Workshop on discrete choice models 2014


10 presentations ordered by alphabetical order of the presenter.


Modeling Anger and Aggressive Driving Behavior in a Dynamic Choice-Latent Variable Model
A hybrid choice – latent variable model combined with a Hidden Markov model is developed in order to analyze the causes of aggressive driving and forecast its manifestations accordingly. The model is grounded in the state-trait anger theory; it treats trait driving anger as a latent variable that is expressed as a function of individual characteristics, or as an agent effect, and state anger as a dynamic latent variable that evolves over time and affects driving behavior, and that is expressed as a function of trait anger, frustrating events, and contextual variables (e.g. geometric roadway features, flow conditions, etc.). This model may be used in order to test measures aimed at reducing aggressive driving behavior and improving road safety, and can be incorporated into micro-simulation packages to represent aggressive driving. An application of this model to data obtained from a driving simulator experiment performed at the American University of Beirut is presented. The results derived from this application indicate that state anger at a specific time period is significantly affected by the occurrence of frustrating events, trait anger, and the anger experienced at the previous time period. [joint work with Mazen Danaf and Isam Kaysi]
Börjesson Maria, KTH
Explaining changes in attitudes to congestion pricing before and after introduction
Several cities have reported that public support for congestion charges has increased substantially after charges have been introduced. Several alternative explanations have been suggested, such as larger benefits or smaller adverse effects than anticipated, status quo bias, or reframing processes. We study attitudes to congestion pricing in Gothenburg before and after they were introduced in January 2016, using a two-wave postal survey to measure changes not only in attitudes to congestion charges and beliefs about their effects, but also changes in attitudes to a variety of other issues, such as environment, equity, taxation and price instruments in general. Factor analysis with varimax rotation was used to categorize attitudes according to how they correlate within respondents. Four resulting factors, used as latent variables, together with variables reflecting self-interest and socio-economic status were used to predict the attitude towards congestion pricing before and after the introduction of charges applying an ordered logit model. Attitudes to the Gothenburg congestion charges did indeed become more positive after the introduction. The dominating reason for the change in attitudes seems to be status quo bias, rather than any substantial changes in beliefs about effects, changes in general attitudes or reframing processes, although these factors also contribute to some extent.
Bierlaire Michel, EPFL
Specification of the cross nested logit model with sampling of alternatives for route choice models
A novel approach is proposed to use the Cross Nested Logit (CNL) model in route choice when sampling of paths is considered. It adopts the Metropolis-Hasting algorithm to sample the choice sets for the model. A new expansion factor and an approximation method are put forward to calculate the sampled probabilities of alternatives. We build on state-of-the-art results for the Multivariate Extreme Value models and extend then to the route choice context. Case studies on both synthetic data and a real network demonstrate that the new method is valid and practical. This paper thus provides an operational solution to use the CNL model in the route choice context, where the number of alternatives is particularly large. [joint work with Xinjun Lai]
Cirillo Cinzia, University of Maryland
Modelling car ownership and use in the US

The raising cost of the energy, the heavy congestion on metropolitan motorways and the pollution in urban areas caused by the use of private cars, enormously influence our economies and our lifestyles. Increasing awareness about these problems is expected to affect mobility habits and to create opportunities for changes in the automotive industry in the near future.

An integrated modeling framework for vehicle ownership and use that accounts for several choice dimensions and that is based on a number of policy variables will be presented. The problem that is solved includes a mixture of discrete and continuous variables; therefore, the resulting modeling structure requires the simultaneous analysis of different variables that are not from the same family. The modeling framework proposed is applied to predict behavioral changes in response to the evolution of the society (income), built environment (density), transportation policies (fuel cost) and public transportation infrastructures (metro and bus).

Danalet Antonin, EPFL
Choice set generation for activities using importance sampling
In spatiotemporal choices, such as destination choice, the number of alternatives is very large. Our model decomposes the choice of destinations in time into a first model of choice of activity type in time, aggregating destinations into activity types, and into a second model of choice of destinations, conditional on the activity type and time of day. This decomposition is motivated both behaviorally (people tend to choose first the "pattern" of their day, and then only the specific destination) and computationally (in order to manage the large cardinality of the choice set). The choice set for the first model is still very large: the sequence of activity types in time is combinatorial. Traditional approaches usually propose an ad-hoc and deterministic list of assumptions on the size, composition and variability of the choice set (modeling the consideration set). We propose to consider the universal choice set, containing all possible sequences of activities in time, and to sample sequences of activity types in time according to a given distribution. By performing importance sampling, the goal is to avoid misspecification of choice sets in choice models. We assume that the bias of including unconsidered alternatives is smaller than the bias of forgetting important alternatives for the decision maker. The unconsidered alternatives just have a very low choice probability, while forgetting an important alternative generates endogeneity. The importance sampling is performed by representing each sequence as a path in a network and by using a Metropolis-Hastings sampling of paths algorithm.
de Lapparent Matthieu, IFSTTAR
Decision under uncertainty and dynamic discrete choices: application to single-car owners in France
We develop a finite-horizon (optimal stopping problem) conditional Logit dynamic discrete choice model of car holding duration and use for French households owning only one car over the 2000-2007 period. Given initial conditions on purchase price and fuel consumption, we consider three stochastic state variables (income, fuel prices and mileage) and two deterministic ones (age and cumulated mileage) to explain ownership duration. Accounting for forward-looking economic agents greatly improve the understanding of the underlying logic that drives such choices despite theoretical assumptions and used data may appear as a sketchy description of a more complex reality.
Fernandez Antolin Anna, TRANSP-OR, EPFL
Choice Probability Generating Functions
Choice probability generating functions (CPGF) allow to characterize all (finite mean) random utility models with additive location vectors (ARUM). In this presentation, the concepts needed to show the full characterization of ARUMs by CPGF will be introduced and the main results obtained from this will be discussed, following Fosgerau, McFadden and Bierlaire (2013). Choice probability systems (CPS) will also be introduced and it will be shown that under some properties a CPS, a CPGF and an absolutely continuous finite-mean ARUM are equivalent. As a consequence it can be seen that given the choice probability for one alternative, under some conditions, the corresponding CPGF can be derived and from the CPGF the remaining choice probabilities can be obtained.
Johansen Bjørn Gjerde, Institute of Transport Economics
The effects of travellers' attitudes and perceptions on the demand for high speed rail in Norway

A large-scale study was recently conducted to evaluate the feasibility of high speed rail (HSR) in Norway (Jernbaneverket, 2012). While the study indicated that building HSR in Norway is far from economically feasible, the vast data collected in this context provides excellent possibilities for in depth analyses of the heterogeneity of travellers and the importance of the latent variables that influence mode choice.

During the last decades, much research has been done to better capture heterogeneity among consumers. One approach to address the heterogeneity is "hybrid choice modelling” (see Walker, 2001; Ben-Akiva et al., 2002). The method focuses on estimating the decision making process behind modal choice by including personality traits as latent variables in the utility functions and identification of different latent segments of travellers using latent classes. In this paper, this method is utilized to describe the mode choice between all available modes and HSR in two corridors; Oslo-Bergen and Oslo-Torndheim in Norway.

These personality traits are mainly revealed through indicator variables in the form of questions regarding attitudes and behaviours in daily life. This can for instance be information regarding recycling behaviour to reflect environmental consciousness, or information regarding safety attitude and behaviour in traffic to reflect the preference for safety. The obvious advantage of such indicators is that the information that are not inferable from market behaviour can be included in the decision-making process. If these latent variables are able to capture underlying personality traits, this may account for some of the unobserved heterogeneity and hence make forecasting more reliable.

In addition to capturing individual heterogeneity, the model framework makes it possible to understand how different individual specific characteristics affect the personality traits. This allows for predicting different personality traits for different segments of individuals, and hence one should be able to predict the distribution of personality traits over the whole population. This is of particular interest in the context of forecasting.

By integrating latent variables, the personality traits "comfort" and "global environmental consciousness" can be included as latent variables in the decision-making process to explain the choice between presently available modes and HSR in Norway in the two corridors of interest. In addition, modelling latent classes identifies the discrete segments and the preferences of the travellers in their mode choice.

The modelling framework, a latent class and latent variable model, is applied to data collected on travellers using different modes, car, bus, conventional train and air, in the two corridors. In addition to socio-economic and psychometric data, a stated preference experiment was used to draw inferences on the preferences of the travellers for travel with HSR (see Halse, 2012).

In this paper, we will present the modelling framework and briefly describe the data used in this study.

We conclude that the two latent variables described earlier are significant and increase the explanatory power of the models. Moreover, they affect the choice probability for HSR positively and seem to do a better job in explaining mode choice than the available observable individual specific characteristics. We also identify latent classes of travellers. A multinomial logit model with a conventional separation between business travellers and leisure travellers seem to be outperformed by a joint business/leasure estimation with latent classes in combination with latent variables.

Kazagli Evanthia, TRANSP-OR, EPFL
Revisiting Route Choice Modeling: A Multi-Level Modeling Framework for Route Choice Behavior
The use of random utility models for route choice analysis involves challenges stemming from the large size of the choice set and the physical overlap of paths, the latter resulting in complex correlation structures, but also from the high requirements in data. These factors increase the complexity of the models significantly. Given the difficulty and complexity of designing and estimating route choice models, it is desirable to try to simplify them and facilitate their estimation in large-scale applications. The modeling framework proposed in this paper strives to satisfy this need while maintaining a behaviorally realistic approach. Within the proposed framework, the trade-off between complexity and realism can be explicitly controlled by the analyst, depending on the availability of data and the needs of the application. The innovation and the importance of the approach lies in the potential to break down the combinatorial complexity of the route choice models by replacing the current route representation -and subsequently modeling- which is based on paths consisting in sequences of links on the network, with an aggregate and more abstract representation. The key feature of the framework is the concept of the Mental Representation Item (MRI). This concept is employed to support the new approach for representing the routes in a practical and realistic manner. We argue that the new representation simplifies the choice set problem and reduces the complexity with respect to the correlation among the alternatives. We provide an illustration of the framework using the city of Borlänge in Sweden as a case study.
Sharif Azadeh Shadi, EPFL
Introducing a non-parametric method for customer behavior modeling
In revenue management for transportation companies, accurate forecasts play a critical role to precisely predict demand of each product at a given time. However, the only available information, i.e., registered booking, is constrained to booking limits. Most commonly used uncensoring methods in this context usually try to compensate the missing information by using parametric techniques such as EM. They are usually costly to be implemented when there is a stochastic process assumption for customer arrival rates at different bookings intervals. In this presentation, we address the problem of demand modeling in two different contexts. Then, we propose an algorithm that takes availability constraints into account via a non-parametric mathematical representation. To solve the problem, first, we linearize the mathematical representation by implementing different levels of relaxation. Then, by introducing a tailored branch and bound algorithm and by the help of global optimization techniques we estimate daily potential demand of each product as well as product utilities. The impact of such approach has been tested on revenue performance.