Title: On variants of the iterative scaling algorithm
Authors: Tamas Rudas - Hungarian Academy of Sciences Centre for Social Sciences (Hungary) [presenting]
Anna Klimova - IST Austria (Austria)
Abstract: The Iterative Scaling Algorithm has been present in statistics for a long time and is routinely used in many applications, including small area estimation in official statistics, post-stratification in survey analysis and maximum likelihood estimation of log-linear models. The main focus of interest is the latter area of application. First, a unified treatment of the original IPS and two of its modifications, the Generalized Iterative Scaling, and of the Improved Iterative Scaling, are given. Then, it is shown that these algorithms cannot deal with the problem of maximum likelihood estimation in a coordinate-free generalization of the log-linear model, called relational models. In these models, the sample space does not have to be a Cartesian product, the multiplicative effects are not necessarily associated with cylinder sets and an overall effect may not be present. Maximum likelihood estimates have many surprising properties. A new variant of IS is described, which may be applied to relational models, and it is shown that a generalization of it also works in the presence of zero observed frequencies.