Bayesian analysis of matched categorical data
Matched categorical data arise, for instance, from repeated measurements in clinical trials or in panel surveys for opinion swings. Such data can be arranged
in a square contingency table, a typical hypothesis of interest being that of marginal homogeneity (equality of row and column distributions). When the variables are binary, marginal homogeneity coincides with symmetry (equality of the off-diagonal cell probabilities) and this can be exploited to carry out Bayesian inference, as reported in the literature on correlated proportions. Unfortunately, the latter procedure is ad hoc and cannot be readily extended to the case when the two variables are polytomous. We deal with this problem by devising an alternative method based on a recently developed theory of marginal models for categorical data.
This is joint work with: Guido Consonni (University of Pavia, Pavia) and Eleonora Baviera (L. Bocconi University, Milan)