On the performance of small-area estimators: Fixed vs. random area parameters