I would like to conduct an ROI decoding analysis using SPM12 and TDT. I generated some spherical ROI masks using peak coordinates from previous studies (MNI template). If I understand correctly, I need to do the ROI analysis on data in native space (i.e., before normalization + smoothing). I am not sure if I can use these spherical masks simply after resampling them to the brain size of the data? If the data are not normalized, how can I make sure that the masks (based on MNI coordinates) correspond to the same regions in the native space?
Yes, you are correct: if the ROIs are in the same space as your functional data, then you are good. If the ROIs are not, then you would either use inverse normalisation (you can find out how this works in the SPM manual, I guess), or you could simply conduct analyses in normalised space. I don’t see anything wrong with doing analyses on normalised data, it typically just takes longer because there tend to be more in-brain voxels. I like to do things in native space but I think it’s totally fine to work in MNI space.