Hi Gang,

I ran the analysis, and it looks like study/scanner has a big impact everywhere (voxelwise p < .005):



This map looks too meaningful, and I want to wrap my head around it.

To give some background on the processing stream, I followed Ashburner's 2010 VBM tutorial (warning: PDF), which specifies the following preprocessing steps:

• Segment the images into GM, WM and CSF
• Create a group GM/WM template by iterated nonlinear warping of each subject's a posteriori GM/WM probabilities to the group mean
• Warp each subject's GM probabilities to match the template, dividing by the determinant of the Jacobian to account for volume changes in the warp
• Blur the resulting images with an isotropic 10mm FWHM gaussian filter

Then I analyzed the data with 3dLME, using a model which I'm not sure was appropriate:

3dLME -mask mask+tlrc -prefix LME -model 'Group*Age+Study+ICV' \
    -qVars Age,ICV -ranEff \~1 -SS_type 3 -num_glt 3 \
    -gltLabel 1 AA-BB -gltCode 1 Group : '1*AA' '-1*BB' \
    -gltLabel 2 second-first -gltCode 2 Study : '1*second' '-1*first' \
    -gltLabel 3 Age -gltCode 3 'Age :' -dataTable \
    Subj    Group   Age     Study   ICV     InputFile            \
    001     BB      35      first   1.42373 smwc1s-yy-BB-001.nii \
    004     AA      40      first   1.30568 smwc1s-yy-AA-004.nii \
    005     AA      48      first   1.28727 smwc1s-yy-AA-005.nii \
    007     AA      42      first   1.44345 smwc1s-yy-AA-007.nii \
    ...

In all, there were 117 subjects in the first study and 24 subjects in the second, with 23 subjects in both (i.e., only one subject in the second study was not in the first).

I don't feel like I fully understand how to specify the variance model in 3dLME, and I'm wondering if I've done something dumb with it. Thoughts?

Isaac