Dear Nia,
I would recommend to fit your SEM model to data from each subject
individually, instead of extracting a group time series from a fixed
effects analysis where the chosen voxel may be driven by a few or
even a single outlier. Make sure you standardise the selection of
voxels such that they are anatomically and functionally comparable
across subjects. After fitting the model to each subject, enter the
parameters into a 2nd level analysis - see papers by James Rowe for
examples of this - and test for significant differences, e.g. using a
two-sample t-test, separately for each parameter of interest. This
should give you a better insight into whether the two groups are
different or not.
Best wishes,
Klaas
At 15:35 20/10/2006, you wrote:
>Dear Klaas
>
>I have data for two control groups performing the same task. I have
>been trying
>to perform a connectivity analysis using structural equation modelling but I
>don't get the same result from both groups. I'd appreciate it very much if you
>could tell me if I'm doing anything wrong.
>
>The TR used is 2.1s with a 1.5T scanner and the voxel size is 3.5mm x 3.5mm x
>4.5mm with a 0.5mm gap. The task is an nback task with a block length of 60
>seconds.
>
>First of all I realign the images. Should I be using the resliced images from
>this step? I then normalise and smooth with a FWHM of twice the voxel size, so
>7 x 7 x 10mm.
>
>The high pass filter that I apply is 1.5 times the difference between block
>onsets, so 180 seconds.
>
>Having chosen the regions that I want to include in the model, I use a random
>effects analysis to choose the most appropriate voxels. I then perform a fixed
>effects analysis and extract the data from there, using an F contrast and
>adjusting for effects of interest. Would individual subject analyses be more
>appropriate? For each region, I combine the data for each subject
>into a matrix
>and perform principal component analysis. I then use the first principal
>component for each region as input for structural equation modelling. How is
>the principal component analysis usually performed? There are
>different ways of
>performing this analysis so which is most appropriate? Is it necessary to
>transform the data in some way in order to satisfy the assumptions that
>structural equation modelling makes?
>
>Is there something that I can do differently to produce more
>consistent results?
>Should I change the task or use more subjects? Is it a problem with the voxels
>that I'm using? How do other researchers control for subject variability and
>get consistent results? How do I know what sort of results I should
>be getting?
>
>Any help would be very much appreciated.
>
>Many thanks
>Nia
|