No one really knows. I'm doing a training experiment at the moment and the spread of figures is so great it's really difficult to get any kind of idea what's going on.
With moral it could make something like a 5% difference to a players performance. Now, if that were the case, then it'd be almost impossible to collect enough data to back that up. You'd need to run hundreds of games and try and discount all the other variables, it's just not gonna work.
Best bet is keep it topped up and hope for the best.