Machine Learning-Augmented Entropy Balancing
Machine Learning-Augmented Entropy Balancing for Causal Inference · Also known as: ML-EB, augmented entropy balancing, ML-augmented EB, doubly-robust entropy balancing
Machine learning-augmented entropy balancing (ML-EB) combines Hainmueller's entropy balancing reweighting scheme with a machine-learning outcome model to produce a doubly-robust causal estimator. By jointly optimising covariate balance weights and a flexible predicted-outcome adjustment, ML-EB delivers consistent treatment-effect estimates even when either the weighting or the outcome model is misspecified, and it handles high-dimensional covariate spaces that classical entropy balancing cannot easily balance.
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-EB when you have observational data with a binary treatment and a rich covariate set, and you want a doubly-robust average treatment effect on the treated (ATT). It is especially valuable when the number of covariates is large or their functional relationship to the outcome is nonlinear, so a parametric propensity-score or outcome model would be unreliable. A sample of at least 200 units per arm is advisable for the ML outcome model to generalise. Do not use it when the treatment is multi-valued or continuous without adaptation, when sample size is very small (n < 100 total), or when the strong ignorability assumption (no unmeasured confounders) is implausible — no weighting method, however refined, can eliminate unobserved confounding.
Strengths & limitations
- Doubly robust: consistent ATT estimate if either the balancing weights or the ML outcome model is correct.
- Exact moment balance on specified covariate moments is guaranteed by the entropy-balancing step, eliminating iterative propensity-score trimming.
- Machine-learning flexibility lets the outcome model capture nonlinear and interaction effects that misspecify linear regression adjustment.
- Scales to high-dimensional covariate spaces where classic propensity-score models struggle with overlap and curse of dimensionality.
- Cross-fitting ensures valid semiparametric inference without parametric assumptions on the nuisance models.
- Requires specifying which covariate moments to balance; higher-order moments (skewness, kurtosis) must be added manually and can make the optimisation infeasible with limited control units.
- The ML outcome model and cross-fitting add computational and implementation complexity relative to simple inverse-probability weighting.
- Like all observational methods, validity rests on the unconfoundedness assumption; unmeasured confounders bias every weighting approach.
- Bootstrap variance estimation can be slow for large datasets; influence-function methods require differentiable ML models.
Frequently asked
How does ML-EB differ from standard entropy balancing?
Standard entropy balancing reweights control units to match treated-group covariate moments but relies on an implicit linear outcome model for its bias correction. ML-EB replaces that implicit model with an explicit flexible machine-learning outcome model, producing a fully doubly-robust estimator that is consistent even when the weighting model captures imbalance imperfectly, provided the ML model is well-specified.
What does 'doubly robust' mean in this context?
An estimator is doubly robust if it is consistent when either one of two nuisance models is correctly specified. In ML-EB, those two components are (1) the entropy-balancing weights and (2) the ML outcome model. If at least one is correct, the treatment-effect estimate converges to the true ATT as sample size grows, providing insurance against a single modelling error.
Why is cross-fitting necessary?
Cross-fitting (sample splitting) prevents the ML outcome model from overfitting the residuals it is supposed to correct. Without it, the model memorises the training sample, so its predictions on those same units appear too accurate, which biases the augmentation term. Cross-fitting ensures the model's predictions are evaluated on held-out data, preserving valid semiparametric inference.
What if I cannot balance all covariate moments exactly?
The entropy-balancing optimisation can become infeasible if you impose too many high-order constraints with limited control units. In practice, restrict constraints to means and, where theoretically important, second moments or key interactions. The ML outcome model compensates for remaining imbalance on higher-order terms it detects in the data.
Which machine-learning algorithms work best in the outcome model?
Gradient-boosted trees and lasso (or elastic net) are the most common choices due to their variable selection and nonlinearity handling. Stacking (super learner) ensembles that average multiple learners are theoretically appealing because they adapt to the true data-generating process without committing to a single algorithm. Avoid deep neural networks unless the sample is very large, as their variance can dominate the bias correction.
Sources
- Hainmueller, J. (2012). Entropy balancing for causal effects: A multivariate reweighting method to produce balanced samples in observational studies. Political Analysis, 20(1), 25-46. DOI: 10.1093/pan/mpr025 ↗
- Zhao, Q., & Percival, D. (2017). Entropy balancing is doubly robust. Journal of Causal Inference, 5(1), 20160010. DOI: 10.1515/jci-2016-0010 ↗
How to cite this page
ScholarGate. (2026, June 3). Machine Learning-Augmented Entropy Balancing for Causal Inference. ScholarGate. https://scholargate.app/en/causal-inference/machine-learning-augmented-entropy-balancing
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
- Entropy BalancingCausal inference↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare