Policy Evaluation Entropy Balancing
Entropy Balancing for Causal Policy Evaluation · Also known as: Entropy Balancing, EB Weighting, Maximum-Entropy Reweighting, Hainmueller Balancing
Entropy balancing is a maximum-entropy reweighting method that assigns weights to control-group units so that their weighted covariate moments exactly match those of the treated group. Introduced by Hainmueller (2012), it provides exact balance on specified moments without iterative propensity-score trimming, making it a powerful preprocessing tool for causal policy evaluation in observational studies.
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 entropy balancing when you have a binary treatment in observational data and want exact balance on covariate moments without specifying a propensity-score model. It is well-suited to policy evaluations where the number of covariates is moderate, the treated group is not too small, and overlap between treated and control units is reasonable. It is especially useful when repeated propensity-score trimming and re-estimation cycles are inconvenient. Avoid it when overlap is very poor (few comparable control units exist), when the treatment is continuous or multi-valued (use generalized entropy balancing instead), or when covariates are high-dimensional without dimension-reduction — in those settings the balance constraints may be infeasible or weights may be highly variable.
Strengths & limitations
- Achieves exact balance on specified covariate moments by design, eliminating the need for iterative model refinement.
- Does not require correct specification of a propensity-score model; the optimization directly targets balance.
- The estimator is doubly robust: consistent if either the weighting scheme or the outcome model is correctly specified.
- Convex optimization guarantees a unique global solution, making the procedure fast and deterministic.
- Can balance means, variances, and higher-order moments simultaneously by adding constraints.
- Requires adequate overlap: if very few control units resemble the treated units, weights become extreme and variance inflates.
- Balance is only guaranteed on the moments explicitly constrained; unspecified moments or unmeasured confounders remain unbalanced.
- Standard implementation handles only binary treatments; continuous or multi-valued treatments need generalized extensions.
- With many covariates, the user must decide which moments to constrain, introducing discretionary choices.
Frequently asked
How is entropy balancing different from propensity-score weighting?
Propensity-score weighting estimates the probability of treatment and uses it to weight units, but balance depends on correct model specification and is only approximate. Entropy balancing directly enforces exact balance on chosen covariate moments through a convex optimization, without ever estimating a propensity score. If the propensity model is misspecified, PSW can fail; entropy balancing cannot, for the moments it constrains.
What moments should I constrain?
At a minimum, constrain the means of all pre-treatment covariates you believe are related to both treatment and outcome. Adding variance constraints improves balance on the covariate distributions beyond their central tendency. Skewness constraints are sometimes added for highly skewed variables. Start with means; add higher moments if post-weighting SMDs remain large.
Is entropy balancing doubly robust?
Yes. Zhao and Cooney (2017) proved that the entropy-balanced weighted mean estimator is doubly robust: it gives a consistent estimate of the ATT if either the weighting correctly balances the covariates or the outcome model is correctly specified — provided at least one of the two conditions holds.
What sample size do I need?
The control group must be at least as large as the treated group (and ideally several times larger) to give the optimizer enough flexibility to satisfy the balance constraints without extreme weights. If the control pool is small relative to the number of balance constraints, the problem may be infeasible or yield highly variable weights.
Can I use entropy balancing with panel data?
Yes, though the standard implementation balances at the unit level in a cross-sectional sense. For panel designs, entropy balancing is typically applied to pre-treatment baseline characteristics, and the balanced weights are then carried into a DiD or panel regression for the outcome analysis.
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., & Cooney, D. (2017). Entropy Balancing is Doubly Robust. Journal of Causal Inference, 5(1). DOI: 10.1515/jci-2016-0010 ↗
How to cite this page
ScholarGate. (2026, June 3). Entropy Balancing for Causal Policy Evaluation. ScholarGate. https://scholargate.app/en/causal-inference/policy-evaluation-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.
- Difference-in-DifferencesEconometrics↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare
- Synthetic Control MethodCausal inference↔ compare