Estimating Social Network Models with Link Misclassification

Arthur Lewbel; Xi Qu; Xun Tang

REVIEW 3 major objections 5 minor 32 references

Estimating Social Network Models with Link Misclassification

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that an adjusted 2SLS estimator, which rescales the noisy network matrix using estimated misclassification rates and instruments with the transpose of that matrix or a second independent report, consistently estimates peer

desk verdict Adjusted 2SLS is a real fix for misclassified links, but it inherits the standard exogeneity of link formation (A3), and the 'no link-formation model' claim doesn't cover that. read the letter →

arxiv 2509.07343 v1 pith:6FPLV53L submitted 2025-09-09 econ.EM

Arthur Lewbel , Xi Qu , Xun Tang This is my paper · ORCID

classification econ.EM

keywords socialnetworkspeereffectslinkmisclassificationtwo-stageleastsquaresinstrumentalvariablesmeasurementerrormicrofinanceparticipation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Survey-based network links are often wrong: people forget or misreport who they connected with, and data entry lapses flip zeros and ones. This paper shows that such two-sided misclassification is not a nuisance that 2SLS can ignore: it makes every covariate endogenous, so conventional network estimators are inconsistent. The proposed fix is an adjusted two-stage least-squares estimator that rescales the noisy adjacency matrix with estimated misclassification rates, then builds instruments from the transpose of the noisy matrix or from a second independent report. The estimator needs no model of how links form, only that link errors are random and independent of outcome shocks, and it recovers misclassification rates in closed form. Applied to household participation in a microfinance program in Indian villages, it finds that an additional linked participant raises a household's participation by about 5.1 percent, while ignoring misclassification biases the peer effect upward.

What carries the argument

The adjusted adjacency matrix W(H,p0,p1) and its conditional-expectation property E(W|G,X)=G. This property is what makes Lemma 1 hold — the adjusted structural error is mean-zero given the true network and covariates — and it is what turns the usual invalid instrument HX into the valid instrument H'X, because the problematic conditional covariances vanish in the transpose. For row-normalized linear-in-means models, a row-specific invertible probability transformation fW plays the same role.

What would settle it

Simulate a many-group design with the same structure as the paper but let dyad link formation depend on the outcome shock, e.g., Pr(G_ij = 1) = Phi(a + b|eps_i + eps_j|), generate two noisy measures with known p0 and p1, and run the adjusted 2SLS. If the peer-effect estimate does not converge to the true lambda as the number of groups grows, the exogeneity assumption E(eps|G,X,H)=0 is the point of failure.

Watch

Extended reading notes

Core claim

The central claim is that the corrected adjacency matrix W = (H - p0(11' - I))/(1 - p0 - p1), built from the observed noisy matrix H and unknown rates p0 (false links) and p1 (missed links), satisfies E(W|G,X)=G, so the composite error in the adjusted outcome equation is mean-zero conditional on G and X. That restores exogeneity of the covariates X. The remaining endogeneity of the peer outcome Wy is handled by new instruments: with a single unsymmetrized report, H'X is valid even though HX is not; with two conditionally independent reports, one noisy report can instrument the other. The paper also identifies p0 and p1 from covariate-driven variation in link formation and from joint moments

Load-bearing premise

The load-bearing premise is that the true network links and the reported noisy links are unrelated to unobserved shocks that also affect outcomes; if people form or report links because of exactly those outcome shocks, the correction does not remove the bias.

Editorial extensions

If this is right

Applied researchers can estimate peer effects from misclassified survey links with a closed-form two-step 2SLS, without specifying a link-formation model or a likelihood for the true network.
When two conditionally independent network reports are available, one report can be used as an instrument for the other, regardless of whether the reports are symmetrized.
In the microfinance application, the peer effect is estimated at about 0.051, meaning one additional linked participant raises participation by roughly 5.1 percent; naive 2SLS that ignores misclassification produces upward-biased estimates.
The method extends to linear-in-means social interaction models through a finite invertible transformation of the noisy row, as long as p0 + p1 differs from 1.
Asymptotic normality with group-level clustering is established for the adjusted 2SLS estimators as the number of independent groups grows.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the transpose-instrument mechanism should generalize beyond social networks to any dyadic or interaction regressor measured with independent Bernoulli errors; the practical rule is to use the transpose of the noisy design as an instrument rather than the raw noisy design.
Editorial inference: because the method corrects reporting error but maintains exogeneity of true link formation, a natural next step is to pair W with instruments that are valid for G itself when link formation is endogenous.
Editorial inference: the authors themselves note one uncovered cell — an asymmetric true network with a single unsymmetrized report has valid instruments but no identified misclassification rates without an auxiliary link-formation model; researchers in that setting need a second report or an additional model.
Editorial inference: the covariate phi used to recover misclassification rates must actually shift link formation (pi1 != pi0); checking this first-stage relationship in the observed noisy reports is advisable, since a flat phi would leave the rates weakly identified.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick read of Lewbel-Qu-Tang. The headline: the paper gives applied network people a simple closed-form correction for misclassified links in linear-in-sums peer-effects models, and the misclassification-rate estimator from two noisy measures is genuinely new. It deserves refereeing.

The core trick—replace H by W=(H-p0(ιι'-I))/(1-p0-p1) and use H'X or the other noisy measure as instrument—is clean. Lemma 1 holds under A1–A3, and the proof that the H'X moments cancel using A4 is right. The closed-form moment equations in Appendix A2 for (p0,p1) are a real contribution: they don't require modeling Pr(G|X), only a binary φ(X) that shifts link propensity. Simulations confirm the bias correction, and the application to microfinance is a nice demonstration, even if the bias there is small.

The soft spot is A3: E(ε|G,X,H)=0. That's not a minor regularity condition; it says actual link formation is exogenous to the outcome error. The paper's "does not require structural modeling of link formation" is true only in a narrow sense—it avoids specifying Pr(G|X), but it does not relax exogeneity of G. If links are endogenous (common shocks that form friendships and affect outcomes), the adjusted estimator is inconsistent and the misclassification correction can't rescue it. The simulations always draw G from X plus independent ε, so they impose A3. Since the whole method is about fixing data problems, a referee worth her salt will ask for a discussion of when A3 is plausible, and possibly a diagnostic or partial-identification statement when A3 fails.

Other concerns are smaller: Proposition 5 relies on REG conditions in an online appendix that isn't in the preprint; the standard errors for the two-measure S2SLS are omitted ('for brevity'); and A4' between the visit-out and visit-in measures in the application is plausible but not tested. None of these are fatal.

Bottom line: serious paper, well-executed, with a real new result. The main caveat is that 'no link formation model' is not the same as 'no link endogeneity.' I'd cite it if I worked on measurement error in networks, and I'd send it to a referee with instructions to focus on A3.

Referee Report

3 major / 5 minor

Summary. The paper proposes a two-step adjusted 2SLS estimator for linear social network models y = λGy + Xβ + ε when the observed adjacency matrix H misclassifies the true links at fixed, unknown rates (p0, p1). The central device is the adjusted matrix W(H,p) = (H − p0(ιι′−I))/(1−p0−p1), which satisfies E(W|G,X)=G under assumptions (A1)-(A2). The authors show that the adjusted structural form y = λWy + Xβ + v, with v = ε + λ(G−W)y, has E(v|G,X)=0 under A3 (Lemma 1), and that conventional instruments such as HX are invalid while H′X (single unsymmetrized measure) or the other noisy measure H(2)X (two measures) are valid. They provide closed-form identification of the misclassification rates from moments of H(1), H(2), and X only, propose feasible two-step estimators, state an asymptotic normality result (Proposition 5), report Monte Carlo evidence, and apply the method to estimate peer effects in microfinance participation using Banerjee et al. (2013) data.

Significance. If the results hold, the paper makes a useful and practical contribution: it extends 2SLS estimation of peer effects to the empirically common case of non-vanishing link misclassification without specifying a full parametric model of network formation. The identification of the misclassification rates is based on network-report moments only, not on outcome moments, so the procedure is not circular in the outcome equation. The estimator is closed-form, simple to compute, and the proposed instruments (H′X for a single unsymmetrized measure, and cross-measure instruments for multiple measures) are new and intuitive. The simulation evidence supports the main consistency claims. The principal limitations are that exogeneity of the true network (A3) is maintained, and that the asymptotic distribution is stated under regularity conditions relegated to an Online Appendix that is not included in this preprint; both need attention before the paper is fully acceptable.

major comments (3)

[Section 3.1, Lemma 1, Eq. (6)] Assumption (A3), E(ε|G,X,H)=0, is load-bearing for the central consistency claim. If actual link formation G is correlated with unobserved shocks that also affect outcomes, Lemma 1 fails and the adjusted estimator is inconsistent; W corrects for the difference between H and G, not for endogeneity of G itself. The paper's repeated statement that the method 'does not require structural modeling of link formation' should be qualified: it avoids specifying Pr(G|X), but it does not relax exogeneity of G. The simulations in Section 5 generate G from X plus independent ε, so they impose A3 by construction and cannot reveal the failure. I recommend adding an explicit scope discussion and, if possible, a diagnostic or robustness discussion for the case where A3 is doubtful, especially since many network applications involve link formation driven by unobserved shocks.
[Section 4.2, Proposition 5 and Section 3.6, Proposition 4] The asymptotic distribution in Proposition 5 is stated under 'regularity conditions (REG) in the Online Appendix,' but the Online Appendix is not part of this arXiv submission. Since the application in Section 6 reports clustered standard errors based on this proposition, the inference part of the paper is not verifiable from the submitted manuscript. The authors should either include the regularity conditions in the main text or provide the Online Appendix as part of the submission. Similarly, Proposition 4, which underlies the linear-in-means extension in Section 3.6, has its proof only in the Online Appendix; at least a sketch of the induction argument should be in the main text.
[Section 3.4.1 and Section 6.2] Identification of the misclassification rates relies on a binary partition ϕ(X) with π1 ≠ π0. In the application this is same-caste versus different-caste pairings. While this is a weaker assumption than a full link-formation model, it is still a substantive assumption about link formation, and the 'no structural modeling of link formation' claim should be tempered accordingly. Moreover, the application also assumes conditional independence of the misclassification errors in the two symmetrized measures (A4′). Since the two measures come from different survey questions about visits, correlated recall errors across questions could violate this assumption and bias both the misclassification-rate estimates and the IV exogeneity. The paper should discuss the plausibility of this independence in the application and, ideally, provide some robustness evidence.

minor comments (5)

[Section 5, Table 1(b)] The reported 'Expected # of peers' values do not match the DGP. For n=25 with π1=0.2, π0=0.1, and P(X_i1=1)=0.5, the expected degree is 24×(0.5×0.2+0.5×0.1)=3.6, not 3.75; for n=50 it is 49×0.15=7.35, not 7.5. Please correct or explain the calculation.
[Section 4.2/4.3 and Tables 1(b), 1(c)] The text says the adjusted 2SLS uses W(t′)X as the instrument, while the tables report H(t′)X. Both are valid under Proposition 2, but the notation should be consistent.
[Section 5] The claim that the estimator is 'unlikely to suffer from weak instrument issues' is not backed by first-stage F-statistics or a formal strength analysis. Reporting first-stage F statistics would strengthen the finite-sample section.
[Section 3.3, Proposition 3] The rank conditions in Eq. (9) are called 'primitive,' but they involve M=(I−λG)−1, which depends on the unknown λ and G. They are better described as high-level moment-rank conditions; a more detailed discussion of their content would be helpful.
[Section 6.3 and Table 5] The caption of Table 5 refers to columns (a)-(e), but it does not define that these correspond to the estimators in Table 4. Please make the cross-reference explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: misclassification rates are identified from network-measure moments that exclude outcomes, and the adjusted-2SLS moment conditions are derived from stated assumptions rather than imposed by construction.

full rationale

The paper's derivation chain is not circular. The misclassification rates (p0,p1) are identified in Section 3.4 from moments of the noisy measures H(1), H(2) and covariates X (e.g., equations (11) and (24)), none of which involve the outcome y. The adjusted matrix W is a deterministic function of H and p, and the key property E(W|G,X)=G is derived from Assumptions (A1)-(A2), not assumed as the conclusion. Lemma 1, which restores exogeneity of X, is proved from (A3) and E(W|G,X)=G; the proof is a direct calculation. The instruments H'X (or W'X) are shown in Proposition 2 to be uncorrelated with the composite error v using (A3), (A4), and the derived identities E(W|G,X)=G and E(W^2|G,X)=G^2; this is a substantive moment condition, not an identity imposed by defining the instrument. The 2SLS estimand in Proposition 1 follows from the standard IV moment equation. The first-stage estimation of p is accounted for in the asymptotic distribution (Proposition 5) through the influence-function correction F0 τs; it is not treated as known. The only self-citations (Lewbel, 2007; Lewbel, Qu, and Tang, 2023) appear in the literature review to delineate contributions and are not load-bearing for any theorem. Assumption (A3), requiring E(ε|G,X,H)=0, is an explicit exogeneity condition; if it fails the estimator is inconsistent, but that is a limitation/robustness concern, not a circularity. No step in the paper reduces to fitting a parameter to the outcome and then predicting that outcome, nor does any uniqueness claim rest on a self-citation. The simulations impose the assumptions by design and confirm the algebra, but they are not needed for the identification argument. Therefore the central consistency claim has independent content.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard misclassification assumptions (A1, A2), exogeneity of the true network (A3), and conditional independence of misclassification errors (A4/A4'). The misclassification rates are identified from the data and are not free in the ad hoc sense, but they are estimated inputs to the main estimator. No new entities are postulated.

free parameters (6)

p(1)_0 = 0.0020
Estimated probability that a non-link is recorded as a link in the visit-go measure in the application (Section 6.2). This is an input to W(1).
p(1)_1 = 0.1425
Estimated probability that a real link is missed in the visit-go measure. Input to W(1).
p(2)_0 = 0.0001
Estimated false positive misclassification rate for the visit-come measure.
p(2)_1 = 0.1079
Estimated false negative misclassification rate for the visit-come measure.
pi_1 = 0.0357
Estimated probability of a true link between same-caste households; nuisance parameter in the identification of p (Section 3.4.1).
pi_0 = 0.0144
Estimated probability of a true link between different-caste households; nuisance parameter.

assumptions (6)

domain assumption A1: E(H_ij | G, X) = E(H_ij | G_ij, X)
Misclassification of a given link does not depend on the true status of other links. Invoked in Lemma 1 and Proposition 2 (Section 3.1).
domain assumption A2: E(H_ij | G_ij=1, X) = 1-p1, E(H_ij | G_ij=0, X) = p0, p0+p1<1
Fixed misclassification rates, positive correlation with true link. Invoked in Lemma 1 and identification.
domain assumption A3: E(eps | G, X, H) = 0
Links are exogenous to outcome errors. Necessary for Lemma 1 and all instruments; weakest assumption.
domain assumption A4/A4': conditional independence of misclassified entries (or across measures)
Needed for E(W^2 | G,X) = G^2 and instrument validity in Propositions 2 and Section 3.3.2.
domain assumption pi_1 != pi_0: the chosen function phi(X) affects link formation
Required for identification of misclassification rates in Section 3.4. If the observed covariate does not shift link probabilities, the moments are degenerate.
domain assumption Rank conditions (IV-R) and non-singularity in (9)
Ensures instruments are relevant; rules out the reflection problem and knife-edge cases (Proposition 3).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Estimating Social Network Models with Link Misclassification." pith.science (2026). https://pith.science/paper/6FPLV53L

@misc{pith2026250907343,
  author       = {Pith},
  title        = {Pith review of: Estimating Social Network Models with Link Misclassification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6FPLV53L}},
  note         = {Machine review of arXiv:2509.07343}
}

read the original abstract

We propose an adjusted 2SLS estimator for social network models when reported binary network links are misclassified (some zeros reported as ones and vice versa) due, e.g., to survey respondents' recall errors, or lapses in data input. We show misclassification adds new sources of correlation between the regressors and errors, which makes all covariates endogenous and invalidates conventional estimators. We resolve these issues by constructing a novel estimator of misclassification rates and using those estimates to both adjust endogenous peer outcomes and construct new instruments for 2SLS estimation. A distinctive feature of our method is that it does not require structural modeling of link formation. Simulation results confirm our adjusted 2SLS estimator corrects the bias from a naive, unadjusted 2SLS estimator which ignores misclassification and uses conventional instruments. We apply our method to study peer effects in household decisions to participate in a microfinance program in Indian villages.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 30 canonical work pages

[1]

Advani, A. and B. Malde (2018). Credibly identifying social effects: Accounting for network formation and measurement error. Journal of Economic Surveys\/ 32\/ (4), 1016--1044

work page 2018
[2]

Aigner, D. J. et al. (1973). Regression with a binary independent variable subject to errors of observation. Journal of Econometrics\/ 1\/ (1), 49--59

work page 1973
[3]

Auerbach, E. (2022). Identification and estimation of a partially linear regression model using network data. Econometrica\/ 90\/ (1), 347--365

work page 2022
[4]

Banerjee, A., A. G. Chandrasekhar, E. Duflo, and M. O. Jackson (2013). The diffusion of microfinance. Science\/ 341\/ (6144), 1236498

work page 2013
[5]

Blume, L. E., W. A. Brock, S. N. Durlauf, and Y. M. Ioannides (2011). Identification of social interactions. In Handbook of social economics , Volume 1, pp.\ 853--964. Elsevier

work page 2011
[6]

Bollinger, C. R. (1996). Bounding mean regressions when a binary regressor is mismeasured. Journal of Econometrics\/ 73\/ (2), 387--399

work page 1996
[7]

Jochmans, and J.-M

Bonhomme, S., K. Jochmans, and J.-M. Robin (2016). Non-parametric estimation of finite mixtures from repeated measurements. Journal of the Royal Statistical Society: Series B: Statistical Methodology\/ , 211--229

work page 2016
[8]

Boucher, V. and A. Houndetoungan (2020). Estimating peer effects using partial network data . Centre de recherche sur les risques les enjeux \'e conomiques et les politiques

work page 2020

Show all 32 references

[9]

Djebbari, and B

Bramoull\'e, Y., H. Djebbari, and B. Fortin (2009). Identification of peer effects through social networks. Journal of econometrics\/ 150\/ (1), 41--55

2009
[10]

Butts, C. T. (2003). Network inference, error, and informant (in) accuracy: a bayesian approach. social networks\/ 25\/ (2), 103--140

2003
[11]

Chandrasekhar, A. and R. Lewis (2011). Econometrics of sampled networks. Unpublished manuscript, MIT.[422]\/

2011
[12]

Hong, and E

Chen, X., H. Hong, and E. Tamer (2005). Measurement error models with auxiliary data. The Review of Economic Studies\/ 72\/ (2), 343--366

2005
[13]

Graham, B. S. (2020). Network data. In Handbook of Econometrics , Volume 7, pp.\ 111--218. Elsevier

2020
[14]

Griffith, A. (2022). Name your friends, but only five? the importance of censoring in peer effects estimates using social network data. Journal of Labor Economics\/ 40\/ (4), 779--805

2022
[15]

Griffith, A. and J. Kim (2023). The impact of missing links on linear reduced-form network-based peer effects estimates

2023
[16]

Hardy, M., R. M. Heath, W. Lee, and T. H. McCormick (2019). Estimating spillovers using imprecisely measured networks. arXiv preprint arXiv:1904.00136\/

2019 arXiv
[17]

Hausman, J. A., J. Abrevaya, and F. M. Scott-Morton (1998). Misclassification of the dependent variable in a discrete-response setting. Journal of econometrics\/ 87\/ (2), 239--269

1998
[18]

Hu, Y. (2008). Identification and estimation of nonlinear models with misclassification error using instrumental variables: A general solution. Journal of Econometrics\/ 144\/ (1), 27--61

2008
[19]

Hu, Y. and Z. Lin (2018). Misclassification and the hidden silent rivalry

2018
[20]

Hu, Y. and Y. Sasaki (2017). Identification of paired nonseparable measurement error models. Econometric Theory\/ 33\/ (4), 955--979

2017
[21]

Kelejian, H. H. and I. R. Prucha (1998). A generalized spatial two-stage least squares procedure for estimating a spatial autoregressive model with autoregressive disturbances. The Journal of Real Estate Finance and Economics\/ 17\/ (1), 99--121

1998
[22]

Klepper, S. (1988). Bounding the effects of measurement error in regressions involving dichotomous variables. Journal of Econometrics\/ 37\/ (3), 343--359

1988
[23]

Lee, L.-F. (2007). Identification and estimation of econometric models with group interactions, contextual factors and fixed effects. Journal of Econometrics\/ 140\/ (2), 333--374

2007
[24]

Lewbel, A. (2007). Estimation of average treatment effects with misclassification. Econometrica\/ 75\/ (2), 537--551

2007
[25]

Qu, and X

Lewbel, A., X. Qu, and X. Tang (2023). Ignoring measurement errors in social networks. The Econometrics Journal\/ , utad028

2023
[26]

Li, T. (2002). Robust and consistent estimation of nonlinear errors-in-variables models. Journal of Econometrics\/ 110\/ (1), 1--26

2002
[27]

Lin, X. (2010). Identifying peer effects in student academic achievement by spatial autoregressive models with group unobservables. Journal of Labor Economics\/ 28\/ (4), 825--860

2010
[28]

Liu, X. (2013). Estimation of a local-aggregate network model with sampled networks. Economics Letters\/ 118\/ (1), 243--246

2013
[29]

Mahajan, A. (2006). Identification and estimation of regression models with misclassification. Econometrica\/ 74\/ (3), 631--665

2006
[30]

Manski, C. F. (1993). Identification of endogenous social effects: The reflection problem. The review of economic studies\/ 60\/ (3), 531--542

1993
[31]

Molinari, F. (2008). Partial identification of probability distributions with misclassified data. Journal of Econometrics\/ 144\/ (1), 81--117

2008
[32]

Shalizi, C. R. and A. Rinaldo (2013). Consistency under sampling of exponential random graph models. Annals of statistics\/ 41\/ (2), 508

2013

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Advani, A. and B. Malde (2018). Credibly identifying social effects: Accounting for network formation and measurement error. Journal of Economic Surveys\/ 32\/ (4), 1016--1044

work page 2018

[2] [2]

Aigner, D. J. et al. (1973). Regression with a binary independent variable subject to errors of observation. Journal of Econometrics\/ 1\/ (1), 49--59

work page 1973

[3] [3]

Auerbach, E. (2022). Identification and estimation of a partially linear regression model using network data. Econometrica\/ 90\/ (1), 347--365

work page 2022

[4] [4]

Banerjee, A., A. G. Chandrasekhar, E. Duflo, and M. O. Jackson (2013). The diffusion of microfinance. Science\/ 341\/ (6144), 1236498

work page 2013

[5] [5]

Blume, L. E., W. A. Brock, S. N. Durlauf, and Y. M. Ioannides (2011). Identification of social interactions. In Handbook of social economics , Volume 1, pp.\ 853--964. Elsevier

work page 2011

[6] [6]

Bollinger, C. R. (1996). Bounding mean regressions when a binary regressor is mismeasured. Journal of Econometrics\/ 73\/ (2), 387--399

work page 1996

[7] [7]

Jochmans, and J.-M

Bonhomme, S., K. Jochmans, and J.-M. Robin (2016). Non-parametric estimation of finite mixtures from repeated measurements. Journal of the Royal Statistical Society: Series B: Statistical Methodology\/ , 211--229

work page 2016

[8] [8]

Boucher, V. and A. Houndetoungan (2020). Estimating peer effects using partial network data . Centre de recherche sur les risques les enjeux \'e conomiques et les politiques

work page 2020