Dear Nia,
>1. realign, normalise and smooth my data in SPM5.
>2. perform a random effects analysis to decide on
>coordinates to use within my chosen brain regions.
On a general note, since sem model design is necessarily hypothesis driven,
it should be possible to specify your regions and plan the network in
advance. The permitted moderator variables should also be specifiable in
advance. Over-elaborate models should be avoided, because of the risk of
under-identification, and the problems associated with loops or reciprocal
connections. Well motivated attempts to incorporate all possible
'SPM-blobs' into a network model may create suboptimal or even invalid
models, that are less able to test one's prior hypotheses.
More specifically, I would suggest choosing individual subjects' timeseries
data based on a F-contrast of all effects of interest for each
subject. This will pick out voxels in your planned regions that are
related to the tasks in some way, without specifying a particular contrast
vector that might bias the selection of voxels that are only
differentially active for one aspect of your design. It will also
preferentially select GM voxels, since CSF and WM would not normally be
activated by you experimental task. You can also correct for motion
artifacts, physiological noise or spikes if these are represented by other
regressors in your DM.
>3. perform a fixed effects analysis and extract data
>from the coordinates chosen in the random effects
>analysis.
This is a hard way to get the timeseries'. You could try something like
Marsbar to pull out time series for voxels of interest, or if you use the
SEM toolbox written by Christian Buechel (cited in his papers, also those
of mine which have used sem, amongst others) it includes a function called
spm_sem_regions that allows one to specify multiple regions and pull out
their time series representing the first eigenvariate or the mean or the
peak voxel's timeseries. It would avoid having to repeat a large fixed
effects analysis, and keeps all the data together in simple .mat files.
>4. for each subject, take an average of all the time
>series within each region of interest to analyse.
>5. for each brain region, perform principal component
>analysis on a matrix containing the time series for
>each subject.
There are risks in using the first component. With good quality scans from
still subjects and smoothed data I have found that it makes no meaningful
difference whether one uses the first eigenvaraite, the mean of a region or
peak voxel. I have preferred the first eigenvariate for a variety of reasons.
However, if you have motion artifacts, occasional spikes or other
artifacts, these might be large enough to dominate your first eigenvariate
and relegate you experimental effects to the second or third components.
>6. take the first principal component as an input to
>AMOS to perform structural equation modeling.
I have not used AMOS. I have extracted timeseries and used LISREL, but
found the same results were much easier to reach with an SEM toolbox!
Good luck,
James
|