EXPLORATORY ANALYSIS OF ASSOCIATION MODELS USING LOGLINEAR MODELS AND SINGULAR VALUE DECOMPOSITIONS

A method of approximating maximum likelihood fits of multivariate association models is explored. The proposed procedure is based on singular value decompositions of matrices of interaction terms from hierarchical loglinear models estimated by the method of maximum likelihood. Potential applications include use in exploratory data analyses, in screening procedures to select models to be estimated by maximum likelihood, and in generating initial estimates for algorithms which compute maximum likelihood estimates. Two examples are used to illustrate features of the estimation procedure and its application to model screening.