A0720
Title: On the identifiability under dependent censoring or in a competing risk setting using divergence measures
Authors: Roel Braekers - Hasselt University (Belgium) [presenting]
Abstract: In a competing risk setting or under dependent censoring, latent random variables are often assumed to model the different times until certain failure events. However, this type of data is observed only for the minimum event time and the cause of failure. Therefore, it needed to make assumptions on the association between the latent random variables to identify the joint survival function of the latent random variables from the observed quantities. For example, assuming independence or a known association function (in the form of a copula function) between these random variables is a common assumption in a competing risk setting to avoid identifiability. When a parametric family of copula functions is assumed for the association and parametric marginal distributions for the different event times, it is harder to check whether the assumed model is identifiable from the observed quantities. Divergence measures are used to develop a method to verify this identifiability in a competing risk setting or under dependent censoring. It is shown that this method works for various parametric copula functions for the association between the different event times and their parametric marginal distributions, even when the number of parameters increases to a very large number.
