Hi Stephen and other FSL experts,
As mentioned in my previous email (below), I'd like to know how I can get one value per voxel per subject that contains an unbiased estimate of the residual variance in the voxel signal across multiple runs after model fitting. This is in replication of an analysis that used the values stored in the SPM ResMS.img to match subjects from two groups. As you suggested, I've looked up which info this SPM image contains:
The ResMS.hdr/img contains an image of estimated residual variance which can be described as (e = Y - Xb --> ResMS = e'e/(n-k), where n = number of scans, k = number of factors). However, in reality, SPM divides e'e by trace(R*V), which is a more accurate approximation of the degrees of freedom in models with more complicated error structures. What trace(R*V) stands for is explained mathematically on page 37 of the HBF book Chapter 7: http://www.fil.ion.ucl.ac.uk/spm/doc/books/hbf2/pdfs/Ch7.pdf
I hope this clarifies a bit more what I'm looking for. I *think* - but might be wrong - that the FSL stats/sigmasquareds.nii.gz image is the closest to what is described above but I've only found this image in the first level analysis output and not at the second level in which I average across runs within subjects, so am not sure how I should obtain an overall subject average of residual variance.
thank you for your help,
Tessa
|