For me it depends previous experience with the problem domain.
If I have done this sort of app before, I am more likely to take the time to clarify the data model first and then start building code on top of that.
On a first-time project I am more likely to just jump in with coding, whipping up dummy data as needed and learning about hidden dimensions of the problem domain as I go. Not uncommonly there are data categories that would have been difficult to anticipate. When I discover these, I revise the data model and continue on. Often this approach begins with coding up a script to build the database and populate it. That way, on subsequent iterations, I just modify the db-build script, run it, and I'm back in business.
