Route Choice Variability: a Case Study

Route Choice Variability: a Case Study


V Benezech, Universite Paris-Est, ENPC, FR; A Kippelen, ENPC, FR


Passengers route choices in the Paris transit network are studied using smart card data. Three distinct types of behaviours are identified and a discrete choice model for route choice integrating quality of service is developed.


In urban transit networks, passengers are often faced with a variety of options for their journeys. These options are sometimes limited to the choice of a stop at which to board or alight a fixed bus service. In more complex networks, they can involve completely different sets of paths or even a choice of mode (light rail versus underground, for instance). A better understanding passenger choices is crucial both for future planning since in many cities, new developments in transit intend to relieve overcrowded lines and there is no clear understanding of the effect of redundant lines, and for transit assignment where the usual hyperpath model is commonly used even though its foundations have seldom been tested.

These choices can be studied via field or online surveys (see for instance an international email survey interrogating the existence of "hypertravelers"). More recently, Automated Fare Collection data have been used to infer passenger choices. In this paper, we use three types of complementary data on a selection of routes of the Paris region, namely field survey results, smart card data and train timetables, both realised and theoretical.

The study was carried out in two phases. In a preliminary phase, commuters from the Juvisy rail station, where two distinct options to reach Paris are available, were asked about their preferred itineraries and factors that may impact their choices. Combined with the train timetable from that day, three different types of passengers were identified corresponding to different types of response to risks. "Risk-averse passengers" (22 percent of surveyed passengers) always use the first service to come, whatever the remaining waiting time for the other option. For destinations in Paris where both options are attractive (in the sense of the optimal strategy model), these passengers can be assumed to travel on a hyperpath. "Smart passengers" (23 percent) try to adopt the quickest option using the departure board on display in the station and their knowledge of the network. It has to be noted that around half of those passengers actually travelled via the longest option. Finally, "routine passengers" (47 percent) always follow the same route, except maybe in extreme circumstances such as strikes or major line disruptions. The remaining 8 percent change their itineraries according to other factors. This survey demonstrates that when two lines run on parallel, the resulting transit option is not just what would be expected from the combined frequency model, with factors such as comfort being preponderant over travel times for many passengers.

In a second phase, smart card data were used concurrently with observed and theoretical timetables. Because in the Paris region, smart card data are recorded upon exiting only in some suburban rail stations and usually not recorded at transfer points, specific journeys where different itineraries leave different smart card traces were selected. For these journeys, the different passenger types described above were identified. The proportions of routine passengers vary greatly according to the journey considered and some factors, such as quality of service, which may explain the differences are brought forward. Comparisons are made between observed traffic flows and flows that would be expected from the hyperpath model. Again, the discrepancies are explained using line and service characteristics. These include crowdedness, transfer difficulty, regularity and mode reputation.

A binary discrete choice model describing users' itinerary choices is then specified and tested. In this model, the indicators describing the services and options are all defined as bilevel. For instance, transfer difficulty was set to 1 for all those options that necessitate leaving the subway network (to a train station, for instance). The results achieve predictions with some accuracy, or at least the results are improved from a model that would only consider travel and transfer times, but there remain several shortcomings on which we are currently working. The flow proportions are calculated for all passengers and not specific types, so that the model does not actually describe passengers decision process. The indicators remain a very vague description of lines and services, and a scientific procedure (rather than an ad hoc one) to calculate them has yet to be found. Finally, for the model to work within a transit assignment framework, a lot of path enumeration is theoretically necessary, which is costly in time.


Association for European Transport