Abstract:
|
Survival models are commonly used for data with censored information. However, parametric or semiparametric survival models require strong assumptions which often are not satisfied in real application and are also limited in the presence of group adjustments. Motivated from our breast cancer gene expression data, which exhibit the "small n and large p" problem, we propose a nonparametric mixture of semiparametric Bayesian hierarchical survival models to simultaneously study the effects of clinical covariates and expression levels of genes in a pathway on survival time, as well as to detect potential presence of different clusters. Given the existence of different patterns for the survival time of patients at different breast cancer stages, our model is able to not only detect the subtle changes of pathway effect on survival time, which could be missed using gene-based analysis, but also cluster similar patients across different disease stages. The advantages of our approach are illustrated with simulated data and a breast cancer gene expression dataset.
|