3.1. Results
No information is available as to the reasons for retraction of 178 out of 2,041 articles, or 8.7% of articles analyzed (
Table 1). Papers with an unknown reason for retraction appeared in journals with a lower IF than papers with a known reason for retraction (
p < 0.0001). Often, articles retracted for unknown reasons appeared in journals that were not among the subscription holdings available at several large universities. Papers retracted for an unknown reason were less often written by an author with other retractions (χ
2 = 41.57;
p < 0.0001), and were retracted in less time than other retracted papers (
p < 0.0001).
Table 1.
Comparison of articles for which a retraction notice was available and the reason for retraction could be determined (“Known”) vs. articles for which no retraction notice was available or the reason for retraction could not be determined (“Unknown reasons”).
Table 1.
Comparison of articles for which a retraction notice was available and the reason for retraction could be determined (“Known”) vs. articles for which no retraction notice was available or the reason for retraction could not be determined (“Unknown reasons”).
| Known reasons | | Unknown reasons | | χ2 or T | |
---|
| Mean | SD | Mean | SD | value | p value |
Sample n | 1,863 | - | 178 | - | - | - |
Journal IF | 7.71 | 9.79 | 3.12 | 4.76 | 6.169 | <0.0001 |
Repeat offenders (%) | 709 (38.1) | - | 24 (13.5) | - | 41.568 | <0.0001 |
# Authors per paper | 5.04 | 3.31 | 4.29 | 2.96 | 2.907 | <0.004 |
Months to retract | 33.95 | 34.81 | 22.03 | 25.42 | 4.448 | <0.0001 |
Comparing papers retracted for misconduct to papers retracted for reasons other than misconduct (
Table 2), journal IF was higher among fraudulent papers (
p < 0.0001). Roughly 57% of fraudulent papers were written by a first author with other retracted papers, whereas only 21% of papers retracted for reasons other than misconduct were written by an author with other retractions (χ
2 = 246.3;
p < 0.0001). Average number of authors was significantly higher for fraudulent papers (
p < 0.0001), and fraudulent papers were retracted significantly more slowly than papers retracted for reasons other than misconduct (
p < 0.0001).
Table 2.
Comparison of articles for which retraction was explained as a result of known or possible fabrication or falsification (“Misconduct”) vs. articles for which retraction was explained as the result of any other cause (“No misconduct”).
Table 2.
Comparison of articles for which retraction was explained as a result of known or possible fabrication or falsification (“Misconduct”) vs. articles for which retraction was explained as the result of any other cause (“No misconduct”).
| Misconduct | | No misconduct | | χ2 or T | |
---|
| Mean | SD | Mean | SD | value | p value |
Sample n | 881 | - | 982 | - | - | - |
Journal IF | 8.75 | 10.12 | 6.77 | 9.40 | 4.376 | <0.0001 |
Repeat offenders (%) | 500 (56.8) | - | 209 (21.3) | - | 246.345 | <0.0001 |
# Authors per paper | 5.53 | 3.34 | 4.60 | 3.23 | 6.106 | <0.0001 |
Months to retract | 43.23 | 37.47 | 25.66 | 29.91 | 11.234 | <0.0001 |
A plot of the demographics of papers retracted for unknown reasons (
Figure 1) suggests that they are more like papers retracted without evidence of scientific misconduct.
3.2. Discussion
This work replicates and extends an earlier study [
4], using a larger and more authoritative database. [
6] The demographics of papers retracted for misconduct are substantially different from papers retracted for other issues (
Table 2), as predicted by the “deliberate fraud” hypothesis:[
4] authors of fraudulent retracted papers appear to target high-IF journals (
Table 2); to have other retracted papers (
Table 2); to diffuse responsibility across more co-authors (
Table 2); and to delay retracting fraudulent papers (
Table 2). These results confirm that papers retracted for misconduct represent a deliberate effort to deceive.
Figure 1.
Comparison of parameters for papers retracted for “Unknown” reasons to papers retracted for either “Misconduct” or “No misconduct”; all points shown are averages. Papers retracted for “Unknown” reasons (n = 178) are more similar to papers retracted for “No misconduct” (n = 982) than to papers retracted for “Misconduct” (n = 881).
Figure 1.
Comparison of parameters for papers retracted for “Unknown” reasons to papers retracted for either “Misconduct” or “No misconduct”; all points shown are averages. Papers retracted for “Unknown” reasons (n = 178) are more similar to papers retracted for “No misconduct” (n = 982) than to papers retracted for “Misconduct” (n = 881).
This database improves upon the first database in several important ways. The database used here spans the years from 1973 to 2012 [
6], rather than the more limited span of years reported in the first study, from 2000 to 2010 [
4]. The new database contains 2,047 retracted articles [
6], rather than 788 retracted articles [
4], so the new database is 2.6-fold larger. The original database relied only upon published retraction notices to determine reasons for retraction, and such notices can be cryptic [
4]; this led to errors in determining why some articles were retracted [
6]. The new database used information gleaned from a wide range of sources, in addition to the published retraction notices, with a focus on reports from the ORI, to determine why articles were retracted [
6]. This new information led to a reevaluation of the reason for retraction [
6] of a substantial number of papers. Overall, 15.9% of retractions from the original study [
4] were reclassified as being due to misconduct.
Our results are substantially different from an earlier study that found that, compared to papers retracted for error, papers retracted for misconduct have fewer authors and appear in low-IF journals [
8]. This earlier study evaluated 395 papers retracted between 1982 and 2002 [
8], so the period of overlap between the two studies is complete, but we evaluated an additional 1652 retracted papers. The earlier analysis [
8] concluded that only 27.1% of papers were retracted for misconduct (falsification, fabrication, or plagiarism), whereas we found that 67.4% of papers were retracted for misconduct (falsification or fabrication) [
6]. Nevertheless, the studies concur in finding that erroneous studies are withdrawn more rapidly than fraudulent studies [
8].
Our results appear to differ from newly-reported findings that retracted randomized clinical trials (RCTs) have significantly fewer authors than case-matched unretracted RCTs. The number of authors of retracted RCTs averages 5.0 (±3.2 SD), whereas unretracted RCTs average 6.7 (±5.8 SD) authors (Steen and Hamer, this journal). In contrast, retracted fraudulent papers have an average of 5.5 (±3.3 SD) authors, whereas papers retracted for other reasons have an average of 4.6 (±3.2 SD) authors (
Table 2). Two possibilities could explain the apparent discrepancies. First, RCTs may generally have more authors than other types of published studies; RCTs often involve multiple sites and may require the time and effort of more investigators. Alternatively, it is possible that retracted articles in general have fewer authors than unretracted articles; it would require a case-control matching of retracted to unretracted articles to address this question, and this research has not been undertaken to our knowledge.
Conclusions here may be controversial because plagiarism is treated as a lesser offense than either fabrication or falsification of data. However, a clear distinction is drawn between plagiarism of words and plagiarism of data. Word plagiarism can be inadvertent, careless, or even innocent [
9], meant to flatter not deceive [
10], whereas data plagiarism must be considered misconduct. Plagiarism of data requires either that plausible circumstances be fabricated under which the allegedly “new” data could have been acquired, or that old data be altered—and so falsified—to appear new. Word plagiarism may be less harmful than data plagiarism, in that word plagiarism is unlikely to have an impact on patient treatment. This is because word plagiarism alone cannot affect the results of a meta-analysis, whereas data plagiarism could potentially lead to the same data being counted twice in a meta-analysis. Such “double-counting” would give inordinate weight to one set of experimental results and could result in an unrealistic between-study homogeneity [
11].
It is a controversial decision to lump all word plagiarism together, whether extensive copying of whole paragraphs or minor use of a few words. However, retraction notices virtually never provide detail as to how extensive the plagiarism was in a particular retracted paper. It would be interesting to compare papers retracted for extensive plagiarism and those retracted for minor plagiarism, but it is not clear how such a study could be undertaken.
A limitation of the present study is that retraction of a paper for fraud probably makes it more likely that other papers by the same author will be examined closely. Therefore, other papers that are tainted by misconduct are more likely to be identified and retracted. In contrast, papers retracted for error are unlikely to lead to reexamination of an author’s published opus. Hence, we cannot distinguish between two possibilities: that fraudulent authors are more likely than other authors to produce multiple fraudulent papers; or that fraudulent authors are more vigorously expunged from the literature. Both possibilities may be true.
Another limitation of the present work is that we cannot really address the motivation of authors who commit fraud. We can only hypothesize what trends might correlate with a deliberate effort to deceive and test those hypotheses. However, confirming the hypotheses does not prove an effort to deceive. The only way to prove such an effort is for authors to confess it and we do not anticipate such an outcome.
A third limitation of this work is that we cannot be certain that the causes of retraction cited in the retraction notices are true. Some retractions attributed to error could actually be due to fraud; it is not in an author’s interest to be open about having committed fraud. Retraction notices are often cryptic or ambiguous, which may be motivated by the retracting author’s desire to deny fraud. Most authors with multiple retractions have probably committed fraud, except in a limited number of retractions that were due to an error so pervasive that it discredits several linked papers or a long period of research.