Some other metrics which are sometimes used include capacity, outage probability, mutual coupling etc. Is there any marked advantage that can be achieved in analyzing the symbol error rate (SER) of MIMO system over the other metrics?
I wouldn't necessarily say "symbol error rate instead of" capacity, or mutual coupling. A high symbol error rate is the way a capacity limit would manifest itself. For example, you can only try higher order constellations until the symbol cannot be accurately decoded anymore, which would be evidenced by bit error rate, resulting directly from not having been able to decode the symbol correctly.
It's true that there are more and less robust RF modulation types. For instance, PSK is usually more robust than ASK (or QAM), because with PSK, each symbol is transmitted using the max power of the transmitter. So for equal spectral efficiency, you would expect PSK to achieve results closer to the Shannon limit, than if you used ASK for that same spectral efficiency. How would you test this in the real world?
If you use a fixed value of spectral efficiency, then one way to test this would be to lower transmitter power until your symbol error rate reaches a certain threshold level. (These things tend to degrade rapidly at the limit, so that picking different threshold levels won't change the results very much.)
MIMO doesn't change this. If the separate propagation paths are too highly correlated, the result will be that you will be unable to decode the symbols correctly, until you reduce the constellation size (and consequently the spectral efficiency).
The SER is an important metric, but many of the "simple" formulas that you can find in the literature give the SER for uncoded transmission. Since most communication systems use channel coding with pretty long codewords, it is the coded SER that is of interest and not the uncoded one.
By the way, with channel coding and sufficiently long codewords the SER will be pretty small. It is then the spectral efficiency that is of interest.