Articles in high-impact journals are, on average, more frequently cited. But are they cited more often because those articles are somehow more “citable”? Or are they cited more often simply because they are published in a high-impact journal? Although some evidence suggests the latter, the causal relationship is not clear. We here compare citations of preprints to citations of the published version to uncover the causal mechanism. We build on an earlier model of citation dynamics to infer the causal effect of journals on citations. We find that high-impact journals select articles that tend to attract more citations. At the same time, we find that high-impact journals augment the citation rate of published articles. Our results yield a deeper understanding of the role of journals in the research system. The use of journal metrics in research evaluation has been increasingly criticized in recent years and article-level citations are sometimes suggested as an alternative. Our results show that removing impact factors from evaluation does not negate the influence of journals. This insight has important implications for changing practices of research evaluation.
Journals play a central role in scholarly communication, yet their role is also contested. The journal impact factor in particular has been criticized on several accounts (Larivière & Sugimoto, 2019). The main critique is its pervasive use in the context of research evaluation, for example in tenure decisions (McKiernan, Schimanski et al., 2019). Scientists shape their research with impact factors in mind (Müller & de Rijcke, 2017; Rushforth & de Rijcke, 2015). In a meeting in San Francisco in 2012, cell biologists called for a ban on the impact factor from research evaluation, and conjoined the “San Francisco Declaration on Research Assessment”1 (DORA). A group of researchers and editors called for publishing entire citation distributions instead of impact factors, to counter inappropriate use (Larivière, Kiermer et al., 2016). More recently, a group of editors and researchers came together and called for “rethinking impact factors” (Wouters, Sugimoto et al., 2019).
At the same time, journal impact is one of the most clear predictors of future citations (Abramo, D’Angelo, & Felici, 2019; Callaham, 2002; Levitt & Thelwall, 2011; Stegehuis, Litvak, & Waltman, 2015). The question is why. Possibly, high-impact journals select articles that somehow tend to be cited frequently. Another possibility is that articles are cited more frequently because they are published in a high-impact journal, not because they tend to be cited frequently per se. Neither citations of an article nor the journal in which it is published needs to be representative of “quality.” Here, we simply study whether citations of an article are influenced by the journal in which it is published, not their relationship to “quality.”
Answering this question is not straightforward. In rare cases, publications appear in multiple journals, and researchers found that the version in a higher impact journal was more frequently cited than its twin in a lower impact journal (Cantrill, 2016; Larivière & Gingras, 2010; Perneger, 2010). However, duplicate publications are quite special, limiting the generalizability of this observation. Some other earlier work claimed that citations were not affected by the journal (Seglen, 1994).
We answer this question by comparing citations of preprints with citations of the published version. The number of citations C may be influenced by both the latent citation rate ϕ and the journal J in which the article is published (Figure 1). Possibly, high-impact journals perform a stringent peer review of articles, selecting only articles with a high latent citation rate, so that ϕ influences the journal J. The latent citation rate itself may be influenced by many factors and characteristics (Onodera & Yoshikane, 2015) and motivations for citing the paper (Bornmann & Daniel, 2008). These factors are not limited to the characteristics of the paper itself, but may also include author reputation (Petersen, Fortunato et al., 2014) or institutional reputation (Medoff, 2006). Regardless of which factors influence the latent citation rate, the number of citations of the preprint before it is published in a journal C′ is unaffected by where it will be published and is affected only by the latent citation rate ϕ. We rely on this insight to estimate the causal effect of the journal on citations Pr(C | do(J)). The identification of the causal effect is possible because of the so-called “effect restoration” (Kuroki & Pearl, 2014), provided we can estimate Pr(C′ | ϕ). We construct a parametric model that provides exactly such an estimate.
We gathered information about 1,341,016 preprints from arXiv, and identified the published version for 727,186 preprints (54%; see Supplementary Material for more details). We extracted citations of both the preprint version and the published version from references in Scopus. Preprint dates, publication dates, and citation dates are all extracted from Crossref, using a daily granularity. We used the major subject headings of arXiv as field definitions. The impact of journals is calculated as the average number of citations received in the first 5 years after publication for all research articles and reviews in Scopus. We perform our analysis per year (2000–2016) and field, as the journal effect may vary per year and field. Moreover, we restrict our analysis to journals that have at least 20 articles that were published at least 30 days after appearing as a preprint on arXiv (Figure S1). Clearly, our data has a selection bias (Bareinboim & Pearl, 2012) on papers being submitted to arXiv or not (A). However, we can show that this does not affect our estimate of the causal effect Pr(C | do(J)) (see Supplementary Material).
Time complicates our analysis. The time T′ before a preprint was published, the preprint duration, will clearly affect the number of prepublication citations C′, while the total time since publication T will affect the postpublication citations C. Preprints with a higher latent citation rate may perhaps be more quickly published, thus affecting T′. To tackle this problem, we model the full temporal dynamics of both pre- and postpublication citations.
The numbers of pre- and postpublication citations are not clearly related (Figure 3, panel A). The numbers of prepublication citations also do not clearly relate to journal impact (Figure 3, panel B). The relation between preprint duration and the number of prepublication citations is also not clear (Figure 3, panel C). The ratio of postpublication citations and prepublication citations is higher for high-impact journals (Figure 3, panel D). Articles in high-impact journals accumulate more postpublication citations relative to prepublication citations compared to articles that have appeared in lower impact journals. These results are possibly obfuscated by two counteracting effects: Higher latent citation rates lead to higher prepublication citations, but perhaps also to shorter preprint durations, reducing the time to attract prepublication citations. The model that we constructed is intended to address this issue.
We here report results from our model for the five largest fields and the publication year 2016. Other fields and years show qualitatively similar results (see Figures S2 and S4). Our model presents a good fit of both pre- and postpublication citations (Figure S5).
The journal citation multiplier is consistently higher than 1 (Figure 4, panel A). Publishing in journals, compared to being available on arXiv only, multiplies the citation rate substantially, as expected. For example, Nature shows a multiplier of 6.0–9.9 (95% CI) for papers published in 2016 in the subject of Condensed Matter and Science shows a multiplier of 7.5–12.0 (95% CI) for such papers. Using the median estimates and the approximation in Eq. 2, this implies that a Condensed Matter article published in Nature in 2016 that obtained about 200 citations would not have obtained even 10 citations had it been available on arXiv only. Had it been published in Science instead, it would have obtained almost 350 citations. This is only an illustration: Both parameter estimates and the citation dynamics themselves exhibit considerable uncertainty (see Supplementary Material).
Most relevant to our question, higher impact journals tend to show higher citation multipliers. The correlation between the (logarithm of) the journal impact and the (logarithm of) the median journal citation multiplier θj is on average 0.45 for each combination of field and year. It ranges from 0.063 for High Energy Physics in 2002 to 0.79 for Astrophysics in 2012. Interestingly, the correlation grows stronger for High Energy Physics and Astrophysics over time, hovering around 0.6–0.7 for recent years (Figure S3).
At the same time, the median latent citation rate eΦj is also clearly increasing with journal impact (Figure 4, panel B). For example, the U.S.-based Physical Review Letters has a relatively high journal impact and shows a latent citation rate of 0.15–0.17 (95% CI) for Condensed Matter in 2016. Its lower impact European counterpart Europhysics Letters shows a latent citation rate of 0.013–0.027 (95% CI) in the same field and year. Overall, the correlation between the (logarithm of) the journal impact and Φj is on average 0.54 for each combination of field and year. For High Energy Physics in 2002 the correlation is 0.72, while for Astrophysics in 2012 the correlation is 0.050. The highest correlation of 0.85 is observed for Astrophysics in 2006. This correlation grows weaker for High Energy Physics and Astrophysics over time (Figure S3). The median effective citation rate of a journal is eΦjθj, which aligns closely with the observed journal impact (Figure S6).
The latent citation rates also vary within journals, and are controlled by ϵj. Journals with a higher ϵj tend to publish articles with a larger variety of latent citation rates. For example, Europhysics Letters shows an ϵj of 0.7–1.1 (95% CI), while Science shows an ϵj of 0.2–0.3 (95% CI), resulting in a broader distribution of ϕi for Europhysics Letters than Science. In general, high-impact journals show more narrow distributions of latent citation rates than lower impact journals (Figure 4, panel C).
Why articles in high-impact journals attract more citations is a fundamental question. We have provided clear evidence that articles in high-impact journals are highly cited because of two effects. On the one hand, articles that attract more citations are more likely to be published in high-impact journals. On the other hand, articles in high-impact journals will be cited even more frequently because of the publication venue. This amplifies the cumulative advantage effect for citations (Price, 1976).
A recent publication (Kim, Portenoy et al., 2020) took a similar approach and compared citations of preprints with citations of the published version. Using a more rudimentary model they obtained similar results and also find an influence of the journal on citations, although they do not address the causal mechanism. They also find that preprints with more citations are more likely to be published, but do not analyze in what journals they are published.
Several mechanisms may play a role in the causal effect of journals on citations. High-impact journals tend to have a higher circulation (Peritz, 1995), and reach a wider audience. In addition, researchers may prefer to cite an article from a high-impact journal over an article from a low-impact journal, even if both articles would be equally fitting. Both mechanisms are consistent with our results and earlier results (Cantrill, 2016; Kim et al., 2020; Larivière & Gingras, 2010; Perneger, 2010). Distinguishing between these two causal mechanisms is difficult (Davis, 2010) and should be investigated further.
An alternative explanation may be that published preprints are more highly cited because the preprints were improved by high-quality peer review in high-impact journals. We deem this an unlikely scenario. Differences between the preprint and the published version are textually minor (Klein, Broadwell et al., 2016). Those modifications can of course be substantively important. Peer review may substantially improve and strengthen a manuscript. Nonetheless, we think it is unlikely to alter a paper’s core contribution so as to affect its citation rate considerably.
Our analysis is limited to mostly physics and mathematics because of our reliance on arXiv. We expect to see similar effects in the medical sciences and the social sciences, in line with earlier results (Cantrill, 2016; Larivière & Gingras, 2010; Perneger, 2010). It would be interesting to replicate our analysis on younger preprint repositories, such as bioRxiv or SocArxiv, once they have had more time to accumulate citations. Another limitation is that we considered references from published articles only. It would be interesting to include also the references of preprints. This presumably increases the number of prepublication citations (Larivière, Sugimoto et al., 2014), which may decrease the overall inferred journal causal effect.
In our model we assumed that the effect of publishing in a journal is identical for all articles published in that journal. However, the effect of publishing in a journal may possibly vary for different articles. For example, articles from well-known authors may be cited frequently regardless of the exact journal in which they are published, while articles from more junior authors may benefit more from publishing in high-impact journals. Teasing out these different effects is not straightforward, but presents an interesting avenue for future research.
The latent citation rate itself may be influenced by many factors and characteristics of the paper (Onodera & Yoshikane, 2015) and motivations for citing the paper (Bornmann & Daniel, 2008). Overall, our results suggest that the characteristics (X1, X2, … ) that drive citations (C) overlap or correlate with factors that drive journal (J) peer review (Figure 5). For example, novelty, relevance, and scientific breadth (X2 to X4) may affect both journal evaluation and citations directly, while methodological aspects affect journal evaluation (X1) and authors’ reputation (X5) only affects citations. Because the journal also affects citations, methodological aspects would have an indirect effect on citations in this example. What factors drive journal evaluation and what factors drive citations is not clear and should be further investigated.
We hypothesize that a subset of factors that are used in journal evaluation are also used in postpublication research evaluation, such as the UK REF (Traag & Waltman, 2019). This means that research evaluation (E) tends to correlate with journals (J) because of underlying common factors (Figure 5). Even if factors that influence research evaluation do not influence citations directly, they will still correlate because of the mediating effect of the journal. For example, if methodological aspects (X1) affect research evaluation (E), it would correlate with citations (C) only because methodological aspects affect the journal (J). If our hypothesis holds true, citations would be indicative of the evaluation of articles only because they were published in a particular journal. In that case, citations should not be normalized based on the journal in which they are published, as was attempted by Zitt, Ramanana-Rahary, and Bassecoulard (2005). Doing so would effectively control for the journal, thereby blocking these causal pathways. Indeed, Adams, Gurney, and Jackson (2008) find that journal-normalized citations do not correlate with evaluation. Similarly, Eyre-Walker and Stoletzki (2013) report an absence of various correlations with evaluations when controlling for the journal. These results provide some evidence for our hypothesis. Journal metrics might even be a more appropriate indicator than citations to individual articles, as was suggested by Waltman and Traag (2020), although our results neither affirm nor refute this possibility.
Possibly, evaluation itself is also affected directly by the journal in which an article is published, and depending on the context, perhaps also by its citations. Indeed, the proposed causal diagram only captures part of a larger web of entanglement.
The use of citations and journals in research evaluation is often debated. Removing the use of journal metrics from research evaluation, as for example advocated by DORA, may decrease the pressure on authors to publish in high-impact journals. The use of article-level citations for evaluation could be condoned by DORA, but the use of journal metrics could not. Even if journal metrics were to be removed from research evaluation, journals would continue to play a role in research evaluation, albeit indirectly. Evaluating researchers based on citations then may still reward authors who publish in high-impact journals. This may effectively exert selective pressures that drive the evolution of the research system (Smaldino & McElreath, 2016). Simply removing impact factors from research evaluation therefore does not negate the influence of journals.
I thank Rodrigo Costas, Ludo Waltman, Jesper Schneider, and other colleagues from CWTS. I gratefully acknowledge use of the Shark cluster of the LUMC for computation time.
The author has no competing interests.
Handling Editor: Staša Milojević