Robust Fuzzy Regression Discontinuity Design
Robust Bias-Corrected Fuzzy Regression Discontinuity Design · Also known as: Robust Fuzzy RDD, Fuzzy RD with robust inference, bias-corrected fuzzy RD, CCT fuzzy RDD
Robust Fuzzy Regression Discontinuity Design estimates a local average treatment effect (LATE) at a threshold where crossing the cutoff raises — but does not guarantee — treatment receipt. Introduced by Calonico, Cattaneo, and Titiunik (2014), the robust framework applies bias-corrected local polynomial estimation with a robust variance estimator, correcting the coverage failures of conventional bandwidth-optimal inference in both the sharp and fuzzy cases.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Robust Fuzzy RDD when a known threshold assigns units to treatment with imperfect compliance — for example, test-score cutoffs for programme eligibility where not all eligible applicants enrol, or income thresholds where some below-cutoff individuals still receive benefits. The method is appropriate when the running variable is continuous, manipulation of the score is implausible, and the first-stage jump in treatment probability is substantial (F-statistic above 10). Do not use it when compliance is perfect (sharp design applies), when the running variable is discrete with few mass points near the cutoff, when there is evidence of heaping or manipulation at the threshold (check with density tests), or when the researcher-chosen bandwidth is far from the data-driven optimum without a clear justification.
Strengths & limitations
- Delivers valid causal inference (LATE) even when the threshold only partially determines treatment, by using the assignment as an instrument.
- The CCT bias-correction and robust variance estimator provide confidence intervals with correct asymptotic coverage at the MSE-optimal bandwidth, removing the guesswork of bandwidth over-smoothing.
- Data-driven bandwidth selection removes analyst discretion, improving reproducibility and reducing specification searching.
- Exploits near-random variation near the threshold, so results are credible without requiring full covariate balance across the full sample.
- Well-supported by software — the rdrobust package in R and Stata implements the full CCT pipeline for both sharp and fuzzy designs.
- Estimates a local effect (LATE for compliers near the threshold) that may not generalise to units far from the cutoff or to always-takers and never-takers.
- Requires a sufficiently strong first stage: a weak jump in treatment probability at the cutoff leads to a large IV ratio variance and unreliable estimates.
- A continuous running variable without manipulation is essential; discrete running variables or self-sorting near the threshold violate the identification assumption.
- Small samples near the threshold reduce precision sharply, since only local observations contribute to estimation.
- Interpretation is more complex than a sharp design, requiring separate reporting of the first-stage jump, reduced-form effect, and the LATE ratio.
Frequently asked
What makes this design fuzzy rather than sharp?
In a sharp RDD, crossing the threshold perfectly determines whether a unit receives treatment. In a fuzzy RDD, crossing the threshold only changes the probability of treatment — some units below the cutoff still get treatment and some above do not. This partial compliance means the threshold is used as an instrument rather than as a direct assignment rule.
Why is the robust bias-correction necessary?
The bandwidth that minimises MSE of the point estimate is not the bandwidth that gives confidence intervals with correct coverage. At the MSE-optimal bandwidth, conventional intervals under-cover because they do not account for the remaining bias. The Calonico-Cattaneo-Titiunik (2014) robust interval adds a bias-correction step and inflates the variance to account for estimation uncertainty in the bias, restoring nominal coverage.
How large does the first stage need to be?
As a rule of thumb, a first-stage F-statistic above 10 is required for the IV ratio to have well-behaved finite-sample properties. With a weak first stage (F < 10), the denominator of the LATE ratio is imprecisely estimated, leading to wide or distorted confidence intervals. Reporting the first-stage jump in treatment probability alongside its confidence interval is mandatory.
Can I use this method with a discrete running variable?
Only with caution. A discrete running variable with few mass points near the threshold means there are limited observations in the bandwidth and local polynomial smoothing is unreliable. Standard RDD density tests for manipulation also perform poorly in the discrete case. Specialised methods for discrete running variables (e.g., Lee and Card 2008) should be considered instead.
What software implements robust fuzzy RDD?
The rdrobust package, available in both R and Stata, implements the full CCT pipeline — data-driven bandwidth selection, local polynomial estimation, bias correction, and robust variance — for both sharp and fuzzy designs. Specify the endogenous treatment variable with the fuzzy option to run the IV ratio automatically.
Sources
- Calonico, S., Cattaneo, M. D., & Titiunik, R. (2014). Robust Nonparametric Confidence Intervals for Regression-Discontinuity Designs. Econometrica, 82(6), 2295-2326. DOI: 10.3982/ECTA11757 ↗
- Imbens, G. W., & Lemieux, T. (2008). Regression discontinuity designs: A guide to practice. Journal of Econometrics, 142(2), 615-635. DOI: 10.1016/j.jeconom.2007.05.001 ↗
How to cite this page
ScholarGate. (2026, June 3). Robust Bias-Corrected Fuzzy Regression Discontinuity Design. ScholarGate. https://scholargate.app/en/causal-inference/robust-fuzzy-regression-discontinuity
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Difference-in-DifferencesEconometrics↔ compare
- Fuzzy Regression DiscontinuityCausal inference↔ compare
- Instrumental Variables in Health ResearchHealth Economics↔ compare
- Local Average Treatment EffectCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare