The Importance of Right and Left Censoring in the Analysis of Retracted Articles Over Time.

To the Editor: We read with interest the systematic evaluation of reasons for retractions in the cardiothoracic and vascular anesthesia field.1Rong LQ Audisio K Rahouma M et al.A systematic review of retractions in the field of cardiothoracic and vascular anesthesia.J Cardiothorac Vasc Anesth. 2022; 36: 403-411Abstract Full Text Full Text PDF PubMed Scopus (7) Google Scholar A key part of science is self-correction, and the authors are contributing by highlighting this process and visualizing its results over the past 30 years. However, we have some criticisms regarding the methodologic approaches to estimate the “rate” of article retraction and the analysis of the change in impact factor (IF) of retracted articles in this field over a 3-decade period. The methodologic flaws imply that the final conclusion that retractions “have plateaued in recent years, and the latency time from publication to retraction, as well as the journal IF have decreased significantly over the last 30 years,” is not justified. Our primary concern is the false conclusion that the time from publication to retraction (their “latency time”) has decreased over time (ie, that articles are retracted faster). This conclusion, however, is based on a biased comparison between old and new retractions due to the absence of right censoring in the analyses of these follow-up data. More recently published articles have a much smaller time window during which they can be retracted than older articles. In other words, articles published in the last decade may be retracted 30 years after publication, but this event lies in the future and is not yet observable. Because differences in follow-up time were not considered, the representation of the results in Figure 1—the more recent the publication, the shorter the interval between publication and retraction—is incorrect. Our secondary concern relates to data collection. The authors used data from Retraction Watch, founded in August 2010. Retractions before that date were included afterward, and the completeness of data collection was not ensured.2Oransky AMI. Retraction Watch FAQ 2010. Available at: https://retractionwatch.com/the-retraction-watch-faq/. Accessed: March 21, 2023.Google Scholar Hence, retractions before 2010 were often missed, which leads to immortality bias: An article from 1995 was less likely to enter the database when retracted in 2000 than in 2010. This leads to bias in the same direction. Therefore, the conclusion that the time from publication to retraction has decreased is not supported by these analyses, as varying follow-up time and time at risk (right- and left-censoring) should have been considered. We can alter Figure 1 to demonstrate this: Data points cannot exist in the red area of the graph and are likely underrepresented in the yellow area. Applying the logic to Figure 4 and leaving out the outlier in 2011, we concluded that there had been no substantial change in IF of retracted articles. In clinical studies, differences in follow-up duration and right-censoring can be handled by taking personal time instead of persons in the denominator of interest to estimate the incidence rate. Person time was usually the sum of calendar time when each subject was at risk of the outcome (making it a rate). In this case, that would translate into “article time,” when articles published 30 years ago contribute 30 years to this sum, whereas an article from last month only contributes 30 days. However, the authors only have data on retracted articles, although the number of other articles at risk for retraction is needed to provide the true retraction rate. When this is known, standard epidemiologic methods can be applied to obtain, for example, rate differences, time-to-event analysis (eg, a Kaplan-Meier curve [with right-censoring]), and analyses of changes in rates over time. However, time only can be counted over time windows during which events could be observed; therefore, if an article was published 30 years ago, but observation only began 10 years ago, the maximum observation time contributed by this article is 10 years (left-censoring).3Kurtzke JF. On estimating survival: A tale of two censors.J Clin Epidemiol. 1989; 42: 169-175Abstract Full Text PDF PubMed Scopus (21) Google Scholar Our criticism of the method does not mean we refute the hypothesis that retraction rates have changed over time. Given the changes in the scientific enterprise, it is still quite possible that the time from publication to retraction has changed. However, the data and figures from the authors do not support this. None. A Systematic Review of Retractions in the Field of Cardiothoracic and Vascular AnesthesiaJournal of Cardiothoracic and Vascular AnesthesiaVol. 36Issue 2PreviewNo systematic studies on retractions in cardiothoracic and vascular anesthesia exist. The aim of this analysis was to identify characteristics and trends of retractions in this field over the past three decades. Full-Text PDF