Dear Victoria,
when it comes to (f)MRI I think it is definitely better to exclude potentially problematic data. Subject groups are typically quite small and some affected data set might result in artificially high/low beta estimates, which could have a huge impact on the group results. Concerning head motion, there's quite some research going on how to better control for that issue, but the best solution is you have stable subjects. If you pay attention at the scanner you might already detect fast head motion online and then you could try to reposition the subject / repeat the session (if it doesn't conflict with the nature of the paradigm), you could accustom subjects at the scanner environment, ... Sounds easy in theory, in practice, well ;-)
I never used AFNI, but as far as I understand this SNR calculation depends on specific model settings (if the regressors of interest are modified, the mean baseline estimate might change, I also wonder what the "residual time series" refers to exactly, so does it include signal changes explained by the regressors or not) and the preprocessing (SNR seems to be calculated on the smoothed data), at least to some extent. Thus I would rely on some parameter obtained during your own preprocessing pipeline.
Global normalisation means that all the volumes of a time series can be scaled to some specific value to correct for global signal changes (like scanner drifts) within that session, which traces back to PET studies. However, the scaling might introduce artificial (de)activations. One illustrative example: Say two equally sized regions A and B have values of 100 during baseline, followed by a task introducing strong activations in region A, say 200, and no changes in B (whether this is a realistic assumption is another issue). This results in a task-related increase of the global signal (100 becomes 150 on average). However, with scaling enabled the values during task might be scaled to 133 for A and to 67 for B, resulting in an average of 100 again. Thus activations in A might be underestimated whereas one might falsely discover some deactivations in B. Therefore people have suggested that global scaling should not be used http://dbic.dartmouth.edu/wiki/index.php/Global_Scaling , at least not if the global signal correlates with the expected time courses / "task-related activations". You could check your data with SPMd http://www-personal.umich.edu/~nichols/SPMd/ for example, unfortunately it's not (not fully?) compatible with SPM8 as far as I remember.
If you don't find any correlations I would still go with the default = no scaling, as slow signal drifts should be removed with the high-pass filter anyway. Rapid changes of the global signal on the other hand might reflect heavily distorted data (signal loss within some slices, intensity changes due to spin-history related motion artefacts), in this case a simple scaling procedure won't help.
Best,
Helmut
|