This question is really fascinating (especially since I suspect I have inkling of the sort of texts you're interested in). Let me just dump a handful of links which I think might be helpful from my recent delicious/pinboard (hmm... delicious pinboard).
While I can't point to specific results/papers. Aditi Muralidharan (@silverasm)'s WordSeer project is promising interesting results about slave narratives, though it does not rely on TEI. In this talk, though, she does a good job of answering Sean's question directly--how can metadata improve on generic full-text search. WordSeer (as I understand it) works mostly by supplementing raw text with part of speech tagging and other grammatical metadata.
With respect to TEI specifically one obvious place to start, I think, is MONK; as Dorothea notes, Martin Mueller is involved with the MONK project. You can find him summarizing MONK here and at greater length here. Unlike WordSeer, the Monk Workbench gives you metadata about author (including gender) and publication date. Matt Kirschenbaum mentions MONK in this article [PDF].
(How relevant it will be to you, I don't know, but when I heard John Unsworth speak recently, he talked about the work flow and complications of large text collections with different encoding standards--with reference to MONK in particular. You can hear him here.)
There was a recent discussion on the TEI list about eXist from which I learned a lot--again, about what's possible rather than specific results/reports.
Finally--this one is also not TEI specific--if you haven't seen Ben Schmidt's blog, he is doing some pretty impressive stuff. This post, on vocabulary use and age cohort gives some idea.
