A team from the Berkeley Seismological Lab (plus a summer intern at LLNL) have completed a review of the first motion data for the event and have derived a mechanism which differs significantly from the automated solution derived shortly after the quake. Importantly, they calculated a solution which allows for volume change at the energy source, which the automated solvers disallow.

They've concluded that the observed waveforms are most consistent with an "anti-crack", which is to say the reduction of a long, thin gap.

The most relevant figure from their work:

Figure 6. Source type plot based on Bowers and Hudson (1999). In this plot earthquake, explosion and collapse data from Ford et al. (2007) is compared to the August 6th event (red star). The August 6th event plots in the general moment tensor space that defines an closing crack, or collapse. The event is located well outside the region occupied by tectonic earthquakes.

Link to the full report (omitted from my comment above):

Seismic Moment Tensor Report for the 06 Aug 2007, M3.9 Seismic event in central Utah

Prepared by: Sean Ford, Douglas Dreger, Peggy Hellweg and Bob Uhrhammer

They additionally model the energy released at 1.92*10^22 dyne*cm (about 2*10^15 joules for the SI world) which is roughly ~500kilotons of TNT. That corresponds to a moment magnitude of 4.2. The magnitude of 3.9 that has been widely reported is the assigned Local Magnitude, which I assume the USGS has been reporting due to the initial difficulties in calculating the mechanism and the associated uncertainty in the moment...

It should also be noted that although the conclusion that this was a collapse-type event is fairly robust, not too much detail about the exact motion or the exact mechanism of stress release can be developed with this sort of analysis.

We can't, for example, say where in a modeled cavity the collapse initiated and in what direction it propagated. Given engineering data from the mine and a detailed model of the propagation paths to the stations which recorded the events, a priori models of the collapse could be evaluated against the observed waveforms.

None of these exercises can provide much information of use to the rescuers, unfortunately :( Here's hoping there's promising news from the drill team soon...

I heard on TV the epicenter of the quake was about 1000 feet below surface and near the back end of the mine. A natural earthquake epicenter was usually from a much deeper level. It seems the roof caved in on top of the miners. They are drilling to determine if there is some airpocket or intact chamber down there.

Some of this information suggests that one of the overlying beds, that may not have dropped earlier, suddenly broke and dropped down onto the underlying rock pile. That hanging up, and then failing to drop over a large area is what might happen if there were a thick sandstone layer in the roof, but if the cavity transferred upwards I would have thought that it might have been detected in the drilling activity, since intersection with a cavity would lead to lost circulation in the drilling fluid. (Though I don't know what type of drill they are using).