Designing completely randomized experiments to detect ordered treatment effects: an empirical study