All too often, the problems with comparing sample rates is the use of not-linear phase filters, and the various group delay characteristics of the filters. A not-linear phase brickwall can have effects well into the audio spectrum, but they are not level differences, they are TIMING and propagation differences. Linear phase (constant delay) are absolutely necessary for controlling the experiment, unless all group delay errors have been compensated. (If this very general, and almost non-technical statement isn't clearly understood with nuance, then don't bother trying to do an experiment. This can be subtle stuff, and strange/erroneous results are very easy to achieve.)
The measurements must be controlled, and done with very careful engineering (I mean, DSP/human factors researcher level and defended), because there are so many variables (including human hearing changing vs time, and a 5-15 second accurate memory for many people.) Gotta understand some statistics also -- making sure that when there ARE random errors, they are at least partially filtered. Biased results (e.g. ham-handed choice of rate conversion software), doesn't help increase the knowlege base for anyone.
I don't mean to sound critical, but I have seen a lot of arguments that even I (just a fairly/reasonably knowlegable DSP person, writing really innovative software, but not the biggest expert in the world) can legitimately be critical and argue away many/most claims about high res sample rates. That doesn't mean that high res sample rates are never useful -- because they are, but not really in linear applications like presentation/playout. As long as you have 48k (I don't like 44.1k for emotional bias reasons), you are able to provide whatever human ears can receive. (Well, I don't like hiss either, and prefer wiggle room, so at least 16bits is best.) If you are doing ANYTHING nonlinear, then it is a good thing to have wiggle room (almost literally) in the sample rate, and usually 96k is a good choice, but I can almost always get by with 66.15k or 72k to save CPU and filter taps.
John