Double Machine Learning
Double/Debiased Machine Learning (DML) · Also known as: Debiased Machine Learning, Neyman Orthogonal Score Estimation, Partialing-Out Lasso, Çift Makine Öğrenmesi
Double/Debiased Machine Learning (DML), introduced by Chernozhukov et al. (2018), is a semiparametric framework for estimating causal or structural parameters in the presence of high-dimensional controls. It uses flexible machine learning methods to model nuisance functions—the conditional expectations of the outcome and the treatment given covariates—and then constructs a debiased estimator of the target parameter that achieves root-n consistency and valid inference despite the regularization bias inherent in high-dimensional settings.
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 DML when you need a point estimate and confidence interval for a causal or structural parameter in observational data with a large number of potential confounders. It requires a partially linear or instrumental-variable-type structural assumption: the target parameter enters linearly while the high-dimensional nuisance can be nonlinear. DML assumes unconfoundedness (no unmeasured confounding) and sufficient smoothness in the nuisance functions for ML convergence. It is not appropriate when the causal parameter itself is heterogeneous across units without further extension; for heterogeneous effects, consider Causal Forests or R-learner variants built on the same orthogonality principle.
Strengths & limitations
- Achieves root-n consistent, asymptotically normal estimation of causal parameters even with hundreds of controls, overcoming the Regularization-Induced Confounding bias of naive lasso or ridge approaches.
- Agnostic to the choice of ML learner for nuisance estimation: any method achieving sufficiently fast convergence rates can be plugged in, giving practitioners flexibility to use the best available tool for each nuisance problem.
- Cross-fitting prevents overfitting bias from contaminating inference, so the method does not require sample splitting that wastes data—all observations contribute to both nuisance learning and final estimation.
- Produces valid standard errors, confidence intervals, and hypothesis tests under standard semiparametric regularity conditions, making causal claims interpretable and falsifiable.
- Assumes a partially linear or similar low-dimensional parametric structure for the target parameter; it does not directly estimate heterogeneous treatment effects across units without extension to the R-learner or Causal Forest framework.
- Validity depends on the unconfoundedness assumption—all relevant confounders must be observed and included in the covariate set. DML cannot correct for unmeasured confounding.
- Nuisance learners must achieve convergence rates faster than n^{-1/4} for the target parameter's root-n consistency to hold; in very high dimensions with weak signal, this condition can fail in finite samples.
- The method is designed for estimation and inference on a scalar or low-dimensional structural parameter; applying it to very high-dimensional parameter spaces requires additional theoretical extensions.
Frequently asked
Why do I need cross-fitting? Can I just use the whole sample for nuisance estimation and then run the partialing-out regression?
Without cross-fitting, the nuisance predictions carry in-sample overfitting that creates a correlation between the residuals and the estimation error, introducing first-order bias in the final estimate. Cross-fitting removes this by ensuring that the nuisance model used to compute a residual was never trained on that observation, making the residual and the nuisance error asymptotically uncorrelated.
What machine learning methods can I use for the nuisance functions in DML?
Any ML method that achieves a mean-squared prediction error converging faster than n^{-1/2} is theoretically valid; in practice this includes lasso, ridge, elastic net, random forests, gradient boosting (XGBoost, LightGBM), and neural networks. The choice should be guided by cross-validated prediction performance on each nuisance problem separately, not by the final causal estimate.
Does DML estimate heterogeneous treatment effects, or only an average treatment effect?
In its basic form, DML targets a single scalar or low-dimensional structural parameter—typically a weighted average treatment effect under partial linearity. For heterogeneous treatment effect estimation across subgroups or individual levels, extensions such as the R-learner or Causal Forest are required; these build on the same Neyman orthogonality principle but add a second stage for heterogeneity.
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 ↗
How to cite this page
ScholarGate. (2026, June 2). Double/Debiased Machine Learning (DML). ScholarGate. https://scholargate.app/en/causal-inference/double-machine-learning
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.
- Doubly Robust EstimationCausal inference↔ compare
- Heterogeneous Treatment EffectsCausal inference↔ compare
- Random ForestMachine learning↔ compare