With our resting state data, we regress out motion, CSF, and white matter signals. We aren't making any statements about individual maps and the data seem to suggest little difference between the group analysis maps both quantitatively and qualitatively.
In my peabrain, I would expect if using REMLfit whitened the data too much, there would be little fluctuation in the signal. Yet, performing ICA on timecourses that have been REMLfitted give very nice components that mesh well with the resting state literature.
To respond to Kyle, converting to z-scores should normalize the data and lead to similar results unless REMLfit causes the r-values to change across subjects or regions. But if the reduction is quantitative, this is still quite an issue that leads to the question of where the truth lies. And to that, I have no opinion.