Hi Sandra,
so you reduce the 2 thickness measures within subject to a rate of change (mm/time), then you have a single map for each subject and now your question is about your glm modeling. So the question is basically about cross sectional modeling in you GLM.
Your equ 7 and 8 test if there is an overall effect (thinning or thickening) not considering any co-variates.
Your other three models 1,3,5 include different amounts of covariates in the model. Of course you will get different results as you are basically 'regressing out' these variables, when placing a zero in the contrast. The variables are used to build the model and fit the functions, but not for the testing.
Now let's look at model 5. You have two groups, scanner 1 and scanner 2. The constrast [1 1] tests if the rate in group 1 plus the rate in group 2 is different from zero. In order to better interpret the gamma map, I'd probably do [ 0.5 0.5 ] to have gamma show the average rate between these two groups, sig won't change. You could also test [ 1 -1 ] to test if there is a difference between the groups (to see if the scanner has an effect on the atrophy rates, assuming groups are matched, or when including (regressing out) appropriate co-variates such as age, gender, etc).
So to answer your question: the different models should not give the same results as they test different things. If you get very different results by including / removing a co-variable, that may mean that that variable is important for your analysis. There is a large body of work on parameter selection. I am not a statistician, so not the best person to discuss this with.
Hope my explanations already help. Best, Martin
On 01/22/2014 05:40 AM, Sandra Preissler wrote:
Dear freesurfer experts and users
At the moment our group is working on longitudinal analyses for two different time points in one group. Our analyses worked well and we tried to calculate the effects of six different covariates and one fix moderator variable (we had two different MRI Scanner and tried to consider this effect) with the function mri_glmfit. We did pre-processing with mri_preproc:
mris_preproc --fsgd extra_komplett.fsgd --cache-in long.thickness-rate.fwhm10.fsaverage --target fsaverage --hemi lh --out lh.extra_komplett.thickness.10.mgh
With the fsgd-File designed like the following example:
GroupDescriptorFile 1
Title G1V2
Class Scannertyp-Main
Variables z1 z2 z3 z4 z5 z6
Input P01 Scannertyp-Main T3 z11 z21 z31 z41 z51 z61
...
Because of the consideration of the variable Scanner we designed the design matrix on our own. Before calculating the effect of the covariates we calculated the main effect (for the assumed underlying regression model please refer to equation number 1 in the attached pdf). For doing this we built a contrast file (please refer to equation 2). In a longer discussion we decided to delete different covariates and get the regression model shown in equation 3 (and the contrast file for the main effect in equation 4). Surprisingly both main effects not even were similar. Therefore we tried two different models (please refer to equation 5 and 7 for the two regression models). If the underlying model is a normal linear model the first three Contrasts and regression models should gave the same results. In our analyses we have really different results, therefore we have problems in interpreting and understanding the underlying general linear model. There seems to be still an influence of the covariates even if there are modeled with 0 (which in our understanding means they were mathematically not included in the model). Is the probability dependent on the distribution of the covariates even if they are modeled with "no influence"? If so, how is the belonging regression model?
The information about the freesurfer version and linux version used are: FREESURFER_HOME: /opt/freesurfer Build stamp: freesurfer-Linux-centos4_x86_64-stable-pub-v5.1.0 Kernel info: Linux 2.6.34.10-0.6-default x86_64
Can you help us in fixing this "theoretical" problem?
Thanks in advance!
Sandra
This message was sent through https://webmail.uni-jena.de
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer