Fitting Log-Linear Models to Contingency Tables With Incomplete Observations

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


The study of contingency tables is prevalent throughout the biological and physical sciences. A common difficultly occurs, however, when some or all of the data can not be completely classified. This paper shows how the ECM algorithm (Meng and Rubin, 1993) along with iterative proportional fitting (Bishop, Feinberg, and Holland, 1975) can be used to fit any hierarchical log-linear model to a table with any combination of partially and completely classified observations. Beyond this, we develop the supplemented ECM (SECM) algorithm which can be used to calculate asymptotic variance matrices of the maximum likelihood estimates fit with a log-linear model. An infant survival example is given which demonstrates both the precision of the SECM algorithm, and a powerful internal checking system which detects errors in the implementation of both ECM and SECM.

Key words: Contingency Table; Convergence Rate; EM Algorithm; Fisher Information; Incomplete Data; IPF; Missing Data; Observed Information; SEM Algorithm

Return to David van Dyk's homepage.