Some articles are highly tweeted, usually in the first few days after publication, while many others receive zero ‘tweetations’. How is the number of tweets related to the number of citations an article receives?

Some scholars have expressed a positive opinion about the association between tweets and citations (Eysenbach 2011; Shuai et al. 2012). Eysenbach (2011), for example, found fairly strong correlations (0.40–0.70) between tweets and citations for 55 articles published in the Journal of Medical Internet Research, and concluded that “social media activity either increases citations or reflects the underlying qualities of the article that also predict citations, but the true use of these metrics is to measure the distinct concept of social impact.” Others have expressed more reserved opinions. Thelwall et al. (2013) evaluated over 100,000 PubMed articles and concluded that tweets do predict citations, but also observed that the correlations were quite weak, and that tweets do not necessarily reflect academic impact (see also Costas et al. 2014; Haustein et al. 2014; Priem et al. 2012; Zahedi et al. 2014a). Yet others are outright negative. For example, Colquhoun and Plested (2014) and Beall (2013) argued that Twitter corrupts science.

There are several mechanisms through which a strong positive correlation between the number of tweets and the number of citations could occur. For example, it is possible that after a scientifically important article is published, this is noticed by scientists and/or the general public, who will then start tweeting. Obviously—through mechanisms unrelated to Twitter—high-impact articles will be well cited several years later, and a strong positive correlation arises between tweets and citations. It is also possible that the Twitter community reinforces citation rates by making an article known to a large scientific audience.

However, several mechanisms could cause the correlation between tweets and citations to be close to zero, or even negative. According to Haustein et al. (2014), it is possible that “the general public is not interested in the same topics as scientists cite but is interested in articles that have low scientific impact”. Their qualitative analysis of the 15 most highly tweeted articles in the biomedical literature showed that these articles are either curious or funny, have potential health applications, refer to a catastrophe, or are topics that concern scientific publishing or scientists’ behavior. In a related vein, it has been suggested that highly tweeted articles are about “catchy” topics (Van Noorden 2012), about “offbeat topics, current events, and general curiosities” (Liu 2014), and about “climate change, human health and diet, and online information and privacy” (Taylor and Plume 2014). Yet another possibility is that Twitter users hardly read the articles they tweet about. A recent revelation by Tony Haile, the chief executive officer of Chartbeat (a company that provides real-time analytics to websites and blogs) is consistent with this possibility: he stated that they “found effectively no correlation between social shares and people actually reading” (Haile 2014). Furthermore, Twitter could be used for “pathological” self-promotion (Buela-Casal 2014; Fenner 2014; Lin 2012), a phenomenon which could undermine the predictive validity of tweetations.

PLOS ONE is a large open access journal that facilitates various altmetrics. I focused on the 27,856 articles published between 1 July 2012 and 31 June 2013. Citations were counted using Crossref. The mean number of tweets per article was 2.66 (SD = 20.0). 59.6 % of the articles received zero tweets.

I entered the publication date and the number of tweets in a linear regression analysis in an attempt to predict the number of citations (variables in all analyses were rank transformed to account for non-normality). The obtained standardized β coefficients were −0.27 for the publication date (this number is negative since citations accumulate over time) and 0.10 for the number of tweets.

Next, I repeated the regression analysis, using the publication date, the number of tweets, as well as the number of article views as predictors. The obtained β coefficients were −0.19 for publication date, −0.06 for tweets, and 0.41 for article views on the PLOS ONE website. So, for a given number of article views, tweets negatively predict citations.

Figures 1, 2 and 3 provide further insight into the underlying mechanisms between tweeting, article views, and citations. Figure 1 shows that a clear association exists between the number of tweets and the number of PLOS ONE article views (β = 0.38 when controlling for publication date). PLOS ONE articles can be viewed not only on the PLOS ONE website, but also on PubMed Central (PMC), a website without altmetrics services. The association between tweets and article views on PMC is close to zero (β = 0.01). These results suggests that tweets attract article views on the PLOS ONE website because people click on the hyperlink embedded in the tweet, and that the number of tweets is unrelated to the interests of non-Twitter (PMC) users. The number of tweets was associated with activity on other social media (β = 0.34 for the number of Mendeley posts and β = 0.41 for Facebook activity; Fig. 1). Figure 2 illustrates that articles without tweets are unlikely to have many article views. This suggests that Twitter is the primary medium for becoming a viewing hit. Figure 3 shows that highly-viewed articles are better cited when they are ignored on Twitter. This suggests that articles which have acquired their views from non-Twitter sources are more highly cited than articles which have acquired their views via Twitter.

Fig. 1
figure 1

Number of PLOS ONE article views, PubMed Central article views, Facebook likes/shares/posts/comments, Mendeley additions, and citations as a function of the number of tweets, for PLOS ONE articles published between 1 July 2012 and 31 June 2013. The number of tweets were counted in logarithmically spaced bins. Data were measured on 10 June 2014

Full size image
Fig. 2
figure 2

Number of articles as a function of number of tweets and number of PLOS ONE article views, for PLOS ONE articles published between 1 July 2012 and 31 June 2013. Articles were divided into 100 groups based on the percentile rank of the number of views. Article views and tweets were measured on 10 June 2014

Full size image
Fig. 3
figure 3

Number of citations as a function of number of tweets and number of PLOS ONE article views, for PLOS ONE articles published between 1 July 2012 and 31 June 2013. Articles were divided into 100 groups based on the percentile rank of the number of views. Citations, article views, and tweets were measured on 10 June 2014

Full size image

The top 15 of the most highly tweeted PLOS ONE articles are listed in Table 1. Consistent with the above literature, highly-tweeted articles were about ‘curious’ topics or about human health and well-being. Specifically, several highly-tweeted articles were about psychological/neuroscience research related to gender, political orientation, and cognitive performance (articles 2, 4, 6, 11, 13, 14). These highly tweeted psychological articles often contained unique or spectacular results, but it is worth mentioning that null findings were highly tweeted too (article 4). Other popular articles were about health and weight loss (articles 7, 8, 9), consequences of the Fukushima accident (articles 3 and 15), the discovery of a new animal species (article 10), an analysis of Twitter data (article 1), and research about the media coverage of scientific findings and the success of research grants (articles 5 and 12).

Table 1 Statistics of the most highly tweeted PLOS ONE articles published between 1 July 2012 and 31 June 2013
Full size table

In conclusion, yes: tweets do predict citations, but the effect is very weak. Article views obtained from Twitter hyperlinks have less academic substance than article views obtained from other referral sources. It seems that the scientific citation process acts relatively independently of the social dynamics on Twitter.

The virtual lack of correlation between tweets and citations does not imply that Twitter is inconsequential for scientists. It is of course possible that the number of tweets represents something else than academic impact, for example ‘hidden impact’ (i.e., academic impact that is not detected using citation counts), ‘social impact’, or relevance for practitioners (e.g., Darling et al. 2013; Taylor 2013). Furthermore, it is possible that tweets influence science in indirect ways, for example by steering the popularity of research topics, by faming and defaming individual scientists, or by facilitating open peer review (Liu 2014; Mandavilli 2011). The exact causal mechanisms between tweets and citations require further study, and this may require longitudinal rather than cross-sectional research designs.

One limitation of the present analysis is that it focused on PLOS ONE only, a large open-access journal covering a variety of research topics. The obtained β coefficient between tweets and citations of 0.10 is in line with Haustein et al. (2014) who reported zero-order Spearman correlations between tweets and citations in the range of 0.10–0.20 for 134,929 biomedical articles that were mentioned on Twitter at least once. However, there exist large differences in Twitter use between scientific disciplines, with tweets being relatively uncommon in for example economics, sociology, and history of science (Holmberg and Thelwall 2014). Furthermore, some journals have an official tweeting account, which increases the probability that their articles receive tweetations (Haustein et al. 2014). The PLOS and PLOS ONE Twitter accounts have posted 6,398 and 7,594 tweets, respectively, as of 16 August 2014. Altmetrics services are not perfectly reliable. For example, it has been found that the PLOS application programming interface (API) does not recover the same tweets as Altmetric.com (Zahedi et al. 2014b). Another limitation is that the use of Twitter for the dissemination of scholarly information is still in its infancy. The social media landscape might change dramatically in the years to come (Van Noorden 2014).

The present analyses were based on citation data which can be downloaded from the PLOS website (http://www.plosone.org/static/almInfo). MATLAB code can be found in the Supplementary Material.