On Wed, 9 Dec 2009 18:41:09 +0100, swann pichon
<[log in to unmask]> wrote:
>Dear all,
>
>It seems that a majority of papers publishing ROI analyses does not apply
>correction for multiple comparisons.
>
>The rationale behind it, I suppose, being that ROI analyses correspond to
>planned comparisons and therefore are exempt from usual corrections applied
>to post-hoc tests.
>
>- Is there a relevant reference that supports that point ?
Which multiple comparison problem are you referring to?
I assume you mean the following: if one tests (for example) 10 ROIs, then
there is a mult comp problem with 10 tests.
AFAICT, the truth of the matter is that, according to textbook statistical
practice, such ROI analyses are _not_ exempt from the requirement that a
correction be done.
Why isn't the correction done?
(a) In many scientific communities, multiple comparison corrections are
routinely not done. (I'm not saying I think this is OK, but am saying it appears
to be true.)
(b) If the number of ROIs is pretty limited, the "sin" of not performing a
correction are pretty minor compared to the mult comp problem that usually
draws everyone's attention in neuroimaging: that over voxels, where the
number of tests is orders of magnitude higher.
>- Can this logic also be applied to ROIs that are dependent of the dataset
>tested ie when ROIs have been generated using main effects of a factorial
>design rather than using a functional localiser ?
>
>Concerning this last question, I am considering Friston et al 2006's remark
>(A critique of functional localiser) which states :
>*"The key thing to appreciate is that a contrast testing for a particular
>effect can be used as a localiser for the remaining [orthogonal] effects. In
>this sense, any factorial fMRI study has as many functional localisers,
>embedded within it, as there are orthogonal contrasts. A typical two-by-two
>design has three orthogonal contrasts. The natural conclusion is that all
>fMRI experiments are simply collections of functional localisers."*
The discussion in that paper and the one it argues with is very interesting.
Friston's point, IIRC, is that the orthogonality of the contrasts means that
there is underlying statistical independence, so the worries in the paper he's
arguing with don't apply. (I can't remember if the independence of contrasts
means that the underlying variables are truly independent, or only perhaps
nearly so---depends on which quadratic form is used to define orthogonality,
if my memory is correct.)
>
>Thanks a lot for helpful comments,
>
>--
>Swann Pichon, PhD
>Laboratory for Behavioral Neurology and Imaging of Cognition
>Department of Neuroscience, University Medical Center
>1 rue Michel-Servet, 1211 GENEVA 4, Switzerland
>Tel: +41 (0)22 379 5979
>Fax: +41 (0)22 379 5402
>Gsm: +33 (0)6 26 43 83 61
>http://labnic.unige.ch/
>
|