Hi Federico,
Thanks for the minimax ref.
> Practically it does not give any different results from the usual
> approaches but the theoretical aspects are interesting.
In what context? In PET and FMRI datasets, comparing to FWE-corrected
p-values (that is, p-values created in the usual way to control
Familywise Type I error), I have found FDR to be considerably more
powerful.
(Were you referring to the minimax/wavelet-shrinkage problem?)
I'm still getting a sense of how the power of (voxel-wise) FDR
compares to the usual cluster size p-values, but it's hard to compare
cluster- and voxel-wise inference.
-Tom
|