{"id":"4db80744-c0fb-4983-922f-d096741fdaae","arxiv_id":"2511.15891","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Action-specific conformity costs for smoking (β_h≈1.08, β_l≈3.98) are identified and estimated in a binary network game; the homogeneous model is rejected for smoking but not drinking.","lead":"This paper lets the cost of peer pressure differ between choosing the high action (smoking) and the low action (not smoking), and shows both costs can be estimated separately from school friendship networks. In U.S. data, non-smokers pay a much larger conformity cost (β_l≈3.98) than smokers (β_h≈1.08), while drinking shows no such asymmetry.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Separate identification of β_h and β_l relies on the quadratic distance function; without testing this form, the empirical heterogeneity could be an artifact.","rationale":"The paper's core contribution is the separate identification of action-specific conformity parameters β_h and β_l. The identification argument (Prop. 2) is formally correct under Assumption 4, but the regressor that separates the two parameters—g_i Σ g_i'—is generated by the quadratic social distance function. The paper itself acknowledges in Appendix B that with a linear distance function, identification requires isolated players, which are rare in the Add Health data. This makes the quadratic functional form load-bearing: if the true disutility is not quadratic, the gΣg' term is a misspecification artifact, and the estimated heterogeneity (especially the large and significant Δβ for smoking) may be spurious. The paper's specification test cannot rule out this alternative because it only distinguishes spillover versus conformity, not quadratic versus linear conformity; the linear conformity model also has β3=0 in the generalized reduced form (A.3). Moreover, the rank condition in Assumption 4(ii) is on equilibrium quantities and is not verifiable before estimation; Proposition 3's sufficient conditions require an additional, untested parameter condition (γ2,κ≠0). A Monte Carlo simulation drawing from a linear-distance model on realistic networks would directly show whether the quadratic estimator produces false evidence of heterogeneity. This concern does not invalidate the theoretical contribution, but it strengthens the need for robustness checks, which the paper lacks. The reader's CONDITIONAL verdict is appropriate, and my read does not change it.","tokens_in":25785,"tokens_out":9811,"duration_ms":102604,"concrete_test":"Monte Carlo: Using the Add Health network structure (or similar) with no isolated players, simulate y from the linear-distance model (Eq. B.2) with β_h≠β_l and known α. Estimate the quadratic heterogeneous model (Eq. 7) by NPL. If the quadratic model yields estimates of β_h and β_l that differ significantly and the LR test rejects homogeneity (as in the paper's smoking result) even though the DGP is linear, then the empirical asymmetry is not evidence for the quadratic model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central identification result (Prop. 2) rests on the presence of the term 0.5 Δβ g_i Σ g_i' in the best response (Eq. 7). This term arises only because the social distance function is quadratic and action-specific (Eq. 3). If the true disutility of deviation is linear (e.g., S_lin in Appendix B.2), the best response is p∗ = F(α + (β_h+β_l)¯p − β_h 1{d>0}), which is linear in ¯p; separate identification of β_h and β_l then requires isolated players (Appendix B.2). In Add Health, isolated students are rare or absent, so the model would be unidentified. The paper's specification test (Appendix A.3) has null β3=0 for both spillover and linear conformity, and the alternative β3≠0 only picks up the nonlinear term gΣg'; it does not test the quadratic form against other nonlinearities. Hence, the reported heterogeneity (β_h=1.077, β_l=3.980) may be an artifact of the assumed quadratic distance. The rank condition (Assumption 4(ii)) is imposed on functions of the unobserved equilibrium p, so it is not directly checkable; Proposition 3's alternative conditions require an exclusion-like condition (γ2,κ≠0) that is itself untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":26177,"tokens_out":4342,"duration_ms":48242,"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":[{"comment":"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","section":"§4.1 and Appendix B, Eq. (7)"},{"comment":"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.","section":"Abstract"},{"comment":"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.","section":"§5.2 and footnote 20"},{"comment":"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.","section":"Assumption 4(ii) and Proposition 3"}],"minor_comments":[{"comment":"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.","section":"Acknowledgements"},{"comment":"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.","section":"§5.2"},{"comment":"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).","section":"§5.2"},{"comment":"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.","section":"Corollary 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and contains a useful theoretical result, but the empirical section's headline claim of action-specific heterogeneity for smoking is conditional on the quadratic distance functional form. The referee report asks for either robustness work or a clear downgrading of the empirical claims. The missing policy simulations promised in the abstract also need to be resolved. This is a fixable issue rather than a fundamental error, so major revision seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nThe citation pattern is fine; the one self-citation is to a related paper on network endogeneity and does not carry the identification.\n\nI 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.","headline":"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.","tokens_in":777,"tokens_out":1276,"would_cite":true,"duration_ms":37805,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["peer effects","conformity","binary outcomes","network game","incomplete information","identification","heterogeneity","social norms"],"falsifier":"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.","tokens_in":25679,"feed_emoji":"🚭","tokens_out":5336,"duration_ms":57423,"temperature":0.7,"pith_summary":"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.","feed_headline":"Smokers face weaker peer conformity than non-smokers","feed_subtitle":"Action-specific network model shows homogeneous peer-effect estimates mislead anti-smoking policy.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Action-specific conformity: smoking vs non-smoking peer effects","Heterogeneous conformity: smokers are less influenced than non-smokers","Conformity varies with behavior: smoking shows weaker peer effects","Anti-smoking policy gets it wrong: conformity isn't uniform","Action-specific conformity changes anti-smoking policy predictions"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Action-specific conformity: smoking vs non-smoking peer effects","Heterogeneous conformity: smokers are less influenced than non-smokers","Conformity varies with behavior: smoking shows weaker peer effects","Anti-smoking policy gets it wrong: conformity isn't uniform","Action-specific conformity changes anti-smoking policy predictions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00127,"raw_usage":{"total_tokens":4945,"prompt_tokens":567,"completion_tokens":4378,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":311,"completion_tokens_details":{"reasoning_tokens":4295}},"tokens_in":311,"tokens_out":4378,"duration_ms":31614,"temperature":1.0,"reasoning_tokens":4295,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:15:16.025800+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}