Inverse Probability Censoring Weights for Routine Outcome Monitoring Data

Many researchers who work in survival studies encounter censored data during their career. Unfortunately, censored data makes analysis more complicated, since exact event times are not observed. One type of censoring is interval censoring, occurring in longitudinal studies where patients are observed at repeated visits. If a patient experiences an event, it is detected at the next visit. In this case analysis is more difficult because no precise event times are observed. Packages are developed for R to handle interval censored data. However, the results produced by these packages are not satisfactory. The parameters of the Cox proportional hazards model can be estimated, but no information about the precision of these estimates is returned. This prevents drawing conclusions about the significance of the covariates. In this thesis, a bootstrapping method is proposed to assess the precision of the parameter estimates, such that better conclusions can be drawn from them. Standard survival methodology assumes that patients’ withdrawal from a study is independent of patients’ characteristics. However, clinical experience may suggest dependent censoring. The Inverse Probability Censoring Weighted (IPCW) Estimator was developed to take a censoring mechanism into account when performing survival analysis. Application of this method is complicated, because it involves many mathematical formulas and programs must be written by the researcher. In this thesis a special way to prepare the data is proposed such that standard survival packages in R can be used to perform the IPCW method. In this way, IPCW becomes more available for researchers with limited mathematical and programming skills. Furthermore, a method to generate data including dependent censoring is proposed. This method is used to perform a simulation study to assess the performance of the IPCW method compared to standard survival analysis in the presence of dependent censoring.