ABSTRACT
Methods for survival analysis described in earlier chapters are only valid if any censoring is independent of the observed survival time. Circumstances where this assumption cannot be made often arise, such as when an individual is withdrawn from a study because of their deteriorating condition. This chapter opens with a summary of how dependent censoring can be detected. The sensitivity of inferences of primary interest, such as estimates of hazard ratios the median survival time, to dependent censoring are then considered. A sensitivity analysis that enables the impact of various degrees of dependent censoring on summary statistics to be established is then outlined. In the presence of dependent censoring, Cox regression models can be fitted using a process that leads to inverse probability of censoring weighted estimates of the model parameters. The method is described in some detail and illustrated in an example.
