I have always kept mine below -3dBfs for anything going to CD. If i'm mixing using the likes of ableton then -12dB is reference. If i'm recording from an analog source again -12dB. Once recorded, peak levels to -3dbfs.
It has been well known for a long time in mastering circles(the ones who do not adhere to the loudness wars) that peak levels should be kept below -3dBfs. This is due to the fact the most consumer level DA will distort above that level. There are a number of stages in the circuit that will distort inside a DA, the converter chip, analog gain stage. It is the intersample peak that causes this
This is made worse when we convert to MP3, convert from WAV to MP3 rates above -3dbfs distortion and pitch changes occur in the MP3 file.
Bob Katz used to go on about this years ago, they would be endless posts on the subject on the old George Massenburg forums. Neilson was probably the authority and you can read some of his papers below.
Stop Counting Samples
0 dBFS+ Levels in Digital Mastering
by Soren Nielsen & Thomas Lund