Complex Model Structures Versus Fine Disaggregation of Travel Demand: the Specification of an Urban Mode Choice Model for the French Urban Area of Rouen
T Anselme, Centre d'Etudes Techniques de l'Equipement/Ministry of Transport; F Combes, Universite Paris-Est, LVMT, FR
In passenger mode choice models, there is a complex interaction between the detail of the description of supply and demand and the adequate model structure and specification. This issue is explored in the frame of a French case study.
Mode choice models are essential decision support tools for urban passenger transport stakeholders. In general, a mode choice model is a discrete choice model, in which a certain utility is associated to each mode. These utilities consist of a deterministic part and of a random part, which are most often combined additively. The specification of both components is a critical driver of the empirical and behavioural quality of the model, and as such, should be chosen with great care.
When choosing the specification of the utility functions, there are four issues: first, the representation of transport alternatives (travel time, mode, cost, comfort, frequency, availability of car-park places etc.); second, the representation of passengers (gender, age, income, etc.), as well as their segmentation into categories, or even the distribution of a coefficient within their population; third, the formal specification of the deterministic utility function, including non-linear transformations; and finally the specification of the random part of the utility function, or, equivalently, the structure of the choice model (logit, probit, nested or cross-nested logit, etc.).
These four issues cannot be considered independently; on the contrary, they share complex relationships. For a given set of variables, some model structures may be more adequate than others, and conversely. This makes the prescription of specification guidelines difficult, passed a certain level of detail.
In order to investigate this topic more closely, this paper examines a substantial number of specifications on a case study. Based on a travel survey of the medium-sized French urban area of Rouen, several mode choice models are tested statistically. These mode choice models distinguish walking, private car, and public transport. All of them take into account road congestion, five time periods per day, and hyperpath assignment for public transport. They are estimated by likelihood maximisation; their capacity to reproduce the observed mode shares is also taken into account to compare them between one other.
Starting from a basic model with both a simple structure (multinomial logit) and standard variables, the model complexity is iteratively increased. First, more complex formal specifications and model structures are considered, including nested logits and Box-Cox transformations, both improving the model?s statistical fit. Second, the representation of travel demand is refined, notably by the distinction of trip purposes, then by taking into account socio-economic variables. Third, the representation of the transport alternatives is also improved substantially by including a qualitative variable representing the availability of car-park places at destination. This variable is a synthesis of many indicators describing the difficulty for car drivers to park at their destinations. Finally, captivity to public transport is considered, on the basis of car and driver license availability in households.
This sequential approach leads to a mode choice model comprehensive in terms of data and satisfying statistically. Interestingly, it also leads to the conclusion that in the context of this study, the more precise the representation of transport alternatives and of travellers is, the less complex the models' specifications need to be. As a matter of fact, in the final model, the benefit of hierarchical structures and non-linear transformations of the utility function compared to a simple multinomial logit model is statistically non significant. Of course, this result comes at a cost, as variables related to car ownership become exogenous: this reduces the model?s sensitivity to transport policy changes, and a car ownership model would then be necessary to overcome that issue. But, if general conclusions are to be drawn, it advocates for a finer representation of travellers and of transport alternatives in mode choice models, as complex specifications may provide inadequate solutions to a lack of accuracy in the representation of transport supply and demand.
Association for European Transport