|
Abstract:
|
Joint models are commonly used in clinical studies for analyzing survival data with time-dependent covariates or biomarkers. It is often assumed that the latent processes that are used to describe the association between longitudinal and survival outcomes follow a multivariate normal distribution. While a joint likelihood analysis may provide valid inferences under correctly specified latent processes or random effects distributions, but the maximum likelihood estimators can be biased under misspecified random effects and hence may provide invalid likelihood inferences. In this work, we explore the empirical properties of the maximum likelihood estimators in joint models under various types of random effects distributions, and propose a robust and efficient skew-normal distribution to address uncertainties in the latent random effects distributions. An extensive Monte Carlo study indicates that our proposed method provides consistent and efficient estimators of the joint model parameters under various types of model misspecifications. We also present an application of the proposed method using a large clinical dataset obtained from the genetic and inflammatory markers of sepsis (Ge
|
