{"id":"5c67d309-2b2f-49be-a7d3-619e48bee78c","arxiv_id":"2501.00473","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Citing a retracted paper is associated with a growing citation deficit that is larger for indirect citations and in lower-impact journals, a pattern the authors call 'attention escape'.","lead":"This paper quantifies the 'harm' of retracted papers as a citation deficit: papers that cite retracted work receive fewer citations than comparable papers. The deficit grows over time, is larger for indirect citations, and is strongest for papers in journals with impact factor below 10.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The distance-gradient result may be an artifact of comparator contamination: C4–C6 papers comprise a large share of their own same-venue/field comparator sets, biasing baseline citations downward and inflating 'harm'.","rationale":"The reader rejected the paper on causal-overreach grounds; my read supports that verdict but with a sharper internal-validity problem. Comparator contamination can manufacture the central distance gradient even if retracted papers have no causal effect, so the 'attention escape' claim fails not only because of unmeasured confounders but because the baseline for comparison is itself contaminated by the treatment. I credit the authors for providing public code and data, and the descriptive apparatus is coherent, but the load-bearing headline mechanism is not established. The paper could be rescued by redefining comparators, adding propensity-score matching or negative controls, and reframing the contribution as a correlational pattern. Because the concern directly undermines the headline mechanism, the existing REJECT verdict stands; no change in verdict is needed.","tokens_in":33146,"tokens_out":4638,"duration_ms":50940,"concrete_test":"Recompute Figures 3b–e and Extended Data Table 3 after redefining the comparator as D'(c) = same venue, year ±1, overlapping field, excluding any paper in the union of C1 through C6 (or, more conservatively, any paper with a citation path to a retracted paper of length ≤ 6). If the median harm for C4–C6 drops to near C1–C2 levels or the monotonic gradient disappears, the attention-escape mechanism is not supported. As a secondary check, run a placebo analysis using random non-retracted source papers matched on publication year, field, and venue size; if a similar distance gradient appears, the metric is measuring base-rate citation differences rather than harm caused by retraction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central 'attention escape' claim rests on the monotonic increase of median harm with citation distance (Figs. 3b–e; Extended Data Table 3) and, secondarily, on the temporal delay (Fig. 2). The harm metric (Eq. 4) compares each paper to a comparator D(c) defined as same venue, publication year ±1, and overlapping field (Eq. 1). D(c) is never restricted to exclude papers that are themselves in the C_k chain. Given the reported sizes—C4 = 8.32M, C5 = 10.73M, C6 = 12.50M against a final matched population of 29.75M—a C6 paper's comparator will, in many venues and fields, be dominated by other papers that indirectly cite retracted work. If those papers have depressed citations for reasons unrelated to the specific retracted source (e.g., quality, topic, author team), the comparator mean is biased downward and the harm score is inflated. This inflation grows with k, because C_k occupies an increasing share of the population, so the main 'harm grows with distance' pattern can be generated even under the null hypothesis of no causal effect. The same contamination can distort the IF comparison if low-IF venues have higher C_k penetration. This is an internal-validity problem distinct from the external confounders (author team, article quality, topic popularity) emphasized by the reader, because it can manufacture the headline pattern without any effect of retracted papers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a citation-based framework to quantify the \"harm\" caused by retracted papers. Using Semantic Scholar and Retraction Watch data, the authors construct citation chains C1–C6 (papers directly or indirectly citing retracted work) and define harm as 1 minus the ratio of a citing paper's citation count to the average citation count of a comparator group defined by the same venue, publication year ± 1, and overlapping field of study (Eqs. 1–5). The main empirical claims are that harm to citing papers grows over time, increases with citation distance up to six steps, and is larger for papers published in journals with impact factor below 10; these patterns are interpreted as an \"attention escape\" mechanism through which retracted papers inflict damage that evades stakeholder attention. The manuscript also reports analyses of harm before versus after retraction.","tokens_in":33372,"tokens_out":3209,"duration_ms":35412,"significance":"If the central causal claim were supported, this would be an important contribution to the study of retracted research, providing quantitative evidence on how retracted papers affect the citation performance of citing papers and suggesting policy implications for tracking indirect citations. The study uses a very large corpus, and the authors make data and code publicly available, which is commendable. However, the validity of the central claim hinges on the harm metric being an unbiased measure of the effect of retracted papers, and the current design does not establish that. The comparator-based deficit is susceptible to a serious contamination bias that can generate the reported distance gradient even under a null hypothesis of no effect, so the significance of the findings is conditional on major re-analysis.","major_comments":[{"comment":"This is the central load-bearing issue.","section":"Methods 4.2, Eq. (1) and Fig. 3 / Extended Data Table 3"},{"comment":"","section":"Eqs. (4)–(5) and Sections 2.2–2.3, Discussion"},{"comment":"","section":"Section 2.4 and Extended Data Table 6"}],"minor_comments":[{"comment":"There is a typo in \"Figrue 5\" that should be corrected to \"Figure 5\".","section":"Section 2.4"},{"comment":"The text contains \"citing prc\" and \"paperc\" where \"paper prc\" or \"paper c\" is intended; please revise for clarity.","section":"Methods 4.2"},{"comment":"The column headers list \"Y8 Harm (IQR)\" twice, with the first occurrence likely meant to be \"Y9\"; also, Table 3 and Table 4 headers should be checked for consistency.","section":"Extended Data Tables 3 and 4"},{"comment":"The term \"harm\" is used both as a metric (a real-valued ratio) and as a normative concept; the paper would benefit from explicitly distinguishing the operational definition from the interpretive label, and from noting that negative harm values are possible.","section":"Throughout"},{"comment":"The figures report median harm and interquartile ranges, but no confidence intervals or significance tests are provided for the comparison across citation distances; adding such measures would make the claims more robust.","section":"Results 2.2 and Fig. 3"}],"recommendation":"reject","confidential_remarks":"The paper addresses an important topic and uses a large dataset with public code, but the central empirical claim is undermined by a contamination bias in the comparator definition that can manufacture the headline distance gradient. I see no way to accept the manuscript in its current form. The authors might be able to rescue the paper with a substantially revised analysis that excludes chain members from the comparator, adds proper controls, and either fixes the causal identification or reframes the claims as associational; however, that would be a major new analysis, not a local revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece here is the relative citation deficit metric applied across six layers of indirect citation, on a very large Semantic Scholar sample. Prior work mostly counted citations to retracted papers; this paper asks whether papers that cite them underperform same-venue, same-field peers, and finds a pattern that grows with citation distance. That is a real descriptive contribution, and shipping code plus data on GitHub is a plus.\n\nThe main soft spot is the causal language. Equation (4) defines harm as a citation deficit, and the paper then talks about retracted papers \"inflicting\" harm and an \"attention escape\" mechanism. The design is observational; papers that cite retracted work differ from non-citers in many ways (topic, author team, quality, open access), and none of those are controlled. So the causal interpretation is not justified. The paper would be much stronger reframed as a correlational pattern with clearly stated confounder limitations.\n\nI checked the stress-test concern about comparator contamination. It does not hold up. If C4–C6 papers are lower-cited for unrelated reasons, then the comparator for a C6 paper contains many such papers, so the comparator mean is also pulled down. Since the paper's own citations are drawn from the same contaminated conditional distribution, the bias is toward zero, not away from it. The distance gradient could be attenuated by contamination, not manufactured by it. So the headline result is more credible than the stress-test suggests.\n\nA separate methodological worry is the temporal alignment of the comparator. The comparator includes papers published in yc-1 and yc+1, and yearly citations are aligned by year-since-publication rather than calendar year. That can create spurious time trends in the \"harm grows over time\" claim. Worth asking the authors to address.\n\nAlso, claims like \"first direct quantitative evidence of destructive mechanisms\" overreach. The novelty is in the apparatus, not in demonstrating mechanism.\n\nWho is this for? Researchers in research integrity and bibliometrics who want a large-scale, openly reproducible map of citation deficits around retracted papers. The descriptive findings are interesting enough to deserve a serious referee, but the paper needs heavy revision: drop or substantially soften the causal framing, add confounder controls or a natural experiment, and fix the temporal alignment. I would send it to peer review rather than desk reject.","headline":"A large-scale descriptive study of citation deficits around retracted papers; the distance gradient survives the comparator-contamination worry, but the causal 'harm' framing is not supported by the design.","tokens_in":34000,"tokens_out":3778,"would_cite":true,"duration_ms":43383,"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":"The paper claims that retracted papers reduce the citation performance of papers that cite them through an 'attention escape' mechanism: harm is delayed, intensifies with indirect citation distance, and is concentrated in journals with…","keywords":["retracted papers","attention escape","citation harm","indirect citations","research integrity","impact factor","scientometrics","citation networks"],"falsifier":"Take the set of papers that cite a retracted work and, for each one, construct a control paper from the same journal, year, and field with similar author-team size, topic keywords, open-access status, and reference-list length; if the citation shortfall disappears or reverses under this matching, the reported harm is an artifact of confounding rather than an effect of the retracted citation.","tokens_in":32870,"feed_emoji":"📉","tokens_out":8256,"duration_ms":72150,"temperature":0.7,"pith_summary":"This paper tries to establish that retracted papers measurably reduce the citation performance of papers that cite them, through a mechanism it calls 'attention escape.' Defining harm as a citing paper's citation shortfall relative to comparable papers from the same journal, year, and field, it finds the shortfall is small in the first years after publication but grows to a median of about 40 percent by the tenth year for direct citers. The shortfall is larger for papers that cite those direct citers, and it grows as the citation distance increases. The effect is strongest for papers in journals with an impact factor below 10, and it persists even after retraction. If correct, this means current alerting tools that focus on direct citations are aimed at the wrong place, because the worst harm happens outside the attention of authors and publishers.","feed_headline":"Retracted research costs citing papers up to 68% of citations","feed_subtitle":"Harm is delayed, grows with citation distance, and hides in lower-impact journals.","key_machinery":"The machinery is a citation-based harm metric. For each citing paper, harm is defined as $1$ minus the ratio of the paper's citation count to the average citation count of a comparator group, where the comparator group consists of papers from the same venue, published in the year before, the same year, or the year after the citing paper, and sharing at least one field of study. The metric is computed for total citations and for each of the ten years after publication. The paper then builds six citation-distance layers by iteratively expanding to papers that cite the previous layer, and compares median harm across layers, time, and impact-factor bands. The comparator-group ratio is what turns 'harm' into a measurable shortfall.","core_discovery":"The central claim is that retracted papers propagate harm through citation networks by an 'attention escape' mechanism: their most damaging effects occur where authors and publishers are least likely to notice. The paper reports three signature patterns in citation data. First, papers that directly cite a retracted work show no significant citation shortfall in their first four years, but the shortfall then grows steadily, reaching a median of about 40 percent by the tenth year. Second, papers connected only indirectly, through up to six citation steps, suffer larger shortfalls than direct citers, and the severity increases with each extra step. Third, the shortfall is concentrated in journals with impact factor below 10; papers in journals with impact factor above 20 may even show a surplus in early years. The paper measures harm as the ratio of a paper's citations to the average of papers in the same venue, a one-year publication window, and an overlapping field, and it concludes that retraction alone does not stop the damage because the harm escapes the attention these interventions rely on.","pith_inferences":["Editorial extension: the causal interpretation would be strengthened or weakened by matching citing papers to controls with similar author-team size, topic keywords, open-access status, and reference-list length; the paper does not perform this matching, so the reported effect sizes are upper bounds if confounding exists.","Editorial extension: the attention-escape mechanism predicts that high-visibility retractions, such as prominent fraud cases, should produce a different harm curve from quiet retractions; comparing those two groups would test the mechanism more directly.","Editorial extension: the same metric could be used as a pre-publication screening score that warns authors when their manuscript's reference chain reaches a retracted paper, but the paper does not assess how often such warnings would be false alarms."],"forward_implications":["Retraction-alerting tools that flag only direct citations will miss the majority of the harm, because indirect citations carry larger citation shortfalls than direct ones.","Promotion and evaluation criteria that emphasize recent publications will systematically overlook the delayed harm that retracted citations cause, matching the paper's finding that harm grows after the early years.","Journals with impact factor below 10 are the highest-priority target for retroactive citation screening, since their papers absorb the largest citation losses.","Post-retraction intervention is needed in addition to the retraction notice itself, because citing papers continue to experience significant harm after the retraction date.","The harm metric can be applied to any individual retracted article to estimate its citation cost to the papers that follow and to rank retractions by damage."],"supporting_citations":[{"why":"Supplies the large scholarly citation database from which all citing papers and comparator groups are drawn.","marker":"[35]"},{"why":"Supplies the set of retracted papers that seed the citation chains and define the treatment of interest.","marker":"[7]"},{"why":"Supplies journal-level metadata used to assign papers to impact-factor bands for the analysis.","marker":"[42]"},{"why":"Provides the six-degrees framework that sets the maximum citation distance explored in the harm propagation analysis.","marker":"[39]"},{"why":"Provides trial evidence that alerting authors to retracted citations produces limited corrective action, the intervention the paper argues is aimed at the wrong layer.","marker":"[30]"},{"why":"Establishes the chain-retraction concept that motivates studying indirect citation propagation and that this paper extends quantitatively.","marker":"[24]"},{"why":"Earlier estimate of the epistemic cost of retracted papers to meta-analyses, complementing the paper's citation-based harm measure.","marker":"[21]"},{"why":"Earlier quantification of financial costs of research misconduct, broadening the context for the paper's citation-harm framing.","marker":"[22]"}],"fun_headline_variants":["Retracted papers harm indirect citations more than direct ones","Delayed citation harm from retractions escapes author attention","Retraction damage persists through citation networks, low-impact journals","How retracted research evades detection and keeps costing citations","Citation harm from retractions grows with distance from source"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The study assumes the citation shortfall it measures is caused by the fact that the paper cites a retracted work, rather than by other traits of the citing paper such as topic popularity, author team, article quality, or open-access status, and it applies causal language without controlling for those traits.","fun_headline_variants_meta":{"raw":{"variants":["Retracted papers harm indirect citations more than direct ones","Delayed citation harm from retractions escapes author attention","Retraction damage persists through citation networks, low-impact journals","How retracted research evades detection and keeps costing citations","Citation harm from retractions grows with distance from source"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000276,"raw_usage":{"total_tokens":1620,"prompt_tokens":890,"completion_tokens":730,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":652}},"tokens_in":506,"tokens_out":730,"duration_ms":7666,"temperature":1.0,"reasoning_tokens":652,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:49:57.929211+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the set of papers that cite a retracted work and, for each one, construct a control paper from the same journal, year, and field with similar author-team size, topic keywords, open-access status, and reference-list length; if the citation shortfall disappears or reverses under this matching, the reported harm is an artifact of confounding rather than an effect of the retracted citation.","supporting_citations":[{"cited_title":"Accessed: 2024-11-22","cited_arxiv_id":null,"evidence_quote":"Supplies the set of retracted papers that seed the citation chains and define the treatment of interest."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies journal-level metadata used to assign papers to impact-factor bands for the analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the six-degrees framework that sets the maximum citation distance explored in the harm propagation analysis."},{"cited_title":"Accountability in Research 31(1), 14–37 (2024)","cited_arxiv_id":null,"evidence_quote":"Provides trial evidence that alerting authors to retracted citations produces limited corrective action, the intervention the paper argues is aimed at the wrong layer."},{"cited_title":"Nature 632(8027), 977–979 (2024)","cited_arxiv_id":null,"evidence_quote":"Establishes the chain-retraction concept that motivates studying indirect citation propagation and that this paper extends quantitatively."},{"cited_title":"Accountability in Research 29(7), 442–459 (2022)","cited_arxiv_id":null,"evidence_quote":"Earlier estimate of the epistemic cost of retracted papers to meta-analyses, complementing the paper's citation-based harm measure."},{"cited_title":"Elife 3, 02956 (2014)","cited_arxiv_id":null,"evidence_quote":"Earlier quantification of financial costs of research misconduct, broadening the context for the paper's citation-harm framing."}],"review_version":1}