{"id":"e6ed9941-f1c7-4d2c-a7b4-6e4af1e2be47","arxiv_id":"2411.17447","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Retracted authors' collaboration networks appear more centralized than their non-retracted networks, but the analysis is confounded by network size and redundant metrics.","lead":"This study compares the co-authorship networks of retracted and non-retracted papers from 30 prolific retracted authors, reporting that retracted networks are more centralized and hierarchical. The comparison is compromised because all networks have diameter 2 and the two groups differ greatly in size, so the reported differences may be artifacts.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The retracted-vs-nonretracted comparison is confounded by corpus/network size: non-retracted networks are systematically larger (Table 1), and all networks have diameter 2, making average path length a mathematical transform of degree centrality; the reported t-tests cannot isolate retraction…","rationale":"The headline claim has two load-bearing premises: (i) that observed network differences are due to retraction status, and (ii) that the reported metrics are independent evidence of hierarchy/centralization. Premise (i) fails because corpus size is a strong confound; premise (ii) fails because diameter-2 networks make APL a deterministic function of DC. The paper's own tables demonstrate this. Therefore the central claim is not yet supported. My proposed size-matched resampling directly tests whether retraction status has any topology effect after removing the size confound. The reader's weakest assumption is the same concern, so I agree. I do not add a new objection about author selection or data availability; those reinforce the rejection but are not the single most load-bearing condition. Since the reader's verdict is REJECT and my concern supports it, no change is needed.","tokens_in":13942,"tokens_out":4542,"duration_ms":43116,"concrete_test":"Conduct a size-matched resampling test: for each author, randomly sample (without replacement) exactly the author's retracted count of papers from their non-retracted paper set, rebuild the non-retracted ego-network, and recompute the paired t-test and Cohen's d for Degree Centrality, Average Path Length, and Closeness Centrality across 100 resamples. If the significant differences disappear or flip sign, the effects in Tables 2-3 are artifacts of corpus size rather than retraction status; if they persist with comparable effect sizes, the central claim survives this specific objection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison is not identifiable because retraction status is strongly confounded with the size of the corpora from which the ego-networks are built. Table 1 shows that for nearly every author the non-retracted count is several times the retracted count (e.g., Ashok Pandey 44 vs 679; Fazlul H. Sarkar 53 vs 531). Average degree centrality, average path length, and closeness centrality all depend on the number of nodes and edges. More importantly, every network in Tables 4 and 5 has diameter 2, which implies the exact identity APL = 2 - DC for the tabulated normalized degree centrality. The rows satisfy this (e.g., author 1 retracted: 2 - 0.1496 = 1.8504), and it is why the t-statistics for Degree Centrality and Average Path Length in Table 2 are exactly opposite (-3.2052 and +3.2052). Thus the significant DC/APL/CL differences are not independent evidence of 'hierarchical and centralized structures'; they are the same signal, and that signal can be produced purely by smaller, sparser ego-networks for retracted papers. No Freeman centralization index or size control is reported, and 'hierarchical' is inferred from degree-eigenvector correlations rather than measured. The paper's limitation section does mention the small 30-author sample, but it does not address this size confound, which is the load-bearing condition for the headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript compares collaboration networks built from retracted and non-retracted publications of 30 authors. The authors report network properties such as degree centrality, average path length, assortativity, and clustering, and use t-tests and Cohen's d to claim that retracted networks are more hierarchical and centralized, while non-retracted networks show distributed collaboration and stronger clustering. The paper concludes that these structural differences could inform policies for research integrity.","tokens_in":14218,"tokens_out":8079,"duration_ms":71161,"significance":"The research question is relevant to the study of scientific misconduct and collaboration dynamics, and a credible finding that retracted co-authorship networks are topologically distinct could inform early-warning systems and research-integrity policy. The manuscript's strengths include its use of public data from Retraction Watch, CrossRef, and Scopus, and the computation of a broad set of network metrics for a relatively large set of authors. However, the central claim is not supported by the evidence as presented: the retracted and non-retracted networks differ systematically in size, several of the reported metrics are mathematically redundant, and the statistical tests ignore the paired structure of the data. No code is provided, and data are only available upon request.","major_comments":[{"comment":"Every network reported in Tables 4 and 5 has network diameter 2, and the tabulated values exactly satisfy the identity APL = 2 - DC (for example, Ali Nazari retracted: 2 - 0.1496 = 1.8504). For diameter-2 ego-networks, average path length is a deterministic function of average degree centrality, so the t-statistics for Degree Centrality and Average Path Length in Table 2 are exact opposites (-3.2052 and +3.2052). The paper therefore double-counts a single structural signal as two independent significant differences; the APL comparison should be either removed or explicitly acknowledged as redundant.","section":"§5.4, Tables 2, 4, 5"},{"comment":"The comparison between retracted and non-retracted networks is confounded by corpus and network size: Table 1 shows that for nearly every author the non-retracted publication count is several times larger (e.g., Ashok Pandey 44 vs 679; Fazlul H. Sarkar 53 vs 531). Because the constructed ego-networks all have diameter 2 (the focal author connects all co-authors), the average degree centrality and closeness values are strongly driven by the number of nodes and edges, which in turn are determined by the number of publications. The unpaired t-tests in Section 5.4 do not control for network or corpus size, so the reported significant differences in DC, APL, and closeness may be entirely a size artifact. A size-controlled analysis (matching, covariate adjustment, or permutation with fixed node counts) is essential before the headline claim can be assessed.","section":"§3.1 and Table 1 vs §5.4"},{"comment":"The t-tests treat the 30 retracted networks and 30 non-retracted networks as two independent groups, but the two networks associated with a given author are paired observations from the same researcher. An unpaired test ignores this dependence and inflates the effective sample size; a paired t-test or mixed-effects model with author as a random effect is required. The paper should also report the degrees of freedom and explicitly state which test was used.","section":"§5.4"},{"comment":"The conclusion that retracted networks are 'hierarchical and centralized' is inferred from correlations between degree centrality and eigenvector centrality, but no direct network centralization index (e.g., Freeman's centralization) is computed or tested. The observed correlations may arise from degree distributions or network size rather than from a genuine difference in centralization, and they are not accompanied by significance tests. The authors should measure centralization directly and compare it between groups after controlling for size.","section":"§5.3 and §6"}],"minor_comments":[{"comment":"The first paragraph contains a duplicated phrase, 'These findings These findings'; please fix the typo.","section":"§1"},{"comment":"The data selection explanation is confusing: six authors are excluded, then a reduced set of 20 is mentioned, then 10 additional authors are added to reach 30; please clarify the final selection process and whether any of the original 30 were replaced.","section":"§3.2"},{"comment":"The average path length formula uses ordered pairs, n(n-1); the text should clarify whether the network is treated as directed for this metric and how isolated nodes (if any) are handled.","section":"§4.1, Eq. (1)"},{"comment":"The column header for closeness centrality is 'CoC' in Table 4 and 'CL' in Table 5; please use consistent abbreviations throughout.","section":"Tables 4 and 5"},{"comment":"The reference list has inconsistencies in author names (e.g., 'Barabasi' vs 'Barabási') and duplicate entries for Barabási (2016); also, some URLs appear incomplete.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a timely topic and the authors have assembled a substantial dataset, but the central comparison is currently not identifiable because of the size confound and the redundancy between several metrics. I recommend major revision rather than rejection because these issues are in principle addressable through a size-controlled reanalysis and paired tests; the APL/DC redundancy should be fixed regardless."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline finding — retracted networks are more hierarchical and centralized than non-retracted ones — does not survive contact with the data. The comparison is confounded by the much larger size of the non-retracted corpora, and the paper's own tables show that every network has diameter 2, which forces average path length to be a simple transform of degree centrality. The t-tests for those two metrics are therefore two looks at the same signal, not independent evidence.\n\nWhat is actually new: this is a fresh application of standard ego-network metrics to retracted versus non-retracted publication sets for the same authors, using Retraction Watch and Scopus. I'll credit the data-collection effort; matching 30 authors across sources is real work. The paper also reports Cohen's d and includes all per-author metric tables, which is more transparent than much of the literature.\n\nWhere it falls short: (1) The size confound. In Table 1, non-retracted paper counts are often an order of magnitude larger (e.g., Ashok Pandey 44 vs 679). Ego-network metrics like degree centrality, average path length, and closeness centrality all scale with network size, and the authors never control for it. (2) The diameter-2 identity. In a diameter-2 graph, APL = 2 − DC for the normalized degree centrality used here. Every row in Tables 4 and 5 satisfies this, and it is why the t-statistic for APL is exactly the negation of the one for DC. Calling both \"significant structural differences\" double-counts a single effect. (3) The statistical testing is poorly documented: no summary statistics, no mention of paired vs independent t-tests, and no justification for treating the 30 authors as independent when each contributes both a retracted and a non-retracted network. (4) The sample-selection narrative is internally inconsistent: the text says six authors were excluded and later \"10 additional authors\" were added to get back to 30, but those 10 names are already in the original Table 1. That undermines confidence in the data handling.\n\nTo be fair, the descriptive part — correlations and CCDF plots — is fine as a first pass, and the limitations section acknowledges the small sample. But the central inferential claim is not supported.\n\nWho it's for: readers working on retraction dynamics may want a quick look at the descriptive metrics, but I wouldn't hang a policy or monitoring recommendation on this. It deserves a peer-review round, not a desk reject, because the question is legitimate and the data are real. A referee should send it back for major revision and require the authors to (a) control for network size or match retracted and non-retracted sets by number of publications, (b) drop or reinterpret the redundant APL/DC pair, and (c) clarify the sample selection. If those can't be fixed, the paper should stay an exploratory report, not a comparative claim.","headline":"A well-intended descriptive study whose central comparison is undone by a network-size confound and a redundant metric pair.","tokens_in":14778,"tokens_out":4127,"would_cite":false,"duration_ms":36164,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Retracted coauthorship networks are more hierarchical than non-retracted ones, a comparison of 30 authors' collaboration graphs shows.","keywords":["retractions","collaboration networks","co-authorship","network centrality","scientific misconduct","network dynamics","statistical validation","ego networks"],"falsifier":"Subsample or match the non-retracted papers to the same count and year distribution as the retracted papers for each author, then rerun the t-tests and Cohen's d; if the significant differences in degree centrality, assortativity, and path length disappear or flip, the central claim is not supported.","tokens_in":13716,"feed_emoji":"🕸️","tokens_out":6701,"duration_ms":56625,"temperature":0.7,"pith_summary":"The paper sets out to show that the co-authorship networks formed by retracted publications are structurally different from those formed by the same authors' non-retracted publications. It constructs ego-centered collaboration networks for 30 heavily-retracted researchers, one network from retracted papers and one from non-retracted papers, and compares nine network metrics between the two sets. The central finding is that retracted networks are more hierarchical and centralized, with strong correlations among centrality measures and longer average path lengths, while non-retracted networks are more distributed, with stronger clustering and connectivity. Statistical t-tests and Cohen's d effect sizes are used to argue that the differences in degree centrality, weighted degree, path length, assortativity, eigenvector centrality, and closeness centrality are unlikely to be chance variations. The authors present this as evidence that network topology can flag retraction-prone collaboration patterns, which could inform research-integrity monitoring.","feed_headline":"Retracted coauthorship networks are more hierarchical","feed_subtitle":"Retracted papers show centralized graphs; non-retracted ones cluster and spread connections.","key_machinery":"The carrying mechanism is the paired ego-network comparison: for each of the 30 authors, two collaboration networks are built—one from retracted publications and one from non-retracted publications—and nine standard network metrics are computed for each network: degree centrality, weighted degree, average path length, assortativity, transitivity, clustering coefficient, eigenvector centrality, betweenness centrality, and closeness centrality. The structural difference is then quantified by t-tests on the difference between the retracted and non-retracted metric distributions and by Cohen's d, which measures the effect size in standard-deviation units. The correlation heatmaps of these metrics supply the qualitative picture of hierarchy versus distributed clustering.","core_discovery":"The paper's central claim is that retracted and non-retracted collaboration networks of the same authors are measurably different in structure: retracted networks are organized around a few highly central, hub-like nodes, while non-retracted networks spread influence across many nodes and form more clustered sub-networks. This claim is supported by correlation analyses among metrics and by two-sample t-tests showing statistically significant differences in six of nine metrics, with large Cohen's d values for assortativity, average path length, and eigenvector centrality. The authors interpret the pattern as showing that a hierarchical, centralized collaboration structure is characteristic of retraction-prone research, whereas distributed collaboration with strong clustering is characteristic of healthier publication records.","pith_inferences":["If the size imbalance between the retracted and non-retracted corpora (e.g., 44 vs 679 papers for one author) is not controlled, the observed differences in degree centrality and path length could partly reflect network size rather than retraction status; a matched or size-normalized analysis would settle this.","The same ego-network construction could be applied to authorship records before any retraction occurs, making it possible to test whether high centralization predicts future retractions prospectively rather than retrospectively.","Breaking the analysis down by retraction reason (fabrication, plagiarism, duplication) might reveal that the structural signature is specific to certain kinds of misconduct, a distinction the current aggregate analysis does not make."],"forward_implications":["If the claim holds, monitoring metrics such as degree centrality and assortativity in a researcher's collaboration network could serve as an early-warning indicator for retraction-prone patterns.","Hierarchical, hub-dependent networks would be more fragile: the deletion of a few central nodes would disproportionately disrupt a retracted network compared with a non-retracted one.","The metrics that do not differ significantly—transitivity, clustering coefficient, and betweenness centrality—suggest that local cohesion and bridge roles are not what distinguishes retraction-prone collaborations.","Institutional research-integrity policies could use these structural signatures as a screening tool, though the paper notes external factors such as funding and policy are not included."],"supporting_citations":[{"why":"Supplies the co-authorship network construction and the analytic approach for measuring collaboration structure.","marker":"Newman, 2004"},{"why":"Provides the scale-free and hub-based view of scientific collaboration networks that underpins the interpretation of hierarchical structures.","marker":"Barabási et al., 2001"},{"why":"Justifies the use of centrality measures in co-authorship networks, which the present study applies to compare retracted and non-retracted networks.","marker":"Yan and Ding, 2009"},{"why":"Defines Cohen's d, the effect-size measure the paper uses to quantify the magnitude of metric differences beyond p-values.","marker":"Cohen, 1988"},{"why":"Prior work on how retraction affects an author's collaboration network, which this study directly extends.","marker":"Sharma and Mukherjee, 2024"}],"fun_headline_variants":["Retracted papers show hub-heavy collab networks","Collaboration style flags retraction risk","Hierarchical collabs tied to retraction","Retracted networks centralize, healthy ones cluster","Distributed collabs beat centralized for integrity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the retracted and non-retracted networks of each author are directly comparable, even though the retracted and non-retracted corpora differ greatly in size; if network size drives the metric differences, the claimed structural signature would not be specific to retraction.","fun_headline_variants_meta":{"raw":{"variants":["Retracted papers show hub-heavy collab networks","Collaboration style flags retraction risk","Hierarchical collabs tied to retraction","Retracted networks centralize, healthy ones cluster","Distributed collabs beat centralized for integrity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000439,"raw_usage":{"total_tokens":2155,"prompt_tokens":797,"completion_tokens":1358,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":413,"completion_tokens_details":{"reasoning_tokens":1291}},"tokens_in":413,"tokens_out":1358,"duration_ms":11130,"temperature":1.0,"reasoning_tokens":1291,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:06:06.504355+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Subsample or match the non-retracted papers to the same count and year distribution as the retracted papers for each author, then rerun the t-tests and Cohen's d; if the significant differences in degree centrality, assortativity, and path length disappear or flip, the central claim is not supported.","supporting_citations":[],"review_version":1}