Plausible GMM - Overview

Summary

Chernozhukov, Hansen, Kong & Wang (2026) extend the quasi-Bayesian posterior (QBP) framework of GMM to settings where moment conditions are not assumed to hold exactly. Instead of the dogmatic restriction ${\mu }_{\ast }\equiv 0$, a researcher places a proper prior $\pi (\mu )$ over the degree of misspecification ${\mu }_{\ast }$ (the “plausibility characteristic”), turning misspecification into a partial-identification problem driven by parameter over-parameterization. The resulting quasi-posterior stays informative under both global and local misspecification, and — under local Gaussian priors — is approximately Gaussian centered on a GMM estimator whose weighting matrix endogenously trades off moment precision against plausibility. The central lesson is “no free lunch”: relaxing the exact-moment assumption necessarily yields wider, less precise inference.

Overview

Structural estimation in economics is built on moment conditions — IV exclusion restrictions, unconfoundedness, parallel trends, Euler equations — argued from institutional knowledge and economic reasoning. Classical IV estimation proceeds as if these restrictions hold exactly. But it is rarely possible to be confident they do, and estimates built on exactly-true moments can be substantially distorted when the moments are even slightly wrong.

This paper formalizes the idea that moment restrictions are plausible but not certain. A researcher encodes subjective beliefs about possible violations via a proper prior over a misspecification vector ${\mu }_{\ast }$, then performs quasi-Bayesian inference on the structural parameter ${\theta }_{\ast }$. See Plausible Moment Restriction Model for the formal setup and Quasi-Bayes for Plausible Moment Restrictions for the inferential machinery.

Main Content

Research question

How can a researcher perform credible inference on a structural parameter ${\theta }_{\ast }$ defined by moment conditions, while explicitly allowing the moment conditions to be only approximately valid, and while encoding subjective economic beliefs about the size of the violation?

Key contributions

1. Plausible GMM (PGMM) framework — extends quasi-Bayesian posteriors (QBPs) to a non-dogmatic prior over ${\mu }_{\ast }$. The structural model becomes $m(\theta )=\mu$, with $\mu$ governed by prior $\pi (\mu )$. Because ${\theta }_{\ast }$ and ${\mu }_{\ast }$ are not jointly identified, the prior has a non-negligible effect even asymptotically — this is a genuine partial-identification regime. See Plausible Moment Restriction Model.

2. Quasi-posterior concentration / new Bernstein–von Mises results — the quasi-posterior converges to a mixture of Gaussians whose weights and components depend heavily on $\pi (\mu )$. With a dogmatic prior $\mu \equiv {\mu }_{\ast }$ this reduces to the classical Chernozhukov–Hong (2003) result (collapse to a Gaussian centered at $\theta ({\mu }_{\ast })$ with efficient-GMM variance). (Formal statements live in §4 / Supplemental Appendix; see Gaussian Local Prior Approximation for the tractable special case.)

3. Approximately optimal Bayesian decisions — quasi-Bayes decision rules approximate Bayes-optimal rules under the maintained prior over misspecification (extending Andrews & Mikusheva 2022 to the non-dogmatic case).

4. Frequentist coverage — Bayesian credible regions from the quasi-posterior have correct frequentist coverage under a two-stage sampling thought experiment (nature first draws ${\mu }_{\ast }\sim \pi (\mu )$, then data is generated with $m(\theta ({\mu }_{\ast }))={\mu }_{\ast }$); equivalently, an ex ante coverage guarantee. A union-of-confidence-intervals construction ${\cup }_{{\mu }_{\ast }\in C}CI({\mu }_{\ast },\alpha )$ delivers uniform coverage when ${\mu }_0$ is a fixed unknown in a known set $C$.

5. Endogenous robust weighting — under Gaussian priors with variance $\propto 1/T$, QBPs are centered on a GMM estimator whose weighting matrix trades off moment precision against misspecification, emerging from the quasi-posterior itself rather than being imposed via a minimax criterion. This connects to Armstrong & Kolesár (2021): in the Gaussian limit experiment the Bayes credible interval under a two-point least-favorable prior coincides with their robust confidence interval. See Gaussian Local Prior Approximation.

Relation to the literature

Strand	Relation
Quasi-Bayes / LTE (Chernozhukov–Hong 2003, Kim 2002, Gallant 2016, Florens–Simoni 2021, Andrews–Mikusheva 2022)	PGMM extends QBPs from the dogmatic ${\mu }_{\ast }\equiv 0$ case to a non-dogmatic prior over ${\mu }_{\ast }$.
Local-misspecification / sensitivity (Armstrong & Kolesár 2021; Bonhomme & Weidner 2022; Andrews, Roth & Pakes 2024)	Armstrong–Kolesár fix $m({\theta }_{\ast })=C_T=c/\sqrt{T}$ and do minimax-robust inference. PGMM obtains an analogous trade-off from a quasi-Bayesian perspective via a prior on ${\mu }_{\ast }$.
Bayesian partial identification (Chib, Shin & Simoni 2018; Gustafson 2015; Andrews, Marmer & Yu 2024)	PGMM allows all elements of ${\mu }_{\ast }$ to be free (no pseudo-true value), so posteriors need not concentrate on a unique point.
Frequentist-coverage-for-identified-sets (Chen, Christensen & Tamer 2018; Conley, Hansen & Rossi 2012)	PGMM imposes a prior on ${\mu }_{\ast }$ rather than profiling it out, which changes the asymptotics.

Examples

See Plausible GMM - Institutions and GDP Application — a revisit of Acemoglu, Johnson & Robinson (2001) on institutions and GDP using linear IV, demonstrating that inference on the institutions coefficient ${\beta }_X$ is relatively robust to the prior over misspecification while honestly reflecting added uncertainty. (A second application to IV quantile regression of 401(k) participation, revisiting Chernozhukov & Hansen 2004, is in the Supplemental Appendix — see Gaps.)

Connections

• Generalizes the dogmatic GMM/QBP setup (moment conditions exact, ${\mu }_{\ast }\equiv 0$) to a plausible one (proper prior over violations).
• Sits between frequentist sensitivity analysis (Sensitivity Analysis in Observational Studies) and Bayesian partial identification — it is subjective-Bayesian in motivation but delivers a notion of frequentist coverage.
• The criterion function is the continuous-updating GMM objective; the quasi-posterior is the MCMC-based Laplace-type estimator of Chernozhukov–Hong.

See Also

• Plausible Moment Restriction Model — the formal setup and the IV running example
• Quasi-Bayes for Plausible Moment Restrictions — criterion function, quasi-posterior, decisions
• Gaussian Local Prior Approximation — the tractable Gaussian-prior special case (Eq. 5)
• Plausible GMM - Institutions and GDP Application — empirical illustration
• Asymptotics and Frequentist Connections — Bernstein–von Mises background

Gaps

The downloaded arXiv PDF (v2, 14 pp.) contains only the main body through the start of §3.1. Not captured from primary source (described only via the introduction): the formal §4 results (the Bernstein–von Mises concentration theorems, the precise decision-theoretic and coverage statements, the increasing-$(k,q)$ asymptotics), and the second empirical application (Chernozhukov–Hansen 2004 IV quantile regression of 401(k) participation on assets). These live in the Supplemental Appendix (“SA”) and Online Materials (“OM”). Ingest those to complete the theorem-level coverage.