Workshop on discrete choice models 2013

Schedule

Thursday

10:15Welcome
10:30Tutorial: Prof. E. Cherchi
12:00Lunch
13:30F. Ramjerdi
14:00E. Kazagli
14:30Z. Yan
15:00Break
15:20O. Blom
15:50A. Stathopoulos
16:20S. Varotto
16:50Break
17:10A. Danalet
17:40J. Knockaert

Friday

09:00G. Reich
09:30A. Glerum
10:00B. Farooq
10:30Break
10:50I. Saman
11:20X. Lai
11:50D. Efthymiou
12:20Lunch
14:30Workshop discussions

Saturday

09:00Hiking
12:00Lunch
18:00BBQ

Presentations

14 presentations ordered by alphabetical order of the presenter.

Abstracts

Blom Oskar, KTH Royal Institute of Technology
A dynamic discrete choice approach to activity scheduling
We propose a dynamic discrete choice approach to activity scheduling, where an agent in every time step takes the action that maximizes the sum of the direct utility and the expected maximum future utility. By connecting each state in the end of one day with a state in the beginning of the next, long term between day planning of, e.g, shopping can be modelled. The large number of possible states and actions within a day makes a direct usage of common estimation methods infeasible. To overcome this problem, we estimate the model in two steps: first, using sampling of alternative action sequences to capture the within-day choices conditioned on the between day planning; and second, using a nested fixed point algorithm to estimate the parameters determining the between day planning. Monte Carlo evidence shows the computational feasibility for realistic-sized applications.
Danalet Antonin, EPFL ENAC INTER TRANSP-OR
Revealed preference data from WiFi traces for pedestrian activity scheduling
We use communication network infrastructure, in particular WiFi traces, to detect activity-episodes sequences in a pedestrian facility. Due to the poor quality of WiFi localization, a probabilistic method is proposed that infers activity-episodes locations and durations based on WiFi traces and calculates the likelihood of observing these traces in the pedestrian network, taking into account prior knowledge. The output of the method consists in generating lists of activity-episodes sequences with their likelihood. Results show that it is possible to predict the number of episodes, the activity-episode locations and durations, using activity locations on the map, WiFi measurements and capacity information. The output of our model is useful for modeling pedestrian activity scheduling and the impact of schedules on pedestrian travel demand.
Efthymiou Dimitrios, EPFL TRANSP-OR
Modeling the willingness to join carsharing using latent class discrete choice models and mixed internet/paper survey data
Carsharing forms an alternative, sustainable transportation mode that can be easily implemented at a local, municipal level. Its service presents characteristics from both private and public transportation, something that makes it unique and attractive to specific proportion of the population. The question usually been raised before the implementation of a carsharing scheme, is about the characteristics of the future users. Many researchers have been tried to answer the question so far, however the restrictions of using biased, stated preference data remain unsolved, while the survey cost is high. In this paper, we model the willingness of people to join a carsharing scheme using mixed Internet and paper survey data. We developed an ordered logit model with two latent classes: one for the satisfaction about the current travel patterns, and a second for the environmental consciousness. The aim of combining the two datasets is to measure the bias generated by the Internet respondents, who are more prone to be positive in their responses in order to satisfy the interviewer. This is verified in the current research by the positive sign of the scale parameter applied at the utility function of the Internet-based data. The use of latent classes enhances the model estimation, by measuring the parameters that determine the respondents' latent behaviors.
Farooq Bilal, TRANSP-OR EPFL
Generation of synthetic population: A Simulation and Graph Theoretic Approach
Data on the entire population is almost never publicly available. Moreover, there is an alarming trend of discontinuing the exercise of conducting full Census in many countries (Belgium, Switzerland, etc.). In this context, population synthesis techniques have been developed for policy analysis and forecasting. Currently, the focus is on treating synthesis as a fitting problem. For instance, Iterative Proportional Fitting (IPF) and Combinatorial Optimization based techniques. The key shortcomings of fitting based procedures include: a) synthesis of only one weighting scheme, while there can be many solutions b) due to cloning rather than true synthesis of the population, losing the heterogeneity that may not have been captured in the microdata c) over reliance on the accuracy of the data to determine the cloning weights d) poor scalability and convergence with respect to the increase in number of attributes of the synthesized agents. In order to overcome these shortcomings, we propose a Markov Chain Monte Carlo (MCMC) simulation based approach. Partial views of the joint distribution of agentąs attributes that are available from various data sources can be used to simulate draws from the original distribution. The problem of association of different types of agents (person-households) is then treated as a maximum weight problem of a bipartite graph. The real population from Swiss census is used to compare the performance of simulation based synthesis with the standard IPF. The standard root mean square error statistics indicated that even the worst case simulation based synthesis (SRMSE=0.35) outperformed the best case IPF synthesis (SRMSE=0.64).
Glerum Aurélie, EPFL TRANSP-OR
Enhanced measurement equations for latent class choice models
Integrating psychometric indicators in latent class choice models allow to enhance the characterization of the latent classes. In this research, we consider measurement equations that include socio-economic indicators of the decision-makers. We show that such a specification increases the significance of the parameters relative to the class-membership relation and leads to a better interpretability of the behavior of individuals in the latent classes. The method is applied to a transportation mode choice case study in Switzerland.
Kazagli Evanthia, TRANSP-OR, EPFL
Incorporating "Mental Maps" in Route Choice Modeling
We are interested in advancing the understanding and modeling of route choice behavior. Route choice analysis consists in identifying the route that a traveler would take/ choose to go from one location (origin) to another (destination) in a transportation network. We focus on car route choice. Among the dimensions of travel behavior, route choice is probably the most challenging to be understood, modelled and predicted. This is due to the high requirements in data, the physical overlap of paths - resulting in complex correlation structures - and the large choice set. Finding ways to facilitate route choice modeling, and rendering it at the same time consistent with how people perceive their paths, arises as an issue of great interest. This motivated the idea of incorporating mental maps in route choice modeling. More specifically, we are interested in building a modeling framework where the route choice decisions take place in a higher/ conceptual level. Path alternatives are constructed as -replaced by- sequences of mental geo-marked components ("anchor points"). This is in contrast to the current route choice modeling approach that is based on link-by-link representation of paths on the real physical network. The proposed approach requires to formally define and operationalise the concept of "anchor point" that serves as the fundamental element in order to give structure to the mental map. Our intention is to use theory to guide the modeling and a smartphone dataset to support the modeling assumptions. We are planning to conduct a survey on travelers to illustrate the concept and gain intuition in how the mental maps are formed as well as how they relate to choice behavior. Furthermore, we intend to exploit the smartphone dataset from which we obtain observed followed paths to derive the "anchor points". For instance, the most frequently traversed points/ segments of the network that are used by many users can signify the commonly recognised anchors.
Knockaert Jasper, VU University Amsterdam
Consumer Heterogeneity in Adopting Electric Vehicles: a Latent Class Approach
While alternative fuel vehicles have strong potential advantages in reaching energy security, environmental and climate goals, their adoption is yet slow. In this study, we use data from an SP experiment among ca. 3,000 Dutch drivers to elicit individual preferences for full electric vehicles. We estimate a panel latent class model based on the total costs of ownership approach that includes monetary and non-monetary costs, and sketch psychological profiles of potential car user types. We identify 4 classes with distinctly different preferences for EVs and their attributes that are largely determined by individual knowledge, attitudes and environmental identity. Further we find that one of the important barriers for EV adoption is time spent on fast (station) charging. Implications for policy and practice are discussed.
Lai Xinjun, EPFL TRANSP-OR
Paired Route Impedance Correction for Multinomial Logit Model Based on Equivalent Impedance
A closed-form logit-style formulation basing on route impedance correction is proposed to alleviate defeats caused by the independence of irrelevant alternatives (IIA) and the homoscedasticity properties of multinomial logit (MNL) route choice model. The algorithm utilizes the traditional correction method to add addition impedance to each route by route pair combination with improved correction algorithm. For each route pair, a binary logit model with the concept of Logit Equivalent Impedance is utilized to calculate the route impedance correction, which is derived from two assumptions: (1) the route choice probabilities are independent of the overlapping part so that IIA property can be alleviated; and (2) the variances of route impedances are proportional to route impedances to resolve the homoscedasticity issue. The closed-form structure and easy computation of original MNL model remain unchanged. Numerical examples show that the proposed approach produces more reasonable results than traditional models with same complexity of computation, and is more stable when the number of routes in reasonable route choice set changes.
Ramjerdi Farideh, Institute of Transport Economics
Heterogeneity of preferences for infrastructure and services among cyclists
Cyclists are a very heterogeneous group. The focus of this paper is on the heterogeneity of value of time as well as the monetary values of infrastructure and services among cyclists with implications for policy design and appraisal. Using cluster analysis and latent class modeling approach, we identify distinct classes of cyclists with different values of travel time and different valuations of cycle infrastructure and services. We compare the results with those based on mixed logit models. We use Cycle Study data from the 2009 Norwegian value of time study.
Reich Gregor, Business Department, University of Zurich
A test of the extreme value type I assumption in the bus engine replacement model

Nearly every dynamic discrete choice model relies, for computational simplicity, on the assumption that underlying utility shocks are distributed extreme value type I. In this note we test this assumption in the context of the Rust (1987) classical model of bus engine replacement and find that, for most specifications tested, extreme value type I errors cannot be rejected. In the cases where extreme value type I errors can be rejected, our more flexible estimation yields significantly different choice probabilities at some regions of the state space.

This is joint work with Bradley J. Larsen, Florian Oswald and Dan Wunderli.

Sarman Igor, Institute for Economic Research (UniversitĂ  di Lugano)
Acceptance of life-threatening hazards among young tourists: a stated choice experiment
In this work we analyze directly the role of potential life-threatening events in the choice decision to take a leisure travel or not, as well as to calculate the trade-offs between various attributes characterizing a leisure travel and risky events at the destination. The presence of life-threatening events is a phenomenon of contemporary tourism, especially in certain areas of the world. While such improbable events when happening have massive consequences on tourism in the short run, they seem to be implicitly accepted and evaluated by tourists visiting potentially risky destinations. Tourism literature has produced various works on the issue and the interest on the topic is significant (e.g. Soenmez and Graefe (1998), Lepp and Gibson (2003), Kozak et al (2007), Jonas et al (2011)). However, such studies in general do not focus on the preferences and behavior of individuals when they actually face the choice of where and how to spend holidays in the presence of this kind of risks for which no probability equivalent exists, and where recommendations, personal travel experience, attitudes and risk perception play a role. In order to attribute a monetary value to the acceptance of such life threatening events, we propose to apply a stated choice experiment in which participants face the decision to travel to holiday destinations in which such hazards are present. The framework of the research is confined to four types of life-threatening hazards (which in the tourism-and-risk literature are commonly comprised in the "physical risk" category): in our case such events are terrorist acts, political insurrections, natural catastrophes and epidemics. Moreover, we decided to focus on the South-East Asia region which represents an interesting case because it is a continuously growing tourist destination where all the negative events we are interested in are simultaneously present. To analyze our data we used an Integrated Choice and Latent Variable (ICLV) model which is a special case of the family of Hybrid Choice Models (HCMs). This methodology allows us to implement the classical discrete choice framework and simultaneously to test the significance of attitudes and risk perception on the decision process To pursue the research objectives a structured questionnaire is being administered to a sample of university students, natives of different countries but currently living and studying in Lugano, Switzerland. This heterogeneous sample where different national and cultural backgrounds are present will permit to test for variations in acceptance of the presence of life threatening events. The choice of focusing on young people is driven both by the importance that such a segment has for modern tourism and by the need to restrict the research to a context in which personal travel experience is still limited and therefore the assessment of improbable hazardous events is less influenced by it.
Stathopoulos Amanda, EPFL TRANSP-OR
Dynamic vehicle ownership forecasting: a framework to model inter-temporal renewal decisions
We develop a model of vehicle transaction decisions that account for the behavioral principles that govern renewal. The proposed model framework sheds light on the transaction timing and type of vehicle ownership as a function of several time-varying determinants. Assuming a first-order process a transition probability matrix of shifts across aggregate ownership of different vehicle types is derived empirically. Modeling draws on a rich repeated cross section dataset with disaggregate observations of new and replacement car ownership decisions over two decades. The model jointly accounts for behavioral variables, institutional factors (e.g. buyer-incentives) and vehicle market dynamics (launches and diffusion) to forecast timing of acquisition and type of vehicle. The findings will aid the prediction of vehicle transactions controlling for the impact of critical policy variables such as scrapping incentives as well as pin down the diffusion patterns of novel vehicle technologies or models.
Varotto Silvia Francesca, EPFL ENAC INTER TRANSP-OR
Modeling travel time perception: an integrated choice and latent variable approach
Travelers usually overestimate and underestimate the actual travel time they experience and this perception error influences their travel decision. In a transportation mode choice case study in Trieste (Italy), it could be noticed that the mean of objective travel time distribution (i.e. measured by devices such as an assignment model and Google Maps) does not matches the mean of subjective travel time distribution (i.e. travel time reported by respondents for the chosen alternative). In addition the 93 % of the users indicated a travel time that is multiple of 5 minutes, demonstrating a rounding effect. The aim of the inclusion of different travel time indicators (reported and calculated travel times) is to investigate the underlying travel time perception of travelers and how this perception influences the modal choice, increasing eventually the forecasting power of the choice model. A latent variable model for the true travel time is integrated into the discrete choice model. The calculated travel time and the reported travel time are used as indicators of the true travel time, assuming that the reported one is affected by socio-economic variables. Eventually a mixed discrete-continuous distribution of travel time is introduced in the latent variables model, explicitly addressing the rounding of travel time.
Yan Zifei, KTH Royal Institute of Technology
Estimation of travel choice models with unobserved attributes that are imputed from traffic assignment models
The increasing availability of detailed, individual-level mobility data enables the estimation of more and more complex discrete choice models of travel behavior. The corresponding choice contexts, however, may not be as easy to observe and hence need to be imputed. In this work, we focus on the effect of uncertainty in the modeling of travel time attributes for very simple route choice models. Such uncertainty results in particular if the decision maker is exposed to a changing and/or stochastic environment and the analyst is unware of the concrete information acquisition and learning protocol implemented by the decision maker. The characteristics of two different estimators of model parameters (specifically, the travel time coefficient) and of attribute uncertainties (specifically, a parametrized travel time covariance matrix) are analyzed. Travel times are assumed to be unavailable to the analyst and hence need to be imputed in the estimation. In brief summary, the joint estimation of travel times and their (co)variances appears feasible. This estimation will be used in real traffic network in our future work. We also believe it could be used to estimate many other models from different areas.