I obtained fmri data from the Human Connectome Project. For each of the fmri tasks (nback, gambling, social, emotion) I set up a first level model, estimated the beta parameters and computed one-sample T-images (e.g. 2back, win, loss, social, etc.). For any further analyses it would be important that the statistical images are comparable across different tasks. I am not sure if one can simply interpret the values as is, that is, if voxel 123 has a value of 2 in the first statistical image and the value of 1 in the other that the assoication between activation and condition is twice as high? Is there any standard procedure for this type of problem?
Beta parameters represent effect sizes , which should correspond to BOLD increase wrt baseline. Statistical maps represent the same effect, but divided by the noise level, so that it represents the significance of the effect. The statistic thus represents a normalized effect size.
If you want to make a statement on the difference between conditions, you need to do a formal test (contrast) between these images. This makes more sense if you do that across subjects: a paired t test will allow you to conclude that there is more activity in a given location for contrast A vs contrast B.
But would such a contrast test be valid when I haven’t modeled the beta images in one single first-level model? I am struggling a little bit with the interpretation of ‘baseline’. When I model each task in a separate model, isn’t each beta value (or T-value if I want to ‘average betas’) automatically only interpretable relative to its ‘own’ baseline?