Because of the sheer number of articles published by PNAS, tracking the performance of each article was considered too onerous. As a result, I tracked the first and last 6-month cohort of articles (June-Dec 2004; and June-Dec 2006). By choosing the first and last cohort, I could estimate a temporal trend in the data. Please remember that PNAS was only one of the 11 journals analyzed in this study, and that Gunter Eysenbach’s study (PLoS Biology, 2006) analyzed only a 6-month cohort in one journal (PNAS, June-Dec, 2006). Granted, a full dataset from PNAS would have been ideal, and I encourage Prof. Harnad to gather and share the intervening years if he feels that the missing data points would change significantly the results of this study. My sense is that they won’t, but will challenge Prof. Harnad to prove me wrong.

Secondly, while aggregating the journals resulted in increased statistical power, we are combining articles published in different scientific fields (biology, medicine, bioinformatics, plant sciences, and multi-disciplinary sciences), which is why journal impact factors are not used as an explanatory variable. Please note that I did include the variable Journal as either a random variable (Table 2) or a fixed variable (Table S2), so journal-to-journal variation is being accounted for in the model.

Full text: http://openaccess.eprints.org/index.php?/archives/451-guid.html

SUMMARY: Davis [2008] — http://arxiv.org/pdf/0808.2428v1 — analyzes citations from 2004-2007 in 11 biomedical journals. For 1,600 of the 11,000 articles (15%), their authors paid the publisher to make them Open Access (OA). The outcome, confirming previous studies (on both paid and unpaid OA), is a significant OA citation advantage, but a small one (21%, 4% of it correlated with other article variables such as number of authors, references, and pages). The author infers that the size of the OA advantage in this biomedical sample has been shrinking annually from 2004-2007, but the data suggest the opposite. In order to draw valid conclusions from these data, the following five further analyses are necessary:

(1) The current analysis is based only on author-choice (paid) OA. Free OA self-archiving needs to be taken into account too, for the same journals and years, rather than being counted as non-OA, as in the current analysis.
(2) The proportion of OA articles per journal per year needs to be reported and taken into account.
(3) Estimates of journal and article quality and citability in the form of the Journal Impact Factor and the relation between the size of the OA Advantage and journal as well as article “citation-bracket” need to be taken into account.
(4) The sample-size for the highest-impact, largest-sample journal analyzed, PNAS, is restricted and is excluded from some of the analyses. An analysis of the full PNAS dataset is needed, for the entire 2004-2007 period.
(5) The analysis of the interaction between OA and time, 2004-2007, is based on retrospective data from a June 2008 total cumulative citation count. The analysis needs to be redone taking into account the dates of both the cited articles and the citing articles, otherwise article-age effects and any other real-time effects from 2004-2008 are confounded.

The author proposes that an author self-selection bias for providing OA to higher-quality articles (the Quality Bias, QB) is the primary cause of the observed OA Advantage, but this study does not test or show anything at all about the causal role of QB (or of any of the other potential causal factors, such as Accessibility Advantage, AA, Competitive Advantage, CA, Download Advantage, DA, Early Advantage, EA, and Quality Advantage, QA). The author also suggests that paid OA is not worth the cost, per extra citation. This is probably true, but with OA self-archiving, both the OA and the extra citations are free.