How long have you got? A full answer would take pages and graphs and dull stuff...
The real key is the source impedance of the interference - how well it can 'couple' into your wanted signal. There are basically two regimes - magnetic (fields coupled from transformers, and heavy currents flowing in large loops locally) and electrostatic - better known as capacitative coupling.
The former 'couples' most effectively into low impedances; the latter most-effectively into high impedances. and for the purposes of argument, the break point between the two turns out to be ,
very roughly, only around 300ohms - a pretty low impedance.
Shielding magnetic fields is therefore a difficult thing, but this impedance break-point is a useful fact: since sources are generally 100ohms or less these days it means - for a lot of purposes short of fundamentally-broken design - interconnect shielding works (mostly) in the electrostatic domain. and that's easy, you just interpose a conductor between your source interference and your conductors to protect, and tie this shield to a low-impedance point, and for this the source 0v point is the best option. Then, ideally, the source of noise current cannot effectively couple any noise into the signal conductors: there are limits, as davidsrdb has pointed -out, depending on what you need to achieve and how 'good' the shield/ cable geometry can be made.
Now - capacitative impedances are very high, the source might only be a few picofarads, so a source of small RF current coupled-in more than anything else. Since the noise source impedance is effectively very large, it doesn't often make things worse by inducing so much current that the signal conductor resistance converts it into a noise current in series with the wanted signal. This is how a twisted-pair alone might be 'good-enough' for audio. It is also why shielded cables for audio are tied only one end on single-ended interconnects - it's effectively a 0v-referenced shield, and all that is needed to 'protect ' the signal send-return pair - but connect it both ends and the noise current
does appear added to the signal (because the signal 0v within, like any wire at RF, has a non-zero impedance)
tl;dr:
- Twisted-pairs work by ensuring the pair of conductors present, as near as can be, the same cross-section to magnetic fields in all directions - so the induced currents largely cancel, and therefore induced noise voltage is minimised. But have no protection against capacitative coupling of noise currents. Low signal source impedance helps here.
- Shielded twisted-pairs add shielding against capacitative coupling, and are usu. used with the shield bonded to signal-0v at the source end.
- 'Star-quad' takes this geometric effect one step further against magnetically-induced noise sources
- Shielded star-quad puts things pretty-much beyond contention.
NB for shielding to be effective, it does
not have to be bonded to mains Earth. It only has to be rather low-impedance compared with your target unit's input impedance in parallel with the source's output impedance. it effectively lets the signal send-return pair work
within an environment defined as '0v at the send end'. To the nth degree for audio the capacitative- coupled shield noise current has
zero effect on the inner signal loop, because it is not in series with it. If you allow it to become so, then it
can be a problem (For contrast - Earth Loops are firmly in 'unwanted-noise-currents in cables causing problems' category!)
(The difference between Mains Earth and signal 0v is the root of so much misunderstanding in audio a different, likely longer, rant is required)
Afterthought for clarity: Balanced connections work mostly, because the function of the shield as a separate conductor is to 'bond' the
chassis together to ensure there is no noise potential voltage between them, leaving the signal conductors unmolested; and
balanced connections rely on balanced impedances for best effect, not balanced voltages (a subject for another day)