Stevan Harnad harnad at ECS.SOTON.AC.UK
Thu Jul 29 14:17:24 EDT 2004

Dear Sigmetrics colleagues:

Your feedback on these cumulating results would be very welcome.

Stevan Harnad

Date: Thu, 29 Jul 2004 19:09:00 +0100 (BST)
From: Stevan Harnad <harnad at>

On Thu, 29 Jul 2004, "Frederick Friend" wrote:

> Congratulations to Tim and his colleagues for this very valuable work! And
> it is clearly presented even to somebody like myself who is not used to
> looking at graphs.
> Can I check out one conclusion? Do the graphs show that in most cases the OA
> advantage is greater when there are few OA articles? If so, do we know -
> apart from the nuclear and particle physics example - what will happen as
> the number of OA articles grows?


It's probably premature to draw any conclusion on a correlation between
the OA impact advantage and the proportion OA until Tim has done the
actual correlations. I believe he is doing them now.

(It is tricky, statistically, because sample-size is varying by year
and field, and in general it is not really effect-magnitude or direction
that covaries with sample-size, but effect detectability. There is also
a covariate effect of time that has to be taken into account: The
proportion of OA grows across the years; a second time effect is that the
OA advantage seems to be biggest in the first year or two. But that is
also where the N's are smallest, because the data are not yet complete.
So a separate analysis for the latest 3 years is probably also

Hence Tim will need to look at the correlations among the following

(1) Number of articles
(2) Number (and proportion) of OA articles
(3) OA citation advantage
(4) year
(5) field

These are all different correlations, and to understand some of them
pairwise, we have to partial out the effects of the others.

The design (it seems to me) is a kind of 3-way analysis of co-variance
(articles (1) by year (4) by field (5)) with 2 dependent variables
(proportion OA (3) and OA advantage (4)) and then proportion (and/or perhaps
number of articles) used as a covariate to be partialled out.
(Tim, please consult with Mac to confirm this stat design and consult
on how to apply it.)

Cheers, Stevan

More information about the SIGMETRICS mailing list