Separate modeling approaches for survival analysis with inaccurate covariate measurements

To fit the Cox regression model with time-varying covariate, ideally we need values of the covariate available for all patients in the risk set at each event time. But in reality, the covariate measurements are collected at limited patient-specific time points, and the data used for analysis are typically inaccurate with imputation errors. Four separate modeling-based methods for survival analysis with inaccurate covariate measurements are evaluated through extensive simulation studies. It is found that our proposed new method, which is based on the maximum likelihood principle, performs the best and is superior to a well-known method proposed by Prentice. These findings aid practitioners in the decision of which method to choose in practice and shed light on the direction of further research work.