I believe that I may have asked this question sometime in the past, but I would like to re-visit this issue.
When exporting a stereo mix (for mastering, etc.), is it preferable to export a mix [with audio recorded at 24 bit] to 24 bit audio, or to 32 bit float?
My understanding is that Sonar processes a mix in 32 float and then dithers on export. Doing a test with two files exported to 24 and 32 reveals a very slight difference in sound files WAY down below the level of hearing perception, but still, if there is a better way to go here, I would like to hear persons weigh in on this issue.
Thanks.

If you're sending the exported file out for mastering, or doing further mastering yourself, Sonar's dithering should be turned off, as dithering needs to be done on the very last stage of the mastering process.
 
I have had great results exporting to 32 bit, then using the 32 bit file to master from.
If the final objective of your process is a CD, then 24 bit is probably enough . . . then you dither it down to 16 bit.
 . . . but exporting it as 32 bit leaves some flexibility, if a CD is not the intended final outcome.
 
I think it's more important to have the overall levels right when exporting for mastering purposes, as the mastering process needs to have some headroom to actually be able to apply mastering type effects, such as limiting and equalization.

Yeah, use 32 bit float.
 
I do.

If there's going to be further processing then 32-bit float makes sense since 32-bit FP is what most audio processing is done in (or sometimes 64-bit FP). So if you take it down to 24-bit integer you are just rounding off things unnecessarily. With iterative math, which is what audio processing is, you want to maintain precision until you have done all the calculation you are going to do and you have the result, only then do you round.
 
Now does it actually make an audible difference? Probably not, but unless space is a concern just stick with floating point data.

The end format is MP3.
I suppose 32 bit float retains more information and does not dither(?).
I use the LAME encoder from there to create a 44.1khz/256KPS MP3.

Though it likely makes no difference in the real world, 32bit float ensures that there is no clipping (not that there should be anyway, but...) and preserves any low level information. 

I don't use 32 bit float for a file. I can't say if there is any sonic difference due to the fact no audio interface supports it. So all my exported audio is 24 bits. If I wish to go to CD it is 16 bit. 

John, we're not talking about the ability to play a 32 bit file, it's about having the ability to preserve as much detail as possible with virtually no errors in the maths when exporting/bouncing.
 
64 bit is of course even more accurate and I always engage this option.

I understand fully. I just don't see or hear how it matters. In a way this is simply theoretical. What I mean, without a way to monitor this we can only assume it has significance. At least enough to warrant the added bits. I think 24 bits is plenty for most audio processing purposes. Yet 32 bits FP will use up a lot more disk space. If it is a noticeable improvement than it will be worth it. I just have no way to know that for sure. If someone compares the two there is no actual way to make a comparison. Both files will have be at 24 bits to hear them.  
 
I know this is contrary to the ideas CW has been promoting I just have no evidence it actually matters. I also think a lot of the errors we see listed here on this forum is due to systems that really can't do full 32 bit FP on large projects. But that is just a notion I have. 
 
 

I understand fully. I just don't see or hear how it matters. In a way this is simply theoretical. What I mean, without a way to monitor this we can only assume it has significance. At least enough to warrant the added bits. I think 24 bits is plenty for most audio processing purposes. Yet 32 bits FP will use up a lot more disk space. If it is a noticeable improvement than it will be worth it. I just have no way to know that for sure. If someone compares the two there is no actual way to make a comparison. Both files will have be at 24 bits to hear them.  
 
I know this is contrary to the ideas CW has been promoting I just have no evidence it actually matters. I also think a lot of the errors we see listed here on this forum is due to systems that really can't do full 32 bit FP on large projects. But that is just a notion I have.