Abstract
The Cox model was originally built upon the proportional hazards (PH) assumption that is often violated in reality. When the PH assumption for a covariate is not satisfied, the Cox model estimates the overall effect of the covariate within the follow-up period, but the estimate may be strongly affected by censoring. For a clinical trial, interim analyses may be inconsistent with the final analysis because staggered entry creates different censoring patterns for analyses at different time points. A method based on multiple imputation and Bayesian piece-wise constant survival models is proposed to adjust interim analyses under nonproportional hazards to improve the consistency. A simulation study is applied to demonstrate and compare the performances of the proposed method and two existing methods. The applications of these methods are illustrated with a gastric cancer trial.
Acknowledgment
The data used as an application example is from the Dutch Gastric Cancer Trial. The authors thank Dr. Hans C. van Houwelingen for providing the data.
Notes
The pH assumption is satisfied when the study time is 2 or 4 years.