Richard:
You'll have to ask a statistician to arbitrate the fine points. However, I
don't think using an F-test which is designed to span all likely subspaces
as a way to exclude voxels from being saved (which is what spm still does
in the initial calculations) is the same as using a subject's own data to
define what is significant. My impression is that one would need a
different dataset to do this. I agree that the contrasts are orthogonal but
the underlying dataset is the same, and so this biases your result. Anyway
I invite comments if others disagree.
Darren
At 10:46 AM 12/6/00 +0000, you wrote:
>Dear Darren,
>
>Many thanks for your speedy reply!
>
>>1) If you're essentially using the same dataset to both identify the
>>cluster of interest and use that cluster to further threshold you data,
>>I'm not sure that would be correct.
>
>My feeling was that if the contrasts were orthogonal, then (under the null
>hypothesis) choosing a cluster using contrast X should not in any way bias
>whether those voxels should differ from zero in contrast Y, so the
>statistics should still be OK.
>
>The kind of situation would be something like this. Let's imagine that
>you know, from previous experiments, that a certain region should show up
>in the contrast 1 1 -2 for conditions A, B and N (N being a 'rest'
>condition). However, you now want to find out whether this region is
>significantly more active in condition A than in condition B. Clearly
>there is no reason, under the null hypothesis (i.e. just noise) that a
>voxel which shows up in A+B-2N should show up in A vs B (or in B vs A for
>that matter). So if the cluster does show up, this is interpretable.
>
>Surely using the same dataset is OK? We used to use the F contrast in
>SPM96 to reduce the number of voxels over which we corrected (as well as
>reducing the computational demands), and as I understood it we could do
>that because the F contrast for 'effects of interest' is orthogonal to
>individual t contrasts.
>
>Does this seem reasonable to you?
>
>Anyway, I'll try it with the recipe which you have kindly given me. I am a
>little puzzled now as to what the 'cluster' feature offered in SVC in
>SPM99 is really for! Surely correcting across a cluster identified using
>the same contrast is not particularly useful.
>
>Thanks again,
>
>Richard.
>
>--
>from: Dr Richard Perry,
>Clinical Lecturer, Wellcome Department of Cognitive Neurology, Institute
>of Neurology, Darwin Building, University College London, Gower Street,
>London WC1E 6BT.
>Tel: 0207 679 2187; e mail: [log in to unmask]
-------------------------------------------------------------------------
Darren R. Gitelman, M.D.
Cognitive Neurology and Alzheimer¹s Disease Center
E-mail: [log in to unmask]
WWW: http://www.brain.northwestern.edus
Voice: (312) 908-9023
Fax: (312) 908-8789
Northwestern Univ., 320 E. Superior St., Searle 11-470, Chicago, IL 60611
-------------------------------------------------------------------------
|