A1531
Title: Bayesian analysis of competing risks data with dependent left truncation via a trivariate Gaussian copula
Authors: Kota Izumi - Kitasato University (Japan) [presenting]
Hirofumi Michimae - Kitasato University (Japan)
Abstract: For competing-risks data subject to left truncation, independence between the latent failure times and the truncation time is often assumed. If this assumption is violated, ignoring the dependence may lead to biased inference. A Bayesian model for competing risks with dependent left truncation is developed by specifying the joint distribution of the two latent failure times, corresponding to competing causes, and the truncation time using a trivariate Gaussian copula. Weibull distributions are assumed for the marginal distributions of the latent failure times, and a uniform distribution is assumed for the marginal distribution of the truncation time, so that the truncation mechanism and the competing-risks process are modeled jointly. Bayesian estimation is performed in Stan via Markov chain Monte Carlo, and uncertainty is quantified through the posterior distributions of the Weibull and dependence parameters. For comparison, maximum likelihood estimation under the same joint model and an alternative model that assumes independence between the failure and truncation times are also considered. In simulation studies, data are generated under prespecified sample sizes, truncation rates, and dependence structures, and the methods are compared in terms of estimation accuracy and interval coverage. The inferential impact of ignoring dependent truncation and the practical value of explicitly modeling it are examined.
