The implicit masking is not optimal
Currently, the implicit masking removes all constant voxels like in randomise. While this is enough for a classic linear model, for the SwE, simply having constant voxels within a subject can be problematic when we compute the within-subject covariance matrix (e.g., correlations of 1). This has been solved in the SPM toolbox. I will try to fix this tomorrow for the fsl version.
I can also see that there is an option --norcmask
in randomise (and included also in swe) which prevents the masking of constant voxels. There is probably a good reason for this, but I cannot figure what it is. If somebody knows it, I would appreciate to learn about it.