Performance of Cox regression models for composite time-to-event endpoints with component-wise censoring in randomized trials.