Dear Mike et al.
If I may add to Cyril's answer: There is no statistical ground for
fixing arbitrary the cluster size threshold to k=10.
Imagine you resampled you original data to either 2x2x2 mm3 or 3x3x3
mm3 voxel resolution, keep all else equal and assume this is
minimally affecting the underneath signal (which is true). Then for
the same activation pattern (from you original data) you would
consider as "interesting" a cluster of 80mm3 or 270mm3 volume
depending on how you resampled your data!!!
AFAIK this "p<.001 uncorrected and k=10" approach was used for
*display purposes only* in order to remove from the picture any "too
small to even consider" cluster of voxels and to make the image look
prettier.
What you actually need to do is:
- take a voxel level cluster forming threshold, say p=.001 (or lower
to be on the safe side)
- derive the minimal cluster size (in voxels) for a cluster to be
significant, say at FWE p<.05.
For the latter you can use this little bit of code:
https://github.com/CyclotronResearchCentre/SPM_ClusterSizeThreshold
Note though that the results are valid only when smoothness is
assumed to be homogeneous all over the brain volume, i.e. when the
cluster extent threshold is the same across the brain!
Best,
Chris
Le 11/07/2016 à 14:01 :
[log in to unmask]"
type="cite">
Actually I have a
puzzle. If an activation blob is identified via 0.001, k=10,
uncorrected in my fMRI experiment,
not all those blobs are correct - now you can only keep the ones
with cluster size or height FWE p<=0.05 (or FDR, but that
doesn't control type 1 error rate per se)