Why test for proportional hazards-or any other model assumptions?