On Mon, Nov 12, 2012 at 7:18 AM, Gabor Oederland <[log in to unmask]> wrote:
> Okay, now I understand.
>
>
> About the "neuropsychological test battery", this should have served as an example for running an F-test first and post-hoc tests afterwards... So I would use post-hoc tests only in case the F-test showed a significant group effect, but not for all the individual neuropsychological tests.
>>> In this case, you'd need to correct for the number of neuropsychological tests, so the initial F-test and post-hocs would need to be divided by N.
>
> Dunno, but as far as I understood you meant that this would not be the case with fMRI data? That is running an F-Test, getting some significant clusters, and running post-hoc tests inside these clusters (and not across the whole brain) would be invalid? But that's what post-hoc tests do, focusing on effects found in previous analyses?
>>> Yes and No. Here is an example, say I have 500,000 voxels and I find a cluster with 100 voxels for the F-test across all groups. Now, I test each comparison but only in 100 voxels. The search region is much smaller and the correction for multiple comparisons across voxels is much less severe. You might falsely conclude that two groups are different, when they are not different at the whole brain level. This is the problem with double-dipping. Do an initial analysis to limit the search regions of the next test.
In traditional statistics, you don't have to worry about the
correction for the number of voxels.
|