Introduction

High-quality scientific research requires time, skill, and effort. The mantra to “publish or perish” threatens quality, as it motivates some researchers to cut corners to improve their track-record [1]. If quantity is rewarded over quality, then the researchers who work carefully are outcompeted [2]. Rewarding quantity favours those willing to compromise their content and sacrifice rigour for speed [3–5].

Investigating whether the quality of research has declined over time requires examining the quality of published papers. However, examining quality is labour intensive, as it requires a detailed post-publication peer review to examine aspects of quality, such as reproducibility, methodological quality, and usefulness for practice [6–9]. The time costs of assessing quality make it difficult to examine robust measures of quality in thousands of papers over time.

In this paper, we examine spelling errors as a sign of research quality. Spelling errors are unintentional and not as serious^† as intentional attempts to mislead, such as plagiarism and data fabrication [11]. However, spelling errors are a sign that the researchers have not carefully checked their work. An advantage of examining spelling errors is that they can be checked across many publications, allowing the creation of a large and generalisable data set. Some spelling errors are inevitable, and typographical errors have existed since the first printing press [12].

However, if the relative rate of errors is increasing, then that would be a sign that the average quality of research is declining.

More spelling errors over time might not only be due to decreased quality, but could also occur because of an increase in non-native English speakers. Maintaining a career in science as a non-native English speaker is challenging [13]. Compared to native speakers, non-native speakers need more time to write a paper, need more time to proofread, and are more likely to have had papers rejected due to their writing [13, 14]. This puts non-native speakers at a disadvantage in the highly competitive science system [4]. The scientific literature has become more geographically diverse over time. China has led this increase, with other large increases over time from India, Japan, Germany, and the Republic of Korea [15]. We attempt to account for the increase in non-native English speakers by including the first author’s country.

We aimed to examine a range of spelling errors in the medical and health literature and to examine their trends over time. We tested whether these errors were associated with the characteristics of the paper, e.g., the number of authors. We also tested whether spelling errors are an indicator of research quality by examining their association with citation counts.

Notes

^†Typographical errors in drug names can have serious consequences [10].

Methods

We started with a list of 4,310 spelling errors that are used to correct common errors on Wikipedia [16]. These errors do not include differences between UK and US spellings. Wikipedia covers much more than health and medical research; however, the writing style is formal and many of the errors were applicable to writing for health and medical research. Four common error types with examples are:

• missing a double-letter, e.g., “ocasion” instead of “occasion”

• wrongly adding a double-letter, e.g., “developped” instead of “developed”

• jumbled letters, e.g., “recrod” instead of “record”

• homophones, e.g., “convertion” instead of “conversion”

We excluded 250 errors that were names or words in medicine or potential acronyms (see Figure S.1). We added 30 common spelling errors related to statistical methods from a preliminary analysis of trends in spelling errors [17]. The final number of spelling errors was 4,095. The full list of errors is available on our GitHub repository [18].

We searched for spelling errors using the PubMed database of medical and health publications (https://pubmed.ncbi.nlm.nih.gov). We originally searched every year from 1970 to 2024 (the latest full year available); however, in random checks we found that some spelling errors in the earlier years were not in the publisher’s version of the abstract and were likely made by PubMed transcribers. Hence, we decided to examine only articles from 2008 onwards, from when 88% of the data were supplied to PubMed electronically by publishers [19]. The final study period was 2008 to 2024 (17 years).

The search was restricted to abstracts in English and was case-insensitive (e.g., including the errors “Prevalance” and “prevalance”). We searched the titles and abstracts only, as the full text is not openly available for all papers. Each spelling error was only counted once per abstract, but multiple different errors in an abstract were counted. The searches were made on 8 February 2025.

We randomly selected 200 spelling errors detected by our algorithm and manually verified that there was an error in the abstract and/or title. We also checked the 100 most common spelling errors to verify that they were spelling errors and not acronyms or valid words. We examined 200 randomly selected abstracts with no errors detected by our algorithm to check for false positives.

Random sample of papers without spelling errors

To create a comparison group of abstracts without the listed spelling errors, we matched two “control” abstracts without detected spelling errors to each “case” abstract with an error [20]. The control abstracts were randomly sampled from all PubMed abstracts in English that were published within two days of each case abstract. We used this matching to create a comparable set of abstracts, but did not use individual matching in the statistical models.

For all abstracts, we extracted the following data:

• Abstract and title word count

• Article type, e.g., editorial

• First author’s country

• Number of authors

• Publisher

• Number of citations (as measured by OpenAlex)

We used the country of the first author rather than all authors to simplify this variable and because the first author is often considered the most important position and is expected to lead the work [21]. We used publisher rather than journal as the journal would likely be too sparse to be usefully modelled, and we assumed that publishers would have similar quality control across their journals (e.g., use of copy editors).

Statistical models

1. Trends over time

An increase in the absolute number of spelling errors over time could be caused by: 1) an increase in the number of abstracts over time, 2) an increase in word counts over time. To adjust for these increases in volume, we use two versions of the trend: 1) the annual rate of abstracts with one or more spelling errors scaled to per 10,000 abstracts; 2) the annual rate of spelling errors per million words. We did not have word counts for all PubMed abstracts, hence we estimated this by multiplying the annual total number of abstracts by the average annual word count, which was estimated using our random sample.

2. Modelling error counts

For the analysis comparing “case” abstracts with spelling errors to “control” abstracts without errors, we fitted the error count using a Poisson model [22]. To control for word count, we used the combined abstract and title word count as an offset. We used the predictors of: number of authors, first author’s country, and publisher. We did not add date as a predictor, as the control abstracts were matched by date. The results are presented as rate ratios, which are a multiplicative scale showing the relative change in spelling error rates.

We suspected that the association between the number of authors and spelling errors would be non-linear. To examine this, we used the fractional polynomial approach to test seven non-linear transformations and a linear association and selected the transformation with the smallest AIC [23].

We used random effects for country and publisher and used horseshoe priors as we expected the effects for some countries and publishers to be close to zero [24].

3. Modelling citation counts

We examined whether spelling errors were associated with citation counts. Our theory is that spelling errors are a symptom of poor practice and therefore will be associated with citations, which we assume is a measure of the quality of the article [25], although an imperfect measure [26, 27].

We modelled citation counts using a Poisson model. We included the years since publication as a predictor, as papers generally accumulate citations over time. We included the potential confounders of article type, number of authors, first author’s country, and publisher.

We used a Bayesian approach for all models, as the parameter estimates and credible intervals are easier to interpret than standard statistical methods. For a complete introduction to Bayesian methods, see [28]. Full details of our models and priors are in Supplement S.2 and all our statistical code and data are available on GitHub [18].

Software and data

All analysis were conducted using R version 4.4.1 [29]. We extracted bibliographic data from PubMed and OpenAlex [30]. We used the ‘rentrez’ [31] and ‘openalexR’ [32] packages to access these databases. We used the ‘countrycode’ package to standardise country names [33]. The data and R code are available online [18].

Ethical approval

The study used publicly available data and did not require ethical approval.

Results

Common errors and time trends

There were 1,916 unique spelling errors that were detected 48,420 times in over 21 million abstracts (0.2%). The spelling errors studied are relatively rare, with one error per 445 abstracts.

The ten most common errors are shown in Table 1. The most common error was “odd ratio”. The correct spelling is “odds ratio” because although this is a single number, it is calculated as a ratio of two odds.

Table 1. Ten most common spelling errors in health and medical abstracts and titles from 2008 to 2024. Considering over 4,000 potential spelling errors.

Spelling error	Total
odd ratio	3,904
demographical	3,063
publically	2,987
principle component analysis	2,144
occured	1,644
fourty	1,160
indispensible	1,131
Fischer’s exact	1,102
occuring	1,086
statically significant	1,066

The trends in spelling errors are shown in Figure 1 and are relatively stable over time. There was an increase in errors from 2008 to a peak in 2016, followed by a general decrease until 2023 and a slight increase in 2024. The general trend remained similar after excluding the top ten errors (Figure S.2), so the trend is not driven by the most common errors.

Predictors of spelling errors

Some descriptive statistics on the included abstracts are in Table 2. Most of the abstracts were Articles, which included observational studies and randomised trials. The median number of authors was five.

After controlling for word count, spelling errors were most common in letters and were less common in reviews compared to articles (Figure 2).

There was a large difference between countries in the rate of spelling errors (Figure 3). Japan had the fewest spelling errors, while Serbia and Turkey had the most. The UK had a below average rate of spelling errors, but was not the lowest.

There was a large difference between publishers in spelling error rates (Figure 4). The eight publishers with the lowest mean error rates were all from the US. The International Union of Crystallography had the highest rate of spelling errors, with the most errors from the journal Acta Crystallographica Section E: Crystallographic Communications, which had generally short abstracts. The two most common errors in this journal were “futher” and “futhermore”.

We found a strong average reduction in the rate of spelling errors for papers with more authors (Figure 5). There was a non-linear association, with diminishing improvements when adding more authors to already large teams.

Spelling errors associated with citations

We found a strong association between spelling errors and citation counts (Figure 6). The rate ratio for two spelling errors was 0.92, which is an 8% reduction in citations compared to articles without spelling errors. The complete multiple regression model is shown in Table S.2.

Table 2. Descriptive statistics for the abstracts with and without spelling errors. The statistics are the number and percentage for categorical variables, and the median and first to third quantile for continuous variables.

		With errors (case)	Without errors (control)
Number		48,420	110,028
Year (median [Q1 to Q3])		2018 [2014, 2021]	2018 [2014, 2021]
Type (%)
	Article	43,225 (89)	89,784 (82)
	Review	4,596 (10)	12,115 (11)
	Letter	341 (0.7)	5,216 (4.7)
	Editorial	87 (0.2)	1,649 (1.5)
	Erratum	26 (0.1)	797 (0.7)
	Preprint	117 (0.2)	235 (0.2)
	Retraction	9 (<0.1)	152 (0.1)
	Other	19 (<0.1)	80 (0.1)
Open access (%)		24,436 (51)	54,260 (49)
Number of authors (median [Q1 to Q3])		5 [4, 8]	5 [3, 7]
Word count (median [Q1 to Q3])		258 [205, 303]	219 [147, 274]
Citations (median [Q1 to Q3])		11 [4, 28]	12 [3, 31]

Discussion

We predicted that spelling errors would increase over time due to an average decline in research quality. However, after an increase from 2008 to 2016, the error rates decreased (Figure 1). The increase in error rates was relatively small, going from 18 per 10,000 abstracts in 2008 to 27 per 10,000 in 2016.

The overall trend may be a mix of competing trends, with a decline in quality as people rush to complete papers, but with better software available to remove spelling errors. Some researchers are now using large language models (LLMs) to write entire sections of their papers [34]. Many abstracts are now likely to be written in whole or in part by LLMs, which should reduce spelling and grammatical errors [35, 36]. However, since LLMs reuse published text, they may mimic some common spelling errors (Table 1) [37]. The first LLMs for writing articles appeared in 2017 [38], so an increased use of LLMs may explain the decreasing trend from 2016 (Figure 1).

Although LLMs may reduce spelling errors, they could become another tool to perform research faster rather than better, creating other language problems, such as vague and derivative text.

Some errors will not be detected by spell checkers, for example “odds ration”, but should be detected by a proofread, which most authors should do since the approval of every author is a recommended criterion for authorship [39]. We found a strong association between more authors and fewer spelling errors (Figure 5), which shows the benefits of increased oversight from co-authors. Similarly, we found a higher error rate in letters, which often do not go through the same editorial checks as articles.

Some publishers appear to make less effort to vet the abstracts they publish [40], as there was a large variance in the rates of spelling errors between publishers (Figure 4). Some publishers may be neglecting quality to maximise profits, and the recent exponential increase in papers may be reducing the time that journal staff have to check articles [41]. Some publishers have been targeted by paper mills that create low-quality papers en masse [42]. One publisher, Impact Journals LLC, with an above-average spelling error rate (Figure 4), is on a list of predatory publishers [43]. Impact Journals LLC publish the journal Oncotarget which accounted for 88% of the spelling errors of this publisher and was de-listed by Medline in 2017 and by Clarivate Analytics in 2018 [44].

As expected, we found a large variance in spelling error rates by the author’s first country (Figure 3). Most of the countries with the highest error rates were non-English speaking or were lower- or middle-income countries. In contrast, the countries with lower rates were often high-income countries or English speaking. Japan had a strikingly low rate of spelling errors, which we cannot explain.

Two recent suggestions for non-native English speakers to improve their writing are to use AI tools to proofread their articles [13, 45] and for institutions to provide more support for English language editing services [13]. Copy editing reduces spelling errors and increases clarity, which improves the authors’ chances of passing peer review [46] and increases the value of the article for practice [47]. However, these services have financial and time costs and can be cut if research budgets are tight [48]. Reported turnaround times for copy editing services are 2 to 3 weeks, with costs starting at USD $360 (inflation adjusted from 2010) [49].

We found that more spelling errors were associated with fewer citations (Figure 6). Some researchers may not cite a paper after reading a spelling error in the abstract, as this may undermine their confidence in the paper’s robustness. Researchers who make spelling errors may be more likely to make other more serious errors, such as incomplete reporting of the results, and hence the reduction in citations could be due to wider quality concerns.

Related work

Other observational studies have examined changes in scientific language over time. The spelling error “myocardial infraction” increased between 1953 and 2016 [50]. Other trends in scientific writing over time include a greater use of acronyms [51] and jargon [52].

Spelling errors have recently been found in cell lines, which can be sloppy practice or research fraud [53].

Limitations

We only examined spelling errors in the titles and abstracts, as the full text of all papers is not freely available. We did not check grammar because we could not automate grammar checks on millions of papers. We only examined the health and medical literature, meaning our results are not generalisable to all scientific fields.

Our model of citation counts probably missed some important predictors (e.g., field and author seniority), meaning that we may have missed confounders of the association between spelling errors and citation counts.

There will be measurement error in our results, as some text that we have flagged as an error will not be an error. For example, some spelling errors may have been errors made by the publishers and not by the authors. However, we believe that these errors are relatively small (Table S.1) and do not impact our general conclusions.

Acknowledgements

The computational resources and services used in this work were provided by the eResearch Office, Queensland University of Technology, Brisbane, Australia.

Thanks to Mark Hooper for directing us to the Wikipedia data on common spelling errors. Thanks to Tasha Kitano for proofreading this paper.

Thanks to the National Library of Medicine for making the PubMed data openly available.

Funding

There was no funding for this study.

Author contributions

Conceptualisation, Methods, Software, Validation, Data curation, Writing – Original Draft: AB

Supplementary material

S.1 Included and excluded errors

A flow chart of the included and excluded errors is shown in Figure S.1.

The list of errors from Wikipedia were opened in Microsoft Word and words that were not marked as spelling errors were checked in Wiktionary (https://en.wiktionary.org/). Some errors were alternative spellings or rarely used words and hence were excluded from our list of errors.

Some errors were excluded because they had a correct meaning in the health and medical literature, e.g., “miliary” is a prefix for “tuberculosis” and not always a misspelling of “military”.

During manual checking of our automated flagging of spelling errors, we discovered that errors with four letters could sometimes be acronyms, and hence were not errors. For example, “agin” as a typo of “again” was also an acronym “AGIN”. Hence, we removed all errors of four letters or fewer, as these could sometimes be acronyms or initialisms, whose use has increased over time in the health and medical literature [51].

Words that could be correct with an apostrophe or hyphen were excluded because the search ignored these grammatical marks, for example the error “criterias” – which could be “criteria’s” and the error “noone” which could be “no-one”.

Some errors were excluded because they were from an abstract that was given in multiple languages. We excluded spelling errors that were only in the keywords.

S.2 Details on the statistical models

The Bayesian regression model and priors to examine differences the counts of spelling errors were:

Errors_i is the number of errors in the ith title and abstract (Errors_i = 0, 1, 2, . . .). The number of words in the abstract and title was an offset scaled to 100 words.

There were T article types, C countries, and P publishers. Some countries and publishers were sparsely represented, so to avoid over-fitting we used a horseshoe prior, which shrinks the estimates to zero to give a parsimonious model [24]. We also combined publishers and countries that had fewer than 200 observations into an “Other” category. The random effects for the country and publisher were centred using the mean to improve the convergence of the chains. For the paper type, which had eight categories, we used “Article” as a reference category to improve convergence (β₁ = 0).

The log-transformation (base 2) for the number of authors was selected using the fractional polynomial approach [23].

The percentage of missing data was 6.0% for publishers and 3.5% for first authors’ countries. We modelled “missing” as a separate category instead of using multiple imputation. No other independent variables were missing.

The Bayesian models were fitted using the Nimble package [54].

S.3 Sensitivity analysis for trends

S.4 Verifying our data

In a random sample of 200 spelling errors detected by our algorithm, we confirmed that all 200 were errors. During these checks, we noticed that many spelling errors were also correctly written in the same abstract, indicating that many errors were typos rather than a consistent misunderstanding.

In a random sample of 200 control abstracts that ideally would not have spelling errors, 564 spelling errors were flagged using the Microsoft Word spell checker. We individually verified each potential error and found eight confirmed errors. Many potential spelling errors flagged by Word were UK spellings, plant names in Latin, specialist medical words, acronyms and symbols; none of which were counted as spelling errors in this study. The low success rate for Microsoft Word in finding spelling errors indicates that this spell-check software cannot be used to accurately count spelling errors, which is why we searched for specific errors.

We used the error counts in a Bayesian calculation to estimate an upper limit for the false positive and false negative rates per abstract (Table S.1). We estimate a 90% probability that the false positive rate is below 1.0% and a 90% probability that the false negative rate is below 5.8%. These results give us confidence that almost all spelling errors flagged by our algorithm are genuine errors. We have less confidence that the control abstracts without errors are completely error-free. However, the false negative error rate is still relatively low and should be ameliorated by our large sample size, meaning that we can still make inferences about which abstracts are more likely to have spelling errors.

Table S.1. Estimated false positive and false negative error rates from 200 randomly selected spelling errors and 200 randomly selected control abstracts without errors. We estimated the upper 90% limit for the error rates using a Bayesian calculation starting with a non-informative Beta(1,1) prior.

Checked (error type)	Number incorrect	Mean incorrect	Upper 90% limit
Spelling errors from algorithm (false positive errors that were not spelling errors)	0	0.0%	1.0%
Randomly selected controls (false negative abstracts that contained spelling errors)	8	4.0%	5.8%

S.5 Regression model of citation counts

Table S.2. Multiple regression model for predicting citation counts. Mean rate ratio, 95% credible interval (CI), and posterior probability that the rate ratio is not equal to one.

Predictor (scale)		Mean	Lower CI	Upper CI	Posterior probability
Number of authors (+5)		1.1558	1.1553	1.1563	< 0.001
Time (inverse)		0.0004	0.0004	0.0004	< 0.001
Number of spelling errors (squared)		0.9890	0.9883	0.9897	< 0.001
Article type
	Article (reference)	1
	Editorial	0.3446	0.3382	0.3510	< 0.001
	Erratum	0.1267	0.1209	0.1329	< 0.001
	Letter	0.2536	0.2511	0.2561	< 0.001
	Other	0.5634	0.5259	0.6034	< 0.001
	Preprint	0.8895	0.8470	0.9337	< 0.001
	Retraction	0.0787	0.0665	0.0930	< 0.001
	Review	2.2390	2.2338	2.2443	< 0.001

References

[1]       Génova, Gonzalo, Astudillo, Hernán, and Fraga, Anabel. “The Scientometric Bubble Considered Harmful”. In: Science and Engineering Ethics 22.1 (Feb. 2016), pp. 227–235. doi: 10.1007/s11948-015-9632-6.

[2]       Smaldino, Paul E. and McElreath, Richard. “The natural selection of bad science”. In: Royal Society Open Science 3.9 (Sept. 2016), p. 160384. doi: 10.1098/rsos.160384.

[3]       Macleod, Malcolm R et al. “Biomedical research: increasing value, reducing waste”. In: The Lancet 383.9912 (Jan. 2014), pp. 101–104. doi: 10.1016/s0140-6736(13)62329-6.

[4]       Binswanger, Mathias. “Excellence by Nonsense: The Competition for Publications in Modern Science”. In: Opening Science: The Evolving Guide on How the Internet is Changing Research, Collaboration and Scholarly Publishing. Ed. by S¨onke Bartling and Sascha Friesike. Cham: Springer International Publishing, 2014, pp. 49–72. isbn: 978-3-319-00026-8. doi: 10.1007/978-3-319-00026-8_3.

[5]       Bohorquez, Natalia G et al. Researchers are willing to trade their results for journal prestige: results from a discrete choice experiment. July 2024. doi: 10.31219/osf.io/uwt3b.

[6]       Begley, C. Glenn and Ellis, Lee M. “Raise standards for preclinical cancer research”. In: Nature 483.7391 (Mar. 2012), pp. 531–533. doi: 10.1038/483531a.

[7]       Wynants, Laure et al. “Prediction models for diagnosis and prognosis of covid-19: systematic review and critical appraisal”. In: BMJ (Apr. 2020), p. m1328. doi: 10.1136/bmj.m1328.

[8]      Héroux, Martin et al. “Poor statistical reporting, inadequate data presentation and spin persist despite Journal awareness and updated Information for Authors”. In: F1000Research 12 (Nov. 2023), p. 1483. doi: 10.12688/f1000research.142841.1.

[9]       Jones, Lee, Barnett, Adrian, and Vagenas, Dimitrios. “Linear regression reporting practices for health researchers, a cross-sectional meta-research study”. In: PLOS ONE 20.3 (Mar. 2025), pp. 1–23. doi: 10.1371/journal.pone.0305150.

[10]       Veronin, Michael A., Schumaker, Robert P., and Dixit, Rohit. “The Irony of MedWatch and the FAERS Database: An Assessment of Data Input Errors and Potential Consequences”. In: Journal of Pharmacy Technology 36.4 (June 2020), pp. 164–167. doi: 10.1177/8755122520928495.

[11]        Bolland, Mark J., Avenell, Alison, and Grey, Andrew. “Publication integrity: what is it, why does it matter, how it is safeguarded and how could we do better?” In: Journal of the Royal Society of New Zealand 55.2 (Mar. 2024), pp. 267–286. doi: 10.1080/03036758.2024.2325004.

[12]       Okrent, Arika and O’Neill, Sean. “Blame the Printing Press”. In: Highly Irregular. Oxford University PressNew York, Sept. 2021, pp. 116–146. isbn: 9780197616017. doi: 10.1093/oso/9780197539408.003.0004.

[13]       Amano, Tatsuya et al. “The manifold costs of being a non-native English speaker in science”. In: PLOS Biology 21.7 (July 2023), pp. 1–27. doi: 10.1371/journal.pbio.3002184.

[14]       Sezer Yamanel, Rabia Gönül et al. “Barriers to writing research papers and getting them published, as perceived by Turkish physicians – a cross sectional study”. In: European Science Editing 47 (Dec. 2021). doi: 10.3897/ese.2021.e69596.

[15]       Schneegans, Susan, Straza, Tiffany, and Lewis, Jake. UNESCO Science Report: The race against time for smarter development. 2021. url: https://www.unesco.org/reports/science/2021/en.

[16]       Wikipedia. Lists of common misspellings. 2024. url: https://en.wikipedia.org/wiki/Wikipedia:Lists_of_common_misspellings     (visited on 11/11/2024).

[17]       Barnett, Adrian and White, Nicole. Casual inference and pubic health – What a rise in common spelling errors says about the state of research culture. 2024. url: https://blogs.lse.ac.uk/impactofsocialsciences/2024/10/28/casual-inference-    and-pubic-health-what-a-rise-in-common-spelling-errors-says-about-the-state-of-research-culture/ (visited on 11/15/2024).

[18]       Barnett, Adrian. Code and data for the analysis of spelling errors in PubMed abstracts. 2025. doi: 10.5281/zenodo.15107029. url: https://github.com/agbarnett/spelling_wikipedia.

[19]      National Library of Medicine. NLM Technical Bulletin. May 2008. url: https://www.nlm.nih.gov/pubs/techbull/mj08/mj08_mla_dg.html (visited on 01/08/2025).

[20]      Schulz, Kenneth F and Grimes, David A. “Case–control studies: research in reverse”. In: The Lancet 359.9304 (Feb. 2002), pp. 431–434. doi: 10.1016/s0140-6736(02)07605-5.

[21]       Demaine, Erik D. and Demaine, Martin L. Every Author as First Author. 2023. arXiv: 2304.01393 [cs.DL]. url: https://arxiv.org/abs/2304.01393.

[22]      Dobson, Annette J and Barnett, A. G. An Introduction to Generalized Linear Models. 4th. Texts in Statistical Science. Boca Raton, FL: Chapman & Hall/CRC, 2018.

[23]      Royston, P., Ambler, G., and Sauerbrei, W. “The use of fractional polynomials to model continuous risk variables in epidemiology”. In: International Journal of Epidemiology 28.5 (Oct. 1999), pp. 964–974. doi: 10.1093/ije/28.5.964.

[24]      Carvalho, Carlos M., Polson, Nicholas G., and Scott, James G. “Handling Sparsity via the Horseshoe”. In: Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics. Ed. by David van Dyk and Max Welling. Vol. 5. Proceedings of Machine Learning Research. Apr. 2009, pp. 73–80.

[25]      Bornmann, Lutz and Daniel, Hans-Dieter. “What do citation counts measure? A review of studies on citing behavior”. In: Journal of Documentation 64.1 (Jan. 2008), pp. 45–80. doi: 10.1108/00220410810844150.

[26]      Pavlovic, Vedrana et al. “How accurate are citations of frequently cited papers in biomedical literature?” In: Clinical Science 135.5 (Mar. 2021), pp. 671–681. doi: 10.1042/cs20201573.

[27]      Wren, Jonathan D. and Georgescu, Constantin. “Detecting potential reference list manipulation within a citation network”. In: bioRxiv (2020). doi: 10.1101/2020.08.12.248369.

[28]      Gelman, Andrew et al. Bayesian Data Analysis. Chapman and Hall/CRC, Nov. 2013. isbn: 9780429113079. doi: 10.1201/b16018.

[29]      R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing. Vienna, Austria, 2024. url: https://www.R-project.org/.

[30]     Priem, Jason, Piwowar, Heather, and Orr, Richard. OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts. 2022. arXiv: 2205.01833 [cs.DL].

[31]       Winter, David J. “rentrez: an R package for the NCBI eUtils API”. In: The R Journal 9 (2 2017), pp. 520–526.

[32]      Massimo, Aria et al. “openalexR: An R-Tool for Collecting Bibliometric Data from OpenAlex”. In: The R Journal 15 (4 2024), pp. 167–180. doi: 10.32614/RJ-2023-089.

[33]      Arel-Bundock, Vincent, Enevoldsen, Nils, and Yetman, CJ. “countrycode: An R package to convert country names and country codes”. In: Journal of Open Source Software 3.28 (Aug. 2018), p. 848. doi: 10.21105/joss.00848.

[34]      Strzelecki, Artur. “‘As of my last knowledge update’: How is content generated by ChatGPT infiltrating scientific papers published in premier journals?” In: Learned Publishing 38.1 (2025), e1650. doi: https://doi.org/10.1002/leap.1650.

[35]      Ren, Dennis and Roland, Damian. “Arise robot overlords! A synergy of artificial intelligence in the evolution of scientific writing and publishing”. In: Pediatric Research 96.3 (Apr. 2024), pp. 576–578. doi: 10.1038/s41390-024-03217-0.

[36]      Binz, Marcel et al. “How should the advancement of large language models affect the practice of science?” In: Proceedings of the National Academy of Sciences 122.5 (Jan. 2025). doi: 10.1073/pnas.2401227121.

[37]      Snoswell, Aaron J., Witzenberger, Kevin, and El Masri, Rayane. A weird phrase is plaguing scientific papers – and we traced it back to a glitch in AI training data. Apr. 2025. url: https://theconversation.com/a-weird-phrase-is-plaguing-scientific-papers- and-we-traced-it-back-to-a-glitch-in-ai-training-data-254463.

[38]      McCook, A. Newly released AI software writes papers for you — what could go wrong? Nov. 2017. url: https://retractionwatch.com/2017/11/09/newly-released-ai- software-writes-papers-go-wrong/.

[39]      International Committee of Medical Journal Editors. Defining the Role of Authors and Contributors.  2024.  url:  https://www.icmje.org/recommendations/browse/roles-and-responsibilities/defining-the-role-of-authors-and-contributors.html (visited on 11/15/2024).

[40]      Editorial. In: Nature Chemical Biology 7.10 (Sept. 2011), pp. 649–649. doi: 10.1038/nchembio.683.

[41]      Hanson, Mark A. et al. “The strain on scientific publishing”. In: Quantitative Science Studies 5.4 (2024), pp. 823–843. doi: 10.1162/qss_a_00327.

[42]      Bishop, Dorothy. “Red flags for paper mills need to go beyond the level of individual articles: a case study of Hindawi special issues”. In: (Feb. 2023). doi: 10.31234/osf.io/6mbgv.

[43]      Predatory Journals. The Predatory Publishers List. 2025. url: https://www.predatoryjournals.org/the-list/publishers (visited on 03/24/2025).

[44]      Groneberg, David A. et al. “The story behind Oncotarget? A bibliometric analysis”. In: Scientometrics 117.3 (Oct. 2018), pp. 2195–2205. doi: 10.1007/s11192-018-2949-6.

[45]      Fontelo, Paul and Liu, Fang. “A review of recent publication trends from top publishing countries”. In: Systematic Reviews 7.1 (Sept. 2018). doi: 10.1186/s13643-018-0819-1.

[46]      Lee, Carole J et al. “Bias in peer review”. In: Journal of the American Society for information Science and Technology 64.1 (2013), pp. 2–17.

[47]      Glasziou, Paul et al. “Reducing waste from incomplete or unusable reports of biomedical research”. In: The Lancet 383.9913 (Jan. 2014), pp. 267–276. doi: 10.1016/s0140-6736(13)62228-x.

[48]      Rahal, Rima-Maria et al. “Quality research needs good working conditions”. In: Nature Human Behaviour 7.2 (Feb. 2023), pp. 164–167. doi: 10.1038/s41562-022-01508-2.

[49]      Kaplan, Karen. “Publishing: A helping hand”. In: Nature 468.7324 (Dec. 2010), pp. 721–723. doi: 10.1038/nj7324-721a.

[50]      Dilley, Rodney J. and Ash, Oliver G. “Addressing the cost of infractions in the online literature and databases”. In: PLOS ONE 12.11 (Nov. 2017), pp. 1–6. doi: 10.1371/journal.pone.0188761.

[51]       Barnett, Adrian and Doubleday, Zoe. “Meta-Research: The growth of acronyms in the scientific literature”. In: eLife 9 (July 2020). Ed. by Peter Rodgers, e60080. doi: 10.7554/eLife.60080.

[52]      Baram-Tsabari, Ayelet et al. “Jargon use in Public Understanding of Science papers over three decades”. In: Public Understanding of Science 29.6 (2020), pp. 644–654. doi: 10.1177/0963662520940501.

[53]      Oste, Danielle J. et al. “Misspellings or “miscellings”—Non-verifiable and unknown cell lines in cancer research publications”. In: International Journal of Cancer 155.7 (May 2024), pp. 1278–1289. doi: 10.1002/ijc.34995.

[54]      de Valpine, Perry et al. “Programming with models: writing statistical algorithms for general model structures with NIMBLE”. In: Journal of Computational and Graphical Statistics 26 (2 2017), pp. 403–413. doi: 10.1080/10618600.2016.1172487.