Hi there,
> Please look this up in the mailbase. This is not a highly recommended
> method of multiple correction. Please read Karl and/or Will's e-mail on
> the mail base. My suggestion, based on what I have learn from will, is
> that you use FWE instead.
I wouldn't be quite so hard on FDR, and, actually, neither is Will:
> As to what you *should* use, I can't say. Historically, the advice has been
> to always try for voxel-level FWE 0.05. But there are many papers out these
> days using FDR. As long as you adequately describe the approach you've used and
> what it means, you're on safe ground.
[www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind04&L=SPM&P=R673451&I=-3]
Perhaps for being so easy a concept to understand, I quite like to use
FDR. As it controls voxels in the original implementation, not clusters,
I mostly use an additional extent threshold of a couple of voxels. Also,
as far as I know, combined cluster size and voxel-wise inference is in
the making, at least for SnPM. What it comes down to is your preference:
do you aim for specificity (I want to be sure I have pretty much NO
false positives) versus sensitivity (I want to be sure to not overlook
true positives, which FWE also controls). Additionally, FWE requires a
minimum ammount of smoothing (about 2 voxel sizes, if I remember
correctly), which FDR does not.
Perhaps I should add that, no, I do not get money for advocating FDR,
and I declare no financial relationship with Tom Nichols (quite sorry
about this, actually ;)
Best,
Marko
=====================================================================
Marko Wilke (Dr.med./M.D.)
[log in to unmask]
Universitäts-Kinderklinik University Children's Hospital
Abt. III (Neuropädiatrie) Dept. III (Pediatric neurology)
Hoppe-Seyler-Str. 1, D - 72076 Tübingen
Tel.: (+49) 07071 29-83416 Fax: (+49) 07071 29-5473
=====================================================================
|