Machine Learning-Augmented Doubly Robust Estimation (ML-DR)
Machine Learning-Augmented Doubly Robust Estimation · Also known as: ML-DR, AIPW with ML, Double/Debiased ML doubly robust, DML-DR
Machine learning-augmented doubly robust (ML-DR) estimation combines the classical doubly robust (AIPW) identification strategy with flexible machine learning models for the nuisance functions — the propensity score and the outcome regression. The result is a causal estimator that is consistent if either ML component is correctly specified, and that achieves valid, root-n inference even when the nuisance models are estimated with high-dimensional regularisation or nonparametric learners.
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-DR when you need a credible average treatment effect from observational data with a large number of covariates that would saturate a classical parametric model. It is particularly appropriate when the form of the propensity score or outcome model is unknown, making flexible ML estimators preferable to logistic regression or OLS. The method suits settings with at least a few hundred observations so that cross-fitting is stable. Avoid it when the sample is small (n < 200), when treatment is not strongly ignorable given observed covariates, when the propensity score is near zero or one for many units (extreme overlap violations), or when the data have strong temporal dependence that cross-fitting fails to respect.
Strengths & limitations
- Doubly robust: consistent if either the outcome regression or the propensity score model is correctly specified, providing two chances to get the right answer.
- Accommodates high-dimensional covariates through regularised or nonparametric ML learners without inflating bias in the causal estimate.
- Cross-fitting eliminates regularisation bias and restores root-n convergence and valid confidence intervals.
- Flexible choice of ML algorithm — LASSO, random forests, XGBoost, neural networks — with model selection by cross-validation.
- Neyman orthogonality makes the causal parameter insensitive to small errors in nuisance functions, improving robustness in finite samples.
- Requires strong ignorability: all confounders must be observed. Unmeasured confounding cannot be handled by any observed-data balancing method.
- Extreme propensity scores (near 0 or 1) lead to extreme AIPW weights and inflated variance; overlap must hold across the covariate distribution.
- Performance depends on the quality of the ML nuisance estimates; if both nuisance models are substantially misspecified, the estimator is not consistent.
- Cross-fitting requires a reasonable sample size to be stable; with very small samples the held-out folds are too small to yield reliable nuisance predictions.
- More complex to implement and interpret than classical doubly robust estimation with parametric nuisance models.
Frequently asked
Why is cross-fitting necessary when I am already using doubly robust estimation?
ML estimators reduce bias by fitting complex functional forms but introduce regularisation bias that contaminates the causal estimate when nuisance predictions are made on the same data used for the AIPW score. Cross-fitting generates out-of-sample nuisance predictions, removing this contamination and restoring root-n convergence.
What ML algorithms can I use for the nuisance functions?
Any supervised learning algorithm is permissible — LASSO, ridge regression, random forests, gradient boosting (XGBoost, LightGBM), neural networks, or ensembles (super learner). The choice should be driven by cross-validated predictive performance for each nuisance function separately.
Does ML-DR work when there is unmeasured confounding?
No. Like all observational causal methods based on selection on observables, ML-DR requires that all confounders are measured and included in the covariate set. Unmeasured confounding cannot be addressed by improving the nuisance models.
How many folds should I use in cross-fitting?
Five or ten folds is common and usually sufficient. With very large samples, two folds can work. With small samples, more folds reduce variance in nuisance estimation but make each fold smaller. The estimator is asymptotically equivalent regardless of the number of folds, so the choice is largely practical.
How does ML-DR differ from standard doubly robust estimation?
Standard doubly robust estimation uses parametric models (logistic regression for the propensity score, OLS for the outcome regression). ML-DR replaces these with flexible, data-adaptive ML learners and adds cross-fitting to remove bias introduced by regularisation, enabling valid inference in high-dimensional settings where parametric models are misspecified.
Sources
- Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., & Robins, J. (2018). Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal, 21(1), C1-C68. DOI: 10.1111/ectj.12097 ↗
- Farrell, M. H., Liang, T., & Misra, S. (2021). Deep Neural Networks for Estimation and Inference. Econometrica, 89(1), 181-213. DOI: 10.3982/ECTA16901 ↗
How to cite this page
ScholarGate. (2026, June 3). Machine Learning-Augmented Doubly Robust Estimation. ScholarGate. https://scholargate.app/en/causal-inference/machine-learning-augmented-doubly-robust-estimation
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
- Inverse Probability WeightingCausal inference↔ compare
- Machine Learning-Augmented Propensity Score MatchingCausal inference↔ compare
- Marginal Structural ModelCausal inference↔ compare
- Propensity Score WeightingCausal inference↔ compare