Workshop on discrete choice models 2016

Presentations

18 presentations ordered by alphabetical order of the presenter.

Abstracts

Abou Zeid Maya, American University of Beirut
Happiness and Choice Models
There has been a lot of interest in measuring, tracking, and modeling happiness among researchers and policy makers recently in various domains (economics, psychology, transportation, life overall, etc.). This tutorial will discuss the concept of happiness or subjective well-being and its relevance for choice models. The tutorial will cover different definitions or conceptualizations of happiness, different measurement approaches and their advantages/disadvantages, causes and correlates of happiness, relationship between happiness and utility, dynamics of happiness, and integration of happiness measures in choice models in static and dynamic contexts using transportation applications as examples.
Antoniou Constantinos, National Technical University of Athens
Measuring the value of *, for * in {privacy, preventing a fatality}
One of the applications of discrete choice models is the indirect calculation of the willingness to pay for various measures (as marginal rate of substitution). In this presentation, two case studies, using stated-preference data from Athens, Greece, will be presented, for the estimation of distributions of (i) the value of privacy and (ii) the value of preventing a fatality (often also referred to as the value of statistical life).
Becker Henrik, IVT, ETH Zurich
Free-floating Car-Sharing Mode Choice Model based on simulated non-chosen Alternatives
The objective of this research is to learn more about customer's actual trade-offs when deciding to use free-floating car-sharing. To that end, all free-floating car-sharing trips from a three-months period were made by a car-sharing operator in Basel, Switzerland, and were combined with trips from a smartphone-based travel diary. For each of the trips in both data-sets, the non-chosen alternatives were simulated using the Google Maps Directions API. Eventually, a multinomial logit model has been estimated to better understand free-floating car-sharing mode choice.
Dubernet Thibaut, ETH Zürich
Models of Friendship Formation for the Generation of Synthetic Social Networks
The topic of the presentation is the estimation and usage of a statistical model of friendship formation, with the aim of generating synthetic friendship networks for the synthetic populations used in multi-agent travel demand simulations. The model is an improvement on the model from Arentze et al. (2013), itself largely based on the logit model. The data used to estimate the model comes from a snowball sample of leisure contacts for Switzerland, which gives precious insights on the structure of the actual social network (Kowald, 2013). Though the authors presented satisfying results on a small synthetic population, using their approach for a reasonably sized population comes with problems. In particular, it requires heuristic calibration during the generation phase to get reasonable clustering and degree distribution. The work presented in the presentation aims at generating a social network for a full Swiss synthetic population. To this end, alternative specifications of the model are being designed, estimated, and used to generate synthetic social networks, with the aim to get transitivity and size of ego-centric social networks directly embedded into the statistical model, rather than coming from an ad-hoc calibration procedure. The aim of the presentation is not only to present the state of the work, but to submit it to criticism and feedback from the audience, to guide further development.
Ehreke Ilka, IVT ETHZ
Investigating field effects on empirical countrywide data from Germany
In transportation so called field effects try to capture social influences on decision makers in behavioral models. Respondent's choices can for example be influenced by people with a similar socio-economic situation (e.g. income class) or who live close to them. Nowadays it is well known that the income elasticity of the value of time is not constant but an increasing function of income, an effect which was also found in recent the German VOT data. Furthermore from a spatial proximity perspective income in Germany is likewise not equally distributed and different patterns in between e.g. rural and urban areas, FRG (Federal Republic of Germany) and GDR (German Democratic Republic) regions or economic disparities in-between certain areas can be recognized. In addition the purchasing power varies in the different regions according to the described differences. The goal of the proposed paper is to analyze patterns of income distribution and regional differences by incorporating field effects in the utility function using countrywide empirical data. In a next step the model needs to be corrected for endogeneity to avoid an upward bias of the field effect.
Fernandez Antolin Anna, TRANSP-OR, EPFL
I would like a new car, but which one do I choose?
We focus on new car purchases using discrete choice models. The first challenge is to define the choice set. We assume that the choice of a new car consists in the choice of a market segment and a fuel type. Aggregating the alternatives in this way has the advantage that the choice set size is treatable, but the drawback that it is not straightforward to impute the attributes of each of the non-chosen alternatives. We use bootstrapping techniques to do so. We compare the estimation results between a multinomial, a nested and a cross-nested logit model. Our dataset contains information about purchases of new cars in France during 2014. It consists of around 43,000 observations, 19,000 of which do not contain missing values. A large percentage of the missing values correspond to non-reported purchasing price. To address this issue, we use an auxiliary regression equation for price.
Floetteroed Gunnar, KTH Royal Institute of Technology
Attempting consistent CBA with complicated transport model systems
Consider a layered transport model system. The upper level is mainly a nested logit model describing location and model choice; no representation of time here. The lower level is a DTA describing route and departure time choice; it is fully dynamic. Both levels are person-based, but there is no individual-level congestion feedback from the lower to the upper level. Question: How to turn this into a model system that can be used to consistently account for population heterogeneity in CBA.
Frejinger Emma, CIRRELT and Université de Montréal
Dynamic programming approaches for estimating and applying large-scale discrete choice models
Dynamic programming (DP) is a powerful technique that can be used to compute optimal policies for sequential decision making problems. Algorithms for solving deterministic and stochastic shortest path problems are based on DP. In this talk we present how DP can be conveniently be used to model route choice behaviour in transport networks when travellers' path choices are given by a discrete choice model. Key features of the resulting so-called recursive models are (i) that they can be consistently estimated without generating any choice sets of paths and (ii) prediction is straightforward and computationally efficient. In this talk, we provide an overview on different ways to account for correlated utilities. The talk is presents part of the dissertation of Tien Mai with the same title, the dissertation is based on seven papers co-authored by Fabian Bastin, Mogens Fosgerau and Emma Frejinger.
Hurtubia Ricardo, Pontificia Universidad Católica de Chile
Identification of relevant attributes for the design of cycling infrastructure using discrete choice and latent variable models
A method to identify which physical and design attributes of cycling infrastructure make it more appealing to different types of users is proposed. Since many of the relevant features of public spaces are subjective (safety, beauty, etc.), a latent variable approach will be used to relate perceived qualitative features with actual physical attributes of the design alternatives and characteristics of the users. A stated preferences survey was conducted, intercepting more than 700 cyclists in Santiago, Chile. In the survey, different design alternatives of cycling infrastructure were presented to respondents in the form of 2D images, including information describing travel time and other attributes of the cycling route for each alternative. Additional questions regarding perceived qualitative attributes (safety, comfort, beauty), mobility habits and socioeconomics were included. An integrated choice and latent variable model is estimated, where the questions about perceived qualitative attributes are used as indicators to measure latent variables that can be structurally related to physical attributes of the design alternative and socioeconomic characteristics of the decision makers
Johansen Bjørn Gjerde, Institute of Transport Economics
Conceptual model of the shippers' choice between sea, rail and road transport
The primary objective is to better understand the factors that influence the shippers' decisions to ship goods by sea or by road and how they interact to produce a certain outcome. We do this by constructing a generic model of the single shipper's choice situation, and embedding it in an equilibrium model with many shippers and with network externalities, in the sense that aggregate volumes by each mode determine the offered freight rate and the freight rates influence the decision of each single shipper. The model takes full account of uncertainty in demand for the goods and in lead times. One of its features is a very detailed representation of transport costs. Another is the upper and lower bounds on vehicle size for the distribution stages and line haul stages. We assume that for each of the transport modes, the shippers choose vehicle size, shipment size (shipments may be smaller than the minimal vehicle size) and reorder point deterministically, in such a way as to minimize the sum of transport costs, non-transport ordering costs, inventory holding costs and stock-out costs. Furthermore, we assume that the probability for choosing a particular mode is determined by a multinomial logit model, in which the (mode specific) logistics costs are the most important explanatory variables. Aggregating over shippers, the expected total transport volumes for each mode can be predicted. Due to economies of scale the transport volume for each mode will affect the freight rate. The procedure is repeated for the new freight rate, until convergence. The model is formulated theoretically, and we are in the process of fitting it to available data.
Kazagli Evanthia, TRANSP-OR, EPFL
Assessing complex route choice models using an abstracted network based on mental representations

In Kazagli et al. (technical report), we aimed at simplifying the route choice problem by modeling the strategic decisions of people --represented by the mental representations of their itineraries-- instead of the operational ones --represented by paths. We introduced the concept of Mental Representation Item (MRI) as a modeling element and we presented the methodology for the derivation of operational random utility models based on MRIs. The use of the MRI as a modeling element enables us to obviate the need for choice set generation and sampling of choice sets.

In this work, we extend the MRI approach in order to apply it to more complex models such as the cross-nested (CLN) and the recursive (RL) logit. The CNL and RL models are of interest as they tackle the two main challenges related to route choice modeling; namely the correlation of the alternatives and the choice set generation. Their estimation however is cumbersome. We are interested in (i) investigating the potential of the MRI approach to break down the combinatorial complexity of these models and (ii) comparing their performance under two representational approaches (MRI and path), using a real case study. For this purpose we use the network of Borlange in Sweden, for which a GPS dataset is available.

Loder Allister, ETH Zürich
How accessibility shapes the landscape of car and season-ticket ownership: A bivariate probit approach
The generalized cost of travel, home location and income are repeatedly named as the main factors in determining the ownership of cars and season-tickets. For revealed preference data, the application of the generalized cost of travel is often challenging. However, we propose to use the Hansen measure of accessibility as an indicator for the generalized cost of travel in the analysis with revealed preference data. We analyze the impact of accessibility to population and employment by private and public transport on the choice of car and season-ticket ownership with the cross-sectional Swiss mobility micro-census 2010 and the National Transport Model 2010. We derive four accessibility variables on the municipality level from the National Transport Model. As this four variables correlate highly with each other, we carry out a principal component analysis in which the first component describes general aspects of accessibility, the second better access by public transport and the third better access to employment. The impact of accessibility on the choice of car and season-ticket ownership is analyzed with two bivariate probit models. One model for the choice of car and Swiss national-wide season-ticket (Generalabonnement) ownership and the second one for the choice of car and local season-ticket ownership. We control with socio-demographic variables, the spatial typology and household’s local access to public transport. The results show that for city center and agglomeration residents increasing levels of general accessibility discourage car ownership and encourage the subscription to the public mode. A similar pattern is observed for better access by public transport and for better access to employment. In the countryside, the overall impact of accessibility is weak. The effect of accessibility on municipality level on the local season-ticket is weaker than on the GA.
Motz Alessandra, Università della Svizzera italiana
A generalized multinomial logit model for the preferences of Swiss households towards the risk of an electricity blackout
Objective of this research is studying the preferences of Swiss households towards the risk of experiencing an electricity blackout, while accounting for their tastes with respect to selected electricity generation techniques. By means of a generalized multinomial logit model, we measure the willingness-to-pay for avoiding an increase in the risk of long and short blackouts, as well as the willingness-to-accept for accepting an increase in the same risk. We detect a significant heterogeneity in these values, suggesting a positive potential for demand response from Swiss households. By computing the posterior distributions of the estimated random parameters, we try to assess whether the observed heterogeneity can be linked to any observable characteristics of the respondents.
Mueller Sven, Karlsruhe University of Applied Sciences
The maximum capture problem with flexible substitution patterns
We consider the maximum capture problem with random utilities. The basic assumption is that a firm wants to locate a given number of facilities in a competitive market where customers choose the facility that maximizes their utility. Utility is treated as random. In the location science literature so far, the corresponding choice probabilities of the customers are given by the multinomial logit model (MNL). There exist several exact mixed-integer linear reformulations to the original NP-hard, non-linear program. Unfortunately, the MNL exhibits the independence from irrelevant alternatives property, i.e., constant substitution between facility locations. In contrast, the so-called mixed multinomial logit model (MXL) allows for flexible substitution patterns. Moreover, the MXL is able to approximate any random utility model arbitrarily close. In this paper we present an intelligible mixed-integer linear program for the maximum capture problem with customer demand modeled by the MXL. Empirical and managerial insights are discussed based on a real world case study that shows the applicability of our approach.
Pacheco Paneque Meritxell, TRANSP-OR EPFL
Choice-based optimization for the integration of supply and demand
During the last decade, the trend has been that of combining customer behaviour models in optimization, since it provides a better understanding of the preferences of clients to policy makers while planning for their systems. These preferences are formalized with discrete choice models, which are the state-of-the-art for the mathematical modelling of demand. However, their complexity leads to mathematical formulations that are highly non linear and non convex in the variables of interest. On the other hand, we are also interested in discrete optimization models where supply and demand closely interact, which is typically the case in transportation. Such models are associated with (mixed) integer optimization problems, whose discrete variables are used to design and configure the supply. The goal of this research is to develop a general methodology which integrates both supply and demand under the framework of discrete choice models whose associated mixed integer linear problems are scalable and solvable within reasonable time.
Picard Nathalie, THEMA, University of Cergy-Pontoise
Homogamy in risk aversion
We compare different models of risk aversion using the stated preference survey MIMETTIC, in the context of transportation, in Paris region. We estimate simultaneously the determinants of risk aversion, the quality of the answers and the relation between spouses' risk aversion levels. Our results indicate that risk aversion is larger for commuting trips that for other purposes, or for women with young children than for men or women without children. We also find a significant positive assortative mating. (with André de Palma and Sophie Dantan)
Schmid Basil, ETH Zurich
Explaining the choice between in-store and online shopping

As part of a comprehensive three-stage survey of 339 participants in the Zurich area in Switzerland, this paper aims at explaining the choice of ICT versus traditional shopping channels for groceries and electronic appliances. A stated preference experiment requests respondents to trade-off different attributes related to their choice between in-home and out-of-home shopping activities, which are partly based on previous information obtained from the travel diary. Explicit experimental assumptions have been made to place the participants in a coherent choice situation.

The relatively high value of travel time savings (VTTS) of about 40 Swiss Francs per hour indicates a large potential for new ICT services, especially when comparing to the relatively low value of delivery time savings (VODT) of less than 16 CHF/time unit (depending on the shopping purpose and reference delivery time). Including a latent variable covering the acceptance level of online shopping leads to a large improvement in the model fit. A higher acceptance level implies a significantly higher shopping cost sensitivity, which can be explained by the larger choice set when considering both online and in-store shopping as possible shopping channels. In such a case, ceteris paribus, the choice between the two channels is mainly price driven.

Sharif Azadeh Shadi, EPFL
Choice-based routing problem in the context of flexible mobility on demand
One of the main challenges of operation managers is to decide about how to offer a mix of products to the customers at a given time with the objective of maximizing the expected revenue as well as maximizing the customers' satisfaction. The expected revenue from an offer set is defined by the price and the demand of each of the offered products. Smart phones and new applications have revolutionized urban transportation and mobility. For example, Uber market share has grown so fast during the last couple of years calling for new challenging optimization models to be solved efficiently. For the first time in this field, we introduce an integrated choice-based optimization model that takes into account the cost of routing for taxis and shared taxis while accounting for customer choice behavior.