Hello FSL Experts,
I am interested in running a 3 group analysis using cortical thickness files (.gii) derived and converted from Freesurfer format to GIFTI. Specifically, with recent papers regarding issues with GRF theory estimations of noise for clusterwise analysis (gaussian distribution vs gaussian + heavy tails), I am interested in trying out FSL's permutation based testing to see how results compare. As I am new to using the stats packages on here, any guidance would be useful. Does FSL handle surface data, and if so is GIFTI an acceptable format?
Specifically, I want to compare cortical thickness across controls and 2 patient groups. For my covariates there are 2 fixed factors (group membership and gender), as well as 3 quantitative measures (Age, Years of Education and disease severity...aka DS). I am interested in DODS (different offset different slope to borrow from Freesurfer terminology). The model I am interested in is the following : Group*Gender+Age+YrsEdu+DS where * specifies an interaction term.
I want to perform an F-test and then second level ttest. If only the latter is available (ttest), then I would like to make sure centering/demeaning is handled correctly. From my understanding, the ttest is the second level test and so centering should be performed across all three groups for quantitative variables (Age, Yrs Edu and DS), and not centered just using the two groups in the comparison (which may not be statistically valid). Furthermore, fixed factors such as group and gender should not have "centering" associated with it, since this does not make sense. I know some software packages suggest dummy coding (-1 female +1 male) in order to approximate a zero center mean, however my three groups are unequal in N (40,34,75) as well as gender breakdowns. Also, I would rather have it treated separately for fixed factors as opposed to dummy coding if this is possible. Finally, I would like to run this using permutation based simulations which can approximate the false positive rate < 5%, since it uses the data itself instead of making assumptions regarding the smoothness of the residuals from the GLM. One note is that when I tried using Freesurfer's permutation function I received and error that my covariates were not orthogonal. This may be due to the fact that I manually centered my covariates across 3 groups, even when comparing just two groups, and so the means were not 0 for all covariates when comparing just two groups....I am not sure if this could be related to the error or not. Another potential reason is that two of my covariates may somehow be partially correlated and therefore not orthogonal.
Which program is recommended to perform all of the above, including the 3 group comparison. I am guessing randomise or PALM might be the answer, however with such a complicated model, I was not sure how to design all of the files needed. Most of the examples I found online for FSL was in regards to a two group comparison. Any help would be much appreciated including which tool to use, how to specify the design matrix, parameters in FSL convention and maybe a sample line command/design matrix etc to get started.
Thanks in advance for all of the help.
Sincerely,
Ajay
|