Flexible and Efficient Data Augmentation Schemes

• Activity Number: 227
• Type: Contributed
• Date/Time: Tuesday, August 9, 2005 : 8:30 AM to 10:20 AM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Taeyoung Park*+
• Companies: Harvard University
• Keywords: Data Augmentation ; Marginal Data Augmentation ; MCMC ; Missing Data ; Working Parameter

Abstract

The method of data augmentation (Tanner and Wong 1987) is a powerful method that constructs iterative sampling schemes by introducing missing data or latent variables into a statistical model. In the context of posterior fitting or sampling, augmenting the observed data to a larger dataset is a common strategy to simplify the convolved model structure and/or computation. Although the resulting algorithms are often easy to implement, they sometimes suffer from slow convergence. The method of marginal data augmentation suggested (Meng and van Dyk 1999) is a technique that can lead to significant improvement in the convergence of the resulting sampler. Here, we propose a more flexible strategy that that aims to maintain the efficiency of marginal augmentation by partially marginalizing out (working) parameters or missing data in some of the sampling steps, which may allow us to avoid computational difficulties or to reduce the total computation time required for convergence. We apply this strategy to a highly structured, multilevel spectral model used in high-energy astrophysics to describe the energy spectrum of the high-redshift quasar, PG 1634+706.