{"id":"a894e3d2-9000-4587-a5b3-4441e1c2ec50","arxiv_id":"2411.18433","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A latent space model with a reciprocity log odds ratio linear in squared latent distance can infer heterogeneous reciprocity patterns, and it finds opposite patterns in an information-sharing network and a high-school friendship network.","lead":"A new statistical model lets researchers see whether the tendency to return a directed connection in a network changes as people in a hidden social space get closer or farther apart. Applied to three real networks, it finds that reciprocity is uniform in one, stronger at distance in another, and weaker at distance in the third.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inference of φ assumes dyad-wise conditional independence (Eqs. 1–2) and that a d=2 Euclidean distance drives both edge and mutual-tie odds; unmodeled community or triadic structure can bias the reciprocity slope, and Section 4 simulations only fit the model, so the application typologies are not…","rationale":"The central modeling claim is coherent and the nesting of edge-independent models in Section 2.4 is a genuine contribution; the simulation recovery under the true model is real evidence for the inference procedure. However, the paper's empirical conclusions—different reciprocity types in three networks—depend on φ being identified under correct specification. Since the latent positions are unobserved and inferred from the same data used to estimate φ, misspecification in the dependence structure is the principal threat to the central claim. The reader's conditional verdict already captures this concern, and I agree with it. The proposed misspecification simulation is the minimal check that would settle whether the concern actually lands: if it shows no bias, the applications are much stronger; if it shows bias, the empirical typology should be downgraded until a model with additional structural terms or a robustness analysis is provided.","tokens_in":12716,"tokens_out":5926,"duration_ms":61105,"concrete_test":"Run a misspecification simulation: generate 50 networks of n=75 from a stochastic block model with two or four communities, within- and between-block edge probabilities matched to the densities in Section 5, and with dyad-independent homogeneous reciprocity (constant ρ, true φ=0). Fit the proposed d=2 model with the Section 3.1 priors and HMC settings, and record for each replicate whether the 95% posterior interval for φ excludes 0. If this occurs in more than 10% of replicates, the Section 5 typology is not robust to the dyad-independence assumption. A second variant should include triadic closure (e.g., an ERGM with mutual and triangle terms, φ=0) to test the same bias.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is dyad-wise conditional independence in Equations (1)–(2) combined with the same latent distance entering both the edge logit and the reciprocity log-odds ratio in Equation (5). If mutual-tie odds are driven partly by triadic closure, community membership, or other structure not expressible in the d=2 Euclidean latent space, the fitted coefficient φ can absorb that misspecification and produce a spurious distance-dependent reciprocity claim. The simulation study in Section 4 only generates data from the fitted model, so it cannot detect this bias. In the information-sharing application, the four office clusters visible in Figure 8c imply that between-office dyads are far in latent space; if those dyads are reciprocated for reasons unrelated to distance (task dependence, reporting lines), the model will report φ>0. The same mechanism could bias the friendship-network φ downward if unmodeled within-school subgroups reciprocate at short fitted distances.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a latent space model for binary directed networks in which the reciprocity log odds ratio for each dyad is a linear function of squared latent distance, ρ_ij = ρ + φ||z_i - z_j||², while the conditional log odds of an unreciprocated edge is s_i + r_j - ||z_i - z_j||². The model nests the edge-independent latent space model of Krivitsky et al. when ρ = φ = 0, and it nests a homogeneous-reciprocity dyad-independent model when φ = 0. Bayesian inference is carried out with Hamiltonian Monte Carlo in NumPyro. A simulation study reports decreasing estimation error as n grows, and three applications are used to conclude that the advice network has homogeneous reciprocity, the information-sharing network has reciprocity increasing with latent distance (posterior mean φ = 0.54, 95% CI 0.37–0.71), and the friendship network has reciprocity decreasing with latent distance (posterior mean φ = -1.19, 95% CI -2.12 to -0.46). The central derivation and nested structure are clear and correct.","tokens_in":12899,"tokens_out":5753,"duration_ms":55097,"significance":"If the empirical conclusions hold, the paper addresses a real gap: existing latent space models for directed networks, including AMEN, assume reciprocity is homogeneous across dyads. The proposed parameterization is simple, interpretable, and makes edge-independent LSMs a nested special case, which is a useful property for model comparison. The simulation evidence that parameters are recovered as n increases is encouraging, and the model is a natural candidate for analyzing directed network data with heterogeneous reciprocity. The main value of the paper depends on the reliability of the empirical findings and on the adequacy of the inference diagnostics, both of which need strengthening before the application-level claims can be accepted.","major_comments":[{"comment":"The local log odds ratio plots are presented as empirical verification of the inferred reciprocity type, but they are not independent of the model: the distances r̂_ij are posterior mean latent distances from the distance-dependent model being tested, and the windows are then used to aggregate the observed dyads. The plot therefore reflects the model's own latent-space reconstruction and cannot serve as a confirmatory check. Please either replace this with a measure that does not use the fitted distance-dependent model, or explicitly label the plots as descriptive summaries of the fitted model and remove the claim of empirical verification.","section":"§5.3–§5.4, Figures 8a and 10a"},{"comment":"The central identifying assumption is that dyads are conditionally independent given latent positions and sender/receiver effects, and that the same squared Euclidean distance controls both edge formation and the reciprocity odds ratio. In the information-sharing application, Figure 8c shows that the latent positions separate into office clusters; under the model, between-office dyads are necessarily far apart, so the positive φ can be interpreted as 'between-office ties are more likely to be reciprocated.' But if reciprocity between offices is driven by task interdependence or reporting structure rather than by distance, φ̂ = 0.54 will absorb that misspecification. The simulation in Section 4 draws from the fitted model and therefore cannot detect such bias. I recommend a sensitivity analysis that includes office membership or another block/covariate term and reports how φ changes; alternatively, the empirical conclusions should be stated more cautiously.","section":"§2.2–§2.3, Eqs. (1)–(5)"},{"comment":"No MCMC convergence diagnostics are reported. The paper states that all chains were run for 5,000 post-burn-in iterations (Sections 4 and 5.1), but it does not give R-hat statistics, effective sample sizes, or trace plots for the simulations or applications. For a model with n(d+2)+2 parameters and hierarchical priors, 5,000 iterations may be too few to assess mixing of the latent positions. The '95% point-wise confidence intervals' in Figures 8a and 10a also have no stated construction method. Please add diagnostics and describe how the intervals are computed.","section":"§3.2 and §4"},{"comment":"The information criteria are used to select among models whose parameter dimension grows with n, and the paper itself acknowledges that the standard assumptions are not met. For the information-sharing network, the AIC difference between the distance-dependent and homogeneous models is only 18 (2158 vs 2176), and the DIC difference is 4 (1941 vs 1945); no posterior predictive checks are shown for this network. Please report how the DIC effective number of parameters is computed, and provide posterior predictive or cross-validation comparisons for all three networks before concluding that the data support distance-dependent reciprocity.","section":"Table 3 and §5"}],"minor_comments":[{"comment":"The phrase 'distant-dependent' appears in Section 3.1 and in the captions of Figures 8 and 10; it should be 'distance-dependent'.","section":"§3.1 and Figure captions"},{"comment":"The text says 'means squared errors' where 'mean squared errors' is intended.","section":"§4"},{"comment":"The package is referred to as 'NumPyro' in the text but 'Numpyro' in one place; please use a consistent spelling.","section":"§3.2"},{"comment":"The sliding windows in Figures 8a and 10a are overlapping and only windows containing all four dyad types are plotted; this should be stated in the methods or captions so that the plotted trend is not overinterpreted.","section":"§5.3–§5.4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable fit for stat.ME and the core model formulation is sound. The main concerns are the empirical robustness of the application-level conclusions, the circularity in the 'empirical verification' plots, and the absence of MCMC diagnostics. I did not find evidence of a data or code availability statement; adding one would strengthen the paper. No concerns about the authorship or citation patterns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate modeling contribution. The parameterization ρ + φ||z_i - z_j||² is natural, the model nests the edge-independent LSM and the p1 model, and the posterior inference via HMC is clearly described. The simulation study (50 replicates per n, mixture latent positions) shows the expected decline in error and is a real check. The paper convinced me the model works when the model is true.\n\nWhat's new: no existing LSM in the cited literature lets reciprocity vary with endogenous latent distance. The paper demonstrates that edge-independent LSMs can mimic some heterogeneity but not the case where reciprocation is likely even when edge formation is unlikely; the information-sharing example is a good illustration. That distinction is the actual contribution.\n\nSoft spots, in order:\n\n- The dyad-wise conditional independence assumption (Eqs. 1–2) plus the same squared distance driving both edge and reciprocity odds is the load-bearing assumption. If mutual ties are driven by community structure or triadic closure beyond the d=2 Euclidean space, φ will absorb that misspecification. The simulation study is performed entirely under the fitted model, so it cannot detect this. This is the main reason to treat the application-level typologies (advice homogeneous, info-sharing φ>0, friendship φ<0) as suggestive rather than definitive.\n- The \"empirical verification\" plots (Figures 8a and 10a) use posterior-mean latent positions and sliding windows; that's a fitted-value check, not independent validation. The paper does not overclaim, but the wording \"empirically verified\" is stronger than what was done.\n- No MCMC convergence diagnostics are reported (5000 iterations after 2500 burn-in is thin for a model with n(d+2)+2 parameters). The information criteria are non-standard for LSMs with growing parameters; the posterior predictive check partly compensates.\n- I would have liked code/data artifacts. The paper names NumPyro but ships nothing.\n\nNone of these is fatal. The math checks out (Equation 4 from the dyad exponential family; Equation 5 is the right log odds ratio), the nesting is clean, and the identifiability issue for latent positions is handled by Procrustes matching. Citation pattern is standard and appropriate.\n\nWho it's for: statisticians working on latent space network models, and applied researchers who want reciprocity heterogeneity in directed networks. It deserves a serious referee. Recommendation: send to peer review; ask for code, convergence diagnostics, and at least one robustness check that does not rely on fitted latent positions (e.g., posterior predictive distribution of local reciprocity, or a split-sample validation).","headline":"A clean latent-space model that makes reciprocity a function of latent distance and nests existing LSMs, with sound math and simulations; the empirical typologies are plausible but the dyadic-independence assumption needs robustness work.","tokens_in":13461,"tokens_out":1908,"would_cite":true,"duration_ms":18288,"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":"Reciprocity in a directed network need not be uniform: the paper models the log odds of a mutual tie as a linear function of the squared distance between actors in a latent social space, and builds Bayesian inference for the slope…","keywords":["Bayesian Inference","Hamiltonian Monte Carlo","Latent space network model","Reciprocity","Directed networks","Distance-dependent reciprocity","Social networks","p1 model"],"falsifier":"Simulate directed networks from a data-generating process with no distance-dependent reciprocity but with a strong triangle-closing mechanism or latent community structure, fit the proposed model, and check whether the posterior of $\\phi$ is centered away from zero; if it is, the inferred distance-dependent reciprocity can be an artifact of misspecification rather than a genuine property of the process.","tokens_in":12481,"feed_emoji":"🔄","tokens_out":8123,"duration_ms":58906,"temperature":0.7,"pith_summary":"This paper asks whether the tendency for directed relationships to be reciprocated is the same everywhere in a network, and argues it is not: reciprocity can be stronger or weaker between actors who sit close together in a latent social space. The authors propose a dyad-independent latent space model in which the log odds ratio measuring reciprocity for a pair of actors is $\\rho_{ij} = \\rho + \\phi\\,\\lVert z_i - z_j\\rVert^2$, so a single slope parameter $\\phi$ controls how reciprocity changes with squared latent distance. Existing edge-independent latent space models are nested in this model as the special case $\\rho = \\phi = 0$, which makes model comparison meaningful. Using Hamiltonian Monte Carlo, the paper estimates the model and reports that three real networks show three distinct patterns: homogeneous reciprocity in a lawyer advice network, reciprocity increasing with distance in an employee information-sharing network, and reciprocity decreasing with distance in a high school friendship network.","feed_headline":"Reciprocity varies with latent distance in directed networks","feed_subtitle":"The model separates no, homogeneous, rising, and falling reciprocity, and finds one of each in three real networks.","key_machinery":"The load-bearing object is the dyad-independent exponential family built on the $p_1$ model, with dyad probability written in terms of natural parameters $\\mu_{ij} = s_i + r_j - \\lVert z_i - z_j\\rVert^2$ and $\\rho_{ij} = \\rho + \\phi\\lVert z_i - z_j\\rVert^2$. The first parameter controls the log odds of an unreciprocated edge and carries the standard latent space distance effect; the second is the log odds ratio that measures reciprocity for the dyad, and it is this parameter that the paper allows to vary linearly with squared latent distance. The same sender ($s_i$) and receiver ($r_i$) effects plus latent positions drive both quantities, so the model adds only two scalar parameters ($\\rho$ and $\\phi$) over the edge-independent latent space model. Hamiltonian Monte Carlo with hierarchical priors provides posterior inference, and Procrustes matching resolves the rotational and translational invariance of the latent positions.","core_discovery":"The central claim is that reciprocity is an edge-level property that can depend on actors' latent similarity, not a single global number. The paper captures this with the log odds ratio $\\rho_{ij} = \\rho + \\phi\\lVert z_i - z_j\\rVert^2$, where $z_i, z_j$ are latent positions and $\\lVert z_i - z_j\\rVert^2$ is their squared Euclidean distance. A negative $\\phi$ means mutual ties are less likely at larger distances relative to unreciprocated ties; a positive $\\phi$ means the opposite; $\\phi=0$ recovers homogeneous reciprocity; and $\\rho=\\phi=0$ recovers an edge-independent latent space model. The paper further shows that the posterior of $\\phi$ can separate these regimes: for the advice network the homogeneous model is preferred, while the information-sharing network has posterior mean $\\hat\\phi = 0.54$ with 95% credible interval $(0.37, 0.71)$, and the friendship network has posterior mean $\\hat\\phi = -1.19$ with 95% credible interval $(-2.12, -0.46)$.","pith_inferences":["A natural extension would be to let the reciprocity slope $\\phi$ vary with dyad-level covariates or latent cluster membership, which would test whether reciprocity differences track observable group boundaries rather than continuous distance.","If the distance-dependent reciprocity signal is real, it suggests substantive hypotheses about social mechanisms: cross-office information-sharing ties may be reciprocated because long-distance communication is more formalized, while close friendships may be reciprocated because of homophily.","The model's use of the same squared Euclidean distance for both edge formation and reciprocity could be tested against a two-distance-scale version; comparing the two would reveal whether reciprocity responds to similarity differently from edge formation.","Because the inference is Hamiltonian Monte Carlo and the parameter count grows with $n$, practical scaling to large networks would likely require variational methods, as the paper itself notes."],"forward_implications":["Because the edge-independent latent space model is nested at $\\rho = \\phi = 0$, standard information criteria can distinguish networks with no reciprocity, homogeneous reciprocity, and distance-dependent reciprocity.","The model can detect networks where reciprocation is likely even when edge formation is unlikely, a pattern edge-independent latent space models cannot represent.","In the three real applications the inferred reciprocity types are qualitatively distinct, supporting the paper's claim that distance-dependent reciprocity is a measurable feature of directed social networks.","The simulation study shows consistent recovery of sender and receiver effects, reciprocity parameters, and latent positions as the number of nodes grows from 50 to 250."],"supporting_citations":[{"why":"Defines the p1 dyad-independent exponential family that the proposed model generalizes; supplies the parametrization of reciprocity as a log odds ratio.","marker":"Holland and Leinhardt (1981)"},{"why":"Introduces latent space network models and the Euclidean distance mechanism for edge formation; also provides the Procrustes matching approach for posterior identifiability.","marker":"Hoff et al. (2002)"},{"why":"Provides the edge-independent latent space model that is nested in the proposed model and is used for comparison and initialization.","marker":"Krivitsky et al. (2009)"},{"why":"The AMEN model, the main existing latent space model with homogeneous reciprocity, which the proposed model contrasts with.","marker":"Hoff (2021)"},{"why":"The No-U-Turn Sampler variant of Hamiltonian Monte Carlo used for posterior inference.","marker":"Hoffman and Gelman (2014)"},{"why":"Sparse Bernoulli model with reciprocation that motivates the offset structure used for distance-dependent reciprocity.","marker":"Krivitsky and Kolaczyk (2015)"}],"fun_headline_variants":["Reciprocity depends on latent distance in directed networks","Latent distance explains non-uniform reciprocity in networks","New model: reciprocity varies with actor similarity","Friendship, advice, info networks show distance-based reciprocity","Latent distances reveal who reciprocates in directed networks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes that, given latent positions and sender and receiver effects, the dyads are conditionally independent and that the single squared Euclidean distance $\\lVert z_i - z_j\\rVert^2$ simultaneously governs edge formation and the reciprocity log odds ratio; if unmodeled structure such as triadic closure or community membership beyond Euclidean distance drives mutual ties, the posterior of $\\phi$ is biased.","fun_headline_variants_meta":{"raw":{"variants":["Reciprocity depends on latent distance in directed networks","Latent distance explains non-uniform reciprocity in networks","New model: reciprocity varies with actor similarity","Friendship, advice, info networks show distance-based reciprocity","Latent distances reveal who reciprocates in directed networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000434,"raw_usage":{"total_tokens":2193,"prompt_tokens":909,"completion_tokens":1284,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":1209}},"tokens_in":525,"tokens_out":1284,"duration_ms":8882,"temperature":1.0,"reasoning_tokens":1209,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:13:04.518663+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate directed networks from a data-generating process with no distance-dependent reciprocity but with a strong triangle-closing mechanism or latent community structure, fit the proposed model, and check whether the posterior of $\\phi$ is centered away from zero; if it is, the inferred distance-dependent reciprocity can be an artifact of misspecification rather than a genuine property of the process.","supporting_citations":[{"cited_title":"N., Handcock, M","cited_arxiv_id":null,"evidence_quote":"Provides the edge-independent latent space model that is nested in the proposed model and is used for comparison and initialization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The AMEN model, the main existing latent space model with homogeneous reciprocity, which the proposed model contrasts with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The No-U-Turn Sampler variant of Hamiltonian Monte Carlo used for posterior inference."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sparse Bernoulli model with reciprocation that motivates the offset structure used for distance-dependent reciprocity."}],"review_version":1}