Dr Gitelman-- I think you raise some crucial questions which span all
imaging modalities.
>
>I need help understanding how to compare variability between subjects.
-this is a big request. You outline the approach for usign the REfx
kit, but note that the result is only evidence that the test subject is the
same or different from the reference group, and, to some extent, where-- but
not HOW they differ. It is your last question which raises the most
chalenging point:
>Are there other ways to show how much difference there is between subjects-
>that is, a measure of the heterogeneity in the activations? I am referring
>not just to whether one group activated more or less in a particular area
>but whether the overall pattern of activations is significantly similar or
>dissimilar.
-- here you are addressing the identification of regional covariance
patterns and there expression in a scan or group of scans, or difference
image. There have been several approaches to this type of modeling [SSM,
Moeller and Strother; PLS, McIntosh and Bookstein; MANoPET, Friston et al-
all in NeuroImage]. My understanding of these models is rudimenatry at best,
and I am likely to misrepresent them. However, I believe that the common
elements are: the data are reduced to spatial factors-- Eigenimages or the
like-- which define regions whose covariance-- positive and negative--
define orthogonal principle components. The expression of such a factor can
then be tested in a new image, or dataset. As with any PCA-type analysis,
how these are set up determines what you will get-- whether including test
and refernce data together, and then probing for which resultant patterns
segregate with group, or an external [behavioral or clinical] measure, or
defining components in teh refernce group only,to test for their presence in
new images, etc.
It can be considered a limitation of the GLM in SPM that it cannot
give any result like this-- which only speaks to having a clear idea of hwat
you are testing and how to use a model. With a "massive univariate
approach"-- SPM-- the constellation of significant regions found are too
often, I believe, erroneously interepreted as a pattern of covarying or
interacting regions. This is the result of wishful thinking on the part of
investigators, not of lack of effort on the part of the authors to shed
light on what is too often applied as a "black box". One outstanding example
of how these models differ when they are directly compared comes from the
FIL, in fact [Fletcher et al, Neuroimage 3:209-215, 1996]. The basic SPM
model requires that you are testing for stationary factors-- so, each
cluster in an SPM must be viewed as an independent "factor" which differs
between conditions; what the relationship *among* these regions is cannot be
determined from the SPM itself-- although secondary analysis of covariance
has been implemented [can't tell you where, though]. So, at least, the
results can guide further analysis. On the other hand, I am increasingly
convinced that applying a covariance analysis from the outset is
preferable-- this is beginning in fMR [eg, Ellmore et al, NeuroImage], but I
gather the problems of noise, signal drift, coupling to blod
pulsation/respiration/etc, are significant stumbling blocks. I think what
you are after is the holy grail; if any of this helps you in your quest, so
much the better. To the extent that I have grossly blundered here, I hope
that other helpline montors will chime in.
>Could one also use statistical non-parametric mapping to say whether the
>individuals from the 2 groups differed?
>
I believe this would at least tell you hwere they differ based on
teh actual distributions in each group. But again, the result is not, if I
understand it, a covariance pattern.
Christopher Gottschalk, MD
Assistant Professor of Neurology & Psychiatry
Yale School of Medicine
Mailing Adress:
VAMC [116-A]
950 Campbell Avenue
West Haven, CT 06516
tel [203] 932-5711 x4329
FAX 937-4937
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
