Questions about "Decoding with SpaceNet: face vs house object recognition"

The SpaceNet is a great method that improve the interpretation of the MVPA. However, I have a question for this. One activation map (i.e., weights or a set of brain regions) for each subject. And, the number of brain regions for each subject may be not same theoretically. So, how to perform some group test?
Also, how to perform the group analysis on the “anova+svm”?

I am a bit uneasy in replying to such a question, because it feels that you are employing concepts from standard analysis and trying to project them to multivariate methods. Standard analysis affects a p-value to each voxel. Decoding does not. Hence the notion of group test is not clearly defined. At least not a the level of a voxel.

If you really need a test at the level of a voxel, you could run a SpaceNet model on each subject, and then do a standard 2nd level analysis, using an FDR to correct for multiple comparison. The interpretation of such analysis would be that you are testing for voxel that appear in common in the decoding patterns across subjects.

Thanks for your kind reply. It is helpful. Further questions I have based on your reply.
1st, I looked through the script of SpaceNet regression, and found that this analysis was performed with all 16 subjects. Then the weight for voxels were plotted in the figure. Is this correct: voxels/brain regions were related to the behavioral index/stimuli as long as voxels/brain regions were specified with a weight in the figure?
2nd for the 2nd paragraph. Did you mean I can use whole brain weights to perform the group analysis?
Looking forward your kind reply.

For your first question: the example in nilearn is on a dataset where there isn’t enough data to perform an analysis at the subject level. Hence the analysis is done by pooling subjects. If you can perform trial-level decoding and have enough trials per subjects you can run an analysis per subject and then do a second-level analysis.

For the second-level analysis, you can use whole-brain weights.