International Conference on Computational Statistics edition:21 location:Genève date:19-22 August 2014
Using a basic latent class model for the analysis of binary three-way three-mode data (i.e. raters who judge whether or not objects have certain attributes) to cluster raters is often problematic because the number of conditional probabilities increases rapidly when extra latent classes are added. To solve this problem we propose a constrained multilevel latent class model in which object-attribute associations are explained on the basis of binary latent features. In addition, rater differences are introduced by assuming that raters only consider each of the latent features with a class-specific probability. For parameter estimation, an EM-algorithm is developed to estimate the posterior mode(s) of the model and a Gibbs sampling algorithm is derived to compute a sample of the posterior distribution. As an illustration, the model is applied to two real data sets: First, the models are used to study individual differences in hostile behavior. Second, the models are used to analyze patient-symptom judgments of different clinicians to study the structure of psychiatric syndromes.