Heterogeneous Treatment Effect Doubly Robust Estimation
Doubly Robust Estimation of Heterogeneous Treatment Effects · Also known as: DR-HTE, augmented IPW for HTE, doubly robust CATE estimation, semiparametric HTE estimation
Doubly robust estimation of heterogeneous treatment effects (HTE) estimates how the causal effect of a treatment varies across subgroups or individual covariate values. By combining an outcome model and a propensity score model, it retains consistency if either model is correctly specified, and supports flexible machine learning nuisance estimators through cross-fitting to produce valid conditional average treatment effect (CATE) estimates.
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 method when your research question asks not just whether a treatment works on average but for whom it works and how strongly. It is appropriate with observational data where treatment assignment depends on observed covariates, and when you have a large enough sample to estimate both nuisance functions and the CATE surface reliably — typically n > 500 and ideally n > 1,000. It is the preferred approach when you suspect meaningful effect heterogeneity across a rich covariate set and want protection against misspecification of any single model. Do not use it when the sample is very small (doubly robust gains vanish and variance explodes), when the treatment is randomised and homogeneous effects are expected (a simple regression is sufficient), or when propensity scores approach 0 or 1 for many units (extreme positivity violations make the correction unstable).
Strengths & limitations
- Doubly robust: consistent if either the outcome model or the propensity score model is correctly specified, not both.
- Accommodates flexible machine learning nuisance estimators through cross-fitting, avoiding the bias-variance penalty that plagues naive plug-in approaches.
- Directly targets the full CATE function, enabling principled subgroup discovery and policy tailoring.
- Produces pseudo-outcomes that behave like ordinary regression targets, so any downstream learner can be applied with standard tools.
- Asymptotically efficient under the semiparametric efficiency bound when nuisance functions are estimated well.
- Requires large samples: nuisance function estimation, cross-fitting, and CATE regression each consume data, making small-n settings unreliable.
- Extreme propensity scores (near 0 or 1) make the augmentation term numerically unstable and inflate variance severely.
- The CATE surface can overfit if the final regression step is not regularised or cross-validated appropriately.
- Interpretation of a high-dimensional CATE function is non-trivial; naive subgroup comparisons from a fitted surface can be misleading without multiple-testing corrections.
Frequently asked
What makes this estimator 'doubly robust'?
The pseudo-outcome construction combines two models: the outcome model and the propensity score. If either model is correctly specified, the pseudo-outcome is unbiased for the true CATE, so you need only one of the two models to be right rather than both.
Why is cross-fitting necessary?
Cross-fitting (sample splitting) ensures that the nuisance functions are estimated on held-out data and not on the data used to compute pseudo-outcomes. Without it, flexible ML estimators can overfit in ways that invalidate the asymptotic theory and lead to biased CATE estimates.
How is this different from the standard doubly robust ATE estimator?
The standard doubly robust estimator targets a single average treatment effect (ATE). The HTE version regresses the doubly robust pseudo-outcome on covariates, so it targets a function tau(x) that maps covariates to conditional average effects. The ATE is just the average of tau(x) over the covariate distribution.
What sample size is reasonable?
A common rule of thumb is at least 500 observations for a simple CATE surface and 1,000 or more when covariates are numerous or propensity scores are extreme. Very small samples (n < 200) typically make the doubly robust augmentation term dominate with noise rather than reduce bias.
What if positivity is violated?
If some units have propensity scores very close to 0 or 1, trim those units from the analysis or use overlap weights. Running the full estimator with extreme propensity scores produces very large variance in the pseudo-outcomes and unstable CATE estimates.
Sources
- Kennedy, E. H. (2023). Towards optimal doubly robust estimation of heterogeneous causal effects. Electronic Journal of Statistics, 17(2), 3008-3049. DOI: 10.1214/23-EJS2157 ↗
- 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 3). Doubly Robust Estimation of Heterogeneous Treatment Effects. ScholarGate. https://scholargate.app/en/causal-inference/heterogeneous-treatment-effect-doubly-robust-estimation
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
- Inverse Probability WeightingCausal inference↔ compare
- Machine learning-augmented doubly robust estimationCausal inference↔ compare
- Marginal Structural ModelCausal inference↔ compare
- Propensity Score WeightingCausal inference↔ compare