Workshop on discrete choice models 2017



14 presentations ordered by alphabetical order of the presenter. Click on the title to access the PDF file (if available).


Belgiawan Prawira F, IVT, ETH Zurich
Context-dependent models comparisons: Swiss and German SP, RP data sets
When facing alternatives, people are often assumed to choose the alternative which maximizes their utilities. This concept is widely known as random utility maximization (RUM). In transportation research, one of the most famous modeling techniques based on this idea, e.g. for modeling mode choice, is the multinomial logit (MNL) approach. Recently there is a growing interest in an alternative modeling approach, random regret minimization (RRM). In RRM, an individual, when choosing between alternatives, is assumed to minimize anticipated regret as opposed to maximize his/her utility. There are three variants of RRM, the classical CRRM, the µRRM, and the P-RRM. There is also a further approach called relative advantage maximization (RAM) turning the idea around and focusing on the gains. We compare MNL with the four mentioned alternatives. The data used are stated choice data sets collected by the IVT, ETH Zurich which comprise of mode choice, location choice, parking choice, carpooling, car-sharing, etc experiments. We compare the performance of those five models by their model fit (Final LL, hit rate, and prediction). We also present a comparison of their VTTS, travel time and cost elasticities.
Cherchi Elisabetta, Newcastle University
Habitual latent behaviour and dynamic effect of inertia.
The influence of habits, giving rise to inertia effect, in the choice process has been intensely debated in the literature. Typically inertia is accounted for by letting the indirect utility functions of the alternatives of the choice situation at time t depend on the outcome of the choice made at a previous point in time. However, according to the psychological literature, inertia is often the results of a habit, which is formed along a long process where many past decisions (not only the immediately previous one) dynamically influence individual’s behaviour. In this study we use panel data gathered over a continuous period of six weeks, to try to distinguish the tendency to stick with the same alternative, that refers only to the previous trip performed, from the individual propensity to undertake habitual trips, that refers to a longer experience and can change dynamically. Tendency to stick with the same alternative is measured using lagged variables that link the current choice of mode with the previous trip made with the same purpose, mode and time of day. Individual propensity undertake habitual trips, which is captured by the individual specific latent variable, and its effect changes over the 6-week period. We also assume that the lagged effect of the previous trips is not constant but its impact is higher or smaller depending on the individual propensity toward habitual behaviours. We found that the impact of habitual behaviours increases over weeks, though the marginal effect decreasing over weeks.
de Lapparent Matthieu, HEIG-VD
Business location choices in the Paris region: modeling and estimating a static discrete game
This presentation is about how to explicitly account for strategic interactions in discrete choice models in presence of imperfect information. Once detailed the baseline analytical framework and computational challenges, e.g. equilibrium selection and nested fixed point estimation, an application to location choices of newly created establishments in the Paris region is presented. It is found that the latter are willing to share locational profits when entering a location while they are defiant about already installed competitors. As expected, they also are attracted by complimentary industrial sectors.
Fernandez Antolin Anna, TRANSP-OR, EPFL
Discrete Continuous Maximum Likelihood
When using discrete choice models, we often face unobserved correlations between alternatives, where it is appropriate to use nested logit models. Given a finite set of nesting structures, the traditional approach is to estimate the models corresponding to each of them and select a posteriori the most appropriate one based on some fit statistics and informal testing procedures. We propose to integrate the selection of the optimal nesting structure to the maximum likelihood framework of the parameter estimation. We call this discrete-continuous maximum likelihood (DCML). We are able to linearize the logarithm in the objective function so that it results in a mixed integer linear problem.
Hillel Timothy, University of Cambridge
A comparison of classification methods for modelling urban mode choice
Current transport models used to inform policy and network decisions in an urban area typically rely on multinomial logit (MNL) models to predict passenger mode choice. These models make efficient use of sparse data, and can be calibrated to aggregate survey data and passenger/vehicle counts. However, the recent adoption of several notable technologies, including online journey-planners, contactless payment cards, and mobile-phone-based location services, has enabled a step-change of several orders of magnitude in the availability of passenger movement data. This data presents the possibility to make use of machine-learning based classifiers to predict mode choice at an individual level, with higher accuracy than the currently employed techniques. We present a new dataset, covering on the one hand complete trip diaries with personal profile data from the London Travel Demand Survey, and on the other hand the path, distance and duration reconstructed from an online directions service for each transport mode (walking, cycling, combined public transport, bus, rail, and driving). We use this dataset to compare the predictive performance of a complete suite of machine-learning classification algorithms with MNL. We then use the highest performing models to determine the key factors driving urban passenger mode choice.
Kazagli Evanthia, TRANSP-OR, EPFL
Application of the MRI framework to a large network: Québec city
The objective of this work is the application of the MRI framework to a large network. It is motivated by (i) the additional complexity in the definition of the model due to the size of the city of interest, and (ii) the lack of a detailed disaggregate network model. We present a new paradigm for the definition and operationalization of the aggregate graph and discuss concrete specifications that are compatible with the standard estimation procedures. We demonstrate the capability of the framework in dealing with the additional complexity --while remaining computationally affordable-- using revealed preference data from the city of Québec, in Canada. The proposed model provides an insightful understanding and description of the aggregate route choice of individuals. It can be readily applied to the prediction of flows on the major segments of the network. Its specification incorporates the effect of the departure time on the travel time parameters, in an attempt to implicitly capture the effect of congestion.
Knockaert Jasper, VU
Estimating a latent class model
The latent class model has become a preferred model specification for analysing discrete choice behaviour when there is unobserved heterogeneity. Its main advantage over alternative specifications is that it allows for a great flexibility in behavioural specification while the LL still has a closed form expression. However, the estimation of the model is often challenging because of the presence of multiple local optima, even when the number of latent classes is small. In this presentation I discuss an algorithm for random initialisation of the model estimation, its implementation with pythonbiogeme, and I present some insights collected by using the algorithm.
Lurkin Virginie, EPFL
Using non-traditional data sources to understand travel behavior
Many researchers have been exploring ways to use non-traditional data sources to understand travel behavior. This interest has been driven in part by the belief that compared to traditional survey data collection methods, internet-based marketplaces enable one to collect survey data cheaper and faster from a larger, more diverse participant pool. However, many have questioned whether models based on survey data from these online marketplaces are similar to models based on survey data from more traditional platforms. To investigate this research question, we conducted a survey of air travelers on MTurk (a crowdsourcing Internet marketplace) and Qualtrics (a traditional marketing survey panel). Results show that MTurk and Qualtrics respondents exhibit similar air trip characteristics, but distinct socio-demographic characteristics and that after controlling for socio-demographic characteristics, itinerary choice models estimated from the MTurk and Qualtrics survey data are statistically equivalent.
Maione Salvatore, Università della Svizzera italiana
Integrating Travellers’ Heterogeneity in Subscription Choice Processes Through Hybrid Choice Modelling: An Application to the Swiss Railway Market
Public transportation companies often classify their customers into only two classes, i.e. first and second class. Such a rough segmentation largely ignores travellers' specific needs and habits and may thus leave significant heterogeneity within classes. Individuals feel more attracted to social groups that are perceived similar to the self (Davis, 1984). Once part of a group, individuals have the tendency to evaluate their own social group (vs. others) more positively (Efferson, Lalive, & Fehr, 2008; Hewstone, Rubin, & Willis, 2002). As many social groups could be present in the train, a 2-class structure might not be optimal. An increasingly specific segmentation could reduce such issues and increase traveller's utility and overall satisfaction. In this project, we investigate if the introduction of dedicated sections based on travellers’ specific habits (vs. the traditional two-class structure) can provide value to travellers. We argue that the key latent trait influencing travellers' preference for dedicated sections is individuals' degree of outgroup derogation (Hewstone et al., 2002; Turner, Oakes, Haslam, & McGarty, 1994; Vanhoomissen & Van Overwalle, 2010). Outgroup derogation, defined as the perception that outgroup members could be considered as a “threat” for the members of the ingroup (Levin & Sidanius, 1999), is measured by a 12-item 7-point Likert scale. First, in a qualitative pre-study, we classify travellers based on the heterogeneity of their traveling needs and design sections that meet those needs. Then, in the quantitative main study, we measure consumers' preferences for these sections using a discrete choice experiment. More specifically, in the hybrid choice framework, we integrate the degree of outgroup derogation as latent variable in a multinomial logit model (MNL). By doing this, we can analyse the impact of the psychological motivations in the choice process (Bolduc & Alvarez-Daziano, 2010). Moreover, due to the panel data structure, we include an error component (EC) in the MNL to consider the heterogeneity that may be present in the choices (Bierlaire & Fetiarison, 2009).
Oyama Yuki, Tokyo Institute of Technology
Structural estimation for route choice models considering link specificity of measurement error variances
For estimating parameters of discrete choice models, observations corresponding to the models are required. In the context of route choice models, we need the information of paths, which are sequences of links and connect between the origin-destination pairs. Passive monitoring with Global Positioning System (GPS) is more and more used to observe trip data, because it contributes to facilitating to observe trip data automatically. However, data from monitoring with GPS is not consistent, in formats, with network and it has the heteroscedasticity of measurement errors dependently on devices and locations. These errors cause the biased observation of route choices, and as the result, the parameter estimation results of route choice models ca be biased. In this study, we propose a sequential link measurement method, which is a bayesian approach and incorporates a Markovian route choice model as the prior. It allows one to infer links based on both measurements and behavioral mechanisms, and at the same time, to estimate the variance of GPS measurement error on each link. Moreover, we propose a structural estimation method for a route choice model to remove biases regarding the initial parameter settings of the prior. The performances of these methods are examined through a numerical example and a case study of applying in a real pedestrian network (joint work with Eiji Hato)
Pacheco Paneque Meritxell, TRANSP-OR EPFL
Integrating advanced discrete choice models in mixed integer linear optimization
The integration of discrete choice models in optimization is appealing to operators and policy makers (the supply) because it provides a better understanding of the preferences of clients (the demand) while planning for their systems. Notwithstanding the clear advantages, the complexity of discrete choice models leads to mathematical formulations that are highly nonlinear and nonconvex in the variables of interest, and therefore difficult to be included in mixed integer linear problems (MILP), which are the common optimization models considered to design and configure a system. In this research, we present a general framework that integrates discrete choice models within MILP. The abovementioned limitations are overcome with simulation. We illustrate a concrete application on benefit maximization and we test the resulting model on a case study from the recent literature in which a mixtures of logit model is estimated. The results show that this approach is a powerful tool to characterize features of the systems based on the heterogeneous behavior of customers.
Ramjerdi Farideh, Institute of Transport Economics
How to influence the public acceptance of road pricing? The Trondheim experiment
The Trondheim scheme was introduced in 1991, with 12 toll stations and a toll fee of 10 NOK during the peak periods. It was abandoned in 2005. By then the number of toll stations had increased to 29. In 2010, a scheme, with 8 toll stations, was reintroduced. In March 2014 the scheme was expanded by 14 toll station. In this presentation, we focus a public opinion survey that was conducted in the early summer of 2014 in Trondheim. In this survey, in addition to the socioeconomic data and a travel diary, data on attitudes, travel habits, perceptions of taxes, equity and other social policies, perceptions of traffic, parking, public transport, environment, etc. were collected. Before any of the attitudinal questions, we asked respondents to state their stands to the scheme that were introduced in 2010 and 2014 (for, against, or neutral) and the reasons for their stands. After the attitudinal questions, we ask the respondents to state how they would vote for the scheme if there were a referendum today (yes, no, no opinion). Most respondents who were for or against the scheme did not change their stands. Those who were neutral towards the scheme were most likely to change their views. The respondents’ stands on the scheme in 2014 that were stated before the attitudinal questions and their vote on the scheme after these questions were used in modelling the ordered logit mode with latent variables to evaluate how likely is a respondent to change view and the type of variables that persuade them to do so. (joint work with Jasper Knockaert)
Schmid Basil, IVT, ETH Zürich
Investigating suppressed demand effects for increasing car travel costs: A latent variable random effects Poisson (LVREP) approach
The Post-Car World study focuses on intra-household restrictions in travel behavior for changes in generalized transport costs, investigating suppressed demand effects from an activity-based perspective. The main research question addressed in this paper is how respondents change their mileage driven, given the increase in car travel cost. The data analyzed comprises 228 respondents who completed the stated adaptation interviews in the last stage of the survey. For each respondent, one day of the previously conducted travel diary was selected to create the personalized choice sets, showing travel characteristics for each activity conducted and offering different mode alternatives. Travel costs were calculated for current Swiss market prices and were increased over the 4 presented choice situations, with private cars experiencing the largest increase. A latent variable random effects Poisson modeling approach was used to account for the non-negative and highly right-skewed dependent variable. The main explanatory variable is the average car kilometer cost, including interaction terms with socio-demographic characteristics and simultaneously estimating one latent variable capturing environmental sensitivity: If, on average, costs increase by 1%, distance traveled by car decreases by 0.5%. Results indicate heterogeneity in behavior, showing that low income and pro-environmental respondents exhibit a stronger adaptation pattern.
Varotto Silvia, TU Delft
Modeling Choices of Control Transitions and Speed Regulations in Full-Range Adaptive Cruise Control
Driving assistance systems such as Adaptive Cruise Control (ACC) and automated vehicles can contribute to mitigate traffic congestion, accidents, and levels of emissions. Recently, an on-road study have shown that drivers are likely to deactivate full-range ACC when approaching a slower leader and to overrule it by pressing the gas pedal a few seconds after the system has been activated. Notwithstanding the influence of these control transitions on driver behavior, most mathematical models assessing the impact of ACC on traffic flow efficiency and safety do not describe appropriately these events. This research aims to develop a choice modelling framework that describes the underlying decision-making process of drivers in control transitions. Based on previous studies, we propose two levels of decision-making: risk feeling and task difficulty evaluation, and system state and speed regulation choice. Drivers evaluate whether the actual level of risk feeling and task difficulty falls within the range which is considered acceptable to maintain the ACC active and the current target speed. If the actual level falls outside the acceptable range, the driver will choose to resume manual control or to regulate the desired speed maintaining the system active. The model is estimated using a dataset collected in an on-road experiment with full-range ACC. Preliminary estimation results showed that the actual level of risk feeling and task difficulty is higher when speeds are higher, when time headways are shorter, and when approaching a slower leader. Interestingly, the acceptable range is influenced by driver characteristics such as driving styles. The key implication of this study is that, to assess the effects of ACC on traffic flow accounting for control transitions, we need a conceptual framework permitting a rigorous model estimation and suitable for implementation into a microscopic simulation.

Tentative schedule

10:45Farideh Ramjerdi
11:10Virginie Lurkin
11:35Silvia Varotto
13:30Basil Schmid
13:55Meritxell Pacheco Paneque
14:40Timothy Hillel
15:05Eva Kazagli
15:50Anna Fernandez Antolin
16:15Yuki Oyama
09:30Elisabetta Cherchi
09:55Matthieu de Lapparent
10:20Jasper Knockaert
11:10Salvatore Maione
11:35Prawira F. Belgiawan
13:30Workshop discussions
Click here


June 22-24, 2017


EPFL GC B1 10 (click here for a map)


CHF 220


Mila Bender

Transport and Mobility Laboratory (TRANSP-OR)
Station 18
CH-1015 Lausanne

Tel: +41 (0) 21 693 24 08
Fax: +41 (0) 21 693 80 60