Dear statistics specialists,
we are struggling with an analysis of ICA-maps that we would like to compare
across four different conditions (similarly, two independent groups of such
maps have been compared in SPM using a 2nd level two-sample t-test, see
Greicius et al. Biol Psych. 2006).
Stack sizes would be 28, 24, 24 and 20 maps per level of the condition of
interest.
The problem now is that these measurements are
- in part dependent as they stem from the same subjects across levels;
however, not 100% to make a real repeated measurements design.
- contain repeated measurements of some subjects within one level.
- example:
level 1: subjects A A B C D E F F G
level 2: subjects A B C D D E H I J K K
level 3: subjects A C C D E H L
and so on.
We now set up a one-factorial design, assuming that (a) independency is not
fulfilled and (b) variances are unequal. Subject-IDs were not encoded
explicitly so far.
We found a reference that Greenhouse-Geisser-correction could (somewhat) be
capable of adjusting the DFs in a situation of multisample sphericity
violation (if groups are about equally sized) - would the methods
implemented in SPM5 be comparable to Greenhouse Geisser?)
We would highly appreciate your advice on what model would be adequate here,
or if the way chosen is acceptable.
Many thanks in advance for opinions,
Philipp Saemann
Michael Czisch
|