Seeking Efficient Data Augmentation Schemes Via Conditional and Marginal Augmentation

Biometrika

Xiao-Li Meng
Department of Statistics, The University of Chicago, Chicago, IL 60637, U.S.A.
David A. van Dyk
Department of Statistics, Harvard University, Cambridge, MA 02138, U.S.A.

Data augmentation is a powerful tool for constructing deterministic and stochastic algorithms for likelihood and Bayesian computation. Constructing an efficient data-augmentation scheme requires a balance of simple implementation and quick convergence. Meng & van Dyk (1997) reported the success of the ``working parameter" approach for constructing fast and simple EM-type algorithms for some common models, and suggested that efficient data-augmentation schemes for EM-type algorithms could also be effective for implementing the corresponding Gibbs samplers. This paper investigates this possibility by providing a theoretical study of two working parameter approaches, the {conditional augmentation} approach and the {marginal augmentation} approach. The former was the focus of Meng & van Dyk (1997) and the latter builds on the parameter-expanded EM approach of Liu et al. (1998). We use the t model as a running example to illustrate the potential of these methods. The t model also turns out to be an excellent springboard for examining the convergence behavior of the Gibbs sampler with improper limiting distributions. A number of open theoretical questions are also discussed, in particular with respect to marginal augmentation which in some cases allows one to obtain a fast-mixing positive recurrent Markov chain by purposely constructing a non-positive recurrent Markov chain in a larger space.

Some key words: Incomplete data; Markov-chain Monte Carlo; PXEM Algorithm; Rate of convergence; Working parameter.

Return to David van Dyk's homepage.