I would much rather focus on the mastering quality than file compression.
MP3 320k, or better - V0 VBR (since it relies more on the psychoacoustic modelling) are transparent for most people.
This means most people cannot detect differences between them and the uncompressed sources.
This is because MP3 and the newer AAC use psychoacoustic model to detect situation during which masking would occur.
To put it shortly - a louder sound with wide spectrum would mask softer sounds in the same band.
A loud sound following a quiet one will make the first one inadubile.
More info at:
https://en.wikipedia.org/wiki/Auditory_masking
Psychoacoustic models divide the sound into multiple bands, trying to mimic the critical bands in our auditory system.
Then they detect the masking and remove information which we would not hear anyway.
The removing is done via decreasing their resolution (by quantizing the MDCT coefficients), e.g. the number of bits to describe the sound level of a given frequency in a given critical band.
Decreasing resolution is more or less decreasing the SNR at that specific frequency, in that specific time slot in the recording by 6dB for a lost bit.
Then the MDCT coefficients are coded statistically and using RLE, similar to what you have in an ordinary file compression.
File compression does not work well with .wav files, but the numerical data of quantised MDCT coefficients can be aligned in such a way so that the efficient coding is possible.
So while 256kbps or 320kbps seem like a lot of information missing from 1.4mbps on CD, it does not mean that you only keep ~20% of the original information.
The AAC and recently developed OPUS codecs have even better psychoacoustic modelling, achieving transparency at lower bitrates than MP3s.
I can't remember where to find it, but there was a whitepaper where both causual listeners and audiophiles were to discern between lossless and high quality MP3s on a high-end system.
Only one person could reliably discern MP3 from lossless, and he was tested to have severe hearing deficiencies. It turned out his brain's psychoacoustics were not working according to the model because of that.