Structural Double Robustness in Estimation

• Structural double robustness is a property ensuring an estimator remains consistent when at least one nuisance component is correctly specified, with bias reducing to a second-order product term.
• It is grounded in semiparametric and information-geometric methods, where local orthogonality and convex contour set properties guarantee global invariance.
• Applications include average treatment effect estimation, partially linear models, and instrumental variable frameworks, while also addressing limitations like double fragility.

Searching arXiv for papers on structural double robustness and related double robustness theory. Structural double robustness denotes a class of results in which the global double robustness of an estimator is not treated as an ad hoc algebraic accident, but as a consequence of the geometry or structural organization of a statistical model and its parameterization. In its most explicit formulation, the property means that the global DR behavior of an estimator—consistency if either nuisance component is correct, together with second-order bias of the form $E[\Delta_1(Z)\Delta_2(Z)]$—is guaranteed by variation independence, convexity or $m$-flatness of contour sets, $\theta$-connectedness, and mean-square differentiability; under these conditions, local influence-function orthogonality transports to global invariance along nuisance submanifolds (Ying, 2024). Related papers use the same label, or closely allied constructions, for orthogonality-based root-$n$ inference with $n^{-1/4}$-rate nuisances, high-dimensional rate and model double robustness, doubly-robust asymptotic linearity in hybrid smoothness classes, and inference procedures that remain valid under misspecification and weak identification (Lok, 2024, Smucler et al., 2019, Bonvini et al., 2024, Kleibergen et al., 2021).

1. Core definition and bias structure

In the semiparametric formulation of "A Geometric Perspective on Double Robustness by Semiparametric Theory and Information Geometry," the model $M$ is a set of probability laws $p$ for observations, the target parameter is $\psi(P)$, and estimation is organized through population estimating functions $\phi(X;\theta,\gamma_1,\gamma_2)$. An adaptive estimating function satisfies $E[\phi(X;\theta(p),\gamma(p))]=0$ and identifies $m$0. It is doubly robust if it remains unbiased when either nuisance component is fixed at truth: $m$1 for any $m$2. At the estimator level, this yields consistency if either nuisance is correctly specified or consistently estimated. Under regularity, the first-order bias cancels and the leading bias takes the product form

$m$3

so the bias vanishes if either $m$4 or $m$5 (Ying, 2024).

A later note, "Demystified: double robustness with nuisance parameters estimated at rate $m$6," distinguishes two senses of double robustness. The first is the classical misspecification notion: for a moment function $m$7 with nuisance partition $m$8,

$m$9

The second is structural double robustness: orthogonality of the score with respect to nuisance directions, expressed as

$\theta$0

This condition removes the first-order effect of nuisance estimation error on the estimator for $\theta$1, so only second-order nuisance errors enter the asymptotics (Lok, 2024).

These two formulations are aligned. The global unbiasedness conditions describe the estimator’s behavior over entire nuisance submodels, while the derivative condition describes local insensitivity at the truth. Structural double robustness is the set of conditions under which the latter implies the former, and under which product-bias remainders become the operative asymptotic object rather than a merely heuristic slogan (Ying, 2024, Lok, 2024).

2. Semiparametric and information-geometric foundations

The semiparametric starting point is the tangent-space calculus. For a model $\theta$2, the tangent space $\theta$3 is the closure of the span of scores $\theta$4 from smooth curves through $\theta$5. The nuisance tangent space $\theta$6 is the closed span of scores along which the target parameter does not change to first order. If $\theta$7 is mean-square differentiable at $\theta$8, an influence curve $\theta$9 represents the pathwise derivative through

$n$0

Local Neyman orthogonality is then the condition

$n$1

Proposition 1 states that any doubly robust estimating function $n$2 lies in the orthogonal complement $n$3 (Ying, 2024).

The central theoretical problem is that influence curves are local objects, whereas double robustness is global, since it requires invariance of the estimating equation along entire nuisance directions. The 2024 geometric paper resolves this by introducing contour sets

$n$4

and analogously $n$5, together with variation independence of $n$6. Theorem 1 states: if the parameterization is variation independent and both contour sets are convex for every $n$7, then any influence curve is doubly robust. Convexity is therefore the structural property that permits local orthogonality to imply global DR "for free" (Ying, 2024).

The paper then gives necessary and sufficient existence results. Under mean-square differentiability and $n$8-connectedness—smooth path-connectedness of each contour set—Theorem 2 shows that any DR estimating function must be orthogonal to the tangent spaces of both contour sets at all their points: $n$9 Theorem 3 shows that, under variation independence and $n^{-1/4}$0-connectedness, this condition is also sufficient for double robustness (Ying, 2024).

The information-geometric reformulation sharpens the same point. The paper defines $n^{-1/4}$1-parallel transport

$n^{-1/4}$2

and $n^{-1/4}$3-parallel transport

$n^{-1/4}$4

together with the duality

$n^{-1/4}$5

A model is $n^{-1/4}$6-flat with respect to $n^{-1/4}$7 if $n^{-1/4}$8-parallel transport preserves the contour-set tangent spaces, and $n^{-1/4}$9-curvature free if the efficient influence curve is orthogonal to the transported tangent spaces. In this language, a doubly robust estimating function is precisely one that is $M$0-parallel-transport invariant along both contour sets. Convexity implies $M$1-flatness, and under variation independence and $M$2-connectedness, $M$3-flatness ensures that any influence curve is doubly robust (Ying, 2024).

3. Orthogonality, rates, and asymptotic linearity

The orthogonality-based explanation of the $M$4 threshold is now standard in the structural DR literature. In the unbiased Z-estimation setup

$M$5

the note (Lok, 2024) applies a componentwise Mean Value Theorem expansion and obtains a schematic first-order representation

$M$6

with

$M$7

Under orthogonality, $M$8, and the remainder satisfies

$M$9

Hence if $p$0, then $p$1, giving

$p$2

The asymptotic variance is therefore the same as if the nuisance were known, and the sandwich variance estimator that ignores nuisance variability is consistent (Lok, 2024).

A high-dimensional version appears in "A unifying approach for doubly-robust $p$3 regularized estimation of causal contrasts." There the one-step estimator

$p$4

is studied for parameters with bilinear influence function

$p$5

The structural identity is the mixed-bias formula

$p$6

This yields rate double robustness: root-$p$7 consistency and asymptotic normality whenever the nuisance rates satisfy $p$8. It also yields model double robustness: asymptotic normality can persist when only one nuisance working model is approximately sparse, provided the other nuisance’s probability limit obeys the required approximate sparsity conditions (Smucler et al., 2019).

A different extension appears in "Doubly-robust inference and optimality in structure-agnostic models with smoothness." That paper begins from the observation that, in purely structure-agnostic classes with only rate bounds $p$9 and $\psi(P)$0, the AIPW rate $\psi(P)$1 is minimax-optimal. It then introduces hybrid classes that retain agnosticism about $\psi(P)$2 and $\psi(P)$3 themselves but impose Hölder smoothness on certain generated-regressor regressions such as

$\psi(P)$4

and analogues for $\psi(P)$5. The proposed estimator augments the DR/AIPW estimator by a kernel-based $\psi(P)$6-statistic correction in the space of generated regressors $\psi(P)$7. In the Lipschitz case $\psi(P)$8 with $\psi(P)$9, the paper shows asymptotic linearity and valid Wald-type inference as long as either $\phi(X;\theta,\gamma_1,\gamma_2)$0 or $\phi(X;\theta,\gamma_1,\gamma_2)$1, with mild sup-norm consistency of the other nuisance; the achieved rate matches the hybrid minimax lower bound $\phi(X;\theta,\gamma_1,\gamma_2)$2 (Bonvini et al., 2024).

Taken together, these results show that structural double robustness is not only a statement about misspecification protection. It is also a statement about the order at which nuisance estimation enters the expansion of the target estimator. In all three formulations—the geometric, the Z-estimation, and the high-dimensional one-step formulation—the decisive structural fact is that the first-order nuisance term is eliminated, so that second-order, product-rate, or squared-error terms determine the remainder (Ying, 2024, Lok, 2024, Smucler et al., 2019).

4. Canonical model classes and applications

The average treatment effect is the canonical example. In the geometric paper, the ATE is parameterized by nuisance components

$\phi(X;\theta,\gamma_1,\gamma_2)$3

with AIPW estimator

$\phi(X;\theta,\gamma_1,\gamma_2)$4

Its bias is approximately

$\phi(X;\theta,\gamma_1,\gamma_2)$5

and the paper identifies convexity of contour sets and variation independence as the structural reason the local efficient influence curve implies global DR in this parameterization (Ying, 2024). The $\phi(X;\theta,\gamma_1,\gamma_2)$6 note gives the same example in orthogonal-score form,

$\phi(X;\theta,\gamma_1,\gamma_2)$7

and uses it to illustrate why the asymptotic variance is unaffected by nuisance estimation when the score is orthogonal and the nuisances are estimated at $\phi(X;\theta,\gamma_1,\gamma_2)$8 (Lok, 2024).

The partially linear model is the other standard benchmark. The geometric paper studies

$\phi(X;\theta,\gamma_1,\gamma_2)$9

with doubly robust estimating function

$E[\phi(X;\theta(p),\gamma(p))]=0$0

where $E[\phi(X;\theta(p),\gamma(p))]=0$1 and $E[\phi(X;\theta(p),\gamma(p))]=0$2. The $E[\phi(X;\theta(p),\gamma(p))]=0$3 note presents the orthogonal score

$E[\phi(X;\theta(p),\gamma(p))]=0$4

again emphasizing that the nuisance derivative vanishes at the truth (Ying, 2024, Lok, 2024).

Instrumental-variables and LATE problems furnish a distinct realization of structural DR. "Double Robustness for Complier Parameters and a Semiparametric Test for Complier Characteristics" studies Abadie’s class of complier parameters under instrument exogeneity, exclusion, overlap, and monotonicity. The paper introduces $E[\phi(X;\theta(p),\gamma(p))]=0$5-weights, the balancing weight

$E[\phi(X;\theta(p),\gamma(p))]=0$6

and doubly robust scores of the form

$E[\phi(X;\theta(p),\gamma(p))]=0$7

It proves

$E[\phi(X;\theta(p),\gamma(p))]=0$8

establishes Neyman orthogonality, and combines the score with cross-fitting and regularized Riesz estimation in the Auto-$E[\phi(X;\theta(p),\gamma(p))]=0$9 procedure (Singh et al., 2019).

Difference-in-differences gives a structurally different DR construction. In "Double-Robust Estimation in Difference-in-Differences," the ATT counterfactual mean $m$00 is estimated by hybridizing propensity weighting and outcome regression for control trends. The proposed estimator is

$m$01

equivalently

$m$02

Proposition 1 states that $m$03 if either the propensity score is consistent or both control outcome regressions are consistent (Li, 2019).

Macroeconometric local projections provide yet another manifestation. "Double Robustness of Local Projections and Some Unpleasant VARithmetic" shows that the LP regression moment

$m$04

is orthogonal in the same sense as a partially linear model. In the locally misspecified VARMA setup, both nuisance errors are $m$05, so their product is $m$06 whenever $m$07. This yields valid LP confidence intervals even when misspecification is statistically detectable, whereas conventional VAR intervals can severely undercover unless the lag length is large enough that the interval becomes as wide as the LP interval (Olea et al., 2024).

5. Parameterization dependence, impossibility, and generalizations

A recurring theme is that structural double robustness depends on parameterization, not merely on the underlying statistical model. The geometric paper’s semiparametric odds-ratio example makes this explicit. In the canonical parameterization with $m$08 and $m$09, doubly robust estimating functions do not exist; the paper attributes this to failure of variation independence and/or non-convex contour sets. Under the alternative parameterization $m$10, $m$11, influence curves are doubly robust. The structural change is therefore a change in geometry rather than a change in the underlying estimand (Ying, 2024).

The same theme appears in continuous updating GMM. "Double robust inference for continuous updating GMM" studies testing hypotheses about the pseudo-true value

$m$12

of the population continuous-updating objective. The proposed DRLM statistic is built from the CU score and a recentered Jacobian,

$m$13

and has a bounding $m$14 limit under misspecification, weak identification, or both. In this setting, "double robust" means robust to two structural failures—moment misspecification and weak identification—rather than to two nuisance regressions. The paper describes this as validity under either or both failures, without pretests, and the target is the pseudo-true value rather than a correctly specified structural parameter (Kleibergen et al., 2021).

These examples show that structural DR includes both existence theorems and impossibility theorems. When the intersection

$m$15

is empty, no DR estimating function exists under that parameterization (Ying, 2024). A plausible implication is that the practical design problem is often a parameterization-design problem: one seeks nuisance coordinates in which contour sets become variation independent, convex, $m$16-flat, or otherwise orthogonally transportable.

6. Limitations, failure modes, and later safeguards

The structural DR literature is explicit about its assumptions. The geometric theory requires mean-square differentiability, tangent-space calculus, and smooth path-connectedness of contour sets; it also emphasizes that $m$17-flatness and $m$18-curvature freeness are sufficient conditions and "may not be necessary in all cases" (Ying, 2024). The $m$19 analysis requires smoothness, interchange of differentiation and expectation, invertibility of the score Jacobian $m$20, and classical CLT conditions; cross-fitting is not required by the note itself, though it is recommended in the broader literature when highly flexible machine learning is used (Lok, 2024). The high-dimensional and hybrid-smoothness extensions impose approximate sparsity, Hölder smoothness, or generated-regressor smoothness conditions that are nontrivial to verify in practice (Smucler et al., 2019, Bonvini et al., 2024).

A more substantive limitation is that classical DR can become fragile under complete nuisance misspecification. "Rescuing double robustness: safe estimation under complete misspecification" formalizes this as double fragility. In the missing-at-random setup, the paper writes the product remainder as

$m$21

and argues that the same structure that is beneficial under partial misspecification can compound errors under joint misspecification (Testa et al., 26 Sep 2025).

Its proposed response is adaptive correction clipping (ACC), which replaces the DR correction by a clipped version lying in the interval spanned by the outcome-regression and IPW estimators: $m$22 The resulting safety theorem yields

$m$23

with $m$24. Under correctness of at least one nuisance, the paper proves consistency of ACC; when nuisances are well specified, it proposes parametric-bootstrap inference based on the joint Gaussian limit of the OR, IPW, and correction terms (Testa et al., 26 Sep 2025).

This later development clarifies a common misconception. Structural double robustness does not mean universal stability under arbitrary misspecification. Rather, it identifies the conditions under which orthogonality, geometry, or related structural devices eliminate first-order nuisance effects. When those conditions fail globally, the product-bias structure can become a liability. The modern literature therefore treats structural DR both as a source of principled robustness and as a diagnostic framework for determining when robustness is impossible, parameterization-dependent, rate-limited, or in need of additional safeguards (Ying, 2024, Testa et al., 26 Sep 2025).