Permuting CM Steps Within the ECM Algorithm

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

The ECM Algorithm (Meng and Rubin, 1993), is a generalization of the EM algorithm in which the maximization (M) step is replaced by several conditional maximization (CM) steps. The order that the CM steps are performed is completely arbitrary, trivial to change, and generally effects how fast ECM converges. Moreover, the same order of CM steps need not be used at each ECM iteration. In this paper, we use the contingency table problem to illustrate the effect of changing the order of CM steps on the actual number of ECM iterations required for convergence. Our examples both show that the effect can be substantial and illustrate that the problem of finding efficient orderings is generally quite difficult because both theory and common sense can be misleading.

Key Words: Contingency Table; Convergence Rate; Gibbs Sampler; EM Algorithm; Incomplete Data; IPF; Missing Data

Return to David van Dyk's homepage.