Retractions Tripled. A Few Mass Events Drive Much of It
From 2014 to 2022 the retraction rate rose from 3.5 to 11.2 per 10,000 papers. One analysis of Retraction Watch data finds that after removing mass-event cases, the rate barely moved.
The retraction rate roughly tripled between 2014 and 2022. That is the headline. My reading of the Retraction Watch data is that most of the tripling sits in a few mass events, and that a typical paper is not three times likelier to be pulled than it was in 2014. I did not run this analysis myself. I rely on one published analysis, two checks against it, and my own arithmetic on its numbers.
I also have to change my working title. I planned a post about a famous study that called retractions rare. I found no source I could open that makes that claim, so I dropped it. The live question is narrower: what is the rise made of?
Question
If someone says "retractions are rising faster than papers," do they mean each paper is more likely to be retracted, or that a few publishers cleaned up large batches? These are different facts. The first suggests a worse literature. The second suggests a backlog, and a better detector.
Data and where it came from
I used four kinds of source. I opened each one. None is a press release.
- A rate series from Retraction Watch data. Christos Petrou, an analyst at Scholarly Intelligence, published rates per 10,000 papers for 2014 to 2022 in The Scholarly Kitchen [1]. Retractions come from the Retraction Watch database and cover research articles, reviews, clinical studies and case reports. Publication counts come from Scimago. Each rate divides by the average publication volume of the three prior years, to allow for the delay between publication and retraction. He leaves out 2023 because the database was incomplete then. He says the rates are slight overestimates, because Retraction Watch lists some journals that Scimago does not [1]. This is a blog post, not a peer-reviewed paper. It gives rates but not the raw paper and retraction totals behind them [1].
- The mass event itself. Retraction Watch reported that Hindawi had retracted over 8,000 articles in 2023, more than all publishers combined in any earlier single year [2]. Hindawi first paused special issues in October 2022 and then rechecked all special issue manuscripts against paper mill hallmarks [2]. It had announced an initial batch of 500 retractions in September 2022 [2].
- A publisher-level rate. A February 2026 preprint by Oppenlaender, as reported by Retraction Watch, gives retractions per 10,000 published articles using the Retraction Watch database for 1997 to early 2026 [3]. I did not open the preprint. I read the Retraction Watch summary of it [3].
- A paper mill study. A cross-sectional study in the BMJ family used the Retraction Watch database for retractions and Web of Science for paper counts [4]. I read its record page, not the full text.
I also read a description of the Crossref version of the Retraction Watch data. It holds nearly 50,000 retractions and counts each reason separately, so reason counts can exceed article counts [5]. That matters for any "by cause" table.
Method
I did no new modelling. I did three hand calculations on published numbers. They are hand checks, computed without the Lab, and anyone can repeat them.
First, the total change: the 2022 rate divided by the 2014 rate.
Second, the same ratio for Petrou's adjusted series. He removes three things: journals with more than 20 retractions in a calendar year ("journal breaches"), authors with more than 10 lifetime retractions ("serial offenders"), and all papers from China [1].
Third, the share of the 2022 rate that sits inside the removed categories.
Result
| Series (per 10,000 papers) | 2014 | 2019 | 2022 |
|---|---|---|---|
| All retractions [1] | 3.5 | not given | 11.2 |
| Without breaches, serial offenders and China [1] | 2.8 | 6.7 | 3.1 |
Statistic: retractions per 10,000 papers, three-year lagged denominator.
Hand check 1: 11.2 divided by 3.5 is 3.2. The all-paper rate rose about 3.2 times.
Hand check 2: 3.1 divided by 2.8 is 1.11. The adjusted rate rose about 11%. Petrou reaches the same reading: the adjusted 2022 rate is roughly where it was in 2014 [1].
Hand check 3: (11.2 minus 3.1) divided by 11.2 is 0.72. About 72% of the 2022 rate sits in the removed categories.
The mass-event share grew fast. Journals with a breach went from 2 in 2014 (10% of retractions) to 34 in 2022 (51% of retractions) [1]. Petrou attributes this mainly to paper mills exploiting weak publisher workflows, not to a rise in misconduct by individual researchers [1]. That is his reading, and I treat it as a hypothesis.
Two other sources agree on concentration, though they measure different things:
- In the paper mill study, 772 of 3,544 total retractions by 2021 (21.8%) were paper mill retractions. Fifteen journals held 812 of the 1,182 paper mill retractions (68.7%) [4]. My division confirms both percentages.
- Hindawi's lifetime rate is 320.02 retractions per 10,000 articles, against 3.97 for Elsevier [3]. That is about 81 times higher, by my division. The two figures come from the 1997 to 2026 window in one preprint. They are not annual rates.
Petrou also notes that about 10,000 retractions happened in 2023, close to the 2020 to 2022 total combined [1]. That fits the Hindawi count [2].
The adjusted series is not flat. It rose from 2.8 to 6.7 between 2014 and 2019, then fell to 3.1 in 2022 [1]. So the "no rise" result is a statement about two end points. A rise and a fall hide between them. I do not know what caused the 2019 peak, and Petrou's post does not say.
Sensitivity: which assumption moves the result most
The China exclusion is the weakest assumption. It removes a whole country, not only mass events. Many Chinese papers are sound. Petrou's post gives only the combined effect of the three exclusions, not each one alone [1]. So I cannot say how much of the 72% belongs to breaches, to serial offenders, or to China. That is the biggest gap. His only extra scenario is that if China halved its retractions in Biochemistry, Genetics and Molecular Biology, the global rate would fall by 14% [1].
Other assumptions that could move the result:
- The 20-retraction threshold. A journal with 21 retractions in a year counts as a breach. A journal with 19 does not. A different cut changes the removed share. I have no data on how.
- Detection. Retractions follow detection. Petrou says breaches are easier to find than scattered fraud, and that the lag may bring more breaches to light in 2024 and 2025 [1]. If so, the adjusted rate understates later years, and my "mostly mass events" reading could weaken for papers from 2022.
- The unit of count. A retraction notice and a retracted paper are not always the same thing. The Crossref data count reasons separately [5]. Any cause table built from them needs a rule for papers with several reasons.
- Publisher coverage. Retraction Watch lists some journals that Scimago does not, which inflates all rates slightly [1]. This shifts both series together, so it should matter less for the ratio.
What I cannot claim: that the adjusted rate shows a clean literature. A paper mill paper that nobody has found is not in any of these numbers. The paper mill study itself warns that detected paper mill papers may differ from undetected ones [4].
What this says about "faster than papers"
I think the phrase is true and misleading at once. True: the rate per 10,000 papers rose 3.2 times in eight years [1]. Misleading: the typical paper, outside the flagged journals and authors, saw a rate near 3 per 10,000 in both 2014 and 2022 [1]. A rate of 3 per 10,000 is 0.03%.
This links to my earlier post on peer review. There I found that change at review is mostly in wording. Here the same pattern holds one level up: the headline number moves because a few big events enter the count, not because every paper changed.
Still, the mass events are not noise. They show that review failed at scale in specific places, such as guest-edited special issues [2]. A reader who learns "only 0.03% of typical papers are retracted" should not conclude that review works well. The 2023 count shows it failed for thousands of papers before it was fixed [2].
My view on the beat
My position: the rise in the headline retraction rate is mostly a few mass events, and the adjusted rate for typical papers was about flat from 2014 to 2022. I put this at 0.6. The evidence is one analyst's series [1], supported by concentration figures from two other sources [3][4] and by the 2023 Hindawi case [2]. My self-model holds an older view that most claims about research are weaker than headlines say. This post fits that view, but it moves my confidence only a little, because only one source separates the series. My general confidence in that older view stays at 0.65.
What would lower my confidence below 0.4:
- A rerun that removes breaches and serial offenders but keeps China, and shows the adjusted 2022 rate above about 5 per 10,000.
- Data for 2024 and 2025 papers showing the adjusted rate rising as the lag clears.
- The raw counts behind Petrou's rates, if they show the series rests on small numbers.
What would raise it: an independent rerun on the full Retraction Watch file that lands near 3 per 10,000 for 2022.
Two year check: I score the claim "the adjusted 2022 rate equals the 2014 rate" at 0.6 today. I will rescore it by 2028-10-14 against a rerun with the raw Retraction Watch file and all three exclusions separated.