Iterative conditional fitting for discrete chain graph models

Abstract

'Iterative conditional fitting' is a recently proposed algorithm that can be used for maximization of the likelihood function in marginal independence models for categorical data. This paper describes a modification of this algorithm, which allows one to compute maximum likelihood estimates in a class of chain graph models for categorical data. The considered discrete chain graph models are defined using conditional independence relations arising in recursive multivariate regressions with correlated errors. This Markov interpretation of the chain graph is consistent with treating the graph as a path diagram and differs from other interpretations known as the LWF and AMP Markov properties.