Route Choice Models with Subpath Components

The problem of route choice is critical in many contexts, for example in intelligent transport systems, GPS navigation and transportation planning. In order to capture the complexity of the decision process, disaggregate models, such as discrete choice models are required. In the Multinomial Logit (MNL) model, the alternatives are assumed to be independent. This assumption is not valid in a route choice context due to overlapping paths. Several adaptations of the MNL model have therefore been proposed in the literature, thereof the Path Size Logit model. In this paper we show that, except the original formulation, all Path Size formulations presented in the literature show counter intuitive results regarding the correction of the independence assumption. Furthermore, the generalized Path Size formulation fails its original purpose of penalizing longer paths in favor of shorter ones. There is however an interesting behavioural interpretation of the Path Size attribute. Namely, overlap ping paths are attractive since travellers have the possibility of switching between routes. A Path Size attribute (original formulation) could therefore be included in the deterministic part of the utility with a behavioural interpreta- tion, but is not sufficient for correcting the independence a ssumption. Considering subpaths instead of links in route choice modelling has two main advantages. First, it is behaviourally more realistic and second, it reduces th e complexity of the models. In this paper we have proposed different definitions of subpaths, ea ch with its own specific purpose. Moreover, we have presented a factor analytic specification of the Logit Kernel (LK) model in- cluding subpath components. Model estimations show very promising results of the LK model combined with a Path Size attribute. The increase in model fit is remarkable, and the covariance parameter estimates suggest that this formulation captures an important correlation structure. Furthermore, using subpath components compared to links in the correlation structure consid- erably decreases the complexity of the model, while its capacity of capturing the correlation structure seems as promising the formulation including links.