What's the best way to record?

A recent task was carried out to try and answer the following question:
What is the best way to record - low gain setting to minimise hiss from the pre-amp and to provide a large amount of headroom? Or, a higher gain setting to give a higher signal level, less headroom and higher noise floor?

The following equipment was used:
Laptop running Sony Soundforge 7
Edirol UA25 USB Powered 24bit/96kHz Audio Interface
Sennheiser K6 Shotgun Condenser Microphone
Beyer DT100 Headphones
Duvet to help minimise reflections and "room noise"

With Soundforge set to record at standard CD-A quality (i.e. fs=44.1kHz and bit depth = 16), the microphone was connected to the audio interface and Phantom Power was switched on (on the Edirol interface). The microphone and subject were placed under a duvet(!) to help reduce background noise and reflections so that the noise floor was as low as possible; the subject was approx. 30cm from the microphone. With Soundforge recording, the gain setting on the audio interface was firstly set to its minimum position and a count was carried out at normal conversation level, then the same was carried out for a gain of approx. 9 o'clock, 12 o'clock (where clipping started to occur) and finally 3 o'clock which caused severe clipping. Each recording was then analysed (using the Statistics tool) for the peak sample value, RMS level (important as the human hear responds to average level, not peak) and approximate noise floor level (as observed on the dBFS meters) for both non-normalised and normalised versions.

The results are were as follows:
Gain of zero (minimum position)
non-normalised peak sample value: -24.07dBFS
non-normalised RMS value: -39.09dB
non-normalised noise floor level: -54dBFS (approx.)

normalised peak sample value: 0dBFS
normalised RMS value: -16.97dB
normalised noise floor level: -37dBFS (approx.)

Gain set to 9 o'clock position
non-normalised peak sample value: -8.3dBFS
non-normalised RMS value: -27.62dB
non-normalised noise floor level: -54dBFS

normalised peak sample value: 0dBFS
normalised RMS value: -20.33dB
normalised noise floor level: -47dBFS

Gain set to 12 o'clock position
non-normalised peak sample value: -1.24dBFS
non-normalised RMS value: -18.2dB
non-normalised noise floor level: -40dBFS

normalised peak sample value: 0dBFS
normalised RMS value: -17.81dB
normalised noise floor level: -40dBFS

Observations and Conclusions
It must be appreciated that some of these results are only approximate observations. In brief, it can be seen that the best compromise appears to be with the gain setting (in this case) to the 9 o'clock position which gave a good amount of headroom (peak, non-normalised value of -8.3dBFS) with a low level noise floor. Normalising these values gives a S/N ratio of approx. 47dB which isn't too bad considering the acoustic environment of the space used. It is also interesting to note that at a glance, the results aren't that much better than the results for the previous post entitled Quick VX2000 Audio Test Results, however, audibly the recordings were far far superior - much cleaner and with far less noise. This backs up the advantage of using separate high quality recording devices when recording sound for picture.