With 3dMVM, there's a difficulty in that (at present) it does not write out residuals to be analyzed. If you analyze the statistics sub-brick(s), the ACF will be biased by the presence -- you hope! -- of activations.
One option would be to compute the 3 ACF parameters for each subject's residuals (errts) time series dataset, then take the mean or median of each of these, and use those 3 values in the 3dClustSim -acf run.
