Okay, so 'shiftiness' is independent of appearance as measured other ways, but what is it exactly? The next step is quantifying what cues people pick up on that identify trustworthiness, and that's what I'm really interested in.
An excellent way to figure this out would be to apply some sort of machine learning technique. Perhaps an Active-Appearance-Model or some other statistical measure to mathematically represent the shape and texture of an "untrustworthy" face as a lower dimensional vector. Then train classifier on this data which can discriminate between "trustworthy" and "untrustworthiness". Of course, there are going to be ethical concerns, but this is definitely one interesting way of doing it.
Might this be a self-fulfilling prophecy? If the 'shifty' group were offered higher interest rate, then this increases the probability that they will bail on their loans, no?