Dear Tomas,
Let me post my answer on the SPM mailing list. Other people might have
the same question as you do.
> When I use p<0.001 I get very "tiny" results and only at the voxel
level -
> see fig. 1
> I prefer to set the primary threshold to p<0.01 which gives more
> meaningfull results at the cluster level in the context of my
hypothesis -
> see fig. 2.
>
> But I doubt if it is valid to use such low threshold and -
unfortunatelly
> -
> I can not find any hints to decide it. Statements such as "experience
> with other datasets may guide you" do not help me not being a
> statistician. I just have "the feeling", that I do not have to run in
> problems with false positives when correcting for it at the cluster
> level...
As far at the cluster defining threshold is concerned, there is no magic
threshold which works for all analyses. However, cluster p-values are
still valid (statistically speaking) regardless of the threshold you
use. A lower threshold produces larger clusters, but cluster p-values
are adjusted for that threshold.
Choosing a right threshold is up to the investigator. A high threshold
lets you localize your cluster very well, but can only find the tip of
your signal. On the other hand, a low threshold lets you find more
voxels per cluster, but has limited ability to localize the signals. So
if you feel p<0.01 gives you meaningful results, you should use it.
Hope this helps!
-Satoru
|