Hello FSL experts,
I have two competing GLM models for a higher-level analysis. In both cases, I have a clinical and a neurotypical group for which variances are calculated separately. One model uses mean, age, and IQ as predictors (with each regressor specified independently by group) and the more complex model adds a measure of social functioning as an additional predictor. When social functioning is added as a predictor, a significant group difference in mean response to the experimental contrast of interest emerges. What I would like to do is be able to choose between these models in a principled way. All my higher level analyses were conducted using flameo.
My initial thinking was to calculate an R-squared map for each model and compare the two, using:
fslmaths res4d.nii.gz -Tstd -sqr var_res4d
fslmaths ../cope3_4D.nii.gz -Tstd -sqr var_cope3_4D
fslmaths var_res4d.nii.gz -div var_cope3_4D.nii.gz VarRes_div_VarCope
fslmaths var_res4d.nii.gz -bin ResMask1
fslmaths ResMask1.nii.gz -sub VarRes_div_VarCope Rsquared
(Which I'm not 100% sure would be right in and of itself, but...)
I'm not sure that this appropriately handles the fact that the two models estimate separate variances for the two groups.
What suggestions do you have for model comparison/model selection on higher-level models, particularly when you have separate estimates of variance by group?
Thanks!
Allison
|