Retraction Watch Database Overview
- Retraction Watch Database is a comprehensive repository of retracted scholarly works, offering detailed metadata including DOIs, publication dates, and up to 111 coded retraction reasons.
- It functions as a critical link between bibliometric platforms like Web of Science, Scopus, OpenAlex, and more, enhancing cross-system analyses of retraction trends.
- Its granular, machine-readable reason coding supports nuanced studies on misconduct, error propagation, and citation impact, informing policy and scientometrics.
The Retraction Watch Database is a comprehensive, freely accessible database of retracted scholarly works maintained by Crossref and used as a cross-publisher infrastructure for identifying which works were retracted, when they were retracted, and, in many cases, why (Oppenlaender, 22 Feb 2026). In empirical research it functions both as a registry of correction events and as a linkage hub: studies use its DOIs, PubMed IDs, article metadata, dates, and reason codes to connect retracted works to Web of Science, Scopus, Microsoft Academic Graph, OpenAlex, Altmetric, Crossref Event Data, and Wikipedia revision histories (Zheng et al., 23 Jul 2025, Memon et al., 2023, Shi et al., 22 Sep 2025). Snapshot studies illustrate its scale: one 2021 snapshot reports 26,504 retracted publications across 5,844 journals and conferences (Memon et al., 2023), and a 2022 snapshot reports 33,955 total records (Shepperd et al., 2022).
1. Institutional role and dataset structure
Retraction Watch began as a Retraction Watch project and, after Crossref acquired it in 2023, formally registered retraction DOIs became integrated into the dataset, reducing the likelihood of systematic omissions for publishers that register retractions with Crossref (Oppenlaender, 22 Feb 2026). In this role, the database operates as shared integrity infrastructure rather than as a field-specific index.
Studies using the database describe a core record structure that includes article title, journal or venue, DOI when available, year of publication, year of retraction, and coded reasons for retraction (Memon et al., 2023). One large-scale study reports that each record could contain up to 104 coded reasons in the 2021 snapshot (Memon et al., 2023), while a later publisher-comparison study describes 111 distinct reason codes and four “retraction natures”: retraction, expression of concern, correction, and reinstatement (Oppenlaender, 22 Feb 2026). This combination of bibliographic, temporal, and reason metadata is what makes the database unusually useful for downstream scientometrics.
The database is also broader than conventional journal-only bibliographies. COVID-19 work built from Retraction Watch included standard journal articles, conference papers, and preprints, including records associated with Cold Spring Harbor Laboratory Press and venues with no discernible Scopus impact (Khurana et al., 2024). That breadth is one reason the database is repeatedly treated as the most comprehensive or most extensive public source for retraction research (Memon et al., 2023, Zheng et al., 23 Jul 2025).
2. Linkage workflows and interoperability
A defining technical feature of the Retraction Watch Database is linkability through persistent identifiers. In a gender-and-retractions study, 21,976 retracted Web of Science articles from 2008–2023 were manually verified and then matched to Retraction Watch by DOI to extract detailed retraction reasons that Web of Science does not systematically provide (Zheng et al., 23 Jul 2025). In that design, Retraction Watch did not define whether a paper was retracted; it supplied the richer characterization layer.
The same pattern appears in career-level and network-level work. A study of publishing careers linked 7,906 Retraction Watch papers to Microsoft Academic Graph by DOI and added 15,363 more through fuzzy title matching, yielding 23,269 of 26,504 Retraction Watch papers matched to Microsoft Academic Graph, or about 88% (Memon et al., 2023). A collaboration-network study began from the Retraction Watch Leaderboard, retrieved retracted-paper identifiers including PubMed IDs, then used those PubMed IDs as bridges into Scopus to obtain Scopus Author IDs and non-retracted publication profiles (Sharma et al., 2024).
Retraction Watch records also support cross-domain linkage outside bibliometrics. A Wikipedia study started from Retraction Watch DOIs, queried Crossref Event Data, and identified 833 retracted articles cited on 900 English Wikipedia pages, forming 1,181 retracted paper–Wikipedia page citation pairs (Shi et al., 22 Sep 2025). An evaluation of offline LLMs built a 161-paper test set of high-profile retractions directly from a Retraction Watch-derived list (Thelwall, 18 Apr 2026). These examples show that the database is not merely a catalog of notices; it is a machine-linkable event layer for broader information systems.
3. Reason coding and analytical taxonomies
Retraction Watch’s reason metadata are granular and multi-label. The database can distinguish items such as “plagiarism of text,” “fabrication of data,” “lack of IRB approval,” “fake peer review,” and “concerns/issues about data,” but downstream studies almost always collapse these codes into broader analytical schemes (Zheng et al., 23 Jul 2025, Oppenlaender, 22 Feb 2026). This is both a strength and a methodological obligation: the raw codes are rich, but most quantitative analyses require reason consolidation.
One influential mapping grouped Retraction Watch reasons into nine categories: mistakes, fabrication/falsification, duplication, plagiarism, ethical issues, authorship issues, single reason, multiple reasons, and uncategorizable or no available reason (Zheng et al., 23 Jul 2025). Another study mapped 104 coded reasons into four broad classes—misconduct, plagiarism, mistake, and other—after manual annotation of 1,250 retraction notices and label propagation (Memon et al., 2023). A publisher-comparison study consolidated 111 codes into 17 categories and then analyzed the ten most frequent (Oppenlaender, 22 Feb 2026).
These codings materially change what can be studied. In the gender-and-retractions literature, reason coding made it possible to separate honest error from misconduct-type retractions and show that gender differences were concentrated in plagiarism, authorship issues, duplication, fabrication/falsification, and ethical issues rather than in mistakes (Zheng et al., 23 Jul 2025). In publisher analysis, consolidated reasons revealed that globally the most frequent categories were Results and/or Conclusions at 39.2%, Third Party at 36.0%, Plagiarism at 30.8%, Data Concerns at 25.5%, and Peer Review Concerns at 22.7% across the ten-publisher subset (Oppenlaender, 22 Feb 2026).
A recurrent implication is that “retraction” is not a single causal class. Retraction Watch enables reason-sensitive analysis, but only after explicit decisions about mapping, aggregation, and treatment of multi-reason notices.
4. Research enabled by the database
Retraction Watch supports production-normalized and population-normalized indicators when linked to publication denominators. In one study, the database was combined with Web of Science and Leiden Ranking classifications to define a retraction rate,
and a Male/Female Retraction Ratio,
allowing field-, country-, year-, and reason-specific comparisons (Zheng et al., 23 Jul 2025). In another, Retraction Watch flags embedded in OpenAlex supported incidence models normalized by total publications and active authors, yielding evidence of exponential growth in retraction incidence with a doubling time of about 5.6 years at the paper level and 5.7 years at the author level, while absolute incidence in 2021 remained 0.12% (Venturini et al., 2 Apr 2026).
Career studies use Retraction Watch to anchor event time. One analysis linked retractions to Microsoft Academic Graph and Altmetric, then showed that high online attention around a retraction increased the probability that an author exits publishing, while early-career authors were especially vulnerable (Memon et al., 2023). Because Retraction Watch includes dates and reasons, the study could compare misconduct, plagiarism, mistake, and other retractions rather than treating all retractions as equivalent events.
Network studies use the database to identify retraction-prone actors and collaboration structures. A comparison of retracted and non-retracted collaboration networks for 30 highly retracted authors found that retracted networks were more hierarchical and centralized, while non-retracted networks showed more distributed collaboration with stronger clustering and connectivity (Sharma et al., 2024). Related work using Web of Science rather than Retraction Watch reached a different conclusion—little evidence that stigmatization strongly alters coauthorship structure—illustrating that the network consequences of retraction depend on design choices and data sources (Sharma et al., 2023).
Citation-propagation studies use Retraction Watch reason labels to examine how problematic work spreads. A Scopus-linked study on plagiarism and fake peer review found 4,924 plagiarism retractions and 6,420 fake peer review retractions after mapping Scopus records to Retraction Watch; plagiarism papers received 141,891 citations versus 55,272 for fake peer review, and the total number of retracted citations to plagiarized papers was 2,270 versus 1,343 for fake peer review (Sharmaa et al., 2 Feb 2025). In public knowledge systems, Retraction Watch metadata made it possible to show that 71.6% of Wikipedia citations to retracted papers were problematic and that the median time to correction was 1,344 days, or over 3.68 years (Shi et al., 22 Sep 2025).
The database is also indispensable for evaluating AI systems that interact with literature. Using a Retraction Watch-derived set of 161 high-profile retracted papers, an offline-LLM study found false-negative rates of 82% for GPT OSS 120B, 84% for Gemma 3 27B, and 88% for DeepSeek R1 70B, confirming that offline models do not reliably internalize retraction status and should not be treated as substitutes for authoritative retraction registries (Thelwall, 18 Apr 2026).
5. Limitations, biases, and interpretive cautions
Retraction Watch is a database of observed retractions, not a database of all misconduct or all error. Several studies explicitly state that some problematic articles never result in formal retractions, so the database captures only observed retractions and therefore only the visible part of a larger phenomenon (Zheng et al., 23 Jul 2025). A plausible implication is that measured patterns mix underlying misconduct prevalence with detection, editorial capacity, institutional willingness to investigate, and public notice practices.
Coverage is also shaped by the systems with which the database is combined. Web of Science–based studies note English-language and STEM bias, exclusion of non-indexed journals, and difficulty with some regional naming conventions (Zheng et al., 23 Jul 2025). Humanities work shows a different limitation: among 343 humanities-only retractions identified through Retraction Watch, only 218 had identifiable retraction notices, 122 cases used the same DOI for the original article and the notice, and only 24 notices had one or more citations in COCI, indicating substantial problems of notice identifiability and findability (Heibi et al., 2023).
Reason metadata are not uniformly informative. The ten-publisher study reports that IEEE alone had 9,526 retractions tagged “Notice – Limited or No Information,” making reason comparisons asymmetrical across publishers (Oppenlaender, 22 Feb 2026). Computer Science work similarly found that about 56–65% of retracted papers provided little or no information about the reasons, far above the level seen in other disciplines (Shepperd et al., 2022). These gaps mean that reason-frequency tables partly reflect notice-writing practices, not just underlying causes.
Attribution is another recurrent problem. Retraction notices rarely identify which author was responsible, so many studies use first author, corresponding author, or all authors as proxies for responsibility (Zheng et al., 23 Jul 2025, Memon et al., 2023). Gender studies add further limitations: name-based gender inference creates large “unknown” groups for some regions, especially Chinese names, and cannot capture non-binary identities (Zheng et al., 23 Jul 2025). These are not minor technical issues; they shape substantive conclusions.
6. Complementary infrastructures and emerging uses
Retraction Watch increasingly operates as one layer in a broader retraction-data ecosystem. RetraLytix, for example, treats Retraction Watch as one of its three core data pillars alongside Crossref and OpenAlex, using it as the primary source of retraction events while external systems add affiliations, subjects, and benchmarking views for countries, institutions, journals, and authors (Singh et al., 25 May 2026). This reflects a general pattern: Retraction Watch supplies retraction identity and reason metadata; other platforms supply denominators, citation graphs, or entity disambiguation.
A complementary preprint analogue has also emerged. WithdrarXiv is explicitly described as being for arXiv what the Retraction Watch Database is for the journal literature: a structured, large-scale view of withdrawn work, but focused on preprints and withdrawal comments (Rao et al., 2024). That comparison clarifies the conceptual scope of Retraction Watch: it remains the principal open registry for formal journal- and publisher-level retractions, while adjacent infrastructures are extending similar logic to preprint platforms.
Future directions repeatedly converge on machine-readable integration. The Wikipedia study argues for edit-time alerts, dashboard-based triage, citation-template integration, and bots powered by retraction metadata (Shi et al., 22 Sep 2025). Career work calls for more standardized, hierarchical reason coding and structured information about who initiated the retraction (Memon et al., 2023). Publisher analysis argues that the value of Crossref’s stewardship of Retraction Watch depends on publishers actually registering retraction DOIs and avoiding non-public “dark archive” removals that bypass shared infrastructure (Oppenlaender, 22 Feb 2026).
Taken together, these lines of work position the Retraction Watch Database less as a static list than as a foundational integrity layer for modern scholarly systems. Its core contribution is not merely to enumerate retractions, but to make retraction status linkable, reason-sensitive, and computationally actionable across bibliometrics, research assessment, literature discovery, public knowledge platforms, and AI-assisted scholarship.