A large majority of mixing and sound modifying programs today use 32bit float internally when processing sound. It is a very good format format for this and it keeps a really large part of the sound information intact. Quite a few playback programs handles 32 bit today, not all though.
When converting between formats the signal should as a general rule be dithered. If you go to 24 bit the choice is not sensitive at all in my experience, simply select one. If you go to 16 bits, it can in my experience make a tiny difference when working on finished mixes, but I let my ears judge there.
What you have to look out for however when going from 32bit float to 24 or 16 bit is that the signal never goes above 0dB. The float format allows signals to go above that level, but the fixed formats cannot handle it. The simple solution is to normalize the files first, leaving a small margin up to 0dB, I generally leave 1 dB. There are a lot of technical reasons for leaving a margin, a bit too involved to describe right here. Regardless one or a few dB-s in lowered volume really does not make much of a difference.
// Gunnar
Gunnar