Hello,
I'm hoping someone can help me with quandry - I'm trying to figure out
the best way to analyze a study.
I have a PET study, with two (unbalanced, N=21 and N=23) groups. Each
subject was scanned (PET) before and after treatment. Initially, I modeled
this using a full factorial, with only group and treatment factors. I
specified that the condition should be a dependent factor. Thus, the
resultant covariance matrix had compound symmetry. It was my understanding
that the degrees of freedom would be adjusted to reflect the dependencies
between scans, but the effective degrees of freedom (SPM.xX.erdf) is equal
to (2*(21+23))-4 = 84 - which is what I would expect for the degrees of
freedom if there were no dependencies. Does that mean that the degrees of
freedom *aren't* corrected, and I must use a flexible factorial with
subjects as a factor?
I also tried modelling this with a flexible factorial, incorporating
subject as a factor. I have two issues here. First, the covariances
calculated between the before and after treatment images are much smaller
than what was calculated in the full factorial model, which I found to be
very odd, and I'm hoping someone can explain this to me. (I think this was
pointed out in an earlier post:
http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0701&L=SPM&P=R47197).
Which covariances are correct?
Also, I can't seem to enter many of the contrasts I'd like. For
example, for the group contrast, since I have 21 in group 1 and 23 in group
2, the first 44 elements in the contrast would be:
[ones(1,21)/21 ones(1,23)/23]
(ala Jan Glascher's very helpful guide to contrasts)
However, I get an error - because when matlab evaluates all those 1/21
entries minus all the 1/23 entries, it produces something not quite zero -
because of rounding errrors - since the contrast elements add up to 1E-15
instead of 0, I'm told the contrast in unestimable. Does anyone know of a
way around this?
Thanks for all your help!
Allison
|