Hello SPMers,
As we work through our first try of the optimized protocol of VBM, we have
run into many questions and we very much appreciate the help we've received
so far. Our questions this time concern the statistical analyses.
We are analyzing normal variability in frontal lobe gray matter. We have
followed the optimized protocol described by Good et al, and created our
study-specific whole brain template based on 20 of our 45 total subjects,
and then our gray and white matter templates based on normalizing all 45
subjects to our study-specific whole brain template.
One thing we are analyzing, to be sure we understand SPM statistics, is
simply right hemishpere vs left hemisphere (hemispheric asymmetry). Our
hypothesis, of course, is that right frontal lobes should be larger.
However, when we run analyses, we cannot find this difference.
The Good paper describes a 2 x 2 x 2 factorial analysis of
flipped/unflipped x sex x handedness and we do not understand how to set up
factorial analyses in SPM. (If you can tell us how, we would very much
appreciate it!) So, we decided just to look at hemispheric asymmetry
without accounting for sex (all our subjects were right-handed, so this was
not an issue for us). Our design matrix was simply: Group 1, unflipped
images, Group 2, flipped images, and the covariate (Basic models, AnCova)
was global mean intensity. We find absolutely no differences in right
hemisphere vs left hemisphere no matter how liberal we set the p values.
We obviously are not understanding something about our design or results
process. Can anybody give us suggestions?
Some reasons we have considered for finding no difference are the following:
1) When we created our template, we smoothed with an 8mm kernel. Should we
have used a smaller kernel?
2) In our final smoothing, we smoothed with a 12mm kernal. Again, should
it have been smaller?
3) This question is more theoretical: Because our group is so small (N=45)
and we created our template from 20, is it possible that all the brains, in
the process of normalization, were simply warped to be too similar, washing
out any differences that may exist? We do not fully understand how one can
expect to find variability among a group when that group was used to create
the template to which each image was normalized. We understand the value
of creating a study-specific template so that you are sure your differences
aren't just due to a different population being used for the template, but,
theoretically, how are differences among subgroups of normals to be found
when those very normals are used for the template? Clarification of this
issue would be very much appreciated. Is it a matter of sample size, and
having enough variability due to a large number of individual brain images?
Thank you very much for considering our questions and for your advice on
our analyses.
Sincerely,
Caitlin Fausey
Maria Vittoria Spampinato, M.D.
|