Why Growing Retractions Are (Mostly) a Good Sign

Retractions of scientific papers have recently been in the spotlight. Unfortunately, the interpretation of statistics about them is often flawed. The realisation that most retractions follow from scientific misconduct [1] seems to have reinforced, in the minds of both scientists and journalists, the idea that data on retractions, and generally data on findings of misconduct, provide information about the prevalence of fraud itself [2]. The recent growth in retractions, for example, is often invoked as evidence that scientific misconduct is increasing [2–4]. Similarly, findings that more papers are retracted by high-ranking journals, in biomedical fields, and in certain countries, and that more men than women are found guilty of misconduct are used to suggest possible risk factors for scientific misbehaviour [5–9]. The obvious alternative interpretation—that these statistics are proportional not to the prevalence of misconduct but to the efficiency of the system that detects it—is given equal or lower attention [3,4,9–11]. I will present four lines of evidence to suggest that retractions have grown not because of rising misconduct—an explanation that I call the ‘‘growing misconduct’’ hypothesis (GMH)—but because scientists have become more aware of and responsive against fraudulent and flawed research. I call this second explanation the ‘‘stronger system’’ hypothesis (SSH), although this is partially a misnomer, because a recent strengthening of measures against misconduct is not just a hypothesis, but a historical fact (Box 1). The data I present in this essay to support my argument were retrieved from the Thomson Reuters Web of Science (WoS) database, which is unique in covering over a century of publications. The WoS database marks both errata (corrections to previous papers) and retractions as ‘‘correction’’ or ‘‘correction, addition’’ (total n = 304,000 circa). Retractions can be retrieved from all these ‘‘corrections’’ by selecting those that include the term ‘‘retraction’’ in their title (total n = 2,294). Notably, most previous studies on retractions have used the PubMed database. Unlike WoS, PubMed has a specific category for retractions. However, PubMed restricts its coverage mostly to biomedical research and only started recording errata in 1987. This limitation may have caused some of the misunderstandings that this essay aims to debunk.