A retraction is not only a paper-level event. It is also a map event.
Every country listed in a retraction record is a small coordinate in the geography of scientific correction. But that map is tricky. A country tag does not mean “this country caused the problem.” It usually means at least one author affiliation was linked to that country. A China-United States paper, for example, counts as both China-linked and United States-linked. That makes the country column less like a passport stamp and more like a collaboration fingerprint.
Using the uploaded Retraction Watch CSV, I analyzed 70,589 records with valid publication and notice dates. For the country analysis, I focused mainly on 56,472 non-conference records, because conference-proceedings batches strongly distort timing patterns. I excluded records tagged as conference abstracts/papers or with conference-like journal titles.
When a paper listed multiple countries, I used an exploded-country approach: one China-United States paper counts once for China and once for the United States. That produced 73,172 non-conference country-paper occurrences.
The central result:
Country-specific retraction patterns are real, but they are mostly explained by publication ecosystems: subject mix, journal clusters, publisher pipelines, multinational collaboration patterns, and reason categories. Country is the visible flag; the machinery underneath is journal, publisher, subject, and failure mode.
1. The first map: countries differ strongly in retraction timing
Among countries with at least 500 non-conference country-paper occurrences, the median time from publication to retraction notice varies widely.
Japan has the longest median lag, 4.60 years, followed by France, Russia, Italy, the United States, Canada, the United Kingdom and Germany. Countries such as China, Pakistan, Saudi Arabia, South Korea, Turkey, Ethiopia and India have shorter medians.
Non-conference records only. Countries shown are among the largest by retraction-record count.
Calculated from the uploaded Retraction Watch CSV. Multi-country papers are counted once for each listed country.
A Kruskal-Wallis test across countries with at least 500 records confirmed that country-associated lag distributions differ strongly: H = 2493.7, p < 1e-300. That is statistically thunderous.
But significance is not explanation. The country label bundles together subject mix, publishers, journals, collaboration patterns and reason types. The real question is: what kind of retraction ecosystem is each country attached to?
2. Multinational versus single-country papers: the raw story is misleading
At first glance, multinational papers seem slower in the full dataset:
| Dataset | Single-country median | Multinational median | Test |
|---|---|---|---|
| All dated records | 1.29 years | 1.70 years | Mann-Whitney p = 4.0e-146, Cliff’s delta = 0.153 |
| Non-conference records | 1.71 years | 1.77 years | Mann-Whitney p = 0.088, Cliff’s delta = 0.011 |
Once conference records are removed, the difference almost disappears. In journal-like records, multinational status alone is not a strong raw predictor of retraction delay.
Even more interesting: after adjustment for country presence, notice year, society-linked status, broad subject, broad reason tags, and top journal or publisher buckets, multinational papers were associated with shorter, not longer, retraction lag. In the most adjusted journal-bucket model, multinational status was associated with about 12.9% shorter log-lag. In a logistic model for very late retraction, defined as more than 10 years after publication, multinational papers had lower odds: OR = 0.52, p = 2.8e-24.
That does not mean international collaboration protects papers from long-lag problems. It means multinational records in this database are often concentrated in recent, publisher-detected clusters and fast correction pathways. The country count is not the cause; it is a shadow cast by the publication ecosystem.
3. Multinational share has changed over time
The proportion of multinational records among non-conference retractions has not been stable. It was modest through much of the 2000s and 2010s, dipped during some batch-retraction years, then rose sharply in the most recent years of the uploaded database.
Share of non-conference records listing more than one country. The year 2026 is partial.
Calculated from the uploaded Retraction Watch CSV.
The 2023 spike in total retractions was not especially multinational: only 15.9% of non-conference records listed more than one country. But 2024, 2025 and partial 2026 show much higher multinational shares, 26.2%, 31.1% and 33.5%.
That suggests a shift in the correction landscape. Recent corrections include more internationally networked papers, or at least more records with multinational affiliation footprints.
4. The country map has two axes: multinational share and retraction lag
Some countries in the dataset are mostly single-country retraction ecosystems. Others are overwhelmingly multinational.
Saudi Arabia, Pakistan, Malaysia and Ethiopia have very high multinational shares, above 80%. China and Russia have low multinational shares, about 13%. Japan also has a relatively low multinational share but a long median lag. France, Canada, Australia and the United Kingdom have high multinational shares and moderate to long lags.
Each point is a country with at least 500 non-conference country-paper occurrences.
Calculated from the uploaded Retraction Watch CSV.
This plot punctures a simple assumption: multinational does not automatically mean slow. Pakistan, Saudi Arabia, Malaysia and Ethiopia are highly multinational but have short median lags. Japan is less multinational but much slower. France is both highly multinational and slow.
The explanation is not collaboration size alone. It is which collaboration networks are attached to which journals and reasons.
5. Country-specific reason signatures are sharp
Reason tags were grouped into broad themes: paper mill/peer-review/AI, image concerns, fraud/misconduct, plagiarism/duplication/copyright, and data/results/method concerns. Categories overlap, so percentages do not add to 100.
Selected countries. Reason categories overlap, so percentages do not sum to 100.
Calculated from the uploaded Retraction Watch CSV.
The differences are not cosmetic. Chi-square tests with Benjamini-Hochberg correction showed strong reason enrichment patterns.
Examples:
| Country | Strong enrichment signal | Approx. odds ratio versus rest |
|---|---|---|
| Japan | Fraud/misconduct | OR ≈ 8.46 |
| Russia | Plagiarism/duplication/copyright | OR ≈ 7.76 |
| Ethiopia | Paper mill/peer-review/AI | OR ≈ 6.39 |
| China | Paper mill/peer-review/AI | OR ≈ 5.72 |
| United States | Fraud/misconduct | OR ≈ 3.60 |
| Germany | Fraud/misconduct | OR ≈ 3.31 |
| United States | Image concerns | OR ≈ 2.15 |
| Italy | Plagiarism/duplication/copyright | OR ≈ 2.18 |
| Pakistan | Paper mill/peer-review/AI | OR ≈ 2.06 |
| India | Paper mill/peer-review/AI | OR ≈ 1.62 |
This is the first real explanatory layer. China, India, Pakistan, Saudi Arabia and Ethiopia have large paper-mill or peer-review-process signatures. Japan and the United States show stronger fraud/misconduct and image/data signals. Russia is dominated by plagiarism/duplication/copyright. Italy has a strong plagiarism plus image profile.
Different countries in the retraction database are not merely faster or slower. They fail through different channels.
6. Country pairs: not all collaborations have the same correction clock
The most common country-pair co-occurrence was China + United States, with 902 records, median lag 2.31 years. But other large pairs, such as India + Saudi Arabia, Pakistan + Saudi Arabia, China + Pakistan and China + Saudi Arabia, have much shorter medians, around 1.5 to 1.6 years.
Pairs are co-occurrences in multi-country non-conference records. A paper with three countries contributes to three pair counts.
Calculated from the uploaded Retraction Watch CSV.
Pair-level Mann-Whitney tests compared each country pair with all other multinational records, with FDR correction. Several pairs had significantly shorter lags:
| Faster-than-background pair | Records | Median lag | Cliff’s delta |
|---|---|---|---|
| Pakistan + United States | 105 | 0.93 years | -0.375 |
| Saudi Arabia + United Kingdom | 73 | 1.24 years | -0.293 |
| Jordan + Saudi Arabia | 106 | 1.39 years | -0.252 |
| Pakistan + United Kingdom | 83 | 1.20 years | -0.236 |
| Ethiopia + Saudi Arabia | 119 | 1.42 years | -0.235 |
| China + South Korea | 323 | 1.37 years | -0.188 |
And several pairs had significantly longer lags:
| Slower-than-background pair | Records | Median lag | Pattern |
|---|---|---|---|
| France + Saudi Arabia | 101 | 7.17 years | Very long-lag biomedical cluster |
| Japan + United States | 197 | 5.61 years | Image/fraud-heavy biomedical profile |
| Italy + United States | 236 | 3.84 years | Image/plagiarism-heavy life-science profile |
| Spain + United States | 123 | 3.00 years | Longer biomedical/data profile |
| South Korea + United States | 116 | 2.88 years | Mixed but slower |
| United Kingdom + United States | 293 | 2.26 years | Medicine/biomedicine-heavy, long tail |
| China + United States | 902 | 2.31 years | Mixed, more image and biology than China-only clusters |
The China-United States pair is especially important because it is the largest pair and does not resemble the rapid paper-mill/peer-review clusters that dominate some other China-linked records. It has a higher image-concern share, 38.6%, and a much higher society-linked share, 15.2%, than China’s overall country profile.
7. Country-pair trends have surged recently
Many large multinational pair clusters are recent. The 2020-2026 era dominates for China-Pakistan, China-Saudi Arabia, India-Saudi Arabia, Pakistan-Saudi Arabia and China-India. China-United States was already present earlier, but it also increased sharply after 2020.
Selected country pairs, counted as co-occurrences in non-conference multinational records.
Calculated from the uploaded Retraction Watch CSV. The 2020-2026 era includes partial 2026.
This is one of the strongest time-specific signals in the dataset.
Older multinational retraction clusters often involve the United States, United Kingdom, Canada, Germany, Italy and Japan. Newer multinational clusters increasingly involve China, India, Pakistan, Saudi Arabia, Ethiopia, Egypt and other countries in large publisher-audit or paper-mill-linked networks.
Again, this is not a national guilt map. It is a map of how publication pipelines globalized.
8. Exact country combinations sharpen the story
Pair co-occurrence is generous: a five-country paper contributes ten country pairs. Exact country sets are stricter.
The largest exact multinational country set is China;United States, with 668 records, median lag 2.50 years, image concerns 42.8%, fraud/misconduct 15.7%, and plagiarism/duplication 37.3%.
That is very different from exact China;South Korea, with 237 records, median lag 1.30 years, paper-mill/peer-review/AI tags 86.9%, and image concerns only 4.2%.
Other exact combinations:
| Exact country set | Records | Median lag | Main signature |
|---|---|---|---|
| China;United States | 668 | 2.50 years | Image/data/fraud, mixed biomedicine |
| China;South Korea | 237 | 1.30 years | Peer-review/paper-mill-heavy |
| China;Pakistan | 203 | 1.68 years | Peer-review/paper-mill-heavy |
| Ethiopia;India | 200 | 1.57 years | Very high peer-review/paper-mill signal |
| Egypt;Saudi Arabia | 189 | 2.51 years | Mixed, image and plagiarism |
| Italy;United States | 132 | 6.84 years | Slow image/plagiarism-heavy profile |
| Japan;United States | 124 | 6.02 years | Slow image/fraud profile |
| Canada;United States | 118 | 2.89 years | Biomedical/data long-tail profile |
This exact-combination view shows that the same country can participate in different retraction worlds. China + United States behaves unlike China + Pakistan. United States + Japan behaves unlike United States + Pakistan. Pair identity matters because it captures networks, journals, subjects and institutions better than single-country labels.
9. Subject effects: countries do not retract in the same disciplinary universe
Subject mix is a major confounder.
China-linked retractions are spread across biology/life sciences, medicine and physical sciences/engineering, but with a strong paper-mill/peer-review signal. India-linked records are heavily physical sciences/engineering. Russia-linked records are unusually social-science-heavy and plagiarism-heavy. Japan-linked records are medicine and biomedical-heavy and long-lag. France is strongly biology/life-science-heavy and long-lag.
Country-level subject summaries show the pattern:
| Country | Strong subject signal | Median lag interpretation |
|---|---|---|
| China | Physical/engineering, biomedical, computing-heavy clusters | Shorter, many publisher-audit and paper-mill/peer-review records |
| India | Physical/engineering and computing clusters | Shorter to moderate, process-heavy |
| Russia | Social sciences and humanities-heavy | Plagiarism/duplication signature, longer median but low after-10-year share |
| Japan | Medicine and biomedical-heavy | Long median and long tail |
| United States | Biology/life sciences and medicine-heavy | Image, fraud and long-tail clusters |
| France | Biology/life sciences-heavy | Long median, fewer process-heavy records |
| Saudi Arabia/Pakistan/Ethiopia | Highly multinational, many physical/engineering and publisher-audit clusters | Short median, low after-10-year share |
This is why country comparisons must be subject-aware. A country whose retracted records come from computational special issues will look different from a country whose retracted records come from decades-old cancer biology, anesthesiology, or molecular medicine.
10. Publisher effects: countries travel through different publishing pipelines
The largest country-publisher clusters are deeply asymmetric. China + Hindawi alone has 9,960 records, median lag 1.28 years, and 98.4% paper-mill/peer-review/AI tags. China + IEEE is largely conference-driven and therefore excluded here, but China still dominates several non-conference publisher clusters.
Non-conference country-paper occurrences. Counts are database records, not rates relative to total output.
Calculated from the uploaded Retraction Watch CSV.
This plot explains a lot of the country pattern.
China looks fast partly because China-linked records are heavily concentrated in fast publisher-audit clusters: Hindawi, Springer, IOS Press/Sage, SAGE, and several journal families with near-total paper-mill/peer-review tagging.
But China is not uniformly fast. China + Spandidos has a median lag of 5.64 years, with high image and plagiarism/duplication signals. China + PLoS has a median lag of 3.92 years and a mixed image/data profile.
Likewise, the United States does not have one pattern. United States + Elsevier has a median lag of 1.33 years, while United States-linked Journal of Biological Chemistry and PLoS One clusters have much longer medians.
So publisher explains country signal, but not completely. The same country changes character when it moves through a different publisher pipeline.
11. Journal effects: country-journal clusters are the real engines
Country-journal clusters are even more revealing than country-publisher clusters.
The largest clusters are dominated by China-linked papers in specific journals with high paper-mill/peer-review tags:
| Country-journal cluster | Records | Median lag | Paper-mill/peer-review/AI |
|---|---|---|---|
| China, Computational and Mathematical Methods in Medicine | 997 | 1.24 years | 99.6% |
| China, Journal of Healthcare Engineering | 982 | 1.62 years | 99.7% |
| China, Journal of Intelligent & Fuzzy Systems | 974 | 1.90 years | 94.8% |
| China, Computational Intelligence and Neuroscience | 888 | 1.22 years | 99.9% |
| China, Security and Communication Networks | 857 | 1.36 years | 99.4% |
| China, Arabian Journal of Geosciences | 767 | 0.30 years | 100.0% |
| India, Journal of Intelligent & Fuzzy Systems | 429 | 1.56 years | 99.8% |
| United Kingdom, Cochrane Database of Systematic Reviews | 321 | 8.90 years | 0.3% |
| China, PLoS One | 655 | 4.01 years | Mixed image/data profile |
| India, Soft Computing | 303 | 3.54 years | 99.0% |
This is a crucial finding:
Countries do not retract. Journal-country pipelines retract.
China in Arabian Journal of Geosciences behaves very differently from China in PLoS One. India in Journal of Intelligent & Fuzzy Systems behaves differently from India in Elsevier biomedical journals. The United Kingdom in Cochrane reviews behaves nothing like the United Kingdom in ordinary research-article clusters.
12. Society versus non-society country patterns
Using a conservative high-confidence society-linked publisher classification, society-linked records are unevenly distributed across countries.
Among large country groups:
| Country | Society-linked share of non-conference records |
|---|---|
| United States | 27.7% |
| Japan | 23.9% |
| South Korea | 21.2% |
| Russia | 17.6% |
| Italy | 16.4% |
| France | 16.3% |
| Canada | 15.5% |
| Spain | 15.3% |
| China | 5.1% |
| India | 6.1% |
| Saudi Arabia | 3.3% |
| Pakistan | 4.0% |
| Ethiopia | 1.0% |
| Malaysia | 1.7% |
This partially explains the long-tail geography. The United States and Japan are more represented in society-linked biomedical and life-science journals, where image/data and misconduct investigations often take longer. China, India, Saudi Arabia and Pakistan are more represented in non-society/unclassified publisher-audit clusters, where paper-mill or peer-review problems can be corrected in batches.
But society status is not enough. In adjusted models, country effects persisted even after controlling for society-linked status, subject, reasons, notice year and top journal/publisher buckets. That means society versus non-society is one ingredient, not the whole recipe.
13. Adjusted models: country signal persists, but shrinks into ecosystem signal
I fitted robust OLS models using log(1 + lag years) as the outcome. These models included country-presence indicators for the largest countries, multinational status, notice year, society-linked status, broad subject flags and broad reason flags. I then added publisher buckets and journal buckets.
The model R² increased as publication ecosystem terms were added:
| Model | Controls added | R² |
|---|---|---|
| Base | Country + year + society + subject + reasons | 0.229 |
| Publisher model | Base + top publisher buckets | 0.258 |
| Journal model | Base + top journal buckets | 0.302 |
Adding journal effects explains substantially more variation, confirming that journal pipelines are central.
In the journal-adjusted model, several country effects persisted:
| Country presence | Approx. adjusted effect on log-lag | Interpretation |
|---|---|---|
| Japan | +51.4% | Much longer lag even after controls |
| France | +31.8% | Longer lag |
| Germany | +23.6% | Longer lag |
| Russia | +22.5% | Longer lag |
| United Kingdom | +16.3% | Longer lag, but shrinks strongly after journal control |
| Italy | +13.3% | Longer lag |
| United States | +10.4% | Longer lag |
| China | -19.0% | Shorter lag |
| India | -9.2% | Shorter lag |
| Pakistan | -10.6% | Shorter lag |
| Saudi Arabia | -6.6% | Shorter lag |
For very late retractions, defined as more than 10 years, a publisher-adjusted logistic model showed:
| Country presence | Adjusted odds ratio for >10-year retraction |
|---|---|
| United Kingdom | 2.77 |
| Japan | 2.53 |
| Germany | 2.47 |
| Italy | 1.61 |
| China | 0.16 |
| Russia | 0.11 |
| Saudi Arabia | 0.18 |
| Pakistan | 0.08 |
| India | 0.38 |
The United States and France were not statistically strong in this specific logistic model after controls, even though their raw long-tail shares were high. That suggests their long-tail patterns are more heavily explained by journal/publisher/subject/reason mix.
The statistical interpretation:
Country effects remain after adjustment, but much of the country pattern is really journal-publisher-subject-reason structure wearing a country label.
14. Hypothesis testing summary
| Hypothesis | Result | Evidence |
|---|---|---|
| Countries differ in time to retraction | Supported | Kruskal-Wallis H = 2493.7, p < 1e-300 |
| Multinational papers are slower to retract | Not supported for journal-like records | Non-conference Mann-Whitney p = 0.088, Cliff’s delta = 0.011 |
| Multinational status predicts late retraction after adjustment | Opposite direction | Logistic OR for >10-year retraction = 0.52, p = 2.8e-24 |
| Country-specific reason profiles differ | Strongly supported | Multiple FDR-corrected chi-square enrichments |
| Country pairs have specific retraction clocks | Supported | Several pair-level Mann-Whitney tests significant after FDR |
| Journal/publisher effects explain country differences | Partly supported | R² rises from 0.229 to 0.302 when journal buckets are added |
| Society-linked journals explain long-lag country profiles | Partly supported | Higher society-linked share in US/Japan/Europe, but adjusted country effects persist |
| Subject mix explains country patterns | Partly supported | Japan/US/France more biomedical, Russia more social science, China/India/Saudi/Pakistan more process-heavy publisher clusters |
15. The exceptions are the most informative part
Exception 1: China is fast overall, but not always fast
China’s median lag is 1.50 years, but China + PLoS has a median lag of 3.92 years, and China + Spandidos has 5.64 years. So “China-linked retractions are fast” is only true in the aggregate because the aggregate is dominated by fast publisher-audit clusters.
Exception 2: Multinational does not mean long-lag
Saudi Arabia, Pakistan, Malaysia and Ethiopia are highly multinational in this dataset, but their medians are short and their after-10-year shares are tiny. Their multinational records are often in recent, publisher-audit, peer-review, or paper-mill-related clusters.
Exception 3: Japan has low multinational share but very long lag
Japan has only about 25% multinational records but the longest country median, 4.60 years, and the highest after-10-year share among large countries, 26.4%. This points toward older biomedical, clinical, institutional and fraud/misconduct-heavy corrections.
Exception 4: Russia is slow but not long-tail
Russia has a median lag of 3.19 years, but only 0.83% after 10 years. Its signature is not late biomedical correction. It is plagiarism/duplication-heavy, often in social-science or humanities-like spaces.
Exception 5: China-United States is not like China-Pakistan
China + United States has a median lag of 2.31 years, image concerns 38.6%, fraud/misconduct 14.5%, and society-linked share 15.2%. China + Pakistan has median lag 1.58 years, paper-mill/peer-review tags 65.2%, and no after-10-year records in this dataset. Same China label, different retraction ecosystem.
16. The final map: countries are not causes, they are coordinates
The country column is tempting. It invites rankings. It whispers: faster, slower, better, worse. But the database resists that crude reading.
Country-specific differences are real. Japan, France, Russia, Italy and the United States have longer median lags. China, Pakistan, Saudi Arabia, Ethiopia, South Korea and India have shorter medians. Some countries are image-heavy, some plagiarism-heavy, some paper-mill/peer-review-heavy, some long-tail biomedical.
But the deeper conclusion is not national character. It is publication ecology.
Countries appear in different parts of the publishing machine:
- China, India, Pakistan, Saudi Arabia, Ethiopia: large recent publisher-audit and paper-mill/peer-review clusters.
- Japan, United States, Germany, Italy, France, United Kingdom: more biomedical, society-linked, image/data, misconduct and long-tail correction clusters.
- Russia: a strong social-science and plagiarism/duplication signature.
- China-United States, Japan-United States, Italy-United States: slower, more biomedical and image-heavy collaboration clusters.
- China-Pakistan, Pakistan-Saudi Arabia, India-Saudi Arabia, Ethiopia-India: newer, faster, more process-heavy multinational clusters.
The country column is therefore not the verdict. It is the first clue.
The real retraction geography is made of journals, publishers, subjects, collaborations, editorial systems, institutional investigations, paper-mill audits, image forensics and time. It is not a flat political map. It is a weather map with moving storms.
Some storms are old and forensic.
Some are recent and industrial.
Some gather around journals.
Some gather around publishers.
Some cross borders so quickly that the country label becomes less useful than the network itself.
Science corrects itself, but the correction travels through pipes. The country column tells us where the pipes surface. 🔬📍
No comments:
Post a Comment