Heterogeneous Treatment Effect Entropy Balancing
Heterogeneous Treatment Effect Estimation with Entropy Balancing · Also known as: HTE entropy balancing, CATE with entropy balancing, heterogeneous effects EB, subgroup entropy balancing
Heterogeneous Treatment Effect Entropy Balancing combines entropy balancing — a preprocessing step that reweights control units to match the treatment group on covariate moments — with methods that estimate how the treatment effect varies across subgroups or individuals. It produces covariate-balanced weights without parametric propensity models, then uses those weights to estimate conditional average treatment effects (CATEs) across moderating variables.
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 this approach when (1) you have observational data with a binary treatment and suspect that treatment effects differ across subgroups or individual characteristics, (2) the propensity score is difficult to specify correctly — entropy balancing avoids that model — and (3) you have a sufficient sample in both treated and control groups to support subgroup analyses (roughly 100+ per group for two subgroups). It is not appropriate when the treatment is continuous, when the sample is very small, or when you need to adjust for unmeasured confounders — entropy balancing only controls measured covariates.
Strengths & limitations
- Achieves exact covariate balance on specified moments without estimating a propensity score model, reducing model-specification error.
- Naturally integrates with any heterogeneous-effect estimator (causal forests, interacted regression, GATES) that accepts observation weights.
- More transparent than black-box doubly robust estimators: the weights and the balance achieved are directly inspectable.
- Allows balancing on means, variances, and higher-order moments simultaneously, giving richer covariate control than propensity score matching.
- Avoids the common problem of extreme propensity weights by directly constraining the weight distribution.
- Only controls for observed confounders; unmeasured variables that differ between treatment and control remain a source of bias.
- Subgroup analyses require substantially larger samples than a single ATE estimate; power drops quickly with many subgroups.
- Choosing which covariate moments to balance is a researcher decision that affects results and should be justified.
- Entropy balancing may fail to converge when the treatment and control groups overlap poorly in covariate space.
- Combining multiple methods (balancing + heterogeneity detection) multiplies researcher degrees of freedom, increasing the risk of data dredging.
Frequently asked
How does entropy balancing differ from propensity score weighting?
Propensity score weighting estimates the probability of treatment from a regression model, then uses those estimated probabilities as weights. Entropy balancing bypasses that model entirely: it finds weights directly by solving an optimisation problem that enforces exact balance on specified covariate moments. If the propensity model is misspecified, propensity weights are biased; entropy balancing is immune to that particular failure.
Can I use entropy balancing with a continuous treatment?
Standard entropy balancing is designed for binary treatments. Extensions to continuous treatments exist (generalised entropy balancing) but are less widely validated. For a continuous treatment, inverse probability of treatment weighting with a flexible propensity model or matching on the generalised propensity score is more established.
How do I test whether subgroup effects are real or noise?
Include a treatment-by-moderator interaction term in a weighted regression and test its significance. For many subgroups, apply a multiple-testing correction (Bonferroni or Benjamini-Hochberg). Pre-registering which subgroups you will examine before seeing the data is the strongest protection against false-discovery.
What sample size is needed?
For the overall ATT, entropy balancing works with relatively modest samples (50+ per group is common). Heterogeneous effect estimation adds demands: each subgroup needs adequate observations for reliable estimates, so for two subgroups you typically want 100+ per group, and more for finer partitions.
What if convergence fails during entropy balancing?
Convergence failure usually signals poor overlap between treated and control units in the covariate space. Remedies include trimming units with extreme covariate values, relaxing the balance constraints to fewer moments, or switching to a method that explicitly models limited overlap such as overlap-weighted estimation.
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 ↗
- Athey, S., & Imbens, G. W. (2016). Recursive partitioning for heterogeneous causal effects. Proceedings of the National Academy of Sciences, 113(27), 7353-7360. DOI: 10.1073/pnas.1510489113 ↗
How to cite this page
ScholarGate. (2026, June 3). Heterogeneous Treatment Effect Estimation with Entropy Balancing. ScholarGate. https://scholargate.app/en/causal-inference/heterogeneous-treatment-effect-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
- Heterogeneous Treatment Effect Propensity Score MatchingCausal inference↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Propensity Score WeightingCausal inference↔ compare