Advantages of latent class models over continuous mixture models in capturing heterogeneity.

Latent class model structures present a flexible extension of standard choice modelling approaches in the context of the representation of taste heterogeneity. In this study evidence is provided of the flexibility of the latent class approach and the potential advantages over other model structures. As a first contribution, formulae are derived for the correlation between individual taste coefficients in latent class structures and show how this correlation is a function of the socio-demographic attributes used in the class-allocation model. From this, an analyst can for example conclude that for specific subgroups in the population, the time and cost coefficients are negatively correlated, while for others, the correlation may be positive. This is a crucial advantage over the Mixed Logit model, which only produces a fixed measure of the correlation between two randomly distributed coefficients. As a next step, it is shown that the same principle applies to the elasticities in latent class models, where these can again be expressed as a function of the socio-demographic attributes used in theclass-allocation model. In the applied part of the study, use is made of stated choice data for departure time and travel mode collected for the Dutch National model. The results showed that the latent class model obtainssignificant gains over the basic MNL model, where these gains in log-likelihood are comparable to those obtained by a Mixed Logit model with a fully specified covariance structure between random coefficients. However, thereal advantages come at the interpretation stage. Here, the analysis illustrates the relationship between socio-demographic indicators and the covariance structure and elasticities in a latent class model. This information is not only useful for the analysis of taste heterogeneity but can also provide significant advantages in forecasting. Additionally, the results show significant differences in the covariance results between the latent class and Mixed Logit models. As an example, while the Mixed Logit model gives a correlation of 0.41 between the travel time coefficients for car andtrain, in the latent class models, this correlation ranges from 0.43 to 0.9 depending on socio-demographic characteristics. For the covering abstract see ITRD E145999