Here is an open online discussion.
Common average is preferred if you have high-density recording (64 channels or above). But there is a compatibility issue about comparing the results to what has been published before in the literature. One could do a direct comparison with the different approaches (See Zhang et al. (2011) for an example on infant auditory EEG research.).
There are some reference-free approaches using source localization. One such approach involves calculating EEG lead field matrix by assuming a certain number of discrete dipole sources (say, 3000) distributed across the spherical brain space; with this approach, it begs the question: why not calculate the activities directly in the cortical sites of interest? After all, the electrode positions on the scalp are not the cortical/subcortical sites processing the information.
Here is a link for insightful discussion about the EEG reference method issue:
http://www.nitrc.org/forum/forum.php?thread_id=3094&forum_id=2