Back to Feedback: Dynamics and Heterogeneity in Panel Data

Stephane Bonhomme

The paper argues that credible panel analysis requires replacing strict exogeneity with sequential exogeneity—a shift that makes standard two-way fixed-effects and difference-in-differences estimators biased and breaks point identification

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 15:12 UTC pith:P2J7UNGM

load-bearing objection A clear, well-written survey making the case that sequential exogeneity deserves to be the default in panel applications; no new results, but the synthesis and examples are valuable.

arxiv 2512.17576 v2 pith:P2J7UNGM submitted 2025-12-19 econ.EM

Back to Feedback: Dynamics and Heterogeneity in Panel Data

Stephane Bonhomme This is my paper

classification econ.EM MSC 62P2091B82

keywords sequential exogeneitystrict exogeneityfeedbackpanel datadifference-in-differencesheterogeneous treatment effectsidentificationdynamic models

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that strict exogeneity—the assumption that current shocks are uncorrelated with all past and future covariates—is too strong for most empirical settings, since treatments and covariates typically respond to past outcomes. It proposes sequential exogeneity, which allows such feedback, as the more credible benchmark, and shows that standard two-way fixed effects and difference-in-differences estimators become biased and inconsistent in short panels when only sequential exogeneity holds. The paper further shows that, with coefficient heterogeneity, allowing feedback destroys point identification of average treatment effects on changers, so analysts must instead estimate identified weighted averages or construct bounds. Finally, it surveys methods for nonlinear and network models that remain valid under feedback. If true, this reasoning implies a shift in how applied researchers should specify and interpret panel data models.

Core claim

The central claim is that credible empirical work requires meaningfully relaxing strict exogeneity assumptions, replacing them with sequential exogeneity under which current shocks are mean-independent of past and current covariates but may be correlated with future covariates. The paper demonstrates that, under sequential exogeneity, fixed-effects and difference-in-differences estimators have a bias of order 1/T in panel data, that pre-trend checks do not reveal this bias, and that in models with unrestricted coefficient heterogeneity average treatment effects on changers are not point-identified. It then characterizes the identified weighted averages and the bounds that remain available, a

What carries the argument

The central object is the feedback process—the conditional density of current covariates given past outcomes and covariates—which strict exogeneity rules out and sequential exogeneity allows. The argument is carried by: (i) the sequential moment restrictions that lagged covariates satisfy when strict exogeneity fails, which underlie classic GMM and quasi-likelihood estimators; (ii) a linear characterization of point-identified weighted averages of heterogeneous coefficients, leading to a best identified approximation; (iii) moment inequalities that yield finite bounds under conditional sequential exogeneity; and (iv) feedback-and-heterogeneity-robust moment conditions in nonlinear models, wi

Load-bearing premise

The whole toolkit hinges on sequential exogeneity being the right relaxation: it rules out simultaneity and serially correlated time-varying confounders, so if those are present the methods no longer deliver credible estimates.

What would settle it

Simulate a two-period binary-treatment panel where the treatment depends on the lagged outcome shock, and verify that the two-way fixed-effects estimator deviates from the true ATT while the sequential-exogeneity-robust weighted average recovers it. Alternatively, re-examine a job-training evaluation with a randomized benchmark; if the sequential-exogeneity-based estimate does not move closer to the benchmark than the difference-in-differences estimate does, the framework's practical value is weakened.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Two-way fixed effects and event-study estimates should be interpreted as depending on strict exogeneity; under feedback they are biased by an amount that shrinks only like 1/T.
Pre-trend checks cannot establish strict exogeneity; parallel pre-trends can coexist with substantial feedback bias, so they should not be used as validation.
With heterogeneous treatment effects and feedback, the average effect on changers is not identified; researchers should target identified weighted averages or report partial-identification bounds.
Nonlinear panel models admit feedback-robust moment conditions in restricted classes (multiplicative, Poisson, proportional hazard), and their identified sets can be computed via linear programming.
Network models of worker-firm mobility need to relax the exogenous-mobility assumption to allow job changes to respond to past wage shocks; methods for this are at an early stage.

Where Pith is reading between the lines

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

A natural extension the paper leaves implicit: in designs with many periods, feedback bias shrinks as T grows, so combining short-T robust methods with long-T bias corrections could restore some of the point identification lost under heterogeneity.
The linear-programming characterization of identified sets suggests that computational inference approaches—e.g., checking whether a target parameter lies in the identified set—could be implemented with off-the-shelf LP solvers in empirical work.
The paper's logic implies that external instruments matter more than often acknowledged: because sequential exogeneity cannot handle simultaneity or serially correlated confounders, instruments external to the model remain necessary in those settings.
The framework invites a re-examination of published difference-in-differences studies to assess whether feedback is plausible from the institutional context, and to report sensitivity of conclusions to replacing strict exogeneity with sequential exogeneity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A clear, well-written survey making the case that sequential exogeneity deserves to be the default in panel applications; no new results, but the synthesis and examples are valuable.

read the letter

You should know two things up front. First, this is a review/position piece, not a new estimator or theorem. It argues that strict exogeneity in panel and did designs is often implausible and that sequential exogeneity should be the default starting point. That argument is not new, but the paper makes it cleanly and honestly. Second, the most modern material leans on the author's own unpublished working papers, but the core message does not depend on those: Chamberlain's negative result is published and the standard Arellano-Bond/Blundell-Bond machinery is textbook material.

The paper does a lot well. It defines strict versus sequential exogeneity carefully, spells out the feedback-bias mechanism, and shows clearly why TWFE and event-study estimates are inconsistent under SeqE. The three-period example where pre-trends are parallel yet the DID estimand is biased is a nice expository contribution—simple and effective. Section 7 gives a compact characterization of which weighted averages are point-identified under SeqE plus coefficient heterogeneity, and the Lee bounds are laid out correctly without getting lost in technicalities. The network discussion is brief but responsible.

Soft spots are minor and mostly inherent to the genre. There is no new empirical evidence that strict exogeneity fails broadly; the case rests on examples and on the logical point that SE rules out feedback by construction. The unpublished working papers for the nonlinear parts mean a referee cannot fully verify those claims from this manuscript alone, but the main substantive conclusion—that average effects are generally not point-identified under SeqE and heterogeneity—already follows from the published Chamberlain result. Also, the paper itself acknowledges that SeqE still rules out simultaneity and serially correlated time-varying confounders (Section 2.1), so it does not oversell the assumption. That caveat is stated clearly, which I appreciate.

The citation pattern is fine. Self-citations are to working papers, but they are identified as such and the attribution is accurate. No invented entities, no hidden free parameters.

Who gets value from this? Applied microeconomists who want an accessible map of dynamic panel methods and the modern heterogeneity-plus-feedback literature, and graduate students who need context before diving into technical papers. It is not for someone hunting for new results.

My recommendation: send it to peer review as a survey. It deserves a serious referee. If a journal wants a review article, this should be accepted after light-to-moderate revision. I would bring it to a reading group and would cite it when discussing sequential exogeneity.

Referee Report

0 major / 5 minor

Summary. The paper is a survey and position piece arguing that strict exogeneity (SE) is too restrictive as a default assumption in panel data and difference-in-differences practice, and that sequential exogeneity (SeqE)—which allows feedback from past outcomes to future covariates—should be adopted as a less restrictive starting point. It defines SE and SeqE, illustrates the bias of OLS, fixed-effects, and event-study estimators under feedback, and cautions against relying on pre-trend checks. It then reviews classic linear dynamic panel methods (Arellano-Bond, Blundell-Bond, QML, bias correction), and moves to heterogeneous coefficients: Chamberlain's negative result, point-identified weighted averages, best identified approximations, Lee bounds, and scalar-heterogeneity/large-T fixes. The final sections cover nonlinear likelihood models with feedback, linear-programming characterizations of identified sets, restricted feedback processes, and network extensions.

Significance. The paper's main value is synthetic and pedagogical. It draws together classic and recent results on feedback in panel data and presents them in a coherent framework. The three-period pre-trends counterexample in Section 5.2 is particularly effective: it gives a transparent case where pre-trends are parallel yet the DID estimand is biased. The review of Section 7 is also useful, especially the linear-weight characterization of point-identified averages and the contrast between unconditional and conditional sequential exogeneity. The paper is appropriately hedged: Section 2.1 explicitly acknowledges that SeqE rules out simultaneity and serially correlated time-varying confounders, so the advocacy is not overstated. If the survey is accepted as a position piece, it can help redirect applied practice toward dynamic feedback and partial identification rather than strict exogeneity. The main caveat is that several technical claims are deferred to the author's unpublished companion papers, but these are not load-bearing for the paper's central message.

minor comments (5)

[Section 3.1] In the bias decomposition after Eq. (1), the text defines term (I) as (X1X1'+X2X2')^{-1} X1 U1, but then writes E[(I)] = E[(X1X1'+X2X2')^{-1} X2 E[U1|X1,X2]]. The X2 outside the bracket should be X1. The conclusion that the term is generally nonzero under feedback is unaffected, but the displayed equality is false as written.
[Section 7.2] The completeness claim that, absent functions φ_t satisfying (39)–(40), the identified set for μ is the whole real line is stated without proof and attributed to Bonhomme (2025, unpublished). Since this is a survey, the attribution is acceptable, but the 'if and only if' and 'identified set equals R' claims are strong. Please add a short proof sketch or an explicit theorem number from the companion paper so the reader can verify the result.
[Section 9] In model (58), the time index is written as t=1,...,N; presumably this should be t=1,...,T. In addition, after (59), the notation 'X^t denotes the set of all X_{ijs} for ... s ≤ t' would be clearer if written as X^t = {X_{ijs}: s ≤ t}.
[Section 5.2] The sentence 'there is no reason at all for λ1=λ0 to imply that the bias in (23) is zero' is correct but could confuse readers. With λ1=λ0, the pre-trend condition (22) holds exactly, while the bias (23) is generally nonzero whenever λ1≠0 and the conditional means of Z_i differ across groups. A brief clarification would help.
[Minor typos] Section 2.1: 'counfounders' should be 'confounders'. Figure 3 caption: 'Identifed region' should be 'Identified region'. Section 8.1: 'Woutersen (2000) derive' should be 'Woutersen (2000) derives'.

Circularity Check

0 steps flagged

No significant circularity: the survey's central argument rests on published results and in-line derivations; self-citations to unpublished work are not definitionally load-bearing.

full rationale

This is a survey/position piece whose central claim is that strict exogeneity is often implausible and that sequential-exogeneity-based methods are a useful alternative. That claim is anchored in published external results (e.g., Chamberlain 2022, Arellano and Bond 1991, Lee 2020, Hahn and Kuersteiner 2002) and in elementary derivations shown in the text, such as the DID bias decomposition in Section 5.1 and the sufficient conditions for identified weighted averages in Section 7.2. The sharp necessity characterizations are attributed to the author's own unpublished work (Bonhomme 2025 in Section 7.2; Bonhomme, Dano, and Graham 2025 in Section 8.2), but they are not used to define the target quantities, and the paper's main message does not reduce to these citations. No fitted parameter is renamed as a prediction, and no equation is shown to be equal to another by construction. The self-citations are better viewed as a verifiability/rigor caveat about unpublished sources than as a circular derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The paper contains no fitted parameters or estimated constants; numerical choices (π1, λ0, λ1, Bernoulli designs) are illustrative and not load-bearing. The load-bearing assumptions are the model classes: SeqE, the random-coefficient linear structure, support/relevance conditions, and the semiparametric likelihood structure for the nonlinear results. No new entities are postulated.

axioms (5)

domain assumption Sequential exogeneity—Er[Uit | Xi1,...,Xit] = 0, with or without conditioning on Ai—is the maintained exogeneity notion; strict exogeneity is ruled out.
Definitions and equations (3), (11), (12), (15), (34), (43). The entire review evaluates methods under SeqE; if contemporaneous selection or serial correlation is present, the methods do not apply.
domain assumption In heterogeneous-coefficient sections, the model is Yit = Bi Xit + Ai + Ft + Uit with unrestricted (Ai, Bi) and binary or continuous Xit.
Equations (33), (37)-(38), (46). Identification characterizations and bounds are specific to this random-coefficients linear model.
domain assumption Support/relevance conditions hold, e.g., Xi2 - Xi1 bounded away from zero in the two-period mover analysis.
Section 7.1 and footnote 10. Without such conditions, the first-difference identification argument collapses.
domain assumption In the nonlinear section, the outcome density fθ is correctly specified: Yit | Yi,t-1, Xit, Ai ∼ fθ, while feedback and heterogeneity distributions are unrestricted.
Section 8.2 setup. The FHR characterization and LP identified-set formulas are relative to this semiparametric likelihood structure.
ad hoc to paper In the pre-trends counterexample, conditional means are linear and selection depends on an index Zi independent of the current shock Ui2.
Section 5.2. Used only to construct an example where parallel pre-trends hold while strict exogeneity fails; not a general assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 26781 in / 13597 out tokens · 148444 ms · 2026-08-03T15:12:27.667740+00:00 · methodology

0 comments

read the original abstract

Many popular estimation methods in panel data rely on the assumption that the covariates of interest are strictly exogenous. However, this assumption is empirically restrictive in a wide range of settings. In this paper I argue that credible empirical work requires meaningfully relaxing strict exogeneity assumptions. Econometricians have developed methods that allow for sequential exogeneity, which in contrast with strict exogeneity allows for the presence of feedback from past outcomes to future covariates or treatments. I review some of the classic work on linear models with constant coefficients, and then describe some approaches that allow for coefficient heterogeneity in models with feedback. Finally, in the last two parts of the paper I review recent work that allows for sequential exogeneity in nonlinear panel data models, and mention possible extensions to network settings.

discussion (0)

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Inference for Fixed Effects Estimators when Panels are Unbalanced
econ.EM 2026-07 conditional novelty 7.0

In unbalanced panels, two-way fixed-effects M-estimators have incidental-parameter plus feedback bias, and a proposed analytical correction restores correctly centered normal inference without knowing which regressors...
Inference for Fixed Effects Estimators when Panels are Unbalanced
econ.EM 2026-07 accept novelty 6.0

Debiased two-way fixed-effects M-estimators for unbalanced panels remain centered under mixed deterministic/stochastic selection and predetermined regressors or attrition.

Reference graph

Works this paper leans on

5 extracted references · 1 linked inside Pith · cited by 1 Pith paper

[1]

Semiparametric difference-in-differences estimators,

Abadie, A. (2005): “Semiparametric difference-in-differences estimators,” The review of economic studies, 72(1), 1–19. Abowd, J. M., F. Kramarz, and D. N. Margolis (1999): “High wage workers and high wage firms,” Econometrica, 67(2), 251–333. Acemoglu, D., S. Naidu, P. Restrepo, and J. A. Robinson (2019): “Democracy does cause growth,” Journal of politica...

2005
[5]

A distributional framework for matched employer employee data,

Bonhomme, S., T. Lamadon, and E. Manresa (2019): “A distributional framework for matched employer employee data,” Econometrica, 87(3), 699–739. Borusyak, K., X. Jaravel, and J. Spiess (2024): “Revisiting event-study designs: robust and eﬀicient estimation,” Review of Economic Studies , p. rdae007. Botosaru, I., and L. Liu (2025): “Time-Varying Heterogeneo...

Pith/arXiv arXiv 2019
[131]

Functional differencing,

Bonhomme, S. (2012): “Functional differencing,” Econometrica, 80(4), 1337–1385. (2025): “Unrestricted Heterogeneity in Linear Econometric Models,” Working Paper. Bonhomme, S., K. Dano, and B. S. Graham (2023): “Identification in a binary choice panel data model with a predetermined covariate,” SERIEs, 14(3), 315–351. (2025): “Moment Restrictions for Nonli...

2012
[231]

Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations,

Oxford University Press. Arellano, M., and S. Bond (1991): “Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations,” The review of economic studies , 58(2), 277–297. Arellano, M., and S. Bonhomme (2012): “Identifying distributional characteristics in random coeﬀicients panel data models,” The Review of ...

1991
[381]

Panel data models: some recent developments,

Arellano, M., and B. Honoré (2001): “Panel data models: some recent developments,” in Hand- book of econometrics , vol. 5, pp. 3229–3296. Elsevier. 41 Arkhangelsky, D., and G. Imbens (2024): “Causal models for longitudinal and panel data: A survey,” The Econometrics Journal , 27(3), C1–C61. Ashenfelter, O., and D. Card (1985): “Using the Longitudinal Stru...

2001

[1] [1]

Semiparametric difference-in-differences estimators,

Abadie, A. (2005): “Semiparametric difference-in-differences estimators,” The review of economic studies, 72(1), 1–19. Abowd, J. M., F. Kramarz, and D. N. Margolis (1999): “High wage workers and high wage firms,” Econometrica, 67(2), 251–333. Acemoglu, D., S. Naidu, P. Restrepo, and J. A. Robinson (2019): “Democracy does cause growth,” Journal of politica...

2005

[2] [5]

A distributional framework for matched employer employee data,

Bonhomme, S., T. Lamadon, and E. Manresa (2019): “A distributional framework for matched employer employee data,” Econometrica, 87(3), 699–739. Borusyak, K., X. Jaravel, and J. Spiess (2024): “Revisiting event-study designs: robust and eﬀicient estimation,” Review of Economic Studies , p. rdae007. Botosaru, I., and L. Liu (2025): “Time-Varying Heterogeneo...

Pith/arXiv arXiv 2019

[3] [131]

Functional differencing,

Bonhomme, S. (2012): “Functional differencing,” Econometrica, 80(4), 1337–1385. (2025): “Unrestricted Heterogeneity in Linear Econometric Models,” Working Paper. Bonhomme, S., K. Dano, and B. S. Graham (2023): “Identification in a binary choice panel data model with a predetermined covariate,” SERIEs, 14(3), 315–351. (2025): “Moment Restrictions for Nonli...

2012

[4] [231]

Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations,

Oxford University Press. Arellano, M., and S. Bond (1991): “Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations,” The review of economic studies , 58(2), 277–297. Arellano, M., and S. Bonhomme (2012): “Identifying distributional characteristics in random coeﬀicients panel data models,” The Review of ...

1991

[5] [381]

Panel data models: some recent developments,

Arellano, M., and B. Honoré (2001): “Panel data models: some recent developments,” in Hand- book of econometrics , vol. 5, pp. 3229–3296. Elsevier. 41 Arkhangelsky, D., and G. Imbens (2024): “Causal models for longitudinal and panel data: A survey,” The Econometrics Journal , 27(3), C1–C61. Ashenfelter, O., and D. Card (1985): “Using the Longitudinal Stru...

2001