Randomised comparisons of medical tests: sometimes invalid, not always efficient