S. Perra, S. Cabras, M. E. Castellanos Nueda
In this work we consider the problem of covariate selection in Weibull regression analysis for right censored data under an objective Bayesian point of view using the usual improper prior for location-scale models. We compare each candidate model with a null reference model. Model comparison is approached via the intrinsic Bayes Factor (IBF) of Berger and Pericchi (JASA, 1996) and the fractional BF (FBF) of O'Hagan (J. Roy. Statist Soc., 1995). Because of censoring (Berger and Pericchi, Ann.Stat, 2004) we need to define a suitable minimal training sample (MTS) that accounts for censoring. We address this problem using sequential MTS (SMTS) for which the MTS sample size, N, is random. We derive its probability distribution P(N) and show that averaging the conditional (on N) FBF over P(N) may be a feasible alternative to the IBF with SMTS for searching over a large space of models. Comparisons between FBF and IBF are made with a simulation study and an application to a real data set.
Palabras clave: improper priors, regression, training sample
Programado
MC1 Métodos bayesianos 1
17 de abril de 2012 12:00
Salón Madrid
