B0714
Title: Robust prediction of survival outcomes through unified Bayesian analysis of nonparametric transformation models
Authors: Catherine Liu - The Hong Kong Polytechnic University (Hong Kong) [presenting]
Abstract: The linear transformation model is a general semiparametric survival model that includes the widely-used Cox model, proportional odds model, accelerated failure time model, and beyond. It is known to be challenging to predict the survival outcomes when the transformation function and the distribution of the error term are both unspecified owning to model identifiability. Unlike the strategy of identifying the model first and then fitting the model in existing Bayesian literature, we develop a unified Bayesian procedure to estimate the nonparametric functions and the parametric component simultaneously and jointly under quite mild assumptions after exponential transformation. We construct an augmented Gamma process smoothed by I-spline functions as the prior for the monotonic transformation function. We refine the parametric estimator by a posterior projection owing to the constraints of identifiability. The predicted conditional survival function behaves very well whatever the underlying true models are in numerical analysis. The MCMC sampler is built based on the No-U-Turn sampler (NUTS) by Stan. Comprehensive simulations and an application to real data illustrate the methodology broadly using an R packageBuLTM.
