arxiv: 2604.15563 · v1 · submitted 2026-04-16 · 💰 econ.EM

Recognition: unknown

True and Pseudo-True Parameters

Harvey Barnhard, Isaiah Andrews, Jacob Carlson

Pith reviewed 2026-05-10 09:15 UTC · model grok-4.3

classification 💰 econ.EM

keywords misspecified modelspseudo-true parametersBayesian decision makingposterior concentrationaverage coverageconfidence intervalsminimum distance estimation

0 comments

The pith

Bayesian posteriors concentrate on pseudo-true parameters only for specific prior sequences in misspecified models, but simple confidence intervals achieve correct average coverage for the true parameter across all such priors without any (

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

The paper examines how Bayesian decision-makers behave when their models are misspecified so that estimates converge to pseudo-true values instead of the true parameters of economic interest. Within a class of priors motivated by the minimum distance objective in linear population problems, it characterizes the prior sequences that cause posteriors to concentrate on the pseudo-true value. This concentration turns out to be fragile to small prior perturbations, meaning pseudo-true values guide decisions only in limited circumstances. Constructively, the authors derive straightforward confidence intervals that deliver correct average coverage for the actual true parameter for every prior in the class, with no restriction on how large the misspecification can be.

Core claim

Parameter estimates in misspecified models converge to pseudo-true values that minimize a population objective function. For Bayesian decision-makers facing a linear population minimum distance problem, posteriors concentrate on the pseudo-true value only under certain sequences of priors within the class motivated by that objective. This convergence is fragile to small changes in the priors. Simple confidence intervals nevertheless exist that guarantee correct average coverage for the true parameter under every prior in the studied class, with no bound on the magnitude of misspecification.

What carries the argument

The class of priors motivated by the minimum distance objective within a linear population minimum distance problem, used both to characterize posterior concentration on pseudo-true values and to construct average-coverage confidence intervals for the true parameter.

Load-bearing premise

The priors must belong to the specific class motivated by the minimum distance objective in a linear population minimum distance problem; outside that class the concentration results and coverage guarantees need not hold.

What would settle it

An example or simulation, using a prior inside the studied class, in which the proposed confidence intervals fail to achieve correct average coverage for the true parameter or in which posterior concentration on the pseudo-true value occurs for a prior sequence outside the characterized special cases.

Figures

Figures reproduced from arXiv: 2604.15563 by Harvey Barnhard, Isaiah Andrews, Jacob Carlson.

**Figure 1.** Figure 1: Relationship between the pseudo-true value and true θ in an example with scalar θ. The solid diagonal line represents the column space of X(P). The vector η ⊥ perpendicular to the column space captures the detectable component of the misspecification vector, while vector ηˆ parallel to the column space captures the undetectable component of misspecification. We next compare the behavior of the confidence s… view at source ↗

**Figure 2.** Figure 2: Intervals for θ using the norm-bounding approach and the rotation invariant prior approach. When dataset A or B is observed, the identified set or confidence interval for θ is given by [l A, uA] or [l B, uB], respectively, where in the left panel l A = u A = θ A. 5.2 Confidence Intervals in Finite Samples So far, our results have abstracted from sampling uncertainty, considering the setting where P is obse… view at source ↗

read the original abstract

Parameter estimates in misspecified models converge to pseudo-true parameter values, which minimize a population objective function. Pseudo-true values often differ from quantities of economic interest, raising questions of how, if at all, they are relevant for decision-making. To study this question we consider Bayesian decision-makers facing a linear population minimum distance problem. Within a class of priors motivated by the minimum distance objective, we characterize prior sequences under which posteriors concentrate on the pseudo-true value. This convergence is fragile to small changes in priors, implying that pseudo-true values are relevant for decision-making only in special cases. Constructive results are nevertheless possible in this setting, and we derive simple confidence intervals that guarantee correct average coverage for the true parameter under every prior in the class we study, with no bound on the magnitude of misspecification.

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

This paper shows posterior concentration on pseudo-true parameters is fragile to prior choice in a linear min-distance setup, but constructs simple intervals with average coverage for the true parameter across their prior class and without bounding misspecification.

read the letter

The central result here is that, in this linear population minimum-distance problem, you can pick prior sequences so the posterior concentrates on the pseudo-true value, but the concentration breaks with small prior tweaks. That part is cleanly characterized and explains why pseudo-true values are not automatically useful for decisions. The constructive payoff is a set of simple confidence intervals that deliver correct average coverage for the true parameter under every prior in the studied class, with no bound needed on how large the misspecification is. That average-coverage guarantee is the practical piece and avoids the usual worst-case bounds that often make robust methods conservative or empty. The paper does a good job keeping the claims scoped to the linear case and the priors motivated by the minimum-distance objective, so there is no hidden overreach in the abstract. The main limitation is exactly that scope: the coverage result does not extend to nonlinear models or to priors outside this motivated class, and average coverage is weaker than pointwise or uniform coverage. If your work stays inside linear setups and you are willing to average over this prior family, the intervals look usable; otherwise the applicability narrows quickly. The derivations are not visible from the abstract alone, but the logic described is internally consistent and the stress-test found no circularity or unstated dependence on bounded misspecification. This is for econometricians working on Bayesian methods or robust inference under misspecification. A reader who already thinks about minimum-distance problems or prior fragility will get the most out of it. I would send it to peer review; the ideas are focused enough that referees can verify the characterizations and check whether the linear restriction is a useful first step.

Referee Report

2 major / 2 minor

Summary. The paper examines pseudo-true parameters in misspecified models for Bayesian decision-makers in a linear population minimum distance problem. Within a class of priors motivated by the minimum distance objective, it characterizes prior sequences under which posteriors concentrate on the pseudo-true value and shows this convergence is fragile to small prior perturbations. It then derives simple confidence intervals that guarantee correct average coverage for the true parameter under every prior in the class, without any bound on the magnitude of misspecification.

Significance. If the characterizations and coverage results hold, the paper clarifies the limited decision-relevance of pseudo-true parameters except in special cases while providing a practical, robust inference tool. The average-coverage guarantee without misspecification bounds is a notable strength within the linear framework and could aid applied work where misspecification is routine. The explicit scoping to the minimum-distance-motivated prior class keeps the claims precise and falsifiable.

major comments (2)

Abstract and the section deriving the confidence intervals: the central claim that the intervals 'guarantee correct average coverage for the true parameter under every prior in the class we study, with no bound on the magnitude of misspecification' requires an explicit construction of the intervals together with the proof that the coverage property follows from the linear minimum-distance structure and the definition of the prior class.
The section characterizing prior sequences: the fragility result (convergence only for special sequences and sensitivity to small changes) should state the precise conditions on the prior class that produce concentration on the pseudo-true value, so readers can verify the boundary between special and generic cases.

minor comments (2)

Abstract: define 'average coverage' explicitly (e.g., whether the average is taken with respect to the prior, the data-generating process, or both) to avoid ambiguity in the coverage guarantee.
Notation and terminology: maintain consistent distinction between the 'true parameter' and the 'pseudo-true parameter' in all statements about posterior concentration and interval coverage.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our paper. We are pleased that the referee views the characterizations and coverage results as significant and recommends minor revision. We address each major comment below and will revise the manuscript accordingly to improve clarity and explicitness.

read point-by-point responses

Referee: Abstract and the section deriving the confidence intervals: the central claim that the intervals 'guarantee correct average coverage for the true parameter under every prior in the class we study, with no bound on the magnitude of misspecification' requires an explicit construction of the intervals together with the proof that the coverage property follows from the linear minimum-distance structure and the definition of the prior class.

Authors: We agree that greater explicitness will strengthen the presentation. In the revised manuscript we will add a dedicated subsection that constructs the confidence intervals explicitly from the linear minimum-distance estimator and the definition of the prior class. We will also supply a self-contained proof that derives the average-coverage guarantee directly from the linearity of the population objective and the structure of the prior class, without any restriction on the size of misspecification. These additions will be placed in the section on confidence intervals and will make the central claim fully verifiable from the linear minimum-distance structure alone. revision: yes
Referee: The section characterizing prior sequences: the fragility result (convergence only for special sequences and sensitivity to small changes) should state the precise conditions on the prior class that produce concentration on the pseudo-true value, so readers can verify the boundary between special and generic cases.

Authors: We accept the suggestion for added precision. While the manuscript already characterizes the relevant prior sequences within the minimum-distance-motivated class, we will revise the section to state the exact conditions on the prior class (including the required sequence properties and support restrictions) that are necessary and sufficient for posterior concentration at the pseudo-true value. The revision will also delineate the boundary separating these special sequences from generic perturbations, allowing readers to verify the fragility result directly from the stated conditions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained

full rationale

The paper's central results characterize prior sequences leading to posterior concentration on pseudo-true values and construct confidence intervals guaranteeing average coverage for the true parameter, all scoped explicitly to a class of priors motivated by the linear population minimum-distance objective. These steps rely on direct analysis of the stated objective function and prior class without reducing any prediction or coverage guarantee to a fitted parameter by the paper's own equations, without load-bearing self-citations, and without importing uniqueness or ansatzes from prior author work. The abstract and described claims remain independent of the target results and do not exhibit self-definitional or renaming patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The paper relies on standard Bayesian decision theory and econometric minimum distance setups without introducing new free parameters or invented entities; the prior class is motivated by the objective but not fitted to data.

axioms (2)

domain assumption Bayesian updating occurs with priors from the class motivated by the minimum distance objective
Invoked to characterize posterior concentration on pseudo-true values.
domain assumption The population problem is linear minimum distance
Defines the setting in which the coverage result is derived.

pith-pipeline@v0.9.0 · 5425 in / 1325 out tokens · 47094 ms · 2026-05-10T09:15:35.325255+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

7 extracted references

[1]

On the properties of variational approximations of Gibbs posteriors,

Alquier, P., J. Ridgw ay, and N. Chopin(2016): “On the properties of variational approximations of Gibbs posteriors,”Journal of Machine Learning Research, 17, 1–

2016
[2]

Misspecified Moment Inequality Models: Inference and Diagnostics,

Andrews, D. W. K. and S. Kwon(2023): “Misspecified Moment Inequality Models: Inference and Diagnostics,”The Review of Economic Studies,

2023
[3]

The purpose of an estimator is what it does: Misspecification, estimands, and over-identification,

Andrews, I., J. Chen, and O. Tecchio(2025): “The purpose of an estimator is what it does: Misspecification, estimands, and over-identification,” . Andrews, I. and J. M. Shapiro(2021): “A Model of Scientific Communication,” Econometrica, 89, 2117–2142. Angrist, J. D. and G. W. Imbens(1995): “Two-Stage Least Squares Estimation of Average Causal Effects in M...

2025
[4]

A general framework for updating belief distributions,

Bissiri, P. G., C. C. Holmes, and S. G. W alker(2016): “A general framework for updating belief distributions,”Journal of the Royal Statistical Society. Series B (Statistical Methodology), 78, 1103–1130. Carlin, B. P. and T. A. Louis(2000):Bayes and empirical Bayes methods for data analysis, Texts in statistical science series, Boca Raton, Fla.: Chapman &...

2016
[5]

Pac-Bayesian Supervised Classification: The Thermodynamics of Statistical Learning,

ed ed. Catoni, O.(2007): “Pac-Bayesian Supervised Classification: The Thermodynamics of Statistical Learning,”Lecture OPTnotes-Monograph Series, 56, i–163. Chernozhukov, V. and H. Hong(2003): “An MCMC approach to classical estima- tion,”Journal of Econometrics, 115, 293–346. Conley, T. G., C. B. Hansen, and P. E. Rossi(2012): “Plausibly Exogenous,” The Re...

2007
[6]

Define Ic(P) = Z πθ(ϑ)πη,c(g(ϑ;P)|ϑ)dϑ and I ∗(P) = Z πθ(ϑ)π∗ η(g(ϑ;P)|ϑ)dϑ

32 The result is then immediate from the argmax continuous mapping theorem (Theorem 3.2.2 of van der Vaart and Wellner 1996).□ Proof of Proposition 4FixPand write J=J W (P), Q(θ) =Q W (θ;P), g θ =g(θ;P). Define Ic(P) = Z πθ(ϑ)πη,c(g(ϑ;P)|ϑ)dϑ and I ∗(P) = Z πθ(ϑ)π∗ η(g(ϑ;P)|ϑ)dϑ. Also let πc(θ|P) = πθ(θ)πη,c(gθ |θ) Ic(P) and π∗(θ|P) = πθ(θ)π∗ η(gθ |θ) I ∗...

1996
[7]

Also let π∗(θ|P) = πθ(θ)π∗ η (g(θ;P)|θ) R πθ(˜θ)π∗η g(˜θ;P)| ˜θ d˜θ

Define Ic(P) = Z πη,c (g(θ;P)|θ)π θ(θ)dθ and I ∗(P) = Z π∗ η (g(θ;P)|θ)π θ(θ)dθ. Also let π∗(θ|P) = πθ(θ)π∗ η (g(θ;P)|θ) R πθ(˜θ)π∗η g(˜θ;P)| ˜θ d˜θ . 34 Then we can write the posterior as πϕ c (θ|P) =w 1,cπc(θ|P) + (1−w 1,c)π∗(θ|P), wherew 1,c = (1−ϕ)Ic(P) (1−ϕ)Ic(P)+ϕI ∗(P) .By Proposition 2, the posteriorπ c(θ|P)concentrates onθ W (P)asc→0.Thus it suff...

1982