Structural Double Robustness in Estimation

• Structural double robustness is a property ensuring that an estimator remains consistent and asymptotically normal even if only one of the nuisance parameters is correctly specified.
• It leverages variation-independent parameterizations and orthogonalization to achieve efficient inference in complex settings such as missing confounders, data fusion, and strategic models.
• Applications include GMM tests, causal effect estimation, and strategic equilibrium analysis, where structural DR significantly boosts robustness and efficiency against model misspecification.

Structural double robustness (DR) is a property of estimators and tests in semiparametric and nonparametric statistics that ensures consistency and/or asymptotic normality of the target parameter even when only one of two (or sometimes more) nuisance functions is correctly specified. Unlike standard double robustness—which typically concerns mean estimation under missing data or treatment assignment—structural DR addresses complex settings involving identification, functional estimation, high-dimensionality, misspecification, and strategic or structural interactions. Structural DR is pivotal in the construction of efficient estimators under model misspecification, the analysis of causal effects with missing or fused data, inference in generalized method of moments (GMM) with weak identification or misspecification, and in the theory of testing and optimality.

1. Foundations and Definition

Structural double robustness generalizes the classical double-robust (DR) framework: for a parameter $\psi$, dependent on two nuisance components $\eta_1, \eta_2$ (e.g., propensity score, outcome regression, selection, imputation, or other functional components), a DR estimator remains consistent and asymptotically normal if either $\eta_1$ or $\eta_2$ is correctly specified. This property is formalized for estimating functions $\phi(Z; \psi, \eta_1, \eta_2)$ such that, at the truth $P$, $E_P[\phi(Z; \psi(P), \eta_1(P), \eta_2(P))]=0$, and for all $P$, bias vanishes when either nuisance is correctly specified:

• $E_P[\phi(Z; \psi(P), \eta_1, \eta_2(P))]=0$ if $\eta_1=\eta_1(P)$,
• $\eta_1, \eta_2$0 if $\eta_1, \eta_2$1.

Structural DR departs from canonical settings in that it often involves:

• Variation-independent nuisance parameterizations, enabling each nuisance to be fit or misspecified independently (Evans et al., 2020),
• Complexity in the dependence or entanglement among nuisance components, e.g., in models with missing confounders, strategic equilibrium, or data fusion (Evans et al., 2018, Xiao, 17 Oct 2025),
• Tangent space orthogonality and geometric invariance across slices of the statistical model (Ying, 2024).

2. Geometric and Semiparametric Theory

The geometric theory of structural DR centers on the behavior of the efficient influence function (IC) and the structure of tangent spaces:

• If the parameterization $\eta_1, \eta_2$2 yields convex, variation-independent contours in the space of densities, then every IC is doubly robust in the sense that it is orthogonal to all nuisance perturbation directions at each slice (Ying, 2024).
• Information geometry provides further characterization: DR is equivalent to invariance of the IC under exponential (“e-”) parallel transport along contours of fixed nuisance, a consequence of m-flatness of the statistical manifold.
• Necessity and sufficiency conditions for DR rely on the orthogonality of the influence function to the tangent space of the model at every point along the nuisance contours (not just at the truth), requiring path-connectedness and variation-independence (Ying, 2024).

Practical consequence: In many common models (e.g., partially linear regression, average treatment effect estimation), convexity and variation-independence are satisfied, so standard influence function calculations yield structural DR “for free.” However, in settings such as semiparametric odds-ratio models with canonical parameterization, convexity may fail, precluding DR unless the parameterization is modified for orthogonality.

3. Construction and Parametrization of Structurally DR Estimators

The construction of structurally DR estimators exploits variation-independent parameterizations and orthogonalization of nuisance blocks. Key approaches include:

• Chen-type odds-ratio decomposition for models with missing confounders: By factoring the joint law of variables ($\eta_1, \eta_2$3) into primitive blocks (odds-ratio, marginals), independence among nuisance parameters is established. The observed data likelihood can be written so that only two “paths” to identification exist, corresponding to correct specification of either set of nuisances (Evans et al., 2020).
• Doubly robust GMM tests: For continuous-updating GMM, the DRLM test is constructed so as to be uniformly asymptotically valid whether misspecification, weak identification, or both occur. This is achieved by defining sample scores and quadratic forms whose weights reflect the joint variance structure under both failure modes (Kleibergen et al., 2021).
• Strategic Nash equilibrium models: DR estimation in strategic environments demands modeling both the equilibrium state (arising from inter-agent dependencies) and nuisance regression (propensity and outcome) (Xiao, 17 Oct 2025).

The general workflow for estimating a structurally DR parameter typically involves:

1. Fitting nuisance functions (possibly using cross-fitting or sample splitting to ensure independence of estimator errors).
2. Solving an estimating equation based on the efficient influence function, regularized likelihood, or GMM score that is robust to nuisance misspecification along either direction.
3. Propagating inference either via sandwich variance (plug-in) or nonparametric bootstrap (Evans et al., 2020, Kleibergen et al., 2021).

4. Theoretical Properties: Consistency, Efficiency, and Optimality

Structurally DR estimators guarantee:

• Consistency and asymptotic normality when either (but not necessarily both) of two prespecified nuisance components are estimated consistently (Evans et al., 2020, Evans et al., 2018, Robins et al., 2020).
• Efficiency at the intersection: When both models are correct, the estimator achieves the semiparametric efficiency bound, corresponding to the asymptotic variance of the efficient influence function (Evans et al., 2020).
• Robustness to structural failures: In GMM, DR inference remains valid regardless of the presence of weak identification, misspecification, or both (Kleibergen et al., 2021).

A summary of settings and robust properties is provided in the table:

Setting	Nuisance Blocks	DR Holds If	Efficiency Bound
Missing confounders	$\eta_1, \eta_2$4 or $\eta_1, \eta_2$5	Either block correct	Attained at intersection
Data fusion	$\eta_1, \eta_2$6	Either correct	Attained at intersection
Strategic equilibrium	$\eta_1, \eta_2$7	Either correct	Attained at intersection
GMM testing	Residual, Jacobian	Either valid	Valid under both failures

In each setting, the DR construction reduces to unbiasedness of the estimating function under independence of nuisance specifications.

5. Extensions: Smoothness, Cross-fitting, and Minimax Rates

Recent work explores the interaction of structural DR with sample splitting, smoothness, and minimax optimality:

• Cross-fitting and double cross-fit DCDR: Structural DR is preserved under sample splitting, enabling bias-cancellation even in high-dimensional or nonparametrically complex settings. Smoothness assumptions (Hölder regularity) enable estimators to achieve $\eta_1, \eta_2$8-rates or minimax rates depending on the sum of regressor smoothness indices relative to the dimension (McClean et al., 2024).
• Undersmoothing and sub-root-n inference: In low-regularity regimes, deliberate undersmoothing of nuisance fits enables correct normalization, yielding valid asymptotic inference even when the estimator does not achieve $\eta_1, \eta_2$9-consistency (McClean et al., 2024).
• Kernel and functional parameter settings: In distributional and RKHS-based causal effect estimation, structural DR estimators are naturally efficient and minimax-optimal for Hilbert-space–valued parameters, provided rate conditions on nuisance estimation hold (Jain et al., 17 Mar 2026).

6. Applications and Empirical Implications

Empirical performances of structural DR estimators have been demonstrated across domains:

• In missing confounder settings, the DR estimator is unbiased whenever either missingness/propensity blocks or joint marginal components are correct, matching theoretical predictions; all other estimators fail under partial misspecification (Evans et al., 2020).
• In data fusion and extreme missing-data problems (with zero probability of observing complete cases), DR estimators remain consistent under accurate selection or imputation modeling, rendering the approach vastly more robust than either inverse-probability weighting or imputation alone (Evans et al., 2018).
• In strategic games, bias reduction from DR estimators can exceed 25% relative to baselines, with robustness scaling to large agent populations and high-dimensional state spaces (Xiao, 17 Oct 2025).
• In GMM with weak identification or model misspecification, DR inference controls error rates where classical tests break down (Kleibergen et al., 2021).
• Empirical model selection leveraging the DR property (e.g., selecting the least variable ATE estimate over model pairs) outperforms standard cross-validation in high dimensions (Robins et al., 2020).

7. Practical Guidelines and Caveats

Construction and implementation of structurally DR procedures require:

• Careful choice of parameterization to ensure variation independence and convexity, enabling robust influence function structure (Evans et al., 2020, Ying, 2024).
• Routine checking of DR property by omitting nuisance models in simulation or using test statistics sensitive to DR behavior (Evans et al., 2020).
• Use of cross-fitting or sample splitting to mitigate overfitting and dependence among estimated nuisance functions (McClean et al., 2024).
• Application of bootstrap or influence-function–based standard errors for valid coverage.
• Strategic use of functional-analytic theory and information geometry for verifying conditions in novel models (Ying, 2024).

In sum, structural double robustness provides theoretical and practical machinery for robust and efficient estimation amid complex model structure, high-dimensionality, and incomplete data. Its efficacy and optimality are rooted in semiparametric theory, orthogonalization of nuisance directions, and geometric invariance, with a demonstrated range of applications across observational inference, econometrics, and modern causal analysis (Evans et al., 2020, Evans et al., 2018, Kleibergen et al., 2021, Ying, 2024, McClean et al., 2024, Xiao, 17 Oct 2025, Jain et al., 17 Mar 2026, Robins et al., 2020).