Optimal group analysis approach

AFNI Message Board
1

Dear AFNIites,

I am looking to run voxelwise regression models of PET average signal (Y) as a function of a number of covariates, say (age, sex, b0...bm blood biomarkers and k0..ks clinical scales). 3dRegAna looks like the most effective way to run this but give that I have ~1000 subjects, the command line will be rather unwieldy, and specifying certain contrasts such as combinations of clinical scales hard to decipher with 's' on the order of 30.

In your expert opinion, is 3dRegAna the best way to do this despite the above or am I better off with 3dMVM despite the slower execution time?

Cheers,
Z

code text  # or delete if not needed
2

Z,

Could you provide some clarification on a few points?

** Am I correct in assuming that there are no within-individual variables?

** Are both blood biomarkers and clinical scales considered quantitative variables?

** Do you think that age and sex may influence blood biomarkers and clinical scales?

** Could you specify the effects of interest that you intend to estimate from the analysis?

Gang

3

Hi GC!

No within individual variables.
Blood biomarkers and clinical scales are quantitative
Age and sex may have influence on blood biomarkers and clinical scales
I would like to assess which brain regions associate with each of the clinical scales, and for interactions with sex

Thanks!
Z

4

Z,

I would like to assess which brain regions associate with each of the clinical scales, and for interactions with sex

Does "interactions with sex" refer to the interactions between sex and the mentioned clinical scales?

Setting up the model using 3dMVM is probably the simplest approach:

3dMVM  -prefix ... -jobs ...                                   \
  -bsVars  'k0*sex+k1*sex+...+ks*sex+b0+b1+...+bm+age'         \
  -num_glt ?                                                   \
  -gltLabel 1 k0 -gltCode  1 'sex : 0.5*male +0.5*female k0 :' \
  ...
  -gltLabel ? ks -gltCode  ? 'sex : 0.5*male +0.5*female ks :' \
  -gltLabel ? k0v -gltCode ? 'sex : 1*male -1*female k0 :'     \
  ...
  -gltLabel ? ksv -gltCode ? 'sex : 1*male -1*female ks :'     \
  -dataTable @table

If you suspect the presence of interactions between sex and the blood biomarkers, it's worth considering their inclusion in the model.

If memory capacity is an issue, you can use 3dZcutup to divide the data into slices and then combine the outcomes using 3dZcat.

Gang