# Workshop on discrete choice models 2015

### Presentations

9 presentations ordered by alphabetical order of the presenter.

• Random regret and moral decision making: new insights and a research agenda, by Caspar Chorus
• Activity path size for correlation between activity paths, by Antonin Danalet
• Correcting for endogeneity using the MIS method: a case study using mode choice RP data in Switzerland., by Anna FernÃ¡ndez AntolÃ­n
• Heterogeneity in preferences for electric cycle, by BjÃ¸rn Gjerde Johansen
• A Route Choice Model Based on Mental Representations, by Evanthia Kazagli
• The relationship between length of stay and transportation mode in the tourism sector: a discrete-continuous analysis of Swiss data., by Andrea Pellegrini
• Flexible Mixture-Amount Choice Models, by Aiste Ruseckaite
• Testing Efficients Stated Choice Designs, by Basil Schmid
• Model structure and flexible structure for activity based models, by Yoram Shiftan

### Abstracts

Chorus Caspar, TU Delft
Random regret and moral decision making: new insights and a research agenda
Danalet Antonin, EPFL
Activity path size for correlation between activity paths
In the activity path choice approach, the activity-episode sequence is modeled as a path in an activity network defining the activity type, duration and time of day. The IIA property might not be appropriate in the activity path approach since activity paths share unobserved attributes due to overlaps. Overlaps correspond to performing the same activity type at the same time and might be correlated. In this talk, we first show that the traditional Path Size correction term cancels out in the activity path approach and then propose two different deterministic corrections for correlation, the Primary Activity Path Size (PAPS) and the Activity Pattern Path Size (APPS), for the utility of overlapping activity paths.
FernÃ¡ndez AntolÃ­n Anna, TRANSP-OR, EPFL
Correcting for endogeneity using the MIS method: a case study using mode choice RP data in Switzerland.
The estimation of value of time (VOT) plays an important role in transport economics, since investments in transportation systems will depend on it. However, the estimate of this parameter is potentially biased and not consistent if endogeneity is not taken into account. We present a new methodology to correct for endogeneity: the Multiple Indicator Solution (MIS). The MIS relies only in having some suitable indicators, which are in general easy to obtain. This method is then applied to a transportation mode choice case study in Switzerland, which consists of revealed preference (RP) data. Indicators related to a “car-loving” attitude are used to obtain better estimates of the VOT.
Johansen BjÃ¸rn Gjerde, Institute of Transport Economics
Heterogeneity in preferences for electric cycle
The project InnoBike, financed by TRANSNOVA, aims at evaluations of measures for promoting cycling, the potential demand for E-bikes and measures to increase market shares of E-bikes in Norwegian urban areas. To measure the demand for E-bike, a stated preference (SP) study was developed. The SP study comprise of a vehicle type choice experiment among ordinary bike and E-bike, a mode choice experiment among car, public transport and cycle, and two route choice experiment for cycle that focus on the evaluation of parking facilities for cycle, cycle paths and shower and changing facilities at work or school. This paper reports the analysis of the vehicle type choice experiment with focus on heterogeneity of preferences for El-cycle The vehicle type choice experiment includes the price for E-bike, the battery recharging time, the battery range, the battery life time and the battery replacement costs, whereas for an ordinary bike only the price is included. Furthermore, an outside alternative is included that represent not choosing any of the two alternatives described in the choice set. In the first part of the questionnaire, the respondents report their travel behaviour to work or school (for students and employed), their use of different modes of transport in their daily activities, in addition to a set of questions that explores respondents habits, attitudes and values. The SP experiments are followed by questions related to the socio-economic, demographic and locational variables. The data was first analysed using a factor analysis to identify different classes of travellers. Latent variable models are applied to the four SP data in the study. Respondents were recruited from an e-mail database and the survey was conducted on internet in the summer of 2014. Net respondents were about 1100 with a response rate of about 22%.
Kazagli Evanthia, TRANSP-OR, EPFL
A Route Choice Model Based on Mental Representations
The estimation of random utility models for route choice with revealed preferences (RP) data involves several challenges. We present a new approach for modeling and analyzing route choice behavior that is motivated by the need to reduce the complexity of the state-of-the-art models. It is inspired by the simplifications actually done by the travelers, using representations of their surrounding space. The proposed framework is based on elements designed to mimic the mental representations used by travelers, denoted as \emph{Mental Representation Items} ($MRIs$). We show how operational models based on $MRIs$ can be derived and we present estimation results using RP data from the city of Borl\"ange in Sweden.
Pellegrini Andrea, IRE - UniversitÃ  della Svizzera italiana
The relationship between length of stay and transportation mode in the tourism sector: a discrete-continuous analysis of Swiss data.

Despite scientific literature on the relationship between tourism and transport is wide (Alegre and Pou (2006), very few studies take into account the interdependence between time at destination and transportation mode, given the origin and the destination of the trip. The purpose of this paper is to investigate how the length of permanence at destination depends on a specific mode of transport and vice versa.

We implement a discrete-continuous choice model (Hanemann (1984) and Dubin and McFadden (1984)), which is typically applied when the optimal discrete choice depends partially on the result of the continuous choice. For instance, the implementation of the discrete continuous choice models allows modelling jointly the discrete and the continuous consumer choices from the same utility maximisation problem.

The present work is based on a dataset in which information about the behavior of Swiss tourists is collected. Information regards type of destination, type of origin, transport mode (Martin and Witt, 1988) and travel expenditure (Song and Li, 2010) and these elements are the focus of our work. This study represents the first case in which a discrete continuous choice model is applied to Swiss tourism data.

Our preliminary results confirm, as we expected, that tourists from both periphery and rural area tend to move by private transportation to reach their holiday destination. On the other hand, the public transport seems to be preferable, if the destination is a holiday at seaside. Moreover, the model points out that when the number of travellers increases, the probability of choosing public transportation decreases. This result might be driven by the fact that the marginal cost per an extra person for the private transportation is null, whereas for the public transportation is positive. In tourism literature, duration of the journey and the transportation mode are generally chosen as the main explanatory variables of tourism demand. In this study, we change the perspective in the sense that we want to analyse the link among them and how it influences the tourism demand.

Ruseckaite Aiste, Erasmus University
Flexible Mixture-Amount Choice Models

Many products and services can be described as mixtures of ingredients. Examples are the mixture of different fruits composing a fruit salad (e.g. 50% of apples, 30% of wild berries and 20% of grapes) and the mixture of different transportation modes used by an individual on a particular trip (e.g. 70% of travel time by metro and 30% by bike). In some scenarios, the total amount of the mixture may also be relevant to a choice made by an individual. In such cases, the choice of different mixtures depends not only on the proportions but also on the total amount. For instance, the recognition of an advertising campaign not only depends on the advertising media mix (e.g. 30% of the expenditures on TV advertising, 10% on radio, and 60% on internet) but also on the total budget of the entire campaign. Furthermore, the optimal media mix is also likely to be a function of the campaign budget. Such data are typically called mixture-amount data.

To capture the fact that the impact of the mixture ingredients varies with the amount variable, the strategy has usually been to express the parameters measuring the impact of the ingredient proportions as a parametric function of the amount. Such models require us to specify the functional form of this relation which may not be straightforward. Furthermore, for some flexible functional forms, the number of parameters may become very large.

In this paper, we present an alternative approach which is flexible but parsimonious. Our model is based on so-called Gaussian processes and avoids the necessity to a-priori specify the shape of the dependence between the mixture parameters and the amount. The correlation of the parameters across different amount levels is controlled by a parameter that can be estimated. Next to the correlation structure across amount levels, one may also expect correlations between the parameters at a specific amount level. In our model, we also allow for such correlation.

As a result, we obtain a model where the impact of mixture ingredients potentially differs across levels of an amount variable. Through the Gaussian process prior we control the variation of these parameters across the levels of the amount variable. This results in a model which is flexible, but avoids the risk of overfitting the data. We develop a Bayesian estimation strategy for this model and apply it to an advertising dataset. We show that our model encompasses two other commonly used model specifications as extreme cases.

Schmid Basil, ETH Zurich
Testing Efficients Stated Choice Designs
Five D-ecient Stated Choice designs are created and tested for their reliability and precision in parameter estimates and value of time (VOT) calculations for a specific mode choice example. The designs are tested based on RP data and choice simulations, which is repeated 2000 times in order to get insights into deviations from the predefined a-priori parameters. A significant eciency gain in standard errors due to the exclusion of dominant choice sets or the application of MNL designs cannot be found, and the ecient GLM approach with further design conditions solves the design problem very satisfying. All five designs show good performance in estimating the parameters and accurately reproduce the a-priori values which are of special interest: The values of time.
Shiftan Yoram, Technion
Model structure and flexible structure for activity based models