English only
INTER  >  TRANSP-OR > Workshop on Discrete Choice Models

Workshop on Discrete Choice Models

August 30 - September 1, 2007

Ecole Polytechnique Fédérale de Lausanne - Room CO 121 (click here for the interactive map)

The 2007 workshop will be organized in the same spirit as the previous ones: an informal meeting for the exhange of ideas around discrete choice models, with the objective to trigger new collaborations, or strenghten existing ones. At the end of the workshop, a list of potential collaborations will be identified, with specific objectives.



There is no registration fee. Everyone interested is invited to attend. Presentations are upon invitation only. All participants, including speakers, must register with the following form.



Click here for the itinerary between EPFL's main information desk and the room where the workshop will take place.

Thursday Aug. 30, 14:00 - Friday Aug. 31, 12:00
Friday Aug 31, 14:00 - 17:00
Workshop meeting
Saturday Sep. 1, 9:00 - 12:00
Hiking and paragliding
Saturday Sep. 1, 12:00
Genuine swiss fondue in a chalet up in the mountain

Tentative schedule.

Thursday August 30, 2007
14:00 Michel Bierlaire Welcome
14:10 Piet Bovy The Path Size Concept revisited: alternative derivations and alternative measures
14:45 Anders Karlstrom Dynamic discrete choice modelling
15:20 Michel Bierlaire Circumventing the problem of the scale: discrete choice models with multiplicative error terms
15:55 Otto Nielsen Trip-based route choice models – A method to eliminate aggregation bias in activity-based models
16:30 Break
17:00 Shlomo Bekhor Methodological transferability in route choice modeling
17:35 Javier Cruz & Thomas Robin Modeling Human Perception of Facial Expressions by Discrete Choice Models
18:10 Stefan Mabit Studies of willingness-to-pay indicators for public transport
Friday August 31, 2007
9:00 Emma Frejinger Choice set generation for route choice models using a sampling approach
9:35 Elisabetta Cherchi On the use of Mixed Logit models with RP/SP data: accounting for non-linearities, systematic and random heterogeneities
10:10 Coffee break
10:40 Nadine Schüssler Similarities in urban and suburban route choice decisions
11:15 Zachary Patterson Mode and Carrier Choice in the Quebec City - Windsor Corridor: A Random Parameters Approach
12:00 Lunch
14:00 Workshop meeting



EPFL-rates apply to the following hotels:

In Lausanne

Swiss Youth Hotels
Bois-de-Vaux 36, 1007 Lausanne
tél: +41 21 626.02.22
fax: +41 21 626.02.26
adresse email:
single: 78.00
Hôtel Elite
Avenue Ste-Luce 1
1003 Lausanne
tél: +41 21 320 23 61
fax: + 41 21 320 39 63
single: 117.00; double: 174.00
Hôtel Alagare
Minotel Suisse
Rue du Simplon 14
1006 Lausanne
tél: +41 21 617 92 52
fax: +41 21 617 92 55
single 105.00; double: 150.00
Hôtel Alpha-Palmiers
Fassbind Hotels
Rue du Petit.Chêne 34
1003 Lausanne
tél: +41 21 555 59 99
fax: +41 21 555 59 98
single: 158.00; double: 220.00

EPFL area

Hotel Pré-Fleuri***
Rue du Centre 1, 1025 St-Sulpice.
Tél. 021 691 20 21
Fax 021 691 20 20
Price for a single room around CHF 150.-
Motel des Pierrettes** St-Sulpice
, 10 minutes walk to EPFL
Route cantonale 19, 1025 St-Sulpice
It has no web-site but you can call at +41 21 691 25 25.
It has no restaurant.
Price for a single room, around CHF 110.- (special price for EPFL hosts)
Hostellerie du Débarcadère
Chemin du Crêt 7, 1025 St-Sulpice,
It belongs to "Relais& Châteaux"
Price for a single room around CHF 170.- (special price for EPFL hosts)
Novotel Lausanne Bussigny
35, Route de Condémines, 1030 Bussigny
(15 minutes by car, no bus possibilities)
Price for a single room, around CHF 200.- (special price for EPFL hosts)



The easiest way to get to EPFL is to take the train from Geneva Airport to Renens. In Renens, take the light-rail (called M1) towards Lausanne. There is a stop at EPFL. The travel time is about 1 hour.

A map of the bus and metro network can be found here and time tables are available at the Lausanne Transport web page. Note that tickets must be bought before departure in machines only accepting coins. The price for a one-way ticket from the center of Lausanne to EPFL is 2.80 Fr. (two zones) and most machines do not give back change.

Check the Swiss Federal Railways website.

To navigate within EPFL, use

List of presentations (11)


Click on the title to download the slides.

List of abstracts

  1. Methodological transferability in route choice modeling by Shlomo Bekhor (Technion - Israel Institute of Technology)

    The search for the shortest path constitutes common practice in actual traffic studies, as this simplistic route choice model enables a universal implementation of traffic assignment and simulation procedures to every network configuration. Since literature illustrates the large effort employed in moving forward from this simplistic approach and providing a more realistic representation of individual route choice behavior, this paper investigates the methodological transferability and the potential universal application of enhanced choice set generation and model estimation procedures. This analysis concentrates on the identification of a “cost-effective” path generation method, able on one hand to replicate the highest number of routes actually chosen by individuals moving on urban networks, and on the other hand to reduce the computational costs related to the production of unrealistic routes. This analysis focuses also on the joint model estimation for the combined dataset of the two data sources, in order to comprehend differences between utility function parameter estimates and to discuss the potential application of enhanced route choice models in a universal context.

    Click here to download the presentation

  2. Circumventing the problem of the scale: discrete choice models with multiplicative error terms by Michel Bierlaire (TRANSP-OR, EPFL)

    We propose a multiplicative specification of a discrete choice model that renders choice probabilities independent of the scale of the utility. The scale can thus be random with unspecified distribution. The model mostly outperforms the classical additive formulation over a range of stated choice data sets. In some cases, the improvement in likelihood is greater than that obtained from adding observed and unobserved heterogeneity to the additive specification. The multiplicative specification makes it unnecessary to capture scale heterogeneity and, consequently, yields a significant potential for reducing model complexity in the presence of heteroscedasticity. Thus the proposed multiplicative formulation should be a useful supplement to the techniques available for the analysis of discrete choices. There is however a cost to be paid in terms of increased analytical complexity relative to the additive formulations. (joint work with M. Fosgerau)

    Click here to download the presentation

  3. The Path Size Concept revisited: alternative derivations and alternative measures by Piet Bovy (Delft University of Technology)

    The Path Size concept used in route choice modeling is an attempt to capture the correlations among alternative routes due to spatial overlap. Using the Path Size as a factor in an extended MNL model it allows in an approximate way to correct route choice probabilities because of the spatial correlations. Having a Path Size factor included, route choice models appear to perform better than without. The paper will first address various ways of deriving the Path Size measure in the utility function and will show which underlying assumptions are needed in establishing the Path Size measure. It is shown that various Path Size measures can be derived depending on the network type and the assumptions adopted. In a second part, the Path Size measures will be applied in extended MNL models (so-called Path Size Logit models) to estimate parameters of route choice models and to predict choice probabilities both for real-world networks and to small-scale synthetic networks. A comparison will be shown of the impact of the various Path Size measures in Path Size logit models relative to MNL.

    Click here to download the presentation

  4. On the use of Mixed Logit models with RP/SP data: accounting for non-linearities, systematic and random heterogeneities by Elisabetta Cherchi (University of Cagliari (Italy))

    One basic assumption in mixed revealed preference (RP)-stated preference (SP) estimation is that, although different, the two datasets represent basically the same phenomenon. Thus, we should expect individuals to show the same tastes irrespective of the instrument used to elicit their preferences. However, different and highly significant parameters in each data set have often been found in a multinomial logit model. The same differences actually occur when more complex behaviours (e.g. non-linearities in the level of service attributes and systematic and random heterogeneity in tastes) are accounted for in model estimation. These differences are not a problem from an estimation point of view. However understanding why such differences appear is crucial if models are used for forecasting, as the model structure used in this case differs from the estimated one . Moreover, the flexible structure of the Mixed Logit is the source of potentially serious confounding effects and understanding what parameter (either RP or SP) is estimating the “ture” phenomenon is important for properly interpreting the results obtained and using the model correctly for forecasting. The problems associated with the use of Mixed Logit models estimated with joint RP/SP data in forecasting have received little attention in the literature and the case where both RP and SP data show different effects has not been addressed. Using real and simulated data we provide evidence that accounting for more complex behaviours significantly improves the estimation results, but very often (i) it is difficult to disentangle the “true” phenomena, (2) these more complex models have a very narrow range of applicability, (3) the microeconomic conditions on individual behaviour are more difficult to fulfil and (4) an erroneous use of these specifications may have a major impact on the predicted results.

    Click here to download the presentation

  5. Modeling Human Perception of Facial Expressions by Discrete Choice Models by Javier Cruz Mota (TRANSP-OR, EPFL)

    Facial expression recognition is a hard and ambiguous problem in computer vision. It is hard due to the wide variety of faces and the wide variety of ways of representing the same expression. And it is ambiguous because, even though common approaches treat it as a classification problem, actually when looking at the same scene, different people can feel a different expression. In this presentation, we will show preliminary results obtained in an ongoing project of Discrete Choice Modeling of human perception of facial expressions. This new approach allows to exploit the mentioned heterogeneity of perceptions in a population of "experts" interpreting a facial expression, where an expert is somebody able to discern between different facial expressions, i.e. any human.

    Click here to download the presentation

  6. Choice set generation for route choice models using a sampling approach by Emma Frejinger (TRANSP-OR, EPFL)

    In this paper, we present a new point of view on choice set generation. In contrast to existing approaches, we hypothesize that all paths connecting the origin to the destination belong to the "true" choice set. In this context, we view stochastic path enumeration algorithms as importance sampling of alternatives. For this type of sampling protocol it is necessary to correct the path utilities in order to obtain unbiased parameter estimates. We propose a stochastic path enumeration algorithm that makes the definition of such sampling correction possible. Some preliminary numerical results are presented.

    Click here to download the presentation

  7. Estimating dynamic discrete choice models with MCMC and reinforcement learning by Anders Karlstrom (Royal Institute of Technology)

    Theoretical models used for derivation of measures of travel time savings are inherently static in nature. We argue that we need to allow for sequential decision making in an uncertain environment if we are to derive welfare measures for travel time uncertainties, for instance. However, empirical discrete choice models are usually also static in nature. In this presentation we report of ongoing work on estimating dynamic discrete choice models with the state dimension is “large”. In practice, to solve a dynamic programming problem where the state dimension is large is a difficult problem, and to estimate such model is even more difficult by order of magnitude. We discuss the state-of-art methods for estimating dynamic discrete choice models in the context of a activity based model. We report on the progress to solve such problems using MCMC methods in conjunction with reinforcement learning techniques.

    Click here to download the presentation

  8. Studies of willingness-to-pay indicators for public transport by Stefan Mabit (Technical University of Denmark)

    One of the main applications of discrete choice models in transportation research is the estimation of the value of travel time and other relevant willingness-to-pay (WTP) measures. Since WTP indicators are of great importance in transport appraisal it is crucial to avoid bias in the estimates. The newest methodology in estimation is based on mixed logit models, which allows for flexible representation of taste heterogeneity, but at the same time opens an array of pitfalls for the modeller. One of the pitfalls is to assume rigid correlation structures between coefficients based on convenience. Mabit and Nielsen (2006) investigated correlated WTP in public transport. In the conclusion they note that the investigation suffered from two important limitations: small sample size and few background variables.

    The primary objective of this paper is to overcome these two limitations using a new Danish 2004 data set. These data have a size that makes the investigation of correlated WTP possible for the different modes independently. Furthermore, the data contain many background variables so that their interaction with the correlation can be thoroughly investigated.

    The secondary objective of the study is to apply the newest developments within the mixed logit field in an applied model. Three procedures not yet common in application that will be used are

    1. Tests of the distribution using discrete mixture models
    2. Tests of the shape of the distribution using the framework of Fosgerau
    3. Formal tests of cross-section specification versus panel specification
    Preliminary results show that only the access-egress and headway coefficients have significant mass points at zero. The remaining coefficients only have support on the negative half axis. Furthermore, the results show that the primary source of correlation between coefficients can be attributed to random scale in the sample.

    The data used contain the modes bus, s-train, metro, and train. They consist of binary stated preference observations with attributes cost, in-vehicle time, access-egress time, headway, wait time, and number of interchanges. There are 7751 observation for bus, 1855 observations for s-train, 876 observations for metro, and 3455 observations for train.

    Click here to download the presentation

  9. Trip-based route choice models – A method to eliminate aggregation bias in activity-based models by Otto Anker Nielsen (Technical University of Denmark)

    The paper compares trip-based assignment procedures with traditional matrix-based procedures. The main benefits of trip-based procedures are 1) that the full information about trips from the demand modelling can be utilised in the assignment, and 2) that the calculation of Level of Services (LoS) are consistently feed back into the demand model. This is especially beneficial in activity-based models, as the detailed casual relationships in the demand model can then be reflected in the assignment procedure. Traditionally, trip-based assignment models are rejected due to calculation times. In the paper it is shown, that this indeed may not be a valid argument concerning the tendency to increase the number of zones, time of day intervals and trip purposes in demand models, including especially activity-based models. It is shown that the theoretical calculation complexity of large-scale models may indeed be comparable or even smaller in trip-based assignment procedures than in traditional matrix-based. This is exemplified on the Copenhagen traffic model. The core issue concerning calculation complexity is that the trip-based assignment depends on the number of trips and the network size. The zone-based models depend on the number of matrices, zones and the network size. This means, that the trip-based models are slower than zone-based, if the network is small or the number of zones and trip-matrices is low. In small cases calculation time is usually not an issue though. If – however – the model consists of many zones and matrices, then the trip-based assignment seams to be more efficient in terms of calculation time.

    Click here to download the presentation

  10. Mode and Carrier Choice in the Quebec City - Windsor Corridor: A Random Parameters Approach by Zachary Patterson (EPFL - TRANSP-OR)

    The Quebec City-Windsor corridor is the busiest and most important trade and transportation corridor in Canada. The transportation sector is the second largest contributor to greenhouse gas (GHG) emissions in the country. Governments around the world, including Canada, are considering increased mode share by rail as a way to reduce transportation emissions. Evaluating the potential of freight mode shift as a means to reduce transportation emissions requires rigorous analytical models that can predict the effect of government policy on mode split. This paper presents the findings of a random parameters mixed-logit model of shipper carrier choice in the Quebec City-Windsor Corridor. The model itself was developed using the results of a stated preference survey undertaken in the fall of 2005. The survey was designed explicitly to evaluate shipper preferences for the carriage of intercity consignments, and particularly for their preferences for carriers that contract the services of rail companies to carry these shipments via rail. A fixed parameters approach suggests that shippers are very mistrustful of using rail to move their consignments and suggests that increasing rail’s share of freight faces tremendous challenges. This result is not entirely consistent with shipper interviews conducted during survey development that suggested some shippers might prefer rail for environmental-public relations reasons. A random-parameters approach was adopted to test whether preference variation across respondents would be able to explain this inconsistency. This random-parameter analysis suggests that there is some variation in shippers’ preferences for the use of intermodal transportation. In particular, the model suggests that for 20% of shippers, knowledge of a carrier’s use of rail has a positive effect. This appears to be the first attempt at a random parameter approach in the freight choice literature.

    Click here to download the presentation

  11. Similarities in urban and suburban route choice decisions by Nadine Schüssler (IVT ETH Zürich)

    To model route choice in urban or suburban areas, the analyst has to take into account a wide variety of influences. Similarities between alternatives belong to the more significant of those factors. Thus, the question of how to treat similarities among routes is still an ongoing research topic. Yet, as urban route choice is so multi-dimensional, another research issue is the best way of monitoring peoples’ decisions. In recent years GPS studies have became more and more popular, because they provide researchers with a detailed record of the actual behaviour and at the same time reduce participants’ burden to a minimum. However, a lot of post-processing is needed to derive trips which can be used for model estimations.

    The work that will be presented at the workshop is based on such a GPS study. 4878 persons living in the Swiss towns Zurich, Winterthur and Geneva were equipped with an on-person GPS receiver and asked to carry it around wherever they went. Each person carried the GPS logger not only for one day, but for several days. In addition to a few socio-demographic characteristics and the GPS records no further information was collected.

    The presentation will present the post-processing procedures that have been employed to derive trips, activities, modes, locations and trip purposes. However, the focus will be on the route choice models for private and public transport that have been estimated on the basis of this data. Special consideration will be given to the influence of similarities on route choice as well as to the role of the choice set and its generation

    Click here to download the presentation

List of participants (20)

  • Amon Patrick, EPFL
  • Bekhor Shlomo, Technion - Israel Institute of Technology
  • Bierlaire Michel, TRANSP-OR, EPFL
  • Bovy Piet, Delft University of Technology
  • Calio Jean, RFF, Forecasting and Economic Planning, Strategy Department
  • CHARGUI Mondher, RFF , Forecasting and Economical Planning, Strategy Department
  • Cherchi Elisabetta, University of Cagliari (Italy)
  • Cruz Mota Javier, TRANSP-OR, EPFL
  • Frejinger Emma, TRANSP-OR, EPFL
  • Karlstrom Anders, Royal Institute of Technology
  • Mabit Stefan, Technical University of Denmark
  • Nielsen Otto Anker, Technical University of Denmark
  • Patterson Zachary, EPFL - TRANSP-OR
  • Robin Thomas, Transp-or, EPFL
  • Schüssler Nadine, IVT ETH Zürich
  • Sorci Matteo, EPFL
  • Thémans Michaël, EPFL TRANSP-OR
  • van Eggermond Michael, ETH Zurich
  • Wiesel Ehud, SNCF, Innovation & Research Department
  • Zander Kerstin, School for Environmental Research, Charles Darwin University, Australia