Graph Neural Networks for Generalized Mundlak Estimator under Network Confounding

Lianyan Fu; Rui Wang; Zihan Zhang

arxiv: 2605.29238 · v1 · pith:JK2CX4H6new · submitted 2026-05-28 · 💰 econ.EM

Graph Neural Networks for Generalized Mundlak Estimator under Network Confounding

Lianyan Fu , Rui Wang , Zihan Zhang This is my paper

Pith reviewed 2026-06-29 00:22 UTC · model grok-4.3

classification 💰 econ.EM

keywords graph neural networksMundlak estimatornetwork confoundingdouble robustnessasymptotic normalitygroup heterogeneityfixed effectseconometrics

0 comments

The pith

Graph neural network Mundlak estimator controls group confounding with balancing statistics and message passing.

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

The paper proposes a generalized Mundlak estimator that replaces group-specific intercepts with aggregated group-level balancing statistics. This substitution controls between-group confounding, permits cross-group comparisons, and drops the linear additivity requirement of conventional fixed-effects models. Graph neural networks then use message passing to learn nonlinear representations and to account for within-group dependence among individuals. Theoretical analysis establishes that the resulting estimator is doubly robust and asymptotically normal. Simulation and empirical checks confirm these properties hold under network confounding.

Core claim

The GME-GNN estimator uses aggregated group-level balancing statistics to fully control between-group confounding, enabling valid cross-group comparisons, and employs GNN message-passing to adaptively learn nonlinear representations and capture intra-group interaction effects, resulting in double robustness and asymptotic normality.

What carries the argument

Aggregated group-level balancing statistics combined with graph neural network message-passing, which replaces intercepts to control confounding and model interactions without linearity or independence assumptions.

If this is right

Valid comparisons across groups become feasible instead of only within-group contrasts.
Linearity and intra-group independence assumptions can be relaxed.
Nonlinear representations and interaction effects are learned adaptively within groups.
The estimator remains consistent under partial misspecification due to double robustness.
Asymptotic normality supports standard inference procedures.

Where Pith is reading between the lines

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

The same balancing-plus-message-passing structure might apply to other panel or clustered data settings with unobserved group effects.
Empirical work could test whether different GNN architectures improve capture of specific network dependence patterns.

Load-bearing premise

Aggregated group-level balancing statistics fully control between-group confounding and GNN message-passing captures intra-group interactions without introducing new bias.

What would settle it

A controlled simulation in which group-level balancing statistics are incomplete or GNN layers introduce bias, causing the estimator to lose double robustness or asymptotic normality.

Figures

Figures reproduced from arXiv: 2605.29238 by Lianyan Fu, Rui Wang, Zihan Zhang.

**Figure 1.** Figure 1: Watts–Strogatz Note: Demonstrating the network topology with 50 nodes, 100 edges, and an average degree of 4.00. heterogeneity with 𝛼𝑔 ∼ 𝒩(0, 1.52 ) and 𝜇𝑋,𝑔 ∼ 𝒩(0, 1.02 ), or high heterogeneity with 𝛼𝑔 ∼ 𝒩(0, 3.02 ) and 𝜇𝑋,𝑔 ∼ 𝒩(0, 2.02 ). Individual covariates are then drawn as 𝑋𝑖 ∼ 𝒩(𝜇𝑋,𝑔(𝑖), 1), independently across individuals. This specification ensures that the distribution of 𝑋𝑖 varies across grou… view at source ↗

**Figure 2.** Figure 2: Industry Exposure Distribution distribution of 𝑇 = 3 peaks sharply at 0.20 ∼ 0.25, nearly coinciding with that of 𝑇 = 2, and shows limited common support with other groups; therefore, spillover effects on treated units are not separately identified [PITH_FULL_IMAGE:figures/full_fig_p031_2.png] view at source ↗

**Figure 3.** Figure 3: Propensity Score Distribution by Treatment Status [PITH_FULL_IMAGE:figures/full_fig_p031_3.png] view at source ↗

read the original abstract

This paper proposes a generalized Mundlak estimator based on graph neural networks (GME-GNN). The estimator is designed to mitigate bias arising from group-level heterogeneity and to accommodate within-group dependence among individuals. Traditional fixed-effects models handle group heterogeneity via group-specific intercepts, but require overly strict linear additivity and intra-group independence assumptions, and are confined to within-group comparisons. Rather than relying on intercepts, GME-GNN uses aggregated group-level balancing statistics to fully control between-group confounding, enabling valid cross-group comparisons and relaxing linearity constraints. It further employs graph neural network message-passing to adaptively learn nonlinear representations and capture intra-group interaction effects. Theoretical analysis shows that the estimator satisfies double robustness and is asymptotically normal. Simulation and empirical studies confirm its performance.

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 proposes GME-GNN to generalize Mundlak estimators via GNNs but the double-robustness result needs explicit conditions to hold up.

read the letter

The paper introduces GME-GNN, a graph neural network version of the generalized Mundlak estimator for settings with network confounding. It uses group-level balancing statistics instead of intercepts to control between-group effects and GNN message passing to handle nonlinear intra-group interactions.

This approach is new in its specific application to relax the within-group and linearity limits of fixed effects models in econometrics.

The paper does well at identifying the shortcomings of traditional methods and framing the estimator as enabling cross-group comparisons with double robustness.

The soft spots are in the theoretical support. The abstract states that the estimator satisfies double robustness and asymptotic normality but supplies no derivation steps, conditions, or details on how the GNN enters the influence function. The stress-test concern is valid here: aggregated group-level statistics may not fully control confounding under full network propagation, especially with degree heterogeneity or cycles. Without linking conditions to the adjacency matrix, the orthogonality for double robustness is not guaranteed from the description.

The simulation and empirical studies are mentioned as confirmation but lack specifics in the provided text.

This paper targets applied researchers in econometrics working with networked panel data who need more flexible estimators. A reader focused on causal methods for clustered data with networks would get value from the proposal.

It shows honest engagement with the Mundlak literature, so it deserves a serious referee to examine the proofs.

I recommend sending this to peer review.

Referee Report

2 major / 1 minor

Summary. The paper proposes a generalized Mundlak estimator based on graph neural networks (GME-GNN) to mitigate bias from group-level heterogeneity and accommodate within-group dependence under network confounding. It replaces group-specific intercepts with aggregated group-level balancing statistics to control between-group confounding (enabling cross-group comparisons and relaxing linearity), and uses GNN message-passing to learn nonlinear representations and capture intra-group interactions. Theoretical analysis asserts that the estimator satisfies double robustness and asymptotic normality; these claims are supported by simulation and empirical studies.

Significance. If the double-robustness and asymptotic-normality results hold under explicit conditions on the aggregation operator and network structure, the work would extend Mundlak-type estimators to nonlinear networked settings and permit cross-group inference without strict additivity or independence assumptions. The simulation and empirical studies provide concrete evidence of finite-sample performance, which strengthens the practical contribution even if the theory requires further elaboration.

major comments (2)

[Theoretical analysis] Theoretical analysis section: the double-robustness claim is asserted without derivation steps showing how the GNN component enters the influence function or the explicit orthogonality condition that would follow from the aggregated balancing statistics; this is load-bearing for the central theoretical result.
[Abstract and confounding control section] Abstract and section on confounding control: the premise that aggregated group-level balancing statistics fully control between-group confounding is invoked to justify cross-group comparisons, yet no explicit condition is stated linking the aggregation operator to the network adjacency matrix that would guarantee the required conditional exchangeability when dependence can propagate along cycles or degree heterogeneity.

minor comments (1)

[Abstract] The abstract would be clearer if it included a brief statement of the key identifying assumption or the form of the influence function.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the presentation of our theoretical results. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

Referee: [Theoretical analysis] Theoretical analysis section: the double-robustness claim is asserted without derivation steps showing how the GNN component enters the influence function or the explicit orthogonality condition that would follow from the aggregated balancing statistics; this is load-bearing for the central theoretical result.

Authors: We agree that the double-robustness property requires more explicit derivation to show the contribution of the GNN layers. The current manuscript states the result but does not fully unpack the steps linking the message-passing operator to the influence function or the resulting orthogonality. In the revision we will add a detailed derivation in the theoretical analysis section (and appendix) that explicitly derives the influence function under the GNN aggregation and states the orthogonality condition induced by the balancing statistics. revision: yes
Referee: [Abstract and confounding control section] Abstract and section on confounding control: the premise that aggregated group-level balancing statistics fully control between-group confounding is invoked to justify cross-group comparisons, yet no explicit condition is stated linking the aggregation operator to the network adjacency matrix that would guarantee the required conditional exchangeability when dependence can propagate along cycles or degree heterogeneity.

Authors: The manuscript invokes the balancing property of the aggregated statistics to achieve conditional exchangeability for cross-group comparisons. We acknowledge that an explicit link between the chosen aggregation operator and the adjacency matrix (accounting for cycles and degree heterogeneity) is not stated as a formal assumption or lemma. In the revision we will insert a new assumption in the confounding-control section that specifies the required conditions on the aggregation operator relative to the network structure to guarantee the exchangeability result. revision: yes

Circularity Check

0 steps flagged

No circularity: theoretical claims presented as independent analysis

full rationale

The provided abstract and description frame the GME-GNN estimator's double robustness and asymptotic normality as outcomes of theoretical analysis, without any exhibited equations, fitted parameters renamed as predictions, or self-citation chains that reduce the central claims to their own inputs by construction. No self-definitional steps, uniqueness theorems imported from the authors, or ansatzes smuggled via citation appear in the text. The derivation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract; no explicit free parameters, axioms, or invented entities are stated. The balancing statistics and GNN layers are described at a high level without detailing any fitted constants or unproven lemmas.

pith-pipeline@v0.9.1-grok · 5653 in / 1136 out tokens · 21041 ms · 2026-06-29T00:22:38.568595+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references

[1]

G., Elder, T

Altonji, J. G., Elder, T. E. & Taber, C. R. (2005), ‘Selection on observed and unobserved variables: Assessing the effectiveness of catholic schools’, Journal of political economy 113(1), 151–184. Arkhangelsky, D. & Imbens, G. W. (2024), ‘Fixed effects and the generalized mundlak estimator’, Review of Economic Studies 91(5), 2545–2571. Barron, A. R. & She...

arXiv 2005
[2]

G., Saul, B., Clemens, J

Liu, L., Hudgens, M. G., Saul, B., Clemens, J. D., Ali, M. & Emch, M. E. (2019), ‘Doubly robust estimation in observational studies with partial interference’, Stat 8(1), e214. Mundlak, Y. (1978), ‘On the pooling of time series and cross section data’, Econometrica: Journal of the Econometric Society pp. 69–85. Published by JSTOR. Neyman, J. & Scott, E. L...

2019

[1] [1]

G., Elder, T

Altonji, J. G., Elder, T. E. & Taber, C. R. (2005), ‘Selection on observed and unobserved variables: Assessing the effectiveness of catholic schools’, Journal of political economy 113(1), 151–184. Arkhangelsky, D. & Imbens, G. W. (2024), ‘Fixed effects and the generalized mundlak estimator’, Review of Economic Studies 91(5), 2545–2571. Barron, A. R. & She...

arXiv 2005

[2] [2]

G., Saul, B., Clemens, J

Liu, L., Hudgens, M. G., Saul, B., Clemens, J. D., Ali, M. & Emch, M. E. (2019), ‘Doubly robust estimation in observational studies with partial interference’, Stat 8(1), e214. Mundlak, Y. (1978), ‘On the pooling of time series and cross section data’, Econometrica: Journal of the Econometric Society pp. 69–85. Published by JSTOR. Neyman, J. & Scott, E. L...

2019