Hello SPM list-
I am attempting to use SPM8 for a classical 1st-level analysis of an
fMRI experiment with multiple subjects and multiple sessions per
subject (where sessions are scanning blocks separated by a minute or
so). My understanding of the appropriate way to do this is to generate
beta estimates separately for each session, then average the images
together across the sessions for each subject, and using the averaged
images for higher-level analysis.
I am running into difficulty at the model specification and estimation
stages, though. My main concern is that I'm not certain what the
difference is between using a separate SPM.mat for each session, for
each subject, or one for all sessions/subjects. My initial
understanding was that it won't matter (as long as I can keep track of
all the files), since every time I put in a new "Subject/Session" in
the model specification new columns are added to the design matrix
that are not shared with the other subjects/sessions, and I'm not
globally normalizing things. So, based on my understanding of the
design matrix, the beta estimates should be independent of one another
and the same as if I did a separate SPM.mat for each session. However,
that does not appear to be the case. My first clue was that my trial
run processing one session worked fine, but when I made a design
matrix covering all subjects and sessions, MATLAB used all available
memory and quit-- it didn't simply take linearly more time as I
expected. Reducing the sessions in the matrix lets it complete, but I
get different beta values than if each session has its own matrix.
So what is the difference here? My guess is that there is some aspect
of the ReML algorithm (which I don't know much about) that is taking
into account global information outside of what is encoded in the
design matrix. If that is the case, what is the appropriate way to
group things?
Thanks for any help.
-Sam Wintermute
|