You can theoretically take either betas or correlation coefficients to the group level (though in the case of the latter, you would probably want to z-score). As I understand it, it is more common to do group statistics over individual subject betas. See the bold text in Gang's post here: [afni.nimh.nih.gov]

Regarding the issue of understanding directionality, I've used the approach you describe (3dmaskave over a region, extracting betas in my case) and have interpreted positive vs. negative correlations on the basis of the sign of the fit coefficient. However, if you would prefer to do this with z-scored correlation coefficients, I don't personally see any reason you can't. Alternatively, you could consider running a direct comparison over two different PPI beta maps (e.g., run 3dttest++ on the beta maps for factor A vs. factor B).

Hope this helps! I've only recently run PPI analyses, so someone else please reply if this isn't accurate.