Machine Learning-Augmented Regression Discontinuity Design
Also known as: ML-RDD, ML-augmented RD, data-adaptive RDD, nonparametric RDD with ML
Machine learning-augmented regression discontinuity design (ML-RDD) combines the sharp identification logic of classical RDD — exploiting a known assignment cutoff in a running variable — with flexible, data-adaptive ML methods for bandwidth selection, conditional mean estimation, and covariate adjustment. The goal is to recover a more accurate and less assumption-laden estimate of the local average treatment effect at the threshold.
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-RDD when a sharp or fuzzy cutoff assigns treatment, the running variable is continuous and reasonably dense near the threshold, and there is meaningful concern that standard low-order polynomial fits may be misspecified — particularly with complex, nonlinear conditional mean functions or many pre-treatment covariates worth adjusting for. It is particularly valuable in large administrative datasets where algorithmic bandwidth and order selection is more reliable than researcher judgment. Do not use it when the running variable can be precisely manipulated around the cutoff (density test fails), when the sample is very small (fewer than ~200 observations within a sensible bandwidth), or when there is no natural discontinuity in the assignment rule — in those cases, a matching or DiD strategy is preferable.
Strengths & limitations
- Data-adaptive bandwidth and order selection reduces researcher degrees of freedom and the risk of specification searching.
- Flexible ML learners can capture nonlinear conditional mean functions that polynomial fits misrepresent, lowering bias.
- Covariate adjustment via ML residualisation can substantially improve precision without requiring strong parametric assumptions.
- Inherits the sharp local identification of classical RDD: only behaviour in a narrow window around the cutoff is used, limiting confounding.
- Robust bias-corrected inference procedures (CCT) are well-developed and compatible with ML-estimated nuisance functions.
- Identification is still local: the estimated effect applies only to units near the threshold, limiting external validity.
- ML methods require larger samples to perform well; with thin data near the cutoff, parametric local polynomials may dominate.
- The no-manipulation assumption must hold; ML estimation cannot rescue a design where agents sort around the threshold.
- Greater methodological complexity makes results harder to communicate to non-technical audiences than a simple local polynomial RDD.
- Requires careful choice of which ML algorithm to deploy; different algorithms can yield somewhat different estimates without a principled selection rule.
Frequently asked
How does ML-RDD differ from standard local polynomial RDD?
Standard RDD fixes a polynomial order (usually 1 or 2) and uses a rule-of-thumb or MSE-optimal bandwidth based on that polynomial. ML-RDD replaces or supplements these choices with data-adaptive selectors — cross-validation, lasso-based regularisation, or regression forests — that let the data determine the appropriate functional form and window width, reducing reliance on analyst-chosen tuning parameters.
Does using ML estimation weaken the causal interpretation?
No. The causal identification comes from the design — the discontinuous cutoff assignment rule — not from the estimator. ML methods are used only to estimate the nuisance functions (conditional means and bandwidths) more accurately; the identifying assumption (no sorting around the cutoff) is unchanged.
What sample size is needed for ML-RDD to outperform standard RDD?
There is no universal threshold, but simulation evidence suggests that data-adaptive bandwidth selectors begin to dominate fixed-order polynomials reliably once there are several hundred observations within the bandwidth. With very small within-bandwidth samples, a simple local linear estimator with the CCT bandwidth is often more stable.
Can ML-RDD handle fuzzy designs where treatment is not perfectly determined by the cutoff?
Yes. In a fuzzy RDD the jump in the probability of treatment at the cutoff serves as an instrument for actual treatment. ML methods can be applied to both the first-stage (treatment take-up) and reduced-form (outcome) regressions, then combined via an IV-type ratio, in analogy to doubly robust IV estimation.
Which software packages implement ML-augmented RDD?
The rdrobust package (R and Stata) implements MSE- and CER-optimal bandwidth selection with bias-corrected inference. The grf package (R) implements causal forests that can be applied to an RDD neighbourhood. Python users can use rdd or econml for related functionality.
Sources
- Calonico, S., Cattaneo, M. D., & Farrell, M. H. (2019). Optimal mean squared error bandwidth selection for regression discontinuity designs. Bernoulli, 25(4A), 2703-2729. link ↗
- Imbens, G., & Wager, S. (2019). Optimized regression discontinuity designs. Review of Economics and Statistics, 101(2), 264-278. DOI: 10.1162/rest_a_00793 ↗
How to cite this page
ScholarGate. (2026, June 3). Machine Learning-Augmented Regression Discontinuity Design. ScholarGate. https://scholargate.app/en/causal-inference/machine-learning-augmented-regression-discontinuity-design
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.
- Fuzzy Regression DiscontinuityCausal inference↔ compare
- Machine learning-augmented difference-in-differencesCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare