A1478
Title: On the choice of censoring distribution in the simulation of survival data
Authors: Konstantinos Pateras - Athens University of Economics and Business (Greece)
Takis Besbeas - Athens University of Economics and Business (Greece) [presenting]
Abstract: Simulation of censored time-to-event data is important but presents its own challenges. First, the simulation of right-censoring, which is the most common censoring mechanism in medical research, requires, in general, simulation from two processes, one describing time-to-event and the other time-to-censoring. Second, it will often be desirable to control the censoring proportion of the simulated data. And third, there are various characteristics that make survival data more complex than other types of data. We consider the choice of censoring distribution in the simulation of censored data from a known survival model. We show that there are often practical advantages in selecting the censoring distribution to be in the same probability distribution family as the lifetime distribution, and we illustrate using Monte Carlo simulation that the choice of distribution with bounded support, such as the uniform, which is typically used, may have disadvantages compared to distributions with unbounded support, such as the exponential. We also evaluate the performance of more flexible censoring distributions, such as the Weibull and Gamma, under a range of values for their additional parameter. We further illustrate the effect of the censoring distribution on a parametric bootstrap analysis of a real survival data set on shunts in infants with heart disease.
