Title: Non-homogeneous interaction effects in the joint action of binary features
Authors: Anna Klimova - National Center for Tumor Diseases Partner Site Dresden TU Dresden (Germany) [presenting]
Tamas Rudas - Eotvos Lorand University (Hungary)
Abstract: Many studies of register data aim to discover association patterns within a set of binary features characterizing the subjects in a population of interest. Unaffected subjects, who possess none of the given features, do not exist in the population, and, therefore, the sample space can be described using an incomplete contingency table where one cell is absent. A new type of interaction parameters which generalize those obtained from the conventional conditional odds ratios is proposed, and, accordingly, a new class of hierarchical log-linear models specified by setting these parameters equal to zero is described. Because the proposed parameters are logarithms of non-homogeneous generalized odds ratios, the resulting models do not include the overall effect (a normalizing constant), and some of their properties are different from those of the classical log-linear models. An example is given to illustrate that the proposed models may allow for describing the association between features in more detail than it can be achieved using a quasi-variant of a classical log-linear model.