Well, when we use any finite-size representation for real numbers, we're essentially defining a mapping between the infinite, uncountable set of real numbers and a finite and countable set of combinations of say 32 or 64 bits..

When the mapping is made according to floating point representation rules, larger numbers will have less representatives than smaller ones. That is to say, when the numbers get bigger, the representation is less dense, so each sequence maps a larger set of different reals, so everything else being equal approximations due to the representation may become a factor. In theory.

Now, I have absolutely no idea how this would affect perception when computations carried on using these numeric representations is ultimately used to drive a speaker...

I could guess that if one drives things a lot over digital 0dBFS in buses (i.e. the calculations begin to involve very large numbers), before returning back down in the main bus, it might lead to some perceptual difference... but if people use a reasonable gain structure, or just occasionally go a little over 0 db FS, all computations will occur within or near the band of maximum representational density, so it won't make any difference - only the stability and the error size of the algorithms will matter (as we expect and experience).

Perhaps some of the people noticing differences love to push their digital levels in the red.