Dear Brigitte,
> After huge digs in the SPM archives I still didn't find the answers to
> some of my questions...Could you please reexplain to me:
>
> 1. What are the theoretical and practical differences of conjunction
> analyse between spm96 and 99?
Conjunctions in SPM96 were based on getting a significant main effect
(averaged over all the contrasts that entered into the conjunction) in
the absense of any interactions among the contrasts. This was
suboptimal because it relied on accepting the null hypothesis of 'no
interactions'.
Conjunctions in SPM99 are tested by ensuring all the contrasts are
jointly significant. The conjunction SPM would look similar to the SPM
that obtains from exclusively masking all the contrasts with theselves
to reveal the common areas or intersections. The t values are the
smallest among all the component SPM(T) values and the conjunction SPM
is a 'minimum T-field'. Gaussian field theory is now applied to this
SPM{minimum T} to give corrected p values. We could not do this in
SPM96 because the minimum T-field theory had not been devised.
> 2. Why and when should I use second level analyses with PET? If, for
> instance, I want to compare flipped with non-flipped images in one
> group of subjects, I think I can do that without second level analyses.
> If I want to compare the difference between flipped and non-flipped in
> one group compared to another group of subjects it seems to me I can
> still do this without second level analyses but I think in one of the
> mails of the archive I read that you should do that with second level
> analyses, why?
The same applies to PET and fMRI. First-level analyses use
within-subject variance and provide for inferences that generalize to
the subjects studied. Second-level analyses use
between-subject/session differences whose estimator is the correct mix
of within and between-subject error to give a Mixed-effect (i.e.
Random) analysis. This allows you to generalise to the population from
which the subjects came.
In PET (but not fMRI) the similarity of between- and within-subject
variances and between the number of scans per subject and the number of
subjects means that the difference between first and second-level
analyses are much much less severe. Traditionally PET studies are
analysed at the first level. Second-level analyses are usually
employed when you want to make an inference about group differences
given some within-subject replications. In your example you would
collaspe repeated measures of flipped vs. unflipped into one contrast
per subject and then compare the contrasts at the second level. This
would be less sensitive but would allow you to generalise to the
populations from which the subjects came.
I hope this helps - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|