Does a fast acceptance predict retraction for misconduct?
Written against the design standard. The author maintains retraction trackers (XeraRetractionTracker, RetractionPollution) that study post-retraction citation, reasons and geography. Acceptance speed is not among what they cover.
1. The question
A paper mill needs its product accepted, and a manipulated or absent peer review is quick. If that mechanism is real, a short time from submission to acceptance should be associated with retraction for misconduct, and not with retraction for honest error, which has no reason to depend on review speed. The second half is what makes this a test rather than a correlation.
2. Prior work, and the search behind the novelty claim
Log key P17.
corpus regex: (peer review (time|duration|speed)|time to acceptance|acceptance (time|interval)
|submission.to.acceptance).{0,80}(retract|misconduct|paper mill)
PubMed: ("time to acceptance"[tiab] OR "submission to acceptance"[tiab]
OR "peer review duration"[tiab] OR "review time"[tiab] OR "acceptance time"[tiab])
AND (retract*[tiab] OR "paper mill*"[tiab] OR misconduct[tiab])Seven PubMed records, none in the corpus. The closest is a study of publication speed and retraction among COVID-19 articles (PMID 35432503), and a study of retracted chemistry manuscripts by error versus misconduct (PMID 37692206). This design extends the speed question beyond one pandemic topic and adds the reason contrast, which is the part that tests a mechanism.
3. Hypothesis
A shorter submission-to-acceptance interval is associated with retraction for misconduct-class reasons, and not with retraction for honest-error reasons.
Null. The odds ratio per halving of the interval is 1 for misconduct-class retractions.
What would falsify it. An odds ratio at or below 1 for misconduct, or an association of the same size for honest error, which would mean speed marks something about the journal rather than about the fraud.
4. Design and unit
Matched case-control, run twice: once with misconduct-class cases, once with honest-error cases, each with its own controls. One article is one unit.
5. Sample
Cases: articles in Retraction Watch with RetractionNature equal to Retraction, a DOI and a PMID, and PubMed History dates for both received and accepted. Classified into misconduct-class (paper mill, fake peer review, fabrication, falsification, plagiarism, manipulated images) and honest-error class (error in data, analysis or results by the authors, with no misconduct reason) using Retraction Watch’s reason codes, with the mapping published before classification.
Controls: four per case, from the same journal and the same publication month, not in Retraction Watch, with both History dates. Sampled without replacement; a case with fewer than four eligible controls keeps those it has and the set is weighted (S9).
6. Exposure
Days from received to accepted, from PubMed History/PubMedPubDate with PubStatus “received” and “accepted”, using full dates (S7). Analyzed on the log scale.
7. Outcome
Case or control status, by reason class.
8. Validation gate
Reason-class mapping is checked by the author on 150 cases against the retraction notices themselves. Below 90 percent agreement on the two classes, the mapping is revised and the check repeated before any model is fitted.
9. Analysis
In R.
clogit(case ~ log2(days_to_accept) + strata(match_set), data = d)Fitted separately for each reason class; the contrast between the two coefficients is the test. Preregistered alongside: a model with the interval in categories, and one excluding special issues, where editorial handling differs.
10. Threats
| threat | handling | removed or disclosed | standard |
|---|---|---|---|
| History dates are missing for many publishers | Complete-case within matched sets; missingness reported by publisher | Disclosed | |
| Misconduct retractions cluster in a few journals | Matching on journal and month holds the journal constant | Removed | S9 |
| Reason codes are Retraction Watch’s, not the publisher’s | Mapping checked against notices on 150 cases | Disclosed | S8 |
| Retraction lag means recent fraud is not yet retracted | Cases are what has been retracted; biases toward the null | Disclosed | S10 |
| Same-day received and accepted dates are a data artifact | Flagged and analyzed separately rather than entered as zero | Disclosed | S7 |
11. Probes
| probe | computes | goes to | what would change the design | status |
|---|---|---|---|---|
| PA | proportion of retracted and of all PubMed articles with both History dates, by publisher | 5 | under 40 percent coverage among cases makes the sample unrepresentative of retractions | not run |
| PB | case counts per reason class after the date requirement | 5 | fewer than 300 honest-error cases makes the contrast imprecise | not run |
| PC | the author’s 150 reason checks | 8 | below 90 percent the mapping is revised | needs the author |
12. What this cannot show
- That fast review caused anything. Journals that review quickly may simply attract fraud.
- Anything about fraud that has not been retracted.