Conditions when statistical tests for model discrimination have high power