It is presumably very difficult to track in an objective way how the data was used i.e. whether the 'extra' citations actually looked at or used the raw data or rely on the text of the paper. Personally I would cite the paper if I had used the data in any case as a way of acknowledging and defining the source, even if I hadn't utilised the text of the publication in any way.

This is largely because there is no obvious identifier for the data (although this is less of an issue for microarray data). I guess there is a relationship here with Creative Commons licenses. A registered license could serve as an indentifier. Can a license have a doi?

Also be interested in whether this can be extended to other fields.

Yes, the counts are of article-to-article citations. All of the data sets were associated with published articles.

The analysis in this paper does not include citations in the form of "We used the data from Accession Number X in Database Y" or "We used the microarray data located at http://XYZ" if they do not also include an article citation. Nor does it include citations that are listed in a table or figure rather than the references section, as is sometimes done in meta-analyses.

The impact of these exclusions depends on point of view. It does make for a conservative estimate of re-use (and needless to say, one that is much easier to assemble without needing full text and natural language processing).

Do these examples cover what you meant by citing raw data directly?
If not, I'm interested in what you meant, could you please elaborate?