Construction, Implementation, and Theory of Algorithms Based on Data Augmentation and Model Reduction

David A. van Dyk
Department of Statistics, The University of Chicago, Chicago, IL 60637, U.S.A.

In the thesis we provide a general framework for maximum likelihood algorithms based on data augmentation and model reduction. Starting with this theoretical framework, we explore methods of constructing and implementing efficient algorithms. We show how to derive faster algorithms by optimizing the rate of convergence as a function of a working parameter which is introduced into the data-augmentation scheme. We then propose the Alternating Expectation/Conditional Maximization or AECM algorithm which includes the EM, ECM, ECME, and SAGE algorithms as special cases. We also show how the matrix rate of convergence can be used to compute the asymptotic variance-covariance matrix of the maximum likelihood estimates. The relative efficiency of competing model-reduction schemes is explored via permutation of the conditional maximization steps within the ECM algorithm. Finally, we explore the inferential use of the data-augmentation scheme in the context of estimating the number of components in a finite mixture, with possible extensions to other model-fitting problems.

Return to David van Dyk's homepage.