Beta
Key findings
- 59 of 344 preprint–published pairs (17.2%) contained at least one confirmed error; we found 85 errors in total.
- Of the 69 errors already present in the preprint, 47 (68%) were corrected in the published version, 3 (4%) were partly corrected and 19 (28%) were published unchanged.
- The published versions also introduced 16 new errors, in 10 papers; 6 of them were concentrated in one paper.
- The most common types were tables that disagree with the text (14), arithmetic and totals (12), confidence intervals that disagree with the estimate or P value (9), abstracts that disagree with the body (8), and percentages that do not match n/N (7).
- Our method mainly finds errors in sentences whose numbers changed between versions, so errors left unchanged in both versions are probably undercounted.
What counts as an error
We counted an error only when the paper contradicts itself, so that a careful reader could spot it without the raw data. Examples: a stated total that differs from the sum of its groups; a percentage that does not equal n/N; a point estimate outside its own confidence interval; a P value that cannot go with its interval; a number in the abstract that differs from the same number in the results; a table that disagrees with the text.
Changes that reflect new work were not counted as errors: reanalysis, a different cohort, longer follow-up, or different rounding, as long as the earlier text was consistent with itself.
Of the 85 errors, 81 are numeric. The other 4 are wrong labels, symbols or cross-references; they are included in all counts below.
Results
From 344 pairs, a script found 8,477 places where a sentence kept its wording but its numbers changed between versions. Of these, 4,402 were set aside automatically as known non-errors, and the remaining 4,075 were classified one by one. Most turned out to be alignment artefacts or legitimate updates; 80 were confirmed errors. Another 5 errors were found while reading the surrounding text of the same papers, giving 85 in total.
| Step | Count |
|---|---|
| Preprint–published pairs | 344 |
| Sentences whose numbers changed between versions | 8,477 |
| Set aside automatically (renumbered citations 3,596; formatting 479; PDF line numbers 174; small integers such as figure numbers 118; interval notation 35) | 4,402 |
| Classified one by one | 4,075 |
| Confirmed errors among these changes | 80 |
| Further errors found while reading the surrounding text | 5 |
| Confirmed errors in total | 85 |
Were preprint errors fixed?
69 of the 85 errors were already in the preprint. By publication, 47 had been corrected, 3 had been partly corrected and 19 were still there unchanged, so 22 of the 69 remained in the published version in some form. The other 16 errors first appeared in the published versions and were not present in the corresponding preprints.
| Outcome | Errors | Share |
|---|---|---|
| Corrected in the published version | 47 | 68% |
| Partly corrected (one instance fixed, another left) | 3 | 4% |
| Published unchanged | 19 | 28% |
Counting numeric errors only gives 65 preprint errors: 43 corrected (66%), 3 partly corrected and 19 unchanged (29%). Restricting the analysis to the 56 errors rated high confidence gives 49 preprint errors: 37 corrected (76%), 2 partly corrected and 10 unchanged (20%).
The published versions introduced 16 errors that were not in the preprint. They were spread over 10 papers, but 6 of them came from a single paper, so this number depends heavily on one case. Across the sample, about one new error was introduced for every three that were corrected.
Types of error
| Type | All errors | Preprint errors | Corrected | Partly | Unchanged |
|---|---|---|---|---|---|
| Table disagrees with text | 14 | 13 | 9 | 0 | 4 |
| Arithmetic or totals | 12 | 10 | 8 | 0 | 2 |
| Confidence interval disagrees with estimate or P value | 9 | 7 | 5 | 0 | 2 |
| Abstract disagrees with body | 8 | 6 | 5 | 1 | 0 |
| Percentage does not match n/N | 7 | 6 | 3 | 0 | 3 |
| Other | 35 | 27 | 17 | 2 | 8 |
“Other” includes the same quantity reported two ways (6), counts that do not match the items listed (5), typographical slips (3), wrong labels (3), text that disagrees with a figure legend (3), definitions that do not match the data (2), impossible values (2) and 11 single cases such as a reversed direction, a wrong sign or unit, or a percentage confused with percentage points. Each type has few errors, so differences between types should not be read as differences in how often they get fixed.
Most papers with errors had only one: 44 papers had one error, 10 had two, and 5 had three to six.
Three examples
The three examples below are rewritten. Numbers, outcomes, populations and wording have been changed so that the source papers cannot be identified. Each error was corrected in the published version.
Preprint“A total of 186 patients were enrolled.” Table 1: 97 patients at site A, 79 at site B, 20 at site C.
Published“A total of 196 patients were enrolled.” Table 1 unchanged.
97 + 79 + 20 = 196, not 186.
Preprint“Diabetes was associated with 30-day readmission (OR 1.62; 95% CI 1.71 to 2.20).” Next row of the same table: chronic kidney disease, OR 1.94 (95% CI 1.71 to 2.20).
Published“Diabetes was associated with 30-day readmission (OR 1.62; 95% CI 1.38 to 1.90).”
An estimate cannot lie outside its own interval. The interval had been copied from the next row, where 1.94 sits at its centre.
Preprint“Follow-up data were available for 412 of 437 participants (84.3%).”
Published“Follow-up data were available for 412 of 437 participants (94.3%).”
412 / 437 = 94.3%.
How this compares with earlier studies
Earlier comparisons of preprints with their published versions mostly asked whether conclusions or estimates changed, and found that they rarely change much. This study asks a different question: whether the numbers in each version agree with each other, and whether those that do not are fixed.
- Conclusions: in 72,644 bioRxiv preprints matched to journal articles, the primary claim in the abstract was unchanged in 39.8%, minorly revised in 50.0% and substantially revised in 10.2% (Yin et al., 2026, a preprint not yet peer reviewed).
- Conclusions during the pandemic: in early 2020 bioRxiv and medRxiv preprints, 7.2% of non-COVID-19 and 17.2% of COVID-19 abstracts had a discrete change in conclusions by publication, mostly without changing them qualitatively (Brierley et al., 2022).
- Estimates: across more than 1,000 epidemiological estimates in 100 COVID-19 preprints, point estimates changed by 6% on average during peer review, and intervals narrowed by 7% (Nelson et al., 2022).
- Reporting: published versions of bioRxiv preprints reported, on average, 4.7 percentage points more of the items on a quality-of-reporting checklist (Carneiro et al., 2020).
- Internal consistency of P values: in psychology journals, about half of articles that used significance testing contained at least one P value inconsistent with its test statistic and degrees of freedom (Nuijten et al., 2016). That study looked at published articles only and at one kind of number.
Methods
- Sample: 344 pairs of a preprint and its published version (medRxiv 284, bioRxiv 60). Each preprint had a CC BY or CC0 licence and its published version had full text in PubMed Central. Most published versions appeared between October 2025 and September 2026; a smaller set of short papers goes back to November 2023. The median time from preprint to publication was 217 days (interquartile range 147 to 338).
- Versions compared: the latest preprint version against the published version; when a preprint had several versions, the first against the latest as well (69 pairs had more than one preprint version).
- Finding changes: a script split both texts into sentences, kept sentences containing numbers, matched each to the most similar sentence in the other version by its words, and flagged pairs whose numbers differed.
- Filtering: changes caused by citation renumbering, formatting, PDF line numbers, small integers such as figure numbers, and interval notation were set aside automatically.
- Classification: each remaining change was classified as an error, an update, an artefact or unclear, by reading the passage in both versions. An error had to be visible from the version itself.
- Errors were identified and checked with AI models, and arithmetic was recomputed by script. A person then checked a random sample of 12 high-confidence errors, stratified by type, against the original papers; all 12 were confirmed. Ratio confidence intervals were checked on the log scale. Each error was rated high, medium or low confidence (56, 26 and 3).
Limitations
- No independent human review yet. The errors were checked with AI models, and a person has checked only a random sample of 12. No independent reviewer has yet checked all 85, and agreement between reviewers has not been measured. Six borderline items await a clinical or statistical opinion.
- Likely undercount. The method finds an error mainly when a sentence’s numbers changed between versions, or when it was noticed while reading the surrounding text. Errors present in both versions with unchanged wording, errors in rewritten passages, and errors only in figures are mostly missed.
- Tables. Preprint tables were read from the PDF text layer, where columns can merge or shift, so table errors are more likely to be missed than over-counted.
- Sample. Only open-licensed preprints whose published versions are in PubMed Central were included. Preprints that were never published, or were published elsewhere, are not represented and may have more errors.
- Cause of correction. A corrected error was not necessarily found by peer review; many were fixed during a wider reanalysis. Only a few published review reports show a reviewer pointing one out.
- Timing. Some papers were published shortly before the analysis, and later corrections are not counted.
Independent review and updates
These results are preliminary. An independent reviewer will check all 85 errors and the borderline items. We will then update the counts on this page, report agreement between reviewers, and release the list of pairs and errors. Each update will be recorded below.
- Version 0.1, 20 October 2026: preliminary results; checked with AI models, with a human spot check of 12 errors.
How to cite
MediCrab. Do preprint errors in numbers get fixed before publication? Preliminary results, version 0.1. 20 October 2026. https://www.medicrab.com/research/preprint-number-errors/
Frequently asked questions
Are preprints reliable?
This study measures one thing: whether the numbers in a paper agree with each other. By that measure, about one preprint–published pair in six had at least one error, and most preprint errors were corrected by publication, but more than a quarter were not. Earlier studies found that conclusions rarely change much between preprint and publication.
Does peer review catch numeric errors?
This study cannot say how reliably peer review detects them. It shows that 19 of the 69 errors present in preprints (28%) were published unchanged. A correction does not show who found the error: many were fixed while the authors reanalysed their data.
Why only open-access papers?
Comparing two versions in full requires both full texts. We used preprints with open licences whose published versions are in PubMed Central.
Can I see the list of papers?
Not yet. We will release the list of pairs and errors after independent review. We do not name papers on this page.
Will the numbers change?
Possibly. They are preliminary until an independent reviewer has checked them. Changes will be recorded in the version history on this page.
References
- Yin H, Ahn W, Forster PM, Rust R. Tracking claim changes from preprint to publication across 72,644 biomedical studies using large language models. bioRxiv preprint, 2026. doi:10.64898/2026.06.30.735556.
- Basu M. Think preprints are unreliable? Analysis of 70,000 studies might change your mind. Nature (news). 2026. doi:10.1038/d41586-026-02167-3.
- Brierley L, Nanni F, Polka JK, et al. Tracking changes between preprint posting and journal publication during a pandemic. PLOS Biology. 2022. doi:10.1371/journal.pbio.3001285.
- Nelson L, Ye H, Schwenn A, et al. Robustness of evidence reported in preprints during peer review. Lancet Global Health. 2022. doi:10.1016/S2214-109X(22)00368-0.
- Carneiro CFD, Queiroz VGS, Moulin TC, et al. Comparing quality of reporting between preprints and peer-reviewed articles in the biomedical literature. Research Integrity and Peer Review. 2020. doi:10.1186/s41073-020-00101-3.
- Nuijten MB, Hartgerink CHJ, van Assen MALM, Epskamp S, Wicherts JM. The prevalence of statistical reporting errors in psychology (1985–2013). Behavior Research Methods. 2016. doi:10.3758/s13428-015-0664-2.