Machine Learning-Augmented Fuzzy Regression Discontinuity Design
Also known as: ML-augmented fuzzy RDD, ML fuzzy RD, double ML fuzzy RDD, nonparametric fuzzy RDD
ML-augmented fuzzy RDD extends the classical fuzzy regression discontinuity design by replacing parametric polynomial approximations with flexible machine learning estimators. Where standard fuzzy RDD uses IV-style estimation at a threshold with imperfect compliance, the ML-augmented variant leverages nonparametric learners — such as random forests or neural networks — to model both the outcome and the first-stage treatment probability near the cutoff, reducing misspecification bias while preserving causal identification.
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 ML-augmented fuzzy RDD when a threshold rule increases but does not perfectly determine treatment assignment, and you suspect the outcome-running-variable relationship is nonlinear or that many covariates confound the functional form near the cutoff. It is particularly valuable with large datasets where ML estimators can exploit high-dimensional covariate information. Do not use it when compliance is essentially sharp (everyone just above the cutoff is treated), as standard sharp ML-RDD suffices. Avoid it with very small samples near the cutoff — ML estimators require adequate data density to produce reliable limit estimates — and when interpretability of the nuisance model is required by stakeholders.
Strengths & limitations
- Avoids the polynomial misspecification bias that afflicts classical fuzzy RDD by using flexible nonparametric first-stage and reduced-form fits.
- Handles high-dimensional covariates naturally through ML feature selection, improving precision of the LATE estimate.
- Double ML debiasing ensures root-n consistent and asymptotically normal inference despite slow convergence of the ML nuisance estimators.
- Retains the clean causal identification logic of fuzzy RDD: only local variation around the cutoff identifies the effect.
- Cross-fitting prevents overfitting from inflating the apparent precision of the estimate.
- Requires a sufficient density of observations near the cutoff; if few units are close to c, ML estimators cannot reliably estimate the limit values.
- The causal estimate (LATE) applies only to compliers at the cutoff — a narrow subpopulation — and does not generalise to the full population.
- ML black-box models reduce interpretability of the nuisance functions, which may be a concern in regulatory or policy settings.
- Bandwidth choice and sample-splitting strategy introduce additional researcher degrees of freedom that must be reported transparently.
- Implementation is substantially more complex than classical fuzzy RDD, requiring familiarity with double ML procedures and careful software choices.
Frequently asked
How does ML-augmented fuzzy RDD differ from classical fuzzy RDD?
Classical fuzzy RDD uses local linear or polynomial regression near the cutoff, assuming a simple parametric form. ML-augmented fuzzy RDD replaces those parametric fits with flexible machine learning estimators for both the outcome surface and the treatment-probability surface, reducing functional form bias. Double ML debiasing then corrects for the slower convergence rate of ML estimators so that inference on the LATE remains valid.
What is the complier local average treatment effect (LATE) in this context?
It is the average causal effect for units at the cutoff whose treatment status is determined by whether they cross the threshold — i.e., those who would be treated if just above c but not if just below c. The estimate does not apply to always-takers or never-takers.
What is cross-fitting and why is it necessary?
Cross-fitting splits the sample into folds, fits the ML nuisance models on held-out folds, and predicts on the remaining observations. This prevents the nuisance models from overfitting the same data used to estimate the LATE, which would otherwise bias the estimate and deflate standard errors.
Do I still need to check the manipulation of the running variable?
Yes. The continuity/no-manipulation assumption (tested with McCrary's density test or similar) is required regardless of how the nuisance functions are estimated. If agents can precisely manipulate the running variable to sort just above or below the cutoff, the causal identification fails even with perfect ML fits.
What sample size is needed near the cutoff?
There is no fixed rule, but ML estimators need enough data close to the cutoff to estimate the limit values reliably. Sparse data near c — even with large total samples — can make the limit estimates noisy. Kernel-weighted approaches that focus on observations near the cutoff help, but you should report sensitivity to bandwidth choice.
Sources
- Hahn, J., Todd, P., & Van der Klaauw, W. (2001). Identification and estimation of treatment effects with a regression-discontinuity design. Review of Economic Studies, 68(1), 201-209. DOI: 10.1111/1468-0262.00183 ↗
- Semenova, V., & Chernozhukov, V. (2021). Debiased machine learning of conditional average treatment effects and other causal functions. The Econometrics Journal, 24(2), 264-289. DOI: 10.1093/ectj/utaa027 ↗
How to cite this page
ScholarGate. (2026, June 3). Machine Learning-Augmented Fuzzy Regression Discontinuity Design. ScholarGate. https://scholargate.app/en/causal-inference/machine-learning-augmented-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
- Doubly Robust EstimationCausal inference↔ compare
- Fuzzy Regression DiscontinuityCausal inference↔ compare
- Instrumental Variables in Health ResearchHealth Economics↔ compare
- Machine learning-augmented regression discontinuity designCausal inference↔ compare