Regression Discontinuity Designs: A Guide to Practice
Causal Question / Estimand
The local treatment effect at the threshold for sharp RD, and the fuzzy-RD LATE (allowing heterogeneous effects, following Hahn–Todd–Van der Klaauw) when the cutoff changes treatment probability rather than treatment itself.
Identification Strategy
A practical companion to the theory: identification via Continuity-at-Cutoff of the outcome regression at the threshold. The paper’s emphasis is on estimation and inference: estimate with local linear regression using only observations near the cutoff (not high-order global polynomials), choose the bandwidth by cross- validation, and fit the two sides separately. For fuzzy RD, the ratio of the outcome jump to the treatment-probability jump is a Wald/IV estimator delivering a LATE. Provides a checklist of graphical and falsification diagnostics.
Key Assumptions
Continuity-at-Cutoff, No-Manipulation (density continuity, à la McCrary), and for fuzzy RD the IV assumptions (Exclusion-Restriction, Monotonicity). SUTVA.
Threats to Validity
Discontinuities in covariates or in the density of the running variable (manipulation); sensitivity to bandwidth and functional form; jumps at non-cutoff values. Recommends: (a) covariate-continuity tests, (b) McCrary density test, (c) placebo-cutoff tests.
Setting / Data
n/a — methodological guide to practice (econometrics).
Key Claims
- Use local linear regression near the cutoff and choose bandwidth by cross-validation; global high-order polynomials are discouraged.
- Fuzzy RD is an IV problem at the cutoff and estimates a complier-specific LATE.
- A standard set of falsification checks (density, covariates, placebo cutoffs) should accompany every RD analysis.
Connections
- Companion survey: LeeLemieux2010-RDDInEconomics
- Superseded on inference by robust bias-correction in CattaneoEtAl2020-RDDHandbook
- Fuzzy RD shares machinery with IV/LATE (IV). See also RDD
Citation
Imbens, G. W., & Lemieux, T. (2008). Regression Discontinuity Designs: A Guide to Practice. Journal of Econometrics, 142(2), 615–635.