Using a basic latent class model for the analysis of three-way three- mode data (i.e. raters by objects by 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, Meulders et al. (J Classification 19:277–302, 2002) proposed a constrained latent class model in which object-attribute associations are explained on the basis of latent features. In addition, qualitative rater differences are introduced by assuming that raters may only take into account a subset of the features. As this model involves a direct link between the number of features F and the number of latent classes (i.e., 2F), estimation of the model becomes slow when many latent features are needed to fit the data. In order to solve this problem we propose a new model in which rater differences are modelled by assuming that features can be taken into account with a certain probability which depends on the rater class. An EM algorithm is used to locate the posterior mode of the model and a Gibbs sampling algorithm is developed to compute a sample of the observed posterior of the model. Finally, models with different types of rater differences are applied to marketing data and the performance of the models is compared using posterior predictive checks (see also, Meulders et al. (Psychometrika 68:61–77, 2003)).