Oori,
> will you be using T values estimated by 3dREMLfit?
Yes, exactly.
> I am curious to know if you are concerned about heteroskedasticity between
> beta and residuals on
>
> (a) the single subject level (i.e., in subject Joe's brain, higher betas are
> associated with larger var(residuals) -- I've seen this in a few cases, and it's
> interesting [sometimes]) and/or
Yes, within-subject variability is considered in the model in the sense higher precision of percent signal change estimate will have more say in the final result.
> (b) heteroskedasticity on the group level -- voxels with higher mean beta on
> the group level being associated with higher var(beta) or higher
> var(residuals) in that voxel
The cross-subjects variability is modeled together with the within-subject variability in the analysis.
> In any case, if you have a reference to the approach you're implementing I'd
> appreciate a pointer.
The basic theory part is the same as in FSL. Most of the heterogeneity stuff is from my own ideas, and I'm still trying to work out a few details and need to test their robustness.
Gang