Doubly Robust Estimation Methods

Updated 22 June 2026

Doubly robust estimation procedures are statistical methods that combine propensity score and outcome models to ensure consistent estimates if at least one model is correctly specified.
They utilize orthogonalization and influence-function theory to effectively reduce bias and achieve semiparametric efficiency across diverse applications like causal inference and off-policy evaluation.
Practical implementations employ sample splitting, cross-fitting, and regularization to address high-dimensional challenges and maintain reliability in estimation.

Doubly robust (DR) estimation procedures provide a class of methods for estimating statistical and causal functionals that are consistent if at least one of two nuisance models—typically a propensity or density ratio model, and an outcome or regression model—is correctly specified. Originating in causal inference, missing data, off-policy evaluation, covariate shift adaptation, and survey integration contexts, the DR property allows estimators to remain consistent and, under favorable conditions, efficient, even when high-dimensional or nonparametric models are employed for one or both nuisances. The DR approach fundamentally exploits orthogonalization and influence-function theory to ensure that the leading bias term vanishes when either nuisance estimate is accurate, and is central to modern semiparametric and double/debiased machine learning frameworks (Guo et al., 2020, Smucler et al., 2019, Tang et al., 2019, Díaz et al., 2017).

1. Conceptual Foundations and Definitions

A doubly robust estimator for a target parameter θ is constructed by combining two sets of models (e.g., propensity score and outcome regression), such that consistency and asymptotic normality of the estimator are achieved if either nuisance model is consistently estimated, but not necessarily both (Guo et al., 2020, Tang et al., 2019, Díaz et al., 2017). Classical motivation comes from the average treatment effect (ATE), where the DR estimator for Δ = E[Y(1)] − E[Y(0)] is: $\hat\Delta_{DR} = \frac{1}{n} \sum_{i=1}^n \left[ \frac{D_i Y_i}{\hat e(X_i)} - \frac{(1-D_i) Y_i}{1-\hat e(X_i)} + \left(1 - \frac{D_i}{\hat e(X_i)}\right)\hat m_1(X_i) - \left(1 - \frac{1-D_i}{1-\hat e(X_i)}\right)\hat m_0(X_i) \right]$ where $\hat e(\cdot)$ estimates the propensity score $P(D=1|X)$ and $\hat m_t(\cdot)$ estimates $E[Y|X,D=t]$ , $t=0,1$ (Guo et al., 2020). The DR property is characterized by a bias of order $\epsilon_e\epsilon_m$ (the product of nuisance estimation errors), while singly robust methods have linear bias (Tang et al., 2019).

The DR principle generalizes to settings with non-binary treatments, continuous exposures, missing data (via missingness modeling and imputation), survey integration, off-policy RL, strategic equilibrium causal systems, and causal panel or high-dimensional difference-in-differences (Kennedy et al., 2015, Abadie et al., 2024, Kato et al., 2023, Tang et al., 2019, Xiao, 17 Oct 2025, Ning et al., 2020). In recent Bayesian and ensemble approaches, DR functionals are evaluated by probabilistically synthesizing multiple nuisance models; consistency is attained provided at least one synthesized channel is correct (Babasaki et al., 2024).

2. Theoretical Properties and Efficiency Comparisons

When both nuisance models are correctly specified and estimated at appropriate rates (parametric or sufficiently strong nonparametric consistency), the DR estimator attains the semiparametric efficiency bound for the given problem. For ATE,

$\sqrt{n}(\hat\Delta_{DR}-\Delta) \to_d N\left(0, V^*\right)$

with

$V^* = E\left[ \frac{\text{Var}(Y(1)|X)}{e(X)} + \frac{\text{Var}(Y(0)|X)}{1-e(X)} + (m_1(X)-m_0(X) - \Delta)^2 \right]$

(Guo et al., 2020).

Double robustness holds: when only one model is correct (either propensity or outcome), the estimator remains consistent, but typically with inflated variance compared to the efficiency bound. Critically, when at least one nuisance is estimated nonparametrically, the estimator can "absorb" misspecification from the other, and efficiency is often preserved (Guo et al., 2020, Ye et al., 2020). However, if both are parametric/semi-parametric and one is misspecified, the variance may increase—or, in the case of propensity misspecification but correct regression, can be unexpectedly smaller than the efficiency bound (super-efficiency) (Guo et al., 2020). The precise behavior is captured in a taxonomy of nine PS×OR combinations, covering parametric, nonparametric, and semiparametric fits (Guo et al., 2020, Ye et al., 2020).

In the infinite-horizon off-policy setting, the bias of the DR estimator is $\mathbb E_{d_{\pi_0}}[\epsilon_w(s)\epsilon_V(s)]$ , second-order in the errors $\hat e(\cdot)$ 0 (density ratio) and $\hat e(\cdot)$ 1 (Bellman residual), and vanishes if either is zero (Tang et al., 2019). Similar bilinear bias structures occur in panel data (Arkhangelsky et al., 2019), survey integration (Seaman et al., 7 Aug 2025), and covariate-shift adaptation (Kato et al., 2023). In missing data, consistency holds under MAR if either outcome or missingness model is correct (Díaz et al., 2017, Molina et al., 2017).

3. Methodological Realizations Across Domains

Causal Inference

In standard observational studies, the DR estimator combines inverse probability weighting (IPW) and regression adjustment: $\hat e(\cdot)$ 2 Extensions include conditional treatment effects (Shin et al., 2021), instrumental variables (LATE) with IPWRA (Słoczyński et al., 2022), and quantiles (Molina et al., 2017). For high-dimensional data, sample splitting, cross-fitting, and $\hat e(\cdot)$ 3-regularization yield root-n consistent estimators when product sparsity is controlled (Smucler et al., 2019).

Off-Policy Evaluation in RL

For infinite-horizon policy evaluation in Markov Decision Processes (MDPs), DR estimators correct the bias of density‐ratio based importance sampling by leveraging a learned value function, achieving low variance and second-order bias (Tang et al., 2019).

Covariate Shift and Domain Adaptation

Doubly robust estimators for covariate shift correct bias from errors in estimated density ratios $\hat e(\cdot)$ 4 by augmenting the loss with a regression model, ensuring consistency if either nuisance is estimated accurately (Kato et al., 2023).

Panel Data and Difference-in-Differences

Doubly robust estimators for panel or dynamic treatment assignment settings model both the outcome (possibly with fixed effects) and the treatment (possibly with unit effects), obtaining identification under either approach and efficiency if both models are correct (Arkhangelsky et al., 2019, Ning et al., 2020).

Survey Sampling and Data Integration

For population mean or prevalence estimation using nonprobability and probability survey data, DR estimators combine IPW (sample selection model) and mass imputation (outcome model). Consistency and variance results hold under correct specification of at least one model (Seaman et al., 7 Aug 2025, Chen et al., 2018).

Missing Data and Targeted Learning

DR estimators are essential in targeted minimum loss estimation (TMLE), where targeted updates produce n^{{1/2}-consistent} and asymptotically normal estimators under correct estimation of either the missingness or outcome regression, even with data-adaptive/flexible nuisance fits (Díaz et al., 2017).

Strategic and Latent Factor Causal Systems

In strategic games, the strategic doubly robust (SDR) estimator corrects for endogenous treatment assignment induced by Nash equilibrium, generalizing DR structure by conditioning propensity and regression models on the equilibrium state (Xiao, 17 Oct 2025). For latent factor models, DR estimators combine matrix completion-based regression and IPW components, achieving parametric rates if at least one low-rank structure is learned well (Abadie et al., 2024).

4. Estimation, Inference, and Implementation

The core components of doubly robust estimation are:

Nuisance Estimation: Flexible ML or semi-/non-parametric models for propensity/density ratio and regression/outcome, often with sample splitting (cross-fitting) to reduce bias from overfitting and remove empirical-process restrictions (Smucler et al., 2019, Kato et al., 2023, Ning et al., 2020).
Plug-in DR Formula: Evaluation of the influence-function estimator using out-of-sample fitted nuisance functions.
Variance Estimation:
- Influence-function-based variance estimators are only doubly robust when derived from the empirical M-estimation system (Shook-Sa et al., 2024).
- Empirical sandwich variance estimators and the nonparametric bootstrap are doubly robust: they consistently estimate variance for Wald-type confidence intervals when either nuisance is correctly specified (Shook-Sa et al., 2024).
- Plug-in/IF-based variance estimators may be conservative or anti-conservative under misspecification.
Bias-Reduction/Augmentation: In off-policy RL and strategic settings, corrections (such as via Bellman residual or strategic equilibrium state) further mitigate bias by leveraging problem structure (Tang et al., 2019, Xiao, 17 Oct 2025).
High-dimensionality/Sparsity: Use of $\hat e(\cdot)$ 5-penalized regression for one or both nuisances, with double robustness preserved so long as the product of sparsities is small (Smucler et al., 2019).
Practical Recommendations: Preference for nonparametric/sufficiently flexible estimation of at least one nuisance, sample splitting, regularization, and care with positivity/overlap enforcement.

5. Applications, Extensions, and Limitations

Doubly robust estimators have been empirically validated across randomized trial missing data (Díaz et al., 2017), policy evaluation in RL (Tang et al., 2019), survey integration (Seaman et al., 7 Aug 2025), strategic causal inference (Xiao, 17 Oct 2025), and covariate shift (Kato et al., 2023). Key extension domains include:

Continuous treatments: Nonparametric kernel DR estimators for dose–response surfaces (Kennedy et al., 2015).
Panel and difference-in-differences settings: DR estimation for dynamic or staggered adoption designs (Arkhangelsky et al., 2019, Ning et al., 2020).
Missing data and quantiles: DR estimation for population quantiles under MAR and robustified for outliers (Molina et al., 2017).
Latent factor and high-dimensional models: DR estimators leveraging low-rank matrix completion under unmeasured confounding (Abadie et al., 2024, Smucler et al., 2019).

Practical limitations include the necessity for overlap (positivity) in the distributions of treatment, strategic, or sampling mechanisms, the computational burden of cross-fitting and equilibrium-finding, and potential inefficiency or variance inflation if both nuisance models are severely misspecified. When both are misspecified, DR estimators may exhibit nonvanishing bias; bias-correction is sometimes possible using analytical formulas for the second-order bias (Ye et al., 2020).

Recent advances propose Bayesian model synthesis—combining multiple agents for propensity and outcome via adaptive weights—achieving DR consistency even if neither model is individually correct (Babasaki et al., 2024).

6. Central Insights and Ongoing Directions

Doubly robust estimation unifies distinct identification and modeling strategies across causal inference, machine learning, reinforcement learning, and survey analysis. Its defining bias and variance decompositions clarify the regimes for root-n consistency and efficiency, and inform the design of robust, flexible estimators under weak assumptions. The DR paradigm continues to be extended to high-dimensional and non-standard settings via advanced regression/device learning, Bayesian ensembling, and strategic equilibrium modeling (Xiao, 17 Oct 2025, Babasaki et al., 2024).

The DR approach has played a foundational role in the recent shift toward plug-in and orthogonalized (Neyman-orthogonal) machine learning methods for estimate and inference on target statistical/causal functionals, particularly in the presence of high-dimensional and complex nuisance structures, solidifying its centrality in modern semiparametric and nonparametric statistics (Smucler et al., 2019, Shin et al., 2021, Guo et al., 2020).