Odd that this should come up right now, as we're just doing the same thing. Our goal was to be able to functionally define ROIs according to how well the model fit the data. We have overall F's (not including nuisance terms) for each subject and wanted to submit them to a random-effects analysis.
One thought was to transform the F's into z-scores to be able to do t-tests (these ROIs will be interrogated later, so I'm not astoundingly concerned about false positives). Does this sound reasonable?
Craig