Davis et al’s 1-year Study of Self-Selection Bias: No Self-Archiving Control, No OA Effect, No Conclusion


Davis, PN, Lewenstein, BV, Simon, DH, Booth, JG, & Connolly, MJL (2008) Open access publishing, article downloads, and citations: randomised controlled trial British Medical Journal 337: a568

Overview (by SH):

Davis et al.‘s study was designed to test whether the “Open Access (OA) Advantage” (i.e., more citations to OA articles than to non-OA articles in the same journal and year) is an artifact of a “self-selection bias” (i.e., better authors are more likely to self-archive or better articles are more likely to be self-archived by their authors).

The control for self-selection bias was to select randomly which articles were made OA, rather than having the author choose. The result was that a year after publication the OA articles were not cited significantly more than the non-OA articles (although they were downloaded more).

The authors write:

“To control for self selection we carried out a randomised controlled experiment in which articles from a journal publisher?s websites were assigned to open access status or subscription access only”

The authors conclude:

“No evidence was found of a citation advantage for open access articles in the first year after publication. The citation advantage from open access reported widely in the literature may be an artefact of other causes.”

Commentary:

To show that the OA advantage is an artefact of self-selection bias (or of any other factor), you first have to produce the OA advantage and then show that it is eliminated by eliminating self-selection bias (or any other artefact).

This is not what Davis et al. did. They simply showed that they could detect no OA advantage one year after publication in their sample. This is not surprising, since most other studies, some based based on hundreds of thousands of articles, don’t detect an OA advantage one year after publication either. It is too early.

To draw any conclusions at all from such a 1-year study, the authors would have had to do a control condition, in which they managed to find a sufficient number of self-selected, self-archived OA articles (from the same journals, for the same year) that do show the OA advantage, whereas their randomized OA articles do not. In the absence of that control condition, the finding that no OA advantage is detected in the first year for this particular sample of 247 out of 1619 articles in 11 physiological journals is completely uninformative.

The authors did find a download advantage within the first year, as other studies have found. This early download advantage for OA articles has also been found to be correlated with a citation advantage 18 months or more later. The authors try to argue that this correlation would not hold in their case, but they give no evidence (because they hurried to publish their study, originally intended to run four years, three years too early.)

(1) The Davis study was originally proposed (in December 2006) as intended to cover 4 years:

Davis, PN (2006) Randomized controlled study of OA publishing (see comment

It has instead been released after a year.

(2) The Open Access (OA) Advantage (i.e., significantly more citations for OA articles, always comparing OA and non-OA articles in the same journal and year) has been reported in all fields tested so far, for example:

Hajjem, C., Harnad, S. and Gingras, Y. (2005) Ten-Year Cross-Disciplinary Comparison of the Growth of Open Access and How it Increases Research Citation ImpactIEEE Data Engineering Bulletin 28(4) pp. 39-47.

(3) There is always the logical possibility that the OA advantage is not a causal one, but merely an effect of self-selection: The better authors may be more likely to self-archive their articles and/or the better articles may be more likely to be self-archived; those better articles would be the ones that get more cited anyway.

(4) So it is a very good idea to try to control methodologically for this self-selection bias: The way to control it is exactly as Davis et al. have done, which is to select articles at random for being made OA, rather than having the authors self-select.

(5) Then, if it turns out that the citation advantage for randomized OA articles is significantly smaller than the citation advantage for self-selected-OA articles, the hypothesis that the OA advantage is all or mostly just a self-selection bias is supported.

(6) But that is not at all what Davis et al. did.

(7) All Davis et al. did was to find that their randomized OA articles had significantly higher downloads than non-OA articles, but no significant difference in citations.

(8) This was based on the first year after publication, when most of the prior studies on the OA advantage likewise find no significant OA advantage, because it is simply too early: the early results are too noisy! The OA advantage shows up in later years (1-4).

(9) If Davis et al. had been more self-critical, seeking to test and perhaps falsify their own hypothesis, rather than just to confirm it, they would have done the obvious control study, which is to test whether articles that were made OA through self-selected self-archiving by their authors (in the very same year, in the very same journals) show an OA advantage in that same interval. For if they do not, then of course the interval was too short, the results were released prematurely, and the study so far shows nothing at all: It is not until you have actually demonstrated an OA advantage that you can estimate how much of that might due to a self-selection artefact!

(10) The study shows almost nothing at all, but not quite nothing, because one would expect (based on our own previous study, which showed that early downloads, at 6 months, predict enhanced citations at a year and a half or later) that Davis’s increased downloads too would translate into increased citations, once given enough time.

Brody, T., Harnad, S. and Carr, L. (2006) Earlier Web Usage Statistics as Predictors of Later Citation ImpactJournal of the American Association for Information Science and Technology (JASIST) 57(8) pp. 1060-1072.

(11) The findings of Michael Kurtz and collaborators are also relevant in this regard. They looked only at astrophysics, which is special, in that (a) it is a field with only about a dozen journals, to which every research astronomer has subscription access — these days they also have free online access via ADS — and (b) it is a field in which most authors self-archive their preprints very early in arxiv — much earlier than the date of publication.

Kurtz, M. J. and Henneken, E. A. (2007) Open Access does not increase citations for research articles from The Astrophysical Journal. Preprint deposited in arXiv September 6, 2007.

(12) Kurtz & Henneken, too, found the usual self-archiving advantage in astrophysics (i.e., about twice as many citations for OA papers than non-OA), but when they analyzed its cause, they found that most of the cause was the Early Advantage of access to the preprint, as much as a year before publication of the (OA) postprint. In addition, they found a self-selection bias (for preprints — which is all that were involved here, because, as noted, in astrophysics, as of publication, everything is OA): The better articles by the better authors were more likely to have been self-archived as preprints.

(13) Kurtz’s results do not generalize to all fields, because it is not true of other fields either that (a) they already have 100% OA for their published postprints, or that (b) many authors tend to self-archive preprints before publication.

(14) However, the fact that early preprint self-archiving (in a field that is 100% OA as of postprint publication) is sufficient to double citations is very likely to translate into a similar effect, in a no-OA, no-preprint field, if one reckons on the basis of the one-year access embargo that many publishers are imposing on the postprint. (The yearlong “No-Embargo” advantage provided by postprint OA in other fields might not turn out to be so big as to double citations, as the preprint Early Advantage in astrophysics did, because at least there is some subscription access to the postprint; but the counterpart of the Early Advantage for the postprint is likely to be there too.)

(15) Moreover, the preprint OA advantage is primarily Early Advantage, and only secondarily Self-Selection.

(16) The size of the postprint self-selection bias would have been what Davis et al. tested — if they had done the proper control, and waited long enough to get an actual OA effect to compare against.

(17) We had reported in an unpublished 2007 pilot study that there was no statistically significant difference between the size of the OA advantage for mandated (i.e., obligatory) and unmandated (i.e., self-selected) self-archiving:

Hajjem, C & Harnad, S. (2007) The Open Access Citation Advantage: Quality Advantage Or Quality Bias?Preprint deposited in arXiv January 22, 2007. 

(18) We will soon be reporting the results of a 4-year study on the OA advantage in mandated and unmandated self-archiving that confirms these earlier findings: Mandated self-archiving is like Davis et al.‘s randomized OA, but we find that it does not reduce the OA advantage at all — once enough time has elapsed for there to be an OA Advantage at all. 

Stevan Harnad
American Scientist Open Access Forum