I have been involved in developing XMLSchema models for a community domain (science) and find the variable content models very complicated and probably too complex for my community. I'd be very interested in successful public deployment of XML Schemas with variable content models (E.g. "salami slice", "venetian blind", "russian doll", "hermaphrodite", "garden of eden"). Roger Costello has done a first class job in systematising this (http://www.xfront.com/BestPracticesHomepage.html) - and I'm personally flattered that the xml-dev mailing list has created this document (in the same way as SAX was created).
There's some SO discussion at: http://stackoverflow.com/questions/231553/what-are-best-practices-for-designing-xml-schemas
My current question is which of these approaches (or none) is actually used, so I'd be very interested in real examples of communities which have used this (or communities which have found any or all of the approaches too complex).