Doubly-Robust Estimation Framework

• Doubly-robust estimation is a framework that integrates outcome and propensity models to yield consistent estimates when at least one model is correctly specified.
• It enhances methodological rigor in causal inference, survey integration, and regression discontinuity without strict reliance on overlap assumptions.
• Recent advances include robust variance estimation and machine-learning extensions that improve bias correction and efficiency in practical applications.

Doubly-robust estimation is a statistical framework in which an estimator combines two nuisance components—most commonly an outcome model and an exposure or propensity model—so that consistency is retained if either component is correctly specified or consistently estimated. In causal inference this property is classically associated with augmented inverse probability weighting and targeted procedures for the average causal effect, but the same logic now appears in regression discontinuity designs, heterogeneous treatment effect estimation, nonprobability survey integration, left-truncated survival analysis, counterfactual distribution estimation in RKHSs, and even label aggregation and search relevance estimation (Shook-Sa et al., 2024).

1. Canonical formulation and core estimands

In observational causal inference, the standard nuisance pair is the conditional outcome regression and the exposure model. The outcome model is $E(Y \mid X, Z)$, the exposure model or propensity score is $e(Z)=\Pr(X=1\mid Z)$, and singly robust procedures include IPTW, which is consistent if the propensity score model is correct, and g-computation, which is consistent if the outcome model is correct. Doubly robust estimators such as AIPW and TMLE combine both models so that consistency and asymptotic normality of the estimator is achieved if either, not necessarily both, of the two models is correctly specified (Shook-Sa et al., 2024).

For the average treatment effect, one standard doubly robust form is the augmented inverse probability weighting estimator

$$\hat{\tau}_{AIPW} = \frac{1}{n}\sum_{i=1}^n \left[ \hat{\mu}_{(1)}(X_i)-\hat{\mu}_{(0)}(X_i) + \frac{W_i}{\hat{e}(X_i)}(Y_i-\hat{\mu}_{(1)}(X_i)) - \frac{1-W_i}{1-\hat{e}(X_i)}(Y_i-\hat{\mu}_{(0)}(X_i)) \right],$$

and, in the multiple-imputation setting, an equivalent AIPW representation is

$$\begin{aligned} \hat{\Delta}_{\text{DR}} = &\frac{1}{n} \sum_{i=1}^n \left[\mu_1(\mathbf{Z}_{C,i})-\mu_0(\mathbf{Z}_{C,i})\right] \ &+ \frac{1}{n} \sum_{i=1}^n\left[ \frac{X_i\left\{Y_i-\mu_1(\mathbf{Z}_{C,i})\right\}}{\pi(\mathbf{Z}_{C,i})} - \frac{(1-X_i)\left\{Y_i-\mu_0(\mathbf{Z}_{C,i})\right\}}{1-\pi(\mathbf{Z}_{C,i})} \right]. \end{aligned}$$

In both formulations, the augmentation term corrects the inverse-weighted component by subtracting residuals relative to the outcome model (Mayer et al., 2019, McGowan, 13 Oct 2025).

A distinct theoretical perspective is the structure-agnostic framework, which imposes no structural properties on the nuisance functions other than access to black-box estimators with specified $L^2$ error rates. Within this framework, doubly robust estimators for the ATE, ATT, and weighted ATE are minimax-optimal. The paper establishes squared-error rates

$$|\hat{\theta}^{WATE} - \theta^{WATE}|^2 \leq C\left[ \max(e_n,e_n')\,f_n\,\|w\|_{P_X,\infty}^2 + \frac{1}{n}\|w\|_{P_X,2}^2 \right]$$

for WATE and

$$|\hat{\theta}^{ATT} - \theta^{ATT}|^2 \leq C\left[e_n f_n + \frac{1}{n}\right]$$

for ATT, with matching lower bounds, so that no estimator using only black-box regression and classification oracles can improve those rates in the stated regime (Jin et al., 2024).

2. Orthogonality, augmentation, and efficient influence functions

A defining technical feature of many doubly robust constructions is an orthogonal or debiased estimating equation. In the automatic Doubly Robust Random Forest, the debiased moment is

$$\psi(Z;\alpha,g) := m(Z; g(X, W)) + \alpha(X, W)(Y - g(X, W)),$$

where $\alpha$ is the Riesz representer. Its conditional bias satisfies

$$E[\psi(Z; \hat{\alpha}, \hat{g}) \mid X=x] - \theta_0(x) = - E[(\hat{\alpha}(x,W)-\alpha_0(x,W))(\hat{g}(x,W)-g_0(x,W))\mid X=x].$$

This formula makes the product-error structure explicit: first-order errors in either nuisance component vanish, and consistency follows if the product term is controlled (Chen et al., 2024).

Semiparametric efficient influence functions provide a parallel route to doubly robust construction. For Counterfactual Policy Mean Embeddings, the efficient influence function for the RKHS-valued target $e(Z)=\Pr(X=1\mid Z)$0 is

$e(Z)=\Pr(X=1\mid Z)$1

yielding a doubly robust estimator that is consistent if either the outcome embedding or the logging propensity is correctly specified. The same paper derives improved uniform convergence rates and an asymptotically normal doubly robust kernel test statistic for hypothesis testing (Zenati et al., 3 Jun 2025).

An analogous one-step logic appears in conditional distributional treatment effects. There the estimand is an RKHS element

$e(Z)=\Pr(X=1\mid Z)$2

and the efficient influence function is

$e(Z)=\Pr(X=1\mid Z)$3

A cross-fitted one-step estimator based on this influence function is doubly robust and locally minimax optimal in the local asymptotic sense (Jain et al., 17 Mar 2026).

Regression discontinuity designs admit a related two-stage correction. In the DR-RD estimator for a sharp RD design, treatment is assigned by $e(Z)=\Pr(X=1\mid Z)$4, the target is

$e(Z)=\Pr(X=1\mid Z)$5

and the augmented observation-level term is

$e(Z)=\Pr(X=1\mid Z)$6

Consistency holds if either the conditional outcome regression or the cutoff-localized residual regression is correct (Kato, 2024).

3. Variance estimation, confidence procedures, and posterior uncertainty

Point-estimator double robustness does not automatically imply variance double robustness. For nonrandomized exposures, the influence function based variance estimator

$e(Z)=\Pr(X=1\mid Z)$7

is only consistent when both working models are correctly specified. By contrast, the empirical sandwich variance estimator obtained from unbiased estimating equations,

$e(Z)=\Pr(X=1\mid Z)$8

and the nonparametric bootstrap,

$e(Z)=\Pr(X=1\mid Z)$9

are demonstrated to be doubly robust variance estimators. In simulations, empirical sandwich and bootstrap track empirical standard error and deliver nominal confidence interval coverage whenever either the outcome or exposure model is correct, whereas the influence function estimator is conservative when the outcome model is misspecified and anti-conservative when the propensity model is misspecified (Shook-Sa et al., 2024).

For heterogeneous treatment effects, a posterior-averaging frequentist procedure uses posterior distributions for both the propensity score and outcome regression models, forms pseudo-outcomes for each posterior draw, and estimates the CATE by averaging over those draws. Its variance estimator is decomposed into a bootstrap-based sampling term and a posterior-variance term,

$$\hat{\tau}_{AIPW} = \frac{1}{n}\sum_{i=1}^n \left[ \hat{\mu}_{(1)}(X_i)-\hat{\mu}_{(0)}(X_i) + \frac{W_i}{\hat{e}(X_i)}(Y_i-\hat{\mu}_{(1)}(X_i)) - \frac{1-W_i}{1-\hat{e}(X_i)}(Y_i-\hat{\mu}_{(0)}(X_i)) \right],$$0

and is conservative in finite samples or under model misspecification while being consistent when both nuisance models are correctly specified and contract faster than $$\hat{\tau}_{AIPW} = \frac{1}{n}\sum_{i=1}^n \left[ \hat{\mu}_{(1)}(X_i)-\hat{\mu}_{(0)}(X_i) + \frac{W_i}{\hat{e}(X_i)}(Y_i-\hat{\mu}_{(1)}(X_i)) - \frac{1-W_i}{1-\hat{e}(X_i)}(Y_i-\hat{\mu}_{(0)}(X_i)) \right],$$1 (Shin et al., 2021).

A Bayesian semiparametric extension replaces full-data modeling with a targeted modeling strategy based on summary statistics. In DRDB, sample splitting and cross-fitting separate nuisance estimation from target-parameter inference, explicit Bayesian debiasing estimates nuisance-induced bias via weighted observables, and the marginal posterior for the ATE satisfies a Bernstein-von Mises theorem when both nuisance models are correctly specified. The posterior remains consistent and robust when only one nuisance model is correct, which the paper terms Bayesian double robustness (Sert et al., 19 Nov 2025).

4. Research designs and data complications

The doubly-robust framework has been adapted to a wide range of designs in which the nuisance pair differs from the standard outcome-model/propensity-score decomposition. In regression discontinuity, the DR-RD estimator combines any regression method for $$\hat{\tau}_{AIPW} = \frac{1}{n}\sum_{i=1}^n \left[ \hat{\mu}_{(1)}(X_i)-\hat{\mu}_{(0)}(X_i) + \frac{W_i}{\hat{e}(X_i)}(Y_i-\hat{\mu}_{(1)}(X_i)) - \frac{1-W_i}{1-\hat{e}(X_i)}(Y_i-\hat{\mu}_{(0)}(X_i)) \right],$$2—including Nadaraya–Watson, local linear regression, series regression, and neural nets—with a Nadaraya–Watson kernel estimator of the conditional mean residual at the cutoff. A notable distinction from standard doubly robust treatment-effect estimation in observational studies is that the DR-RD estimator does not require overlap or common support around the cutoff (Kato, 2024).

Weak overlap motivates a different modification. A new class of doubly robust estimators trims observations with extreme propensity scores and then corrects trimming bias by polynomial approximation of the conditional expectation around zero. The bias-corrected trimmed mean

$$\hat{\tau}_{AIPW} = \frac{1}{n}\sum_{i=1}^n \left[ \hat{\mu}_{(1)}(X_i)-\hat{\mu}_{(0)}(X_i) + \frac{W_i}{\hat{e}(X_i)}(Y_i-\hat{\mu}_{(1)}(X_i)) - \frac{1-W_i}{1-\hat{e}(X_i)}(Y_i-\hat{\mu}_{(0)}(X_i)) \right],$$3

preserves the original estimand rather than redefining the target population, while accommodating unconfoundedness, local treatment effects, and difference-in-differences (Ma et al., 2023).

Missingness produces several further variants. With missing covariates, the generalized propensity score conditions on both observed covariate values and missingness patterns, and the corresponding doubly robust estimator is

$$\hat{\tau}_{AIPW} = \frac{1}{n}\sum_{i=1}^n \left[ \hat{\mu}_{(1)}(X_i)-\hat{\mu}_{(0)}(X_i) + \frac{W_i}{\hat{e}(X_i)}(Y_i-\hat{\mu}_{(1)}(X_i)) - \frac{1-W_i}{1-\hat{e}(X_i)}(Y_i-\hat{\mu}_{(0)}(X_i)) \right],$$4

Simulation results reported in that study show that doubly robust modifications repeatedly perform better than non-doubly robust baselines (Mayer et al., 2019).

For nonprobability survey samples, a standard form is

$$\hat{\tau}_{AIPW} = \frac{1}{n}\sum_{i=1}^n \left[ \hat{\mu}_{(1)}(X_i)-\hat{\mu}_{(0)}(X_i) + \frac{W_i}{\hat{e}(X_i)}(Y_i-\hat{\mu}_{(1)}(X_i)) - \frac{1-W_i}{1-\hat{e}(X_i)}(Y_i-\hat{\mu}_{(0)}(X_i)) \right],$$5

which is consistent for the finite population mean if either the propensity model or the outcome model is correctly specified. The same literature also emphasizes that plug-in variance estimators are not doubly robust and recommends the Kim-Haziza doubly robust variance estimator based on joint estimating equations (Chen et al., 2018).

In the later survey-integration framework, the doubly robust estimate from a nonprobability sample can be combined with an estimate based on the auxiliary probability sample alone: $$\hat{\tau}_{AIPW} = \frac{1}{n}\sum_{i=1}^n \left[ \hat{\mu}_{(1)}(X_i)-\hat{\mu}_{(0)}(X_i) + \frac{W_i}{\hat{e}(X_i)}(Y_i-\hat{\mu}_{(1)}(X_i)) - \frac{1-W_i}{1-\hat{e}(X_i)}(Y_i-\hat{\mu}_{(0)}(X_i)) \right],$$6 with optimal weight

$$\hat{\tau}_{AIPW} = \frac{1}{n}\sum_{i=1}^n \left[ \hat{\mu}_{(1)}(X_i)-\hat{\mu}_{(0)}(X_i) + \frac{W_i}{\hat{e}(X_i)}(Y_i-\hat{\mu}_{(1)}(X_i)) - \frac{1-W_i}{1-\hat{e}(X_i)}(Y_i-\hat{\mu}_{(0)}(X_i)) \right],$$7

This turns variance and covariance estimation into a central part of efficient combination (Seaman et al., 7 Aug 2025).

Time-to-event settings with dependent left truncation require yet another nuisance pair. There the efficient influence function contains an IPW component plus an augmentation term involving a reverse-time martingale for the truncation process, and the resulting estimator is consistent if either the event model $$\hat{\tau}_{AIPW} = \frac{1}{n}\sum_{i=1}^n \left[ \hat{\mu}_{(1)}(X_i)-\hat{\mu}_{(0)}(X_i) + \frac{W_i}{\hat{e}(X_i)}(Y_i-\hat{\mu}_{(1)}(X_i)) - \frac{1-W_i}{1-\hat{e}(X_i)}(Y_i-\hat{\mu}_{(0)}(X_i)) \right],$$8 or the truncation model $$\hat{\tau}_{AIPW} = \frac{1}{n}\sum_{i=1}^n \left[ \hat{\mu}_{(1)}(X_i)-\hat{\mu}_{(0)}(X_i) + \frac{W_i}{\hat{e}(X_i)}(Y_i-\hat{\mu}_{(1)}(X_i)) - \frac{1-W_i}{1-\hat{e}(X_i)}(Y_i-\hat{\mu}_{(0)}(X_i)) \right],$$9 is correct. The same work extends the framework to right censoring before truncation and right censoring after truncation (Wang et al., 2022).

5. Machine-learning, latent-structure, and domain-specific expansions

Recent work has embedded doubly robust estimation into highly adaptive nuisance-learning pipelines. In causal latent factor models, outcome imputation and propensity estimation are both performed by matrix completion under a low-rank latent-factor structure, and a novel cross-fitting procedure ensures that each imputed value is estimated without using its own noise term. The resulting estimator

$$\begin{aligned} \hat{\Delta}_{\text{DR}} = &\frac{1}{n} \sum_{i=1}^n \left[\mu_1(\mathbf{Z}_{C,i})-\mu_0(\mathbf{Z}_{C,i})\right] \ &+ \frac{1}{n} \sum_{i=1}^n\left[ \frac{X_i\left\{Y_i-\mu_1(\mathbf{Z}_{C,i})\right\}}{\pi(\mathbf{Z}_{C,i})} - \frac{(1-X_i)\left\{Y_i-\mu_0(\mathbf{Z}_{C,i})\right\}}{1-\pi(\mathbf{Z}_{C,i})} \right]. \end{aligned}$$0

is shown to have finite-sample guarantees and asymptotic Gaussian error at a parametric rate (Abadie et al., 2024).

Bayesian model synthesis provides a different response to nuisance misspecification. In Bayesian regression synthesis for doubly robust estimation, multiple candidate models for both the propensity score and the outcome are combined through unit-level, covariate-dependent weights, with Bayesian updating applied to the ensemble model weights. The propensity synthesis model is logistic,

$$\begin{aligned} \hat{\Delta}_{\text{DR}} = &\frac{1}{n} \sum_{i=1}^n \left[\mu_1(\mathbf{Z}_{C,i})-\mu_0(\mathbf{Z}_{C,i})\right] \ &+ \frac{1}{n} \sum_{i=1}^n\left[ \frac{X_i\left\{Y_i-\mu_1(\mathbf{Z}_{C,i})\right\}}{\pi(\mathbf{Z}_{C,i})} - \frac{(1-X_i)\left\{Y_i-\mu_0(\mathbf{Z}_{C,i})\right\}}{1-\pi(\mathbf{Z}_{C,i})} \right]. \end{aligned}$$1

and the outcome synthesis model is linear-Gaussian,

$$\begin{aligned} \hat{\Delta}_{\text{DR}} = &\frac{1}{n} \sum_{i=1}^n \left[\mu_1(\mathbf{Z}_{C,i})-\mu_0(\mathbf{Z}_{C,i})\right] \ &+ \frac{1}{n} \sum_{i=1}^n\left[ \frac{X_i\left\{Y_i-\mu_1(\mathbf{Z}_{C,i})\right\}}{\pi(\mathbf{Z}_{C,i})} - \frac{(1-X_i)\left\{Y_i-\mu_0(\mathbf{Z}_{C,i})\right\}}{1-\pi(\mathbf{Z}_{C,i})} \right]. \end{aligned}$$2

The paper shows consistency of the synthesized propensity and outcome estimators and argues that the doubly robust estimator is consistent even if no single candidate model is correctly specified, provided the truth is represented by the ensemble (Babasaki et al., 2024).

Outside conventional treatment-effect estimation, the same architecture reappears in several domains. In crowdsourcing, the doubly robust crowdsourcing estimator combines worker-imitating supervised learners with subsampled real labels,

$$\begin{aligned} \hat{\Delta}_{\text{DR}} = &\frac{1}{n} \sum_{i=1}^n \left[\mu_1(\mathbf{Z}_{C,i})-\mu_0(\mathbf{Z}_{C,i})\right] \ &+ \frac{1}{n} \sum_{i=1}^n\left[ \frac{X_i\left\{Y_i-\mu_1(\mathbf{Z}_{C,i})\right\}}{\pi(\mathbf{Z}_{C,i})} - \frac{(1-X_i)\left\{Y_i-\mu_0(\mathbf{Z}_{C,i})\right\}}{1-\pi(\mathbf{Z}_{C,i})} \right]. \end{aligned}$$3

and the paper proves that the estimator is always unbiased and that even a random surrogate grants $$\begin{aligned} \hat{\Delta}_{\text{DR}} = &\frac{1}{n} \sum_{i=1}^n \left[\mu_1(\mathbf{Z}_{C,i})-\mu_0(\mathbf{Z}_{C,i})\right] \ &+ \frac{1}{n} \sum_{i=1}^n\left[ \frac{X_i\left\{Y_i-\mu_1(\mathbf{Z}_{C,i})\right\}}{\pi(\mathbf{Z}_{C,i})} - \frac{(1-X_i)\left\{Y_i-\mu_0(\mathbf{Z}_{C,i})\right\}}{1-\pi(\mathbf{Z}_{C,i})} \right]. \end{aligned}$$4 variance reduction (Liu et al., 2019).

In search relevance estimation, the doubly robust estimator unifies counterfactual click correction and semantic imputation: $$\begin{aligned} \hat{\Delta}_{\text{DR}} = &\frac{1}{n} \sum_{i=1}^n \left[\mu_1(\mathbf{Z}_{C,i})-\mu_0(\mathbf{Z}_{C,i})\right] \ &+ \frac{1}{n} \sum_{i=1}^n\left[ \frac{X_i\left\{Y_i-\mu_1(\mathbf{Z}_{C,i})\right\}}{\pi(\mathbf{Z}_{C,i})} - \frac{(1-X_i)\left\{Y_i-\mu_0(\mathbf{Z}_{C,i})\right\}}{1-\pi(\mathbf{Z}_{C,i})} \right]. \end{aligned}$$5 Its bias vanishes if either the bias model is correct or the imputation model is correct, and its production instantiation uses click behavior tracking, an examination model, and approximated neural components for online relevance estimation (Zou et al., 2022).

6. Misconceptions, limitations, and methodological cautions

A persistent misconception is that double robustness protects against arbitrary misspecification. It does not. When both nuisance models are wrong, doubly robust point estimators can be biased and confidence intervals can be poorly calibrated; simulation studies in several papers explicitly report misleading results under dual misspecification (Shook-Sa et al., 2024).

A second misconception is that valid point estimation automatically yields valid uncertainty quantification. In the parametric working-model setting, the standard influence function based variance estimator is not doubly robust, and its behavior changes qualitatively with the direction of misspecification: conservative when the outcome model is misspecified, anti-conservative when the propensity model is misspecified (Shook-Sa et al., 2024).

A third caution concerns missing data. When multiple imputation is combined with doubly robust causal estimation, the imputation model must include all variables used in either the propensity score model or the outcome model, and these variables must appear in the same functional form as in the final analysis. Violating these conditions can lead to biased treatment effect estimates even when both components of the doubly robust estimator are correctly specified. The same paper further states that imputation models should be fit separately within each exposure group when outcome models are stratified by treatment, and that complete case analysis is usually unsafe (McGowan, 13 Oct 2025).

Finally, overlap and support requirements are design-specific rather than universal. Standard observational doubly robust estimators are sensitive to weak overlap, prompting trimming and bias correction devices, whereas the DR-RD estimator for sharp regression discontinuity is notable precisely because it does not require overlap or common support around the cutoff (Ma et al., 2023, Kato, 2024).

The doubly-robust estimation framework is therefore best understood not as a single estimator but as a design principle: combine two nuisance representations of the target, construct an orthogonal or influence-function-based correction, and exploit the fact that first-order bias disappears when at least one nuisance component is correct. The modern literature shows that this principle extends from ATE estimation to CATEs, policy mean embeddings, distributional treatment effects, survey inference, truncation-adjusted survival functionals, crowdsourcing, and web search, while simultaneously clarifying the narrow conditions under which its robustness claims remain valid.