I’m working with a dataset with a very large amount of data collected from each participant over many sessions. I’d like to run 3dFWHMx in order to perform cluster correction on an analysis using all of a participant’s data, but the errts file produced by 3dREMLfit is hundreds of gigabytes, which is creating logistical issues. Would it be reasonable to use residuals from running 3dREMLfit on a subset of the data (e.g., the first and last session, or maybe the first run of each session) in order to estimate the smoothness of the noise? Or is it pretty important to use all of the data in order to get a faithful estimate?
Practically speaking, it has been my experience that the smoothness estimates from 3dFWHMx tend to be quite similar across a group of subjects acquired on the same scanner using the same software version and scanner sequence (NB: changing any of these things can make a big difference, though, so apply changes can make a big difference, though, so apply this suggestion only if it applies to your case!). Thus, I would guess that, indeed, using a representative subset of your data would give you good smoothing estimate—we do typically average smoothness to get cluster size, in large part because of the above consideration.
You can verify the homogeneity by estimating the noise smoothness (with ACF parameters!) and seeing how variable that is.