cutoff changes only the probability of treatment, because eligibility does not guarantee take-up or because the rule is applied imperfectly. Sharp RD compares outcomes directly; fuzzy RD scales the outcome jump by the jump in treatment probability, delivering a local effect for those whose treatment status is actually shifted by the threshold.
Why is the RD estimate only 'local' to the cutoff?
RD identifies the treatment effect by comparing units just on either side of the threshold, where they are comparable. It therefore estimates the effect for units at the margin of eligibility, not for those far above or below the cutoff who may respond very differently. A scholarship's effect on a borderline student tells you little about its effect on a top scorer. Extrapolating the local estimate to the whole population requires extra assumptions that the design itself does not justify.
How do you know units did not manipulate their score to get the benefit?
You test for it. The standard check is a density test (originally McCrary, now commonly the Cattaneo-Jansson-Ma test) for a discontinuous jump in the number of units just on the favourable side of the cutoff, which would signal sorting. You also verify that predetermined characteristics are balanced across the cutoff, just as in a randomised experiment. If the density is smooth and covariates do not jump, the as-good-as-random assumption near the threshold is supported.
Sources
- 1.Thistlethwaite, D. L., & Campbell, D. T. (1960). Regression-discontinuity analysis: An alternative to the ex post facto experiment. Journal of Educational Psychology, 51(6), 309–317.
- 2.Imbens, G. W., & Lemieux, T. (2008). Regression discontinuity designs: A guide to practice. Journal of Econometrics, 142(2), 615–635.
- 3.Lee, D. S., & Lemieux, T. (2010). Regression discontinuity designs in economics. Journal of Economic Literature, 48(2), 281–355.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 22). Regression Discontinuity in Policy Evaluation. ScholarGate. https://scholargate.app/public-policy/regression-discontinuity-policy