Dear all,
thanks for the replies.
Here is my recap of the last few points:
1. As most of you said, the FWER can be affected by this and it need not go down:
According to Gaussian RFT, FWER is proportional to the number of voxels in the volume divided by the product of the smoothness in the 3 directions (i.e. x, y and z).
What I am doing is reducing V (the numerator), but also changing the denominator.
So ... the effect could go either way. Furthermore using a GM mask can be motivated in cases where we are not interested in performing hypothesis testing in what I call irrelevant voxels anyways.
Also as Bruno pointed out this would not affect the non-parametric tests (e.g. using SnPM) and perhaps also FDR.
2. I agree with the last point that activations in irrelevant regions might be a hint towards poor pre-processing or GLM in some subjects at least.
3. I totally agree with the point about smoothing.
Bests
Hamed