keep in mind that what we are talking about is linear quantization here, so that no matter how many bits you use for recording, the "step size" is still 6 dB/bit. 24-bit doesn't take the same S/N and divide it into more steps that 16-bit, instead it increases the S/N by reducing the noise floor from the A/D conversion process. this gives you the alternative of running lower levels but still getting the same or better S/N than if you ran higher levels at 16-bit. from a purely S/N perspective you are still better off running as close to 0 dB as possible, but as soon as you go over 0 dB your S/N drops, sometimes dramatically depending on how far over and whether it was just a few samples (like a drum hit) or something more substantial. so the basic idea is that 24-bit allows you to run levels more conservatively, avoiding overs with a reasonable safety margin, but still getting S/N that is as good or better than 16-bits run hotter.
when we talk about post-recording processing, there is no doubt that doing this at the highest bit level possible is an advantage (mathematically at least, at what point the differences become audible is another question).
I would be careful when comparing recordings run at different levels, it is a well-known phenomenon that even a level difference as small as 0.5 dB can be enough to make the hotter source sound better, even if it's identical otherwise.
Steve