Hi everybody,
I want to compare the surface data of 3 groups (GroupA, GroupB and Controlls) but have the problem that they were acquired from 4 different scanner sites. As I can see it, there are three ways how I could tackle this problem:
1. I could use mri_glmfit and create a qdec table / fsgd-file with 12 classes: Class GroupA_site1; Class GroupA_site2,... And then use the contrasts [0.25 0.25 0.25 0.25 0 0 0 0 -0.25 -0.25 -0.25 -0.25] to compare GroupA to the Controlls. My Problem with this approach is, that the sites don't contribute the same amount of subjects to the analysis. I'm not sure if this could be handled by simply using a weighted contrast. Meaning, if Site1 and Site2 had twice as many subjects than Site3 and Site4, I could modify the contrast to [0.33 0.33 0.17 0.17 0 0 0 0 -0.33 -0.33 -0.17 -0.17].
2. I could create dummy variables to account for the variability between sites. In this case, I only need to specify 3 classes (Class GroupA; Class GroupB; Class Controlls) in my fsgd-file. And I use a design matrix that has 4 dummy variables at the end, which specify to which site a subject belongs. This approach might work, but I'm not confident that it is the right one.
3. I could use a mixed effect model approach and specify site as a random effect.
If I understand it correctly, the mixed effect model approach would be the best one, as it accounts for the variability within sites. Is that correct or are there other issues/better approaches?
I tried to implement a mixed effect model by using Bernal's Linear Mixed Effects (LME) Models (http://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffectsModels) but run into some problems. I'm not sure if LME can only be applied on longitudinal data or if my implementation is wrong. I have a design matrix X that specifies the characteristics of each subject per row as follows:
Intercept GroupA GroupB Controll Age IQ Site1 Site2 Site3 Site4 1 1 0 0 11.1 99 0 0 1 0 1 0 1 0 11.1 101 0 0 1 0 1 1 0 0 11.4 95 1 0 0 0 1 0 0 1 12.4 100 1 0 0 0 ...
As I have no repeated measures, 'ni' is just a vector with length X containing '1's. If I do now the vertex-wise linear mixed-effects estimation, I get the following output:
stats = lme_mass_fit_vw(X,[7 8 9 10],Y,ni,lhcortex);
Starting matlabpool using the 'local' profile ... connected to 8 workers.
Starting model fitting at each location ...
Location 24994: Index exceeds matrix dimensions. Location 24994: Algorithm did not converge. Initial and final likelihoods: -10000000000, -10000000000. Location 62484: Index exceeds matrix dimensions. Location 62484: Algorithm did not converge. Initial and final likelihoods: -10000000000, -10000000000. ...
I've checked the matrix dimensions of X, Y, ni and lhcortex and compared them to the LME mass_univariate example stored in ADNI_Long_50sMCI_vs_50cMCI.mat but couldn't find any divergence.
Has anybody encountered similar problems? Is my approach of specifying 'ni' as a vector of'1's even legitimate?
Thanks, Michael