Results
In general, the OA articles are, as predicted, more cited then nonOA ones. More specifically, for OA articles, those that are mandated receive relatively more citations then nonmandated ones. The following chart (Figure 2) illustrates the results for the 4 institutions together. Appendix 1 details the charts for each institution separately.
Figure 2: Comparisons between articles groups for all institutions
The charts relative to each institution are almost similar to that for global chart (the 4 institutions together). The comparison ratios between couples of groups are positive. This shows a citation advantage for the first group (numerator) compared to the second one (denominator) for all these comparisons. To verify whether the difference between citations means for these different article groups is statistically significant, we conducted statistical tests. The technique used (under SPSS) is « PairedSamples T Test ». This choice is justified by the fact that we compare citation means of articles for each journal. Moreover, this test is robust even for moderated violation of distribution normality condition. As the last institution to adopt the mandate was in 2004, the test was run on a sample of articles M and NM from 2004 to 2006.
The PairedSamples T Test results (Table 10) shows that there is indeed a significant difference (p <0.05) in favor of the first group over the second one, for each of the following group pairs (Appendix 2):

OA articles over NonOA ones (pair 1),

OA Mandated articles over OA NonMandated ones (pair 2),

OA NonMandated articles over OA Mandated ones (pair 3),

OA Mandated articles over NonOA Mandated ones (pair 4),

OA Mandated articles over NonOA ones (pair 5),

OA NonMandated articles over NonOA ones (pair 6),

OA Mandated articles over NonOA NonMandated ones (pair 7).
Some have argued that the OA Advantage might be all or mostly just quality (SelfSelection) bias (Eysenbach, 2006; Kurtz and al, 2004; Davis, 2008). However, among OA articles, mandated ones have a citation advantage compared to nonmandated ones. As a consequence, the citation advantage of OA is not necessarily the effect of a selfselection of best articles by their authors. This can never explain the citation advantage at least, for mandated articles, as authors deposit their articles in order to respect their mandate and not based on a quality criteria selection. The spontaneous selfarching rate of 15% is in fact far from the compliance rate of 60%.
Logistic regression
The number of citations an article receives can be influenced by or correlated with a variety of variables. A logistic regression analysis has been conducted to study the correlation between citation counts (as dependent variable) and the following set of potential correlator/predictor variables:

OA: Is the article Open Access (1 if OA and 0 otherwise)?

M: Does the author's institution Mandate Open Access (1) or Not (0)?

Age : How old is the article (articles published from 2002 to 2006)?

Auth_N : How many coauthors does the article have?

Ref_N : How many references does the article cite?

IF : What is the Thompson/ISI "Impact Factor" (average citations per article in 2year window) of the journal in which the article was published (from 0 to 30)?

Sci : Is the article classified by ISI as Science (1) or Social Science (0)

Page_N : How many pages in the article?

USA : What is the country of the first author (USA 1, other 0)?

Review : Is the article classified by ISI as a "review" article (1) or not (0)?

CERN : Is the first author from CERN (1/0)?

South : Is the first author from Southampton (1/0)?

Minho : Is the first author from Minho (1/0)?

Queens : Is the first author from Queensland University of Technology (1/0)?

Age*OA : The interaction between Age and OA
About 32% of the articles in our sample have at least 1 selfcitation with an average of about 2 selfcitations per article. We accordingly excluded all selfcitations from the citation counts.
Citation counts are not normally distributed, particularly because of the many articles having zero citations and they cannot be successfully transformed into a normal distribution. Figue 3 shows the citation counts (minus selfcitations) distribution. So we used binary logistic regression analysis, with a dichotomous dependent variable.
Figure 3: Citation count distribution
We used stepwise logistic regression, for each test selecting the model that maximizes the chisquare likelihood ratio. To make the interpretation of the coefficients easier, we exponentiated the ß coefficients (Exp(ß)) and interpreted them as oddsratios. For example, we can say for the first model that for a one unit increase in OA, the odds of receiving 15 citations (versus zero citations) increased by a factor of 0.957. The following figure^{4} (figure 4) reports Exp(ß) values for each model having "Cit_a_xy&yz" as dependent variables ((x,y,z) {1, 2, 3, ..., 20}), where Cit_a_xy&yz = 1 if citation count (minus selfcitations) is between y and z and 0 if between x and y. Models are referred to as "M_r". The Exp(ß) values of variables turned out to have the same polarity and to be quite similar, with and without selfcitations.
Figure 4: The Exp(ß) values for logistic regressions
The figure (Figure 4) shows that citation count is positively correlated with IF, Age, Ref_N, Auth_N, OA, USA and M. In other words;

The higher the IF of the journal in which it was published, the higher an article's citation count.

The longer since an article was published, the higher its citation count.

The more references an article cites, the higher its citation count.

The more coauthors an article has, the higher its citation count.

Articles that are made OA have higher citation counts, and this small but significant independent OA effect is present in every citation range but it is greatest in the highest citation range (15 citations vs 20+ citations): The OA advantage is strongest for highly cited articles.

Articles from authors at institutions that have Mandates have higher citation counts; this effect is present only in the mediumhigh citation ranges (and is of course confounded with the level of author compliance with the institutional Mandate, discussed further below).

Review articles have higher citation counts; the effect is greater, the higher the citation range.
CERN articles have higher citation counts in the lowest and especially the highest citation range. However, when all CERN articles are excluded from our sample, there is no significant change in the other variables.
There is a significant interaction between Age and OA (Age*OA) for low citation interval (between 1 and 5) as well for high citation interval (20 citations and more). Both the linear main effect of age and OA, and this nonlinear interaction are significant. The following figure (Figure 5) shows the citation mean (Cit_a_15&20+) for OA and NOA articles corresponding to each Age value. This figure confirms the OA advantage. The difference between the two lines corresponding to OA and NOA is higher for older articles.
Figure 5: The citation count means of Age and OA
