Efficient GMM and Weighting Matrix under Misspecification

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.

• The paper introduces a novel misspecification-efficient (ME) estimator that minimizes asymptotic variance under model misspecification.
• Enhanced inference methods, like the Bootstrap ME-GMM, offer valid and tighter confidence intervals compared to standard techniques.
• Theoretical insights redefine GMM efficiency frontiers, providing robust estimators for models with inherent misspecifications.

Efficient GMM Estimation under Misspecification: Theory, Implementation, and Implications

Introduction and Context

The generalized method of moments (GMM) is a central estimator in econometrics, operationally robust to overidentification by allowing more moment restrictions than parameters. Under correct model specification, the choice of weighting matrix $W$ in GMM solely affects efficiency, and the optimal (efficient) weighting is well-established as the inverse of the moment covariance. However, empirical models often suffer from moment misspecification—no parameter vector solves the moment conditions exactly. In these settings, conventional GMM theory becomes insufficient: the choice of $W$ not only impacts estimator variance but fundamentally alters the estimand, since each weighting matrix targets a distinct pseudo-true parameter minimizing the criterion $g(\theta)'Wg(\theta)$. This paper, "Efficient GMM and Weighting Matrix under Misspecification" (2605.04961), provides a comprehensive semiparametric efficiency analysis of GMM under misspecification, characterizes the influence of weighting choices, introduces a new misspecification-efficient (ME) estimator, and proposes operational inference methods robust to practical violations of the moment restrictions.

Augmented Moment Approach and New Efficiency Frontier

The key technical observation is that, under misspecification, the GMM estimator's influence function depends on both the original moments and their Jacobian. By augmenting the moment conditions with their Jacobian and recentring appropriately, a new class of estimators emerges. The standard GMM estimator is shown to be a special, generally suboptimal, case of this class.

The main result is the construction of the misspecification-efficient (ME) estimator, which, by optimally weighting the augmented system, achieves the smallest asymptotic variance for the same GMM pseudo-true value as targeted by any fixed $W$. In the linear model, the ME estimator’s asymptotic variance reduces to $(G'W^*G)^{-1}$, where $W^* = \Sigma_{11,2}^{-1}$ is the inverse of the variance of moment residuals after projection on the Jacobian. Critically, for linear models, the ME variance is invariant to $W$; while $W$ determines the estimand, the estimation precision is bounded below by the same uniform efficiency frontier.

The paper argues that, in many policy-relevant settings (e.g., heterogeneous treatment effects in IV models), researchers should report not only misspecification-robust standard errors for their preferred estimator, but also the ME efficiency bound as a transparent measure of attainable precision under existing moment and identification conditions.

Feasible Inference Procedures: Bootstrap and Split-Sample Approaches

A major practical challenge is that the "oracle" ME estimator is not directly feasible, as its computation requires knowledge of population-level moments and Jacobian recentering terms. Three feasible procedures are developed:

1. Bootstrap ME-GMM: A GMM estimator with double recentering—recentering both moments and Jacobians at their sample analogues—using the optimal weighting from the sample covariance structure. The proposed bootstrap correctly mimics the combined sampling variation in both the moments and the Jacobian, and its limiting distribution matches that of the oracle ME estimator.
2. Double-Recentered (DR) Bootstrap: Jointly perturbs the original moments and the Jacobian, delivering valid percentile confidence intervals for the standard GMM estimator under both correct specification and misspecification. This extends and robustifies classical recentered bootstrap approaches (e.g., Hall and Horowitz 1996).
3. Split-Sample ME Estimator: Uses sample splitting to estimate recentering terms on a holdout subsample, constructing the efficient estimator and variance on the other half. In linear models with $W=W^*$, recentering the original moments can be bypassed due to first-order orthogonality conditions.

Extensive simulation and empirical validation show that the DR bootstrap and sample-split methods yield valid inference and tighter confidence intervals compared to analytic misspecification-robust standard errors.

Semiparametric Efficiency: Uniform Minimax Bounds

An important theoretical contribution is the establishment of local asymptotic minimax (semiparametric efficiency) bounds uniform over $W$ in a pre-specified class $W$0. Under misspecification, the pseudo-true value $W$1 varies with $W$2, complicating the classical notion of semiparametric efficiency. This paper clarifies that, for each fixed $W$3, the efficiency bound is $W$4, and provides a uniform bound across $W$5—serving both researchers with strong substantive priors and those seeking robust, design-agnostic inference.

Numerical and Empirical Results

Monte Carlo experiments (normal and weak instrument regimes) demonstrate that:

• The ME estimator consistently achieves lower standard deviation than standard or robust GMM estimators for the same target, with efficiency gains most pronounced when $W$6 is far from efficient weighting or under moderate identification.
• Empirical illustrations in prominent IV applications (Card 1995; Angrist & Krueger 1991) show that the ME efficiency bound can be 10–40% lower than conventional or robust standard errors, especially with "naive" choices like $W$7. The DR bootstrap confidence interval is competitive, often shorter yet valid.

Implications and Future Directions

This work advances both theoretical and applied econometrics by (1) clarifying the dual role of the GMM weighting matrix under misspecification, (2) providing a constructive means to quantify attainable estimator precision, and (3) enabling rigorous robust inference procedures that are efficient in the presence of misspecification.

Key implications include:

• Transparency: Researchers should routinely report the ME efficiency bound alongside estimated effects and robust standard errors to inform conclusions under model approximations.
• Sensitivity to $W$8: Meaningful divergence in results across one-step, two-step, and iterated GMM should be interpreted as evidence of misspecification sensitivity, with $W$9 influencing both the estimand and attainable precision.
• Design of Robust Procedures: The augmented-moment framework and DR bootstrap are immediately adaptable to a wide class of models, including those featuring nonlinearities or many instruments, and guide principled sensitivity analysis.

Future extensions should address ME estimators under weak identification, adaptive selection of $g(\theta)'Wg(\theta)$0, and connections with other minimum distance and empirical likelihood frameworks under global misspecification.

Conclusion

The paper rigorously redefines the efficiency frontier for GMM estimation in the presence of moment misspecification. By leveraging augmented moment conditions and a constructive semiparametric efficiency analysis, it delivers both theoretical insight and practical tools—most notably, the misspecification-efficient estimator and associated bootstrap procedures—that enable economists to conduct valid, optimally precise inference even when conventional moment-based models are misspecified.