Entropy Balancing
Also known as: EB, entropy reweighting, covariate balancing via entropy, Hainmueller balancing
Entropy balancing is a preprocessing method for causal inference that assigns weights to control-group units so that the reweighted control sample matches the treatment group exactly on a chosen set of covariate moments (means, variances, skewness). Introduced by Hainmueller (2012), it replaces trial-and-error propensity-score trimming with a constrained maximum-entropy optimisation that achieves balance in a single step.
Key highlights
- Achieves exact balance on all specified covariate moments in a single optimisation step, eliminating iterative trimming cycles.
- Transparent: any residual imbalance is immediately visible, unlike black-box propensity-score pipelines.
- The entropy objective keeps weights close to uniform, avoiding extreme weights and the associated variance inflation.
- Has been shown to be doubly robust under certain conditions: consistent if either the outcome model or the balancing model is correctly specified.
- Easily extended to balance higher-order moments (variances, skewness) beyond means, which standard IPW does not naturally target.
- Works well with unequal treatment and control group sizes.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use entropy balancing when you have a binary treatment and observational data and want exact covariate balance on specified moments without the manual tuning required by propensity-score matching. It is well-suited when the treatment and control groups are unequal in size, when off-support units are a concern, or when a prior matching step left residual imbalance. It is appropriate for cross-sectional and panel contexts and works with continuous or binary outcomes. Do not use it when the treatment is continuous or multi-valued (use generalised balancing weights instead), when sample sizes are very small (fewer than roughly 30 per group), when you have many more covariates than control units (the optimisation may have no feasible solution), or when unconfoundedness on observed variables is implausible.
Strengths & limitations
- Achieves exact balance on all specified covariate moments in a single optimisation step, eliminating iterative trimming cycles.
- Transparent: any residual imbalance is immediately visible, unlike black-box propensity-score pipelines.
- The entropy objective keeps weights close to uniform, avoiding extreme weights and the associated variance inflation.
- Has been shown to be doubly robust under certain conditions: consistent if either the outcome model or the balancing model is correctly specified.
- Easily extended to balance higher-order moments (variances, skewness) beyond means, which standard IPW does not naturally target.
- Works well with unequal treatment and control group sizes.
- Assumes unconfoundedness (no unmeasured confounding); balance on observed covariates does not guarantee balance on unobserved ones.
- The optimisation may be infeasible if the specified moments cannot be jointly satisfied given the available control units, requiring relaxation of constraints.
- With many covariates or high-order moments, the constraint set grows large, which can produce extreme weights and efficiency loss.
- Standard error estimation is non-trivial; the sandwich formula or bootstrap must be applied carefully to account for weight estimation uncertainty.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does entropy balancing differ from propensity score weighting?
Propensity score weighting first estimates the probability of treatment using a logistic model, then constructs inverse-probability weights. Balance is an indirect by-product of the model, and may require iterative model adjustments. Entropy balancing directly targets balance as a hard constraint, guaranteeing exact moment balance in one optimisation step without needing to estimate a propensity score.
What moments should I balance?
At minimum, balance the means of all pre-treatment covariates. If distributions differ substantially, also balance variances and — for highly skewed variables — third moments (skewness). The choice should be guided by substantive knowledge of which covariate features drive outcome variation.
What if the optimisation reports no feasible solution?
This means there are no weights that can simultaneously satisfy all balance constraints. Common remedies are: reduce the set of constraints (fewer moments or fewer covariates), trim the sample to the region of common support, or relax hard constraints to soft penalties using a regularised variant.
Does entropy balancing handle multiple treatment groups?
The standard method is defined for a binary treatment. For multiple groups, separate pairwise comparisons can be run, or you can use generalised balancing weight frameworks (e.g., CBPS or energy balancing) that accommodate multi-valued treatments.
How should I report standard errors after entropy balancing?
Use the bootstrap (resample units, re-estimate weights and ATT in each bootstrap replicate) or a sandwich (HC) estimator that treats the weights as estimated quantities. Naive OLS standard errors on the reweighted sample ignore weight-estimation uncertainty and are too small.
Sources
- 1.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.
- 2.Zhao, Q., & Coey, D. (2017). Entropy balancing is doubly robust. Journal of Causal Inference, 5(1). (Working paper version widely cited; see also Zhao & Coey 2018, Stanford GSB Research Paper.)
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Entropy Balancing. ScholarGate. https://scholargate.app/causal-inference/entropy-balancing