This is a question that many amateurs ponder. There is a lot of misinformation around as of mid 2010. One school of thought says that the signal level should be approximately 1/3 of full well, while others teach a higher level is needed.
In addition, flats take time to shoot, so how many are really needed?
This report shows experimental data taken using three different signal levels of flats, 80%, 33%, and 22.5% with from 1 to 30 flats combined.
The results are that the same noise is observed using Photon Transfer techniques for 10 flats of 80% versus 20 flats of 33% versus 30 flats of 22.5 % of full well
By shooting the flats at higher signal level fewer flats are needed to attain a given noise level in the calibrated image