Modelling heterogeneity in rank-ordered data

We consider the modelling of rank-ordered data. This data results from questions in surveys which ask to rank alternatives and contains more information than data with only the most-preferred alternative of individuals. Standard models for modelling this kind of data make assumptions which are undesirable, especially when heterogeneity across individuals is present. These models are then less useful. We consider unobserved heterogeneity in ranking capabilities across individuals and preference heterogeneity in this paper. We use different models and evaluate which model performs best under different circumstances. We conclude that the latent-class rank-ordered logit model performs best in case of unobserved heterogeneity in ranking capabilities and in case of moderate preference heterogeneity, otherwise the mixed rank-ordered logit model performs best. We also find that one needs to be careful to conclude which kind of heterogeneity is present in the data, because the used models lead to contrary conclusions in both a simulation study and an empirical application.