The "more correct" procedure would be to export your 'master' as .wav - for ex., I commonly export to 44.1 KHz/24-bit wav so that I have the highest resolution possible for my masters (you can also export to 48, 96 or even 192 KHz if that is your preference). If the Sonar project was recorded at 44.1 KHz 24-bit, I do not apply dither when generating the 44.1/24 wave master file, either through bouncing to a new track or exporting to wave. Dither will be applied when I convert the 24-bit file to 16-bit for cd burning. There is a time when I apply dither during exporting to 24-bit wav, and that is when the master will be converted using an external program such as r8brain (simply because r8brain does not have a dither function). When using editor programs such as Wavelab or Soundforge, I may apply their UV22 dither algorithm during sample rate conversions (I sometimes use these programs for conversion to .wav or .wma. as I find them more efficient for this purpose). Rule of thumb: it is better to record to 44/24 especially if the end product will be for cd replication.
For .wma or .mp3 encoding, I prefer to use the 16- or 24-bit master file since this has the highest possible resolution PRIOR to encoding, which is very important for the overall clarity of the file. I never export from Sonar direct to wma/mp3, then render the wma/mp3 back to .wav since there is already a conversion error when doing that procedure, specifically, certain frequencies are 'lost' because of the wma/mp3 encoding procedure.