A Mixture-Model Approach to the Analysis of Survival Data

In this paper, we study a mixture model for survival data where covariates may innuence both the incidence probabilities and their conditional latency distributions. The data may include the exactly observed, the right-censored, and the interval-censored failure times. We apply the EM algorithm to nd the maximum likelihood estimate. We also carry out a Markov chain Monte Carlo algorithm for Bayesian inference. Model selection methods based on the predictive density for cross-validated data are developed. These methods allow us to assess whether simpler models would suuce as opposed to the mixture models. The potential of the methods is illustrated with our beetle (Tribolium castaneum) data given by Hewlett (1974).