Heterogeneity in peer effects for binary outcomes

Mathieu Lambotte

REVIEW 4 major objections 4 minor 3 references

Heterogeneity in peer effects for binary outcomes

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The paper establishes that the strength of peer conformity can differ between choosing the high and low action in a binary network game, and provides conditions under which both parameters are uniquely identified and estimable.

desk verdict A real theoretical contribution—separate identification of action-specific conformity in a binary network game—with an empirical section that is honest but fragile, and an abstract that overclaims policy simulations that are not in the paper. read the letter →

arxiv 2511.15891 v5 pith:EWNLP55L submitted 2025-11-19 econ.EM

Mathieu Lambotte This is my paper · ORCID

classification econ.EM

keywords peereffectsconformitybinaryoutcomesnetworkgameincompleteinformationidentificationheterogeneitysocialnorms

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

This paper argues that the standard 'one conformity parameter fits all' model is misspecified for binary choices. It introduces a model in which the cost of deviating from one's friends' average behavior is allowed to be different when the agent chooses the low action (beta_l) versus the high action (beta_h), and proves that both parameters can be separately identified. The identification works even when no isolated players exist in the network, which earlier approaches required. Applied to smoking among U.S. secondary school students, the model estimates beta_l=3.98 and beta_h=1.08, showing that not smoking when friends smoke is much more costly than smoking when friends abstain; for alcohol, the two are indistinguishable. If the homogeneous assumption is wrong, peer-effect estimates and anti-smoking policy counterfactuals based on it will be biased.

What carries the argument

The central object is the heterogeneous quadratic social-distance function S_het(y_i,y_-i) = y_i * (beta_h/2)(y_i - bar y_i)^2 + (1-y_i)*(beta_l/2)(y_i - bar y_i)^2, where bar y_i is the expected share of friends choosing the high action. Since the function is quadratic, the expected best response includes the term (1/2)(beta_l - beta_h) g_i Sigma g_i', where Sigma = pp' + diag(p*(1-p)); this curvature term is what breaks the symmetry of the homogeneous model and permits separate identification. The contraction condition |beta_h| + 1.5|beta_l - beta_h| < 1/max f_eta guarantees a unique Bayesian Nash equilibrium used for estimation.

What would settle it

Estimate the reduced-form equation with the social-norm term and the nonlinear gSigma g' term and test whether the marginal effect of the expected norm is constant (no curvature). A dataset with no isolated players in which the norm's marginal effect is flat across the support of the norm would violate the model's predicted convexity and show that the two conformity parameters are an artifact of the quadratic functional form.

Watch

Extended reading notes

Core claim

The discovery is that action-specific tastes for conformity are separately identifiable in an incomplete-information network game with binary outcomes, provided the social-distance function is quadratic and action-specific. The best response of each player then contains the nonlinear term 1/2 (beta_l - beta_h) g_i Sigma g_i', which gives the extra variation needed to separate beta_l from beta_h without relying on isolated players. The paper shows equilibrium uniqueness under a contraction condition, provides a specification test that distinguishes conformity from spillover microfoundations, and estimates the model on school friendship networks. The smoking estimates indicate that the taste f

Load-bearing premise

The load-bearing premise is that the disutility of deviating from friends' average behavior is quadratic in the deviation: if it is linear or absolute instead, the term that separately identifies beta_l and beta_h collapses, and identification fails without isolated players in the network.

Editorial extensions

If this is right

Smoking studies using homogeneous conformity parameters will produce biased estimates of peer effects and misleading counterfactuals.
Policies that target the taste for conformity when choosing the low action can raise the equilibrium, while policies targeting the high action lower it; a global norm nudge can backfire when the high-action conformity taste is weak.
The spillover and conformity models, equivalent under homogeneity, become testable against each other once beta_h and beta_l are allowed to differ; smoking data reject the spillover model.
For behaviors such as alcohol where beta_h is close to beta_l, the homogeneous model remains a good approximation.

Reading between the lines

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

The identification strategy is only as good as the quadratic distance assumption: with a linear distance function, separate identification requires isolated players, which are rare in many school networks, so a non-quadratic disutility would make beta_l and beta_h functional-form artifacts.
A direct falsifying test is to estimate the reduced form with the nonlinear term and check for curvature in the marginal effect of the norm; datasets where that marginal effect is flat across the norm's support would undermine the heterogeneous quadratic conformity model.
If action-specific conformity generalizes, norm-based interventions should be designed separately for each action—for example, anti-smoking messages that emphasize the rarity of smoking may fail to deter smokers who do not conform, while strongly penalizing non-smokers.

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

The core result is worth knowing: Lambotte shows that in an incomplete-information binary network game with a quadratic social-distance function, the conformity parameters for the high and low actions, beta_h and beta_l, can be separately identified, and that the usual equivalence between spillover and conformity microfoundations breaks down once heterogeneity is allowed. That is genuinely new relative to Xu (2018) and Badev (2021), and it is done carefully. The uniqueness proof is a clean contraction argument, the identification proof is standard but correct, and the NPL estimation is appropriate. The paper is honest about its limitations: the rank condition in Assumption 4 is not directly testable, and the alternative conditions in Proposition 3 require an exclusion-like restriction that is itself untested. That is normal for this literature, not a flaw.

The empirical application is a proof of concept, and it delivers an intuitive result: for teen smoking, the conformity cost is larger for non-smokers than smokers, while for drinking the two parameters are statistically indistinguishable. The alcohol null is a nice placebo. But the empirical support is more fragile than the abstract suggests. First, the abstract claims policy simulations; there are none in the body. That should be fixed before publication. Second, the specification test is weaker than it looks: the conformity model is not rejected at the 0.1% level, but it is rejected at conventional 5% or 1% levels. Reporting only the strict threshold is selective. Third, and most important, the separate identification of beta_h and beta_l rests entirely on the quadratic distance function. The g Sigma g' term in the best response exists only under that functional form. With a linear distance function, only the sum of the two conformity parameters is identified unless there are isolated players, and Add Health has almost none. The paper is transparent about this in Appendix B, but it does not test the quadratic form against other nonlinearities. So the smoking heterogeneity could be an artifact of the assumed distance function. This is a genuine soft spot, though not a fatal one: the theoretical result stands as a conditional contribution, and the empirical heterogeneity should be interpreted as “under quadratic distance,” not as a robust finding.

The citation pattern is fine; the one self-citation is to a related paper on network endogeneity and does not carry the identification.

I would send this to a serious referee. The theory is solid and useful for applied researchers who are willing to take the quadratic distance seriously. The revision should remove the policy-simulation claim from the abstract and add a robustness discussion of the functional-form assumption. This is a conditional accept, not a reject.

Referee Report

4 major / 4 minor

Summary. The paper proposes a structural model of binary choices in networks with action-specific heterogeneity in conformity preferences. Agents play a simultaneous incomplete-information game, and the paper derives a sufficient condition for a unique Bayesian Nash equilibrium, a contraction-based proof of uniqueness (Prop. 1), and an identification result for the heterogeneous conformity parameters (Prop. 2) under a high-level rank condition. The model nests the standard homogeneous-conformity model. The empirical section estimates the model on Add Health data for smoking and alcohol using the NPL estimator, finds significant heterogeneity for smoking (β_h=1.077, β_l=3.980) but not for drinking, and reports specification tests suggesting the data are inconsistent with a pure spillover model and, at a very strict significance level, consistent with the conformity model. The paper also discusses alternative distance functions in Appendix B and shows that separate identification under those alternatives requires isolated players, plus it proves a non-equivalence result between spillover and conformity microfoundations when heterogeneity is present.

Significance. If the identification and estimation results are taken at face value, the paper makes a useful contribution to the empirical peer-effects literature: it shows how one can separately identify the taste for conformity when choosing the high action and when choosing the low action in a binary network game, thereby going beyond the homogeneous-conformity assumption of Lee et al. (2014) and related work. The uniqueness proof and the NPL estimation strategy are careful and standard, and the empirical application is a transparent proof-of-concept on an influential dataset. The specification tests that distinguish spillover and conformity microfoundations under heterogeneity are a useful addition. The caveats are real, however: the central identification result depends on the quadratic social-distance functional form, and the empirical claims are conditional on that form. The paper would be strengthened by robustness analysis or a clearer statement of this limitation. Overall, the theoretical core is sound, but the empirical interpretation and the abstract overstate what has been demonstrated.

major comments (4)

[§4.1 and Appendix B, Eq. (7)] The separate identification of β_h and β_l rests on the quadratic, action-specific social distance function, which introduces the term g_i Σ g_i' into the best response (Eq. 7). The paper itself shows in Appendix B that with a linear or aggregate distance function, the two conformity parameters are separately identified only if isolated players exist, and in the Add Health networks isolated students are rare or absent. The specification test in Appendix A.3 tests β_3=0 versus β_3≠0, which detects the presence of the gΣg' term, but it does not test the quadratic functional form against other possible nonlinearities. As a result, the reported smoking heterogeneity (β_h=1.077, β_l=3.980) could be an artifact of the assumed quadratic distance function. The paper should either provide robustness checks using alternative distance functions (where identification is possible) or clearly state th
[Abstract] The abstract states that the paper is 'conducting policy simulations' and that homogeneous conformity leads to biased 'ex ante policy evaluations.' However, the body of the paper contains no policy simulations: Section 5 reports structural estimates and specification tests, and Section 6 only discusses implications qualitatively. This is a material discrepancy between what the paper claims and what it delivers. Either add the policy simulations discussed in the introduction (e.g., the effect of changing β_h or β_l on equilibrium smoking prevalence) or remove the claim from the abstract and conclusions.
[§5.2 and footnote 20] The conclusion that the conformity model 'cannot be rejected' for smoking is threshold-dependent. The paper reports p=0.002 for the conformity specification test; this means the null is rejected at conventional 5% and 1% levels and is only not rejected at the unusually strict 0.1% level. The choice of 0.1% to claim consistency is not justified beyond stating a preference for strict significance thresholds. Since the entire empirical distinction between spillover and conformity for smoking depends on this non-rejection, the paper should report the raw p-value and provide a more balanced interpretation, or justify the threshold with a pre-specified power/type-I error trade-off.
[Assumption 4(ii) and Proposition 3] The rank condition in Assumption 4(ii) is imposed on k_i, which includes the unobserved equilibrium probabilities p*, so the condition is not directly checkable. Proposition 3 offers sufficient conditions but relies on the untested exclusion-like condition γ2,κ≠0. The paper acknowledges this but does not provide empirical evidence that the rank condition holds in the Add Health application. This is not fatal for the identification theorem, which is conditional on the model, but the empirical claim of separately identified β_h and β_l would be stronger if the paper reported diagnostics for the variation in g_i Σ g_i' and for the instrument strength implied by Proposition 3.

minor comments (4)

[Acknowledgements] Typographical errors: 'dentification and estimation' should be 'Identification and estimation'; 'AddHealthdatathroughhischair' is malformed. The paper appears to have LaTeX/PDF extraction artifacts generally; a full proofread is needed.
[§5.2] In the text around Figure 1, 'pβ is estimated at 2.392' and 'xβ_h=1.077' appear to be typographical corruption of 'β' and 'β_h'. These should be corrected.
[§5.2] The notation '0.01% significance level' is inconsistent with the later '0.1%' in the same paragraph. Since p=0.002, the statement 'cannot be rejected at a 0.01% significance level' is mathematically true but confusing; the intended threshold appears to be 0.1% (0.001).
[Corollary 1] The notation p*(1), p*(0), β_h(1), β_l(0) is not defined clearly. It appears to denote counterfactual equilibria when the parameter is changed from 0 to 1, but this should be stated explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; identification is structural and the derivation does not assume its conclusions.

full rationale

The central identification claim (Proposition 2, Section 4.1) is a standard observational-equivalence argument: if two parameter vectors yield the same equilibrium, then k_i'θ = k_i'θ̃ for every player; full rank of plim (1/M)Σk_i'k_i then implies θ = θ̃. The regressors k_i include the equilibrium objects p̄_i and g_iΣg_i', but these are derived from the model's rational-expectations and quadratic-distance assumptions, not from the parameters being 'predicted.' The NPL estimation treats p* as an unknown fixed point and estimates θ from the implied likelihood; this is not a fitted parameter renamed as a prediction. The self-citation to Lambotte (2025) concerns network endogeneity only and does not carry the identification or uniqueness argument; Proposition 1's uniqueness proof is self-contained via a contraction-mapping argument. The paper also explicitly acknowledges the two main limitations: Assumption 4(ii) is imposed on functions of the unobserved equilibrium and is not directly testable, and separate identification without isolated players relies on the quadratic social-distance function (Appendix B discusses linear and aggregate alternatives). These are model-risk/robustness concerns, not circular reductions: the paper does not define β_h or β_l in terms of the quantities it later claims to identify, nor does it invoke a self-citation to force the result. The specification tests in Appendix A are in-sample model comparisons, not predictions of fitted inputs. Overall, the derivation chain is self-contained and no step reduces to its own inputs by construction.

Assumptions & free parameters 5 free parameters · 8 assumptions · 0 invented entities

All structural assumptions are standard for the incomplete-information network-game literature. The main load-bearing cost is the quadratic distance function, which does the separation work for β_h and β_l; the paper itself proves that linear distance needs isolated players, and it does not verify the presence of isolated players or the untestable part of the rank condition in the data.

free parameters (5)

β_h (conformity cost when choosing high action) = Smoking: 1.077 (SE 0.252); Alcohol: 1.772 (SE 0.198)
Main structural parameter; estimated by NPL in the heterogeneous model.
Δβ = β_l - β_h = Smoking: 2.903 (SE 0.401); Alcohol: 0.035 (SE 0.284)
Parameterized directly in θ=(γ,β_h,Δβ); measures action-specific heterogeneity.
β_l (conformity cost when choosing low action) = Smoking: 3.980 (derived as β_h+Δβ); Alcohol: 1.806
Economic parameter of interest, recovered from fitted β_h and Δβ rather than fitted separately.
γ vector (school fixed effects, own characteristics, contextual peer characteristics) = Not reported numerically in Table 1 (marked only by 'X')
Standard linear-index coefficients in α_i; estimated jointly with peer-effect parameters.
β (homogeneous conformity parameter) = Smoking: 2.392 (SE 0.163); Alcohol: 1.788 (SE 0.156)
Estimated only in the comparison homogeneous model; used to assess bias from imposing homogeneity.

assumptions (8)

domain assumption Assumption 1: ε_i(y_i) iid across actions and players, continuous, independent of α and G
Underlies the rational-expectations information structure and the product likelihood; rules out correlated unobservables.
domain assumption Assumption 2: players form rational expectations p_j = E[y_j|α,G]
Behavioral assumption required to derive the equilibrium mapping and to identify the model from conditional choice probabilities.
ad hoc to paper Assumption 3: |β_h| + 1.5|β_l-β_h| < 1/max f_η(u)
Sufficient condition for uniqueness via contraction; not necessary, chosen for tractability and to align with Lee et al. (2014).
domain assumption Assumption 4: F_η strictly increasing and rank condition on plim (1/M)Σ k_i'k_i
Standard identification assumption; the rank part is untestable because it depends on the unobserved equilibrium p.
domain assumption Quadratic social distance function (Eq. 3)
Functional form generates the g_iΣg_i' term that identifies β_h and β_l separately; linear distance functions lose identification without isolated players (Appendix B).
domain assumption Logistic distribution for ε (empirical specification)
Chooses F_η as the logistic CDF, matching Lee et al. (2014); fixes scale and shape of the private shocks.
domain assumption Linear index α_i = m_i'γ0 + x_i'γ1 + GX_i'γ2
Models observed heterogeneity as a linear index of school fixed effects, individual characteristics, and contextual peer characteristics.
domain assumption Network exogeneity and many-network asymptotic
Treats the friendship network G as exogenous and identifies parameters from variation across many independent school networks.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Heterogeneity in peer effects for binary outcomes." pith.science (2026). https://pith.science/paper/EWNLP55L

@misc{pith2026251115891,
  author       = {Pith},
  title        = {Pith review of: Heterogeneity in peer effects for binary outcomes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EWNLP55L}},
  note         = {Machine review of arXiv:2511.15891}
}

read the original abstract

I introduce heterogeneity into the analysis of peer effects arising from conformity by allowing peer-effect parameters to vary across agents' actions. Using a structural model based on a simultaneous network game with incomplete information, I derive conditions that guarantee the uniqueness of the equilibrium and the identification of heterogeneous peer-effect parameters. Applying the model to data on smoking and alcohol consumption among secondary school students, and conducting policy simulations, I show that assuming a homogeneous preference for conformity leads to biased estimates of peer effects and ex ante policy evaluations.

Figures

Figures reproduced from arXiv: 2511.15891 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 1 linked inside Pith

[1]

and Mira, P

Aguirregabiria, V. and Mira, P. (2007), ‘Sequential estimation of dynamic discrete games’,Econo- metrica75(1), 1–53. Akerlof, G. (1997), ‘Social distance and social decisions’,Econometrica65(95),

2007
[7]

and Tamer, E

Ciliberto, F. and Tamer, E. (2009), ‘Market structure and multiple equilibria in airline markets’, Econometrica77(6), 1791–1828. De Paula, A. (2013), ‘Econometric analysis of games with multiple equilibria’,Annu. Rev. Econ. 5(1), 107–131. Fehr, E. and Schmidt, K. M. (1999), ‘A theory of fairness, competition, and cooperation’,The quarterly journal of econ...

arXiv 2009
[1005]

Akerlof, G. A. (1980), ‘A theory of social custom, of which unemployment may be one conse- quence’,The quarterly journal of economics94(4), 749–775. Azmat, G. and Iriberri, N. (2010), ‘The importance of relative performance feedback information: Evidence from a natural experiment using high school students’,Journal of Public Economics 94(7-8), 435–452. Ba...

1980

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

and Mira, P

Aguirregabiria, V. and Mira, P. (2007), ‘Sequential estimation of dynamic discrete games’,Econo- metrica75(1), 1–53. Akerlof, G. (1997), ‘Social distance and social decisions’,Econometrica65(95),

2007

[2] [7]

and Tamer, E

Ciliberto, F. and Tamer, E. (2009), ‘Market structure and multiple equilibria in airline markets’, Econometrica77(6), 1791–1828. De Paula, A. (2013), ‘Econometric analysis of games with multiple equilibria’,Annu. Rev. Econ. 5(1), 107–131. Fehr, E. and Schmidt, K. M. (1999), ‘A theory of fairness, competition, and cooperation’,The quarterly journal of econ...

arXiv 2009

[3] [1005]

Akerlof, G. A. (1980), ‘A theory of social custom, of which unemployment may be one conse- quence’,The quarterly journal of economics94(4), 749–775. Azmat, G. and Iriberri, N. (2010), ‘The importance of relative performance feedback information: Evidence from a natural experiment using high school students’,Journal of Public Economics 94(7-8), 435–452. Ba...

1980