Bayesian nonparametric modeling of heterogeneous time-to-event data with an unknown number of sub-populations

Time-to-event data are a broad class of data widely encountered at different stages of the product life cycle. In practice, time-to-event data often exhibit heterogeneity, due to a variety of design and manufacturing issues, such as material quality inhomogeneity, unverified design changes, and manufacturing defects. Existing time-to-event modeling approaches mainly ignore this heterogeneity or account for it by pre-determining a fixed number of sub-populations. However, neglecting heterogeneity hinders the modeling accuracy, whereas pre-determining the number of sub-populations is often subjective or unjustifiable. In this article, a Bayesian nonparametric model is proposed to model heterogeneous time-to-event data by assuming an unknown number of sub-populations and quantifying the influence of possible covariates. An estimation algorithm is further proposed to achieve the joint model estimation and selection and to deal with the non-conjugate priors. Case studies demonstrate the effectiveness of the proposed work.