Dear Torsten,
well, let's see... With correction for multiple comparisons, you're actually
controlling for the rate of false positive imputable to the huge number
of tests you're doing (one for each voxel) by rejecting the null
hypothesis - so assuming that your two groups, let's say controls vs
patients, have the same distribution.
You can reject it with a p-value
corrected < 0.05 at least.
What differs here, is the nature of the
distribution you're building by permutating the label of each of your
subjects (C vs P): on the voxel-based approach, you're building the distribution of the maximal t-value found accross your t-map, and with a cluster-based approach, you're building the distribution of the maximal size of the suprathreshold cluster
found accross your t-map. For this latter method, you have thus to
specify the t-threshold you want to use (the -c option in randomise),
and it is somewhat arbitrary. It is nonetheless more sensitive than the
voxelwise correction and fully valid.
I would recommend though not to smooth too much your images (or to feed
the -c option of randomise with a too liberal t-value) as you may have
then some problems to interpret your significant results. Indeed, if
you get for instance one large cluster with two local peaks let's say
in the precentral and postcentral gyrus, you won't be able to say which
one is significant or even if both are. You may find your results
easier to interpret if these two "blobs" are indeed separated and not
connected by a too large gaussian smoothing or a t-threshold too low.
You can test it very quickly using only one permutation (-n option) and
looking at your raw t-maps to check if you have chosen the best amount
of smoothing and to decide which t-threshold should be applied for the
cluster-based correction.
Once again, I'm not really familiar with statistics, so Steve, Jesper or whoever, please correct me if I'm saying something awfully wrong or if I'm using a bad terminology!
Hope this will help anyway,
Gwenaelle
----- Message d'origine ----
De : Torsten Ruest <[log in to unmask]>
À : [log in to unmask]
Envoyé le : Vendredi, 26 Janvier 2007, 10h17mn 51s
Objet : Re: [FSL] RE : [FSL] Cluster Based thresholding
Dear Gwenaelle,
thanks very much for your message. Things are indeed clearer now. But what
rule is used to form a cluster? Or in other words, what's the difference
between the voxel based and the cluster based approach?
Merci en avance!
Cheers,
Torsten
___________________________________________________________________________
Découvrez une nouvelle façon d'obtenir des réponses à toutes vos questions !
Profitez des connaissances, des opinions et des expériences des internautes sur Yahoo! Questions/Réponses
http://fr.answers.yahoo.com
|