This is an article dealing with the analysis of categorical data when some sample units are not fully classified. It is intended to review and develop the application of the maximum likelihood approach when based upon a Multinomial sampling model provided with a noninformative structure for the non-response (or censoring) process. The results of such application are described in matrix terms suitable for their computational implementation regardless of the contingency table shape and the data incompleteness pattern.
keywords: Multinomial distribution, data missing at random, ignorable censoring, informative censoring, maximum likelihood methodology, strictly linear models, log-linear models.