Most people report group statistics that are based on the mean of non-thresholded datasets. In such analyses, thresholding is performed based on the overall group statistics. Of course, for this to work, one needs quite a few subjects.

For smaller samples, the question becomes one of what is the question being asked? If it is "where do my 5 subjects show common activation, and in those regions what is the mean percent signal change?", then the appropriate thing would be to do a conjunction. i.e. only calculate the mean percent signal change for voxels which have significant activation for all subjects (i.e. a non-zero percent signal change in your case where you've already thresholded and set non-sig. voxels to 0). With only 5 subjects in the analysis, this is probably the most sensible course of action*

In any case, it would be incorrect to include 0.0 values that have been set to 0.0 because they do not survive thresholding. Just because those voxels/subjects did not survive thresholding, does not mean that there was no signal change - if you're going to include them, you should really include the original values.

*IMHO