Hi Floris & SPMers,
In that situation, then I would simply raise the cluster defining threshold
for the cluster size test and wouldn't run the "restricted" voxel test. If
you use a higher threshold to define clusters, then you will have smaller
clusters, but you should be able to localize your signals better. You might
think that raising the threshold will reduce the cluster size, hence your
clusters might no longer be significant. But the calculation of cluster
extent p-values takes into account that you are using a higher threshold,
and adjusts p-values accordingly.
The sequence of tests suggested doesn't sound very convincing to me. I
guess you will find more activations just by using an intermediate cluster
size test, since you are basically limiting your search volume to brightest
areas of the image. I think you need some sort of justification to limit
your analysis in those areas.
-Satoru
Satoru Hayasaka ==============================================
Post-Doctoral Fellow, MR Unit, UCSF / VA Medical Center
Email: shayasak_at_itsa_dot_ucsf_dot_edu Phone:(415) 221-4810 x4237
Homepage: http://www.sph.umich.edu/~hayasaka
==============================================================
At 10:20 AM 3/21/2005 +0100, Floris de Lange wrote:
>Dear SPM-ers,
>
>I would like to ask your input about the validity of a certain cascade
>of hypothesis testing.
>The issue pertains to testing for significant activations. The
>suggestion was raised to first test for significant clusters (using
>cluster level statistics), setting a more or less lenient intensity
>treshold, and identifying clusters of significant size.
>Then, in order to improve the localization information within those
>significant clusters, it was suggested to make a mask of the significant
>clusters, and test for significant voxels within this mask.
>
>Now, my idea would be that this is invalid, since you first identify
>where to look (using the greater power of cluster-level statistics) and
>then use this information to do a statistical test on the same data to
>get better localization precision.
>
>It was however suggested that testing for extent of activation and
>testing for intensity are only mildly associated, and can therefore be
>regarded as more or less independent tests.
>
>What is your idea about this way of treating the data, first identifying
>significant clusters and using this information to identify significant
>voxels? Is it valid or not?
>
>Many thanks for your input,
>Yours,
>Floris
|