{"id":"3e182088-918b-4209-a171-f477aa7b3424","arxiv_id":"2507.16689","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper proposes and analyzes pooled and equally weighted tetrad logit estimators for ordered dyadic outcomes with category-specific fixed effects, proving consistency of the pooled version under weaker sparsity conditions.","lead":"This paper develops estimators for ordered logit models where connections between people depend on unobserved sender and receiver tendencies, and the method removes those tendencies by comparing groups of four nodes. The result is a practical way to estimate friendship or alliance formation from ordinal data even in sparse networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's consistency proof uses an unsupported moment bound: it asserts Σ E[ℓ̃²] = O(q_N p_N) after only proving E[ℓ²] = O(1), so the O(1/(N p_N)) variance rate and Assumption 6's sufficiency are not established.","rationale":"The reader's verdict identified the additive threshold structure as the weakest assumption and judged Theorem 2 technically sound. My stress-test found a more immediate, internal problem: the proof of Theorem 2 relies on a moment bound that is not derived. The displayed bound on E[ℓ²] is O(1) per tetrad, but the subsequent variance calculation requires the stronger O(p_N) bound. Without it, the claimed rate of convergence of the normalized objective is not obtained, and Assumption 6 may be insufficient as stated. This does not disprove the theorem; a refined argument or an added regularity condition (e.g., a uniform bound on E[ℓ²|S=1]) could restore it. Therefore the verdict remains CONDITIONAL, but the condition should be to repair or restate the consistency proof, not only to address the empirical and inferential caveats the reader listed.","tokens_in":28378,"tokens_out":40873,"duration_ms":418028,"concrete_test":"Take a single-cutoff (M=1) version of the model with a scalar covariate X such that P(X=±R)=1/(2R²) and P(X=0)=1−1/R² for growing R, and choose fixed effects so the informative-tetrad probability p_N ~ 1/√N. Compute E[1{S=1}‖r‖²] directly; if this quantity is Θ(1) rather than O(p_N), the proof's step fails. Then simulate PTLE for N up to 5000 and check whether the scaled score variance decays at rate 1/(N p_N) or stalls, which would indicate that the theorem requires an additional moment or tail regularity condition.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In the proof of Theorem 2 (Appendix A.2), the variance calculation for the normalized objective requires Σ_σ E[ℓ̃_{mσ}²] = O(q_N p_{mN}). The only bound supplied is E[ℓ_{mσ}²] ≤ E[(log 2 + 2‖r_σ‖‖β‖)²] = O(1) per tetrad, which gives Σ_σ E[ℓ̃_{mσ}²] = O(q_N), not O(q_N p_{mN}). The sharper O(q_N p_{mN}) bound is essential: with it the numerator variance is O(N³ q_N p_N) and the scaled objective converges at rate O(1/(N p_N)); without it the variance rate becomes O(1/(N p_N²)), which does not vanish under Assumption 6 when p_N ~ 1/√N (N p_N → ∞ but N p_N² → 1). Thus Theorem 2's conclusion that PTLE is consistent under Assumption 6 is not proven by the supplied argument, and may fail if informative tetrads are rare but have large covariate realizations. This is an internal gap in the central consistency claim, distinct from the reader's concern about the additive threshold structure.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops estimation methods for ordered logit models of directed dyadic/network data with category-specific sender and receiver fixed effects. The authors extend tetrad-differencing conditional maximum likelihood from binary network models to ordered outcomes by binarizing at each threshold. They propose two estimators: ETLE, which weights each threshold equally, and PTLE, which pools informative tetrad-threshold pairs. Theorem 1 derives a conditional probability that eliminates the fixed effects under an additive threshold structure. Theorem 2 claims consistency of ETLE under Assumption 5 (sufficient information at each threshold) and of PTLE under the weaker Assumption 6 (sufficient pooled information). The paper includes Monte Carlo simulations, an empirical application to friendship networks, and extensions to more restrictive fixed-effects structures.","tokens_in":28570,"tokens_out":14047,"duration_ms":141610,"significance":"If the results hold, the paper offers a practical and novel solution to a relevant problem: estimating ordered network formation models with flexible fixed effects under sparsity and rare outcome categories. The PTLE's consistency under weaker pooled-information conditions is a useful theoretical contribution, and the simulation evidence supports the practical preference for PTLE. The empirical application is illustrative, though based on a small network. However, the main consistency proof has a technical gap, and the inference results are explicitly conjectural; these issues currently limit the strength of the paper's claims.","major_comments":[{"comment":"The proof asserts E[(Σ_{m,σ} ℓ̃_{mσ})²] = O(N³ q_N p_N) after only deriving the per-tetrad bound E[ℓ̃²_{mσ}] = O(1). From the displayed inequalities and this per-tetrad bound, the best that follows is O(N³ M q_N) = O(N³ q_N), not O(N³ q_N p_N). The sharper order is essential: with only O(N³ q_N), the variance of the scaled PTLE objective is O(1/(N p_N²)), which need not vanish under Assumption 6 when p_N → 0 (e.g., p_N ~ 1/√N). The same issue affects the ETLE proof in part (a), where the analogous assertion for each cutoff m would require E[ℓ̃²_{mσ}] = O(p_{mN}). The authors should either prove a sharper per-tetrad second-moment bound under the stated assumptions (for instance, using bounded covariates or a fourth-moment condition with an additional argument) or revise the assumptions and proof so that the claimed variance rate follows.","section":"Appendix A.2, proof of Theorem 2"},{"comment":"The asymptotic normality of PTLE and the consistency of the sandwich variance estimator (Eq. 18) are presented only as a conjecture, with 'suitable regularity conditions' left unspecified. Despite this, the empirical application in Section 6 reports standard errors and significance levels based on Eq. (18), and Section 5.3 evaluates its coverage properties. Without a formal theorem establishing the asymptotic distribution, the inferential claims in the empirical section are not rigorously justified. The authors should either provide a complete asymptotic normality result with explicit conditions (e.g., the sixth-moment condition mentioned) or clearly label the standard errors as heuristic and temper the conclusions drawn from them.","section":"Section 4.2"}],"minor_comments":[{"comment":"The sandwich estimator Υ_N sums over all ordered dyads (i,j) of v_{ij} v_{ij}', where each tetrad contributes to four different dyad-level sums. The paper does not explain whether or how this double-counting is accounted for in the sandwich formula, which would be helpful for readers implementing the method.","section":"Section 4.2 and Section 4.3"},{"comment":"The computation requires enumerating all q_N = O(N^4) tetrads, which becomes prohibitive for networks much larger than the N=32 example. The authors should discuss computational feasibility or possible subsampling schemes for larger networks.","section":"Section 4.3"},{"comment":"In the heterogeneous-threshold simulation, the thresholds are defined using node-specific components λ_{im} = δ_{im} (imposing symmetry between sender and receiver effects), but the main model allows λ_{im} and δ_{jm} to differ. The text should clarify that the simulation only covers a symmetric special case and does not explore asymmetric sender/receiver effects.","section":"Section 5.2"},{"comment":"The standard errors for the Binary (3) column appear misaligned; for instance, the Common gender row shows five parenthetical values for six coefficient columns. The table should be reformatted for clarity.","section":"Table 14"},{"comment":"The abstract states that 'standard methods without fixed effects produce counterintuitive results,' but the ordered logit without fixed effects in Table 14 yields positive coefficients for all three homophily variables, which are not counterintuitive. The claim should be reconciled with the reported results or clarified.","section":"Abstract and Section 6.2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a useful problem and the core idea is sound, but the proof of Theorem 2 is incomplete in a way that affects the paper's main theoretical contribution. The moment-bound gap is likely repairable either by imposing a bounded-covariate assumption or by a more careful variance calculation using the indicator S_{mσ}; I would encourage the authors to provide a rigorous fix. The conjectural nature of the inference results is also a concern for a journal publication, particularly because the empirical significance stars rely on the unproven sandwich estimator. With these gaps addressed, the paper could be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper is worth engaging. It extends Charbonneau/Jochmans tetrad differencing from binary to ordered outcomes with category-specific sender and receiver fixed effects, and introduces ETLE/PTLE. The PTLE consistency result under pooled information (Assumption 6) is the genuinely new theoretical piece, and the distinction between requiring information at each threshold vs across thresholds is practically important. The sufficiency results (Theorems 1, 3, 4) are clean algebra and check out. Simulations match the theory; PTLE's robustness to sparse categories is convincing.\n\nI checked the second-pass stress test about the moment bound in Theorem 2. It does not hold up. The contribution lmσ includes the indicator Smσ, so its second moment is O(p_mN), not merely O(1); summing over tetrads and cutoffs gives O(q_N p_N), and the claimed variance rate O(1/(N p_N)) follows. The consistency proof is fine on this point.\n\nSoft spots, in order of importance. First, Section 4.2 is a conjecture, not a theorem. The asymptotic normality of PTLE and the consistency of the sandwich variance estimator are asserted with a hand-wave to Jochmans. For a methods paper whose headline contribution is an estimator, that is a real gap. The simulation evidence on coverage is encouraging, but it is not a proof.\n\nSecond, the abstract says 'standard methods without fixed effects produce counterintuitive results.' Table 14 shows the ordered logit without fixed effects already has positive, significant homophily coefficients. What looks counterintuitive is Binary(3) flipping the sign on common program. That's a different statement, and the current sentence overclaims.\n\nThird, treating 'unknown person' as missing is ad hoc. If missingness relates to relationship quality, the estimates are potentially selected. A sentence or two on the assumption would help.\n\nFourth, the additive threshold structure λ*_ijm = λ_im + δ_jm is load-bearing. Any dyad-specific threshold interaction breaks the cancellation. That is inherent to the approach; it should be stated more prominently as a limitation.\n\nWho is this for: empirical network researchers with ordered relational data (friendship, trade, alliances) and econometricians working on network fixed effects. It deserves a serious referee. I would conditionally accept: require the authors to either prove the asymptotic distribution or label it clearly as a conjecture, correct the abstract, and discuss the missing-data assumption.","headline":"Solid, useful extension of tetrad-differencing CML to ordered dyadic outcomes; PTLE is a real contribution, but the inference section is conjectural and the abstract overstates the empirical contrast.","tokens_in":29136,"tokens_out":2740,"would_cite":true,"duration_ms":30312,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62P20","62F12"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that a pooled tetrad logit estimator consistently estimates $\\beta_0$ in ordered dyadic models with category-specific sender and receiver fixed effects, under only pooled identification across outcome categories.","keywords":["ordered logit","dyadic data","network formation","fixed effects","incidental parameters","tetrad differencing","conditional maximum likelihood","homophily"],"falsifier":"Simulate the ordered dyadic model with $\\lambda^*_{ijm}=\\lambda_{im}+\\delta_{jm}+\\rho_{ijm}$ for nonzero dyad-specific $\\rho_{ijm}$; if the PTLE estimates do not track $\\beta_0$ as $N$ grows, the sufficiency theorem is specific to the additive class. In the additive class, a design with $Np_N\\to\\infty$ but $Np_{mN}\\not\\to\\infty$ for every threshold $m$ should show PTLE centering on $\\beta_0$ while ETLE drifts.","tokens_in":28117,"feed_emoji":"🕸️","tokens_out":8860,"duration_ms":85877,"temperature":0.7,"pith_summary":"This paper shows that ordered relationship data between pairs of agents can be estimated without estimating the many sender and receiver effects as parameters. The device is to binarize every outcome threshold and compare two senders against two receivers in a four-node tetrad; for a given threshold, the conditional probability of a contrasting tetrad pattern is logistic and free of all fixed effects. The paper proposes two estimators that aggregate these tetrad contributions across thresholds: an equally weighted version (ETLE) and a pooled version (PTLE). The central result is that PTLE is consistent under the weak condition that the thresholds together supply enough identifying information, even when no single outcome category does. This matters because networks are often sparse in high categories such as \"best friendship,\" where equal weighting and single-cutoff binary methods become unstable.","feed_headline":"Pooled tetrad logit stays consistent when outcomes are sparse","feed_subtitle":"Equal-weighting fails when rare categories have few informative tetrads; pooling them does not.","key_machinery":"The machinery is tetrad-differencing conditional maximum likelihood. For each threshold $m$, define $D_{ij}(m)=1\\{Y_{ij}\\ge m\\}$ and form the tetrad statistic $Z_\\sigma(m)=\\tfrac12\\big((D_{i_1j_1}(m)-D_{i_1j_2}(m))-(D_{i_2j_1}(m)-D_{i_2j_2}(m))\\big)$. Tetrads with $Z_\\sigma(m)=\\pm 1$ are informative, and their conditional probability is $\\Lambda(r_\\sigma'\\beta_0)$, where $r_\\sigma=(X_{i_1j_1}-X_{i_1j_2})-(X_{i_2j_1}-X_{i_2j_2})$. The additive decomposition $\\lambda^*_{ijm}=\\lambda_{im}+\\delta_{jm}$ makes the fixed effects cancel in the tetrad odds ratio, and cancellation requires the same cutoff $m$ on all four dyads. PTLE pools every informative tetrad-threshold pair into one binary logit objective; ETLE instead normalizes each threshold's contribution by its own number of informative tetrads, which is what makes ETLE vulnerable to category-specific sparsity.","core_discovery":"The core claim is that the incidental-parameter problem in ordered network logit models can be solved by extending tetrad-differencing conditional maximum likelihood from binary to ordered outcomes. After writing $D_{ij}(m)=1\\{Y_{ij}\\ge m\\}$ and imposing the additive threshold structure $\\lambda^*_{ijm}=\\lambda_{im}+\\delta_{jm}$, conditioning on informative tetrads gives $P(Z_\\sigma(m)=1\\mid Z_\\sigma(m)\\in\\{-1,1\\}, X_\\sigma)=\\Lambda(r_\\sigma'\\beta_0)$, with no fixed effects appearing. The pooled tetrad logit estimator, which maximizes the sum of these conditional log-likelihood contributions over all informative tetrad-threshold pairs, is proved to converge in probability to $\\beta_0$ under Assumptions 1 through 4 and Assumption 6, requiring only that the pooled information $Np_N$ diverge and that the pooled Hessian have full rank. By contrast, the equally weighted estimator requires Assumption 5, namely that each cutoff individually provide enough information, which fails when some outcome categories are rare.","pith_inferences":["An extension the paper leaves implicit is that the same tetrad-pooling logic should transfer to other ordered link-strength scales, such as alliance levels or rating data, whenever the additive threshold decomposition holds.","A natural refinement the paper does not derive is an information-weighted pooling scheme that reweights tetrad-cutoff pairs by their precision, which could improve finite-sample efficiency beyond PTLE's implicit weighting by number of informative tetrads.","If empirically relevant thresholds contain dyad-specific interactions, the sufficiency result fails, so applied users should treat the additive decomposition as a substantive assumption rather than a normalization.","Because PTLE only requires pooled information across categories, it opens the door to distribution-regression or semi-parametric versions in which coefficients vary by threshold, a direction the paper mentions only through its discussion of concurrent work."],"forward_implications":["Researchers can estimate ordered network formation models without estimating the $2NM$ incidental sender and receiver effects, so sparse networks and rare outcome categories no longer force a binary cutoff choice.","The pooled estimator PTLE should be preferred over the equally weighted ETLE in applications, since PTLE remains consistent even when identification comes only from pooling information across thresholds.","Single-cutoff binary tetrad estimators are inefficient and highly sensitive to which cutoff is selected; pooling all informative tetrad-cutoff pairs avoids discarding information from the rest of the outcome distribution.","The dyad-clustered sandwich variance estimator corrects the severe under-coverage that naive standard errors produce in both dense and sparse networks, making the method usable for inference.","In the Dutch friendship application, the fixed-effects-adjusted PTLE finds significant positive homophily in gender, smoking, and academic program, whereas methods without fixed effects can produce counterintuitive negative signs."],"supporting_citations":[{"why":"Supplies the binary tetrad-differencing conditional likelihood and the consistency proof template that Theorem 2 adapts to ordered outcomes.","marker":"Jochmans (2018)"},{"why":"Underpins the multiple-fixed-effects binary response technique whose differencing logic is extended to category-varying thresholds.","marker":"Charbonneau (2017)"},{"why":"Provides the network formation model with degree heterogeneity that motivates treating sender and receiver effects as incidental parameters.","marker":"Graham (2017)"},{"why":"Gives the consistent fixed-effects ordered logit estimator in panel data, the ordered-choice baseline this paper generalizes to dyadic dependence.","marker":"Baetschmann et al. (2015)"},{"why":"Shows how threshold differences are identified in fixed-effects ordered logit, informing the alternative specifications in Section 7.","marker":"Muris (2017)"},{"why":"Supplies the Dutch student friendship dataset used in the empirical application.","marker":"Van Duijn et al. (2009)"}],"fun_headline_variants":["Pooled tetrad logit stays consistent under sparse ordered networks","Ordered dyadic data: pooling tetrads beats equal-weighting","Consistent estimation for ordered network models with rare categories","New CML estimator handles sparse ordered dyadic outcomes","Tetrad-differencing extended to ordered outcomes for sparse networks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every dyad's threshold is additively separable, $\\lambda^*_{ijm}=\\lambda_{im}+\\delta_{jm}$, with no dyad-specific interaction: if real thresholds contain terms such as $\\rho_{ijm}$ that depend on the specific pair, the fixed effects no longer cancel in the tetrad odds ratio and the estimator can be inconsistent.","fun_headline_variants_meta":{"raw":{"variants":["Pooled tetrad logit stays consistent under sparse ordered networks","Ordered dyadic data: pooling tetrads beats equal-weighting","Consistent estimation for ordered network models with rare categories","New CML estimator handles sparse ordered dyadic outcomes","Tetrad-differencing extended to ordered outcomes for sparse networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1679,"prompt_tokens":983,"completion_tokens":696,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":612}},"tokens_in":599,"tokens_out":696,"duration_ms":6855,"temperature":1.0,"reasoning_tokens":612,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:04:31.481203+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the ordered dyadic model with $\\lambda^*_{ijm}=\\lambda_{im}+\\delta_{jm}+\\rho_{ijm}$ for nonzero dyad-specific $\\rho_{ijm}$; if the PTLE estimates do not track $\\beta_0$ as $N$ grows, the sufficiency theorem is specific to the additive class. In the additive class, a design with $Np_N\\to\\infty$ but $Np_{mN}\\not\\to\\infty$ for every threshold $m$ should show PTLE centering on $\\beta_0$ while ETLE drifts.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the binary tetrad-differencing conditional likelihood and the consistency proof template that Theorem 2 adapts to ordered outcomes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Underpins the multiple-fixed-effects binary response technique whose differencing logic is extended to category-varying thresholds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the network formation model with degree heterogeneity that motivates treating sender and receiver effects as incidental parameters."},{"cited_title":"E., & Winkelmann, R","cited_arxiv_id":null,"evidence_quote":"Gives the consistent fixed-effects ordered logit estimator in panel data, the ordered-choice baseline this paper generalizes to dyadic dependence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows how threshold differences are identified in fixed-effects ordered logit, informing the alternative specifications in Section 7."}],"review_version":1}