arxiv: 2605.04961 · v2 · submitted 2026-05-06 · 💰 econ.EM · stat.ME

Recognition: unknown

Efficient GMM and Weighting Matrix under Misspecification

Byunghoon Kang

Pith reviewed 2026-05-08 16:15 UTC · model grok-4.3

classification 💰 econ.EM stat.ME

keywords GMM estimationmoment misspecificationefficient estimationweighting matrixasymptotic variancerecentered momentspseudo-true valuebootstrap

0 comments

The pith

Augmenting moment conditions with recentering and optimal weighting produces a GMM estimator with minimal asymptotic variance under misspecification.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper establishes that standard GMM is inefficient when moments are misspecified because its influence function incorporates both the moments and their Jacobian. It constructs an augmented set of recentered moments that neutralizes this dependence while preserving the same limiting pseudo-true value. Optimally weighting the larger system then yields an estimator with strictly smaller asymptotic variance than the usual GMM procedure. The result matters for applied work because economic models are often misspecified in practice, so efficiency gains improve precision without shifting the economic target. In linear cases the variance formula collapses to the textbook efficient-GMM expression after projecting moments orthogonal to the Jacobian.

Core claim

The standard GMM estimator is a special case within a broader class based on augmented moment conditions with recentering. By optimally weighting the augmented system, we obtain a misspecification-efficient (ME) estimator with the smallest asymptotic variance for the same GMM pseudo-true value. In linear models, the asymptotic variance of the ME estimator reduces to the textbook efficient-GMM variance formula (G'W*G)^{-1}, where W* is the inverse of the variance of residualized moments after projection on the Jacobian G. The paper also develops a feasible double-recentered bootstrap estimator and a split-sample ME estimator, and establishes uniform local asymptotic minimax bounds over aclass

What carries the argument

Augmented moment conditions with recentering, which incorporate the Jacobian to adjust the influence function for misspecification while keeping the pseudo-true value fixed.

If this is right

The ME estimator achieves the smallest asymptotic variance among all GMM estimators sharing the same pseudo-true value.
In linear models its asymptotic variance equals (G'W*G)^{-1} with W* defined on residualized moments.
The double-recentered bootstrap provides a misspecification-robust and efficient inference procedure.
Split-sample implementation of the ME estimator is feasible when full-sample computation is costly.
Uniform local asymptotic minimax optimality holds over classes of weighting matrices.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Applied researchers could obtain tighter confidence intervals in misspecified models such as IV or structural equations by adopting the ME estimator without altering the target parameter.
The recentering construction may extend naturally to other moment-based procedures outside classical GMM, including minimum-distance and generalized empirical likelihood estimators.
Finite-sample simulations under varied misspecification patterns could quantify whether the asymptotic efficiency gains appear in realistic sample sizes.

Load-bearing premise

The influence function of the standard GMM estimator under misspecification depends on both the original moment conditions and their Jacobian in a way that recentering can neutralize without changing the pseudo-true value.

What would settle it

A Monte Carlo experiment or analytic calculation showing that the asymptotic variance of the proposed ME estimator exceeds the variance of standard GMM for the same limiting value would falsify the efficiency claim.

read the original abstract

This paper develops efficient GMM estimation when the moment conditions are misspecified. We observe that the influence function of the standard GMM estimator under misspecification depends on both the original moment conditions and their Jacobian, motivating a new class of estimators based on augmented moment conditions with recentering. The standard GMM estimator is a special case within this class, and generally suboptimal. By optimally weighting the augmented system, we obtain a misspecification-efficient (ME) estimator with the smallest asymptotic variance for the same GMM pseudo-true value. In linear models, the asymptotic variance of ME estimator reduces to the textbook efficient-GMM variance formula $(G'W^{*}G)^{-1}$, where $W^{*}$ is the inverse of the variance of residualized moments after projection on the Jacobian $G$. We consider a feasible double-recentered bootstrap estimator, which can be considered as a misspecification-robust and efficient version of Hall and Horowitz (1996) recentered bootstrap GMM estimator, and also consider a split-sample ME estimator. Finally, we establish uniform local asymptotic minimax bounds over a class of weighting matrices. We illustrate the proposed methods in simulation and empirical examples.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a concrete way to shrink the asymptotic variance of GMM under misspecification by augmenting the moments and reweighting them, while keeping the same pseudo-true value.

read the letter

The central contribution is an augmented-moment construction that recenters the system so the population moments stay zero at the original GMM pseudo-true point. Optimal weighting of that system then produces an estimator whose asymptotic variance is smaller than or equal to the usual GMM variance. In the linear case the formula collapses exactly to the textbook efficient-GMM expression after projecting the moments onto the Jacobian, which is a useful sanity check rather than an accident of the setup.

Referee Report

2 major / 3 minor

Summary. The paper develops efficient GMM estimation under misspecification by augmenting the original moment conditions with recentering terms derived from the influence function of the standard GMM estimator. The resulting misspecification-efficient (ME) estimator optimally weights the augmented system to achieve the smallest asymptotic variance among estimators that share the same GMM pseudo-true value; standard GMM is recovered as a special case but is generally suboptimal. In linear models the ME asymptotic variance reduces to the textbook efficient-GMM formula (G'W*G)^{-1} after residualizing moments on the Jacobian G. The paper also proposes a feasible double-recentered bootstrap, a split-sample ME estimator, establishes uniform local asymptotic minimax bounds over a class of weighting matrices, and illustrates the methods with simulations and empirical examples.

Significance. If the central derivation holds, the ME estimator supplies a practical efficiency gain under the common situation of misspecified moments while leaving the target pseudo-true value unchanged. The explicit reduction to the known linear-model formula supplies a useful benchmark, and the uniform local asymptotic minimax result strengthens the optimality claim within the considered class. The double-recentered bootstrap extends Hall-Horowitz recentering in a misspecification-robust direction. These features address a gap between theoretical GMM efficiency results (which assume correct specification) and empirical practice.

major comments (2)

[§3.1, Eq. (12)] §3.1, Eq. (12): the claim that the augmented moments remain zero at the original pseudo-true value for any weighting matrix in the class is load-bearing for preserving the target parameter; the manuscript should display the explicit population identity that verifies this property before proceeding to the optimal-weighting derivation.
[§4.2] §4.2, the uniform local asymptotic minimax result: the precise class of weighting matrices over which the bound is taken is not stated explicitly; without this the scope of the optimality statement relative to all feasible GMM estimators remains unclear.

minor comments (3)

[§3.3] The notation W* for the optimal weighting matrix in the linear-model reduction should be distinguished more clearly from the conventional efficient-GMM weighting matrix to avoid reader confusion.
[Section 5] Simulation design (Section 5): the data-generating processes and misspecification magnitudes should be tabulated so that the reported efficiency gains can be reproduced exactly.
[§4.1] The feasible double-recentered bootstrap algorithm would benefit from a step-by-step pseudocode listing, especially the split-sample version.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our manuscript. We address each major comment below and will revise the paper to incorporate the suggested clarifications.

read point-by-point responses

Referee: [§3.1, Eq. (12)] §3.1, Eq. (12): the claim that the augmented moments remain zero at the original pseudo-true value for any weighting matrix in the class is load-bearing for preserving the target parameter; the manuscript should display the explicit population identity that verifies this property before proceeding to the optimal-weighting derivation.

Authors: We agree that displaying the explicit population identity will strengthen the exposition and make the load-bearing property transparent. In the revised manuscript, we will insert a new display equation immediately after the definition of the augmented moments in §3.1, showing that E[m_aug(θ*)] = 0 holds for any positive definite weighting matrix W. The identity follows because the recentering term is constructed from the influence function of the original GMM estimator evaluated at θ*, which by definition cancels the nonzero expectation of the original moments at the pseudo-true value independently of W. We will present the short algebraic verification before deriving the optimal weighting matrix. revision: yes
Referee: [§4.2] §4.2, the uniform local asymptotic minimax result: the precise class of weighting matrices over which the bound is taken is not stated explicitly; without this the scope of the optimality statement relative to all feasible GMM estimators remains unclear.

Authors: We thank the referee for highlighting this point. The minimax result in §4.2 is taken over the class of all weighting matrices W that are positive definite and converge in probability to a positive definite limit (i.e., the standard class of consistent weighting matrices for GMM). This class is implicitly the one for which the ME estimator remains consistent for the same pseudo-true value. In the revision we will state the class explicitly at the opening of §4.2, noting that the uniform local asymptotic minimax optimality holds within this class but does not claim superiority over GMM estimators that target a different pseudo-true value. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper derives the misspecification-efficient estimator from the standard influence function of GMM under misspecification (a known asymptotic result), constructs an augmented recentered moment system whose population value is zero at the original pseudo-true parameter by explicit design for any weighting matrix in the class, and obtains the ME estimator via the inverse of the asymptotic variance of those recentered moments. This is a conventional efficiency argument within a fixed class of estimators sharing the same pseudo-true value. The linear-model reduction to the textbook formula (G'W*G)^{-1} serves as an external consistency check rather than a definitional loop. The uniform local asymptotic minimax bound further grounds optimality independently. No self-definitional steps, fitted inputs renamed as predictions, load-bearing self-citations, or ansatz smuggling appear in the provided derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract alone supplies insufficient detail to enumerate specific free parameters or invented entities; the construction relies on standard GMM regularity conditions and the stated influence-function observation.

axioms (1)

standard math Standard regularity conditions for GMM asymptotic expansions hold under misspecification
Invoked to derive influence functions and asymptotic variances.

pith-pipeline@v0.9.0 · 5494 in / 1151 out tokens · 28809 ms · 2026-05-08T16:15:40.010362+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

7 extracted references

[1]

Abadir, K.M., and Magnus, J.R. (2005). Matrix Algebra.Cambridge University Press. Acemoglu, D., Johnson, S., Robinson, J. A., and Yared, P. (2008). Income and Democracy. American Economic Review, 98(3), 808-842. Andrews, D.W.K., and Cheng, X. (2012) Estimation and inference with weak, semi-strong and strong identification.Econometrica, 80, 2153-2211. Andr...

2005
[2]

Vuong, Q.H. (1989). Likelihood ratio tests for model selection and non-nested hypotheses,Econo- metrica, 57(2), 307-333. White, H. (1982). Maximum likelihood estimation of misspecified models.Econometrica, 50(1), 1-25. Windmeijer, F. (2005). A finite sample correction for the variance of linear efficient two-step GMM estimators,Journal of Econometrics, 12...

1989
[3]

Since the model is linear,g(X i, θ) =g(X i, θ0) +G(X i)δ, we have ψ(Xi, θ) =ψ i(θ) = g(Xi, θ0) +G(X i)δ vec(G(Xi)′) ! = Im C 0I mp ! ψi(θ0) whereC= (δ ′ ⊗I m)K p,m andK p,m is themp×mpcommutation matrix satisfiesvec(G(X i)) = Kp,m vec(G(Xi)′) (Abadir and Magnus (2005)).L(θ) = Im C 0I mp is block lower-triangular (with identity diagonal blocks) and depends...

2005
[4]

EF0 ∂eψ(X, θfW ) ∂θ ′ #′h EF0 eψ(X, θfW )eψ′(X, θfW ) i−1

Furthermore, we have ∆∗ = Σn(bθW ) p∗ →Σ(θ W ). So that, (Γ∗ n(bθ∗ M E)′∆∗Γ∗ n(˜θ∗))−1Γ∗ n(bθ∗ M E)′∆∗ p∗ →(Γ(θ W )′Σ(θW )−1Γ(θW ))−1Γ(θW )′Σ(θW )−1 Next, we consider √neψ∗ n(bθW ) in (30), √neψ∗ n(bθW ) = √n g∗ n(bθW )−g n(bθW ) vec(G∗ n(bθW )′)−vec(G n(bθW )′)′ ! .(31) Note that each of the two terms has conditional mean zero given{X i}n i=1, sinceE ∗[g...

1987
[5]

This reduces not only computational costs, but also the conservativeness of the CI considered here. For anyγ∈ ΓW , we can define the ME-estimator defined in (18) as follows; bθM E(W;γ) = arg min θ ψn(θ)−γ ′ bΣ(bθW )−1 ψn(θ)−γ (35) whereψ n(θ) = 1 n Pn i=1 ψ(Xi, θ), ψ(Xi, θ) = (g(X i, θ)′, vec(G(Xi, θ))′)′, and bΣ(bθW ) is a consistent estimator of Σ(θ W )...

2012
[6]

SD” is the MC standard deviation of the point estimate across MC draws; for HH and DR bootstrap, “SD

45 Table 6: Monte Carlo results under weak instrument setup (W=I) δ= 0δ= 0.5δ= 1δ= 2 Estimator SD Cov Len SD Cov Len SD Cov Len SD Cov Len n= 200 GMM + Conv. SE 1.00 0.946 0.847 1.51 0.898 1.055 2.92 0.859 2.036 6.28 0.844 4.173 GMM + Robust SE 1.00 0.950 0.873 1.51 0.949 1.275 2.92 0.917 2.485 6.28 0.918 5.135 Oracle ME 0.89 0.920 0.781 0.93 0.936 0.846 ...

1996
[7]

historically strong institutions

γ(Inc t−1) GMM point−0.346−0.303−0.151 Conv. SE (0.200) (0.110) (0.071) Misspec. SE (0.417) (0.121) (0.056) ME efficiency bound (0.081) (0.081) (0.081) ME-GMM Boot SD (0.092) (0.090) (0.085) DR 95% CI [-1.43,0.39] [-0.62,-0.07] [-0.32,0.03] β(Inc t−1 ×c i) GMM point 0.367 0.318 0.299 Conv. SE (0.168) (0.122) (0.061) Misspec. SE (0.176) (0.130) (0.055) ME ...

2014