ESTIMATION IN THE POLYTOMOUS LOGISTIC-REGRESSION MODEL

This paper elucidates the main problems associated with estimation in the polytomous logistic regression model. We review the methods for estimating the parameters in the model and introduce a modified procedure in which all pairwise binary logistic models are fitted and combined to construct a single vector of estimates. The resulting estimators are found to be as efficient as the maximum likelihood (ML) estimators in various cases. Our method requires more computations but has an advantage for large data sets where computer space limitations may render the ML estimation intractable. Also, it enables to detect particular features in the data structure. Examples of real and artificial data are used to illustrate the properties of the estimators.