previous day
next day
all days

View: session overviewtalk overview

09:00-10:30 Session 9A: Mobility as a Service -- Sharing Autonomous Vehicles
Kara Kockelman (The University of Texas at Austin, United States)
Location: Corwin West
Lama Bou Mjahed (Northwester University Transportation Center, United States)
Hani Mahmassani (Northwestern University, United States)
Interest in Private and Shared Autonomous Vehicles: The Role of Millennial-Specific Factors

ABSTRACT. See attached file.

Helen Pinto (Northwestern University Transportation Center, United States)
Michael Hyland (Northwestern University, United States)
Omer Verbas (Argonne National Laboratory, United States)
Hani Mahmassani (Northwestern University, United States)
Integrated Mode Choice and Dynamic Traveler Assignment-Simulation Framework to Assess the Impact of a Suburban First-Mile Shared Autonomous Vehicle Fleet Service

ABSTRACT. See attached file.

Georges Sfeir (American University of Beirut, Lebanon)
Maya Abou-Zeid (American University of Beirut, Lebanon)
Isam Kaysi (American University of Beirut, Lebanon)
Modeling the demand for a shared-ride taxi and shuttle services in an organization-based context
SPEAKER: Georges Sfeir

ABSTRACT. With the continuous growth of cities and the high levels of traffic congestion, transport modes that can lessen the burden of congestion on commuters and the number of private cars have become a necessity. Ridesharing services emerge as a new mode that could alleviate this problem by providing lower costs than private cars or private taxis and more flexibility than public transport modes with fixed routes and schedules. This study investigates the potential market demand of a ‘shared-ride taxi’ and ‘shuttle’ services in an organization-based context. It will develop several model structures for modeling the weekly frequency of commuting by shared-ride taxi and shuttle services, if they were implemented in an organization-based context, and by the current mode of commute. The different modeling structures will be compared based on the variables estimates’ signs and different statistical goodness-of-fit measures such as the adjusted rho-square statistic, the robust t-tests, and the likelihood ratio test. The estimation results will be used for policy analysis to determine market adoption rates of the proposed modes under different pricing scenarios and services’ attributes.

Krishna Murthy Gurumurthy (The University of Texas at Austin, United States)
Kara Kockelman (The University of Texas at Austin, United States)
Shared Autonomous Vehicle Fleet Simulation to Determine Fleet Requirement Using Cellphone Data from Orlando, Florida

ABSTRACT. INTRODUCTION Traffic safety and congestion are key transportation issues for many regions around the world. Driver error remains the leading cause of vehicle crashes (NHTSA, 2015) and rising vehicle-miles traveled (VMT) is worsening traffic congestion (FHWA, 2017). The introduction of autonomous vehicles (AVs) for personal use may dramatically reduce vehicle collisions by eliminating driver error. AVs will also improve mobility options for many travelers, especially those without driver’s licenses. Over the past 10 years, several contributions have been made to optimize and/or implement dynamic ride-sharing (DRS), with various researchers suggesting that DRS is a key method for reducing future roadway congestion (Levofsky and Greenberg, 2001; Berbeglia et al., 2010; Ma et al., 2013; Farhan and Chen, 2017; Levin et al., 2017). More recently, DRS has been successfully demonstrated using agent-based models (see, e.g., Fagnant and Kockelman, 2016; Bischoff et al., 2016; Loeb et al., 2017; and Hörl, 2017, such as MATsim (Horni et al., 2016) and a synthetically generated dataset of people and journeys to simulate dynamic traffic conditions. When it comes to actual trip-making, mode choices, and traffic patterns, DRS has been investigated for cities like Atlanta, Georgia, Taipei, Taiwan, and New York City. DRS applications include the entire U.S. state of New Jersey and the nation and Singapore, using travel demand model trip-making predictions, publicly available taxi datasets, and/or synthetically generated itineraries. Investigations demonstrate system feasibility and/or assess the computational efficiency of different methods for assigning vehicles and/or matching travelers in shared rides. (See Agatz et al., 2011; Santi et al., 2014; Alonso-Moro et al., 2016; Brownell and Kornhauser, 2014; Bhat, 2016; and Tao, 2007.) Agatz et al. (2011) developed a sophisticated algorithm to match riders to their drivers and conducted a simulation using person-trip data obtained from Atlanta’s travel demand model. Their results suggest that DRS works well not only in high-density, high-use settings, but also in sprawling suburbs and at low rates of utilization. However, they focused on driver (and thus TNC vehicle) unavailability, which can hamper sharing and dilute DRS opportunities. Brownell and Kornhauser (2014) focused on SAV system performance for the state of New Jersey. Employing a gridded-network for the entire state, along with synthetic trip-making data, valuable precision, accuracy, and applicability may have been lost in assessing optimal fleet requirements. Santi et al. (2014) and Alonso-Moro et al. (2016) overcame both these issues by using publicly available taxi datasets for New York City and real networks (via OpenStreetMaps, an open-source platform for map data). Alonso-Moro et al. observed that 98% of the City’s 3 million taxi trips could be served with just 2,000 vehicles and low waiting times (averaging just 2.8 minutes), backing DRS capabilities. Bhat (2016) confirmed those New York City taxi results, and added a vehicle repositioning algorithm. Tao (2007) also used a taxi data set, but for the city of Taipei. He developed a heuristic DRS algorithm using real-time taxi movements (not just trip calls by travelers) to test its’ efficiency in a realistic network setting. Tao (2007) achieved 60% ride matches and concluded that a higher matching rate could be obtained across larger networks with greater density of trip-making. Of course, taxis do not represent all person-trips in any region. Such trips tend to be shorter than household-vehicle trips (due to their cost), more often for business reasons or those without parking access (again due to their cost), and for visitors (due to their unfamiliarity with the region). DRS investigations of more representative trip-making are desired. By using a population-weighted cellphone dataset, as done here, one overcomes the drawbacks of faked or taxi-based trip patterns. However, certain details are lost (such as trip-to-trip connections throughout the day), in order to protect travelers’ privacy, over space and time. Thus, cell-phone-based trips or other forms of extensive diary data tend to be aggregated by traffic analysis zones (TAZs) or neighborhoods, to obscure home and work addresses. To keep data size manageable (for dataset sharing), trips are often aggregated into hourly or multi-hour time-of-day bins as well. More detailed trip ends and trip schedules can be simulated/faked and disaggregated, while preserving the population’s basic trip patterns. This process ensures that matches are less obvious (with trips coming from all over a zone and hour, rather than from its centroid or mid-point, for example), and was used here. But it comes at the expense of some accuracy and precision (versus the reality of actual trip locations and times, which are rarely available to anyone, for any large population). CELLPHONE DATASET The cellphone-based dataset employed in this study was generated by AirSage for the month of April 2014 and for travel across the Orlando metropolitan area in Florida. AirSage uses the regular location pings of cell phones that are turned on and carried by customers of its partner companies (like Verizon and Sprint). Cellphone trips observed were aggregated based on six factors: each trip’s inferred origin and destination TAZs, the hour and day in which most of the trip was made (e.g., 0100-0200 on April 4 or 1600-1700 on April 20), inferred trip purpose, and cell-phone subscriber class. All trips (and basic demographics) inferred from phone pings (of the carriers’ cell towers) were then expanded to reflect all trip-making in the region (including travel by persons who do not own cell phones or carry theirs with them, turned on). The Orlando region’s metropolitan planning agency (MetroPlan Orlando) models travel across 1,267 TAZs (with 1,261 of them representing metropolitan area and the remaining 6 representing external TAZs). External-zone trips can be very long, with ambiguity in their true destination or origin, so all external trips were removed from the dataset before seeking matches. Traveler type also is not relevant, so it is not used here, in making matches (though one can imagine a future where some women may prefer to share rides with other women, and/or some people may prefer matches made within their age group, for example). MetroPlan Orlando also provided a detailed network, with nearly 24,000 nodes and around 61,000 links. Shortest-path travel times between each TAZ were used while disaggregating the trips, as discussed in the next paragraph. DATA DISAGGREGATION AirSage provided an anonymized, spatially and temporally aggregate dataset, with trips classified into hourly bins and their origins and destinations by traffic analysis zone (TAZ). Much smaller time steps and much more detailed locations (instead of zone centroids) are needed for a reliable DRS application matching travelers’ intra-regional trips. A time-step of one minute was used here, to facilitate computation while preserving dataset integrity; and origin and destination points (O’s and D’s) were randomly sampled from the list of nodes present within each trip’s origin and destination TAZs (which is on average, less than ¼-mile away). Figure 1 shows Orlando’s network and the nodes utilized for this disaggregation. In this way, the original 30-day 24-hour dataset was disaggregated into 30 different sets of one-minute trip-request files with intersection-based spatial detail. The departure times of these trips were not always in the hourly bin that AirSage indicated for each trip, because trips (within this region) can begin many minutes earlier (or can end many minutes later). This is because only the majority of the trip’s duration had to have occurred in the hour bin to which the trip was assigned by AirSage. As hourly aggregations were made by AirSage, 30-minutes of overflow was permitted into the previous and next hour bin. Once a start time was assigned, the shortest-path travel times for that time of day, as obtained via Caliper Corporation’s TransCAD software, a travel-demand modeling tool, were used to sample individual trip travel times from a normal distribution, whose mean equaled this shortest-path travel time and had a standard deviation of ±2 minutes. The results were smooth, minute-by-minute trip-request files for each of the 30 days, and natural looking departure and arrival time patterns throughout each of the 30 days.

Figure 1: The Orlando Network and Nodes Used for Spatial Disaggregation a) Orlando Network Separated by TAZ Gridlines b) Centroids Used in Aggregated Data c) Nodes Available for Spatial Disaggregation METHODOLOGY & RESULTS A fleet simulation was carried out to assess the optimal SAV fleet requirement for the Orlando metropolitan area to cater to all the trips with pre-specified service characteristics. A framework was developed in MATLAB, a commonly-used programming tool, to simulate a fleet of SAVs for a typical day. The trip request file generated from data disaggregation served as an input to the framework, along with the characteristics that are expected of the fleet. This included: fleet size, maximum allowable waiting time before an SAV is assigned to a passenger, maximum allowable time an SAV can take to reach the passenger, maximum additional time that is imposed on passengers who will be detoured for a new pickup and maximum additional time that a newly picked-up passenger has to wait while the previous occupants are dropped off. Table 1 states all these variables along with their abbreviations and this will be used for the remaining sections of the paper. In addition to this, Orlando’s network was converted into a MATLAB directional graph (digraph) and used to analyze shortest-path routes and times taken by SAVs. TABLE 1 List of Abbreviations Used in Reference to the Simulation Framework Abbreviation Description noOfSAVs Total number of SAVs used in the fleet maxExtraTripTime Minimum time imposed on travelers sharing their trips maxWaitingTime Maximum time that a passenger had to wait before an SAV reached them maxSearchTime Maximum time that a trip was stored on the waitlist before being rejected unserviced Total trips that could not be serviced under the above restrictions ETA Estimated time of arrival for an SAV to either pick up or drop off a passenger

The framework was composed of three distinct blocks: SAV allocation, SAV update and waitlist management. The SAV allocation block allocates the nearest SAV to a trip request based on the maxWaitingTime criterion. If no SAV was found satisfying this criterion, the trip request is stored in the waitlist. If an SAV with an existing occupant is located, the maxExtraTripTime criterion is checked prior to allocation, to minimize delays imposed on the travelers. After all the trips in a particular time step are either allocated to an SAV or stored in the waitlist, the SAV update block for the next time step is executed. In the SAV update block, the current location, destination and ETA of an SAV is monitored. If the SAV has not reached its destination for either a pickup or a drop-off operation, then its current location and ETA are updated. If the SAV has reached its destination for pickup, the drop-off operation is initiated. If a drop-off was executed, the second destination for drop-off of shared rides is processed, or the SAV stays idle, waiting for the next request. Once the update block has been executed, all previously waitlisted trip requests are checked for SAV allocation before moving on to the next time step of trip requests. If the trip requests have been on the waitlist for more than maxSearchTime, they are removed from the waitlist and unserviced is updated to reflect the same. The flowchart for the process described is shown in Figure 2. Fleet sizes varying from 5,000 - 30,000 SAVs, in intervals of 5,000 were used for these simulations.

FIGURE 2 The flowchart describing the main modules of the simulation framework. Results indicated that a fleet size of only 20,000 SAVs were sufficient to cater to more than 80% of the demand and 25,000 vehicles catered to all of the demand when passenger waiting times was as low as 5 minutes, on average, to hail a ride. For higher passenger waiting times, such as 10-15 minutes, a lower fleet size of 15,000 SAVs were sufficient to cater to more than 80% of the demand.

09:00-10:30 Session 9B: Time Use -- General Choices
Andrea Pellegrini (Università della Svizzera italiana, Switzerland)
Location: MCC Theater
Florian Schneider (Delft University of Technology, Netherlands)
Danique Ton (Delft University of Technology, Netherlands)
Lara-Britt Zomer (Delft University of Technology, Netherlands)
Winnie Daamen (Delft University of Technology, Netherlands)
Dorine Duives (Delft University of Technology, Netherlands)
Sascha Hoogendoorn-Lanser (Netherlands Institute for Transport Policy Analysis, Netherlands)
Serge Hoogendoorn (Delft University of Technology, Netherlands)
Trip chains: a comparison among latent mobility pattern classes

ABSTRACT. In this research, we study trip chains from a new perspective. We analyze differences in trip chain complexity regarding the number of trips and the number of transport modes between five latent mobility pattern classes. The first results suggest that differences between the classes exist for both complexity indicators. However, the identified differences are challenging the commonly assumed role of mode choice on trip chain complexity.

Thomas O. Hancock (University of Leeds, UK)
Stephane Hess (University of Leeds, UK)
Andrew J. Daly (University of Leeds, UK)
Charisma F. Choudhury (University of Leeds, UK)
James Fox (RAND Europe, UK)
Model averaging: revisiting our approach to decision rule heterogeneity in travel behaviour models and reinterpreting the results of latent class models

ABSTRACT. Please see attached file for the extended abstract.

Andrea Pellegrini (Università della Svizzera italiana, Switzerland)
Accounting for complementarity and substitution patterns in Multiple Discrete Continuous models with multiple constraints

ABSTRACT. Multiple discrete choices are rather common in consumer decisions. For instance, consumers may undertake several activities within a leisure day or purchase different products while shopping. Along with multiple discrete decisions, consumer typically decide how much of the chosen products to consume (e.g. the time allocated to various activities or the amount of money spent on different products). Bhat (2005, 2008) develops a novel framework called the multiple discrete continuous extreme value (MDCEV) model to analyze multiple discrete continuous (MDC) decisions using a Kuhn and Tucker (1951) first order conditions of optimal utility.  The MDCEV framework employs an additively separable utility function which can accommodate situations where consumers may select either only one product (goods are perfect substitutes) or multiple products (goods are imperfect substitutes) due to diminishing marginal utility. Nevertheless, the assumption of separable preferences implies that the MDCEV approach does not consider the case of complementary goods, which is quite likely in many consumption decisions. Bhat, Castro and Pinjari (2015) propose a non-additively separable (N-AS) utility function within the MDCEV method. Such a N-AS utility includes second order interactions of utility functions along with interaction parameters that allow the marginal utility of consumption of a good to depend on the quantities consumed of other goods. Positive and negative interaction parameters capture complementarity and substitution effects, respectively.  Consumers are assumed to maximize the N-AS utility function subject to a single constraint (i.e. budget constraint). However, in many consumption situations consumers may face multiple constraints, such as monetary constraints and time constraints. Ignoring multiple constraints can potentially entail misleading policy evaluation because of the underestimation of price and/or time sensitivity of consumers (Satomura, Kim and Allemby 2011). On the other hand, Satomura, Kim and Allemby (2011), Castro, Bhat, Pendyala and Jara-Diaz (2012) and more recently Astroza, Pinjari, Bhat and Jara-Diaz (2017) develop MDC modeling frameworks to accommodate multiple constraints (e.g. budget constraint and time availability). Nevertheless, these studies employ additively separable utility functions, and as consequence, do not account for complementarity and rich substitution effects.

The aim of the proposed research is to develop a MDC modeling framework that can simultaneously accommodate (a) non-additive preferences and (b) multiple constraints (specifically, both monetary budget constraint as well as time availability constraint) as determinants of individual time use decisions.

Francisco J. Bahamonde-Birke (DLR and TU-Berlin, Germany)
Establishing the variability of multinomial discrete models. Which proportion of the variability is actually being explained by our models?

ABSTRACT. EXTENDED ABSTRACT (please see attached file for correct format and equations) Econometric models are a key element in travel behavior research, allowing to explain travel decisions on the basis of a set of explanatory variables. Thus, they allow identifying the most important factors triggering a given behavior, as well as to derive trade-offs among different attributes, such as the subjective value of travel time reductions. The most common econometric models are the well-known linear (or linearized) regressions. They are, however, not suitable for the most common applications of econometric models in transportation, namely destination and mode choice models. This is because the outcomes of destination and modal choices are of discrete nature (e.g. an individual travel by a certain mode or another). Hence, linear continuous models such as linear regression do not allow characterizing the phenomenon. Econometric modeling in transportation research is, therefore, mainly based on discrete choice models. These models are based on random utility theory (Thurstone, 1927; McFadden, 1974), and assume the existence of underlying latent variables, called utility functions, associated with each alternative in the choice-set. These utility functions consist of a deterministic (associated with the explanatory variables of the model) and a stochastic component (represented through an error term). Then, it is assumed that a given alternative is chosen if and only if its underlying utility function is the largest utility functions among all alternatives in the choice-set (Train, 2009; Ortúzar and Willumsen, 2011). A disadvantage of this approach is that it does not allow establishing precisely the utility of a given alternative, but merely to establish that the utility ascribed to it is larger than the utility of the remaining ones (if this alternative is the chosen one). Thus, there is no accurate metric to compute the differences between the “actual” underlying utility and the predicted one. First and foremost, this affects the quality of the information available for estimation, increasing the data requirements, but it also hinders the modeler from establishing accurately the goodness-of-fit of the model, as the error cannot be directly measured. This way, it is not possible to construct an equivalent to the coefficient of determination r2 of linear model that indicates the proportion of the observed variability in the dependent variable being explained by the model. Several pseudo-equivalents for the coefficient of determination (called pseudo r2) have been proposed for discrete choice models. McFadden (1974) proposed a likelihood ratio index that basically compares the likelihood of reproducing the observations under the assumptions of the estimated model (LM) with the likelihood of reproducing the observations considering an equiprobable model (L0) through the following expression:

This index is, however, highly dependent on the proportion of the chosen alternatives in the sample. For that reason, Tardiff (1976) suggested contrasting LM with the likelihood of reproducing the observations considering a model that reproduces the proportion of chosen alternatives (LK):

Further indexes based on the likelihood ratio principle have been proposed (e.g. Ben-Akiva and Lerman, 1985), but all suffer from the same problem: while there is a clear interpretation for the extremes (0 and 1), the same cannot be said for intermediate values. Furthermore, it is neither possible to infer anything from them, nor to compare the goodness-of-fit of models estimated with different samples (Scott Long, 1997; Ortúzar and Willumsen, 2011). An alternative to likelihood ratio based indexes has been proposed by McKelvey and Zavoina (1975). Given that the underlying utility functions consist of deterministic (y*) and stochastic () components, it is possible to establish which proportion of the total variability can be ascribed to which of them. Thus, they suggest the following index: , which, basically, is indicative for the proportion of the variance explained by the model at the level of the utility function. Simulation studies have shown that this index most closely approximates the coefficient of determination r2 obtained from regressions on the underlying latent variable, outperforming all likelihood ratio based indexes (Hagle and Mitchell, 1992; Windmeijer, 1995; Scott Long, 1997). Similarly, it is also possible to extend this index in order to establish the proportion of the observed variability that can be ascribed to any variable in the model. The index, however, requires a single value for , while in the praxis, all utility functions may exhibit a different one (i.e. the will depend on the alternative being considered) . In the binominal case (or the ordinal case, which can be considered as an extension of the former), both utility functions can be reduced to a single expression accounting for utility differences. Hence, it is possible to calculate a single value for the index that is representative for the entire model. In the case of multinomial discrete model, however, it is not possible, creating major difficulties to establish the actual variability of the model. The present paper offers an in-depth discussion on different indexes for the adjustment of discrete choice models, as well as on different way to estimate the variability associated with multinomial models. It further proposes a method to estimate the total variability of the model accurately, and therefore, it extends the for multinomial discrete models. Furthermore, it includes a simulation exercise that shows that the proposed index outperforms likelihood ratio based indexes. While proposing a new index for the goodness-of-fit that outperforms existing ones is indeed one of the most relevant contributions of this paper, its implications are more far-reaching. Accurately assessing the actual variability associated with multinomial discrete models is crucial in various transportation problems. For instance, only establishing this variability accurately would allow assessing the relative importance of a given explanatory variable on the model at aggregated level (not only at the level of a single utility function); i.e. which part of the model’s total variability can be tracked down to a certain variable, such as price or travel time. Similarly, it also allows establishing, which proportion of the variability can be ascribed to the heterogeneity in the valuation of an attribute in a mixed (multinomial) logit model. Similarly, in a related application, Bahamonde Birke and Ortúzar (2014a; 2014b) established that the bias associated with the sequential estimation of hybrid choice models depends on the proportion of the total model’s variability due to the latent variables. Summarizing, assessing accurately which proportion of the total variability is actually being explained by our models and which proportion is being explained by stochastic elements will greatly increase the power of our conclusions. Keywords: multinomial discrete choice models, variability, goodness-of-fit

REFERENCES Bahamonde-Birke, F.J. and Ortúzar, J. de D. (2014a). On the variability of hybrid discrete choice models. Transportmetrica A: Transport Science 10(1), 74-88. Bahamonde-Birke, F.J. and Ortúzar, J. de D. (2014b). Is Sequential Estimation a Suitable Second Best for Estimation of Hybrid Choice Models? Transportation Research Record: Journal of the Transportation Research Board 2429, 51-58. Ben-Akiva, M.E., and Lerman, S.R. (1985). Discrete choice analysis: theory and application to travel demand. MIT press, Cambridge, Mass., USA. Hagle, T.M., and Mitchell, G. E. (1992). Goodness-of-fit measures for probit and logit. American Journal of Political Science 36(3) 762-784. McFadden, D. (1974). Conditional logit analysis of qualitative choice behaviour. In Zarembka, P. (ed.), Frontiers in Econometrics, 105-142. Academic Press, New York. McKelvey, R.D., and Zavoina, W. (1975). A statistical model for the analysis of ordinal level dependent variables. Journal of mathematical sociology 4(1), 103-120. Ortúzar, J. de D. and Willumsen, L.G. (2011). Modelling Transport. Fourth Edition, John Wiley and Sons, Chichester, UK. Scott Long, J. (1997). Regression models for categorical and limited dependent variables. In: Advanced quantitative techniques in the social sciences 7. SAGE Publications, London, UK. Tardiff, T.J. (1976). A note on goodness-of-fit statistics for probit and logit models. Transportation 5(4), 377-388. Thurstone, L.L. (1927). A law of comparative judgment. Psychological Review 34, 273 286. Train, K. E. (2009). Discrete Choice Methods with Simulation. Second Edition, Cambridge University Press, Cambridge, UK. Windmeijer, F. A. (1995). Goodness-of-fit measures in binary choice models. Econometric Reviews 14(1), 101-116.

09:00-10:30 Session 9C: Healthy, Happy, and Holistic Living -- Empirics
Lucas Harms (KiM Netherlands Institute for Transport Policy Analysis, Netherlands)
Location: Corwin East
Calvin Thigpen (Arizona State University, United States)
Maarten Kroesen (Delft University of Technology, Netherlands)
Susan Handy (University of California, Davis, United States)
An exploration of the reciprocal relationships between bicycling attitudes, skills, and behavior on a college campus

ABSTRACT. Attitude-behavior studies in the field of travel behavior have traditionally considered attitude as a determinant of behavior without analyzing the reciprocal relationship. This omission has foundational implications for both theory testing and refinement, as well as consequences for attempts to derive policy from research findings. In this study, we examine the reciprocal relationships between bicycling behavior, attitudes, and skill. We seek to compare the relative strength of the reciprocal relationships and provide policy suggestions based on the results. We assembled a longitudinal panel of undergraduate students from the University of California, Davis by connecting their responses across multiple years of an annual campus travel survey. Using a cross-lagged panel model, we find strong, statistically significant stability relationships for all three dependent variables and find that bicycle use has a moderate, positive influence on skill in the subsequent year. Our results also suggest that bicycling attitudes have a moderate, positive relationship with behavior in the following year, while there is limited evidence for the influence of bicycling behavior on attitudes. This research suggests that context and life stage may have moderating effects on the reciprocal relationships; among our sample of undergraduate students, bicycling behavior has limited to no influence on behavior attitudes, in contrast to previous studies’ findings among the general population. Furthermore, bicycling behavior in college helps develop bicycling skills, which may have enduring impacts and supports the notion that educational institutions can play an important role in preparing students for active travel later in life.

Marie-José Olde Kalter (Goudappel Coffeng, Netherlands)
Laura Groenendijk (Goudappel Coffeng, Netherlands)
Mark van Hagen (NS, Netherlands)
Lucas Harms (KiM Netherlands Institute for Transport Policy Analysis, Netherlands)
Marco Ter Brommmelstroet (University of Amsterdam, Netherlands)
Miranda Thush (ThuisraadRO, Netherlands)
Travel time perception of cyclists as policy instrument

ABSTRACT. Background More insight into the experienced waiting time of commuters on train stations has led to a different design of stations, whereby the main focus is to reduce the perceived waiting time by enhancing the station (Van Hagen, 2011). Travelers experience a pleasant time to be shorter and value the travel by train higher. For cyclists, as well, the travel experience could be positively affected by influencing the perceived travel time. This might be more efficient than a focus on objective speed (Van Hagen and Thüsh, 2016).

A presumption is that the experienced travel time is important for cyclists. Particularly when choosing a route, but also in the decision to go cycling or not. The results from a small field survey on the perceived travel time and route choice behavior of cyclists in the Netherlands show that cyclists more often chose a longer route when they perceived that route to be more pleasant (Goudappel Coffeng, 2012). Furthermore, the cyclists often experienced this route to be shorter. Although the research sample is not representative and the wait at traffic lights is not taken into account, the research led to the following question: "Is it possibly more important to design an attractive rather than a fast bicycle route if we want to influence travel behavior?"

If attractive routes are actually perceived as being shorter, then cyclists' perceived travel time could be reduced considerably by consciously leading cycling routes through a pleasant environment. This may not only affect route choice, but also promote cycling as a mode of travel. A focus on increasing the attractiveness of cycling routes could thus be beneficial to cyclists’ perceived travel time and, consequently, their use. Additionally, taking cyclists’ perceived travel time into account leads to an improvement of existing transport models and CBAs.

Based on existing knowledge and insights we are unable to answer the posed question. Suspicions regarding cyclists’ experienced travel time do exist, but they do not suffice for drawing hard conclusions. In order to answer the question, Goudappel Coffeng, NS, ThuisraadRO and the University of Amsterdam initiated a study early this year into cyclists’ perceived travel time. Twelve parties participate in the research: the municipalities of Amsterdam, Rotterdam, Utrecht, The Hague, Maastricht and Nijmegen, the Stedendriehoek region, the metropolitan area of Rotterdam The Hague, the provinces of Gelderland, South Holland and Groningen and the Netherlands Institute for Transport Policy Analysis. Apart from a financial contribution, all parties also supply cases to ensure that the research closely adheres to practice.

The main objective of this research is to gain a better understanding of cyclists’ travel time perception. The following research questions are addressed: - Which characteristics are associated with the attractiveness of cycle routes? - To what extent are these characteristics related to travel time perception? - What is the influence of these characteristics on route choice?

Do cyclists always choose the shortest or fastest route? Despite the fact that cycling is already a very prevalent mode of travel in the Netherlands, it is important to continue to stimulate cycling. This requires insight into the factors that affect the use of bicycles. There is a rich research history into the determinants of cycling, both nationally and internationally. Several studies in the Netherlands show that the travel time ratio between the car and bicycle as well as the speed are important explanatory factors for choosing to travel by bicycle, see for example Rietveld and Daniel (2004) and Research voor Beleid (2006). An evaluation of the Delft cycle route network by Gommers and Bovy (1987) suggests that the route selection of cyclists in an urban network is strongly affected by the travel time. Foreign research shows a similar high preference for short and fast routes (Menghini et al., 2010; Hood et al., 2011; Broach et al., 2012). We therefore see that many cycling measures focus on these two aspects: cycling highways for higher speed and more favorable travel time compared to the car.

However, more recent studies on the route choice of cyclists show that travelers do not always choose the shortest or fastest route. For example, the results of a study by RoyalHaskoningDHV and Eindhoven University of Technology (2016) show that the presence of bicycle infrastructure, road surface quality and slopes have a greater influence on route choice than a four-minute travel time reduction. Moreover, the influence of these aspects increases with increasing distance. In his research into the value of time of cyclists on bicycle highways, Van Ginkel (2014) found that both commuters and recreational cyclists prefer a more comfortable cycling route with a lower value of time. Whether that makes people choose a more comfortable route was, however, not part of the research. Recent research in Copenhagen (Vedel et al., 2017) has shown that travelers are willing to cycle longer distances if the route passes through a green environment. It also showed that delays and traffic density have a negative influence on the preference for a particular route. Cyclists are prepared to take a longer route if they are able to bypass large traffic jams or bypass roads. The extent to which the perception of travel time plays a role here has not been investigated. Other studies indicate that the attractiveness of a route is to some extent influenced by the amount of motorized traffic, route aesthetics and route vivacity (Wahlgren and Schantz, 2012; Gehl, 2010). It is still unknown whether people choose more attractive routes.

As mentioned earlier, it is suspected that the perception of travel time plays an important role in the route choice of cyclists. From the theory of time perception, we know that people are hardly able to estimate the length of a period of time, but usually can indicate whether something takes a long or short time. It is this subjective perception of time that affects travel behavior: the choice for a particular mode of transport, route or time. When people move along a boring, monotonous route, the brain gets only a few stimuli, the person gets bored and the trip is perceived to be longer. This is entirely different from when people travel along an attractive, varying route. The brain receives enough positive stimuli, making the route appear shorter than it actually is, even though the journey takes as much time as the boring and busy route. An important question here is: what determines a person’s perception of a route’s attractiveness and what is the relationship with the subjective time perception?

Data and methodology Based on the collected literature, interviews with experts and focus groups, we have created a conceptual model (Figure 1). The model shows the various factors that affect the attractiveness and the perception of travel time of a bicycle route. The characteristics of the physical environment comes first, where we distinguish between characteristics of the bicycle infrastructure and the characteristics of the environment. How the physical environment is experienced also depends on individual characteristics: sociodemographic aspects such as age, sex and education, attitudes and habits and psychological aspects such as the mood in which someone has gotten on the bicycle. External factors, such as weather conditions, may also affect the experience. Whether, and to what extent, the perception affects the travel behavior (route choice) and vice versa is an important outcome of this study.

All participating parties have provided two cycling routes in the Netherlands. The routes differ from each other in terms of physical and environmental characteristics. Some routes run through rural areas with a lot of vegetation and low traffic, other routes run through urban areas with high urbanization and a lot of interaction with road users. There are also routes that are characterized by a lot of variety in the area. All routes have been filmed by us under similar circumstances. All infrastructural and environmental features are collected of the routes. Based on these features, the different parts of the cycle routes will be clustered at a later stage.

Under a representative sample of the Dutch population (according to the Golden Standard), two questionnaires are distributed in November 2017. In the first questionnaire, we ask 1,500 respondents to score two bicycle routes on attractiveness and also indicate which route is preferred. In the second questionnaire, we ask 3,000 other respondents about the perceived travel time of a cycle route. This way, each cycle route is scored for attractiveness as well as perceived travel time. In the final phase of the research we will bring the research results together. Based on the results of the questionnaires, we know how attractive each route is (and which factors affect this the most), and how the travel time is perceived (and which factors affect this the most). By combining these results we will determine the effect of attractiveness of a route on route selection and perception of travel time. The results are expected in early 2018.

In summary, with the research results: - We provide insight into current trends and developments in cycling travel experience - We use case studies to illustrate the perception of traveling time - We present nationally relevant outcomes on the influence of different factors on the attractiveness and travel time of a cycle route - We determine the effects of this on the route selection - We provide concrete handles to use the insights for bicycle policy

Discussion Stimulating bicycle use is a responsibility of the government. More and more provinces and municipalities in the Netherlands recognize the importance of a good bicycle infrastructure as an alternative to the car, and the quality of cycling facilities is high on the agenda. The ministry of Infrastructure and Environment is also focusing on the promotion of cycling. The program Beter Benutten has since 2011 been working together with the ministry, regional authorities and industry in twelve regions to reduce the congestion on the Dutch roads. Many measures from the Beter Benutten program are bicycle stimulating projects, often behavioral interventions or communication campaigns actively promoting the use of the bicycle and/or e-bike.

Presently, the perception of travel time does not play an explicit role in Dutch cycling policy. With this research we want to make a change. A focus on increasing attractiveness rather than shortening of bicycle routes seems a smart strategy. We see two major practical applications of more knowledge about cyclists' travel time perception: 1. In transport models and CBAs. Cycling through attractive routes should also be able to find a representation in the model system. The strategy of increasing route attractiveness can lead to changes in model outcomes by explicitly including the travel time perception of cyclists in transport models and CBAs. 2. When preparing bicycle policy and carrying out bicycle measures. The focus of policies and measures should not only be on the fastest and shortest routes. Taking the attractiveness of the environment and the perception of travel time into account can thus be included as an additional control mechanism in the cycling policy. Consider, for example, the spatial design and the design of the cycling infrastructure.

References Broach, J., Dill, J. & Gliebe, J. (2012). Where do cyclists ride? A route choice model developed with revealed preferences GPS data. Transportation Research A, 46, 1730-1740.

Gehl, J. (2010). Cities for people. Island Press, Londen. Ginkel, J. van (2014). The value of time and comfort in bicycle appraisal. A stated preference research into cyclists’ valuation of travel time reductions and comfort improvements in the Netherlands. Master Thesis, Universiteit Twente, Enschede.

Gommers, M.J.P.F. & Bovy, P.H.L. (1987). Evaluatie fietsroutenetwerk Delft, Routekeuzegedrag en netwerkgebruik. Eindrapport, Delft.

Hagen, M. van (2011) Waiting experience at train stations. Dissertation, Eburon, Delft.

Hood, J, Sall, E. & Charton, B. (2011). A GPS based Bicycle Route Choice Model for San Francisco. California. Transportation Letters: the International Journal of Transportation Research, 3(1), 63-75.

Menghini, G., Carrasco, N., Schüssler, N & Axhausen, K. (2010). Route choice of cyclists in Zurich. Transportation Research A, 44, 754-765.

Research voor Beleid (2006). Verklaringsmodel voor fietsgebruik gemeenten. Eindrapport. In opdracht van het Fietsberaad, Research voor Beleid, Leiden

Rietveld, P. & Daniel, V. (2004). Determinants of bicycle use: do municipal policies matter? Transportation Research Part A: Policy and Practice, 8(38), 531-550

RoyalHaskoningDHV & TU Eindhoven (2016). Het optimaliseren van fietsgedrag in verkeersmodellen.

Vedel, S.E., Jacobsen, J.B. & Skov-Petersen, H. (2017). Bicyclists’ preferences for route characteristics and crowding in Copenhagen – A choice experiment study of commuters. Transportation Research Part A: Policy and Practice, 100, 53-64.

Van Hagen, M. & M. Thüsh. (2016). Wat als de overheid zich in plaats van op snelheid zou richten op de beleving van de reistijd? In: Toekomstbeelden van het fietsgebruik in 5 essays, pp. 42-49. Kennisinstituut voor Mobiliteit. KIM-16-A01. ISBN/EAN 978-90-8902-142-7.

Wahlgren, L. & Schantz, P. (2012). Exploring bikeability in a metropolitan setting: stimulating and hindering factors in commuting route environments. BMC Public Health, 12(1), 168

Alexander Reichert (TU Dortmund, Germany)
Christian Holz-Rau (TU Dortmund, Germany)
Who is reponsible for more GHG-emissions in transport: Mono- or Multimodal User groups?- A comparison

ABSTRACT. (see upload)

Lucas Harms (KiM Netherlands Institute for Transport Policy Analysis, Netherlands)
Maarten Kroesen (Delft University of Technology, Netherlands)
User characteristics and trip patterns of e-bike use in the Netherlands
SPEAKER: Lucas Harms

ABSTRACT. The sale of electrically assisted bicycles (e-bikes or pedelecs) is growing at a rapid rate across Europe. Within Europe, the Netherlands is one of the biggest markets for e-bike sales with 10.4 sales per 10,000 people, roughly equating to 17 per cent of all Dutch bicycle sales. Around 1 million e-bikes are now in ownership out of a total stock of 22 million bicycles for 17 million Dutch inhabitants. Whereas market data is available describing sales trends of e-bikes, there is limited understanding of the user characteristics and trip patterns.

In our paper we will provide an overview of user characteristics and trip patterns of e-bike use in the Netherlands, based on cross sectional data from the Dutch National Travel Survey (NTS) and longitudinal data from the Netherlands Mobility Panel (MPN). The Dutch NTS and MPN are the first surveys of its kind to include the e-bikes as a separate modality, including user characteristics such as age and gender and trip characteristics such as distances covered and travel speed. The results indicate that e-bikes now account for around 15 per cent of total distance travelled by bicycle in the Netherlands - roughly equivalent to 2 billion kilometres per year. Average journey distance covered by e-bikes is 5 kilometres – a third further than conventional cycling (3.5 kilometres). E-biking is particularly significant among the older population and accounts for over 40 percent of all cycling kilometres travelled by adults age 65 and above and almost 25 percent for adults aged 50 up to 65 years. Older adults report using their e-bikes for leisure and shopping whilst for younger adults commuting plays a more significant role. In recent years a shift seems visible to younger users and an increase in the share of work-related and shop-related trips.

In our paper we will discuss characteristics of e-bike use in the Netherlands, trip patterns and trip purposes as well as some recent trends, based on cross sectional NTS data and longitudinal MPN data since 2013. We will use descriptive statistics to give an overview of recent trends in the Netherlands and to increase understanding in user characteristics and trip patterns. We will also show to what degree e-bikes are substituting regular bike or car trips, and whether e-bikes are facilitating longer trips that people wouldn't have done otherwise. Whether these trends and patterns are indicative of the developments in other western counties will be discussed as well. The contribution builds upon existing scientific knowledge about spatial and social patterns and trends in (e-)cycling within a mature cycling context (as published by Harms et al. 2014 and Jones et al 2016) and recent research focusing on the substitution of other modes by e-bikes (Kroesen 2017).

References Harms, Lucas, Luca Bertolini, and Marco Te Brömmelstroet. "Spatial and social variations in cycling patterns in a mature cycling country exploring differences and trends." Journal of Transport & Health 1.4 (2014): 232-242.

Jones, Tim, Lucas Harms, and Eva Heinen. "Motives, perceptions and experiences of electric bicycle owners and implications for health, wellbeing and mobility." Journal of Transport Geography 53 (2016): 41-49.

Kroesen, Maarten. "To what extent do e-bikes substitute travel by other modes? Evidence from the Netherlands." Transportation Research Part D 53 (2017) 377–387

09:00-10:30 Session 9D: Social Influence
Michael Maness (University of Maryland, United States)
Location: UCEN SB Harbor
Michael Maness (University of South Florida, United States)
New Models of Random Taste Variation and Constraint Change Induced by Social Influence

ABSTRACT. Over the past two decades, travel behavior analysis has begun shifting focus from the individual to the social. Accordingly, discrete choice models, a common modeling technique in travel behavior analysis, have begun to integrate social context into its framework (Dugundji and Walker 2005, Paez and Scott 2007). Social influence, defined as the tendency to engage in behavior similar to others, has been one approach to this integration. Thus far these efforts have resulted in work showing that social influences may be relevant in travel decision making. Current models of social influence in travel behavior assume a direct-benefit effect is generated from conforming to the behavior of others (i.e. utility itself is directly increased by conforming). There have been attempts to add social influence in new way including in the choice model of latent class model (El Zarwi et al. 2017) and through the structural and measurement equations of hybrid choice model (Kamargianni and Polydoropoulou 2014). But these techniques still treat social influence as deterministic taste variation. Only limited work has explored social influence through random taste variation. Specifically, Maness and Cirillo (2016) is the only example, where they develop a latent class model where social influence is used as a covariate in the class membership model of a binary choice model. They found that household’s taste for bicycle ownership varied with the prevalence of cycling in their metro areas with more favorable tastes developing as ownership increased. They conclude their work by mentioning that:

“The latent class framework allows for a great deal of flexibility in modeling social influence and social interactions. The model described in this paper refers to an informational conformity hypothesis and this is done for clarity of exposition and to focus the analysis of model properties. But this modeling framework does not only exclusively describe an informational conformity hypothesis. The model is just a subset of models of social effects causing random taste variation – an area with limited exploration in the literature. By changing assumptions on class number, social influence effects and types, and behavioral variations between individuals, other behavioral theories can be described under this same latent class framework.” (p. 97)

Additionally, Maness and Cirillo (2016) explain that a latent class structure can be used to not only explore taste variation but also expectation changes and constraint changes. This study aims to explore two additional model specifications that center on random taste and constraint variations induced by social effects. The first model extends the indirect informational conformity framework in Maness and Cirillo (2016) to multiple classes through an ordered class membership model. The second model captures constraint changes due to social influence through a two-level class membership model.

People Change Classes which Induces Taste Variation Increasing the number of classes is one possible direction for extending the random social taste variation model introduced by Maness and Cirillo (2016). This formulation will be written under a binary choice decision. Begin by assuming a population N of decision makers where individuals are connected in a social network G. Each individual n is faced with a choice task where the individual must choose between two alternatives y_n={0,1}. The term adoption will be used to refer to choosing y_n=1. In this population, individuals may fit into multiple taste classes (class 1,2,3 in this case) where these classes are assumed to be ordered such that a higher class has a greater propensity toward adoption than the classes below it. This process is unobserved and will be modeled latently with discrete classes. Information class membership is affected by: the individuals’ inherent knowledge level (individual information αz_n) as well as knowledge of the behavior being transferred to individuals through seeing the behavior in the population (social information δy ̅_n). This will be denoted by the information function F_n, which will take the following linear-in-parameter form F_n=αz_n+δy ̅_n+e_n^F c_n={■(1&if F_n≤φ_(1|2)@2&if φ_(1|2)

The thresholds φ_(1|2) and φ_(2|3) serve to delineate which classes the individuals will decided when obtaining their social and individual information. For example, if we assume that the error term e_n^F is normally distributed, then the class membership probabilities follow similarly to a binary probit formulation: π_n^[1] =Φ(φ_(1|2)-αz_n-δy ̅_n ) π_n^[2] =Φ(φ_(2|3)-αz_n-δy ̅_n )-Φ(φ_(1|2)-αz_n-δy ̅_n ) π_n^[3] =1-Φ(φ_(2|3)-αz_n-δy ̅_n )

Due to being exposed to the behavior more often, individuals surrounded by others who adopt may be able to reevaluate their preferences in move to higher information classes under the new information they receive. Thus, the preferences of these individuals may vary compared to individuals in lower information classes. Assuming utility maximizing behavior for individuals, the utility between behaviors y_n={0,1} for an individual n for each class {1,2,3} is given as follows: U_n^[1] =β^[1] x_n+ε_n^[1] U_n^[2] =β^[2] x_n+ε_n^[2] U_n^[3] =β^[3] x_n+ε_n^[3]

The use of an ordered class membership model helps to create parsimony by reducing the number of parameters compared to an unordered class membership model. Additionally, it makes the behavioral explanation of the social information parameter δ easier to interpret as it is difficult to normalize that parameter in the unordered case to explain a progression of classes with greater levels of social information.

Choice Sets of Classes Change Dynamically, but People Remain in Class New information may increase the knowledge of available options for an individual. In discrete choice models, constraints are handled by the choice set. Thus to deal with this in a latent class formulation, the corresponding classes could have different choice sets. The diffusion of innovations literature provides motivation for this approach. Adopters of a new innovation are divided into categories with varying degrees of risk aversion and social status. Rogers (2010) suggests a classification system where adopters are labeled: innovators, early adopters, early majority, late majority, and laggards. For simplification of exposition, we will consider a simplified model with innovators, majority, and laggards. This can be modeled through a latent class structure with 3 classes. But an important difference in this case is that individuals do not change classes as adoption increases. The behavioral theory assumes that people remain in their classes throughout the adoption of the technology. Thus, we intend to introduce an additional latent class representing the likelihood that the new option will be considered. This likelihood which depends on the susceptibility of consideration S depends on the level of adoption / social influence δy ̅_n: S=δy ̅_n+e^S

An individual’s membership to an adopter type then depends on characteristics of the individual z_n. The adoption type membership propensity T_n is denoted as follows: T_n=γz_n+e_n^T

The choice model is completed by assuming either deterministic choice to not adopt if the new alternative is not in one’s choice set (due to their class membership), U_n=-∞. Otherwise, the individual will consider the option as a function of their individual characteristics and attributes of the new alternative x_n. Thus the choice utilities can be denoted as follows: U_n={■(-∞&if c_n is not activated or not considering@βx_n+e_n&if c_n is activated or considering)┤

Probability of adoption the new alternative in this two-level class membership model can be summarized as follows: P_n=∑_classes▒[█(prob a class is considering c_n∙@prob of being in aclass considering c_n ∙@prob βx_n>0)]

A framework like this can create value in using discrete choice models to model the adoption of new technologies. This model form can allow for dynamic changes in the probability of adoption over time as the adoption rate changes. A major benefit is the possibility of learning more about the threshold of adoption that separate innovators/early adopters from the majority of the population. This framework takes a probabilistic take on the value of these thresholds whereas existing work typically must assume some exact cutoff point between classes. This may aid behavioral realism as the cutoffs between classes are a simplification of reality so using “fuzzy” thresholds may lead to improved model fit and predictions.

Expected Results Model estimation is still in progress using simulated and observed data. The simulated data will be used to explore the dynamical properties of the models. In particular, this work will explore equilibrium properties of these models (include location, stability, and quantity of equilibria) as well as the speed at which equilibrium is achieved. Practical application will take place using data on bikesharing adoption over time at the University of South Florida campus and in the cities of Tampa and St. Petersburg. The case study on the city level system will be of particular interest, since that data will include both increases in ridership as well as changes in the service area by station and bike additions. Expected results will show the distribution of social influence across the population and the strength of these influences. Additionally, the models results will be applied for forecasting the adoption rate for bikeshare services in these case study locations. The development of the model formulations and estimation procedures will aid in developing this new area of social effects induced taste and constraint variation. This work will help to determine if this can account for additional unobserved heterogeneity as compared to traditional approaches to including social effects in discrete choice models. Additionally, this effort serves as a stepping stone to explore the possible modeling space introduced through random taste and constraint variation from social effects. Follow-up work can be used to more fully explore the similarities and differences between random taste variation social influence models and deterministic taste variation models. References Dugundji, E. R., & Walker, J. L. (2005). Discrete choice with social and spatial network interdependencies: an empirical example using mixed generalized extreme value models with field and panel effects. Transportation Research Record: Journal of the Transportation Research Board, 1921, 70-78. El Zarwi, F., Vij, A., & Walker, J. L. (2017). A discrete choice framework for modeling and forecasting the adoption and diffusion of new transportation services. Transportation Research Part C: Emerging Technologies, 79, 207-223. Kamargianni, M., Ben-Akiva, M., & Polydoropoulou, A. (2014). Incorporating social interaction into hybrid choice models. Transportation, 41(6), 1263-1285. Maness, M., & Cirillo, C. (2016). An indirect latent informational conformity social influence choice model: Formulation and case study. Transportation Research Part B: Methodological, 93, 75-101. Páez, A., & Scott, D. M. (2007). Social influence on travel behaviour: a simulation example of the decision to telecommute. Environment and Planning A, 39(3), 647-665. Rogers, E. M. (2010). Diffusion of innovations. Simon and Schuster.

Sören Groth (ILS - Institut für Landes- und Stadtentwicklungsforschung gGmbH, Germany)
End of “Feeling the Car”: Altered Emotions Towards More Multimodality?

ABSTRACT. Problem / Purpose In recent years, a transition from an automobile to a multimodal society has been postulated within western transport and mobility research (e.g. Spickermann et al., 2014). With regard to individual mobility patterns, this discussion contains the idea of a shift from exclusive car use for all trips to the flexible use of several modes (Buehler and Hamre, 2015; Kuhnimhof et al., 2012a). In particular, young adults have been titled a "New Generation" (Kuhnimhof et al., 2011) to express the historically unique departure from the exclusive use of the car to more multimodal behaviour. A number of studies have found that car use among young adults has decreased since the turn of the millennium (Delbosc and Currie, 2013; Kuhnimhof et al., 2012b; Kuhnimhof et al., 2012a). On the one hand, the combination of public transportation and new mobility services (free-floating carsharing, modern bike sharing schemes etc.) are increasingly seen as a competitive alternative to private car ownership (Cohen and Kietzmann, 2014; Geels, 2012). But, also an increasing share of young car owners use much more often alternative modes of transport (Kuhnimhof et al., 2012b; Kuhnimhof et al., 2012a). To date, the altered mobility patterns within the “new generation” of young adults have mainly been explained by generational changes in life circumstances such as: lower incomes, more students, a later career start, a later family formation etc. (ibid.). Also greater environmental awareness is assumed (ibid.). However, little attention has been paid to possible changes in the emotional relationship between human and automobile (Bratzel, 2014; Delbosc and Currie, 2014). This omission is problematic, as narratives such as “Feeling the Car” by Mimi Sheller (2004), “Car use: lust and must” by Linda Steg (2005), „For Love of the Automobile” by Wolfgang Sachs (1992) etc. have convincingly explained the relevance of emotions for the broad use of private cars over recent decades. With regard to the discussion of a transition from an automobile to multimodal society, the aim of this contribution is to provide empirical evidence for the “end of the love story between human beings and the automobile”.

Theory / Methodology For the purpose of this research, theoretical assumptions of "symbolic-emotional dimensions of mobility" according to Marcel Hunecke (2000) are used. For a long time, empirical evidence on the relevance of emotions in mode choice existed solely on the basis of qualitative studies (Klühspies, 1999; Praschl et al., 1994). Towards the end of the 1990´s, Marcel Hunecke (2000) was able to translate symbolic-emotional factors into a theoretically founded quantitative model. Here, he differentiates four basic symbolic-emotional dimensions of mobility: autonomy, excitement, status, and privacy. Hunecke with his team express: “All symbolic-[emotional] evaluations of transport modes can finally be reduced to these four dimensions, which are characterized by a functional core of physical or socioeconomic aspects, but depend strongly on processes of social interpretation.” (Hunecke et al., 2008). For example, the evaluation of autonomy concerning to different transport modes is connected to feelings such as freedom, self-determination, flexibility, individuality etc. (ibid.). For the purpose of this research, data were collected by using a household survey of residents in the German city Offenbach (Main) in spring 2013 (Schäfer, 2016). The city is located in the Rhine-Main metropolitan region, southeast of Frankfurt (Main) and has around 115,000 inhabitants. The questionnaire was distributed to 3,212 inhabitants aged 18 and older and selected by a random route sampling method. Ultimately, 620 questionnaires were completed and returned, yielding a response rate of 19.3 percent. Comparing the sample with the population census 2011, it is not representative to the Offenbach population (the sample is older, has a better education or less likely to have a migration background than the population average). Conceptual assumptions on mobility styles were considered using Hunecke´s theory as its core together with some more attitudinal factors identified from the literature. Considering emotions and attitudes on different transport modes (car, public transportation, bicycle etc.), a total of 44 items was placed in the questionnaire. The items were queried on a five-point-Likert scale. Factor and cluster analysis are established methods used to identify homogeneous attitudinal groups. (i) Prefixed factor analyses reveal how the dataset is structured by identifying underlying factors. (ii) The subsequent cluster analysis groups the respondents into homogeneous emotional clusters using the (uncorrelated) factors.

Findings Based on eleven extracted factors, hierarchical and non-hierarchical cluster analyses helped to identify the homogeneous emotional clusters, using the Ward method and subsequently k-means procedure. Squared Euclidian distance was applied as a measure of distance. With the aid of the elbow-criterion, a plausible five-cluster solution was chosen. Here, two emotion-based clusters can be highlighted that reflect the possible transition from an automobile to a multimodal society outlined in the beginning: (i) Car-loving Monomodals (19.1 percent) exhibit a strong and consistent emotionally-loaded car orientation. Also, they reject any ecologically motivated criticism of the car, and refuse any sympathies to the use of other transport modes. Their “love for the automobile” is expressed in an almost exclusive car use. However, Car-loving Monomodals seem to represent the socio-demographic profile of a departing car based society: They are dominated by older men with higher income but lower education. (ii) In contrast Car-rejecting Multimodals (21.6 percent) seems to be diametrically opposed to the Car-loving Monomodals in all emotional matters. They show a strong emotional relation to all transport modes, except to the car. Also, preferences to the flexible and situational mode choice are expressed of the people in this cluster. In fact, Car-rejecting Multimodals realize a highly multimodal behaviour. Unlike the first cluster, they have stereotypical socio-demographics of the above mentioned “New (multimodal) Generation”: Car-rejecting Multimodals are young, show no gender differences and are highly educated.

Research limitations / implications On the basis of quantitative data, this contribution provides empirical evidence that emotions of young adults might have changed in relation to modes of transportation. However, a central challenge for transport and mobility research will be the validation of an “ending love story between human and automobile” by representative data. For example, the national questionnaire on German mobility behavior (“MiD – Mobilität in Deutschland”) is currently being revised. It might be an innovative approach to integrate such emotional / attitudinal constructs into the national datasets. A further important discussion must be conducted with regard to the “stability” of the data. For example, changes in individual life situations (family formation, career start etc.) can lead to emotional changes in the human-machine relationship. Analysis of a more recent survey of the sample could help here, subject to project funding.

References Bratzel, S., 2014. Die junge Generation und das Automobil – Neue Kundenanforderungen an das Auto der Zukunft? In: Ebel, B., Hofer, M. B. (Eds.), Automotive Management. Springer Berlin Heidelberg: Berlin, Heidelberg, pp. 93–108. Buehler, R., Hamre, A., 2015. The multimodal majority?: Driving, walking, cycling, and public transportation use among American adults. Transportation 42 (6), 1081–1101. 10.1007/s11116-014-9556-z. Cohen, B., Kietzmann, J., 2014. Ride On!: Mobility Business Models for the Sharing Economy. Organization & Environment 27 (3), 279–296. Delbosc, A., Currie, G., 2013. Causes of Youth Licensing Decline: A Synthesis of Evidence. Transport Reviews 33 (3), 271–290. 10.1080/01441647.2013.801929. Delbosc, A., Currie, G., 2014. Using discussion forums to explore attitudes toward cars and licensing among young Australians. Transport Policy 31, 27–34. Geels, F., 2012. A socio-technical analysis of low-carbon transitions: Introducing the multi-level perspective into transport studies. Journal of Transport Geography 24, 471–482. Hunecke, M., 2000. Ökologische Verantwortung, Lebensstile und Umweltverhalten. Asanger: Heidelberg. Hunecke, M., Haustein, S., Böhler, S., Grischkat, S., 2008. Attitude-Based Target Groups to Reduce the Ecological Impact of Daily Mobility Behavior. Environment and Behavior 42 (1), 3–43. Klühspies, J., 1999. Stadt - Mobilität - Psyche: Mit gefühlsbetonten Verkehrskonzepten die Zukunft urbaner Mobilität gestalten? Birkhäuser Verlag: Basel. Kuhnimhof, T., Buehler, R., Dargay, J., 2011. A New Generation: Travel Trends for Young Germans and Britons. Transportation Research Record: Journal of the Transportation Research Board 2230, 58–67. 10.3141/2230-07. Kuhnimhof, T., Buehler, R., Wirtz, M., Kalinowska, D., 2012a. Travel trends among young adults in Germany: Increasing multimodality and declining car use for men. Journal of Transport Geography 24, 443–450. 10.1016/j.jtrangeo.2012.04.018. Kuhnimhof, T., Wirtz, M., Manz, W., 2012b. Decomposing Young Germans’ Altered Car Use Patterns. Transportation Research Record: Journal of the Transportation Research Board 2320, 64–71. 10.3141/2320-08. Praschl, M., Scholl-Kuhn, C., Risser, R., 1994. Gute Vorsätze und Realität: Die Diskrepanz zwischen Wissen und Handeln am Beispiel Verkehrsmittelwahl. Bundesministerium für Umwelt Jugend und Familie: Wien. Sachs, W., 1992. For love of the automobile: Looking back into the history of our desires. University of California Press: Berkeley. Schäfer, P. e. a.K., 2016. Elektromobilität als Motor für Verhaltensänderung und neue Mobilität. Abschlussbericht des Gesamtvorhabens „Sozialwissenschaftliche und ökologische Begleitforschung in der Modellregion Elektromobilität Rhein-Main“: Arbeitspapiere zur Mobilitätsforschung Nr. 8: Frankfurt am Main. Sheller, M., 2004. Automotive Emotions: Feeling the Car. Theory, Culture & Society 21 (4-5), 221–242. 10.1177/0263276404046068. Spickermann, A., Grienitz, V., Gracht, H. A. von der, 2014. Heading towards a multimodal city of the future? Technological Forecasting and Social Change 89, 201–221. 10.1016/j.techfore.2013.08.036. Steg, L., 2005. Car use: Lust and must. Instrumental, symbolic and affective motives for car use. Transportation Research Part A: Policy and Practice 39 (2-3), 147–162.

Francesco Manca (Imperial College London, UK)
Aruna Sivakumar (Imperial College London, UK)
John W Polak (Imperial College London, UK)
Jonn Axsen (Resource & Environmental Management, Simon Fraser University, Canada)
An empirical exploration of endogeneity in hybrid choice models with social influence measures

ABSTRACT. During the last decade, experts in transport field have started to investigate how to include social influence in discrete choice models. Social influence is a process involving different types of interactions and also unconscious mechanisms of conformity. It can affect groups of people and sometimes entire societies. Previous transport studies developed measures of social conformity by considering the peers’ choice as explanatory variable or by taking into account the social environment of the individual as latent variable. In our recent work, we considered the interaction between specific attitudes of the individual’s social network and tie strength and built a new conformity measure (peers’ attitude variable) to investigate the electric vehicle purchase preferences of individuals. However, variables capturing behaviours in a social group can potentially be endogenous. In that case, the correlation between variable and error likely leads to biased estimated parameters. By following Walker et al. (2011), we adapt the Berry, Levinsohn, and Pakes (BLP) procedure to the specific case of the hybrid choice model. Indeed, the peers’ attitude variable resulted significant when included in the structural model component. Results show that the inclusion of market-specific constants in the hybrid choice model generates very marginal changes in parameters. The instrumental variable in the first regression is significant and have the correct sign. As expected, the parameter of the social influence variable was biased upward but potential endogeneity seems not to affect the other parameters. This suggests that interdependency in attitudes does exist but needs to be handled carefully as it can also be affected by unobserved factors of the individual’s social group.

09:00-10:30 Session 9E: Choice Experiment Design Part 1
John Rose (University of Technology, Sydney, Australia)
Location: MCC Lounge
Fangqing Song (Institute for Transport Studies & Choice Modelling Centre, University of Leeds, UK)
Stephane Hess (University of Leeds, UK)
Thijs Dekker (University of Leeds, UK)
Comparing and combining best-worst scaling and stated choice data to understand attribute importance in mode choice behaviour
SPEAKER: Fangqing Song

ABSTRACT. (Please see the uploaded file)

Fangqing Song (University of Leeds, UK)
Stephane Hess (University of Leeds, UK)
Thijs Dekker (University of Leeds, UK)
Accounting for intra-respondent preference heterogeneity in repeated stated choice data: when new travel modes come into play
SPEAKER: Fangqing Song

ABSTRACT. (Please see the uploaded file)

John Rose (University of Technology, Sydney, Australia)
Antonio Borriello (Università della Svizzera italiana, Switzerland)
On the separation of perceptual translations of attributes and generalised attitudes in discrete choice
SPEAKER: John Rose

ABSTRACT. Over the past four decades, discrete choice models (DCM) have become the dominant method for modelling and understanding traveller preferences and behaviour. Estimation of DCMs requires analysts to specify separate utility functions for the different alternatives present within a given market of interest. Typically, these utility functions are defined in terms of the attributes of the alternatives being examined, the characteristics of the decision makers making the decisions, or some combination of the two. By observing how variations in the attributes of the alternatives or characteristics of the decision makers systematically vary with the observed choices within the market, the influence on choice behaviour of each attribute describing the alternatives and/or characteristic of the individual decision makers can be determined in the form of utility weights. Based on the overall combined utility values derived for each alternative within the model, the probability that each alternative will be chosen is then calculated, which subsequently provides a direct mapping between the individual specific choice environment described within the data and overall market behaviour.

Despite currently being somewhat less used in practice, early theorists identified that the attitudes and perceptions held by different decision makers towards both the objects being considered, and towards more general life concerns, can play an important role in choice behaviour. For example, McFadden (1980), based on theories of consumer behaviour, discussed the importance of including attitudinal and perceptual data within the DCM framework, stating “The theory of the economically rational utility-maximizing consumer, interpreted broadly to admit the effects of perception, state of mind, and imperfect discrimination, provides a plausible, logically unified foundation for the development of models of various aspects of market behavior.” Citing earlier work on hybrid conjoint models (e.g., Green et al. 1981 and Green 1984) as his inspiration, McFadden (1986) next set out a formal theory of consumer decision related to consumer decision making and choice and in doing so derived the theoretical framework for the hybrid choice model (HCM) (McFadden 1986). Importantly, the framework established by McFadden allowed for the separate inclusion of both general attitudes and alternative specific perceptions within the utility functions of a DCM via a latent variable structure.

The behavioural framework set out by McFadden (1986) uses the definition of attitudes first established by Allport (1935) who defined attitudes as “mental or neutral state of readiness, organized through experience, exerting a directive or dynamic influence on the individual's response to all objects and situations to which it is related. A simpler definition of attitude is a mind-set or a tendency to act in a particular way due to both an individual’s experience and temperament” (quote taken from Pickens 2005). Whilst somewhat related to attitudes, perceptions are defined as being the organization, identification and interpretation of sensory information aiming at representing and understanding the environment (Schacter et al. 2011). In a practical sense, the model framework defines perception as the translation of a stimulus into something meaningful to the subject based on prior experiences, which however, may be substantially different from reality.

The behavioural framework established by McFadden (1986) was first operationalized and tested by Train (1987) to explore the impact of attitudes and perceptions upon rate scheduling for public utilities. Swait (1994) used a similar modelling structure to model preferences for different beauty products. Similar conceptual frameworks were made operational around this time assuming various transportation contexts by Morikawa (1989), Morikawa et al. (1990), Ben-Akiva and Morikawa (1990), Vieira (1992) and Ben-Akiva and Boccara (1995). Hensher (1990) used a comparable model framework to explore user preferences for various bus services. Although formally acknowledged in Louviere et al. (2000) and Ben-Akiva et al. 2002), until the recent appearance of Bahamonde-Birke et al. (2017), the idea of treating separately perceptions and attitudes appears to have been somewhat abandoned in the early to mid-2000s, with the two concepts being used interchangeably (at least in name).

The main focus over the past decade or so has been on the inclusion of attitudinal data (without a clear distinction being made between attitudes and perceptions) within DCMs with a spate of recent papers examining how attitudes influence bicycle choice (e.g., Maldonado-Hinarejos et al. 2014), the relationship between attitudes, mode choice and the built environment (e.g., Van Acker et al. 2011), and how attitudes affect automobile choice (e.g., Daziano and Bolduc 2013). Other recent papers have explored methodological issues related to how best to incorporate attitudes within the DCM/HCM framework (e.g., Ben-Akiva et al. 2002, Danziano and Bolduc 2013a, Raveau et al. 2012, Vij and Walker 2014).

More recently, Bahamonde-Birke et al. (2017) have argued that there exists a difference between attitudes and perceptions, where attitudes are defined as “a mind-set or a tendency to act in a particular way due to both an individual’s experience and temperament” and perceptions the process by which individuals experience their environment (Lindsay and Norman 1972) and depend, therefore, on both the person and the stimuli (Pickens 2005)”. Note that this represents a call to return to the original model framework outlined originally by McFadden (1986).

Within the context of a high-speed rail choice, Bahamonde-Birke et al. (2017) collected a series of indicators, some of which are related to rail, whilst others are related more general issues, classifying the former as being related to perceptions and the later to generalised attitudes. Bahamonde-Birke et al. (2017) then used a HCM to relate the perceptual indicators to a series of latent variables defined by the level of service experienced recently by respondents as well as the characteristics of the respondents, whilst making the general attitudes a function of the socio-economic characteristics of the respondents only. The resulting latent variables are then used to enter into the utility functions of high-speed rail alternatives defined within a stated choice experiment.

The approach adopted by Bahamonde-Birke et al. (2017) represents only a partial representation of the behavioural framework described by McFadden (1986) in that the method seeks to link the perceptual indicators to the levels of service experienced in an actual context and then relate these to a DCM based on stated preference data. The original framework however suggests that the attributes and levels describing the alternatives faced within a decision context are interpreted by individual decision makers at the time that the choice is being made, and it is this interpretation representing the perceptual beliefs of the individuals that impacts upon utility, not the original levels seen by the individuals (at least directly). Consider for example a decision maker faced with a trip where one alternative takes approximately 20 minutes. Whilst traditional DCMs would enter 20 minutes directly into the utility function representing that alternative within the model, the framework offered by McFadden (1986) would suggest that what should be included within the model is how the respondent interprets 20 minutes of travel time. That is, does the decision maker actually perceive 20 minutes as 20 minutes, and qualitatively, how do they experience that time? Bahamonde-Birke et al. (2017) attempt to mediate this perception via a previous experience, whereas McFadden (1986) would suggest that the interpretation occurs at the time of choice, and whilst perception may be influenced by previous experiences, perception occurs at the time of choice, irrespective of whether the choice is made in a stated or revealed preference setting.

In this paper, we compare the results of a series of stated choice experiments in which using a split sample, respondents are given either a traditional choice experiment (control sample), or one in which they are asked to provide their perceptions to each of the attributes presented for each choice task experienced. In the example, respondents are shown two alternatives described by travel time, travel cost, and the number of accidents occurring on the road network over the previous year. Rather than provide a single travel time and cost, respondents are given distributions of both representing uncertainty as to the actual travel time and cost experienced should that route be chosen. For each attribute, respondents are asked to answer a number of questions designed to elicit their perceptions of the attributes experienced. Respondents are also asked to provide information about the overall route, as well as generalised attitudes. The study design allows for a comparison between models estimated on data collected for the control group and models estimated based on the perceptual and actual data presented to the treatment group, with the models estimated on the perceptual indicators being consistent with the framework originally proposed by McFadden (1986). This represents the first true comparative study to test the original McFadden (1986) behavioural framework known to the authors, with the previously cited works either implementing the model framework (or some derivation thereof) or not, without comparing models collected on different types of choice data.

Michiel Bliemer (The University of Sydney, Australia)
John Rose (University of Technology, Sydney, Australia)
Matthew Beck (The University of Sydney, Australia)
Generating partial choice set designs for stated choice experiments

ABSTRACT. See attached PDF file for extended abstract.

09:00-10:30 Session 9F: Land Use -- Self-Selection and Dynamics
Junyi Zhang (Hiroshima University, Japan)
Xiaodong Guan (Hong Kong Baptist University, Hong Kong)
Donggen Wang (Hong Kong Baptist University, Hong Kong)
Exploring the multiplicity of travel-related self-selection
SPEAKER: Xiaodong Guan

ABSTRACT. Please see the attached PDF file.

David Pérez Barbosa (Hiroshima University, Japan)
Junyi Zhang (Hiroshima University, Japan)
Joint influence of built environment and travel behavior of young people on their future life and migration choices: a case of study in rural Japan
SPEAKER: Junyi Zhang

ABSTRACT. It has been of increasing interest among multidisciplinary researchers to further clarify how the adequate designs of neighborhood spaces can enhance the well-being of their residents. Promoting people’s active and independent mobility has been largely discussed as one of the mechanisms of well-being enhancement and promotion.

Japan is a country that faces current issues with a decreasing population and birthrates, together with aging population. These issues have recently drawn significant attention from policymakers and stakeholders in Japan. In addition, it is mostly young people who move to the cities, and that means that as Japan’s population ages, many cities and towns outside the main metropolitan areas (such as Tokyo, Osaka, Nagoya) are getting slowly depopulated to points of partial and total abandonment. That brings other problems such a decrease in the workforce, making necessary for old people to reduce their dependence on the working population as much as possible.

This poses serious challenges for the development of future Japanese society in the mid and long-term. Recently, the development of compact cities has increasingly been attracting interest as a measure for tackling several problems associated to the depopulation. It is considered that in compact cities the necessary functions for daily living could be located within walking distance and the access to these functions and services would be much more convenient. This type of new development is becoming increasingly necessary for regional concentration and revitalization in the context of population change. In addition, a part of the efforts for revitalization should be oriented towards making these areas increasingly attractive for younger generations to settle.

With these issues in consideration, we conducted a survey in the period May 2016 – September 2016 among high school students in several rural locations in Hiroshima Prefecture. For comparative purposes, we distinguished between depopulating and non-depopulating areas for selecting the survey locations. The depopulating areas are municipalities where the decreasing population has already been officially recognized by the government of Japan. This recognition has a purpose in terms of how subsidies from the national government are oriented towards smaller regions and similar policy implications.

The main parts of the survey are listed as follows: information of their daily trip to school, description of their living (built) environment, assessment of well-being and preferences regarding future life plans as well as for future migration.

The living environment is characterized by the distance to key facilities in the corresponding or nearest urban area. Facilities that have been considered include but are not limited to recreational facilities, shopping facilities and facilities that are relevant for the interests of young people. We considered facilities related to educational activities (school, cram school, kindergarten), travel (bus stop, train station), services (Bank, post office, city hall, community center, health center or hospital, Police station), shopping of basic items (convenience store, Drugstore, Supermarket), shopping of other items (clothing shop, shopping mall, bookstore) and leisure and entertainment (Park, Game center, Bowling center, Baseball practice center, Swimming pool and various sport facilities).

Travel behavior is represented by the most frequent trip to school, in terms of travel mode or combination of modes and travel duration. The considered modes are walking, bicycle, train, bus or car (being transported by car); and combinations of them accordingly. It was found that in rural depopulating areas, students depend on buses for commuting in a much higher proportion than students in non-depopulating areas; and this affects negatively their well-being in various aspects.

Future life choices are considered in 3 big groups: family choices, career choices and individual choices. Family choices relate to the willingness to consider options such as marrying and raising children in the future. Career choices related to the willingness for options such as working in a company, creating one’s own business or continue working in a family business. Finally, individual choices relates to life options that can be decided without the involvement of other or future family members such as the future willingness to travel, get a car, travel or go for higher education. In addition, there is an additional section in the survey questionnaire that inquiries into a set of preferred future migration options and the level of intention to do them; such as migrating permanently abroad, migrating permanently to somewhere else in Japan, migrating temporarily, stay temporarily or decided to stay in the current place. Finally, considering the life oriented approach (Zhang, 2014; Zhang, 2017), respondents also do a self-assessment of their well-being (i.e. happiness) regarding several life domains.

With this information available, multilevel structural equation modeling (MSEM) becomes useful for modeling the complex relationships among the variables to be studied on different levels as well as across different levels (see Figure 1). Due to its flexibility, we consider that it can be used for the cross-sectional data of our study. These behavioral models allow us to understand in more detail the main reasons, drivers of migration and motivations for migration-related preferences in rural areas of Japan. We found that satisfaction with finances, achievements in life and standards of life are among the main factors that are negatively influenced by a difficult accessibility. Some urban facilities have also relevant impacts on well-being. Closeness of access to supermarkets, parks, sports facilities and shopping centers are the most influential for increased feelings of social inclusion and happiness among our respondents. In addition, long distances from sport and entertainment facilities (i.e. game center, swimming pool, park, sports facilities, bowling alleys and baseball practice centers) have a considerable negative impact on future health habits.

Regarding the future plans, the most influenced plans are related to career plans, which are one of the main drivers of migration in the long term. Family related plans were not related directly to living environment. Career related plans are confirmed the main driver for out-migration in depopulating areas, whereas there is also more inclination to consider working or owning a family business in depopulating areas than in non-depopulating. On the other hand, pursuing individual plans is also strongly associated with the decision of out-migration for both depopulating areas and non-depopulating areas, with a higher proportion of students in depopulating areas in comparison.

With the application of multilevel structural equation models, we seek to understand with more detail and predict more accurately the behavioral responses to changes in the relevant policy variables related to urban planning. We apply this principle to the most relevant changes in the living environment that would have some influence in migration intentions of young generations, which in many cases will correspond to their migration behavior in a few years.

The phenomenon of out-migration and depopulation cannot be reversed in the short-term, but it is a necessity to consider a smart-shrinking of urban areas should be considered; thus allowing city planners to allocate efforts in order to better and more efficiently control as well as to mitigate the negative consequences of depopulation; with more restrictive budget limitations. Thus, this study contributes with information of interest for policymakers at the local, regional and national level.

Jonas De Vos (Geography Department, Ghent University, Belgium)
Dick Ettema (Utrecht University, Netherlands)
Frank Witlox (Geography Department, Ghent University, Belgium)
Does the previous residential neighbourhood affect travel behaviour of recently moved residents?
SPEAKER: Jonas De Vos

ABSTRACT. A considerable amount of studies have indicated that people to some extent select themselves in specific neighbourhoods allowing them to travel in their desired way. Although a lot of studies analysed the degree to which travel preferences affect the residential location choice, few studies looked at the effect of a residential relocation on people’s travel behaviour and attitudes. A new residential context has the potential to disrupt previous travel choices and could potentially change people’s attitudes. This study – using 1,539 recently relocated residents in the city of Ghent (Belgium) − analyses self-reported changes in mode frequency and travel attitudes after a relocation, and uses a cohort approach to look at mode frequency and attitudes at different times after the relocation took place. Results suggest that (i) travel attitudes often influence the residential location choice, and (ii) both travel attitudes and travel mode choice change after a relocation, albeit in different ways depending on the current (urban versus suburban) and previous residential neighbourhood (more/equally/less urbanised). This study also suggests that a (possible) dissonance between travel attitudes and the residential neighbourhood is partly a temporal situation, as attitudes gradually change in accordance with the new residential environment.

09:00-10:30 Session 9G: Big Data -- Time Space
Andrew Bwambale (University of Leeds, UK)
Location: UCEN Flying A
Tien Mai (Université de Montréal and CIRRELT, Canada, Canada)
Xinlian Yu (University of Massachusetts Amherst, United States)
Song Gao (University of Massachusetts Amherst, United States)
Emma Frejinger (Université de Montréal and CIRRELT, Canada, Canada)
A link-based recursive route choice model for stochastic and time-dependent networks
SPEAKER: Xinlian Yu

ABSTRACT. This paper considers the routing policy choice problem in a stochastic time-dependent (STD) network. A routing policy is defined as a decision rule applied at the end of each link that maps the realized traffic condition to the decision on the link to take next. The problem is formulated as a Recursive Logit model for STD networks, in which the probabilistic choice of the next link is modeled at each link, following the framework of dynamic discrete choice models. An algorithm for solving the value functions that relies on matrix operations is proposed so that the model can be estimated in reasonable time. Estimation and prediction results are presented for a network with 2772 nodes and 5447 links (including 619 stochastic links) situated in the Stockholm region, Sweden.

Andrew Bwambale (University of Leeds, UK)
Charisma Choudhury (University of Leeds, UK)
Stephane Hess (University of Leeds, UK)
Analysing time-of-travel choices using mobile phone data

ABSTRACT. The temporal variations in traffic congestion across the day are a reflection of the departure time choices by different individuals operating within a metropolitan area. The last few decades have been characterised by rapid growth in passive mobility tracking technologies, which has led to a few departure time choice studies based on GPS data. Although the use of smartphone apps has reduced the cost of obtaining GPS data, such studies remain expensive and thus usually involve small samples. On the other hand, the rapid growth in mobile phone penetration rates worldwide has led to the emergence of alternative passive mobility datasets such as Global System for Mobile communication (GSM) data, which can anonymously cover much wider proportions of the population. GSM data reports the IDs of all the GSM cells traversed by an active mobile phone at regular time intervals. However, there has been no study using such data to model departure time choice decisions, which motivates this research. The developed framework takes into account the fact that the desired times of travel are unobserved and estimates distributions to account for this.This is implemented using advanced choice models and tested using data from the Nokia Mobile Data Challenge yielding intuitive results.

Yihong Wang (Delft University of Technology, Netherlands)
Gonçalo Correia (Delft University of Technology, Netherlands)
Bart van Arem (Delft University of Technology, Netherlands)
Harry Timmermans (Eindhoven University of Technology, Netherlands)
Understanding travelers’ preferences for different types of trip destinations based on mobile internet usage data
SPEAKER: Yihong Wang

ABSTRACT. This paper proposes to segment the population using mobile internet usage data and accordingly to better understand travellers’ preferences for different types of trip destination revealed in mobile phone traces. The method is tested for the city of Shanghai, China, by using a special mobile phone dataset that includes not only the spatial-temporal traces but also the mobile internet usage data of the same users. We identify statistically significant relationships between a traveller’s favourite category of mobile internet content and more frequent types of trip destinations that he/she visits. For example, compared to others, people who preferred the content category “tourism”, had a significantly higher preference to visit touristy areas. Moreover, people who used mobile internet more intensively were more likely to visit more commercial areas, and people who used less mobile internet preferred to have activities in predominantly residential areas.

Yousef Maknoon (Delft University of Technology, Netherlands)
Shadi Sharif Azadeh (Erasmus University Rotterdam, Netherlands)
Michel Bierlaire (Ecole Polytechnique Federale Lausanne (EPFL), Switzerland)
Dynamic assortment optimization based on customers’ behavior using transaction data

ABSTRACT. In this paper, we aim at investigating the impact of demand modelling on revenue performance for a shared mobility system. We consider the case where the information associated to the systematic part of the utility function is not available. We use dynamic assortment optimization to estimate utilities. We compare this assumption with the case where the choice probabilities are assumed to be static and homogenously estimated. The results indicate that this additional information helps to improve the service level as well as the revenue performance.

09:00-10:30 Session 9H: Attitudes and Perceptions Part 1
Sara Khoeini (Arizona State University, United States)
Location: UCEN Lobero
Sara Khoeini (Arizona State University, United States)
Shivam Sharda (Arizona State University, United States)
Denise Capasso Da Silva (Arizona State University, United States)
Ram Pendyala (Arizona State University, United States)
Chandra Bhat (The University of Texas at Austin, United States)
Unraveling the Relationship between Attitudes and Behavioral Choices Using a Latent Segmentation Approach
SPEAKER: Shivam Sharda

ABSTRACT. Please see attached PDF document.

Angela Stefania Bergantino (Department of Economics, Management, and Business Law, University of Bari, Italy)
Mauro Capurso (Institute for Transport Studies and Choice Modelling Centre, University of Leeds, UK)
Thijs Dekker (Institute for Transport Studies and Choice Modelling Centre, University of Leeds, UK)
Stephane Hess (Institute for Transport Studies and Choice Modelling Centre, University of Leeds, UK)
Modelling consideration of alternatives as a function of individuals’ attitudes and perceptions: an application to airport access mode choice

ABSTRACT. See uploaded file

Fazilatulaili Ali (Newcastle University, UK)
Dilum Dissanayake (Newcastle University, UK)
Margaret Bell (Newcastle University, UK)
Malcolm Farrow (Newcastle University, UK)
A Bayesian Approach of Multivariate Probit Model to Analyse Car Users’ Attitudes towards Environmental Impact

ABSTRACT. The abstract is submitted as a PDF attachment

Danique Ton (Delft University of Technology, Netherlands)
Lara-Britt Zomer (Delft University of Technology, Netherlands)
Florian Schneider (Delft University of Technology, Netherlands)
Sascha Hoogendoorn-Lanser (KiM Netherlands Institute for Transport Policy Analysis, Netherlands)
Dorine Duives (Delft University of Technology, Netherlands)
Oded Cats (Delft University of Technology, Netherlands)
Serge Hoogendoorn (Delft University of Technology, Netherlands)
Latent classes of daily mobility patterns: The relationship with attitudes towards modes
SPEAKER: Danique Ton

ABSTRACT. Active modes (i.e. walking and cycling) have received significant attention by governments worldwide, due to the benefits related to the use of these modes. Consequently, governments are aiming for a modal shift from motorised to active modes. Attitudes are generally considered to play an important role in determining mode choice and travel behaviour. Therefore, many studies have investigated the relationship between attitudes and behaviour. Understanding the relationship between the attitude towards modes and the daily mobility pattern, can help in providing input for efficient and effective policies that aim at increasing the active mode share. This research investigates patterns in the daily mobility patterns of individuals in the Netherlands and tests the relationship between these patterns and attitudes towards modes. This study identifies five classes of mobility patterns: 1) car and bicycle users, 2) exclusive car users, 3) car, walk, and bicycle users, 4) public transport users, and 5) exclusive bicycle users. Regarding attitudes towards modes, eight factors are identified: five mode related attitudes, two public transport related attitudes, and one related to the prestige of using modes. The results show that the majority of the users is multimodal in their daily mobility pattern. Furthermore, individuals are more positive towards modes that are used in the daily mobility pattern, compared to unused modes. The exclusive car users are most negative to the unused modes. Consequently, when the goal is achieving a higher active mode share, it might be more fruitful to target the multimodal classes and/or classes that contain active mode use, compared to the habitual car users. 

10:30-11:00Coffee Break
11:00-12:30 Session 10A: Mobility as a Service -- More Sharing
Catherine Morency (Polytechnique Montreal, Canada)
Location: Corwin West
Konrad Goetz (ISOE - Institut for Social Ecological Research, Germany)
Georg Sunderer (ISOE - Institut for Social Ecological Research, Germany)
Wiebke Zimmer (Institute for Applied Ecology, Germany)
Freefloating car sharing – acceptance, attractiveness and traffic behavior
SPEAKER: Konrad Goetz

ABSTRACT. Car sharing constitutes an innovation in the field of urban mobility, which is currently being widely acknowledged by the public as well as by science (cf. Shaheen et al. 2009 and Loose 2010). One reason for this are the enormous growth figures not only in Germany but also internationally (cf. Shaheen&Cohen 2007). These numbers comprise both users and vehicles available. The other reason for the high level of attention is the assumption that car sharing, as part of the urban mobility mix, should replace private cars and thus contribute to environmental protection. Finally, car sharing is being attentively observed because there has been a new, additional form called freefloating car sharing. Unlike the classic, station-based car sharing that has been around since the 1980s, cars now no longer have to be picked up and dropped off at the same location. Instead, vehicles within a certain area of a city are available for spontaneous use. They can be located, opened and used with the help of a smartphone app and afterwards be parked in any public parking. Both forms of car sharing have seen a significant growth in recent years. It is remarkable that growth figures for flexible car sharing in Germany were particularly high, indicating that after just a few years this new form of car sharing has clearly surpassed the old, station-based car sharing regarding its number of users (cf. bcs 2017). The research results we are about to present will help to understand the success of flexible car sharing. To this end, firstly the question will be answered as to which population groups use freefloating car sharing, i.e. who are the drivers of this innovation. Here, it is interesting to find out whether the freefloating car sharing is used by the same groups that use the station-based version, or whether certain social groups can be better reached with the new form of car sharing. Furthermore, we want to show what kind of mobility equipment (vehicles, public transport time cards) users have. Secondly, results on the attractiveness and acceptance of freefloating car sharing will be presented. Here, we are focusing on the reasons and motives for use. Finally, the third focal point to be presented during the conference will be the presentation of our results on travel behavior which are not yet available but will have been evaluated by summer 2018. The above-mentioned research questions are answered in a comparative manner focusing on to two engine variants: on the one hand, there is the variant with electric cars and on the other hand we look at the variant with conventional combustion engines. So in addition to general findings on the attractiveness and acceptance of freefloating car sharing, differences between the use of the two types of driving power can be identified. So the analyses also provide information on whether freefloating car sharing is a good way of integrating electromobility into the overall German urban mobility services. The analyses are based on data taken from the project "share - Scientific Accompanying Research of car2go with battery-electric and conventional vehicles", which is funded by the Federal Ministry for the Environment, Nature Conservation, Construction and Nuclear Safety (BMUB). 2. The research design of the share study In the “share” study, a comparative panel design was used. The users of the Freefloating Carsharing system were questioned four times over a period of two years. The strength of this design is, on the one hand, that attractiveness and acceptance as well as traffic behavior can be observed over a certain period of time. On the other hand, both versions of the system (electric cars and conventional ones) can be compared, hence "comparative" . The study was conducted by the Institute for Applied Ecology (lead) and the ISOE - Institute for Social-Ecological Research, Frankfurt/Main. The users are customers of the practice partner car2go, a subsidiary of the Daimler Group and provider of flexible car sharing. The Fieldwork was conducted in cooperation with aproxima Gesellschaft für Markt- und Sozialforschung, Weimar. In order to compare the two product variants, car2go customers were surveyed at two locations that provided vehicles with conventional engines (Frankfurt/Main and Cologne) and one providing electric cars (Stuttgart). Only new customers were admitted as participants. Recruitment was carried out via e-mail invitations shortly after registration at car2go and also via a project flyer for new customers. The particular focus on new customers served to ensure a valid assessment of the forecast situation. This was particularly relevant for the project’s question of a possible change in traffic behavior through participation in the Freefloating Carsharing and for the question if this participation influenced the households’ decision of whether to buy new or sell existing vehicles. For all participants the first interview took place immediately after registration. The following survey phases (2, 3 and 4) took place three months, one year and two years after registration. So the customers’ evaluation of the sharing system could be examined both after their first impressions and after a longer period of time. For all four surveys, the method of standardized online questionnaires was used. In order to retain a sufficient number of study participants, recruitment was carried out over a period of two years from June 2013 to May 2015, so that the total field time lasted until May 2017. During this time, a total of 3,096 new car2go customers were acquired for the survey. Two thirds of them are imputable to the Stuttgart location (electric vehicles), the remaining third to Cologne and Frankfurt am Main (vehicles with internal combustion engines). As a result of the normal panel mortality, the number of participants decreased over time. That is why, 2,237 interviews were conducted in the second survey wave, 1,757 in the third wave and 1,449 interviews in the last survey wave. This means that compared to the previous wave, we had very good response rates of 72 to 82 percent. The results presented below refer to the status of the ongoing evaluation in November 2017, i. e. parts of the project have not been evaluated so far. At the time of the IATBR conference in summer 2018, all of the findings will be available. 3. Results 3a) Sociodemography As a parallel service to classic station-based car sharing, freefloating car sharing has significantly contributed to the increased shared car use in recent years. The empirical results of the share project show that certain population groups can be better reached through free floating carsharing than through station-based carsharing. The groups we are talking about are mainly young adults aged between 18 and 30 years. Other population groups such as people with a low level of education and people aged over 60 have hardly been reached with free-floating car-sharing. The findings on sociodemography are in accordance with the results of other studies on flexible car sharing (WiMobil 2016; Riegler et al. 2016). 3b) Acceptance, attractiveness and barriers An important success factor of freefloating car sharing is to take into consideration and address people’s various motives, some of which are purely emotional. The results show that the system is seen as practical, cool, fun and environmentally friendly. However, environmental considerations are generally not in the foreground. The focus is rather on the perceived practicability for everyday mobility. The study shows that these positive ratings remain stable over time. However, two negative evaluations, which could act as future barriers were clearly stated: on the one hand, there are the costs which are perceived as relatively high (29 Eurocents per Minute) and on the other hand, there are shortcomings in the availability of car sharing vehicles in the respective cities. Both problems could jeopardize further success. When comparing the two propulsion variants, it can be said that the electric version is considered more environmentally friendly and almost as practical as the conventional one though there are small cutbacks with regard to the subject of loading. However, this is compensated for by the customers’ great driving enthusiasm and the advantage of the electric version in terms of coolness and sympathy. Overall, the comparison shows that the use of electric cars as part of flexible car sharing is a good way of familiarising people with electromobility and integrating electromobility into the transport system.

3 c) Traffic Behavior In the project, data on traffic behavior were collected by using the method of remembering a concrete day, analogous to the ‘Mobility in Germany’ study (see Infas 2010). The first survey phase right after the registration was dealing with the traffic behavior prior to the use of free floating car sharing. Waves 2 to 4 dealt with the traffic behavior after registration, i. e. the integration of car sharing into the daily mix of means of transport. The results on traffic behavior are currently being evaluated. They will be available for the conference. It is particularly interesting to see how the inclusion of car sharing has an influence on modal split or the distance-based modal split. It is of particular interest whether as a consequence of using car sharing the overall traffic behavior tends to be more environmentally friendly or not.

References: Bundesverband CarSharing (bcs). 2017. Aktuelle Zahlen und Daten zum CarSharing in Deutschland, Zugegriffen: 13. Oktober 2017 Infas (Hrsg.) (2010): Mobilität in Deutschland 2008. Bonn Zugegriffen 13. Oktober 2017 Loose, Willi. 2010. Aktueller Stand des CarSharing in Europa. Zugegriffen: 13. Oktober 2017 Riegler, S., Juschten, M., Hössinger, R., Gerike, R., Rößger, L., Schlag, B., Manz, W., Rentschler, C., Kopp, J. 2016. CarSharing 2025 – Nische oder Mainstream?. Abschlussbericht Projekt „Neue Nutzungskonzepte für individuelle Mobilität“. München: ifmo Shaheen, Susan A. & Cohen, Adam P. 2007: Worldwide Carsharing Growth: An international Comparison Zugeriffen: 13. Oktober 2017 Shaheen, Susan A./Cohen, Adam P./ Chung, Melissa S. 2009: North Amercian Carsharing-10-Year Retrospective, in: Transportation Research Record: Journal oft he Transportation Research Board, Nr. 2110: 2009:35-44 WiMobil. 2016. Abschlussbericht zum Projekt „Wirkung von E-Car Sharing Systemen auf Mobilität und Umwelt in urbanen Räumen“. Zugegriffen: 31. Juli 2017

Alexandros Nikitas (University of Huddersfield, Huddersfield Business School, UK)
Jana Sochor (Chalmers University of Technology, Division of Design & Human Factors, Sweden)
Analysing the Acceptability and Usage Patterns of Bike-sharing in a City with an Established Pro-cycling Culture

ABSTRACT. Bike-sharing is a rapidly emerging mobility initiative, with almost 1500 operating schemes in over 50 countries worldwide (Meddin & DeMaio, 2017), that genuinely supports transport policy-making focused on creating incentives for voluntary travel behaviour change. A bike-sharing scheme, which is synonymous to a public bicycle or cycle hire programme, is underpinned by the philosophy of sharing economies and its lessons for maximising the potential of available resources going beyond an ownership-bound regime for usage. It can be defined as a system that provides affordable short-term access to a fleet of locally customised bicycles on an “as-needed” basis that could extend the reach of public transit services to final destinations and be a door opener for increased bicycle usage. Bike-sharing, typically ‘framed’ as an ideal first- and last-mile travel solution for congested metropolitan environments (DeMaio, 2009) and less often tested for the context of smaller and medium cities (Caulfield et al., 2017) is now with the very recent and aggressive introduction of station-less, free-floating, GPS-managed schemes more relevant than ever before.

Nonetheless, and despite a growing interest from transport scholars, the existing literature on bike-sharing is still lagging; there is a lack of evidence reported on existing schemes on whether they achieved their objectives (Ricci, 2015) and a paucity of research with large numbers of individuals who are not bike-share users despite the fact that these studies are of vital importance to bike-sharing user growth (Fishman, 2016). The present work is an effort to contribute in addressing this critical research gap. It is based on a primarily quantitative study of 535 online survey responses that was designed to identify perceptions reflecting and affecting the public acceptance of the bike-sharing scheme of Gothenburg, Sweden (known as Styr & Ställ) and to examine bike-sharing usage patterns and behavioural norms.

Gothenburg, the fifth largest city between the Nordic countries with a broader metropolitan area population close to one million, has Sweden’s largest cycle network with a total length of approximately 800 km. The city has been traditionally governed by politicians showing strong support to bike-related initiatives so there is a long-lasting commitment on cycling infrastructure and service investments. Gothenburg has therefore forged over the years a pro-cycling culture and can be classified as a bike-friendly city; this is an identity that could theoretically help the growth and success of a bike-sharing initiative. Styr & Ställ, launched in August 2010 and incrementally expanded from that point forward, is a traditional station-based self-service bike rental system. It is spread across 66 city centre stations offering more than 1000 bicycles to its annual subscribers and three-day pass holders. Thirty-minute trips are free for the users; charges apply only for usage beyond this time threshold. There is also a window of 15 added minutes without a charge if there are no available stands at the return station provided that the customer uses the terminal to locate a nearby station with availability. The scheme operates between March and December due to severe winter weather conditions on January and February.

An online questionnaire, promoted through social media and e-mailing lists, was the data collection tool used for this study. It contained 20 questions (some of them including various sub-questions) administered in six thematic parts namely: the respondents’ daily commuting experience; their perceptions on cycling in general and cycling in Gothenburg in particular; their perceptions about Styr & Ställ; their past and current bike-sharing usage patterns; their bike-sharing experiences from the scheme; and their socio-demographic characteristics. The section about real-life user experience was used to provide the respondents with an opportunity to evaluate some chosen characteristics and qualities of Styr & Ställ; this would help developing an understanding of what works well with the scheme and what needs to be improved. This particular section was completed only from the individuals that had used the scheme at least once. The other survey sections were open to every respondent. Five-point Likert-scales varying from ‘strongly agree’ to ‘strongly disagree’ were used for consistency reasons. Participation incentives, in the form of a prize draw for a new bicycle, and an introductory note explaining the purpose of the study were used as a means of improving response rates. The final sample refers to 535 responses with 368 completions of the usage experience section.

The gender and household splits of the sample were reasonably balanced. There was an over-representation of people aged 20 to 39 and of people using bicycles as their primary means of travelling to work or their most frequent destination, possibly because of the online nature of the survey and because the attitude object (i.e. bike-sharing) being more topical to them respectively. Almost 90% of the respondents declared that they rarely or never experience traffic congestion in Gothenburg. Close to one third of them declared that they have the exact same travel behaviour between winter and summer; one third of them self-reported that they have very different behavioural patterns during winter especially those cycling and walking to work. Cycling had a positive significance to the majority of the respondents. In general, the respondents felt that cycling is a sustainable, cost-saving, enjoyable, healthy travel choice with a strong potential to help efforts meaning to reduce road traffic congestion. The respondents were in favour of more cycling-related investments, even the ones never or rarely cycling. Reduced safety was the only negative aspect associated with cycling; only 46.4% of the respondents positively classified cycling as a safe mode.

When asked about Gothenburg’s bike-sharing scheme per se the majority of the respondents believed that Styr & Ställ is an affordable travel mode with the capacity to promote healthy living, improve road traffic conditions, make cycling more popular, complement the city’s other public transport services and help the city becoming more liveable. Also 85.3% of them recognised bike-sharing’s potential for making people’s travel behaviour increasingly less car dependent. Most importantly though, the survey participants acknowledged the significance of bike-sharing for their city; 92.4% agreed or strongly agreed that the scheme is good for Gothenburg and 93.5% disagreed or strongly disagreed with the notion that Styr & Ställ is a sub-standard transport initiative. Even the respondents that have never used it before or self-reported no (or little) intention to bike-share in the future were positive towards the scheme. Many respondents were also supportive of the scheme’s further expansion through more bike-sharing investments and considered that the scheme is a viable public service for the city; 86.5% and 96.1% agreed or strongly agreed with these notions respectively. Also only 6.5% of the people that have actually used the scheme considered that there is something fundamentally flawed with it.

Despite these high acceptability rates our study found that the majority of the respondents rarely used the scheme even as a secondary travel option; 76.8% of the respondents stated that they never use the scheme while only 2.8% use it as their main mode choice. Nonetheless, longitudinal data collected for the first four full operating years of Styr & Ställ (based on respondents’ self-reporting capacity) indicate a small but distinct annual increase in the number of participants that used the scheme especially for those using it as a main or a secondary travel alternative to their typical modal choice. This underuse indicates that there is, in theory, a massive untapped potential for utilising, in real usage terms, the scheme’s wide acceptance.

Ordinal regression analysis modelling was used as a means of understanding better what factors influence the perceptions referring to the acceptance of bike-sharing and the level of bike-sharing usage. Numerous models were tested, using a variety of combinations of independent variables (each representing one of the key survey elements), that could explain and quantify these relationships. This work discusses only the two best fits; one for acceptance and one for usage. The regularity with which people cycle to work, household income, the perception of how available public bicycles are currently across the city, the most frequent (and important) reason behind respondents’ travel mode choice and perceptions about cycling safety were found to influence the respondents’ acceptance of further bike-sharing investments. This dependent variable was chosen since it represents a very powerful metric for support; people not only accept the scheme per se but they want to see it further expanded. On the other hand, the regularity with which people cycle to work, gender, the level of travel behaviour change between winter and summer time and the availability of a bicycle in the household were all factors influencing the respondents’ bike-sharing usage frequency.

When the respondents were asked to directly identify reasons for not using (or not using frequently) Styr & Ställ the lack of good bike-sharing infrastructure (e.g. inconvenient or too sparsely located docking stations) and road safety concerns for bike-share users were recognised as two main usage barriers. These barriers are in line with the literature; safety concerns have been reported as obstacles to bike use generally and docking station scarcity has been reported as an obstacle referring specifically to bike-sharing (Fishman et al., 2014). The nature of these barriers allows policy-makers and scheme operators to make appropriate capacity and safety-enhancing improvements that could address the associated challenges.

Also only half of the respondents found the public bicycles (53.6%) and the docking stations (49.5%) of the scheme to be attractive. The respondents that have actually used the scheme at least once were not very satisfied with the comfort of the bikes, their availability on the stations and the overall ease of using the scheme; only 40.3%, 36.3% and 50.9% of the respondents were satisfied respectively. So better design enhancing the quality of bikes and stations and efforts to simplify the rental process can potentially improve the scheme’s appeal to road users.

Nevertheless, the most important roadblock for increased bike-sharing usage directly self-reported, was one that could be very relevant for other cities with an established pro-bicycling identity and one that policy-makers and scheme operators cannot do anything about; 41.1% of the respondents were not interested to use the scheme because they owned and preferred using their private bikes. This finding perhaps indicates that bike-sharing success, in usage terms, may not be always realistically achievable; in the context of Gothenburg, or in any bike-friendly city probably, many people will always prioritise the use of their private bicycle over bike-sharing. Expectations for bike-sharing usage should therefore be modest; this is the price to pay for having a scheme in a city that has an established bike-friendly character. So bike-sharing success should be perhaps redefined for these environments and be measured primarily in terms of a scheme’s potential to be accepted by the general public. High willingness to see a scheme being maintained and expanded even if there is little or no likelihood of the individuals showing this type of acceptance to actually ever using it, is a win for bike-sharing in its own right.


1.Caulfield, B., O'Mahony, M., Brazil, W. and Weldon, P. (2017). Examining usage patterns of a bike-sharing scheme in a medium sized city. Transportation Research Part A: Policy and Practice, 100, 152-161.

2.DeMaio, P. (2009). Bike-sharing: History, impacts, models of provision, and future. Journal of Public Transportation, 12(4), 41-56.

3.Fishman, E. (2016). Bikeshare: A review of recent literature. Transport Reviews, 36(1), 92-113.

4.Fishman, E., Washington, S. and Haworth, N. (2014). Bikeshare’s impact on car use: Evidence from the United States, Great Britain and Australia. Transportation Research Part D: Transport and Environment, 31, 13-20.

5.Meddin, R. and DeMaio P. (2017). The Bike-Sharing World Map. Available online: (retrieved 01/11/2017).

6.Ricci, M. (2015). Bike sharing: A review of evidence on impacts and processes of implementation and operation. Research in Transportation Business & Management, 15, 28-38.

Catherine Morency (Polytechnique Montreal, Canada)
Jean-Simon Bourdeau (Polytechnique Montreal, Canada)
Hubert Verreault (Polytechnique Montreal, Canada)
Modelling the interactions between mobility options in the surrounding of bikesharing stations

ABSTRACT. In the recent years, many cities have seen a tremendous increase in the number and variety of transportation systems operating in their area. Bikesharing, station-based carsharing, free-floating cars, ridesharing are among the mobility options that have become more and more available in addition to typical modes such as transit and taxi. Hence, many cities are also investing more for the development of infrastructures for active modes. This diversified set of options is changing the way people travel as they are more prone to combine multiple services to fulfill their needs.

In the Montreal Area, station-based carsharing has been available for more than 20 years, free-floating cars are now available in various zones and an important bikesharing system is provided since 2009. Also, four subway lines provide efficient access in the central part of the region, with multiple bus lines connecting stations and important destinations. This setting provides an interesting ground to observe and model the intensity and patterns of usage of the various modes available to the travellers. Many contributions can be found in the literature on the modelling of the usage of particular transportation modes. For instance, Ciari et al. (2014) and Heilig et al. (2017) model both types of carsharing systems (station and free-floating) while Wang et al. (2016) and Morency et al. (2017) model bikesharing station activity using operational data. Smart card data from transit system are also used often to report on the variability of transit use, as done by Morency et al. (2007), Briand et al. (2017) and Agard et al. (2013).

However, there are still very few researches looking simultaneously at the use of various modes and at their interactions i.e. how the use of one system may impact the use of another system. Still, papers discuss the future of urban mobility that builds on the strength of various modes (Spickermann et al., 2014).

Unfortunately, there is no single representative and large-scale data set containing the travel behaviours of individuals across these various options. In the Montreal region, large-scale Origin-Destination travel surveys are quite helpful in measuring trends for a typical weekday but are not suited to observe the variability of behaviours over time or the use of low-share modes. While there has been an app-based survey conducted recently in the region, it is not neither suited for the analysis since there was no sampling frame nor process to factor the sample to ensure its representativeness. Hence, it does not cover a sufficiently long period of time to understand the various interactions between modes in the presence of changing conditions. Since there is no available dataset to model the interaction at the user level, the analysis is based on the processing of passive streams outputted from operational systems and conducted at the system-level.

This research, which is part of a wider program on the modelling of the interactions between modes, focuses on the interactions between at least 6 modes (bus, subway, taxi, bikesharing, station-based carsharing, free-floating cars) in the surrounding of bikesharing stations. Car, cycling and walking may be added to the analysis of the required relevant data become available (most possibly counts) in the upcoming weeks. Various sets of data, covering at least a 2-year period (2015-2016 and maybe 2017), are processed: • Transit (subway): the use of the subway system is captured through the smart card validations at the entry station. Only entrance points are available. The number of validations per hour per day is the usage indicator. The system generates some 250 millions transaction per year. • Transit (bus): the use of the bus system is also captured using the smart card transactions though for buses, we don’t know where the transaction is occurring but we know a what time and on which line. Again, smart cards are only tap-in. The number of bus validations per hour per day is the usage indicator. There are some 240 millions validation per year on the Montreal buses. • Carsharing (station-based): shared cars are reserved by members and it is possible to measure the number of reservations starting each hour of each day. It is the indicator used to assess the level of usage but it is important to mention that a transaction does not necessarily amount to one trip since a car can be rent to perform a set of trips and that the car can be reserved for multiple days. We count some 360 000 reservations per year. • Carsharing (free-floating): for free-floating carsharing, the number of transactions more or less equals the number of trips. Hence, the usage indicator is the number of transactions starting each hour of each day. There are around 240 000 trips observed in 2015 on the Auto-mobile service. • Taxi: the use of taxi is estimated using the number of trips made using four taxi companies in the areas (accounting for almost 50% of the fleet operation on the Montreal Island). The usage indicator is the number of taxi trips starting each hour of each day. Each taxi car conducts an average of 12 customer trips per day (so we estimate the number of yearly trip on the island at 17 million). • Bikesharing: a 5000 bikes’ service, Bixi, is available in the region. The use of the system is reflected by a database describing each transaction including origin and destination station and timestamp at both points. The number of transaction starting each hour of the day of each day is used as usage indicator. It is worth noting that the system is typically only available from April to November (due to winter). There were 3.5 million bikesharing trips during the 2015 season.

To be able to capture the dynamic of the interactions between all these mobility options, patterns are examined in specific locations, for this paper in the surroundings of bikesharing stations. Mobility interaction analysis zones (MIAZ) are created and daily patterns of all systems are constructed for each one, over 2 years. Various definitions of MIAZ are tested: straight line buffer within 400 m to 1200 m, network distance buffer with similar distances and mutually exclusive areas based on Voronoi surfaces. Data mining techniques are used to understand the typical patterns of use for each mode as well as for all modes combined and cross-analysed with variables describing the MIAZ (population, activity opportunities, level of service for each mode) as well as the context (namely fuel price and weather). Clusters of daily patterns and usage intensity are created and used in a regression tree to identify the most important factors in the allocation of travel days to clusters. The purpose of first to explain typical days using explanatory factors but also to forecast type of day using these same variables for unobserved days (i.e. not used in the clustering process).

Since what is happening in a MIAZ, on a daily basis, is not necessarily independent from what is happening in the surrounding MIAZs, spatial autocorrelation measures are estimated to validate the scale of the spatial interaction. Finally, at the light of the clustering results and the spatial analysis, longitudinal models are developed (first growth models) to model the evolution of usage patterns of all modes and assess the contribution of various classes of explanatory factors.

References • Ciari, F., Bock, B., & Balmer, M. (2014). Modeling Station-Based and Free-Floating Carsharing Demand: Test Case Study for Berlin. Transportation Research Record: Journal of the Transportation Research Board, No 2416, pp. 37-47. • Heilig, M., N. Mallig, O. Schröder, M. Kagerbauer, and P. Vortisch. Implementation of free-floating and station-based carsharing in an agent-based travel demand model. Travel Behaviour and Society, Available online 23 February 2017. • Morency, C., Trépanier, M., Paez, A., Verreault, H., Faucher, J. (2017) Modelling bikesharing usage in Montreal over 6 years. CIRRELT PAPER CIRRELT-2017-33, available online: • Wang, X., Lindsey, G., Schoner, J.E., Harrison, A. (2016). Modeling bike share station activity: Effects of nearby businesses and jobs on trips to and from stations. Journal of Urban Planning and Development, Vol. 142, No. 1, 2016. ttp:// • Morency, C., M. Trépanier, and B. Agard. Measuring transit use variability with smart-card data. Transport Policy, Vol. 14, No. 3, 2007, pp. 193 - 203. • Briand, A.-S., Côme, É., Trépanier, M., Oukhellou, L. (2017), Analyzing Year-To-Year Changes In Public Transport Passenger Behaviour Using Smart Card Data, Transportation Research Part C: New Technologies, Volume 79, June 2017, pp. 274–289. • Agard, B., V. Partovi Nia, and M. Trépanier. Assessing public transport travel behaviour from smart card data with advanced data mining techniques. Presented at the 13th World Conference on Transport Research, Rio de Janeiro, Brazil, 2013. • Spickermann, A., Grienitz, V., von der Gracht, H.A. (2014). Heading towards a multimodal city of the future? Multi-stakeholder scenarios for urban mobility, Technological Forecasting & Social Change, 89 (2014) 201–221.

Shiva Habibi (Chalmers University of Technology, Sweden)
Frances Sprei (Chalmer University of Technology, Sweden)
Predicting the success and demand of free-floating car-sharing services in cities
SPEAKER: Shiva Habibi

ABSTRACT. In recent years, free-float car sharing services (FFCS) have been offered as a more flexible mobility solution. Unlike traditional car sharing where vehicles must be returned to the specified stations, FFCS allows users to pick up and return cars anywhere within a specified area of a city. This flexibility would allow expanding car-sharing as a mobility resource.

In this study, we estimate a discrete-continuous model of viability and demand for FFCS services in different cities.  The discrete variable is a binary variable of whether or not a city is a viable option for these services in terms of city characteristics and the continuous variable is the number of booking per day. To estimate this model, we have a dataset on 32 cities in Europe and North America with access to FFCS. From this dataset, we construct attributes on usage and efficiency of FFCS to be included in the model. On the other hand, we have data on more than 300 cities in Europe and North America. The variables include city demographics and socioeconomics, spatial structure and transport system characteristics.

This model predicts the success of these services in different cities. The results will provide insights to the policymakers to realize the characteristics that help develop or cause failure of these services in their cities and support them to evaluate with the relevant policies.

11:00-12:30 Session 10B: Travel Demand - New Methods
Lei Liu (Dalhousie University, Canada)
Location: MCC Theater
Thomas O. Hancock (University of Leeds, UK)
Stephane Hess (University of Leeds, UK)
Charisma F. Choudhury (University of Leeds, UK)
Quantum probability: a new method for considering travel choices

ABSTRACT. Please find the attached file containing the extended abstract.

Divyakant Tahlyan (University of South Florida, United States)
Parvathy Vinod Sheela (University of South Florida, United States)
Michael Maness (University of South Florida, United States)
Abdul Pinjari (Indian Institute of Science, India)
Improving the spatial transferability of travel demand forecasting models: An empirical assessment of the impact of incorporating attitudes on model transferability

ABSTRACT. 1. Introduction Spatial transferability of travel forecasting models, i.e. the ability to use a travel forecasting model developed in one region for travel forecasting in another region, is of considerable interest due to a variety of reasons (Sikder et al., 2013). First, a model’s performance assessment in various contexts provides an understanding of its ability to provide robust forecasts under different scenarios. Second, the ability to transfer models between regions can save significant cost and time for regions that cannot afford to build a model from the scratch. The issue of spatial transferability is relevant to not just small/mid-sized regions in the United States, who are generally short of funds to conduct an extensive data-collection. It is also relevant to planning agencies in many developing countries, which generally have a meager budget for transportation planning (San Santoso and Tsunokawa, 2005). Spatial transferability of travel forecasting models has been widely studied over the past few decades (Badoe and Miller, 1995; Ben-Akiva and Bolduc, 1987; Sikder et al., 2013; Yasmin et al., 2015) and a review of the literature on spatial transferability can be found in Sikder et al. (2013). In this context, a positive relationship between model specification and model transferability has long been argued (Atherton and Ben-Akiva, 1976; Burnett, 1981; Koppelman and Wilmot, 1986; Yasmin et al., 2015). Further, it has been postulated that travel forecasting models that are more behaviorally realistic and those that better capture heterogeneity in travel behavior are more transferable than traditional models (Sikder et al., 2013).

In the context of heterogeneity, there is increasing recognition that heterogeneity in travel behavior is not just attributable to socioeconomic and demographic differences but may also be due to differences in individuals’ attitudinal factors. Incorporating attitudes and perception variables into discrete choice models is most widely performed using integrated choice and latent variable (ICLV) models. ICLV models offer greater insights into the decision making process by including additional information through measurement equations for the latent variables. It is also believed that ICLVs produces more efficient model outputs (i.e. with less variation), such as demand elasticities and market predictions (Vij and Walker, 2015). Existing work has examined whether ICLV models are more behaviorally sound and offer better predictions, but the spatial transferability of ICLV models has not been fully explored. Does the increased behavioral realism and predictability extend beyond the original spatial context to other areas? We hypothesize that well-specified ICLV models – particularly, ones that better capture the observed/unobserved heterogeneity in the data – will perform better (in terms of spatial transferability) than traditionally estimated choice models. We speculate that this will be due to how ICLV models produce non-linearity in the impact of exogenous variables on the choice outcome. This will allow the model to guide itself to a reduced form choice model specification that may have been unconsidered by the analyst (when developing from a mixed logit base only) and likely better fitted. If this is the case, then it will be beneficial for regions that generally lack detailed attitudinal/perception data to borrow ICLV model from regions which have access to such detailed attitudinal/perception data.

This paper focuses on assessing the impact of incorporating attitudinal and perception variables on the spatial transferability of travel forecasting models. Specifically, this paper compares the spatial transferability, in an empirical setting, for three model structures: 1) multinomial logit (MNL), 2) mixed multinomial logit (MMNL), and 3) integrated choice and latent variable (ICLV) models. Transferability is assessed using both an application-based assessment approach as well as a joint context estimation approach (see Sikder et al. (2013) for details). The data utilized for assessing the spatial transferability in the three contexts comes from a survey conducted among 1091 respondents in the states of Florida and Michigan in the United States (591 observations from Florida and 500 from Michigan). In the survey, respondents were asked about their preferred way of using autonomous vehicles (when available), along with personal and household characteristics, current travel characteristics, and perceptions about benefits and concerns related to autonomous vehicles (AVs). The available data is used to model an individual’s intention for using AVs using three model structures and then spatial transferability assessment is done using various assessment metrics.

2. Transferability Assessment The assessment of spatial transferability of models between geographic regions can be done using various techniques that can broadly be classified into the (1) application based approach and (2) estimation based approach (Bowman et al., 2014). In the application based approach, the model is estimated using the data from one region (base context) and applied to data in another region (application context) to assess how well the model in the base context performs in the application context. The application based approach to assess transferability is believed to be statistically rigorous and generally results in the rejection of spatial transferability. This is because this approach tests the transferability of the model as a whole, without allowing for an investigation of specific parameter estimates. Further, as the application based approach involves dividing the sample into geography-specific datasets, this approach can easily confound the transferability assessment results, if small samples are used. The issues associated with the application based approach can be alleviated by using estimation based approach (also called joint context estimation), where the data from the base and application contexts are combined to estimate a single model. Potential differences between the two contexts are recognized by estimating different parameters. In this estimation based approach, a base model is first estimated by combining data from both base and application contexts and then a context specific dummy variable is interacted, one by one, with each of the variables in the base model to form the difference variable. The difference variables helps in assessing whether the corresponding variables are transferable between the base and application contexts. A particular advantage of this approach is that one can test whether each parameter in the model is transferable or not (Bradley and Daly, 1997; Karasmaa, 2007; Sikder et al., 2014). After the models have been estimated, various measures of transferability assessment such as transferability test statistic (TTS), transfer index (TI), root mean square error (RMSE), and relative aggregate transfer error (RATE) are used (Sikder et al., 2013). These assessment metrics measure model fit and the predictive ability of the transferred models in the application context. Here, it must be noted that while the estimation of the ICLV model that is being transferred will make use of the available measurement indicators, these will not be used while applying the transferred model to the application context.

3. Present Study Transferability assessment for various model structures in this study is done using data collected from a survey conducted among 1091 respondents from the states of Florida and Michigan in the United States, who were asked about their preferred way of using autonomous vehicles (AVs), when available (see (Menon, 2015) for details of the data). The options to choose from included: (1) owning an AV for personal use; (2) sharing an AV or using it as a service (e.g. Uber/Lyft), and (3) not using AVs at all. Apart from the response on preferred way of using AVs, the survey collected information on personal and household demographics, information on current travel behavior, perceptions of various attributes of AV technology and opinion on familiarity and perception of the benefits and concerns with AVs. The collected data is first used to estimate separate models for the state of Florida and Michigan, using all three model structures (three model structures and two geographic regions for a total of 6 models). Then, combined models are estimated by pooling data from each geographic region (one model using each model structure for a total of 3 models). In the combined model estimation, the variables in the base models are interacted with context specific dummy variable, to estimate the difference variable. The estimated models are then used to calculate various transferability assessment metrics.

In terms of modeling, first, the choice – “preferred way of using AVs” is modeled using the multinomial logit model structure. The utility that an individual gains from choosing an alternative is expressed as a function of the individual’s personal and household characteristics. Wherever possible, the non-linearity in the effect of an exogenous variable on the choice utility is also considered. Second, a mixed multinomial logit model is estimated by allowing the parameter estimates corresponding to various exogenous variables to vary (following a normal distribution) across individuals. In the combined mixed logit models, heterogeneity in means and variances of random parameters across geographic contexts is also considered. This is done by expressing the means and variances as a function of context specific dummy variable. Third, an ICLV model is estimated with perceptions of benefits and concerns related to AV technology as latent variables. In the measurement equation of the ICLV model, the response to the observed attitudinal indicator variables, which capture the manifestation of the latent construct, is modeled on an ordinal scale.

To assess the spatial transferability of the models between geographic regions, we first calculate the transferability assessment metrics using the application based approach. The metrics are calculated to assess the transferability of models estimated in Florida to Michigan and vice versa. Then, a transferability assessment is also done using the estimation-based approach, where the models estimated using the pooled dataset are used to calculate the assessment metrics. The calculated metrics are then compared across model structures to investigate the impact of incorporating attitudinal/perception related variables on model transferability.

4. Expected Results From this study, it is anticipated that incorporating altitudinal/perception variables in the choice model will help in making considerable improvements to the spatial transferability of the model. Preliminary transferability assessment using the MNL model suggests poor transferability of model from Florida to Michigan and vice versa. This is not unexpected as mentioned in section 2. But, we expect that incorporating attitudinal/perception variables through ICLV model structure will improve transferability across regions. At the least, it is expected that transferred ICLV model will perform better than the locally estimated MNL model because its reduced form will be able to improve upon a mixed logit specification. If that is so, it will help regions which generally do not have an extensive dataset that includes attitudinal/perception related variables to estimate an advanced ICLV model. Such regions can borrow advanced models from other regions to achieve better prediction accuracy than they achieve from a locally estimated multinomial logit model.

5. References Atherton, T.J., Ben-Akiva, M.E., 1976. Transferability and updating of disaggregate travel demand models. Badoe, D.A., Miller, E.J., 1995. Comparison of alternative methods for updating disaggregate logit mode choice models. Transportation Research Record(1493), 90-100. Ben-Akiva, M., Bolduc, D., 1987. Approaches to model transferability and updating: the combined transfer estimator. Département d'économique, Université Laval. Ben-Akiva, M., Walker, J., Bernardino, A.T., Gopinath, D.A., Morikawa, T., Polydoropoulou, A., 2002. Integration of choice and latent variable models. Perpetual motion: Travel behaviour research opportunities and application challenges, 431-470. Bowman, J.L., Bradley, M., Castiglione, J., Yoder, S.L., 2014. Making advanced travel forecasting models affordable through model transferability, Preseted at the 93rd Annual Meeting of Transportation Research Board, Washington, DC. Bradley, M.A., Daly, A.J., 1997. Estimation of logit choice models using mixed stated preference and revealed preference information. Understanding travel behaviour in an era of change, 209-232. Burnett, K., 1981. Interspatial, intraspatial, and temporal transferability of travel-choice models. Spatial transferability of travel-demand models. In: new horizons in travel-behavior research, Presented at the Fourth International Conference on Behavioral Travel Modeling. Karasmaa, N., 2007. Evaluation of transfer methods for spatial travel demand models. Transportation Research Part A: Policy and Practice 41(5), 411-427. Koppelman, F.S., Wilmot, C.G., 1982. Transferability analysis of disaggregate choice models. Transportation Research Record 895, 18-24. Koppelman, F.S., Wilmot, C.G., 1986. The effect of omission of variables on choice model transferability. Transportation Research Part B: Methodological 20(3), 205-213. Menon, N., 2015. Consumer Perception and Anticipated Adoption of Autonomous Vehicle Technology: Results from Multi-Population Surveys. University of South Florida. San Santoso, D., Tsunokawa, K., 2005. Spatial transferability and updating analysis of mode choice models in developing countries. Transportation Planning and Technology 28(5), 341-358. Sikder, S., Augustin, B., Pinjari, A., Eluru, N., 2014. Spatial Transferability of Tour-Based Time-of-Day Choice Models: Empirical Assessment. Transportation Research Record: Journal of the Transportation Research Board(2429), 99-109. Sikder, S., Pinjari, A.R., Srinivasan, S., Nowrouzian, R., 2013. Spatial transferability of travel forecasting models: a review and synthesis. International Journal of Advances in Engineering Sciences and Applied Mathematics 5(2-3), 104-128. Vij, A. and Walker, J.L., 2016. How, when and why integrated choice and latent variable models are latently useful. Transportation Research Part B: Methodological, 90, pp.192-217. Yasmin, F., Morency, C., Roorda, M.J., 2015. Assessment of spatial transferability of an activity-based model, TASHA. Transportation Research Part A: Policy and Practice 78, 200-213.

Rico Krueger (The University of New South Wales, Australia)
Akshay Vij (Institute for Choice, University of South Australia, Australia)
Taha H. Rashidi (The University of New South Wales, Australia)
A Dirichlet Process Mixture Model of Discrete Choice with an Application to Transport Mode Choice Behaviour
SPEAKER: Rico Krueger

ABSTRACT. Motivation Flexible discrete choice models are the cornerstones of contemporary travel demand modelling methods. A key concern of discrete choice analysis is the representation of unobserved inter-individual taste variation. In that vein, the literature distinguishes two principal approaches to capture random taste heterogeneity in discrete choice models: Continuous mixture models are based on the assumption that individual taste parameters are drawn from a sample-level continuum of tastes. In finite mixture models, individuals are probabilistically assigned to a finite number of segments with homogeneous tastes. Discrete choice models with continuous and discrete mixing distributions are widely used in disciplines studying individual choice behaviour, but are subject to limitations. The distributional assumptions of continuous mixture models may be overly rigid and may not yield convincing representations of inter-individual random taste heterogeneity, as the shape of the estimated taste parameter distribution is constrained to be equal to the functional form of the imposed parametric random distribution. For valid inferences in discrete choice models with continuous mixing distributions, it is thus imperative to correctly specify the mixing distribution of randomised taste parameters (Hess et al., 2005). In practice however, the analyst is unlikely to be able to exhaust the hypothesis space of theoretically-feasible distribution functions (e.g. Keane and Wasi, 2013). Finite mixture approaches such as the latent class discrete choice model free the analyst from making rigid distributional assumptions (e.g. Greene and Hensher, 2003), but are cumbersome to estimate, as the number of mixture components must be exogenously determined based on post-hoc model selection criteria. In response to the limitations of the discrete choice models with continuous and discrete mixing distributions, further semi-nonparametric and nonparametric extensions of the logit model have been proposed (also see Vij and Krueger, 2017; Yuan et al., 2015, for reviews). Semi-nonparametric approaches are conceptually appealing, as unbounded continuous distributions can be closely approximated by a finite mixture of multivariate Gaussians. However, semi-nonparametric approaches are computationally demanding, as the analyst must perform post- hoc model selection to determine the appropriate number of mixture components. In addition, increasing the modality of the mixture distribution incurs the estimation of an additional set of mean and variance-covariance parameters, which may render models with high-dimensional semi-nonparametric mixing distributions intractable. The nonparametric representation of heterogeneity distributions can be achieved with the help of nonparametric discrete mixing distributions. The latent class choice model is the simplest implementation of a nonparametric discrete mix- ing distribution with a finite number of support points, whose locations and associated probability mass need to be estimated. In practice, high-dimensional latent class choice models are often plagued by identification issues due to the multi-modality of the log-likelihood function. Therefore, a stream of literature proposes nonparametric discrete mixing distributions with structured support points (see Dong and Koppelman, 2014; Train, 2008; Vij and Krueger, 2017). Effectively, these approaches implement multi-variate histogram estimators by defining a multi-dimensional grid on the coefficients space such that the het- erogeneity distribution in question can be closely approximated by estimating the amount of probability mass positioned on the vertices of the grid. Logit models with gridded mixing distributions can capture complex heterogeneity distributions (see e.g. Vij and Krueger, 2017), but require the analyst to specify the complexity of the grid of mass points a priori. The defining characteristic of infinite-dimensional nonparametric models is that the model complexity, i.e. the size of the parameter space, is endogenised (e.g. Gelman et al., 2013). Infinite-dimensional models are known as Bayesian nonparametric models, because infinite-dimensional models are Bayesian models, which are defined on an infinite-dimensional parameter space and the use of stochastic process priors results in only a subset of the parameters being employed to explain the data at hand (Orbanz and Teh, 2011). In other words, model complexity is incorporated into the posterior density and estimated conditional on the observed data (Gershman and Blei, 2012).

Objective and contribution The objective of this research is to leverage Bayesian nonparametric methods to capture unobserved taste heterogeneity in a logit-based discrete choice model. Specifically, we propose a Bayesian nonparametric discrete choice model, which leverages the Dirichlet process prior (Ferguson, 1973) as a flexible mixing distribu- tion in a mixed logit model. The proposed model consists of a high-dimensional discrete mixture of logit choice kernels and employs the truncated stick-breaking process representation (Ishwaran and James, 2001) of a Dirichlet process prior to endogenously partition the sample into a finite number of segments with homogeneous tastes. To operationalise the model framework, we derive an expectation maximisation algorithm, which allows for efficient model inference. Finally, we validate the model framework in a simulation study and a case study on route choice behaviour. Our research is related to the works by Kim et al. (2004), Burda et al. (2008) and Li and Ansari (2013): Kim et al. (2004) develop an infinite mixture model of discrete choice, where a Dirichlet process prior is used to infer the number of mixture components required to characterise the unobserved taste heterogeneity distribution of a sample. The work by Burda et al. (2008) work is conceptually similar to the one by Kim et al. (2004), but additionally develops a mixed logit-probit choice kernel to relax the irrelevance of independent alternatives restriction of the multinomial logit model. Both Burda et al. (2008) and Kim et al. (2004) adopt general Markov Chain Monte Carlo sampling algorithms for Dirichlet process mixture models (Neal, 2000). Li’s and Ansari’s (2013) discrete choice model accounts for both heterogeneity and endogeneity. The authors employ a truncated stick-breaking process representation of the Dirichlet process (Ishwaran and James, 2001) and devise a Markov Chain Monte Carlo sampling scheme for model inference. Kim et al. (2004), Burda et al. (2008) as well as Li and Ansari (2013) present applications in the realm of consumer behaviour to validate their model frameworks.

Model validation To validate the proposed model framework, we develop an extensive simulation study and present a case study on transport mode choice behaviour. For the simulation study, we generate multiple synthetic samples of individuals, who are pseudo-observed to make transport mode choices, whereby the alternatives are characterised by the level of services attributes in-vehicle travel time, out-of- vehicle travel time. Figure 1 displays the true and the estimated taste parameter distributions for one of the Monte Carlo experiments, where the true taste parameters are drawn from a two-component mixture of normal distributions. Overall, the simulation study shows that the proposed choice model performs well at capturing differently-shaped taste parameter distributions. By design, the proposed model yields a discrete representation of heterogeneity by allowing for the estimation of the locations and weights of K probability mass points, where K is the truncation level of the truncated stick-breaking process approximation of the Dirichlet process. Yet, the agglomerative clustering properties of the Dirichlet process result in only few probability mass points with non-zero weights. To empirically validate the proposed model framework, we source data from from the Bay Area Travel Survey (BATS) 2000, a regional household travel survey of the San Francisco Bay Area. The transport mode choice data are used to estimate the proposed nonparametric discrete choice model. In addition, a multinomial logit model with fixed coefficients and a mixed multinomial logit models with common continuous mixing distributions are estimated to benchmark the performance of the nonparametric discrete choice model in terms of data fit and predictive ability. To this end, we train the models on 70% of the data and retain the remaining 30% of the data as a hold-out. Table 1 gives the log-likelihood values of the training and hold-out samples for each of the considered models. The empirical application demonstrates that the proposed discrete choice model provides a superior fit to the training and the validation samples in the considered empirical application. The superior performance of the proposed discrete choice model can be attributed to two factors. First, the tail behaviour of the nonparametric mixing distribution prior allows for a flexible representation of taste segments with extreme attribute valuations. Second, the proposed discrete choice model accounts for correlations between random taste parameters.

Conclusion We present an approach for unsupervised taste segmentation by developing a Dirichlet process mixture model of discrete choice. The proposed model leverages a truncated stick-breaking process approximation of the Dirichlet process as a flexible nonparametric mixing distribution for the taste parameters in a multinomial logit model. The mixing distribution endogenously partitions the sample into a theoretically infinite number of segments with homogeneous tastes. However, due to the agglomerative clustering properties of the Dirichlet process, the proposed model can adapt its complexity to the evidence and characterises heterogeneity distributions parsimoniously. The proposed model relies on less restrictive assumptions than mixed logit models with continuous mixing distributions and is easier to estimate than a latent class choice model. In a case study on transport mode choice behaviour, we demonstrate that the proposed model framework outperforms established modelling approaches in terms of in-sample fit and out-of-sample predictive ability. Future work will develop a detailed analysis of the implications of the nonparametric heterogeneity representation of the proposed model on welfare measures such as travel time valuations.

References Burda, M., Harding, M., and Hausman, J. (2008). A Bayesian mixed logit–probit model for multinomial choice. Journal of Econometrics, 147(2):232–246. Dong, X. and Koppelman, F. S. (2014). Comparison of continuous and dis- crete representations of unobserved heterogeneity in logit models. Journal of Marketing Analytics, 2(1):43–58. Ferguson, T. S. (1973). A Bayesian Analysis of Some Nonparametric Problems. The Annals of Statistics, 1(2):209–230. Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., and Rubin, D. B. (2013). Bayesian Data Analysis, Third Edition. CRC Press. Gershman, S. J. and Blei, D. M. (2012). A tutorial on Bayesian nonparametric models. Journal of Mathematical Psychology, 56(1):1–12. Greene, W. H. and Hensher, D. A. (2003). A latent class model for discrete choice analysis: contrasts with mixed logit. Transportation Research Part B: Methodological, 37(8):681–698. Hess, S., Bierlaire, M., and Polak, J. W. (2005). Estimation of value of travel-time savings using mixed logit models. Transportation Research Part A: Policy and Practice, 39(2):221–236. Ishwaran, H. and James, L. F. (2001). Gibbs Sampling Methods for Stick- Breaking Priors. Journal of the American Statistical Association, 96(453):161– 173. Keane, M. and Wasi, N. (2013). Comparing Alternative Models of Heterogeneity in Consumer Choice Behavior. Journal of Applied Econometrics, 28(6):1018– 1045. Kim, J. G., Menzefricke, U., and Feinberg, F. M. (2004). Assessing Heterogeneity in Discrete Choice Models Using a Dirichlet Process Prior. Review of Marketing Science, 2(1). Li, Y. and Ansari, A. (2013). A Bayesian Semiparametric Approach for Endo- geneity and Heterogeneity in Choice Models. Management Science, 60(5):1161– 1179. Neal, R. M. (2000). Markov Chain Sampling Methods for Dirichlet Process Mixture Models. Journal of Computational and Graphical Statistics, 9(2):249– 265. Orbanz, P. and Teh, Y. W. (2011). Bayesian Nonparametric Models. In Sammut, C. and Webb, G. I., editors, Encyclopedia of Machine Learning, pages 81–89. Springer US. DOI: 10.1007/978-0-387-30164-8_66. Train, K. E. (2008). EM Algorithms for nonparametric estimation of mixing distributions. Journal of Choice Modelling, 1(1):40–69. Vij, A. and Krueger, R. (2017). Random taste heterogeneity in discrete choice models: Multivariate nonparametric finite mixture distributions. Yuan, Y., You, W., and Boyle, K. J. (2015). A guide to heterogeneity features captured by parametric and nonparametric mixing distributions for the mixed logit model. In 2015 AAEA & WAEA Joint Annual Meeting, July 26-28, San Francisco, California.

Naznin Sultana Daisy (Dalhousie University, Canada)
Hugh Millward (Saint Mary’s University, Canada)
Lei Liu (Dalhousie University, Canada)
Modeling Activity-Travel Behavior of Non-Workers Grouped by their Daily Activity Patterns

ABSTRACT. This study aims to better understand and model the activity-travel behavior of non-workers through utilizing a cluster-based Multivariate Probit (C-MVP) model along with the estimation of transition matrices of activity episodes. Five non-worker clusters with homogeneous daily activity patterns are identified, and a series of C-MVP models are estimated for each cluster. Most earlier studies assumed activity participation to be an independent phenomenon in multivariate modeling, resulting in either logit or mixed logit models. However, activity participation in one activity is associated with both the previous and next activity. This interdependency between activities can be captured with a correlation matrix in the C-MVP model. Additionally, transition matrices were estimated to understand the correlation among in-home and out-of-home activities.

11:00-12:30 Session 10C: Healthy, Happy, and Holistic Living -- Electrics
Jae Hyun Lee (PH&EV Center, Institute of Transportation Studies, University of California, Davis, United States)
Location: Corwin East
Javier Bas (University of Maryland, United States)
Elisabetta Cherchi (Newcastle University, UK)
Cinzia Cirillo (University of Maryland, United States)
Anders F. Jensen (Technical University of Denmark, Denmark)
Predicting the Diffusion of EV: A Dynamic Approach to Model the Impact of Imitation and Experience
SPEAKER: Javier Bas

ABSTRACT. Driven by environmental awareness and new regulations for fuel efficiency, electric vehicles (EVs) have significantly evolved in the last decade, though its market share is still much lower than expected. Besides understanding the reasons for this slow market penetration, it is crucial to have appropriate models to predict the right diffusion of this innovative products. Recent works in predicting the EV market combined substitution with diffusion models, where advanced discrete choice models are used to measure the substitution effect among alternative vehicle’s engine and a Bass-type model is used to account for the diffusion effect of innovation. However, the most recent substitution/diffusion models (Jensen et al. 2017) are not explicitly dynamic and do not measure the fact that innovation is communicated through certain channels over time among members of a social system (Rogers 2010).

In this paper, we extend this substitution/diffusion models by including explicitly the dynamic effect, by making the EV demand in time t dependent on the EV sold in time t-1, and by measuring explicitly some of the effects of social conformity on the individual choices. The model includes also the impact of policy incentives, in particular in the availability of parking spaces and parking cost strategies.

Data used in the paper comes from different sources. The coefficients for the disaggregate substitution model come from a model estimated using data collected in Denmark (Cherchi, 2017). Data on Market Shares are based on the revealed demand for EV registered in Denmark, covering the period 2013-2017. Results show an initial slow penetration of the EV in the market, that progressively increases in the 2050 horizon.



Bass FM (1969). A new product growth for model consumer durables. Management Sci. 15(5):215–227.

Cherchi E (2017). A stated choice experiment to measure the effect of informational and normative conformity in the preference for electric vehicles. Transportation Research Part A: Policy and Practice 100, 88-104.

Jensen, A., Cherchi, E., Mabit, S. and Ortúzar, J. de D. (2017) Predicting the potential market of electric vehicles. Transportation Science 51, 427-440.

Rogers, E. M. (2010). Diffusion of innovations. Simon and Schuster.

Scott Hardman (University of California, Davis, United States)
Jae Hyun Lee (University of California, Davis, United States)
Gil Tal (University of Califonria Davis, United States)
Investigating the users of PEVs 2010-2017: Are we moving beyond early adopters?
SPEAKER: Scott Hardman

ABSTRACT. Plug-in electric vehicles (PEVs) are attempting to achieve successful market entry for the third time. PEVs first entered the market in the late 1800s and early 1900s. For a while the vehicles were preferred to internal combustion engine vehicles (ICEVs) due to their ease of operation. However, as ICEV technology advanced, partly due to the advent of the electric starter motor, PEVs lost market share. Between the early 1900s and the 1990s PEVs were used only in niche market, for example as milk delivery vehicles in the United Kingdom. In the 1990s PEVs emerged again, with vehicles produced by GM, Honda, Ford, Nissan, Toyota, and smaller OEMs. These vehicles were sold in the 1990s and early 2000s however again did not achieve successful market entry for a variety of reasons. Since 2010 PEVs have been attempting market entry for the third time. The vehicles now benefit from more advanced battery technology, longer driving ranges, and better charging infrastructure. Since their introduction in 2010 over 2 million PEVs have been sold to consumers worldwide (International Energy Agency, 2017; Lutsey, 2017). Though this is a remarkable achievement PEVs still only represent a small percentage of the global vehicle market. For PEVs to achieve successful market entry they will need to be adopted by more consumers.

To date the users of electric vehicles have been a homogenous group of consumers. Studies have found that PEV users are mostly high income, have a high level of education, live in multi-car households, live in suburban locations, and be mostly male (Egbue and Long, 2012; Jakobsson et al., 2016; Lane et al., 2014; Plötz and Gnann, 2011). They have also been found to a high level of empathy, be willing to accept change, and have positive attitudes to technology in general (Hardman et al., 2016). These characteristics suggest that PEV users are innovators or early adopters based on the assumptions in Diffusion of Innovation theory (Rogers, 2003). To date only one study has identified any different in the early adopters of BEVs (Hardman et al., 2016). This study found that adopters of high-end BEVs (e.g Tesla BEVs) were of higher income and higher education than the adopters of low-end BEVs (e.g Nissan Leaf). The high-end adopters were more representative of typical early adopters. However, both groups of BEV adopters were still considered to be early adopters.

Rogers theory includes several groups of adopters. These are innovators, early adopters, early majority, late majority, and laggards. Innovators and early adopters are critical to the market introduction of any new technology as they are always the first consumers to use a product. However, for any technology to achieve mass market entry they need to be adopted by new consumers. For PEVs to achieve successful market entry they need to be used by all consumers, including the early majority, late majority, and eventually laggards. This study will investigate whether the transition to PEVs in California is moving beyond just early adopters. To do this this study will use questionnaire survey data gathered by the Plug-in Hybrid & Electric Vehicle Research Center 2010-2017. Since 2010 the center has administered a questionnaire survey to new car users in California. This survey has used the same methods in 2010, 2011, 2012, 2013, 2014, 2015, 2016, and 2017. Respondents were recruited via sending emails to PEV users who have applied for the California clean vehicle rebate. The survey has yielded more than 3,000 respondents on average per year meaning the data set contains 24,000 responses from consumers who have purchased a PEV.

The survey collects several types of data. First it gathered household sociodemographic information. This includes data on income, level of education, gender, household type, household location, number of vehicles in the household and information about these vehicles, number of people in the household, and number of people who can drive in the household. The survey also measures consumer purchase motivations, including the impact of several incentives on the decision to adopt a PEV. This includes the California rebate, US federal tax credit, and high occupancy vehicle lane access. The survey also gathers in formation on the households PEV including the vehicle purchased and data of purchase.

This data will be analyzed using a variety of statistical methodologies including ANOVA, regression analysis, factor analysis, cluster analysis, and time series analysis. These methods will be used to investigate changes in the socio-economic profile of PEV adopters over time. ANOVA will be used to understand differences in the importance of PEV incentives between different consumer groups, different vehicle makes, and different locations. This will help to identify for which types of people the incentive is less or more important. ANOVA can also be used to investigate whether the importance of incentives differs each year. Cluster analysis can be used to understand whether the current group of PEV adopters are heterogeneous. It may emerge that within the current group of PEV users different sub groups exist. It will therefore be important to track how these have been charging over time. Regression will be used to understand how the socio-economic profile of PEV adopters has changed over time from 2011-2018. This will show the transition to PEVs is moving towards the early majority, or whether more early adopters are buying the vehicles. This analysis will be broken down by different regions and different vehicle makes and models. It can therefore show how the transition to PEVs is changing for each vehicle type, each region, and by different income groups (in addition to other sociodemographic attributes). The data can also be used to explore how the importance of purchase incentives (tax credits, state rebate) are changing over time.

The results from the study will be interesting to diffusion scholars and transportation researchers who are interested in the transition to PEVs and interested in understanding who the users of new transportation technologies are. The paper will answer an important question regarding the transition to PEVs. It will show whether PEVs are being used by more consumers with the same socio-economic profile or whether PEV users with different socio-economic profiles are purchasing the vehicles. The results showing how the importance of purchase incentives has been changing over time will be interesting for policy makers. These incentives have so far been significant factors in encouraging consumers to purchase PEVs (Aasness and Odeck, 2015; Hardman et al., 2017; Hardman and Tal, 2016; Mallette and Venkataramanan, 2010). The incentives will soon begin to phase out, understanding what impact this will have could inform policy decision.

Aasness, M.A., Odeck, J., 2015. The increase of electric vehicle usage in Norway incentives and adverse effects. Eur. Transp. Res. Rev. 7. doi:10.1007/s12544-015-0182-4 Egbue, O., Long, S., 2012. Barriers to widespread adoption of electric vehicles: An analysis of consumer attitudes and perceptions. Energy Policy. doi:10.1016/j.enpol.2012.06.009 Hardman, S., Chandan, A., Tal, G., Turrentine, T., 2017. The Effectiveness of Financial Purchase Incentives for Battery Electric Vehicles - A Review of the Evidence (Article Under Review). Renew. Sustain. Energy Rev. Hardman, S., Shiu, E., Steinberger-Wilckens, R., 2016. Comparing High-End and Low-End Early Adopters of Battery Electric Vehicles. Transp. Res. Part A Policy Pract. 88, 40–57. doi:10.1016/j.tra.2016.03.010 Hardman, S., Tal, G., 2016. Exploring the decision to adopt a high-end battery electric vehicle: The role of financial and non-financial motivations. Transp. Res. Rec. J. Transp. Res. Board 16–1783. International Energy Agency, 2017. Global EV Outlook 2017 Two million and counting. Jakobsson, N., Gnann, T., Plötz, P., Sprei, F., Karlsson, S., 2016. Are multi-car households better suited for battery electric vehicles ? – Driving patterns and economics in Sweden and Germany. Transp. Res. Part C 65, 1–15. doi:10.1016/j.trc.2016.01.018 Lane, B., Sherman, C., Sperl, J., Krause, R., Carley, S., Graham, J., 2014. Beyond Early Adopters of Plug-in Electric Vehicles? Evidence from Fleet and Household Users in Indianapolis. Transportation Research Board 2014 Annual Meeting. Lutsey, N., 2017. The rise of electric vehicles: The second million [WWW Document]. ICCT - Int. Counc. Clean Transp. URL Mallette, M., Venkataramanan, G., 2010. Financial incentives to encourage demand response participation by plug-in hybrid electric vehicle owners. 2010 IEEE Energy Convers. Congr. Expo. ECCE 2010 - Proc. 4278–4284. doi:10.1109/ECCE.2010.5618472 Plötz, P., Gnann, T., 2011. Who should buy electric vehicles ? – The potential early adopter from an economical perspective. ECEEE 1073–1080. Rogers, E.M., 2003. Diffusion of Innovations, 5th Editio. ed. Free Press, New York.

Sadaf Aminmansour (The University of Queensland, Australia)
Zili Li (The University of Queensland, Australia)
Simon Washington (The University of Queensland, Australia)
Jake Whitehead (The University of Queensland, Australia)
Carlo Prato (The University of Queensland, Australia)
Zuduo Zheng (The Queensland University of Technology, Australia)
The Impact of New Mobility Technologies and Government Polices on Vehicle Willingness to Purchase in Australia
SPEAKER: Carlo Prato

ABSTRACT. Development of new vehicle technologies, such as alternative fuel vehicles (AFVs) and autonomous vehicles (AVs), if implemented well, may enable transport system operators to address some of the current mobility challenges such as crash risk, congestion, and vehicle pollution. Several studies have investigated factors influencing people’s intention to purchase AFVs or AVs; however, little research has focused on understanding the relationship between people’s decision to keep or sell an existing vehicle whilst simultaneously considering the purchase of an AV or AFV. This study considers the simultaneity between buying ‘new’ and selling ‘current’ vehicles and aims at explaining and forecasting future vehicle technology demand. Moreover, government policies have been playing a major role in the adoption of ‘new’ technologies. Thus, a better understanding of decision-making for future vehicle technologies will account for existing vehicle holdings and potential government policies to encourage uptake. Using data from a stated preference survey administered in Australia, this paper examines the simultaneity between buying and selling, and the effects of government policies on the propensity to purchase new technologies. The stated preference data were modelled using simultaneous binary probit models, and reveal that Australians are more likely to purchase an AFV or an AV compared to a conventional vehicle, and are more likely to buy a new car if they simultaneously agree to sell an existing vehicle or two-wheeler. The two most effective policies for incentivizing the purchase of AFVs were an upfront rebate/cash subsidy and discounts on road tolls.

Jae Hyun Lee (PH&EV Center, Institute of Transportation Studies, University of California, Davis, United States)
Gil Tal (PH&EV Center, Institute of Transportation Studies, University of California, Davis, United States)
Wei Ji (PH&EV Center, Institute of Transportation Studies, University of California, Davis, United States)
Alan Jenn (PH&EV Center, Institute of Transportation Studies, University of California, Davis, United States)
Exploring neighborhood effects on PEV market penetration in California (2010-2016)
SPEAKER: Jae Hyun Lee

ABSTRACT. 1. Introduction With the introduction of a variety of new Plug-in electric vehicles (PEVs), including battery electric vehicles (BEVs, such as Nissan Leaf, Tesla, etc.) and plug-in hybrid electric vehicles (PHEVs, e.g. Chevrolet Volt, etc.), the market share of PEVs in the US has been rapidly increasing year-to-year. The U.S. National Research Council predicts that approximately 13 million Electric vehicles will be on U.S. roads by 2030, which will account for 4% of the US fleet in 2030 (with the most probable scenario). California is the largest PEVs market in the U.S., accounting for about 48% of cumulative PEV sales in the U.S. between 2011 and 2016. About 0.2 million PEVs have been sold between 2010 and 2016 in California, and its market share in new vehicles sold in California is now 4.8% in 2017 (California Auto Outlook, 2017). However, the State of California wants to accelerate the penetration of PEVs and fuel cell vehicles to 1.5 million vehicles by 2025 (Governor's Interagency Working Group, 2013) and reach 15% of new vehicles sold in California by 2025 (California Air Resources Board, 2014). In order to develop policy tools to accelerate the market penetration of PEV, it is important to better understand/identify the factors relating to individuals’ adoption of PEVs. The most commonly found factors were PEV buyers’ socio-demographic variables such as age, gender, and level of education, household size and income (Carley, Krause, Lane, & Graham, 2013; Egbue & Long, 2012; Higgins, Paevere, Gardner, & Quezada, 2012; Khan & Kockelman, 2012; Musti & Kockelman, 2011; Sierzchula, Bakker, Maat, & van Wee, 2014; Tal & Nicholas, 2013; Türnau, 2015). The PEV buyers’ travel patterns (including mobility) were also revealed as important factors and were measured in various ways, for example travel distance (Egbue & Long, 2012; Tamor, Gearhart, & Soto, 2013), driving distance (Higgins et al., 2012; Krupa et al., 2014), and trip length and duration (Khan & Kockelman, 2012). The adoption of PEVs was also affected by household fleet composition and vehicle characteristics (number of vehicles in households, Musti & Kockelman, 2011; vehicle type, price and usage, Lieven et al., 2011; Sierzchula et al., 2014; Tal & Nicholas, 2013). Moreover, the existing PEV infrastructure affects PEV purchasing behavior because it helps to increase buyers’ awareness of PEVs, and to potentially increase their willingness to buy PEVs (Javid & Nejat, 2017; Sierzchula et al., 2014). In addition, various attitude variables were also identified as important factors to adopt PEVs including pleasure of driving (hedonic attributes), identity derived from owning and using EVs (symbolic attributes) and pro-environmental identity (Schuitema, Anable, Skippon, & Kinnear, 2013) An individual’s frequent exposure to PEVs can also directly increase awareness of PEVs and increase the probability that they will purchase a PEV in the future. In other words, the number of PEVs in an individual’s surrounding area can be a predictor of PEV purchase behavior and PEV market penetration growth models. Moreover, without taking this effect into account, the factors influencing PEV adoption behavior cannot be accurately identified because of the existence of the errors from spatial autocorrelation (Chen, Wang, & Kockelman, 2015). This effect has been well recognized in transportation research, called “neighborhood effects” or “effect of spatial interaction” (Sidharthan, Bhat, Pendyala, & Goulias, 2011; Wang & Chen, 2015; Zegras, 2010), and its theoretical foundation can be found in Tobler’s “First Law of Geography: everything is related to everything else, but near things are more related than distant things”(Tobler, 1970). Particularly in PEV sales studies, this is also related to the adoption and diffusion of innovation; imitators adopt new technology based on their surrounding innovators usages of the technology (Bass, 1969; Rogers, 1995). Therefore, the PEV neighborhood effects can be defined here as the existence of correlating PEV purchase behavior among geographically-close buyers. These effects were recently captured in PEV market penetration modeling with zone-level PEV sales data (Chen et al., 2015). They developed a series of multivariate Poisson-lognormal conditional autoregressive (CAR) models and tested if spatially-correlated latent variables capture neighborhood effects in zone-level PEV sales. They found that neighborhood effects do exist in PEV market penetration models, indicating that nearby households Are more likely to own a PEV than those farther away. In this paper, we expand this research by exhaustively examining the neighborhood effects in PEV market penetration with 7 years (2010-2016) in California. The first objective of this paper is to understand the existence and boundaries of neighborhood effects in PEV market penetration in the largest PEV market in the U.S. By using different distances to define neighbors and testing their effects in PEV market penetration models, it is possible to find the boundary of neighborhood effects in PEV market penetration in California. The second objective is to examine year-to-year changes in neighborhood effects with both cross-sectional spatial models and longitudinal (panel) models. As electric vehicles become more prevalent, it is possible to observe stronger neighborhood effects. The last objective of this paper is to explore spatial heterogeneity of neighborhood effects. The strength of neighborhood effects varies across the state of California: they can be negligible in areas with low level of PEV market penetration. Therefore, we will develop models to test neighborhood effects that take into account the spatial heterogeneity of PEV market penetration. 2. Data Almost 200,000 Plug-in vehicles were sold in California between 2010 and 2016. About 180,000 applied for the state rebate and have their data included in our study. We use a subsample of 172,880 privately owned vehicles categorized by vehicle type and census tract location. Overall we have 7,341 tracts with PEVs (out of a total of 8,057 in the state) or 91 percent of the tracts, representing a population of nearly 33.7 million. PEV sales data by 2016 based on the Clean Vehicle Rebate Program (CVRP) records was aggregated to census tract level and used to explore neighborhood effects in PEV market penetration. Socioeconomic data including median income and employment ratio of each census tract were derived from California household travel survey (CHTS) data (CalTrans, 2013). The weighted average commute distance of residents from each census tract and the ratio of the length of available HOV lanes on the shortest commute route over the total commute distance (HOV share) were calculated based on LODES data (US Census Bureau, 2013). 3. Methods and preliminary results Figure 1 shows the spatial distributions of PEV penetration rate between 2010 and 2016. The market penetration rates are computed annually, and cumulative PEV sales per zone are divided by market limit per zone which is defined based on the ability to charge the car at home. The penetration rate was less than 1% in every census tract in California until 2012. A surge of PEV penetration can be first observed in 2013 in San Francisco area (2-4%), and it continuously increased until 2016. In addition, the areas with higher than 5% of penetration rate are more geographically expanded year-to-year. For example, most census tracts in SF had higher than 5% market penetration in 2016. In greater LA area urban areas with higher income generally have higher PEV penetration. Additionally, the suburban areas in the LA region have generally higher rates than areas in the center of LA; Orange County and the city of Irvine in particular have a relatively higher penetration rate than other areas. In order to examine neighborhood effects, we first computed Moran’s I of PEV sales in each year (Figure 2). The Moran’s I shows the similarity between each zone’s value (PEV sales in each Census tract) and the mean of its neighbors’ values, with the neighbors defined based on the distance between zones (Moran, 1950). Therefore, this Moran’s I illustrates degrees of neighborhood effects (spatial dependency) on PEV sales based on proximity. In 2010, neighborhood effects were not found because of low level of PEV sales (102 PEVs). From 2011, it is possible to observe the neighborhood effects in annual PEV sales, and it has been stronger year-to-year until 2014. From 2011 to 2012, the effects sharply decreased between 2km and 6km ranges (from 0.4 to 0.25), and then it decreased slowly until 80km. In 2013, the neighborhood effects were only strengthened overall, and then stable for four consecutive years (2013-2016). However, this effect has to be also examined with spatial models to test their impact on PEV market penetration with other important factors such as socio-demographic variables, travel patterns and infrastructure. We will use three different models to test the neighborhood effect with the above explanatory variables: 1) spatial lag models for each year’s PEV sales, 2) Latent class spatial lag model (Lee, Davis, McBride, & Goulias, 2016; Vermunt & Magidson, 2015) and 3) spatial lag panel model (Time-space recursive model, Anselin, Le Gallo, & Jayet, 2008). The first model will show the year-to-year differences in neighborhood effects in PEV sales and the second model will show spatially heterogeneous neighborhood effects in California. The last model will show the existence of overall neighborhood effects in 4 recent years. PEV sales in 2013-2016 were used as dependent variables individually (model 1 and 2 – cross-sectional model) and together (model 3 – panel model). Both spatial and time lag variables are included in these models because the PEV sales at each zone and its neighboring zones in the previous years can capture neighborhood effects (These variables are generally used in spatial models to control spatial-temporal autocorrelation but we interpret these as the existence of neighborhood effects). As a preliminary result (from the first model), we found significant neighborhood effects within 16 km of the neighborhood boundary in all four spatial lag models (PEV sales 2013-2016) based on Lagrange multiplier tests (Anselin, 1988; Anselin & Bera, 1998; Anselin & Rey, 1991). The employment ratio and median income were positively correlated to the PEV market penetration although employment ratio was not significant in 2013. On the other hand, negative coefficient was found at average commute distance. Lastly, the mean HOV shares were significantly positively correlated with PEV market penetration in 2013 and 2015. In the simple spatial panel model, all time lag variables have significant coefficients (t-1, squared t-1, t-2, squared t-2) as well as all other explanatory variables, and their directional effects were the same as the cross-sectional models. In order to examine spatial heterogeneity of PEV sales, we use Latent class cluster analysis. As a result, we found 4 clusters showing different growth patterns of the PEV market penetration (Figure 2). The cluster 4 (red line) shows the fastest growth pattern, but the slowest speed of market growth was found in cluster 1 (blue line). Cluster 3 (yellow line) shows slightly higher speed of PEV market growth compared to the average (gray dot line), and cluster 3 (green) shows relatively slow speed of growth. Geographically, the census tracts belonging to cluster 1 were mostly found in the SF Bay area and LA area. SF seems to have a larger PEV market than LA because the center of LA area shows slow speed of PEV market growth. The relatively higher growth pattern (cluster 2, yellow) was mostly observed in surrounding areas of SF and LA; the largest areas of this cluster were found on the north side of the Bay area (Sonoma county area). Clusters 3 and 4 were mostly found in rural areas, and cluster 3 was mostly distributed surrounding areas of cluster 1 and 2. Overall, the preliminary results show significant neighborhood effects on PEV market penetration and spatially heterogeneous patterns of market penetration. These findings are encouraging and highlight the need to further investigate spatially and temporally heterogeneous neighborhood effects on PEV market penetration. By the time of the IATBR 2018 conference we will have models capturing spatial heterogeneity of the effects using Latent class regression models and the spatial panel models to understand overall neighborhood effects in recent 4 years in California.

11:00-12:30 Session 10D: Long Distance Travel
Lisa Aultman-Hall (University of Vermont, United States)
Location: UCEN SB Harbor
Maxim Janzen (ETH Zurich, Switzerland)
Kay Axhausen (ETH Zurich, Switzerland)
Simulating Continuous Long-Distance Travel Demand
SPEAKER: Maxim Janzen

ABSTRACT. We propose a new agent-based simulation that is able to simulate long-distance travel demand for a long period of time. We will introduce the simulation, explain the main challenges and provide implementation details. Finally, we will use the simulation to simulate a  toy scenario, with a realistic, synthesized population.

Hannah Ullman (University of Vermont, United States)
Lisa Aultman-Hall (University of Vermont, United States)
Access to Long-Distance Travel as a Factor in Well-being

ABSTRACT. Abstract uploaded as pdf.

Kazuo Nishii (University of Marketing and Distribution Sciences, Japan)
Hideki Furuya (Toyo University, Japan)
Kuniaki Sasaki (University of Yamanashi, Japan)
Identification of latent classes determining expressway users' decision on their stopover behaviors by a model of market segmentation and location choice
SPEAKER: Kazuo Nishii

ABSTRACT. Tourism and leisure vehicle drivers need not only services accessing to high-speed roadways and safety but also other types of services at stopping places on expressway to ensure their comfortably traveling to tourism destination. In Japan, Nippon Expressway Corporation companies (NEXCOs) have attached increasingly great importance to the improvement of their providing those services at the Service Areas and Parking Areas (denoted SA/PAs) on expressway since privatization of Japan Highway Public Corporation (JH) in 2005. The SA/PAs are consequently required to offer not only traditional rest stop services but also attractiveness of tourism destination as well as additional services, such as refreshments, dining options, and souvenir shopping. It therefore implies the marketers who aim to increase demand to stop over at SA/PAs need to produce marketing strategies with attractive services at the targeted SA/PA so that potential tourists can recognize it as an additional tourism destination on the way to the final one. It is quite important for those marketers to consider diversification of tourist’s decision structure patterns underlying stopover behaviors at SA/PAs on expressway. Such diversification would hint us that there exists heterogeneity among decision structures in the microscopic level and thus intimate the existence of latent classes in macroscopic level so that we can calibrate the market segmentation of expressway users’ stopover behavior patterns. Market segmentation is a technique used to subdivide a heterogeneous market into homogeneous subgroups that can be distinguished by different variables, such as consumer needs, characteristics, or behavior (Katsoni 2013). The bases for market segmentation mainly include individual characteristics related to demography, geography, behavior, attitude, lifestyle, motivation, personality and sense of values. In travel and tourism marketing, the purpose is to facilitate more cost-effective marketing strategy through formulation, promotion, and delivery of the purpose-designed products that satisfy identified needs of the target groups. In other words, segmentation is justified on the grounds of achieving greater efficiency in the supply of products to meet identified demand, and increased cost-effectiveness in the market process (Morrison 2013). This also implies a shift in the direction of meeting the needs of individuals for product experiences that satisfy them and bring them back for more (Middleton 2009). When focusing our attention on methodological aspects of tourist segmentation, we can find there exists a variety of type of segmentation methods. Previous methods have been classified into mainly three types; a posteriori or factor-cluster segmentation, a priori or criterion segmentation, and the modeled types, for example, a neural network model (Katsoni 2013). A priori market segmentation can be less time consuming and more effective for separating markets at less cost. In tourism marketing, the importance of segmentation has been widely acknowledged (Mazanec 1992). To date research has assisted us to understand which bases should be used by tourism destination to effectively segment tourism markets, and these efforts have largely centered upon building tourist profiles for a destination, using visitor data (Frochot 2005). It is however noted that these previous segmentation methods are lack of adequately quantitative explanation of causality between the specific segment and productive services in tourism destination. We here need to identify the causal structure underlying tourists’ decision on their travel and tourism behaviors. It is therefore quite clear that the third type would have great potential as an effective tool for identifying such causality underlying the decision on drivers’ stopover behaviors on expressway.

The objective of this study is to identify market segmentation in tourism and leisure vehicle driver’s behavior patterns of stopping over at SA/PAs on expressway. The latent class analysis (denoted LCA) is applied to driver’s behavior pattern data on two different routes: One is Chugoku Expressway and the other is Sanyo Expressway. Our developed LCA models intend to explore the number of the latent classes determining decisions on both stopover at SA/PA and the location choice. They also intend to represent the structure how the drivers, who would be conditioned by those latent classes, make decision on choosing the location of SA/PA with using the observed variables of the targeted routes. Furthermore LCA models can simultaneously identify the probability of those unmeasured class memberships and its causal relationship among the estimated parameters of covariant variables. In our previous study (Nishii et. al. (2013, 2015)), the a-priori market segmentation method was applied to the nested logit models focusing on the travel distance on expressway. In such a-priori market segmentation, it is assumed that expressway drivers are a single member among the segmented classes and that each of the segmented models can respectively represent the decision on both whether these drivers stopover at a certain SA/PA or not, and where they choose the location. On the contrary, the LCA models are capable of explicitly identifying the number of the latent classes and the causal structures underlying drivers’ stopover behaviors. It means that drivers would belong to the multiple latent classes with their membership probability.

In this paper an analytical framework of tourism related long distance traveler’s behaviors is first introduced reviewing our previous studies on use of SA/PAs on expressway. In the second section, using two web-based survey data sets of driver’s stopover behaviors at SA/PAs on the Chugoku and the Sanyo expressways, the basic cross-section analysis is executed to identify how factors determine these drivers’ stopover behaviors. The analysis intends to specify several key-variables in our LCA modeling. The key-variables would consist of long distance drivers’ attributes, the tour characteristics, information related to the targeted SA/PAs and their locational characteristics. The third section represents calibration of our LCA model with covariate parameter-estimates and introduces the results of applying such LCA models to two different routes: One is the data set covering four rest areas on the Chugoku Expressway; Nishinomia-Nashio SA, Akamatsu PA, Yashiro SA, and Kasai SA. The other is the data set of seven rest areas of the Sanyo Expressway covering the intersection between Kodani SA and Kibi SA. In this section, based on the results of our comparative analysis, we will discuss the effectiveness of our LCA models and refer to future directions in our study.

Let us briefly introduce a basic model structure of LCA, which consists of both the membership probability of the s-th latent class for individual i (denoted πis) and the probability of choosing the j-th category for the i-th individual who is determined by the s-th latent class (denoted Pi|s(j) ). It is here hypothesized that πis would be πs, (i=1,2,…, N). Assuming that such membership is given and that the variable Xij (xij=1, when individual i choose the j-th category of SA/PA and otherwise xij=0) is known, we can define Pi|s(j) as follows:

Pr (Xij = xij|Ci = s) = Pi|s(j)xij (1 - Pi|s(j))1-xij   (1).

Assuming the Pi|s(j) is independently and irrelevantly chosen within the same latent class (s) for the j-th category (j=1,2,…,M), the probability of category choice pattern for individual i (denoted Pr (Xi1 = xi1, …,XiM = xiM) = Pr (θ) ) is expressed as a function of πs and Pi|s(j) as follows:

Pr (Xi1 = xi1, …,XiM = xiM) = ΣWs=1 πs ΠMj=1 Pi|s(j)xij (1 - Pi|s(j))1-xij (2),

where πs = πis (assuming the s-th unmeasured class membership probability is common to individual i) . Using Bayes’s theorem, the conditioned probability Pr (Ci = s|Xi1 = xi1, …,XiM = xiM) can be defined as follows:

Pr (Ci = s|Xi1 = xi1, …,XiM = xiM) = Pr (Xi1 = xi1, …,XiM = xiM|Ci = s) /Pr (θ) (3).

The EM algorism, that consists of two iterative steps, is used to estimate parameters of the observed variables in Pi|s(j) andπs . In this iteration process, the probability that the i-th individual belongs to the s-th latent class; yis, where the value is assumed to be the expected one, yis*, is estimated in the Expectation step. The following equations of the likelihood (L) and log-likelihood (log L) functions are used in the Maximization step of the EM algorism (See Demster et. al (1976) and Okada & Moriguchi (2010)).

L = ΠNi=1 ΠWs=1 [πs ΠMj=1 Pi|s(j)xij (1 - Pi|s(j))1-xij ]yis (4), and

log L = ΣNi=1 ΣWs=1 {yis xij Σj log Pi|s(j) + yis (1-xij) Σj log (1 - Pi|s(j))} + ΣNi=1 ΣWs=1 yis log πs (5).

In this study, the Ordered Binomial Logit model (denoted OBL-typed model) is used to calibrate the model structure of Pi|s(j). This is due to the fact that the parameter estimation is based on the binomial logistic regression model focusing on the odd probability ratio as shown in the following equations of the Rasch model (See Sumi (2012)):

Next, let us introduce results of the parameter estimates of our LCA model in case of the Sanyo Expressway. As shown in Figure 1, seven of the SA/PAs on Sanyo Expressway are targeted for our modeling driver’s stopover behaviors. In the case of Sanyo Expressway data, it is found that four latent classes model is the best practice among the converged LCA ones judging from the results of goodness of fit indices such as BIC, AIC, p-value, and R2-coefficient: This four latent classes model shows that the value of AIC is the lowest and that R2 = 0.844. Table 1 shows results of parameter estimates of Pi|s(j) and those of the significant covariate variables of πs, (s=1,2,3,4) in the model. The upper table indicates that there exist the latent classes determined by specific variables such as, SA/PA dummy, distance from ON-IC, and SA/PA location index. The drivers belonging to Class 1, Class 2 and Class 4 tend to prefer to choose SA/PAs in the upper side. It is however noted that those who belong to Class 1 tend to prefer to choose specific rest areas; Odani SA and Fukuyama SA. On the other hand, those who belong to Class 4 prefer to Michiguchi PA, and those in Class 2 hardly prefer to Odani SA, Takasaka PA and Yahata PA. In the case of those who belong to Class 3, they are not determined by specific SA/PA and the location but by the traffic volume in front of SA/PA. The lower table shows the results of parameter estimates of the covariate variables in the membership probability (πs) . They indicate that each of these latent classes are determined by a few of common covariate variables; distance between ON-IC and OFF-IC, purpose of travel, recognition, and reason for stopover. Both Class 1 and Class 3 are featured with ‘Short-distance traveling’ but they have difference in other covariate variables: In the purpose of travel, ‘Sightseeing and leisure’ is dominant for Class 1 but ‘To return home from hometown’ has majority in Class 3. In recognition, ‘unknown’ in Class 1 and ‘well-known’ in Class 3 are dominant. Furthermore, in the reason for stopover, ‘Toilet’ and ‘Gas station’ in Class 1 and ‘Dining’ and ‘Rest for passengers’ in Class 3. On the other hand, Class 2 is determined by ‘middle length distance traveling’ (100 km-200 km) and Class 3 is determined by ‘long distance traveling’ (200 km-300 km). In addition, both two classes differ from the reason for stopover and the number of visits to the SA/PA. Next, we represent results of applying our LCA models to the data set of the Chugoku Expressway in order to compare with those in the Sanyo Expressway. This analysis intends to identify common latent classes determining expressway users’ decision on their stopover behaviors. Finally, we discuss future direction for improving modeling of market segmentation of SA/PA stopover behaviors through summarizing valuable fact-findings from our developed LCA models.

11:00-12:30 Session 10E: Choice Experiment Design Part 2
David Brownstone (University of California, Irvine, United States)
Location: MCC Lounge
Assel Dmitriyeva (New York University, United States)
Daniel Fay (WSP USA, United States)
Xuebo Lai (New York University, United States)
Joseph Chow (New York University, United States)
Effect of routing constraints on learning in contextual bandit mobility-on-demand destination recommendation systems

ABSTRACT. One of the biggest contributing factors to high operating costs for mobility-on-demand is the presence of disruptive incidents. MOD services can be smarter by interacting with users and recommending activity destinations to them. We explore the use of recommender systems made popular by companies like Amazon and Netflix that can suggest destinations to a user when they are booking a trip or when an incident occurs that would either significantly increase the cost of delivering a passenger or do so at a higher risk of delaying other passengers to their appointments. This would also allow services like Uber, Lyft, Didi, and Via to offer an option for a traveler to book a trip to “a restaurant”, for example, and leave it to the service to recommend specific nearby restaurants to the user that they are likely to enjoy. This is also a first step in making mobility companies act as “physical search engines” for travelers. The problem, particularly with new services with limited initial data, is how to build up the database in an efficient manner. The problem of learning from repeated trials of option selections is called a contextual bandit problem. Destination recommender systems conducted by an MOD service are different from conventional recommender systems. There is a fundamental conflict between trying to minimize regret by learning as efficiently as possible, and trying to minimize operating costs and passenger travel costs. A destination may be highly rated and offers good learning opportunity for a user, but having to re-route the MOD service vehicle (and its passengers) to serve the location might significantly increase the other two costs. By having to account for those routing constraints, we hypothesize there is a spatial effect that degrades the regret minimization bound, and due to differences in spatial distribution of destinations, user travel patterns, and road infrastructure in different cities, this effect will vary from city to city. This study makes two primary contributions to the literature. First, we propose the first destination recommender system for MOD services to learn travelers’ destination preferences over time with minimal regret in recommendation quality, operating cost, and user travel cost. Second, we conduct experiments using user data from Yelp Open Dataset to obtain training and test data for three different cities: Las Vegas, Toronto, and Charlotte. These experiments provide insights on how to balance the trade-offs between recommendation quality with those costs, and how built environments can impact route-constrained recommender systems.

Muhammad Fayyaz (Institute of Transport and Logistics Studies, Business School, University of Sydney, Australia)
Michiel Bliemer (Institute of Transport and Logistics Studies, Business School, University of Sydney, Australia)
Matthew Beck (Institute of Transport and Logistics Studies, Business School, University of Sydney, Australia)
Route choice behaviour in stated choice experiments with and without consequences

ABSTRACT. Stated choice experiments (SCE) are widely used to examine travel choice behavior in hypothetical choice contexts and to derive values of time and reliability for transport project appraisal purposes. Because of the hypothetical nature of these experiments, there are no consequences attached to drivers’ responses, and hence the external validity of SCE outcomes is often questioned. This paper examines drivers’ route choice behavior in a typical SCE and in an SCE where consequences are experienced through driving in a simulator and where money needs to be paid for toll roads. These actual time and money consequences make the choice tasks less hypothetical.  Route choice observations of 55 drivers were analyzed using a heteroscedastic panel mixed logit model. We did not find evidence that the value of time and the value of travel time reliability were different when we added consequences to the tasks in the SCE, except for drivers with low income where a significantly lower value of time was found when drivers experienced time and money consequences of their choices. Further, we found that error variance is much larger when choices were made in the driving simulators.

David Brownstone (University of California, Irvine, United States)
Michael McBride (University of California, Irinve, United States)
Si-Yuan Kong (University of California, Irvine, United States)
Amine Mahmassani (University of California, Irvine, United States)
Experimental Studies for Traffic Incident Management

ABSTRACT. Non-recurring traffic incidents are responsible for nearly 60% of delay caused by roadway congestion, prompting the need for efficient incident management (FHWA, 2000). Network operators can often alleviate congestion and mitigate delays by diverting traffic from affected roadways onto alternate routes. One tool widely available for inducing such diversions is variable message signs (VMS) – programmable electronic roadside displays that can provide travelers with timely information regarding road conditions. Some of the earliest VMS systems in the U.S. were used in Detroit in the 1960’s to direct motorists to alternate routes based on freeway traffic conditions (Dudek, 2002), and field studies by Dudek et al. (1978) and Weaver et al. (1977) have confirmed the ability of VMS to aid incident management on freeways by diverting traffic onto alternate routes. Under most circumstances, however, transportation agencies are hesitant to use VMS to encourage diversions. Although they possess tools to determine the optimal proportion of vehicles to divert (Cragg et al., 1995), they lack a reliable method for achieving the targeted diversion rates. Prior research shows that when a high percentage of travelers are presented with route-choice information, myopic agents can make diversion decisions that, in aggregate, worsen road network performance (Mahmassani and Jayakrishnan (1991)) and that VMS systems deployed in Minnesota don’t lead to significant reductions in travel times (Levinson and Huo (2003)). As a result, both the agencies and city officials share fears of overloading surface streets with an excessive diversion (FHWA 2000). Consider further research to optimize VMS systems. Efficient incident management through VMS necessitates finding a type of public information that, when provided to all drivers, will produce the desired distribution of traffic across available routes. With the limited opportunity for drivers to coordinate, it is unlikely to achieve the desired response. This phenomenon is manifest in highly stylized route-choice games conducted by Selten et al. (2004) and Iida et al. (1992), where an efficient equilibrium distribution was extremely difficult to reach – even across repeated trials with full information and feedback. Instead, one might expect to see non-smooth changes in diversion rates as VMS content is varied. Field studies by Chatterjee et al. (2002) and Horowitz et al. (2003) confirm the unpredictability of diversion rates induced by VMS. Though it is possible to mitigate many of these coordination issues through the selective provision of information privately to drivers via in-vehicle systems, such systems are not yet ubiquitous, and system operators cannot control for users receiving information from third parties. Given the large presence of VMS infrastructure in the US and abroad, it is desirable to improve its effectiveness as a low-cost, readymade tool for incident management. To achieve this objective, we need to gain a better understanding of how VMS information affects time-limited decision-making in scenarios where drivers possess imperfect information of the environment and influence each other’s behavior. We seek to explore how the availability and manipulation of VMS content will affect driver decision-making in real-time. In particular, we focus on how an increase or decrease in the “intensity” of VMS content – that is, message adjustment intended to induce more or fewer drivers to divert – can produce a desired change in the diversion rate. We designed a route-choice experiment to test a variety of different VMS messaging schemes using a 2-dimensional real-time driving simulator with a simple road network. We incentivize subjects with real monetary payments to induce a controlled value of time preference. The experiment and driving simulator were designed to support the following features: 1) vehicles move in real-time and obey simplified Newtonian kinematics, requiring drivers to exert effort to maintain course and speed. 2) A large number of human drivers share the same virtual roadway to create a sense of immersive traffic. 3) Traffic and congestion are generated endogenously from a combination of the large number of vehicles and the effects of reductions in route capacity. 4) Drivers are exposed to limited and dynamic forms of information based on what they would perceive while driving on the freeway in real life, and 5) the road geometry is simplified yet retains a structure relevant to studying the questions of interest. The experiment platform was a 2-Dimensional, real-time, top-down perspective-driving simulator implemented as a browser application using Node.js, JavaScript, and HTML5. Subjects saw a top-down view of the roadway where vehicles were represented as small colored squares - the driver's own vehicle was colored blue while all other vehicles were colored red. The driver's viewport constantly tracked his/her vehicle and presented a fixed window of visibility around it - the driver could see farther ahead than behind to simulate the forward-focused vision of real-world drivers. From top to bottom, the driver's screen contained the following elements: the secondary information area that displayed the current experiment round, the VMS display area, the driver's viewport, and the primary information area that displayed the driver's earnings and percent completion of their itinerary in real-time.

Figure 1: Driver's screen

Using the keyboard Up Arrow or letter W, Left Arrow or letter A, and Right Arrow or letter D keys, drivers’ controlled their vehicles to accelerate or change lanes left and/or right. All vehicles accelerated at the same rate and quickly reached the same maximum speed. If a driver stopped accelerating, their vehicle will decelerate at a constant rate until it reached the minimum speed. The minimum speed was designed to prevent a driver from completely blocking their lane and it slowed enough, such that a driver who always traveled at the minimum speed would never complete their itinerary before their entire endowment was expended. While cruising, vehicles were automatically guided to stay in the center of the nearest lane. A minimum following distance was enforced between cruising vehicles to allow space for lane changes to occur. If another vehicle, when attempting to change lanes, obstructed a driver’s vehicle, their vehicle would have slowed down slightly to allow them to move in behind the obstructing vehicle. Drivers were informed that there were no rewards or penalties for colliding with other objects or vehicles. In addition to human controlled vehicles, the computer-controlled vehicles, which follow simple pre-defined control routines, were used to fill in the front of the driving platoon to create a sense of immersive traffic. We have demonstrated the feasibility of this approach by carrying out a number of experiments using University of California, Irvine undergraduates as experimental subjects. We found that providing any incident-related information via VMS improves system performance relative to the no-information baseline, but also found that more complicated dynamic messaging with feedback did not always improve system performance relative to standard VMS messaging. Given the difficulty of getting a representative sample of drivers to come to the experimental laboratory, we implemented the real-time experiments using the Amazon Mechanical Turk (MTurk) platform. This approach made it affordable to run experiments using a much larger and representative pool of experimental subjects. Unfortunately, the limitations of the MTurk platform, coupled with the challenges of remotely administered sessions, made converting and running experiments much more difficult than anticipated. Nevertheless, enough experiments were completed to show that the results using undergraduates were not substantially altered by using a more diverse and representative experimental subject pool. Dynamic road pricing is another possible tool for managing road congestion. However, optimal pricing requires that system operators know the distribution of the Value of Time (VOT) for road users. It is difficult to measure the VOT distribution using standard transportation survey techniques, and there is evidence that VOT varies across trip purposes and time. The second goal of this project was to investigate the possibility of using preference elicitation methods to elicit the VOT for each road user. This procedure was implemented in the experimental platform and carried out a series of experiments using UCI experimental subjects. Although the method used gives incentives for subjects to truthfully reveal their VOT, the results show that due to the cognitive complexity of the process many subjects reported erroneous VOT values. Nevertheless, the efficiency loss due to these errors was small, demonstrating this is a promising new method of managing congestion. This work is described in Section 5 of this report.

References: Chatterjee, K, N.B Hounsell, P.E Firmin, and P.W Bonsall. “Driver Response to Variable Message Sign Information in London.” Transportation Research Part C: Emerging Technologies 10, no. 2 (April 2002): 149–69. doi: 10.1016/S0968-090X(01)00008-0. Cragg, C.A., M.J. Demetsky, and Virginia Transportation Research Council (1994). Simulation Analysis of Route Diversion Strategies for Freeway Incident Management. VRTC. Dudek, C.L. (2002). Guidelines for Changeable Message Sign Messages. Report No. FHWA-XX-2002-XX. FHWA, U.S. Department of Transportation, December.

Dudek, C.L., G.D. Weaver, D.R. Hatcher, and S.R. Richards. “Field Evaluation of Messages for Real-Time Diversion of Freeway Traffic for Special Events.” Transportation Research Record 682 (1978): 37–45. FHWA (2000). “Traffic Incident Management Handbook.” Horowitz, Alan, Ian Weisser, and Thomas Notbohm. “Diversion from a Rural Work Zone with Traffic-Responsive Variable Message Signage System.” Transportation Research Record 1824, no. 1 (January 1, 2003): 23–28. doi: 10.3141/1824-03. Iida, Yasunori, Takamasa Akiyama, and Takashi Uchida. “Experimental Analysis of Dynamic Route Choice Behavior.” Transportation Research Part B: Methodological 26, no. 1 (February 1992): 17–32. doi: 10.1016/0191-2615(92)90017-Q. Levinson, D. and Huo, H. (2003). Effectiveness of Variable Message Signs. TRB Conference 565. Mahmassani, Hani S., and R. Jayakrishnan. “System Performance and User Response under Real-Time Information in a Congested Traffic Corridor.” Transportation Research Part A: General 25, no. 5 (September 1991): 293–307. doi: 10.1016/0191-2607(91)90145-G. Selten, Reinhard, Michael Schreckenberg, Thorsten Chmura, Thomas Pitz, Sebastian Kube, Sigurður F. Hafstein, Roland Chrobok, Andreas Pottmeier, and Joachim Wahle. “Experimental Investigation of Day-to-Day Route-Choice Behavior and Network Simulations of Autobahn Traffic in North Rhine-Westphalia.” In Human Behavior and Traffic Networks, edited by Michael Schreckenberg and Reinhard Selten, 1–21. Berlin, Heidelberg: Springer Berlin Heidelberg, 2004. doi: 10.1007/978-3-662-07809-9_1. Weaver, G.D., C.L. Dudek, D.R. Hatcher, and W.R. Stockton. “Approach to Real-Time Diversion of Freeway Traffic for Special Events.” Transportation Research Record 644 (1977): 57–61.

Camila Balbontin (Institute of Transportation and Logistics Studies, Australia)
David A. Hensher (Institute of Transportation and Logistics Studies, Australia)
Andrew T. Collins (Institute of Transportation and Logistics Studies, Australia)
Process homogeneity, process heterogeneity, and preference heterogeneity: How to better represent decision making and preferences

ABSTRACT. Discrete choice modelling studies are with rare exception focused on the outcome in decision-making while making (by default) strong assumptions on the process that individuals adopt in reaching a decision. They typically assume a linear in the parameters additive in the attributes (LPAA) approach (with interactions). A growing number of transportation (and other) choice studies have questioned these assumptions by investigating alternative process strategies; and some recent literature has considered the application of multiple process strategies allowing for process heterogeneity. This topic has given rise to new questions, such as whether process heterogeneity is confounded with preference heterogeneity.  A recent study has proposed a way to incorporate two heuristics to allow for process heterogeneity and also allow for preference heterogeneity, referred to as conditioning of random process heterogeneity (Balbontin et al., 2017a). In this research we extend this method to incorporate three process heuristics - LPAA, Value Learning, and Relative Advantage Maximisation - and to incorporate other behavioural refinements, such as risk attitudes and perceptual conditioning, and overt experience. An important findings is that when process heterogeneity is accounted for through specific heuristics, behavioural refinements and overt experience may not be required. This helps in identifying appealing parsimonious preference expressions in choice models. When preference heterogeneity is overlayed in these more parsimonious models through interaction with random parameters, we obtain new insights into the relationship between preference heterogeneity and process heterogeneity. These two phenomena are correlated, and condition each other in behaviourally informative ways. This study shows empirically that there exists (in two datasets at least) a significant attribute-specific relationship between process strategies and random parameters.

Reference: Balbontin, C., Hensher, D.A., Collins, A.T., 2017a. Is there a systematic relationship between random parameters and process heuristics? Transp. Res. Part E Logist. Transp. Rev. 106, 160–177.

11:00-12:30 Session 10F: Land Use -- Expanding the Sphere of Travel Behavior Research
Fatemeh Nazari (University of Illinois at Chicago, United States)
Milad Ghasri (The University of New South Wales, Australia)
Amarin Siripanich (The University of New South Wales, Australia)
Taha Rashidi (The University of New South Wales, Australia)
Integrating the Garing-Lowry Model with a Travel Demand Model

ABSTRACT. [Please refer to the attached document]

Jason Hawkins (University of Toronto, Canada)
Adam Weiss (University of Toronto, Canada)
Khandker Nurul Habib (University of Toronto, Canada)
Modelling Residential Location Choices of Dual-Worker Households by Incorporating Intra-Household Interactions within a Random Utility Maximization Framework
SPEAKER: Jason Hawkins

ABSTRACT. Dual-worker households are an increasingly important subset of household arrangements in the analysis of transportation behaviour. Increasing costs of living relative to wages necessitate additional sources of income for the household. The proportion of dual-worker households in Canada rose from 38% in 1976 to 70% in 2015 [1]. In the context of transportation behaviour, dual-worker households negotiate over limited resources (e.g. the number of household vehicles or mobility tools) to minimize their journey-to-work travel expenditure (time and costs). Capturing such interactions of household members and their inclination towards compromise or shared travel are critical in modelling residential location choices of dual-worker households. Conventional univariate decision theories are insufficient to model such choice contexts. Ho and Mulley [2, 3], as well as Akbari and Habib [4], suggest that short and long-term decisions (e.g. home location, auto ownership, and proximity to rail transit choices) that are conventionally modelled at the level of the household are in fact composite decisions resulting from tradeoffs between members (joint decisions) of the household to achieve a reasonable level of household utility. However, applications of household level composite utility-based joint decision theory in transport modelling have been scarce and it is a pressing issue in household-based travel demand investigation [5]. Consider a dual-worker household with a single vehicle in which the wife has a shorter commute to work (typical of most households [6, 7]). It is assumed that the husband travels to work by public transit due to high congestion on the major road to his place of work. A toll is introduced on the road, which reduces congestion. The husband now sees a higher utility of travel by automobile and uses the family vehicle. The wife suffers from this decision on two fronts: 1) she no longer has access to the vehicle and 2) she suffers a loss of real income due to the toll incurred by the husband. Such situations of household-level decisions are not adequately considered in the most transportation behaviour literature. Neglecting such intra-household interactions could compromise the validity of policy decisions based on model results. The present research contributes to this critical research gap by presenting a Random Utility Maximization (RUM) based joint decision model of home location choices for dual-worker households.

One of the major factors in residential location choice is accessibility by different travel modes. For example, a household with a single vehicle and two workers is likely to place a higher weight on access to public transit to one, or both, work locations than an otherwise similar household with access to two vehicles. Thus, modal accessibility has a strong spatial correlation as the quality of transit and active modes will vary across a region. Likewise, access to a personal vehicle largely will be dependent upon the price level of housing, such that a higher price in one location will restrict vehicle ownership relative to an otherwise similar house at a different location. Similar to Ho et al. [8], in our modelling approach, we propose considering the expected maximum utility of mode choice as a determinant in the residential location model. However, instead of simply using daily travel mode choice utility as a direct input to residential location choice, it is used to explain the choice of auto ownership and that is fed into the joint choice of home location. This process is explained graphically in Figure 1. Residential location choice, auto ownership, and proximity to rail transit tend to be related long-term decisions, whereby households make a decision whether to trade-off auto ownership for transit or walk access to work. As such, mobility tool access is deemed an important factor in the model. Mobility tool access can be generically defined as access to an automobile, train, bus, or any form of mechanical mobility tool. We are restricted in our analysis by data availability to considering only auto deficiency.

[see attached file for Figure 1]

Figure 1 Model nesting structure

For this, an innovative commuting mode choice model is devised, with the objective of capturing the implicit trade-offs made in vehicle allocation and the decision to engage in joint travel. Choice set formation for the commuting mode choices is derived from mode pairs for both workers rather than independently specifying individual choice sets. This model structure allows us to consider the implicit travel time increases and household level savings associated with joint travel. A dummy variable is added to all joint travel utilities to capture the increased likelihood of joint travel in auto deficient households. This provides important insights with respect to detours associated with joint travel, which may influence residential location choices.

Assuming fixed work locations, residential location choices are influenced by the possibility and practicality of the household members travelling jointly to work along with how vehicles are allocated to household members in mobility tool deficient households. Consider the case of a dual-worker household with work locations at a large distance from one another and poor transit access to both work locations. Individual utility maximization would suggest the household choose a residential location at any point on the line between the two work locations to minimize the summation of the travel distance for each worker in the household. When considering the possibility of joint travel, the optimal configuration changes substantially. Instead of any point on the line between the two locations being feasible, a configuration where the residential location is closer to the work location of the household member being dropped off minimizes household travel distance. Figure 2 outlines these concerns. Even in the case of good access to transit, the auto deficiency will influence residential location choice. In these cases, if the modeller assumes uniform access to vehicles for both household members, the residential location that maximizes household utility will again be at any point along the line connecting the two work locations. Given that only one household member can be allocated the vehicle in an auto deficient household, the true preferred home location is much more proximate to the work location of the household member who is not allocated the vehicle, thereby making active travel modes, transit, or joint travel with their spouse more attractive. This hypothetical case suggests that when modelling residential location, it is important to consider both the possibility for joint travel and the allocation of vehicles in mobility tool deficient households. This understanding is furthered by noticing that residential location choice trade-offs are directly tied in with questions of mobility tool access.

Note that while Picard et al [2] have identified the relationship between vehicle allocation and joint travel alongside residential location in their previous work, they did not account for the endogeneity of vehicle ownership in the choice process, which is a concern. They also present a structural model to accomplish the similar outcomes that our innovative choice set model accomplishes. However, within their observed data set there is no information distinguishing between driving and acting as a passenger. Passenger and driver roles are inferred from the travel times for each occupant. The model proposed in this paper aims to build on their work by directly considering the observed travel patterns (auto passenger versus auto driver) and by accounting for the trade-offs made at the household level with respect to vehicle ownership and residential location.

[see attached file for Figure 2]

Figure 2 Shared trip distance for (a) home location equidistance from work locations W1 and W2 and (b) home location near work location of W1

It seems intuitive to suggest that dual-worker households travel more than one-worker households due to increased journey-to-work travel. The risk associated with this would be increased total travel with rising rates of dual-worker households. However, there is no consensus as to whether observation bears out this intuition. Much of the literature suggests that dual-worker households commute the same, if not less distance, relative to one-worker households [9]. This issue boils down to whether partner commuting distances are complements or substitutes. That is, whether households engage in intrahousehold trade-offs of commuting distance. In a study of travel survey data in Montreal, QC, Surprenant-Legault et al. [9] find that commuting distances are complementary in that as the distance travelled by one worker increases by 1%, so does the distance of the other worker. However, they find the elasticity is less than one, suggesting that trade-offs in residential location choice occur such that total household commute distance increases by less than 1%. The structure of our mode choice model lends itself to modeling these conditions via its implicit representation of intrahousehold dynamics. In a traditional husband-wife household with both working, it has been found that wives typically have shorter commutes and therefore carry a stronger weight in the joint residential location choice. Similarly, it has been found that residential location tends to reduce travel distance by transit for women. Therefore, modal accessibility has a gendered bias and intrahousehold dynamics are vital to correctly modeling the decision-making process of the household.

Residential location choice is clearly the result of a suite of factors, with modal accessibility being only one of them. Sociodemographic factors are also important in the decision to locate in a specific area. Monetary budgets constrain households, with housing and transport consuming major portions of household income. Statistics on dwelling prices are available from the Toronto Real Estate Board (TREB), which are used as measures of the weight of relative prices on the determination of residential location choice. Sociodemographic characteristics are also important factors in the decision process. The TTS provides the household characteristics for each person, from which we can ascertain the number of workers, their gender, and the number of children in the family unit. Our previous research in mode choice determined that parents often make sub-optimal decisions to maximize the utility of their children [10]. The same is true of residential location choice, in that proximity to high-quality schools will play a disproportionate role relative to the work location in the decision process. Principles developed in the context of mode choice will be applied to residential location choice to explore how group decision making can be used to improve the fidelity of models of dual-worker households. The TTS and residential price statistics will be combined to develop a model of residential choice. The predictions of the model will be compared against the stated residential locations provided in the TTS. Conclusions will then be drawn as to the role of intra-household bargaining, modal accessibility and choice, and dwelling price level play in the residential location choice process.

Fatemeh Nazari (University of Illinois at Chicago, United States)
Abolfazl Mohammadian (University of Illinois at Chicago, United States)
Sybil Derrible (University of Illinois at Chicago, United States)
Interactions Between Energy/Water Consumption and Travel Decisions in an Activity-Based Framework

ABSTRACT. Given the ever-increasing levels of energy and water consumption worldwide and the consequent economic and environmental impacts, it is of critical importance to understand the factors affecting energy and water consumption. Despite the numerous studies in different fields on modeling energy and water consumption, there is a dearth of research in considering interactions between mobile (i.e., transportation) and stationary (i.e., residential) sources of energy and water consumption. In addition, the impact of individuals’ travel behaviors and activities on energy and water consumption have yet to be investigated. This paper integrates an Artificial Neural Network (ANN) model of residential energy and water consumption with the Agent-based Dynamic Activity Planning and Travel Scheduling (ADAPTS) framework based on in-home activities of agents. While traditional activity-based models mainly focus on out-of-home activities, the updated ADAPTS framework is capable of joint simulation of in-home and out-of-home activities of agents. In fact, agents could update their out-of-home activities according to their in-home activities and vice versa. Given the interaction of in-home and out-of-home activities, this paper is one step forward to link travel decisions, derived from out-of-home activities, and residential energy and water consumption based on in-home activities. An empirical analysis is conducted that relates the energy and water consumption of agents to their in-home activity type and duration, which highlights the roles of socio-economic characteristics and residential factors.

Dena Kasraian (University of Toronto Transportation Research Institute, Canada)
Shivani Raghav (University of Toronto Transportation Research Institute, Canada)
Eric Miller (University of Toronto Transportation Research Institute, Canada)
Long-term transportation and land use developments in the Greater Toronto-Hamilton Area, Canada: a preliminary analysis
SPEAKER: Dena Kasraian

ABSTRACT. Access to transport infrastructure is known as one of the drivers of urbanization, along with factors such as demographic growth, economic development, and spatial policies. Improvements in transportation networks increase accessibility, which results in the redistribution of land use, i.e., the location of activities. Consequently, change in activity locations and the way people use transport infrastructure networks to access them, which constructs their travel behavior, induces the need for improvements in transportation networks. Highways are known to promote urban sprawl, while transit stations are used as a tool to encourage compact urban development in their vicinity. Measuring the impact of transport accessibility on the magnitude and rate of urbanization, which is an almost irreversible process, is critical for evaluating past investments in transport infrastructure and guiding future spatial and transport planning. The impact of transportation on land use is only observable over the long term as it is a slow process. For instance highway and road infrastructure are built over decades to connect growing economic regions. This transport connectivity creates economies of scale and induces locational shift of firms and labour. However, this impact has hardly been measured over long time periods consisting of several decades. Previous long-term studies of transportation and land use have mainly evaluated changes in population density as a proxy for growth. Urbanization, in terms of the area converted from undeveloped to urban, is less investigated mainly due to the unavailability of long-term data on urban growth. From a spatial perspective, it is important to investigate urbanization to gain insight into the location, magnitude and rate of urban sprawl and to measure the share of transport accessibility in it. Furthermore, the impact of transport networks on land use needs to be studied at a regional scale regarding the “network characteristic” of transport infrastructures. Based on this characteristic, a change in a specific part of a transport network not only changes the accessibility of that link or node and the land use in its direct vicinity, but has consequences for accessibility at the network level and induces land use change in other locations. Despite this issue, long-term research at the regional scale has also been scarce. Finally, most long-term studies only investigate the impact of road or transit accessibility, but rarely both. Addressing the gaps above, this paper explores the long-term co-development of urbanization and transport networks in the Greater Toronto-Hamilton Area (GTHA), Canada from 1971 to 2011. The goal is to first provide an overview of the magnitude and rate of the development of urbanization and transport infrastructures separately and then to explore the relation between them. Urbanization is measured as the built-up area, including real estate for housing, services, companies, infrastructure and parks. Investigated transport networks include both transit, i.e., commuter rail, heavy rail subway and their stations, as well as roads, consisting of highways and their exits, arterials and local roads. Urban development and transportation networks are measured respectively at 10-year and 5-year intervals during the study period. The method consists of descriptive graphs of the development of urbanization and transport networks, as well as a GIS-based buffer analysis which explores the relationship between the amount of urbanization and the proximity to regional transit stations and highway exits over time. Data from various sources were gathered and made consistent to create a database which documents the growth of both the built-up area and transport networks over 40 years. The sources for the built-up area include the Canada Land Use Monitoring Plan (CLUMP), for 1971 and 1981 and Agriculture and Agri-Foods Canada (AAFC), for 1990, 2000 and 2010. Road and rail transit networks were derived from Equilibre Multimodal/Multimodal Equilibrium (EMME) networks developed by the Travel Modelling Group (TMG) at the University of Toronto Transportation Research Institute’s for Travel Modelling Group for the GTHA at 5-year intervals from 1986 to 2011. Transport networks for the 1970s and early 1980s were extracted from the 1986 network by consulting historical hard copy transportation maps made by the Ontario Ministry of Transportation. Descriptive results show that the length of the GTHA commuter rail network, the so-called Government of Ontario (GO) Transit, increased during the 1970s and the first half the 1980s, was stable during the following two decades, and again grew rapidly since the mid-2000s due to the addition of two GO rail lines which extended beyond the GTHA boundary. Overall, the GO Transit and subway network both grew more than fourfold in length over the study period. As for the commuter rail stations, GO rail service started in 1967 with 14 stations and expanded to 7 corridors and 61 stations by 2011, while the number of subway stations saw an increase from 45 to 69 in the same duration. Most of the new GO transit and subway stations opened in the beginning and the end of the study period, leaving the period 1986–2001 unchanged for the most part. GTHA underwent substantial urbanization over the study period, witnessing a 125% growth in the built-up areas, mostly in form of ribbon developments and developments on the edges of the main cities. Nevertheless, this growth slowed down in the 2000s. The co-development of urbanization and transport networks is analyzed by measuring the changes in the built-up areas in the vicinity of highway exits and GO transit stations. Five concentric non-overlapping ring-buffers with intervals of 1 kilometer were generated for existing stations and highway exits at each time point. Thus for years 1971, 1981, 1991, 2001, and 2011 the share of urban land within certain buffers of existing stations/highway exits and for the study area as a whole were calculated and compared. Results show that considering the amount of urban land present in the whole region, more is located in the vicinity of highway exits than regional transit nodes. This difference is much more pronounced at the beginning of the study period when not many GO transit stations existed. There is on average 15% more urban land in vicinity of highway exits compared to transit stations since 1981. Comparing urban growth in the five concentric rings, a distance decay trend becomes evident: development generally started close to the railway stations and highway exits and moved outwards, as development occurred less at longer distances from the stations. In other words, controlling for the buffer area, the first ring was always more built-up than the second and so on. When comparing the urbanized percentage of buffers around stations and highway exits, it becomes clear that a higher percentage of the buffers around stations are urbanized compared to buffers around exits. Thus while there is more urban land in the vicinity of highways in general, the urban land around stations is more concentrated. Future research will apply longitudinal analysis methods to investigate the impact of transit and road –measured by gravity-based accessibility indicators– on urban densities at the census tract level while controlling for the reverse impact of urbanization on transport networks.

11:00-12:30 Session 10G: Smartphone Survey Design
Zachary Patterson (Concordia University, Canada)
Location: UCEN Flying A
Mark Bradley (RSG, United States)
Elizabeth Greene (RSG, United States)
Christopher Coy (RSG, United States)
How Will Smartphone-Based Household Travel Surveys Enable Better Analyses of Travel Behavior?
SPEAKER: Mark Bradley

ABSTRACT. Introduction RSG has recently completed two major household travel surveys using a smartphone app as the primary means of data collection for up to seven days of travel for each respondent. Surveys have been completed of over 6,000 households in San Diego, California and over 3,500 households in Columbus, Ohio. The smartphone-based app is the only method offered for households where all adults own smartphones, while most adults who do not own a smartphone respond using more traditional diary-based survey methods, either on-line or via telephone interview. (In the Ohio survey, non-smartphone-owners were also given the option of having a smartphone sent to them for use in the survey.) Smartphone-based travel data collection is superior to the previous GPS-based survey methods where respondents would be sent a “black box” GPS recorder that would record their travel. Since people were not used to carrying around such devices, they often forgot to carry them or keep them charged, and often left them in their car when traveling. That is rarely an issue with smartphones, which most owners carry with them at all times. Also, survey questions can be triggered automatically on the smartphone, asking respondents for all details of the trips that have been recorded. Thus, all the data items that are captured in diary-based interviews are also captured in the smartphone app, without the need to do separate on-line or telephone-based prompted recall surveys. The lower respondent burden of smartphone-based data collection seems apparent from the initial use by respondents. Most respondents provide details of their trips within an hour or two of the time that the trip is recorded, while some others wait and provide details for all trips at the end of the day. Also, the large majority of respondents who provide one full day of travel data go on to provide a full seven days of data, with very little attrition across days. (Our surveys have not yet asked for more than seven consecutive days of data, although it seems likely that many respondents would be willing to keep providing data for more than seven days.) Also, data could be collected over a much longer period if the collection became fully passive at some point—only collecting travel times and locations without asking the extra questions about mode, purpose, co-travelers, etc. We have begun to use the smartphone-based data in various types of modeling and analysis, and the results thus far are very promising. In this paper, we discuss a variety of analyses that we have done, the implications for travel behavior research in the near term, and some possibilities for the longer term. The value of more complete and accurate travel data GPS-based data tends to be more accurate than self-reported data in terms of the trip end locations and times of day. Respondents of diary-based, self-reported methods tend to round off and/or misremember trip departure and arrival times, and often report inaccurate location data. Compared to GPS “black boxes”, smartphone apps have the advantage that respondents are able to view their trips on an on-screen map any time after a trip end is registered, and can easily split or merge trips if the app has missed a very short stop (e.g. to pick up or drop off a passenger) or has identified a false activity stop (e.g. a long stop at a traffic light or in a traffic queue). Smartphone-based data also provides route traces for all trips. Even if one is not interested in modeling route choice behavior, the extra trace data can provide useful information such as the stop locations used for transit trips, the time spent waiting at the transit stop, the location of remote parking, and the time spent walking from the parking location to the destination. Perhaps the largest benefit of smartphone-based data is that it consistently captures 20 to 25% more trips than diary-based methods, even after accounting for demographic differences between smartphone owners and non-owners. There appear to be two main reasons for the higher trip rates. One reason is that the smartphone-based travel days include only half as many person-days with no trips compared to diary-based methods. Many of the no-travel days in diary-based methods are “soft refusals”, where respondents do not report any trips as a quick way to complete the survey. A comparison of smartphone-based and diary-based data suggests that the soft refusals with diary-based methods are particularly prevalent among “Millennials” age 18 to 35. (A conjecture is that these younger adults have less patience manually filling in information that they know is already being tracked automatically by various types of apps on their smartphone.) A second reason is that respondents tend to under-report specific types of trips and activities with diary-based methods. Analysis shows that the types of trip and stops most often underreported are intermediate stops on multi-stop auto tours and short trips to/from work during the workday. (It is often conjectured that short walk trips are most often under-reported in travel diary surveys, but our evidence to date is that short car trips are more likely to be under-reported than short walk trips.) Overall, the evidence is that smartphone-based methods capture more trips than either diary-based methods or GPS “black box” methods, and thus provide a more complete and accurate record of travel patterns. The value of multiple days of data The value of collecting multiple days of travel data from each respondent is not obvious without some analysis. If people are very habitual in their weekday travel patterns, then collecting five weekdays of data would not provide much more useful data than collecting a single weekday. Our analysis of seven days of reported data has indicated, however, that most respondents are not as habitual as one might expect. Even for home-to-work trips, most respondents vary some aspect of their trips from day to day, be it the timing of the trips (more than two hours variation from day-to-day), the number of co-travelers, the mode of travel, or even the work destination visited. We are often asked about the relative value of collecting multiple survey days from the same households compared to getting more households to each do a one- day survey. This can be analyzed by determining what percentage of trips collected on each day are unique compared to the same person’s trips on previous days—in other words, how much new information are we getting on each survey day? Analyzing the smartphone-based trips collected for up to 7 days from the recent San Diego travel survey, a trip was designated as “unique” if the respondent made no other trip on previous survey days between the same two Census blocks, by the same mode, at roughly the same time of day (within a 2-hour gap). Figure 1 shows how many “unique” trips there are as a multiple of trips collected on the first day of the travel period. For home-based work (HBW) and home-based school (HBS), five weekdays of data provide about 2.5 times as many unique trips as one day. For the total trips and for other trip purposes (home-based other=HBO, non-home-based work=NHBW and non-home-based other=NHBO), five weekdays of data collection provide more than four times as many unique trips as one day of data collection. This evidence strongly supports the value of collecting multiple days of smartphone data, particularly given the fact that each additional day of data costs only a fraction of the cost that would be required to recruit another household into the survey. FIGURE 1: CUMULATIVE NUMBER OF UNIQUE TRIPS AS A MULTIPLE OF DAY 1 TRIPS, BASED UPON SANDAG TRAVEL SURVEY DATA

A particular benefit of capturing multiple days of data is that some modes of travel and/or types of trips are not done every day, so many more of them can be captured with multiple days of data collection. For example, over 1,700 Uber and Lyft (TNC) passenger trips were captured in the San Diego survey. Since many of these trips were made on Fridays and weekends, there would not be nearly as many such trips if data had only been collected for a “typical weekday”. In this case, the number of Uber and Lyft trips is enough to analyze the differences between TNC users versus non-users, and to possibly include the TNC mode in mode choice model estimation. Bicycle is another mode that is often used only on some days of the week, and where additional trips in the data can be very useful for modeling.

Possibilities for even smarter data collection and modeling As the smartphone data collection apps are further refined to reduce respondent burden and to collect complete and accurate data (and the data post-processing methods are continually refined as well), we can expect the benefits of smartphone-based data collection to grow even more in the future. One interesting possibility is for the development of better imputation methods that can reduce the need for active respondent participation. Although respondents do not appear to mind providing data actively for the first week or so, an app that could learn to impute a particular person’s travel details would obviate the need to continue asking for travel details over time. The evidence from past studies is that we cannot yet reliably impute travel modes or activity purposes from purely passive data on an aggregate level, but learning an individual’s travel patterns and how to impute them over time based on their responses may be a more feasible context for accurate imputation. Another possibility is the use of multi-day data to expand activity-based travel analyses beyond a single day, to model weekly or even multi-week travel and activity patterns. With traditional 9 to 5, Monday to Friday work weeks becoming less common and more people working at least part-time from home, the substitution between weekday and weekend activities may become a more important type of behavior to include in future travel models. These possibilities, as well as others, will be discussed in the paper and presentation.

Takahiko Kusakabe (The University of Tokyo, Japan)
Toru Seo (Tokyo Institute of Technology, Japan)
Wataru Nakanishi (Tokyo Institute of Technology, Japan)
Yasuo Asakura (Tokyo Institute of Technology, Japan)
Implementation of Interactive Online Machine Learning Approach for Smart Phone based Activity-Travel Survey

ABSTRACT. This study aims to implement a smart phone based interactive activity-travel survey method and examine the method by an empirical survey. This study implements the method proposed by Kusakabe et al (2015), which was aimed to adapt the method for long-term activity-travel survey by using an online machine learning method for adapting an estimation model as well as an online estimation method. In order to apply the method to the actual online survey, this study additionally develops an online stay-point detection method using GPS data in order to immediately ask travellers their trip contexts after their beginning/ending of their trips. By relying on both the conventional offline stay-point detection and machine learning method, the proposed method can detect the switching points of moving/staying more immediate than the conventional ones. The proposed method was examined by actual travel survey over one month.

Zachary Patterson (Concordia University, Canada)
Kyle Fitzsimmons (Concordia University, Canada)
Michael Widener (University of Toronto, Canada)
Jessica Reid (University of Waterloo, Canada)
David Hammond (University of Waterloo, Canada)
Duration Analysis of Length of Participation in Smartphone Travel Surveys

ABSTRACT. As smartphone travel surveys begin to play a more important role in the collection of travel behavior data, there is still a great deal we don't understand about them. One critical aspect about which we know little relates to how long respondents will participate in these surveys, and the factors that might affect this; be it related to the apps themselves, the characteristics (e.g. incentives and respondent burden) of the studies, or characteristics of respondents. This paper will contribute to knowledge in this important aspect of smartphone travel surveys by using data from four different smartphone travel surveys (and 12,000 respondents) administered through different versions of the same app between 2014 and 2017. Duration models will be estimated to statistically evaluate the factors affecting length of survey participation. Conclusions will helpful for better planning of such studies in the future.

11:00-12:30 Session 10H: Attitudes and Perceptions Part 2
Roger Chen (Rochester Institute of Technology, United States)
Location: UCEN Lobero
Roger Chen (Rochester Institute of Technology, United States)
Investigating a Sense of Place and Site Visit Frequency

ABSTRACT. This paper examines the relationships between: (i) Sense of Place (SOP); (ii) non-motorized sustainable travel; and (iii) the adoption and use of information and communication technologies (ICT). A guiding principle in designing the built environment for sustainability and livability is the latent construct of Sense of Place (SOP), which leads visitors to perceive and associate a strong identity or character with a particular location. Visitors’ SOP likely affects their access of sites via walking or other non-motorized travel modes. The widespread adoption of mobile ICT, such as smartphones and tablets, likely shapes their SOP. In an information era, the ability of ICT to provide ubiquitous information and communication across multiple timeframes and geographies has expanded interaction with locations to include both physical and virtual interactions. Visitors can engage with a location pre and post-trip through online information search. Furthermore non-motorized travel access allows more direct exposure to the surrounding environment, possibly influencing sense of place positively relative to private modes, such as personal autos. To investigate these interrelationships, we conduct a visitor intercept survey and analyze the responses to capture the direction and magnitude SOP impacts on non-motorized site visit frequency. The estimation results indicate that while SOP statistically impacts non-motorized travel and visits, in general, the effects of ICT are less pronounced.

Yongsung Lee (Georgia Institute of Technology, United States)
Giovanni Circella (Institute of Transportation Studies, UC Davis, United States)
Patricia Mokhtarian (Georgia Institute of Technology, United States)
Subhrajit Guhathakurta (Georgia Institute of Technology, United States)
Are all Millennials "Urban Hipsters"? A Latent-Class Approach to Modeling Heterogeneous Residential Preferences among Millennials and Members of Generation X in California
SPEAKER: Yongsung Lee

ABSTRACT. The Millennial Generation, loosely defined in the literature as those who were born from 1980 to 2000, is reported to travel differently from previous generations when they were young. On average, Millennials wait longer to obtain a driver’s license, own fewer vehicles, drive or use cars less, make more trips by walking, biking, and public transit, and adopt emerging transportation services more (e.g., Uber/Lyft and Zipcar) (Blumenberg, Brown, Ralph, Taylor, & Voulgaris, 2015; Blumenberg, Ralph, Smart, & Taylor, 2016; Blumenberg et al., 2012; Blumenberg, Wander, Taylor, & Smart, 2013; Delbosc & Currie, 2013; Kuhnimhof, Zumkeller, & Chlond, 2013a, 2013b; McDonald, 2015; K. Ralph, Voulgaris, Taylor, Blumenberg, & Brown, 2016). In response, planners, policymakers, and journalists have speculated on the reasons for which the largest birth cohort of the moment shows unique travel patterns and decisions. After all, their behaviors should guide our efforts to predict future travel demand and invest in infrastructures in coming years. Scholars and practitioners point to several factors that explain the current choices of Millennials, such as economic factors and cultural factors (Delbosc & Ralph, 2017). Still, they do not yet properly factor in a critical element on which Millennials base their everyday decisions: people make various travel-related choices conditional on the attributes of their residence, workplace/school, or places they visit most often (Cao, Mokhtarian, & Handy, 2009; Mokhtarian & Cao, 2008). For example, even if claimed to be economically constrained and culturally oriented towards active modes, Millennials may not be able to walk, bike, or take public transit often if they live in built environments that do not support active modes of travel or provide access to transit services. Thus, instead of explaining their travel behaviors considering the home and work/school locations as exogenously given (Blumenberg et al., 2016; Buehler & Hamre, 2015; McDonald, 2015; K. M. Ralph, 2016), we should focus on better understanding the way in which Millennials choose their current locations, and how this eventually differs from the members of previous generations.

The existing studies on Millennials’ residential location choice focus on several causal factors such as temporal economic conditions, long-term changes in society, and cultural differences between Millennials and previous generations. Among other factors, Millennials tend to postpone marriage, childbearing, and homeownership while they pursue more education and struggle in the job market (Blumenberg et al., 2016; Fry, 2016; McDonald, 2015; Patten & Fry, 2015; Xu, Johnson, Bartholomae, O'Neill, & Gutter, 2015). Scholars who support an economy-centered explanation of Millennials’ behavior expect that once Millennials pay back student loans and earn higher incomes, they will likely switch to low-density suburban neighborhoods and auto-oriented travel patterns. However, the trend of delaying life course events and living in more urbanized areas began before the recent economic recession. Accordingly, some argue that it’s the long-term changes in society (e.g., the transition to a knowledge-based economy and the growing demand for high education) that make young adults achieve higher educational attainment, form their own households at a later point in life, and raise fewer children per household. Thus, many of today’s young adults do not need to buy a home in their 20s or early 30s. Instead, Millennials tend to stay at neighborhoods in central cities because they benefit from skill matching and skill spillovers while solving the colocation problem of “power couples” (Costa & Kahn, 2000; Duranton & Puga, 2004; Millsap, 2016; Peri, 2002). According to this long-term social-change explanation, even as Millennials age, more of them will likely stay in central cities. Further, a group of scholars suggests that values, perceptions, and preferences of Millennials differ from those of previous generations. Millennials prefer neighborhoods with convenient access to urban amenities more than previous generations (Coutour & Handbury, 2016; NAR, 2015; ULI, 2013), and emerging transportation services such as carsharing and on-demand ridehailing help them live in central cities without owning cars (Hallock & Inglis, 2015; Martin, Shaheen, & Lidicker, 2010). Thus, proponents of this culture-centered explanation claim that, even as Millennials succeed economically later in their lives, they will present residential location choices and travel patterns that differ from those of older birth cohorts. Since these explanations provide policy implications in substantively different directions, transportation scholars and professionals are seeking a more comprehensive understanding of whether Millennials’ urban resurgence as of today will be transformative in the future.

In this study, we focus on the third causal factor, the difference in preferences, to account for the residential location choices of Millennials, with members of Generation X (i.e., Gen Xers) as a comparison group. We see that few studies explore the impacts of preferences on travel behavior and location choices because of a lack of data on various qualitative characteristics of individuals. Moreover, while the previous studies focus more on differences across generations, we find few studies analyzing the differences within a generation. Given that Millennials are likely more heterogeneous than their predecessors, some of them may behave as stereotypical urban-oriented Millennials, while others make choices in similar ways to those of Gen Xers. In this context, the main research questions this study aims to answer are: To what extent do residential preferences differ within and between generations, controlling for socio-demographic and other differences? Specifically in what ways do those preferences differ?

To address this question, we employ the California Millennial Dataset (CMD), a dataset collected in fall 2015 with a detailed online survey that asked a wide set of questions to Millennials and Gen Xers residing at various locations across California. The research team adopted a quota sampling approach to obtain sufficient cases from all regions and neighborhood types in California, and developed a set of weights for the final sample (N=1,975) so that weighted analyses can be representative of the population in California. Among other groups of variables, the survey collected individuals’ information on: socioeconomic and demographic characteristics, attitudes and preferences about various topics in life, current travel behavior and mobility choices, and the locations of home and work or school. To measure individual attitudes and preferences, in the survey, we asked individuals to rate their level of agreement with 66 qualitative statements on a 5-point Likert-type scale. We factor-analyzed their responses and extracted 18 factors. Among these, this study employs six attitudinal factors: pro-suburban, car as a tool, pro-environmental policies, materialism, smartphone/internet enthusiasm, and pro-exercise.

As for residential and work/school locations, the participants reported information in the form of full or partial addresses in the survey. We cleaned, validated, and geocoded these addresses using the Google Maps Application Programming Interface (API). For accurate addresses, we identified the US Census block group in which the home and work/school for each participant is located, and we used this census geography to link individuals to their neighborhood attributes by appending information from various external sources. Neighborhood socioeconomic and demographic attributes came from the US Census 2015 American Community Survey 5-year estimate; neighborhood land-use attributes from the US Environmental Protection Agency Smart Location Database, the Google Maps API, and; and public school quality information from For various neighborhood-level attributes collected from various sources, we factor-analyzed them to extract underlying dimensions such as access to various businesses, the balance between jobs and housing, socioeconomic status, development density, and affordability and abundance of rental housing units.

For this paper, we employ a subsample of 902 commuters (either to workplace or school). We focus on regular commuters for whom we can generate hypothetical choice sets for the residential location assuming that they searched for their place of residence within a certain distance from their work/school (which makes the commute experience not too long or unpleasant). We also exclude cases with less accurate geocodes because we cannot ascertain the neighborhood attributes of their residences, and exclude Millennials living with their parents because the neighborhood where they live usually does not reflect the choice of these dependent Millennials, but that of their parents. Weighted summary statistics for the subsample reveal that, compared to Gen Xers, Millennials usually live with a higher number of young children under six, but have fewer older children. They have lower household income, include fewer homeowners, own fewer vehicles, but associated more importance to the use of smartphones and access to the internet.

We estimate a latent-class choice model (LCCM) of the residential location choice of these individuals. In the LCCM approach, two models are jointly estimated: a class membership model, and a model of outcome given class. The membership model estimates the probabilities of individuals belonging to one latent class or another through a multinomial logit model (MNL) with individual and household socioeconomic, demographic, and attitudinal characteristics as explanatory variables. The dependent variable of that model is an unobserved categorical variable indicating the latent class, or unobserved group, in the population. The outcome of interest is residential block group; accordingly the class-specific outcome models, which are also MNL, capture how the members of each class value various neighborhood attributes in their choices. One difficulty of residential location choice models is that researchers usually do not have information on the alternatives that are actually considered by each individual; only the actual choice is known. Accordingly, in line with the approaches proposed in the literature (McFadden, 1978; Salon, 2009), we assume that the workplace or school location is exogenously given to each individual, and build hypothetical choice sets through selecting nine random census block groups located within 45 miles from this location. Thus, including the block group of the actually chosen place of residence, each of the individuals in the sample has a choice set containing ten alternatives. Although the random selection of alternatives may not reflect the actual neighborhood search process that these individuals went through, this approach diversifies the neighborhood attributes of the alternatives in the choice set and helps reveal the residential preferences for each class.

After estimating LCCMs with a varying number of latent classes, we choose a four-class solution based on goodness-of-fit measures and interpretability of the model. We name the four classes by referring to certain socioeconomic, demographic, and attitudinal characteristics that are more pronounced among the members of that class but less so among the members of the other classes. Based on their residential location choice processes: we label the four groups as middle-class families (36.3%), pro-urban renters (29.2%), low-income workers (24.4%), and pro-suburban homeowners (10.2%). Individuals belonging to the middle-class family class tend to live with a partner and children, tend to have medium household income (i.e., between $60,000 to $120,000 per year), own their home and vehicle(s), and present pragmatic attitudes towards car ownership. In comparison, individuals belonging to the pro-urban renter class are more likely to live alone, have a part-time job, not own a home or vehicles, prefer urban lifestyles and are in favor of environmental policies regulating/limiting the use of cars. The members of the low-income worker class tend to be less educated, work full time, and have low household income. Those in the pro-suburban homeowner class are likely to be highly educated, earn more than $120,000 a year, own a home and vehicles, prefer suburban lifestyles, and are more materialistic. In addition, we see substantial variations in the share of the four classes from one age group to another: Younger Millennials are more often found to belong to the pro-urban renter class, while older Millennials more often follow the behavior of middle-class families.

Figure 1. The shares of four latent classes by age group

Compared to a pooled residential location choice model, the LCCM solution with four latent classes produces better goodness-of-fit measures, even under the penalties for its substantially larger number of parameters. The reason for this better fit is clear from examining the coefficients of the outcome models, most of which differ substantially across class. Here are some of the most relevant findings: Middle-class families are found to derive a higher utility for neighborhoods in good school districts than the other groups do; pro-urban renters tend to choose neighborhoods with convenient access to various businesses and affordable rental units more than the other groups do; low-income workers are more likely to locate in moderately dense neighborhoods in the suburbs than elsewhere, all else being equal, than the other groups do; and pro-suburban homeowners value a short commute distance to their work/school and a good balance between jobs and housing units in a neighborhood more than the other groups do.

This paper contributes to the literature about the urban lifestyles of Millennials in several ways, such as modeling residential location choice with a rich set of explanatory variables, identifying unobserved groups of individuals in the population with heterogeneous residential preferences, and exploring the extent to which life course events (e.g., full-time employment, marriage, childbearing, and homeownership) and attitudes (e.g., pro-urban and pro-environment) account for their membership and preferences.

Note that we analyze a cross-sectional dataset and cannot claim any longitudinal changes with certainty. However, based on our findings, we speculate that many pro-urban renters today will switch to middle-class families tomorrow as they establish their own households, raise children, earn a higher income, and buy a home and vehicles. The changes in their socioeconomic situations may affect their preferences for urban lifestyles: i.e., they may become less pro-urban, but more pro-suburban. Given that the data collection for the 2nd wave of CMD is scheduled for the spring of 2018, we plan to examine any longitudinal changes of the same individuals in the sample in later studies.

References Blumenberg, E., Brown, A., Ralph, K., Taylor, B. D., & Voulgaris, C. T. (2015). Typecasting Neighborhoods and Travelers: Analyzing the Geography of Travel Behavior among Teens and Young Adults in the US. Retrieved from Blumenberg, E., Ralph, K., Smart, M., & Taylor, B. D. (2016). Who knows about kids these days? Analyzing the determinants of youth and adult mobility in the US between 1990 and 2009. Transportation Research Part A: Policy and Practice, 93, 39-54. doi: Blumenberg, E., Taylor, B. D., Smart, M., Ralph, K., Wander, M., & Brumbagh, S. (2012). What's Youth Got to Do with It? Exploring the Travel Behavior of Teens and Young Adults. Retrieved from Blumenberg, E., Wander, M., Taylor, B. D., & Smart, M. (2013). The Times Are They A-Changin’? Youth, Travel Mode, and the Journey to Work. Paper presented at the Transportation Research Board 92nd Annual Meeting, Washington, D.C. Buehler, R., & Hamre, A. (2015). The multimodal majority? Driving, walking, cycling, and public transportation use among American adults. Transportation, 42(6), 1081-1101. doi: Cao, X., Mokhtarian, P. L., & Handy, S. L. (2009). Examining the impacts of residential self‐selection on travel behaviour: a focus on empirical findings. Transport Reviews, 29(3), 359-395. doi: Costa, D. L., & Kahn, M. E. (2000). Power couples: changes in the locational choice of the college educated, 1940–1990. The Quarterly Journal of Economics, 115(4), 1287-1315. doi: Coutour, V., & Handbury, J. (2016). Urban revival in America, 2000 to 2010 Paper presented at the Research Symposium on Gentrification and Neighborhood Change, Philadelphia, PA. Delbosc, A., & Currie, G. (2013). Causes of youth licensing decline: a synthesis of evidence. Transport Reviews, 33(3), 271-290. doi: Delbosc, A., & Ralph, K. (2017). A tale of two millennials. Journal of Transport and Land Use, 10(1). doi: Duranton, G., & Puga, D. (2004). Micro-foundations of urban agglomeration economies. Handbook of regional and urban economics, 4, 2063-2117. doi: Fry, R. (2016). For First Time in Modern Era, Living With Parents Edges Out Other Living Arrangements for 18- to 34-Year-Olds. Retrieved from Hallock, L., & Inglis, J. (2015). The Innovative Transportation Index: The Cities Where New Technologies and Tools Can Reduce Your Need to Own a Car. Retrieved from Kuhnimhof, T., Zumkeller, D., & Chlond, B. (2013a). Who Are the Drivers of Peak Car Use? Transportation Research Record: Journal of the Transportation Research Board, 2383(7), 53-61. doi: Kuhnimhof, T., Zumkeller, D., & Chlond, B. (2013b). Who made peak car, and how? A breakdown of trends over four decades in four countries. Transport Reviews, 33(3), 325-342. doi: Martin, E., Shaheen, S., & Lidicker, J. (2010). Impact of carsharing on household vehicle holdings: Results from North American shared-use vehicle survey. Transportation Research Record: Journal of the Transportation Research Board, 2143(19), 150-158. doi: McDonald, N. C. (2015). Are millennials really the “go-nowhere” generation? Journal of the American Planning Association, 81(2), 90-103. doi: McFadden, D. (1978). Modeling the choice of residential location. Transportation Research Record(673). Millsap, A. (2016). Location choice in early adulthood: Millennials versus Baby Boomers. Papers in Regional Science. Mokhtarian, P. L., & Cao, X. (2008). Examining the impacts of residential self-selection on travel behavior: A focus on methodologies. Transportation Research Part B: Methodological, 42(3), 204-228. doi: NAR. (2015). Community & Transportation Preferences Survey U.S. Metro Areas, 2015. Retrieved from Patten, E., & Fry, R. (2015, MARCH 19, 2015). How Millennials today compare with their grandparents 50 years ago. Retrieved from!0 Peri, G. (2002). Young workers, learning, and agglomerations. Journal of Urban Economics, 52(3), 582-607. doi: Ralph, K., Voulgaris, C. T., Taylor, B. D., Blumenberg, E., & Brown, A. E. (2016). Millennials, built form, and travel insights from a nationwide typology of U.S. neighborhoods. Journal of Transport Geography, 57, 218-226. doi: Ralph, K. M. (2016). Multimodal Millennials? The Four Traveler Types of Young People in the United States in 2009. Journal of Planning Education and Research, 0739456X16651930. doi: Salon, D. (2009). Neighborhoods, cars, and commuting in New York City: A discrete choice approach. Transportation Research Part A: Policy and Practice, 43(2), 180-196. doi: ULI. (2013). Americans' Views on their Communities, Housing, and Transportation: Analysis of a national survey of 1,202 adults, 1-60. Retrieved from America in 2013: A ULI Survey of Views on Housing, Transportation and Community website: Xu, Y., Johnson, C., Bartholomae, S., O'Neill, B., & Gutter, M. S. (2015). Homeownership among millennials: The deferred American dream? Family and Consumer Sciences Research Journal, 44(2), 201-212. doi:10.1111/fcsr.12136

Ben Clark (University of the West of England, Bristol, UK)
Kiron Chatterjee (University of the West of England, Bristol, UK)
In search of the ideal commute: The dynamics of commuting, wellbeing and job and housing choices
SPEAKER: Ben Clark

ABSTRACT. Context and study aims Commuting is performed by the labour force, a significant proportion of national populations, on a regular basis. It has been argued that commuting behaviours are habitual since they are performed on a routine basis with little or no conscious consideration of alternative ways of travelling to work. On the other hand, the nature of the commute journey (its origin and destination, the journey schedule and frequency) is derived from longer term lifestyle decisions concerning where to live and type of employment and its location. It has been shown that commuting behaviours are more likely to change at the time of life events, particularly those that alter the origin/destination of the journey to work like moving home and changing jobs (Clark et al 2016). There is growing interest in the impacts of commuting for personal wellbeing since any wellbeing benefits or dis-benefits of commuting will affect a large proportion of the population, implying a significant public health issue. Whilst there is evidence of associations between commuting behaviour and different aspects of personal wellbeing (summarised below), the contribution of this paper is to provide a specific focus on dynamics, i.e. the processes and outcomes of behaviour change. The empirical analysis addresses two objectives: (i) to examine the effects of changes to commuting behaviour on different aspects of wellbeing and to identify whether wellbeing impacts grow or diminish over time; and (ii) to examine whether, and if so how, people adjust their (commuting) situations over time in response to arduous commutes associated with low wellbeing, e.g. by moving home or changing jobs. These objectives are addressed through an analysis of a large scale six wave panel data set drawn from the UK Household Longitudinal Study (UKHLS). Evidence on commuting and subjective wellbeing Subjective wellbeing (SWB) is a broad concept that refers to an individual’s evaluation of how well their life is going. It encompasses (i) frequency of emotional responses, (ii) self-evaluations of satisfaction with life overall and (iii) whether individuals feel they are fulfilling their potential. There are theoretical mechanisms through which commuting could impact on different aspects of SWB. It may be stressful and adversely affect mood during and after the journey. It may consume time and money that workers would rather spend on other activities. On the positive side, a commute may be relaxing and provide time to switch off and if it involves physical activity it could be appreciated for its health benefits. Indeed, a number of previous panel studies have provided evidence of an association between commuting and subjective wellbeing. Stutzer and Frey (2008) found for a sample of workers in Germany, that longer duration commutes were associated with lower life satisfaction, although this was not found to be the case for workers in the UK (Dickerson et al 2014). Longer duration commutes have also been found to be associated with lower satisfaction with leisure time availability (Dickerson et al 2014) and worse mental health (Roberts et al 2011), with the effects particularly pronounced for women (Roberts et al 2011). Walking to work and using the bus have been found to be associated with better mental health than commuting by car (Martin et al 2014). A small number of studies have examined the relationship between the commute and likelihood of changing job or home. Van Ommeren (1996) found from an empirical analysis of Telepanel (1992) Dutch data that every additional 10 kms of commuting distance decreased the expected duration of the current job and current residence by more than two years. Clark et al. (2003) found that households are more likely to decrease commute distance and time in association with home moves (accompanied by job change or not). They estimated a critical value of 8km as the commute distance beyond which the likelihood of decreasing commute distance increases rapidly. Knowledge gaps Although these studies provide evidence of a relationship between commuting and wellbeing, it is important to examine the dynamic nature of this relationship, for three reasons. First, establishing whether a relationship is causal requires it to be demonstrated that the cause – a commuting behaviour change, happens before the effect - a change in SWB. Second, identifying whether a long-term outcome differs from a short-term outcome is necessary for appreciating the long-term consequences of changes in commuting behaviour. For example, it is important to know if a change in SWB endures or is short-lived. And third, examining the process of change over time provides informative insights into how commuters adjust their personal circumstances in relation to their goals and experiences as they move through their lives.


The objectives were addressed through an analysis of a sample of over 26,000 employed people living in England using data from the UKHLS. Six waves of data from the UKHLS were available, covering 2009/10 to 2014/15. The UKHLS captures two measures of commuting every wave: (i) The one way door to door journey time (in minutes) and (ii) the mode of transport normally used to get to work. The UKHLS also captures multiple measures of SWB every wave, including job satisfaction, leisure time satisfaction, self-reported health, strain, mental health and life satisfaction. Analytical approach An incremental approach was taken to build understanding of the nature of the dynamic relationship between commuting, wellbeing and job/home changes. Firstly, the six-wave sample was pooled and the prevalence of wave to wave changes in commute mode, commute time, and changes to the origin (home location) and destination (job location) of the commute journey were examined. This provided the basis for the development of three different models: (i) Conditional change score models, designed to examine the short term effects of behaviour changes on SWB. These models took the change in wellbeing score (for each individual) from one wave to the next as the dependent variable. This was estimated as a function of the base year wellbeing score (since the change in wellbeing score from one wave to the next is conditional on the base score), the commute duration in the base wave, the change in commute duration by the following wave, the change in commute mode by the following wave and various control variables. (ii) Lagged regression models, designed to examine whether the size of the effect of a commuting change on SWB alters up to three years after the change occurs. These models took the wellbeing score as the dependent variable. This was estimated as a function of binary variables to indicate whether a given commuting change (e.g. a mode switch or a change in commute time category) had occurred in the current wave (within 0-12 months), in the previous waves (12-24 months ago), or two waves ago (24-36 months ago). (iii) Binary logit models, designed to examine whether adjustments in life situation are more likely in response to arduous commutes and low wellbeing. These models took a binary variable to indicate change of job or housing as the dependent variable. This was modelled as a function of ‘stressor’ factors that may be expected to alter the probability of the job or housing change occurring, including: base year income, base year measures of wellbeing (such as job satisfaction which might be expected to increase or decrease the probability of changing job), base year commute mode, base year commute duration and control variables.

Key findings

In relation to the prevalence of behaviour change: 18 percent of observations were found to involve a change in commute mode from one year to the next; 22 percent of observations involved a change in commute time category between short (up to 15 minutes), medium (16 to 45 minutes) and long (over 45 minutes) duration commutes; and 20 percent of cases involved a change in either the origin or destination of the commute (corresponding to a home or job location change). In relation to short term effects of these behaviour changes, the conditional change score models indicated that changing from a short to a long duration commute is associated with concurrent wellbeing disbenefits. This included a reduction in job and leisure time satisfaction, an increase in strain, and worsening mental health. Switching from driving to walking/cycling was associated with increased job and leisure time satisfaction, and a reduction in strain. With respect to the lagged effects of behaviour changes on wellbeing, the analysis indicated that the full negative effects on leisure time satisfaction of switching from a shorter to a longer duration commute are not felt within the first 12 months and the effect increases 12 months after the change. This suggests that people take on longer duration commutes without initially realising the full impact or appreciating other benefits that over time become less salient. Finally, the models indicated evidence of feedback between wellbeing and changes in life situation. For example, they showed lower job satisfaction and being a long duration commuter in the base year increases the likelihood of changing jobs by the following year, with the latter increasing the likelihood by around 25%. This implies that people seek to avoid arduous commutes associated with lower wellbeing by changing jobs (Figure 1).


The results imply that longer duration commutes are maintained as long as the benefits of higher income and a satisfying job outweigh the drawbacks of the commute journey. If not, people are more likely to alter their situation by changing jobs. In other words, there is dynamic feedback between long duration commuting, job satisfaction and earnings and the propensity to change jobs.

Figure 1: Dynamic feedback between commute duration, job satisfaction and likelihood of changing jobs

References Clark, B., Chatterjee, K. and Melia, S. (2016) Changes to commute mode: The role of life events, spatial context and environmental attitude. Transportation Research Part A: Policy and Practice, 89. pp. 89-105. ISSN 0965-8564. Clark, Huang and Withers (2003) Does Commuting Distance Matter? Commuting Tolerance and Residential Change, regional Science and Urban Economics 33, 199-221, 2003 Dickerson, A., Hole, A., Munford, L. (2014) The relationship between well-being and commuting revisited: Does the choice of methodology matter? Regional Science and Urban Economic, 49, pp.321-329. Martin, A., Goryakin, Y., Suhrcke, M. (2014) Does active commuting improve psychological wellbeing? Longitudinal evidence from eighteen waves of the British Household Panel Survey. Preventative Medicine, 69, pp.296-303. Roberts, J., Hodgson, R., Dolan, P. (2011) “It’s driving her mad”: Gender differences in the effects of commuting on psychological health. Journal of Health Economics, 30, pp.1064-1076. Stutzer, A., and Frey, B. (2008) Stress that doesn’t pay: The commuting paradox. The Scandinavian Journal of Economics, 110(2), pp.339-366. Van Ommeren, J., Rietveld, P., Nijkamp, P. (1997) Commuting: In search of jobs and residences. Journal of Urban Economics 42, p.402-421.

Jonas De Vos (Geography Department, Ghent University, Belgium)
Do people travel with their preferred travel mode? Analysing the extent of travel mode dissonance and its effect on travel satisfaction

ABSTRACT. Numerous studies have indicated that travel mode choice is affected by travel-related attitudes. A positive stance towards a certain travel mode increases the probability that people will choose this mode for a particular trip. However, not a lot of studies have analysed whether people actually choose their preferred travel mode. In this paper we will look at whether respondents with a preference for car use, public transport use, cycling and walking will actually use these modes. Furthermore, we also analyse whether respondents who use their preferred travel mode (i.e., consonant travellers) are more satisfied with their trips compared to respondents travelling with a non-preferred travel mode (i.e., dissonant travellers). Results from this study, analysing leisure trips of 1,656 respondents from the city of Ghent (Belgium), indicate that about half of the respondents chooses a non-preferred travel mode and that dissonant travellers can be mainly found within public transport users and least within cyclists, partly due to relatively low levels of public transport attitudes and high levels of cycling attitudes. Furthermore, travel mode dissonance seems to have an important impact on travel satisfaction. Consonant travellers have above average travel satisfaction levels, independent of the used travel mode, while dissonant travellers (except dissonant pedestrians) have below average travel satisfaction levels. This suggests that using a preferred travel mode has at least an equally important impact on travel satisfaction than the chosen travel mode itself.

12:30-13:30Lunch Break
13:30-15:30 Session 11A: Mobility as a Service -- Future
Mark Bradley (RSG, United States)
Location: Corwin West
Laura Gebhardt (German Aerespace Center, Germany)
Rebekka Oostendorp (German Aerespace Center, Germany)
From urban mobility practices, strategies and logics of actions to future mobility solutions - an user-centered mixed-methods approach

ABSTRACT. When examining and discussing future development paths of urban mobility we have to deal with complex transformation processes that do not only involve technical and organizational challenges but also include questions regarding the people who are the traffic participants in the urban mobility system. Recent studies have shown that evaluation and acceptance of users towards technological innovations (such as autonomous driving) are bound to their current behavior and how people ascribe meaning to their travel behavior (Fraedrich and Lenz 2014; Watson 201; Zmud and Sener 2017). If we assume that it is not only a technology that is responsible for a transformation in society but rather how a technology is linked and embedded to specific daily life practices, we should target a comprehensive knowledge about daily life practices of the users to better understand sociotechnical transformation processes in relation to technical innovations (e.g. autonomous driving). Hence, it is meaningful to approach users’ perspectives and to understand their mobility practices, underlying reasons for travel behavior and user requirements.

Against this background and based on empirical findings we want to answer the following questions:

• What kind of mobility practices, strategies and logics of action do people representing different mobility user types pursue when being mobile in the city? • What requirements do different mobility types have concerning urban mobility?

In our contribution, mobility is not only understood as an act of physical movement, but as a social phenomenon, rooted in the reality of the people (Eberle 2000; Miebach 2006). For this reason, mobility research is committed to an action-theoretical approach, regarding spatial mobility as a social activity, understanding mobility from the user perspective and knowing how important constructions of mobility are. This is connected to the assumption of action-oriented mobility research that actors produce and reproduce mobility under the conditions of their everyday reality (Hannam et al. 2006). That is the reason why everyday reality is the starting point for our exploration of the practices of people in cities and the related underlying logics of action of different mobility types. This demands for inter- and transdisciplinary approaches as well as the application of adjusted research methods (Fraedrich and Lenz 2014), which allow to capture and understand mobility as a socio-technical phenomenon and to develop – taking that into account – appropriate mobility services and innovative technical solutions for urban mobility.

User group segmentations represent the possibility of reducing the complexity of heterogeneous populations by identifying homogeneous subgroups (Hunecke 2015). They are an established methodological means in social sciences for analyzing daily travel determinants (Bartz 2015; Prillwitz and Barr 2011) and are used by different disciplines, also increasingly in transport sciences (e.g. Haustein and Nielson 2016; Wittwer 2014; Vij et al. 2011). One advantage of segmentation approaches relates to improvingthe possibilities for communication between scientists of different disciplines and practitioners in reducing the complexity of heterogeneous populations (Hunecke 2015; Hunecke and Haustein 2007).

To answer the above-mentioned questions a mixed-methods approach has been chosen with a focus on qualitative methods. Research in these ‘softer’ or more ‘intangible’ topics often demands survey methods and approaches that are either completely new in their design or introducing methods that are mostly unfamiliar within mainstream travel survey methods (Carrasco and Lucas 2015). In our case we identified, developed and sketched different mobility types in a multistage process: On one side the typology is the result of a factor and cluster analysis made with empirical data from a survey conducted in Berlin, Germany. (n= 1.098). It is based on the frequency and trip purposes of used modes and mode combinations and thereby reflects the users’ travel behavior. The identified mobility types can further be characterized by their socio-demographic characteristics and available mobility resources. On the other side the picture of the mobility types was drawn in more detail by qualitative methods to get information on the underlying motives, preferences and requirements leading to the observed travel behavior. More precisely: we conducted narrative interviews and group discussions with representatives of different mobility types where we addressed the underlying motives and requirements of the users for their urban travel behavior. We complemented the group discussions with creative methods and visual elements (Christiansen 2005; Cooper et al. 2007; Degele et al. 2009; Haper 2003; Rhinow et al. 2012) to develop and discuss new ideas and prototypes of vehicles for potential future development paths of urban mobility together with the participants.

Although qualitative methods are increasingly used within transportation research, as a complement to more established quantitative surveys, their potential is often still underrated or poorly promoted (Carrasco and Lucas 2015; Grosvenor 2000). Qualitative methods are particularly suitable to cover aspects that allow determining subjective acknowledgements to a specific topic but they also make it possible studying the role of psychological and social factors to determine people’s travel behaviors and choices (Carrasco and Lucas 2015). And, in our particular case they help to find out in detail about the mobility strategies, logics of action and requirements that define and structure the individuals’ practices.

The paper brings together results from the quantitative survey with the findings of the qualitative co-creation workshops and narrative interviews with different mobility types. Through the chosen mixed-methods approach the variety of practices, strategies and requirements for being mobile in the city of different mobility types are highlighted. For example, several mobility types drive a car, but in different ways and due to different motivations: The “all-purpose car user” is acting according to his personal preferences whereas the “multimodal user” acts much more purposefully and pragmatically. In his everyday life, the all-purpose car user never weighs different means of transport against each other, but has basically opted for his car while the “multimodal user” makes his choice of transport spontaneously and according to his purposes. The mobility types also differ by their socio-demographic characteristics and available mobility resources. For example, the “all-purpose car user” is more likely male, older than average, has a car always available but no public transport pass. By contrast, the “multimodal user” is more likely female, younger than average, has a car and also a public transport pass. It is not surprising that also the prototypes of vehicles that were developed within the workshops differ between the mobility types. The group of “all-purpose car users” developed an individual vehicle with various options for the personal comfort (such as voice control or interior depending on the personal mood), whereas the group of “multimodal users” design a modular-vehicle with a minimal of functions and no comfort, which is shared and thus improves the situation of the transport system in the city.

The presented mixed-methods analyses provide insights in the variety of users and their practices, strategies, logics of action and related requirements. The chosen methods contribute to the understanding of the fundamental elements of travel behavior in cities. Furthermore, the chosen approach allows the development of future mobility solutions on the basis from user’s mobility practices and requirements for moving in the city. Due to the dominance of quantitative methods in transport research, findings of qualitative studies are often not presented and discussed in traditional transport journals and conferences until now. With our presentation we want to encourage an intensified debate and knowledge sharing between quantitative and qualitative transport researchers with mutual interests in understanding the influence of social and psychological factors on people’s travel choices.

Bartz, F. M. (2015): Mobilitätsbedürfnisse und ihre Satisfaktoren. Die Analyse von Mobilitätstypen im Rahmen eines internationalen Segmentierungsmodells. Dissertation. Humanwissenschaftlichen Fakultät der Universität zu Köln.

Carrasco, J.-A. and Lucas, K. (2015): Workshop synthesis: Measuring attitudes; quantitative and qualitative methods. In: Transportation Research Procedia, 11, pp. 165-171.

Christiansen, E. (2005): Boundary objects, please rise! On the role of boundary objects in distributed collaboration and how to design for them. Paper presented at Workshop 10‚ ‘Cognition and Collaboration‘ on Conference for Human Computer Interaction (CHI 2005), April 2-7, 2005, Portland, Oregon.

Cooper, A.; Reimann, R. and Cronin, D. (2007): About Face 3: The Essentials of Interaction Design. New York: Wiley Publishing.

Degele, N. ; Kesselhut, K. and Schneickert, C. (2009): Sehen und Sprechen: zum Einsatz von Bildern bei Gruppendiskussionen. In: Zeitschrift für Qualitative Forschung, 10 (2), pp. 363-379.

Eberle, T. S. (2000): Lebensweltanalyse und Handlungstheorie: Beiträge zur verstehenden Soziologie. Konstanz: Universitätsverlag.

Fraedrich, E. and Lenz, B. (2014): Automated Driving – Individual and Societal Aspects. In: Transportation Research Record: Journal of the Transportation Research Board, Vol. 2416 (2), pp. 64-72.

Grosvenor, T. (2000): Qualitative Research in the Transport Sector. Resource paper for the Workshop on Qualitative/Quantitative Methods. In: Transportation Research Board, Transport Surveys: Raising the Standard. Proceedings of an International Conference on Transport Survey Quality and Innovation, May 24-30, 1997, Grainau, Germany. Transportation Research Circular E-C008, Washington, DC, USA, II-K/1-18.

Hannam, K.; Sheller, M. and Urry, J. (2006): Editorial: Mobilities, lmmobilities and Moorings. In: Mobilities, 1 (1), pp. 1-22.

Haper, D. (2003): Fotografien als sozialwissenschaftliche Daten. In: Flick, U.; Kardorff, E.; von Steinke, I. (Hrsg.): Qualitative Forschung. Ein Handbuch. Reinbek bei Hamburg: Rowohlt, pp. 402-415.

Haustein, S. and Nielson, S. T. (2016): European mobility cultures: A survey-based cluster analysis across 28 European countries. In: Journal of Transport Geography, 54, pp. 173-180.

Hunecke, M. (2015): Ansätze zur Segmentierung von NutzerInnengruppen. In: Mobilitätsverhalten verstehen und verändern. Studien zur Mobilitäts- und Verkehrsforschung. Wiesbaden: Springer VS, pp. 47-74.

Hunecke, M. and Haustein, S. (2007): Einstellungsbasierte Mobilitätstypen: Eine integrierte Anwendung von multivariaten und inhaltsanalytischen Methoden der empirischen Sozialforschung zur Identifikation von Zielgruppen für eine nachhaltige Mobilität. In: Umweltpsychologie, 11 (2), pp. 38-68.

Miebach, B. (2006): Soziologische Handlungstheorie: Eine Einführung. Wiesbaden: Springer VS.

Prillwitz, J. and Barr, S. (2011): Moving towards sustainibility? Mobility styles, attitudes and individual travel behaviour. In: Journal of Transport Geography, 19, pp. 1590-1600.

Rhinow, H.; Köppen, E. and Meinel, C. (2012): Prototypes as Boundary Objects in Innovation Processes. Conference Paper in the Proceedings of the 2012 International Conference on Design Research Society (DRS 2012), Bangkok, Thailand.

Vij, A.; Carrel, A. and Walker, J., L. (2011): Capturing modality styles using behavioral mixture models and longitudinal data. Paper presented at 2nd international choice modelling conference, Leeds.

Watson, M. (2012): How theories of practice inform transition to a decarbonised transport system. In: Journal of Transport Geography, 24, pp. 488-496.

Wittwer, R. (2014): Zwangsmobilität und Verkehrsmittelorientierung junger Erwachsener: eine Typologisierung. Schriftenreihe des Instituts für Verkehrsplanung und Straßenverkehr der Technischen Universität Dresden, Heft 16/2014.

Zmud, J. P. and Sener, I. N. (2017): Towards an understanding of the travel behavior impact of autonomous vehicle. In: Transportation Research Procedia, Vol. 25, pp. 2500-2519.

Michael Tanko (KTH Royal Institute of Technology, Sweden)
Harsha Cheemakurthy (KTH Royal Institute of Technology, Sweden)
Susanna Hall Kihl (KTH Royal Institute of Technology, Sweden)
Water-based public transport passenger perceptions and planning factors: A Swedish perspective
SPEAKER: Michael Tanko

ABSTRACT. Introduction and background

While once playing a major role in cities, transport by waterways has since declined with inland migration and the popularity of land-based transport. However, as land-based transport increasingly reaches capacity and cities begin to redevelopment their urban waterfronts, there are opportunities for the reintroduction of waterways in a transport role. Furthermore, there are also potential benefits in reducing other transport issues such as air pollution, noise and accidents. Cities such as London, New York, Hamburg and Brisbane have created urban water transit networks that work in parallel with existing public transport modes, while other cities are also in the process of developing similar systems.

There currently exists, however, limited research on the development of these transport networks. Current literature in the area has focused on issues of planning and land use implications (Weisbrod and Lawson 2009, Thompson et al 2006, Tanko & Burke 2015, Tanko & Burke 2017), economic benefits and property value effects around terminals (New York City Economic Development Corporation 2013, Tsai et al. 2015), and some preliminary studies looking at passenger travel patterns (Soltani et al. 2015, Rahman et al. 2016). However, specific research into traveller attitudes and perceptions toward water transit and how this may affect passenger experiences and mode choice is so far lacking.

Such research incorporating traveller attitudes is becoming increasing common in other transport modes in order to unpack the reasons for travel behaviour. It has been shown that inclusion of non-traditional variables (such as comfort, perception of safety etc) in the travel experience (further to travel time and cost) can have a significant impact on passenger’s subjective evaluation of the experience and hence their choice of travel mode (see Morikawa & Sasaki 1998, Ben-Akiva et al. 2002, Morikawa et al. 2002, Johansson et al. 2006). Travel for the sake of enjoyment apart from the destination has also challenged the concept of travel as a derived demand, suggesting benefits may accrue to passengers from the travel experience itself (Mokhtarian & Salomon 2001).

Water transport may offer such benefits, such as environmental factors from a scenic waterborne journey or increased comfort on board vessels. Such benefits may add value to the passenger experience and can be influential on mode choice. In a study of the San Francisco San Francisco Bay Area Ferry it was found that individual users favoured specific traits of ferries, such as comfort of vessel and safety, as well as traditional variable such as time and cost (Outwater et al. 2003). In a study of how comfort and safety perceived by inland waterway transport passengers in Colombia, it was found that the latent variable comfort, resulting from indicators space between seats and behaviour of other users, may increase the perceived utility of passengers, in particular younger passengers and those with a higher education level (Marquez et al. 2014). A survey of East River Ferry commuters in New York found that commuters prefer ferries because it is a less stressful way to commute and less crowded than the subway. Other benefits identified include travel time savings, safety benefits and general comfort benefits to passenger (New York City Economic Development Corporation 2012). A recent comparative study of boat and boat users in Brisbane by the authors also found that passengers choose boat journeys despite longer travel times and therefore display excess travel. A hypothesis is that either aesthetic or productivity benefits accrue to passengers, which is the basis of the current research project, and parallel studies being conducted currently in London and Brisbane.

In Nordic conditions there have only been a few research examples. In a study of boat passengers in the Norway archipelago, it was found that working commuters preferred service quality improvements such as comfort and on-board amenities as the preferred focus of service improvements (Mathison & Solvoli 2010). In a small study of boat commuters in Stockholm it was found in addition to reduced travel time, guaranteed seating and service (cafeteria, toilet and WiFi) were significant factors in people choosing to use boats, suggesting that incorporating these factors in mode choice models may be important to predict travel choices, particularly over longer distances (Van Berlekom 2014).

Stockholm County Council has recently expressed interest in expanding its existing modest urban passenger water transit service into a city-wide, year-round service. The intention of this paper is to expand the knowledge of passenger perceptions toward water transport, specifically in the context of a Nordic city with weather and other constraints to shed light on the factors that affect mode choice of users in Stockholm to aid in the planning of the future water transport network. This work is being conducted as part of a larger research effort of water transport development in Sweden, with research into water transport route design, land use and ferry oriented development (FOD) planning, and vessel design being also being studied at the KTH Centre for Naval Architecture in Stockholm.


The study design included a survey of boat users in Stockholm to gather their opinion on what factors contribute to their choice of using boat transport. The survey was completed via distributed surveys on board route 82 vessels on the Nacka route (Figure 1). We apply the theory of planned behaviour (Icek 1991) to conceptualize the decision-making process of passengers to understand contributing factors. People's action in this case is dependent on their intention which is a factor of attitudes, perceived control, social norms, fulfilling personal needs and the contribution of specific land use in the area. Land use implications were firstly addressed measuring the relative accessibility of ferry piers to passengers to assess how much this contributed to a passenger’s choice to use the boat service. Demographic details were also gathered. Personal attitudes toward the journey are gathered through a questionnaire addressing key factors that contribute to travel choice. The survey is based on previous studies by Eboli and Mazzulla 2007 and Eboli and Mazzulla 2010. Passengers were asked a series of questions on a Likert scale from Strongly Agree to Strongly Disagree about the importance of service attributes and the satisfaction with each factor. As noted in the literature review a hypothesis of this research is that latent variables such as comfort and productivity benefits may contribute to mode choice. Therefore, a structural equation modelling (SEM) approach was taken which allows assessment of latent variables after factor analysis. Specifically, a coefficient (C-SEM) was used as this method is more suitable to the testing of a hypothesis as opposed to a Partial Least Squares (PLS-CEM) approach, more suitable to forecasting applications. The total sample size was 722 responses gathered across a two-week period incorporating peak and off peak, and weekday and weekend travel. The analysis is currently ongoing to specify the SEM model and discern the relative importance of service factors. Our expected results are identification of the key factors that boat passengers consider in line with previous research on other modes, which have been widely studied. As mentioned this research is part of a wider program of the planning of water transport in Stockholm, and this initial start on passenger’s perceptions is expected to form a key basis for future work on the planning of new routes and terminals, as well as the design of vessels incorporating the results of this study.

Figure 1 Route 82 boat in Stockholm Source: Sjövägen 2013


Ben-Akiva, M., D. McFadden, K. Train, J. Walker, C. Bhat, M. Bierlaire, D. Bolduc, A. Boersch-Supan, D. Brownstone, D. S. Bunch, A. Daly, A. de Palma, D. Gopinath, A. Karlstrom, and M. A. Munizaga 2002 Hybrid Choice Models: Progress and Challenges. Marketing Letters, Vol. 13, No. 3, 2002, pp. 163-175. Eboli, L & Mazzulla, G. 2007. Service Quality Attributes Affecting Customer Satisfaction for Bus Transit. Journal of Public Transportation, 10 (3): 21-34. DOI: Available at: Eboli, L & Mazzulla, G (2010) How to Capture the Passengers’ Point of View on a Transit Service through Rating and Choice Options, Transport Reviews, 30:4, 435-450, DOI: 10.1080/01441640903068441 Icek, A 1991 The Theory of Planned Behaviour. Organizational Behaviour and Human Decision Processes 50(2):179-211 · December 1991 DOI: 10.1016/0749-5978(91)90020-T Johansson, MV, Heldt, T, Johansson P The effects of attitudes and personality traits on mode choice Transportation Research Part A – Policy and Practice, 40 (2006), pp. 507-525 Marquez, Luis and Cantillo, Viictor and Arellana, Julian 2014. How are comfort and safety perceived by inland waterway transport passengers? Transport Policy Vol 36 p 46-52. Mathisen, T. A. & Solvoll, G.2010 Service quality aspects in ferry passenger transport-examples from Norway. Delft University of Technology. Mokhtarian, P.L. & Salomon, I., 2001. How derived is the demand for travel? Some conceptual and measurement considerations. Transportation Research Part A: Policy and Practice, 35(8), pp.695–719. Morikawa,T., Sasaki, K., 1998 Discrete choice models with latent variables using subjective data. In: Ortúzar, J. de D., Hensher, D.A., Jara-Diaz, S. (Eds.), Travel Behavior. Research: Updating the State of Play. Pergamon, Oxford, pp. 435-455. Morikawa, T., Ben-Akiva, M., McFadden, D., 2002 Discrete Choice Models Incorporating Revealed Preferences and Psychometric Data. Econometric Models in Marketing Advances in Econometrics: A Research Annual, vol. 16. Elsevier Science Ltd New York Economic Development Corporation 2012. East River Ferry Service Summer Survey Executive Summary & Methodology. Performed by Business Research & Economic Advisors for NYC Economic Development Corporation July 2012 New York City Economic Development Corporation, 2013. Waterfront Action Agenda Transforming New York City’s Waterfront, New York. Outwater, M.; Castleberry, S.; Shiftan, Y.; Ben-Akiva, M.; Shuang Zhou, Y. & Kuppam, A. 2003 Attitudinal market segmentation approach to mode choice and ridership forecasting: Structural equation modeling. Transportation Research Record: Journal of the Transportation Research Board, Transportation Research Board of the National Academies, 2003, 32-42 Rahman, S., Wong, J. & Brakewood, C., 2016. Use of Mobile Ticketing Data to Estimate an Origin–Destination Matrix for New York City Ferry Service. Transportation Research Record: Journal of the Transportation Research Board, 2544(2544), pp.1–9. A Sjövägen 2013 , Sjövägen Available at Soltani, A., Tanko, M., Burke, M. I., Farid, R. 2015. Travel Patterns of Urban Linear Ferry Passengers: Analysis of Smart Card Fare Data for Brisbane, Australia. Transportation Research Record: Journal of the Transportation Research Board Tanko, M. & Burke, M., 2015. Innovation and transport planning: Introducing urban linear ferries in Brisbane. In P. Burton & H. Shearer, eds. State of Australian Cities Conference 2015: Refereed Proceedings. Gold Coast: Urban Research Program at Griffith University on behalf of the Australian Cities Research Network Tanko M and Burke M 2017, Transport innovations and their effect on cities : the emergence of urban linear ferries worldwide, Transportation Research Procedia, vol. 25, pp. 3961-3974, Thompson, R., Burroughs, R. & Smythe, T., 2006. Exploring the connections between ferries and urban form: Some considerations before jumping on board. Journal of Urban Technology, 13(2), pp.25–52. Tsai, C.-H. et al., 2015. Exploring property value effects of ferry terminals: Evidence from Brisbane, Australia. Journal of Transport and Land Use, pp.1–19. Van Berlekom, A 2014. Analysis of commuter boat preferences in the Stockholm area: From a passenger's perspective. KTH School of Architecture and the Built Environment. Master’s Thesis. Weisbrod, R.E. & Lawson, C.T., 2003. Ferry systems: Planning for the revitalization of U.S. cities. Journal of Urban Technology, 10(2), pp.47–68.

Ramin Shabanpour (University of Illinois at Chicago, United States)
Nima Golshani (University of Illinois at Chicago, United States)
Kouros Mohammadian (University of Illinois at Chicago, United States)
Investigating adoption behavior of advanced vehicle technologies


Autonomous vehicles (AVs) have been hailed as one of the leading players in the development of smart cities. This technology has taken huge strides forward and many pioneer automakers and technology firms are racing to bring it to the market within the next decade. At the same time, electric vehicles (EVs) have shown promise for enhancing the environmental sustainability of the transportation systems by reducing the vehicle carbon footprint, especially if the electricity has been generated by renewable energy sources. Along with the leap of vehicle technologies associated with autonomous driving control systems as well as electric sources of power, research on them has been quickly maturing over the past few years. One of the major research directions in this area is the analysis of public perceptions about and adoption behavior of AVs and EVs. There exist some studies focusing on the adoption behavior of AVs and EVs, aiming at specifying how consumers’ demographic characteristics, travel habits, and built-environment factors affect their adoption behavior. Since AVs are not yet publicly available, and EVs have not been widely accepted in the market, most of these studies have used stated preference (SP) choice experiments to collect information regarding consumers’ preferences and attitudes towards them and applied various types of discrete choice models to scrutinize consumers’ adoption decision. Autonomous driving and electric fuel technologies are maturing simultaneously, and future vehicles will likely take advantage of both automated control systems and electric fuel sources. However, the review of the literature dealing with adoption behavior of AVs and EVs reveals that most of the related studies have explored their adoption behavior independently. Therefore, there is a substantial need to consider both these technologies in a unified modeling structure to effectively forecast and plan for their adoption by consumers. This study aims to address this need, offering a detailed analysis of consumers’ preferences for electric and autonomous vehicles. All results presented in this study are based on a web-based SP survey which is recently conducted in Chicago, US.

Mark Bradley (RSG, United States)
Ben Stabler (RSG, United States)
Dan Morgan (Caliper, United States)
Howard Slavin (Caliper, United States)
Agent-based Exploratory Modeling and Analysis of Scenarios for Private and Shared Autonomous Vehicle Use
SPEAKER: Mark Bradley

ABSTRACT. Introduction The paper describes a project carried out for the US Federal Highway Administration to demonstrate the concepts of Exploratory Modeling and Analysis (EMA) in the context of the transition to connected and autonomous (CV/AV) vehicle technology. The methodology was to integrate an activity-based model (ABM) with dynamic traffic assignment (DTA) in the Jacksonville, FL region, introducing new features in the models to reflect specific assumptions regarding the demand and network supply for CV/AV. EMA is an approach developed by researchers to deal with “deep uncertainty”. The approach is different from typical scenario analysis in that it is designed to deal with a large number of uncertain model relationships and inputs. While more standard scenario analysis might vary a few of the model inputs (e.g. future population growth and income levels), EMA is more appropriate in a future context where even the fundamental relationships or parameters of the model may be in question. Such a context is a “disruptive” technology such as autonomous and connected vehicles. The paper describes the methods and philosophy behind EMA and compares it to other scenario-based modeling approaches. Methodology The first task of the research was to more tightly integrate the DaySim ABM and the TransModeler simulation-based DTA, both of which had already been implemented for Jacksonville. As part of this work, the feedback between DaySim and TransModeler was enhanced in important ways. First, the activity-based demand model was modified to include a new mode (“Paid Rideshare”, such as Uber or Lyft) and also to include a choice of private auto ownership--the choice between owning conventional or autonomous vehicles. The ABM was also modified to use separate travel time and cost skims for autonomous and conventional vehicles, allowing for the DTA to treat autonomous vehicles as a separate “user class” and pass back AV-specific skim matrices. A key enhancement to the TransModeler DTA was the production of dynamic skims—using the simulated travel times to update the OD travel time matrices that are fed back to the demand model. TransModeler was also updated to accommodate the tour-based output of the ABM, keeping the temporal interdependence and consistency when simulating different trips and activities along a tour. The second task focused on adapting the ABM and DTA models to accommodate key dimensions of uncertainty in the context of AVs. The following model input and parameter assumptions were modeled and analyzed in the Phase 1 of the project, which has now been completed: • The level of private AV ownership among households, and the relationship with household characteristics such as income, age of the head of household, and commute distances of household workers. • The level of usage of AV-based “paid rideshare” services (Uber, Lyft, etc.), and how the usage varies with important factors such as age, travel cost, and land-use density near the trips ends. • The disutility of in-vehicle time in a driverless, autonomous vehicle can be discounted by a specified percentage relative to the disutility of in-vehicle time when driving a conventional vehicle. • The level of vehicle automation, including fully driverless operation and features such as adaptive cruise control. • The allowance for fully autonomous vehicle operation on the network, including provision of AV-only highway lanes or facilities to allow the vehicles to operate more efficiently than in mixed traffic. Summary of the Phase 1 EMA Findings Several different scenarios were run using various combinations of all of the different types of assumptions listed above. Figure 1 shows that the percent of person trips made in AVs varied across the scenarios from 20% with relatively low private AV (“AV low”) and shared AV (“SH low”) use, to a high of about 95% with both private and shared AV use at high levels. Assuming a very low disutility of in-vehicle time in AVs (“VOT low”) relative to conventional vehicles does not increase share of trips by AV substantially, although it does increase the average distance of AV trips substantially due to shifts in destination choice. In the Phase 1 simulations, feedback of AV and conventional vehicle travel times from the DTA does not shift the demand significantly across global iterations. Figure 1

The exploratory modeling and analysis to date also has not shown any substantial reductions in congestion or increases in effective capacity arising from high penetration of AVs. (In a DTA using traffic microsimulation, road capacity is an emergent output rather than a pre-assumed input.) The preliminary results suggest that the common speculation that traffic systems will operate at a higher effective capacity when AVs reach a high level of market penetration may be premature. Although more scenarios must be run to draw firmer conclusions, the use of detailed network-based simulation modeling has helped to focus on some of the challenges that will be faced when designing traffic systems to accommodate both autonomous and conventional vehicles in mixed traffic. Phase 2 Priorities In Phase 2 of the research, which will be completed by April, 2018, a more complete set of scenarios is being run using an experimental design approach, so that the effect of each demand and supply assumption that is varied can be analyzed and attributed independently of other assumptions. In addition to running more scenarios, the Phase 2 research is testing a wider range of input assumptions. On the demand side in DaySim, this includes being able to represent different assumptions regarding: • Changes in intra-household ridesharing/chauffeuring behavior due to AV ownership (and associated changes in the generation of empty vehicle trips). For example, an AV may drop off a household commuter, then return home empty to be available for any non-workers until it has to pick up the commuter at the end of the work day. • It is important that these household decisions also reflect parking availability for AVs at the destinations, as that may influence the relative attractiveness of returning home. On the network side in TransModeler, the Phase 2 updates include being able to represent different assumptions regarding: • The way in which AV-based paid ridesharing services route and locate vehicles when empty. This supply model will produce level-of-service skim matrices of wait times, travel times and fares for these services, and those skims will be passed to the demand model (rather than the Phase 1 approach in which the demand models use land use density as a proxy for the availability and wait time for rideshare vehicles). • Different treatment of empty vehicle trips on the network. For example, empty vehicles could be prohibited from using congested facilities during peak periods. Empty autonomous vehicles will be another “user class” in the DTA, with travel time skims that can influence the generation of empty trips for privately-owned AVs. • The location and supply of parking, including “super-stacked” and/or remote parking for self-parking vehicles. Use of the Research Results The objective of using detailed microsimulation models for this exploratory research is to learn things about the complex demand-supply relationships for autonomous and connected vehicles that could not be learned from more simple modeling approaches. On the other hand, we recognize that other modeling approaches, such as trip-based or activity-based models in combination with static network assignment, or non-network-based strategic models, are more practical and will also be important for modeling future AV scenarios. The results from our detailed microsimulation approach will be used to provide guidance for how the less-complex models can be parameterized to mimic the behavior and relationships that are observed from the microsimulation models.

13:30-15:30 Session 11B: Choice Model Estimation Methods
Karthik Konduri (University of Connecticut, United States)
Location: MCC Theater
Virginie Lurkin (Ecole Polytechnique Fédérale de Lausanne, Switzerland)
Anna Fernandez Antolin (Ecole Polytechnique Fédérale de Lausanne, Switzerland)
Michel Bierlaire (Ecole Polytechnique Fédérale de Lausanne, Switzerland)
Discrete-continuous Maximum Likelihood Estimation

ABSTRACT. The state-of-the-art for the mathematical modeling of disaggregate demand relies on choice theory inspired from micro-economics. In the estimation of discrete choice models, in general, only continuous parameters are considered, although advanced models include also discrete ones. The most typical example of a discrete parameter that is usually disregarded from the estimation process is the nest allocation parameter in nested logit models. Nesting structures are used in discrete choice models when correlation between alternatives is suspected. They are used in a very broad range of transportation contexts such as airline itinerary choice, car-type choice, route choice, and in mode choice among others.

In practice, to determine the most appropriate nesting structure, the analyst has several options: (i) to enumerate all the possible values, and estimate the continuous parameters for each combination, and (ii) to make the problem continuous by relaxing the integrality of the discrete parameters. For instance, a membership indicator becomes a continuous variable between 0 and 1 (like in the cross-nested logit model), or by making the membership probabilistic (like in latent class models). In both cases, however, the likelihood function features several local optima, so that classical nonlinear optimization methods may not nd the (global) maximum likelihood estimates.

In this work, we propose a new mathematical model that is designed to nd the global maximum likelihood estimates of a choice model involving both discrete and continuous parameters. We call our approach discrete-continuous maximum likelihood (DCML) because we introduce into the maximum likelihood framework binary parameters. We rely on simulation to formulate our problem as a mixed integer linear problem (MILP). This is a rst attempt towards a complete MILP formulation of the maximum log likelihood, which results in a problem with high computational complexity. The goal of this presentation is to show under which circumstances our approach is computationally feasible, and to study its strengths and limitations. To do so, we use a stated preferences mode choice case study collected in Switzerland in 1998. The respondents provided information in order to analyze the impact of the modal innovation in transportation represented by the Swissmetro, a mag-lev underground system, compared to the usual transport modes of car and train.

Our contributions are multiple. First, to the best of our knowledge, we are the rst to include discrete parameters estimation in the maximum likelihood framework in the context of discrete choice models. Second, our model is formulated as an MILP. We use simulations and piecewise linear function approximation to dispose of the non-linearity of the log likelihood. We believe that it is the first time that the log likelihood is linearized. Finally, the proposed mathematical model is general and can be used with any choice model, as long as the distribution of the error terms can be simulated (e.g.: cross-nested, logit or latent class models).

Mauro Capurso (Institute for Transport Studies and Choice Modelling Centre - University of Leeds, UK)
Stephane Hess (Institute for Transport Studies and Choice Modelling Centre - University of Leeds, UK)
Thijs Dekker (Institute for Transport Studies and Choice Modelling Centre - University of Leeds, UK)
Stated consideration and attribute thresholds in mode choice models: a hierarchical ICLV approach
SPEAKER: Mauro Capurso

ABSTRACT. See uploaded file

Annesha Enam (University of Connecticut, United States)
Karthik Konduri (University of Connecticut, United States)
Jingyue Zhang (University of Connecticut, United States)
A Simulation Free Analytical Estimation Approach for Integrated Choice and Latent Variable Models (ICLV) with Multivariate Choice Kernels and Hierarchical Relationships between Latent Variables

ABSTRACT. Introduction In the recent years, the integrated choice and latent variable (ICLV) modeling framework, a statistically rigorous framework that allows the appropriate treatment of the latent explanatory variables into choice models (McFadden 1986, Walker and Ben-Akiva 2002), is increasingly being employed in the transportation planning arena to explain heterogeneity in choice process as a function of various latent psychological factors. ICLV framework allows for a rich depiction and understanding of the phenomenon underlying the decision making process and choice behaviors exhibited by individuals. Despite the popularity of the ICLV framework, recently, the approach has been criticized for its comparable predictive ability to other methods of catering unobserved heterogeneity. Researchers have also suggested exercising caution when drawing conclusions about causality and policy impacts (Chorus and Kroesen 2014). Despite the criticisms, researchers have consistently maintained that one of the distinguishing features of the ICLV framework is its ability to identify the structural relationships between different variables: including the influence of latent variables on the choice process and between the latent variables themselves (Hess 2012, Vij and Walker 2016). There are a number of ICLV implementations in literature. However, there are three important limitations. First, most ICLV formulations to date have been implemented with single discrete choice kernel and maximum simulated likelihood (MSL) routine has served as the workhorse for estimating such model formulations (Kim et al. 2014). Recently, the authors proposed one of the first ICLV implementations multiple discrete continuous choice (MDC) kernel (capable of simultaneously modeling the choice of the goods and associated consumption quantity that are imperfect substitutes) (Enam et al. 2017). Transportation planning field is in need of ICLV formulations with other types of multivariate choice kernels. Second, in most ICLV implementations, the specification of the structural relationships among the latent variables are simplified (Kamargianni et al. 2015). The hierarchical relationships between different latent variables are seldom specified. Even when hierarchical relationships between latent variables are specified this is done at the expense of simplifying the specification of the choice kernel (Zhao 2009). Third, simplified assumptions are made to treat the indicators used to construct the latent variables in ICLV model implementations. Indicators are typically collected on a Likert scale of discrete ordered options. But, the indicator variables in most ICLV implementations are often treated as continuous variables (Paulssen et al. 2014). The limitations described above (related to the choice kernel, structural relationships, and treatment of indicators) are in part related to the computational tractability issue of estimation routine. As noted above, MSL serves as the dominant approach for estimating ICLV models. The MSL method is cumbersome due the method’s dependency on numerical simulation. The MSL method is practically infeasible in the presence of large number of latent variables. In order to tap into the full potential of the ICLV models, it is essential to develop alternative estimation routines that allow development of comprehensive ICLV implementations – with desirable statistical properties (such as consistency and asymptotic efficiency) and necessary computational tractability for the deployment in the practical scenarios. There are recent research efforts offering alternative estimation routines that have the potential for overcoming the limitations referenced above. Bayesian (Daziano 2015) and composite likelihood (CL) (Varin 2011, Bhat and Dubey 2014) based techniques have emerged as the two promising alternative estimation methods for developing comprehensive ICLV models. The primary objective of the proposed research is to present a simulation free analytical estimation technique for the ICLV models that allows study of multivariate choice kernels appropriate for the dependent variables of interest, allows the specification of hierarchical structural relationships between latent variables, and accommodates appropriate treatment of the indicator variables. Similar to the approach proposed by Bhat and Dubey (2014), the proposed estimation technique will combine composite marginal likelihood (CML) method proposed by Varin et al. (2011) with analytical approximation of the normal cumulative density function developed by Bhat (2011). Subsequently, the approach will be used to develop comprehensive ICLV formulations to understand the relationship between different psychometric variables and activity travel behaviors of interest.

Methodology In the proposed research, the ICLV implementations will be developed for three different multivariate choice kernels typically encountered in the transportation planning field, namely, multivariate ordered probit (MOP) (e.g. suitable for modeling occurrences of the number of episodes by activity types), multinomial unordered probit (MUOP) (e.g. suitable for modeling mode choice) and multiple discrete continuous probit (MDCP) (e.g. suitable for simultaneously modeling activity participation and time allocation decisions subject to time constraint). Any ICLV implementation consists of three main components: 1) structural equation model of latent variables, 2) measurement equation model of latent variables and 3) structural equation model of the choice kernel. As with any ICLV, in the three model implementations described above, both the measurement equation of the latent variables and the structural equation of the choice kernels will be functions of latent variables. Additionally, to allow for hierarchical relationship between latent variables, the structural equation of the latent variables will also be functions of latent variables. The likelihood function for the ICLV implementation involves evaluating a multidimensional integral representing the joint probability of observing the indicator variables (that are used to measure the latent variables) and the choice indicators. As noted earlier, the CL estimation technique will be used in this research for estimating the models. The CL estimation technique proceeds by constructing convenient lower dimensional surrogates of the joint probability function which are easier to evaluate compared to the high dimensional joint probability function (Varin 2011). In the current research, the joint probability function will be expressed in terms of pairwise marginal probability functions – this particular method of decomposing the higher dimensional integral is also referred to as composite marginal likelihood (CML) method in the literature and has been adopted by Bhat et al. (2013), and Bhat and Dubey (2014). When dealing with the three multivariate choice kernels referenced above using CML technique, different considerations for decomposing the higher dimensional integral is needed. For example, joint probability function arising in the presence of MOP choice kernel can be decomposed into bivariate cumulative distribution functions (CDF) that are straightforward to evaluate. On the other hand, the presence of MUOP kernel with ordinal indicators for latent variable will involve treatment of higher order probability functions that need to be analytically evaluated to circumvent the need for numerical simulation. In the current research, the approximation proposed by Bhat (commonly referred to as MACML approximation) will be utilized. Finally, in the presence of MDCP choice kernel, the size of the integral to be decomposed using CML technique varies from one individual to another. Consequently, weighted CML technique needs to be invoked in order to obtain consistent and efficient estimates of parameters as demonstrated by Enam et al. (2017). Further, for developing ICLV implementations with hierarchical relationships between the latent variables, rigorous exploration of the identification conditions is required. The research will present the identification conditions of the ICLV implementation for each of the three choice kernels in the presence of hierarchical relationships between latent variables. The estimation routines will be tested on simulated datasets to demonstrate feasibility and applicability to recover the true parameters underlying the data generation process.

Empirical Study The ICLV model formulations developed in this research will be applied on data drawn from the 2013 Disabilities and Use of Time (DUST) and 2014 Childhood Retrospective Circumstances Study (CRCS) which are in turn supplements of the Panel Study of Income Dynamics (PSID) surveys (PSID 2016). DUST provides information regarding elderly couples’ satisfaction with life and different domains of it in addition to their daily time use information on a randomly chosen weekday and weekend. The CRCS supplement provides same elderly individuals’ (as the DUST) evaluation of different childhood experiences including their relationship with parents, neighborhood condition, health condition, and economic condition among others. The DUST and CRCS datasets provide a unique opportunity to explore the contribution of childhood experiences (such as trauma, childhood abuse, and exploitive relationships, physical and mental ailments) on the well-being at later stages in life. Moreover, the dataset will allow the exploration of the relationship between childhood circumstances and the adult age activity participation and time use choices (either direct or mediated via satisfaction with life in later stages of life). Using the ICLV formulation with hierarchical latent variables, the empirical study will seek to understand the influence of past experiences (using data from CRCS) and present life circumstances (using daily time use information in addition to current economic and health condition from the DUST) on the elderly couples’ satisfaction with life and different domains of it. In particular, the paper will employ the ICLV formulation with multinomial ordered probit kernel where the satisfaction with life and different domains of it such as job, finance, marriage, memory, and health would serve as the ordinal choice variables. On the other hand, multiple latent variables (such as relationship with parents, presence/relationship with friends, childhood physical health, childhood mental health, and childhood neighborhood experiences among others) would be constructed from the various indicators available in the CRCS dataset. Consequently, structural relationships will be explored between the different measures of well-being at adult age (specified as ordered dependent variables) and the childhood circumstances (casted as latent variables). Further, structural relationships between various latent variables representing the childhood circumstances will also be explored. Additionally, aggregate time use information for different activity types will be used as a measure of present life circumstances and consequently their relative influence on the adult age well-being will be explored.

Anticipated Results The research described in this study attempts to contribute both methodologically and empirically. On the methodological front, the study will develop and demonstrate a simulation free composite marginal likelihood (CML) based analytical estimation technique that allows the application of ICLV models with complex choice kernels, hierarchical structural relationships between latent variables, and rigorous treatment of latent variable indicators. The estimation technique will allow for more rigorous specification of ICLV models in empirical research, thus, enabling explorations with potential for novel behavioral insights. Moreover, the research will demonstrate the feasibility and applicability of the simulation free estimation technique and promote the adoption of comprehensive ICLV implementations in practice. On the empirical front, the study attempts to contribute to the wellbeing literature by disentangling the impacts of past childhood experiences and present life circumstances (measured via daily time use) on the overall well-being of the individuals (Duarte et al. 2010, Abou-Zeid and Ben-Akiva 2012, Archer et al. 2013). The empirical study would allow to disentangle the influence of daily time use of the elderly people on their overall wellbeing from other dispositional influences such as childhood physical health and mental conditions. Further, the empirical study contributes to the study of elderly, the study of whom is becoming increasingly important owing to their growing shares in the US and elsewhere.

Relevance to the Theme of the Conference The proposed research addresses the overall theme of the 2018 IATBR conference and speaks to specific topics of interest to the conference. Related to the overall theme, the proposed research attempts to further the understanding of the motivational, situational and dispositional factors that impact time use behavior and overall well-being of elderly individuals. Related to the specific topics of interest to the conference, the proposed research contributes directly to the psychometrics theme by exploring the heterogeneity in various wellbeing measures of elderly as a function of childhood and adult age experiences.

Romain Crastes Dit Sourd (Choice Modelling Centre, UK)
David Palma (Choice Modelling Centre, UK)
Stephane Hess (Choice Modelling Centre, UK)
Andrew Daly (Choice Modelling Centre, UK)
Christian Holz-Rau (VPL-TU Dortmund, Germany)
Joachim Scheiner (VPL-TU Dortmund, Germany)
New insights into model specification and selection for composite marginal likelihood estimation: application to car availability

ABSTRACT. please see PDF attached.

13:30-15:30 Session 11C: Healthy, Happy, and Holistic Living -- Charging Behavior
Rosaria Berliner (University of California, Davis, United States)
Location: Corwin East
Seheon Kim (Eindhoven University of Technology, Netherlands)
Soora Rasouli (Eindhoven University of Technology, Netherlands)
Harry Timmermans (Eindhoven University of Technology, Netherlands)
Dujuan Yang (Eindhoven University of Technology, Netherlands)
A priority based approach to dynamic multi-day activity generation: illustrated with PEV users’ charging behavior
SPEAKER: Seheon Kim

ABSTRACT. The aim of this paper is to develop a conceptual framework for activity scheduling in the latest version of Albatross (Rasouli, et al., 2018), which can simulate activity-travel patterns across multi-day or multi-week time periods. We suggest a heuristic to make activity generation process dynamic, allowing time-varying, individual-specific, activity priority for scheduling the flexible part of an activity agenda. The proposed framework is illustrated using users’ charging activity for plug-in electric vehicle (PEV) as an example. The model is derived from multi-week activity diary data collected through a smartphone-based prompted recall survey (Geurs et al., 2015). The current version of Albatross consists of two major components that together define an activity schedule for a single day. The first component generates an activity skeleton consisting of fixed activities (i.e., work, school, business and bring/get) and their exact start time and duration. The second component determines the part of the schedule related to flexible activities (e.g., daily shopping, service, non-daily shopping, social, leisure, touring) to be conducted that day, their travel party, duration, time-of-day and travel characteristics. It should be noted that for both components the activities are generated and scheduled according to a predefined priority list. Although such a predefined priority list assumes that all respondents feel the same level of urgency for all activities which cannot be the case, the approach can still be defendable for 1-day activity pattern prediction where any interrelation between the conduct of a certain activity in one day with the need/desire to conduct it in any time in the future is ignored. In this case a predefined priority associated to activities can be attributed to the sequence with the highest probability derived from the data. However, if the one day activity agenda is to be extended to multi-day/multi-week activity patterns, such an approximation may cause substantial error as the priority of each activity in any subsequent days may depend on the urgency of conducting the activity in that specific day which may in turn depend on the time elapsed since the last previous conduct of that activity. In other words and in order to make the model dynamic, it is required to incorporate a memory of when the activity was conducted the last time, and to make scheduling decision sensitive to this history information. Based on the dynamic activity priority deduced in this paper, in the new version of Albatross, the activities will be generated and scheduled in the order of the priority.

Michael Heilig (Karlsruhe Institute of Technology, Germany)
Patrick Plötz (Fraunhofer, Germany)
Tamer Soylu (Karlsruhe Institute of Technology, Germany)
Lars Briem (Karlsruhe Institute of Technology, Germany)
Martin Kagerbauer (Karlsruhe Institute of Technology, Germany)
Peter Vortisch (Karlsruhe Institute of Technology, Germany)
Assessment of fast charging station locations - an integrated model based approach

ABSTRACT. The diffusion of plug-in electric vehicles (PEVs) is a slow starter in most countries – also due to the lacking charging infrastructure. Building fast-charging infrastructure may overcome these issues. Therefore, it is important to assess its future demand in advance for goal-oriented planning processes and investments. We developed a model-based approach to assess the future demand of fast-charging infrastructure (150 kW) relative to a given market diffusion of PEVs for the Stuttgart Region located in south west of Germany with 2.7 Mio inhabitants. First, we use the microscopic travel demand model mobiTopp to assess the demand for charging infrastructure for several market penetration scenarios. Based on the model output, we developed an assessment tool to calculate the number of necessary charging points for each location as well as the amortization period.

Results show that the most profitable location is near to a highway and in the middle of a large commercial area. The amortization will take less than two years. The less profitable location will never redeem and is located in a rural area, where most PEV users are able to charge their car at home anyway. The results show further that up to a market penetration of 200,000 PEVs, four charging points are sufficient for most locations.

Xuekai Cen (The Hong Kong University of Science and Technology, Hong Kong)
Hong K. Lo (The Hong Kong University of Science and Technology, Hong Kong)
Lu Li (The Hong Kong University of Science and Technology, Hong Kong)
Enoch Lee (The Hong Kong University of Science and Technology, Hong Kong)
Modeling Electric Vehicle Charging Behavior from Stated Preference Survey
SPEAKER: Xuekai Cen

ABSTRACT. In this paper, a mixed user equilibrium (MUE) model with Electric Vehicles (EVs) and Gasoline Vehicles (GVs) is proposed to account for the charging behavior of EV drivers in an urban network. The main difference between EVs and GVs lies in that certain EVs with immediate charging need have to traverse a specific station for recharging, while GVs and other EVs without immediate charging need do not have such a requirement. The proportion of EVs with immediate charging need is OD specific, related to their daily commute trip lengths and EV driving ranges, i.e. EVs will need recharging once every few days. The mixed user equilibrium (MUE) conditions state that EV drivers with charging need choose the routes via a charging station while en route to their destinations with minimum travel time cost, electricity cost plus charging station cost; whereas GV drivers and other EV drivers select the routes with minimum travel cost without having to traverse any charging station. The demands for EVs and GVs follow a logit model, whose utility function is derived from an EV stated preference survey conducted in Hong Kong. We formulate a convex mathematical program to capture the MUE conditions, and develop a Frank-Wolfe based algorithm for solution. We first use the Yang-Bell network to demonstrate properties of the formulation, and the Sioux-Falls network to show its solution efficiency.

Rosaria Berliner (University of California, Davis, United States)
Scott Hardman (Institute of Transportation Studies University of California, Davis, United States)
Gil Tal (Institute of Transportation Studies University of California, Davis, United States)
Autonomous vehicle and electric vehicle adoption: Are early adopters of electric vehicles interested in adopting new vehicle technologies?

ABSTRACT. Research on vehicle automation is one of the timeliest topics in transportation – everyone is interested in vehicle automation and how and when it will be adopted. Some of the questions plaguing the research community include vehicle design, cost, and consumer adoption. What will the vehicles look like? How much will they cost? How much are people willing to pay for automated vehicle technology? Who will buy these vehicles? Many of these questions will remain unanswered until automated vehicles are available to the consumer; however, beginning to understand the first adopters of these technologies may be possible by focusing on similar early adopters of transportation technologies such as electric vehicles and “autopilot” technologies.

The market introduction of plug-in electric vehicles has been guided by an abundance of literature on vehicle adoption. Studies have used questionnaire surveys to identify who the early adopters of electric vehicles are finding that they will be highly educated, high income, mostly male, living in households with more than 1 car, be part of large social groups, and be willing to accept change. (Egbue and Long 2012; Jakobsson et al. 2016; Lane et al. 2014; Plötz and Gnann 2011). These findings were in alignment with the expected profile of early adopted form Diffusion of Innovation Theory (Rogers 2003).

Prior research on autonomous and self-driving vehicles focuses on potential users, vehicle safety and perceptions of safety, opinions of consumers, and willingness to pay for vehicle autonomous vehicle technology. Research in these areas is dynamic and growing quickly. Several researchers believe that younger people will be the first to adopted automated vehicle technology (Abraham et al. 2016; Bansal and Kockelman 2016; Deloitte 2014; Lee et al. 2017). A 2016 survey of Americans ages 12 through 64 years old conducted by Kelley Blue Book, the automotive research firm based in Irvine, CA, reports nearly 63% of Americans believe that fully autonomous (or driverless) vehicles are safer and more efficient; however, many believe that total adoption of autonomous vehicles will not be achieved in their lifetimes (Kelley Blue Book 2016).

A 2014 report by Schoettle and Sivak surveyed consumers in the U.S., U.K., and Australia about opinions on autonomous and self-driving vehicles. They found that majority of respondents had some prior knowledge of autonomous vehicle technology but also expressed concerns about riding in self-driving vehicles in terms of safety, security, and performance (Schoettle and Sivak 2014). However, despite hesitations, a majority of respondents indicated a desire to have this technology in their vehicle (Schoettle and Sivak 2014). Furthermore, most of the respondents were unwilling to pay extra for the autonomous technology (Schoettle and Sivak 2014). In a survey about autonomous vehicle and shared autonomous vehicle use among American consumers, (Gurumurthy, Kockelman, and Hahm 2018) found that Americans are willing to pay, on average, $2,073 to own an autonomous vehicle over a conventional gasoline vehicle and pay an additional $1,078 to include a manual driving option.

This study investigates the potential adoption of autonomous and self-driving vehicles through analysis of data collected with an online questionnaire survey with three distinct samples: the first includes new luxury vehicle buyers, the second includes new electric vehicle buyers, and the third includes new electric vehicle buyers with autonomous capabilities (i.e. Tesla buyers with the autopilot hardware and software). For the three samples, we focus on vehicle purchase information (i.e. year, make, model, purchase price, lease, etc.), household vehicle composition, household demographics, commute information, longest road trip, shared mobility adoption, opinions and attitudes about various topics in transportation, the environment, etc., knowledge and awareness of autonomous vehicles, and socio-demographic and socio-economic characteristics. Those in the first and second samples will be asked “what-if” questions in terms of owning autonomous vehicles; whereas in the third sample, more specifically, those who own an electric vehicle with autopilot software (i.e. Tesla vehicle purchasers with the autopilot software), will be asked about how, when, and where they use the automated feature in their vehicle. All three samples will be asked about their intention or desire to use autonomous vehicle features, their understanding of the different levels of autonomy, and their preferred level of vehicle automation.

As a precursor to the study described here, we surveyed plug-in electric vehicle households, throughout June and July 2017, in California and asked them about their knowledge of autonomous vehicles. Most respondents fell in the middle – they did not indicate total ignorance of autonomous vehicles but also did not indicate that they were experts on autonomous vehicles. The graph below, Figure 1, represents the distribution of their responses.

Figure 1. Distribution of responses to autonomous vehicle knowledge (likert scale from -3 to 3 where -3 is no knowledge and +3 is expert knowledge).

Many electric vehicle owners feel that they have some knowledge of autonomous vehicles; however, this differed amongst electric vehicle type. In particular, Tesla owners believed they were more knowledgeable about autonomous vehicles than other PEV owners.

We are currently in the piloting stage of this project and expect to begin the formal data collection by the end of 2017. The survey respondents will be recruited via mail. Approximately 30,000 new vehicle owners in the United States will receive a recruitment letter and we expect a response rate of 10% or roughly 3,000 responses. In the recruitment letter, respondents will be directed to a link of the online survey where they will complete the survey. The survey has an expected completion time of roughly 25 to 30 minutes. We will distribute the survey to 1/3 electric vehicle owners with an autopilot feature available on their vehicle, 1/3 plug-in electric vehicle owners, and 1/3 conventional luxury vehicle owners. We are confident that by the time of the IATBR conference, we will be able to present preliminary results from the analysis of our three samples.

This study will provide insights into the differences in attitudes and the potential adoption of autonomous vehicles of three key groups most likely to be among the first to purchase these vehicle technologies: luxury car buyers, electric vehicle car buyers, and electric vehicle car buyers with the autopilot feature on their new vehicle. Luxury car buyers may seek the convenience of autonomation for lackluster commutes and electric vehicle car buyers are already a group of individuals proven to be early adopters of new vehicle technologies.

References: Abraham, H. et al. 2016. Autonomous Vehicles, Trust, and Driving Alternatives: A Survey of Consumer Preferences. Bansal, Prateek, and Kara M. Kockelman. 2016. “Are We Ready to Embrace Connected & Self-Driving Vehicles? A Case Study of Texans.” Transportation (44): 1–35. Deloitte. 2014. Global Automotive Consumer Study Exploring Consumers’ Mobility Choices and Transportation Decisions. Egbue, Ona, and Suzanna Long. 2012. “Barriers to Widespread Adoption of Electric Vehicles: An Analysis of Consumer Attitudes and Perceptions.” Energy Policy 48(2012): 717–29. Gurumurthy, Krishna Murthy, Kara M. Kockelman, and Hyungseung (Jeffrey) Hahm. 2018. “Deeper Understanding of Americans’ Autonomous Vehicle Preferences: Questions on Long-Distance Travel, Ride-Sharing, Privacy, & Crash Ethics.” 97th Annual Meeting of the Transportation Research Board. Jakobsson, Niklas et al. 2016. “Are Multi-Car Households Better Suited for Battery Electric Vehicles? – Driving Patterns and Economics in Sweden and Germany.” Transportation Research Part C: Emerging Technologies 65: 1–15. Kelley Blue Book. 2016. “Future Autonomous Vehicle Driver Study.” (September). Lane, Bardlay et al. 2014. Beyond Early Adopters of Plug-in Electric Vehicles? Evidence from Fleet and Household Users in Indianapolis. Transportation Research Board 2014 Annual Meeting. Lee, C. et al. 2017. “Age Differences in Acceptance of Self-Driving Cars: A Survey of Perceptions and Attitudes.” In 3rd International Conference on Human Aspects of IT for the Aged Population, Vancouver, British Columbia. Plötz, Patrick, and Till Gnann. 2011. “Who Should Buy Electric Vehicles ? – The Potential Early Adopter from an Economical Perspective.” ECEEE (2009): 1073–80. Rogers, Everett M. 2003. Diffusion of Innovations. 5th Editio. New York: Free Press. Schoettle, B, and M Sivak. 2014. A Survey of Public Opinion About Autonomous and Self-Driving Vehicles in the U.S., U.K., and Australia.

Rick Wolbertus (Delft University of Technology/ Amsterdam University of Applied Sciences, Netherlands)
Maarten Kroesen (Delft University of Technology, Netherlands)
Robert van den Hoed (Amsterdam University of applied Sciences, Netherlands)
Caspar G. Chorus (Delft University of Technology, Netherlands)
Fully charged: An empirical study into the factors that influence connection times at EV-charging stations

ABSTRACT. Statement of contribution: This study is the first to systematically and quantitatively explore the factors that determine the length of charging sessions at public charging stations for electric vehicles in urban areas. We use a unique and large data set – containing information concerning 3.7 million charging sessions of 84,000 (i.e., 70% of) Dutch EV-users – in which both private users and taxi and car sharing vehicles are included; thus representing a large variation in charging duration behavior. Based on a estimation of a series of mixed logit- and latent class-based ordinal regression models, we identify key factors explaining heterogeneity in charging duration behavior across charging stations. We show how these explanatory variables can be used to predict EV-charging behavior in urban areas and to optimize types and numbers of charging infrastructure.

Full abstract in attached PDF file

13:30-15:30 Session 11D: Social Exclusion
Elizabeth McBride (UCSB, United States)
Location: UCEN SB Harbor
Elizabeth Mcbride (University of California, Santa Barbara, United States)
Adam Davis (UCSB, United States)
Konstadinos Goulias (UCSB, United States)
An Analysis of Activity and Travel Fragmentation among Disadvantaged Groups

ABSTRACT. Spatio-temporal analysis of time fragmentation among different groups. Lessons learned about disadvantaged groups.

Hsin-Ping Hsu (Tamkang University, Taiwan)
Domestic Responsibilities Under Constrained Mobility: Household-Serving Trip Characteristics of Immigrant Women

ABSTRACT. This study aims to explore how immigrant women fulfill their domestic responsibilities under constrained mobility since research shows that immigrants have lower access to a car and women do more household-serving travel. By examining the frequency, duration, and distance of grocery shopping and child escorting trips across gender and immigrant status, the travel patterns and barriers of immigrant women can be identified and help inform policy decision-making that can better serve the travel needs of disadvantaged groups.

Benjamin Motte-Baumvol (Université de Bourgogne - CNRS - THEMA, France)
Olivier Bonin (IFSTTAR - AME - LVMT, France)
Broadening the framework of analysis for immobility

ABSTRACT. This study has used SEM with latent variables to process UK NTS data in order to examine the time aspects of immobility by analysing the articulation between days when respondents do and do not make trips. SEM makes a substantial contribution to the work both in terms of data processing and of the resulting knowledge of inequalities in mobility. First the constitution of latent variables proved a good way to process trip characteristics. Trips are characterized by many observed variables, especially in a survey over several days in which the time aspect must also be considered. Lastly, the SEM proposed is distinctive in its capacity to take account of multiple relations, in particular in the context of Confirmatory Factor Analysis (CFA). We have thus been able to show that although immobility on a weekday can be partially offset by more trips on other weekdays for some reasons, no set-off is observed at weekends, whether respondents are in the working population or not.

Caroline Bayart (University Lyon 1, France)
Nathalie Havet (University Lyon 1, France)
Patrick Bonnel (University Lyon 2 - ENTPE, France)
Daily immobility and mobility behaviours: An Application of hurdle models in a French case study

ABSTRACT. Sociodemographic factors involving immobility are not necessarily the same as those involving less mobility among mobile people. As part of two econometric estimates, we will focus on the factors influencing, firstly the decision to travel, and secondly, the level of mobility (number of trips and daily distance budget). Our approach consists in improving the characterisation of the determinants of immobility and mobility by considering gender impacts. The first aim of this article is to explore whether the factors that explain the decision to daily travel and the frequency of travel (number of trips, distance budget) are the same. Can we isolate specific factors for each type of behaviour? Secondly, we examine whether the obtained results differ according to gender.

Clotilde Minster (Karlsruhe Institut for Technology, Germany)
Jimmy Armoogum (IFSTTAR, France)
Bastian Chlond (Karlsruhe Institut for Technology, Germany)
Immobility and low mobility levels: an investigation of a seven-day travel diary panel

ABSTRACT. Immobility in transport, i.e. the non-tripmaking is a challenge for transport authorities, urban planners but also for statisticians. Reported immobility has to be examined in order to determine if the amount of measured immobility is a survey artefact or a real behavior – by whatever reasons. In this paper we describe who the immobile persons and the persons with low mobility levels in Germany are, and to understand the reasons for their immobility. Persons with low mobility levels are defined in our research as those who undertook less than 3 trips a week. Research conducted by statisticians and surveys methodologists suggest, immobility may be analyzed as a “soft refusal” e.g. non-response (Axhausen et al. 2002a). Policy makers, transport and urban planners may also interpret immobility as a real immobility: Therefore past research also highlighted, that some population groups are not or less mobile as barriers may restrain their mobility (Bacqué & Fol 2007; Fol 2010; Motte-Baumvol et al. 2015).

Based on this literature review, we assume, not all population groups defined by their socio-demo and spatial characteristics do need to make out-of-home activities every day. Immobility or low mobility levels will therefore be more likely compared to the general population, for example for older and unemployed persons. Therefore, we aim to identify who are the persons with no or little mobility, and to suggest explanations on their behavior observing mobility behavior all things being equal.

Immobility and low mobility levels need to be analyzed beyond a given observation day (Axhausen et al. 2002b). Therefore, this research is based on data from the German Mobility Panel (MOP) from 2001 to 2015. Our analyses are based on 214´791 unweighted observation days. The MOP is a nationwide, multi-day and multi-period household travel survey that has been conducted in Germany since 1994. Each year persons aged 10 years and older participate in the MOP survey by filling out a 7-day trip diary. The participants take part in the MOP survey three consecutive years. To date, most of the studies on immobility are based on a single day of observation, (as most travel surveys report only one day). For instance, in Madre et al. (2003) and in Axhausen et al. (2007) these two works suspect that the interviewees report less trips than they really perform. The authors say this is probably because of travel surveys questionnaire are burdensome or the period under review period is too long. The seven-day and multi-period design of the MOP is for our research question an asset as it enables to examine intrapersonal variations within one week and beyond this the option to analyses decreasing report levels-of-completeness (Chlond et al. 2013; Wirtz et al. 2013), but also between the three consecutive survey waves (see Kitamura et al. 2001).

Explaining variables of immobility are not documented in travel surveys. In the MOP survey, neither explanation of the daily mobility behavior nor on activities conducted at home are gathered (see Hubert et al. 2008). However, information on weather conditions, on health status, the occasion of non-normal days and on holidays are reported. The same apply for information on age, occupation, possibility of remote work, and health related mobility restrictions which are reported since the beginning of the panel. Therefore we observe from 2001 to 2015 that in average 64.7 % of the persons were mobile every day. Focusing on working days (Monday – Friday), we observe, approx. 83 % of the population is mobile every day. Approx. 11 % of the population is immobile one day, 4 % is immobile two days, and 2.5 % is immobile more than three days. Focusing on the weekend, we observe that much more persons are immobile: approx. 21 % of the population is immobile at least one day, 5.5 % of the population is immobile Saturdays and Sundays. Over the weekend, we observe 73 % of the population is mobile every day.

Moreover, we observe, the less the level of mobility a person over the week or the weekend, lower will be the number of trips per mobile day. Over a whole week, persons mobile every day (from Monday to Sunday) make an average of 3.9 trips a day. Those immobile four days and more make in average less than 2.6 trips a day. The same relationship between and the low mobility and the number of trips made during mobile days. Focusing on week days (Monday – Friday), we observe that the persons who are mobile every day make an average of 3.9 trips a day. Those immobile one day between Monday and Friday make an average of 3.3 trips a day during their mobile days. Those immobile four days or more make an average of 2.6 trips during their mobile day. The same apply over the weekend (Saturday – Sunday), but with a lower difference: those mobile both days make 3.3 trips a day while those immobile one day make 2.8 trips by mobile day. These descriptive results underline different mobility behavior depending of the level of immobility. That is why, our contribution aims to define who are the immobile persons or those with a low level of mobility thanks to a multinomial logit analysis. A multinomial logit analysis give access to a detailed analysis of the profile of persons depending of their mobility levels (immobile five days a week, 4 days a week, 3 days a week, etc.): Similar analysis will be conducted both over the weekdays and over the weekend. Moreover, our last set of analysis will highlight mobility behavior “all things being equal” thanks to a variance analysis. In both the multinomial and the variance analysis, all variables available at the level of the household and the person – especially geographical ones – will be integrated. That means, that variables on the weather will also be considered in order to examine external and internal factors determining the mobility level. Finally, our data set being a panel survey, we will be able to control the continuity of mobility behavior from one year to the next one (considering also any intrapersonal changes in factors explaining the level of troipmaking).

Our contribution will contribute to a better knowledge of mobility behavior, especially for the persons having a low mobility level.


Adey P. (2006) If Mobility Is Everything Then it is Nothing: Towards a Relational Politics of (Im)mobilities, Mobilities, 1(1):75-94. Armoogum J., Hubert J.-P., Roux S., Le Jeannic T. (2010) Plus de voyages, plus de kilomètres quotidiens : une tendance à l´homogénéisation des comportements de mobilités des Français, sauf entre ville et campagne, in : La Mobilité des Français, Panorama issu de l´ENTD 2008, CGDD, La Défense, pp. 5-24. Axhausen K.W., Löchl M., Schlich R., Buhl T., Widmer P. (2007) Fatigue In Long-Duration Travel Diaries, Transportation, 34:143-160. Axhausen K.W., Madre J.-L. (2002a) Share of Immobiles in Travel Diary Surveys: A Review. ETH Report 133, Swiss Federal Institute of Technology, Zürich. Axhausen K.W., Zimmermann A., Schönfelder S., Rindsfüser G., Haupt T. (2002b) Observing the Rhythms of Daily Life: A Six-Week Travel Diary, Transportation, 29:95. Bacqué M, Fol S. (2007) Inequality Regarding Mobility: Observations And Policies. Swiss Journal Of Socioloy 33(1):89-104. Chlond, B.; Wirtz, M.; Zumkeller , D. (2013): Do dropouts really hurt? – Considerations about data quality and completeness in combined multiday and multiperiod surveys. In: Zmud, J.; Lee-Gosselin, M.; Carrasco, J. A.; Munizaga, M.A. (2013): Transport Survey Methods, Best Practice for Decision Making, Emerald Publishing, ISBN 978-1-7819-0287-5 Coutard O., Dupuy G., Fol S. (2004) Mobility of the poor in two European metropolises: Car Dependence Versus Locality DDependence, Built Environment, 302(2):138-145. Fol S. (2010) Mobilités du quotidien. Encouragement ou injonction à la mobilité ?, Projet, 1(314):52-58. Hannam K. Sheller M. Urry J. (2006) Editorial: Mobilites, immobilities and moorings, Mobilities, 1(1):1-22. Hubert J.-P., Armoogum J., Axhausen K.W., Madre J.-L. (2008) Immobility And Mobility Seen Through Trip-Based Versus Time-Use Surveys, Transport Reviews, 28(5):641-658. Kitamura R., Yamamoto T., Fujii S. (2001) The Effectiveness of Panels in Detecting Changes in Discrete Travel Behavior, Transportation Research Part B, 37:191-206. Madre J.-L., Axhausen K.W., Gascon M.-O. (2003) Immobility: A Microdata Analysis, 10th IATBR conference, Lucerne, August 2003. Madre J.-L., Axhausen K.W., Brög W. (2007) Immobility In Travel Diary Surveys, Transportation, 34:107. Motte-Baumvol B., Bonin O., Nassi C.D., Belton-Chevallier L. (2016) Barriers and (im)mobility in Rio de Janeiro, Urban Studies, 53:14, pp.2956:2972. Richardson T. (2007) Immobility in Urban Travel Surveys, 30th Australasian Transport Research Forum. Wirtz, M., Streit, T., Chlond, B., Vortisch, P. (2013), "On New Measures for Detection of Data Quality Risks in Mobility Panel Surveys”, in: Transportation Research Record: Journal of the Transportation Research Board, Volume 2354 / Travel Surveys; Asset Management; and Freight Data 2013, pp.19-28

13:30-15:30 Session 11E: Complex Issues in Models
Yoram Shiftan (Technion, Israel)
Location: MCC Lounge
Einat Tenenboim (Technion, Israel)
Nira Munichor (Bar-Ilan University, Israel)
Yoram Shiftan (Technion, Israel)
Justifying toll payment with biased travel time estimates: Behavioral findings and route choice modeling
SPEAKER: Yoram Shiftan

ABSTRACT. Background

Road pricing has become more and more prevalent in recent decades, yet its effects on travelers’ experience, perceptions and travel-related choices are not sufficiently known. In recent years, concerns were raised regarding the quality of transport demand forecasts, maintaining that estimation accuracy is often very poor (e.g., Parthasarathi & Levinson, 2010; Wee, 2007). Toll road demand forecasts were also found to incorporate large errors and a considerable optimism bias (Bain, 2009). Clearly, various kinds of uncertainties are liable for poor forecasting, yet data quality issues were identified as a major probable cause. The present study joins previous endeavors that argue for the integration of travelers' subjective data into the modeling process, in attempt to gain higher-quality data (e.g., Devarasetty et al., 2014; Vacca et al., 2017). According to this view, objective time data is a simple and technical measure, overlooking valuable aspects of time perception that are likely to directly relate to travelers' choices. The integration of subjective time in travel behavior models is expected to contribute to model estimation, providing this integration is done thoroughly and sensibly. In the present study, we focus on subjective time estimates obtained before a trip is carried out. These refer to travelers’ expectations of how long a certain trip will take (contrary to perceived time, which usually refers to the perception of how long a certain trip took). In essence, pre-trip time estimates involve more than pure perception - they involve the memory of several past perceptions as well as personal expectations and external information. This specific line of investigation was chosen because pre-trip estimates are more closely related to individuals’ behavior with regard to their travel choices, i.e., travelers tend to make various travel-related choices before carrying out a trip, not after.


In a recent study, Tenenboim & Shiftan (2016) showed that travelers respond differently to toll versus non-toll routes. Specifically, it was found that drivers estimated toll trips as shorter than non-toll trips, even when objective times were controlled for (henceforth, the toll road effect). In fact, toll-route choice was identified as one of the main factors accountable for the discrepancy between subjective and objective travel times. The present study was aimed at identifying the underlying psychological mechanisms of the toll road effect on time estimation, specifically focusing on examining the hypothesis that toll payment leads drivers to form false expectations of travel time savings. Two possible sources were identified and examined in this context. According to the payment account, toll route drivers tend to report higher travel time savings because they expect a return for the price paid (see the price-quality relationship, e.g., Rao, 2005). According to the choice account, toll route drivers tend to report higher travel time savings because they strive to justify their toll route choice, to others or to themselves (see the justification bias, e.g., Grisolia & Ortuzar, 2010). As previously mentioned, this study was also aimed at investigating the contribution of subjective time to demand modeling, given prior evidence highlighting its potential in improving estimation accuracy.


In a field experiment, experimenters approached people departing a local shopping mall to its parking lot, asking them for their destination, and whether they intended to drive via the Carmel tunnels, a local toll road located in proximity to the mall. Only drivers for whom the toll road was a feasible alternative were recruited. All participants in the study were paid 10 NIS (approximately 2.5 Euros), a payment which was randomly presented in one of two ways: Participants in the toll-paid group were told this payment was to cover the cost of the toll, whereas participants in the control group were told it was simply a participation reward. In other words, all participants received the same incentive, yet only half of them were led to believe it offered them a free ride via the toll road. As a result, participants in the toll-paid group who initially intended to drive via the toll road got to do so for free, whereas participants in the toll-paid group who did not initially intend to drive via the toll road had changed their route choice following the experimenter’s request. Nevertheless, participants’ initial intention to drive or not via the tunnels was registered. The experimenter then asked participants to indicate the chosen route to their destination, an alternative route, and estimated travel times for both. Participants’ intentions to drive via the toll road and the payment-frame manipulation formed a 2*2 experimental design, enabling us to examine the two possible sources for the toll effect on estimated travel time. Data of various complementary variables were also gathered (e.g., destination familiarity, origin familiarity, number of passengers, pressure for time, frequency of toll road driving, frequency of navigation app usage). Finally, the experimenter installed a route tracker application on participants’ smartphones, which provided us with their actual routes, speed data, and objective times for comparison with their subjective estimates. Data from a total of 386 drivers was collected.

Results and conclusions

An initial examination of the data revealed substantial variance in actual travel times, as trips varied from 5.7 min to 100.8 min (M=27 min, SD=17.1 min), leading us to separately consider short and long trips in some of the analyses. Estimated times ranged from 4 min to 150 min (for chosen routes) and 180 min (for alternative routes). Mean deviation of estimated time from actual time was -7.1%, indicating a general tendency to underestimate travel times. On average, the toll road saved drivers an actual travel time of 5 min. Interestingly, while those who did not initially intend to drive via the tunnels had reasonable time savings expectations, those who did intend substantially exaggerated in their estimation reported estimated time savings of 13 min on average. This finding supports a justification bias account, as those who intended to drive via the tunnels apparently tried to justify their route choice (not wanting to feel as if they made a bad choice and wasted their money). Additionally, a comparison between the two toll-paid groups revealed that the group who freely decided to drive via the toll road exaggerated their time savings more than the group who drove via the toll road following the experimenter’s request. This finding offers further support for the justification bias account, as these drivers apparently strived to justify their toll route choice. However, we failed to find support for the payment account, as drivers rode the tunnels for free significantly exaggerated their toll time savings more than drivers who actually paid the toll. The zero price effect may explain this outcome (see Shampanier et al., 2007), as driving via the tunnels for free was most likely experienced by these drivers as a ‘win’, consequently leading them to overemphasize its benefits. Another interesting finding concerned the time ratio (estimated/actual time) for home destinations, which was found to be significantly lower (.90) than the time ratio for other destinations (.97), suggesting that drivers underestimated travel time when returning home but not when driving elsewhere, a finding that is consistent with the return trip effect (Van de Ven et al., 2011; Tenenboim & Shiftan, 2016). Linear and non-linear regression models were estimated for the prediction of drivers’ time estimates based on actual times and various related variables. Regression models were also estimated for predicting drivers’ estimated travel time savings. These models highlighted the role of drivers’ initial intention to drive or not via the toll road and the payment-frame manipulation, alongside several other variables (e.g., familiarity with the origin surroundings, familiarity with the toll road, number of passengers in the vehicle and trip length). Furthermore, drivers’ (toll) route choice was modeled using binomial logit models based on estimated and actual travel times. In support of our expectation, the estimated time model yielded a better fit for the data compared to the actual time model, suggesting that estimated time has an added value to the modeling process. In an integrated model incorporating both estimated and actual times, we found a significant contribution for estimated time, yet a negligible contribution for actual time, indicating that in the presence of estimated time, the contribution of actual time was practically insignificant. Examining the contribution of other factors to drivers’ route choice, we identified two main components: high familiarity with the tunnels and/or the shopping area (origin), and a high degree of pressure for time. As a whole, this study contributes to the understanding of some of the key factors affecting toll route choice. The present findings suggest that the integration of estimated travel time in the modeling process does in fact improve modeling results, providing this integration is done sensibly, leaning on knowledge gained in research. The present study also promotes a better understanding of the factors that affect bias of travel time estimates, thereby can support the estimation of travelers' subjective estimates. Finally, this interdisciplinary study demonstrates the valuable insights that can be gained from incorporating psychological constructs within travel behavior research.


Bain, R. (2009). Error and optimism bias in toll road traffic forecasts. Transportation, 36 (6), 469-482. Devarasetty, P., C., Burris, M., Chao, H. (2014). Comparing perceived and actual travel time savings on freeways with managed lanes. Journal of Transportation, 6(1), 1–13. Grisolía, J., M., Ortúzar, J., D., D. (2010). Forecasting vs. observed outturn: Studying choice in faster inter-island connections. Transportation Research Part A, 44(3), 159–168. Parthasarathi, P., Levinson, D. (2010). Post-construction evaluation of traffic forecast accuracy. Transport Policy, 17, 428-443. DOI:10.1016/j.tranpol.2010.04.010 Rao, A., R. (2005). The quality of price as a quality cue. Journal of Marketing Research, 42 (4), 401-405. Shampanier, K., Mazar, N., & Ariely, D., 2007. Zero as a special price: The true value of free products. Marketing Science, 26(6), 742–757. Tenenboim, E., Shiftan, Y., 2016. Accuracy and Bias of Subjective Travel Time Estimates. Transportation, 1-25. Published 24 December, 2016. DOI 10.1007/s11116-016-9757-8 Vacca A., Prato, C., G., Meloni, I. (2017). Should I stay or should I go? Investigating route switching behavior from revealed preferences data. Transportation. DOI 10.1007/s11116-017-9788-9 Ven, N., Rijswijk, L., Roy, M., 2011. The return trip effect: Why the return trip often seems to take less time. Psychological Bulletin Review, 18(5), 827–832. Wee, B. (2007). Large infrastructure projects: A review of the quality of demand forecasts and cost estimations. Environment and planning B: Planning and design, 34, 611-625.

Tomas Rossetti (Pontificia Universidad Católica de Chile, Chile)
Hans Lobel (Pontificia Universidad Católica de Chile, Chile)
Ricardo Hurtubia (Universidad Católica de Chile, Chile)
Modeling subjective perceptions of public spaces and their effect on user behavior

ABSTRACT. Perceptions play a central role on the way people use and navigate their cities. Several studies have found significant relations between perceptions and use of public spaces, for example finding that perceived comfort and ease of access increase use of public spaces (Khisty, 1994; Shriver, 1997) or encourage the use of certain transportation modes (Antonakos, 1995; Hyodo et al., 2000; Zacharias, 2001; Hunt and Abraham, 2007) . Jiang et al. (2012) and Tilahun and Li (2015) show that transit users have a greater disposition to walk to stations if the streets leading to them are perceived to be safe, interesting and busy. Latkin and Curry (2003) even established a relation between perceived neighborhood disorder and depression.

Modeling and quantifying these perceptions is particularly challenging because it demands to have perceptual indicators in a wide range of urban settings. To solve this, some studies have used crowdsourced surveys such as Place Pulse ( ) or Urban Gems ( To the best of our knowledge, all of these studies have used machine learning techniques to predict perceptions (Naik et al., 2014; Ordonez and Berg, 2014; Quercia et al., 2014; Porzi et al., 2015; Dubey et al., 2016) . Even though they have reasonable degrees of predictive power, these models work as “black boxes” and do not provide information allowing to understand the drivers behind these perceptions.

In previous work (Rossetti et al., 2017, available at ), we argue that in spite of these models' adequate predictive power, there is important information left out by their uninterpretable parameters. If they were interpretable, valuable insight could be obtained to understand the drivers behind people's perceptions of public spaces. Considering this, we made a case for the use of discrete choice models in this specific case, and successfully estimated perceptual models based on the Random Utility Theory for a massive amount of choice experiments with the Place Pulse dataset (Salesses et al., 2013).

To estimate these models, the images shown in the choice experiments had to be parametrized in ways that could be useful for later parameter interpretation. Two types of features were selected to do this:

1. Low level features: Information contained in the image itself and that do not require any kind of human interpretation, such as contrast, luminosity or color saturation.

2. High level features: Information with semantic meaning. These features were retrieved with a semantic segmentation model (Badrinarayanan et al., 2015) , based on machine learning methods. They label each pixel of the images as belonging to one of eleven classes, such as “ Sidewalk,” “Road” or “Pedestrian.”

A multinomial logit model was estimated for each one the six qualitative attributes considered in the Place Pulse survey. Even though the goodness-of-fit of these models is lower than the fit of machine learning-based models, the parameters obtained are interpretable and allow to draw conclusions that may have incidence in the policy cycle.

Despite the success in the estimation of models with this type of data, several questions regarding the validity of the results remain open. We will validate our models by, in the first place, measuring how well they describe qualitative attributes that can be correlated to actual geographical, socioeconomic and land use data. Later, we will test if these modeled qualitative attributes have any significant explanatory power on actual location choices in urban contexts. Each one of these points is briefly discussed next.

On the relation between perceptions and observed land use

The perceptual models presented in Rossetti et al. (2017) rely solely on psychometric indicators provided by users around the world. The obvious question that arises from this fact is if these indicators have any relation with objective information such as a neighborhood's average income or land use policies. If there were, it would mean that these models can adequately replicate users' perceptions, and that these perceptions are good predictors of neighborhoods' objective attributes. This would allow to use these perceptual models for further applications that can uncover the relations between behavior and perceptions.

[float Figure: [float Figure:

[Sub-Figure a: “Beautiful” ] ] [float Figure:

[Sub-Figure b: “Boring” ] ] [float Figure:

[Sub-Figure c: “Depressing” ] ]

[float Figure:

[Sub-Figure d: “Lively” ] ] [float Figure:

[Sub-Figure e: “Safe” ] ] [float Figure:

[Sub-Figure f: “Wealthy” ] ]

[Figure 1: Perceptual maps for Santiago (Chile) ] ]

[float Figure: [float Figure:

[Sub-Figure a: Poverty rate at the commune level ] ][float Figure:

[Sub-Figure b: Vehicle ownership at the commune level ] ]

[float Figure:

[Sub-Figure c: Percentage of high-income population at the traffic analysis zone level ] ]

[Figure 2: Relation between predicted perception of wealth with socioeconomic indicators ] ]

An initial validation was carried out by the authors for the case of Santiago (Chile). Over 126,000 images for this city were taken from Google Street View and later processed. With this information, perceptual maps were created (see Fig. [fig:Perceptual-maps-for] ). These maps show that the northeastern area of the city seems to be perceived as wealthier, safer, more beautiful and less depressing. This area is also the one that concentrates higher-income households, which is why there could be a relevant relation between these perceptions and this area's objective characteristics. The analyses shown in Fig. [fig:Perceptual-maps-for-1] , that relate how wealthy images are perceived with socioeconomic data, prove there is a significant relation with the expected sign. This leads us to believe that there are more relations between estimated perceptions and objective attributes to be uncovered.

To better establish a relation between objective attributes of cities and perceptions estimated through the models presented in Rossetti et al. (2017) , our first objective is to generate four more perceptual maps, preferably in cities from diverse continents and cultural backgrounds. Then, the relation between socioeconomic information and perception of wealth, and the relation between land use and perceived boringness and liveliness, will be established. This will indicate if there is a link between the estimated perceptions and objective data.

On the relation between perceptions and location choice

Once the perceptual models have been validated with the objectives described in the previous section, showing there is a relation between perceptions and objective attributes of the built environment, the question that remains is if these perceptions play a role on people's decisions. Our hypothesis, informed by the literature that has stated perceptions affect the way people use and navigate public spaces, is that this will be the case. If the perceptual models presented in Rossetti et al. (2017) can replicate perceptions with a reasonable degree of accuracy, then they could be used to verify these relations.

We specifically intend to estimate urban location choice models that include perceptual variables as inputs. They will be estimated including variables usually considered in these contexts, such as accessibility to jobs or dwelling unit size, along with perceptual variables estimated with the models mentioned above. If the parameters related to these perceptual variables are significant, then we will be able to conclude they are relevant drivers behind users' behavior.


In conclusion, our objective is to relate the perceptual models presented in Rossetti et al. (2017) with objective attributes of the built environment and with people's behavior. While the former goal seeks to validate these models, the latter seeks to find if these qualitative attributes can be used in behavioral models.

The model validation will be carried out by relating socioeconomic with the “Wealthy” perceptual model, and land use data with other attributes, such as those obtained from the “ Boring” and “Lively” models. If there are significant and meaningful relations, then we will be able to conclude that these perceptual models capture perceptions with a reasonable degree of accuracy. Initial work done for the case of Santiago (Chile) shows this should be expected.

The relation between people's behavior and perceptions will be analyzed through the estimation of urban location choice models. After controlling for relevant variables, these models will include the perceptual variables quantified in Rossetti et al. (2017) . If the parameters related to them are significant, then we may be able to conclude perceptions affect people's location decisions.

There are other applications we intend to address in the future. We specifically believe users' mode and route choices are affected by perceptions (see Quercia et al., 2014), and are particularly suitable scenarios for modeling with these variables. There are other relations, such as perceptions of safety and use of public spaces during the night, that may be uncovered, and that can help understand the factors that influence active living or intensity of social interactions in urban communities.

We believe these perceptual models have great potential to predict users' behavior in various applications. We intend to present a validation and application from an ongoing project that, if successful, can aid future researchers better understand how people perceive their surroundings.


[Antonakos 1995] Antonakos, Cathy L, "Environmental and travel preferences of cyclists", Transportation Research Part A: Policy and Practice 29, 1 (1995), pp. 85.

[Badrinarayanan et al. 2015] Badrinarayanan, Vijay and Handa, Ankur and Cipolla, Roberto, "SegNet: A Deep Convolutional Encoder-Decoder Architecture for Robust Semantic Pixel-Wise Labelling", arXiv preprint (2015), pp. 5.

[Dubey et al. 2016] Dubey, Abhimanyu and Naik, Nikhil and Parikh, Devi and Raskar, Ramesh and Hidalgo, C??sar A., "Deep learning the city: Quantifying urban perception at a global scale", Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinforma… 9905 LNCS (2016), pp. 196--212.

[Hunt and Abraham 2007] Hunt, J. D. and Abraham, J. E., "Influences on bicycle use", Transportation 34, 4 (2007), pp. 453--470.

[Hyodo et al. 2000] Hyodo, Tetsuro and Suzuki, Norikazu and Takahashi, Katsumi, "Modeling of Bicycle Route and Destination Choice Behavior for Bicycle Road Network Plan", Transportation Research Record: Journal of the Transportation Research Board 1705, 1 (2000), pp. 70--76.

[Jiang et al. 2012] Jiang, Yang and Christopher Zegras, P. and Mehndiratta, Shomik, "Walk the line: Station context, corridor type and bus rapid transit walk access in Jinan, China", Journal of Transport Geography 20, 1 (2012), pp. 1--14.

[Khisty 1994] Khisty, C.J., "Evaluation of pedestrian facilities: beyond the level-of-service concept", Transportation Research Record 1438 (1994), pp. 45--50.

[Latkin and Curry 2003] Latkin, Carl A. and Curry, Aaron D., "Stressful Neighborhoods and Depression: A Prospective Study of the Impact of Neighborhood Disorder", Journal of Health and Social Behavior 44, 1 (2003), pp. 34--44.

[Naik et al. 2014] Naik, Nikhil and Philipoom, Jade and Raskar, Ramesh and Hidalgo, Cesar, "Streetscore - predicting the perceived safety of one million streetscapes", IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops (2014), pp. 793--799.

[Ordonez and Berg 2014] Ordonez, Vicente and Berg, Tamara L., "Learning high-level judgments of urban perception", Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinforma… 8694 LNCS, PART 6 (2014), pp. 494--510.

[Porzi et al. 2015] Porzi, Lorenzo and Buló, Samuel Rota and Lepri, Bruno and Ricci, Elisa, "Predicting and Understanding Urban Perception with Convolutional Neural Networks", Proceedings of the ACM International Conference on Multimedia (ACM-MM) (2015), pp. 139--148.

[Quercia et al. 2014] Quercia, Daniele and Schifanella, Rossano and Aiello, Luca Maria, "The Shortest Path to Happiness: Recommending Beautiful, Quiet, and Happy Routes in the City", in Proceedings of the 25th ACM conference on Hypertext and social media (Santiago, Chile: , 2014).

[Rossetti et al. 2017] Rossetti, Tomas and Lobel, Hans and Rocco, Víctor and Hurtubia, Ricardo, "Modeling subjective perceptions of public spaces through machine learning and discrete choice techniques" (2017).

[Salesses et al. 2013] Salesses, Philip and Schechtner, Katja and Hidalgo, César A., "The Collaborative Image of The City: Mapping the Inequality of Urban Perception", PLoS ONE 8, 7 (2013).

[Shriver 1997] Shriver, Katherine, "Influence of environmental design on pedestrian travel behavior in four Austin neighborhoods", Transportation Research Record: Journal of the Transportation Research Board 1578, 1 (1997), pp. 64--75.

[Tilahun and Li 2015] Tilahun, Nebiyou and Li, Moyin, "Walking Access to Transit Stations: Evaluating Barriers Using Stated Preference", Transportation Research Record 2534 (2015), pp. 16--23.

[Zacharias 2001] Zacharias, John, "Pedestrian Behavior and Perception in Urban Walking Environments", Journal of Planning Literature 16, 3 (2001), pp. 3--18.

Albab Noor (Resource Systems Group, Inc., United States)
Karthik Konduri (University of Connecticut, United States)
Nicholas Lownes (University of Connecticut, United States)
Understanding Mode Choice Behavior in Hartford: A Latent Class Choice Analysis Approach
SPEAKER: Albab Noor

ABSTRACT. A body of literature suggests that transit usage is influenced by a myriad of factors including transit level of service, thereby influencing automobile usage. On the contrary, other studies showed that auto usage may not be decreased by improving public transit. Such contradictions in literature renders the study of this topic non-trivial. This study investigates the aforementioned topic utilizing the concept of market segmentation. The study also explores the use of an alternative source of travel time data and compares it with the traditional source. A latent class choice model was estimated to identify four segments in the sample population of the study area and their mode choice behavior was investigated. The model shows the existence of four distinct segments of which only two segments are responsive to transit level of service. The model also shows that automobile ownership has a strong influence on automobile usage and improving transit service may not nudge the auto-dependent classes to choose transit. Results from the study shows that both views in literature hold, but for different segments of the population. Moreover, it identifies segments in the population who are more likely to respond to transit service improvement compared to others.

Deo Nobel (Almec Corporation, Japan)
Sadayuki Yagi (Almec Corporation, Japan)
Donny Cleo Patra Pakpahan (University of Indonesia, Indonesia)
Sahrial (PT. SAKA ADHI PRADA, Indonesia)
Investigating Previous Travel Demand Forecast Models for Improvement on Accuracy: A Case Study of Jakarta
SPEAKER: Deo Nobel

ABSTRACT. Introduction Japan International Cooperation Agency (JICA) has provided assistance in urban transport planning for over 60 metropolitan cities of developing countries, where various urban transportation problems are being tackled such as rapid urban development, extraordinary growth of population and the number of vehicles, traffic congestion on roads and low service standards of public transport, and endangered traffic safety, by conducting urban transportation master plan (M/P) studies or feasibility studies (F/S). However, JICA has been faced with the following three main issues: namely, large-scale household travel surveys (HTS), travel demand forecast models, and cooperation needs in urban transportation sector. In terms of travel demand forecast models, JICA is concerned about: (1) improvement of accuracy in travel demand forecast, (2) travel demand forecasting methods tailored for study objectives and cooperation needs, and (3) reduction of required time and cost for travel demand forecast. In the context of those issues, this so-called “project research” aims at extracting issues of transportation surveys and travel demand forecasting methods and studying directions for improvement with a view to reorganizing the contents for future cooperation in the urban transportation sector in developing countries. As the first step of the project research, a total of 12 major urban transportation M/P studies in developing countries were reviewed through literature review followed by interviews with the ex-JICA experts and/or the counterpart in charge of transportation surveys and travel demand forecasting at the times in order to find out issues and tendencies in transportation surveys with a focus on the HTS and in travel demand forecasting methods, as well as in cooperation needs of the urban transportation sector.. As part of the findings from the above-mentioned review, gaps of socioeconomic framework between previous projections and actual observations are observed in some M/P studies that were reviewed in this research. The impact of these differences and a travel demand model on accuracy can be analyzed by inputting actually observed socioeconomic and network variables to the previous demand forecast models with parameters unchanged. If the model outputs are similar to the actual observations, it can be concluded that the model is valid and socioeconomic framework caused the gap. In order to find a clue especially to the issues on improvement of accuracy of the demand forecast models, this paper takes the Jakarta metropolitan area as a case study and investigates the accuracy of the demand forecast models by inputting actually observed socioeconomic and network variables to the previous demand forecast models with parameters unchanged. Investigation of the Demand Forecast Models: Case Study of Jakarta In response to the request of the government of Indonesia, JICA conducted “The Study on Integrated Transportation Master Plan (SITRAMP)” in the Jakarta metropolitan area from November 2001 to March 20041). The survey program associated with SITRAMP was of considerable scope; a total of 11 transportation surveys were carried out in order to obtain various data on socioeconomic indicators, person trip movements, freight patterns, transport operations, transport infrastructure, and public opinions. Among others, HTS was a large-scale home interview survey of household daily travel with a sample size equal to 3% of Jakarta metropolitan area population of 21.6 million in 2002. Furthermore, from July 2009 to September 2011, a Japan-Indonesia joint technical cooperation project called “JABODETABEK Urban Transportation Policy Integration (JUTPI)” was also conducted by JICA in order to update the transportation survey database and revise the SITRAMP master plan2). In this project, another large-scale survey, Commuter Travel Survey (CTS), was conducted in 2010 to understand the characteristics of commuting trips (e.g., destination, mode, travel time, cost) of worker(s) and student(s) of each household and to collect the socioeconomic information of the household and household members in the Jakarta metropolitan area. This survey dataset covers as many as 179,000 households which correspond to 3% of the entire population, and provides daily commuting (i.e., home-based work and school) travel patterns and, again, detailed information on household socioeconomic characteristics. The latest dataset obtained from the CTS provides a further opportunity for an in-depth study such as investigation of the demand forecast models by inputting the latest socioeconomic data into SITRAMP models which were developed based on the conventional four-step modeling procedure. Comparing the socioeconomic data from the two studies, namely, SITRAMP (2002) and JUTPI (2010), there are quite significant changes in population and other socio-demographic data as shown in Table 1. While total number of workers from SITRAMP (2002) to JUTPI (2010) increased only slightly (2.6%), total number of students increased by 14.8% and population increased by 29.4%. Meanwhile, SITRAMP (2002) also had set “future” socioeconomic framework for 2010 with an increase in total workers, students, and population by 15.1%, -0.9%, and 11.7%, respectively. Such significant gaps indicate that the increase in population was mostly dominated by the increase in the number of students rather than workers while the forecast for 2010 was the other way around at the time of SITRAMP (2002) projection. The reason behind this may be related to the migration of workers to “newly” developed area outside the Jakarta metropolitan area and the reputation of schools/universities in the Jakarta metropolitan area that had gained more popularity. As for the total trips, though only commuting trips (i.e., work and school trips) can be compared due to the limitation of CTS, observed total number of commuting trips from JUTPI (2010) is 28.92 million trips that are equivalent to an increase by 37.6% from SITRAMP (2002). Meanwhile, SITRAMP trip generation model output with JUTPI (2010) socioeconomic input is only 23.1 million commuting trips that are equivalent to an increase by 10.1% from SITRAMP (2002). Thus, SITRAMP trip generation models could not properly forecast total number of trips even with “correct” socioeconomic input. Structures and/or parameters of SITRAMP trip generation models seem to have drastically changed. Such a large gap in the generated trips must have a significant impact in the following steps, namely, trip distribution, modal split, and network assignment. Comparison of trip distribution model outputs in the form of block OD matrices (on a city/regency level) between synthesized trips from the SITRAMP gravity models with new zonal trips from the above step as inputs and observed trips from JUTPI (2010) shows a significant difference because of the large gap of the number of trips from the trip generation models. However, if percentage compositions of trips in each block OD pair to the total trips are made and compared between synthesized trips from the SITRAMP models with new zonal trip inputs and observed trips from JUTPI (2010), the result shows similarity. Even if the comparison of both OD matrices are made in the same way but on a zone level, the result shows good similarity. This may indicate that, despite the huge gap in trip generation models, SITRAMP trip distribution models including the impedances applied are still valid. For modal split model in SITRAMP (2002), observed person-trip data are stratified into several binary choices together with the corresponding zone-to-zone impedance measures: namely, choice between non-motorized and motorized modes of transport (whereby the share of non-motorized trips varies depending on the distance), choice between public (transit) and private modes of transport (whereby the share of transit trips varies depending on the ratio of generalized time), and choice between motorcycle and automobile mode of transport (whereby the share of motorcycle trips varies depending on the distance). Analysis of HTS database suggests that respondents are not sensitive to the ratio of generalized time between public and private mode. Therefore, it is really difficult to forecast modal split/interchange between public and private modes. Thus, different approach for modal split should have been considered. Though disaggregate mode choice models were developed in JUTPI (2010), modal split model structure in SITRAMP (2002) was followed for the purpose of investigating SITRAMP demand forecast models. Another unexpected factor that affected modal split was the rocketing increase in the number of motorcycles. SITRAMP modal split models with new OD trip matrices from the above step as inputs show 65.9% for public transport and 19.4% for motorcycle (excluding non-motorized transport) for the commuting purpose; whereas, mode shares from observed commuting trips from JUTPI (2010) show the opposite result, that is, 61.7% for motorcycle and only 25.9% for public transport. In fact, increase in the number of registered motorcycles was about three times in the 8 years from 2002 to 2010. Increasing traffic congestion and unsatisfactory public transport services in those days caused people in Jakarta to try to find a more “economical” and fast mode for transport, that is, motorcycles for the reason of the dimension as well as the relatively easy financing scheme. Such a behavior was not incorporated in the modal split models in SITRAMP (2002) at all. As for network assignment, since observed traffic count locations were limited, network assignment results are examined by comparing outputs from the assignment of SITRAMP (2002) network with new synthesized OD matrices from the above previous steps with roadside traffic and passenger counts in JUTPI (2010). Highway assignment result shows that there is some discrepancy between the SITRAMP model output and observed traffic volume from JUTPI (2010). Overall, in terms of trip volume, the former is smaller than the latter, because the former has a smaller number of motorcycle trips as a result of the previous modal split. Transit assignment result also shows some discrepancy as the SITRAMP model output generally has a larger number of public transport trips than the observed passenger volumes in JUTPI (2010) in line with the above-mentioned discrepancy in the SITRAMP modal split model output. Directions for Conclusion For improvement of accuracy in travel demand forecast models, four major causes of gaps between previous forecast from the models and actual observations, namely, 1) socioeconomic framework, 2) travel demand forecast model and 3) other external factors, are discussed based on the above findings from the case study of Jakarta metropolitan area as well as review of several other M/P studies. First, gaps of socioeconomic framework between previous projections and actual observations are observed in in the case of Jakarta. Although it is important to forecast population distribution considering concentration to the capital region and migration, future socioeconomic framework is often determined by agreement with counterpart agencies based on their future vision. It may not be easy to maintain both accuracy and accordance with policy of the counterpart agencies. In any case, continuous update of database and models are essential in the long run. As for travel demand forecast models, according to the analyses on previous studies, errors were observed in trip generation models, trip distribution models, and modal split models of some projects. Meanwhile, based on the findings from the case study of Jakarta, it may be concluded that trip distribution models were well developed and synthesized OD matrices were overall reliable, probably because of the abundant trip data that had been collected in the HTS at the sampling rate of 3%. Review of other M/P studies imply that errors encountered in trip distribution models may have been caused by relatively poor quality of travel survey data such as lack of enough number of samples from high-income households and other biases arising from complex survey system and design. As for trip generation and modal split models, though further investigation is necessary, transferability of those models may not always apply in urban areas of the developing world due to the existence of external factors such as the Jakarta metropolitan area, even though the model structure may remain the same. On the other hand, for modal split models, it is understood that disaggregate model is widely applied due to accuracy, flexibility in analysis and smaller sample size. Disaggregate approach could also be applied for trip distribution models. In addition, disaggregation of first to third steps of the four-step method: namely, trip generation, trip distribution and modal split, that is, development of activity-based modeling of travel demand should also be discussed for application in metropolitan regions of the developing countries.

Jeff Newman (Cambridge Systematics, Inc., United States)
Rachel Copperman (Cambridge Systematics, Inc., United States)
Jason Lemp (Cambridge Systematics, Inc., United States)
David Kurth (Cambridge Systematics, Inc., United States)
Boris Lipkin (California High-Speed Rail Authority, United States)
Matt Henley (WSP, United States)
John Helsel (WSP, United States)
Gaussian Process Regression for Risk Analysis of Travel Demand Forecasts
SPEAKER: Jeff Newman

ABSTRACT. The work describes the use of Gaussian Process Regression (GPR) for risk analysis meta-modeling. GPR is a non-parametric approach that allows for better response surface fit than more traditional linear regression meta-models. The methodology is described, and an application to the ridership and revenue forecasts for the California High Speed Rail system is discussed.

13:30-15:30 Session 11F: Life Course -- Long Term
Margareta Friman (Karlstad university, Sweden)
Margareta Friman (Karlstad university, Sweden)
Jessica Westman (karlstad university, Sweden)
Lars Olsson (Karlstad university, Sweden)
Children’s Life Satisfaction and Satisfaction with School Travel

ABSTRACT. To understand children’s experiences of their daily travel, and the consequences of these experiences, it is essential that we directly address children. The Satisfaction with Travel Scale (STS) is a self-report instrument consisting of nine items divided into three subscales – two reflecting affective travel experiences and one reflecting cognitive travel experiences. The present study has two aims: (i) to examine the psychometric properties of a child version of the STS (referred to as the STS-C), and (ii) to test a potentially positive relationship between travel satisfaction and life satisfaction among children, something which has been found among adults. Three hundred and forty-five children completed the STS-C, life satisfaction scales, and sociodemographic variables. Analyses using Partial Least Square structural equation modelling revealed that the STS-C was internally reliable, had a sound construct validity, and confirmed a one-factor second-order measurement model with three first-order constructs (subscales). Furthermore, children’s satisfaction with school travel was also significantly related to their life satisfaction as measured by their satisfaction with: themselves, school experiences, friendships, family, and living environment. The relationship between travel satisfaction and life satisfaction varied between modes, whereby it was stronger among those who traveled by active modes than those who traveled by school bus or car. Younger children and boys were more satisfied with their travel to school, something which also had an indirect effect on their life satisfaction.

Angelique Bojanowski (Laval University, Canada)
Owen Waygood (Laval University, Canada)
Changes in young adults’ travel behavior
SPEAKER: Owen Waygood

ABSTRACT. Introduction With increasing concern about climatic changes, city planners look to implement measures to reduce car use. Among the possible measures, soft interventions are often used because of the advantages they have such as relatively low costs (as compared to infrastructure), greater acceptability from the public and thus easier for decision makers such as politicians to approve. Soft measures are based on theories of travel behavior and behavior change. Psychological profiling through the understanding of habits, attitudes, norms and intentions of the target population can improve their effectiveness. Travel behavior research has expanded and looks not only at daily travel patterns, but also at changes over one’s lifetime. This macro view is often termed the lifecycle approach. Through this approach, it is possible to find key transition points where an individual is more likely to make changes. For soft interventions, this approach is relevant as it identifies those points of transition where people seek new information, are more open to change, and habits can either be broken or new ones established (Verplanken and Wood, 2006). In this project, information will be “pushed” (i.e. provided, as opposed seeking) to young adults at different points of transition. Their evolving transportation patterns will be assessed to determine what type of information may be more effective in establishing sustainable, low-carbon travel choices. In this first study of the project, we will identify points of transition in the young adulthood (from 15 to 24 years old) of residents of Quebec City. Background The theories behind soft measures are well explained by Bamberg et al. (2011). One of them is the Theory of Planned Behavior (TPB) by Ajzen (1991). Ajzen’s model suggests that behavior is determined by intentions, which are influenced by attitudes, subjective norms (such as social norms and moral norm) and perceived barriers. In contrast to TPB, Triandis (1977) and Verplanken et al. (1994) argue that many routine behaviors such as daily travel are determined more by habit. When a habit is formed, people are more likely not to think about their mode choice. This could be a problem if you want people to change their behavior as they not think about the travel behavior as a real choice. As Verplaken et al. (1998) showed, it was individuals with greater variation in their travel patterns that were more likely to adjust their travel behavior to new conditions. In more recent work, Verplanken and Wood (2006) describe how the use of information might influence choices. One important point was that information is more likely to be effective if people have a context change like a new school, new home or job. Indeed, they tend to reevaluate their choices and then be more perceptive of information given to them. Recently, Lanzendorf (2003) and Scheiner (2007) looked at travel behavior with a new approach: mobility biographies. They showed that travel behavior is not just a repetitive behavior but also a behavior which happens during one lifetime. As such, it is interesting to look at patterns over one’s life. Lanzendorf (2003) found that travel behavior is stable until a key event happens. Those key events could be related to personal events like a new job or to societal changes like a new regulation. However, even though key events are common among different cultures like having a child, the changes in travel behavior seem to depend on the population of study. Objectives Determine the points in time where significant changes in travel patterns occur for young adults between the ages of 15 and 24. Materiel and Methods The data used in this survey is from the Origin-Destination survey of Quebec City from 2011. An analysis of variance (ANOVA) of the percentage of car trips by age between 15 and 24 years old for residents of Quebec City was conducted to identify the transition points. Following that, multiple comparisons were performed using the Tukey post hoc test to determine where statistical differences exist. The percentage of car trips over a day was calculated by summing all car trips and dividing by the total number of trips. The same analysis was performed for individuals who were identified as students. To further develop that, surveys are currently being completed by students in secondary (high) schools and CÉGEPs (like vocational or university-preparation schools) in the Quebec City region. That survey gathers data on attitudes and values related to the different modes, as well as their current habits of travel behavior based on Verplanken’s test. Data is not available for this abstract, but initial results will be presented at the conference. Results There is a statistically significant difference in the mean percentage of car trips between the different age groups for all young adults (p <0.001) and for those that were students (p <0.001). The percentage of car trips is shown in Figure 1 and one can note that it increases with age. Before the age of 17, less than 24% of trips are made by car as a passenger or a driver even though children at the age of 16 are eligible for a driver’s license. Seventeen appears to be a clear point of transition in the life of the residents of Quebec City regarding their travel behavior. It coincides with the end of secondary school (or high school). At 18, 63% of trips are made by car, 56% as a driver. Another point of transition appears to be 20 years old, coinciding with the end of CÉGEP. In Figure 1, the age groups with different letters (e.g. BC, DE, etc.) are significantly different from each other according to a pairwise comparison with the Tukey adjustment at 5% level. We can then categorize the people from 15 to 24 in three groups: before 17, 18 to 20 and 21 to 24 years old.

The percentage of car trips by students increases with age as well (Figure 2). One point of transition was identified at 17 years old. However, contrasting with the general population, students aged 18 and more have the same travel behavior. The maximum percentage of car trips observed is 63% at 18 years old and does not significantly increase from 18 to 24. Although it is the majority of trips, it is less than the average observed for the general adult population of Quebec City (83%).

Discussion and conclusion The percentage of car trips increases with the age of young adults, from as low as 20% up to 75% between the ages of 15 and 24. A very clear point of transition is 17 when the percentage of car trips increases to around 40% before jumping to 60% at 18. This point of transition would be a crucial time to implement soft interventions. At 17 years old, young adults finish secondary school and go to CÉGEPs, a pre-university or vocational school that is unique to Quebec (in Canada). This is also the age when they first move out for some of them, particularly when they attend specific programs of CÉGEPs (e.g. police officer, beauty technicians, mechanics, academics, etc.). It is important for them to be aware about the choices they have for transportation. It is very much the norm to use a car, but this social norm has numerous consequences for society (crashes, climate change emissions, etc.). Another point of transition is around 21 for young adults starting to work after attending CÉGEPs. When they continue school, the percentage of car trips doesn’t really change through time and stay around 60% a little lower than the average for workers in Quebec City (about 80%). It appears to also be a good time to push information about transport mode choices. There are limits to the current analysis. One is that the survey is still strongly based on having a landline that is less and less common for young people. Knowing that, the percentages of car use may not reflect the reality as young people with landlines may be more likely to live at home with access to a parent’s car. For example, previous studies with a general university population (Bojanowski, Craig-St-Louis, & Falardeau, 2014) found more general sustainable habits for the commute to university. The survey being conducted currently in secondary schools and CÉGEPs will allow us to construct profiles and test the influence of different information on the different profiles. References

Ajzen, I. (1991). The theory of planned behavior. Organizational Behavior and Human Decision Processes, 50(2), 179‑211. Bojanowski, A., Craig-St-Louis, C., & Falardeau, D. (2014). Comportements en matière de transport à l’Université Laval étudiés, représentés dans l’espace et inspirant l’élaboration de stratégies informatives efficaces (Essai (M.ATDR)). Université Laval, Québec. From Lanzendorf, M. (2003). Mobility biographies. A new perspective for understanding travel behaviour (p. 10‑15). 10th International Conference on Travel Behaviour Research, Lucerne. Scheiner, J. (2007). Mobility biographies: Elements of a biographical theory of travel demand (Mobilitätsbiographien: Bausteine zu einer biographischen Theorie der Verkehrsnachfrage). Erdkunde, 161‑173. Triandis, H. C. (1977). Interpersonal behavior. Brooks/Cole Pub. Co. From Verplanken, B., Aarts, H., van Knippenberg, A., & van Knippenberg, C. (1994). Attitude Versus General Habit: Antecedents of Travel Mode Choice1. Journal of Applied Social Psychology, 24(4), 285‑300.

Mengqiao Yu (University of California, Berkeley, United States)
C. Anna Spurlock (Lawrence Berkeley National Laboratory, United States)
Tom Wenzel (Lawrence Berkeley National Laboratory, United States)
Joan Walker (University of California, Berkeley, United States)
Two approaches to modeling changes in lifestyle-based behavior: from static to dynamic from exogenous to integrated
SPEAKER: Mengqiao Yu

ABSTRACT. Lifestyle is an important intrinsic motivation of people’s travel and activity behavior. The purpose of this research is to explore state of art methods that can identify and characterize long-term trends in lifestyle from historic integrated travel-related behavior (residential choice, car-ownership choice, car-usage choice, and mode choice). Two approaches are adopted. The first approach is to model the lifestyle trend using Hidden Markov Model (HMM) framework where each lifestyle is represented by one latent state and a multidimensional choice model is constructed for each latent lifestyle. The second approach is to treat the sequential lifestyle-generating decision making process as an inverse reinforcement learning (IRL) problem where we train a reward function inferred from historic behavior and use the function to predict future actions based on a given future expectation. Furthermore, we will discuss the pros and cons of the two approaches and apply them to two datasets, one from DMV registration data and the other one from an online retrospective survey. This research will both contribute to broadening the conceptual scope of lifestyle in the transportation behavior arena to a multi dimensional context, as well as enhance dynamic behavior modeling methodological frameworks with different machine learning approaches.

Bobin Wang (Beijing Jiaotong University, China)
Harry Timmermans (Eindhoven University of Technology, Netherlands)
Soora Rasouli (Eindhoven University of Technology, Netherlands)
Chunfu Shao (Beijing Jiaotong University, China)
Xun Ji (Beijing Jiaotong University, China)
Dynamic Decision-Making Process of Holiday Activity-Travel Behavior in Relation with Life Course Mobility: Using Decision Network Approach
SPEAKER: Bobin Wang

ABSTRACT. 1. Introduction With the expansion of transport infrastructure and improvement of people's living standards in China, holiday travel has become an inevitably demand for people pursuing high quality of life. According to the National Tourism Administration, there are 705 million domestic tourists during the “Golden Week” of National Day in 2017, with a year-to year growth of 11.9% compared with 2016. As a large number of travelers pouring into their destinations, holiday traffic congestion becomes more and more serious in large commercial centers and tourist attractions (People's Daily Online, 2017). Therefore, it is necessary to analyze the characteristics of holiday travel behavior and investigate the dynamic decision-making process, in order to make appropriate travel demand management (TDM) policy and alleviate traffic congestion in holidays. Different with the rigid demand of commuting on workdays, holiday travel demand is more elastic with flexibility in time and space. Thus its destination, activity time, travel mode choices are diverse and changeable. Holiday travel behavior has the following properties. First, a holiday travel union is usually a group, rather than an individual, so holiday activity-travel decision process involves multiple facets or portfolio choices concerning the group needs (Dellaert et al., 1998; Grigolon et al., 2013). Second, holiday travel choices may take a longer decision-making process and establish long-term agendas. Third, different holidays occur at discrete time points with different vacation time, thus it is difficult to investigate the dynamics of holiday activity-travel decisions in the short-term (day to day dynamics or week to week dynamics). Therefore, this study analyzes the dynamic decision-making process of holiday travel behavior from the long-term perspective. Life course approach provides a rich framework for better understanding of dynamic decision-making process. Van der Waerden et al. (2003) provided a conceptual framework for understanding the dynamics of activity-travel behavior, and argued that life course events and critical incidents allow people to reconsider and adapt their current activity-travel patterns. Lanzendorf (2003) put forward the mobility biography approach for longitudinal analysis of travel behavior. This seminal work led to a small but consistent stream of research, culminating in a recent workshop and book, in which the life-oriented approach was provided and expanded to quality of life (Zhang, 2012, 2017). In general, these developments in transportation research provide a new idea for the analysis of travel behavior dynamics. Empirically, there has been a substantial body of qualitative and quantitative research about the effects of life events on activity-travel behavior. However, most models take the utility of a particular activity-travel decision as a function of attributes, but neglect the interrelationships between influencing factors. The interaction between different activity-travel decisions is also seldom considered. Moreover, some studies give a prior specification of the model structure, which is not derived from the longitudinal data. In reality, holiday activity-travel decision-making is a complex process, involving dynamic, interdependent and time dependent relationships in it. Elaborating on Verhoeven et al. (2005), this study utilizes the Bayesian Decision Network approach to investigate the dynamic decision-making process of holiday activity-travel decisions in relation with residential mobility, household structure mobility, and car ownership mobility. The contributions of this paper are three-fold. First, a network model is developed by combining causal and temporal relationships into mobility biography analysis. Second, a decision network is learned from the retrospective longitudinal data to explore the dynamic interrelationships between multidimensional activity-travel decisions and life course events. Third, a new approach is provided to model the dynamic relationships across different time horizons of consecutive mobility decisions.

2. Methodology Life course mobility are the results of people’s life selections. Some mobility may change the spatiotemporal context that activity-travel decisions have to be made in, such as changes of residential location and work location. Other mobility constrain or expand individual’s choice sets, such as car acquisition or car disposal. Moreover, the time horizon of different mobility differs. Long-term mobility are related to changes of residential or work location, while mid-term mobility involve car ownership and household structure changes. Short-term mobility are related to activity-travel adjustment, with respect to travel mode, destination, route and activity time decisions. Dynamics, interdependence, and temporal effects are the properties of life course mobility examined in this study. First, life course mobility focus on the transition from one state to another state. Second, different decisions are rarely made in isolation, but are strongly interrelated. Third, these interrelationships may go forward (lagged effect), backward (anticipated effect), or be synchronic (contemporaneous effect). In order to consider the timing of every life course mobility into the model, “Time ago/future events” are defined as: “how many years ago/in the future a certain mobility occurred/will occur, with respect to target year. While, holiday activity-travel decisions only focus on the target year. Considering the influencing factors of holiday travel behavior, this study focuses on three life course mobility: residential mobility, household structure mobility, and car ownership mobility. Moreover, choice behavior is often context dependent, which is also affected by the external environment. Meanwhile, the external environment can also influence the life course mobility. Therefore, the formal modeling framework underlying this study is presented in Figure 1.

Figure 1 Modeling Framework

3. Data and survey In order to gather information about people’s long-term, mid-term, and short-term mobility over the life course, longitudinal data was necessary. A retrospective approach was used in this study, asking respondents to recall their life course mobility in chronological order. The survey was conducted at the household level, and the questionnaire recorded respondents’ life course from when they were 18 years old. If a respondent moved to Beijing when he/she was older, the respondent was requested to recall his/her life trajectory after arrival. The questionnaire recorded four life trajectories: residential trajectory, household structure trajectory, car ownership trajectory, and holiday travel behavior trajectory. Questions related to the four life trajectories included the total number of mobility, the time (year) when the mobility happened, and what exactly was changed by these mobility. The sample size for present analysis is based on 294 questionnaires with 5251 observation years. After the data cleaning and data reconstruction, the variables used in the decision network are shown in Table 1. The state “0” means the mobility happens in the observation year, and this state only belongs to the time ago events. The state “never” means this kind of mobility never happen. For example, there is no car in the family at the observation year, so this family cannot add or replace a car in the past, nor dispose of or replace a vehicle in the future. Holiday activity-travel decisions only focus on the mobility results on the target year. Because the life course mobility mainly influence the spatiotemporal context within the city, this study only consider the travel mode choices within Beijing.

Table 1 Explanation of Variables

4. Results and analysis Hugin software was used to build and estimate the Bayesian Decision Network. The network structure was learned from the retrospective longitudinal data, and the significance level was set to the standard value of 0.01 for the learning process. The holiday activity-travel decision-making process and dynamics relationships with life course mobility are shown in Figure 2.

Figure 2 Learned Bayesian Decision Network

The direct and indirect relationships between the life course mobility and holiday activity-travel decisions can be found through these links. Based on the time property of the node, the temporal relationships are distinguished by lagged effects, contemporaneous effects, and anticipated effects. These relationships are discussed below: Relationships between outer factors and life course mobility: Age directly links to household structure mobility, which indicates that specific mobility occur at specific age. Family monthly income mainly influence residential location changes in the future, and it is also affected by the household member changes in the past. Relationships between holiday activity-travel decisions and life course mobility: residential location changes, increase of family size, and car acquisition in the past have lagged effects on holiday travel mode choice. While holiday travel distance is mainly decided by the residential location changes. Changes in family size has an effect on the decision of activity time in holidays, and residential location changes in the future have anticipated effects on holiday activity duration. Relationships between holiday activity-travel decisions: there are also relationships between different holiday activity-travel decisions. Holiday travel distance is affected by travel mode choice and activity time, this is also confirmed by the study of Wang et al. (2017).

5. Conclusion Holiday activity-travel decision-making is a complex process, involving dynamic, interdependent and time dependent relationships in it. Very few studies have attempted to address the complexity of these dynamic decisions. This paper uses Bayesian Decision Networks to explore and analyze the dynamic decision-making process of holiday activity-travel behavior in relation with life course mobility, covering three holiday activity-travel decisions (travel mode choice, travel distance choice, and activity time choice) and three life course mobility (residential mobility, household structure mobility, and car ownership mobility). A web-based survey collected retrospective longitudinal data to provide valid information for these life trajectories. The model results provide evidence to confirm that activity-travel decisions are primarily triggered by a set of life course events and critical incidents (Van der Waerden et al., 2003). The causal and temporal relationships between different life course mobility have been explored, and their lagged, contemporaneous, and anticipated effects on holiday activity-travel decisions are analyzed. Moreover, this paper also investigates the interaction between different activity-travel decisions, which are usually ignored by other studies. The decision network model offers a framework for the dynamic analysis of direct and indirect effects of life course mobility on holiday travel behavior. By containing the temporal effects into the model, the probability of dynamic activity-travel decisions can be predicted using this network model.

Reference Dellaert, B. G., Ettema, D. F., Lindh, C., 1998. Multifaceted Tourist Travel Decisions: A Constraint-based Conceptual Framework to Describe Tourist's Sequential Choices of Travel Components. Tourism Management 19 (4), 313-320. Grigolon, A., Kemperman, A., Timmermans, H., 2013. Facet-based Analysis of Vacation Planning Processes a Binary Mixed Logit Panel Model. Journal of Travel Research 52(2), 192-201. Lanzendorf, M., 2003. Mobility Biographies. A New Perspective for Understanding Travel Behaviour. The 10th International Conference on Travel Behaviour Research, Lucerne. People's Daily Online, 2017. Van der Waerden, P., and H.J.P. Timmermans, 2003. Key Events and Critical Incidents Influencing Transport Mode Choice Switching Behavior: An Exploratory Study. Transportation Research Board, Washington, D.C. Verhoeven, M., Arentze, T., Timmermans, H.J.P. and van der Waerden, P., 2005. Modeling the impact of key events on long-term transportation mode choice decisions: Decision network approach using event history data, Transportation Research Record, 1926, pp.106-114. Wang, B., Shao, C., and Ji, X., 2017. Dynamic Analysis of Holiday Travel Behaviour with Integrated Multimodal Travel Information Usage: A Life-Oriented Approach. Transportation Research Part A: Policy and Practice 104: 255-280. Zhang, J., Tsuchiya, Y., Hinohara, H., and Chikaraishi, M., 2012. Citizens' Life Behavior and Quality of Life: Survey and Modeling. The 34th International Association for Time Use Research (IATUR), Matsue City, Japan. Zhang, J., 2017. Life-Oriented Behavioral Research for Urban Policy. Springer, Japan.

Nobuhiro Sanko (Grad School of Bus Admin, Kobe University, Japan)
Age-period-cohort analysis for alternative-specific constants in commuting mode choice models

ABSTRACT. Attached please find a pdf file.

13:30-15:30 Session 11G: New Survey Data and Analysis
Chandra Bhat (The University of Texas at Austin, United States)
Location: UCEN Flying A
Claude Weis (TransOptima GmbH, Switzerland)
Matthias Kowald (RheinMain University of Applied Sciences, Germany)
Kay W. Axhausen (ETH Zurich, Switzerland)
Milos Balac (ETH Zurich, Switzerland)
The Swiss National Stated Preference Study on Transport Behaviour 2015

ABSTRACT. 1. Introduction Every five years, the Swiss Federal Offices for Spatial Development (ARE) and Statistics (BFS) carry out the Microcensus Mobility and Transport (MCMT), a one-day CATI diary survey representative of the Swiss population in terms of socio-economics and trip characteristics (BFS and ARE, 2017). In the year 2015 (for the second time after 2010), a Stated Preference (SP) survey linked to the MCMT was carried out (ARE, 2016). Respondents selected and recruited during the MCMT interviews were asked to answer a follow-up paper and pencil questionnaire. This later survey instrument included items from on a combination of mode and route choice experiments based on one of the trips the respondents had reported during the MCMT CATI interview. The data, in combination with the Revealed Preference (RP) source of the MZMV, are primarily used in transport policy projects and for estimating mode and route choice models. Thus, they allow updates of regional and national transport models to current behavioral tendencies, and serve to obtain valuations of supply variables (travel times, etc.) to be used in cost-benefit analyses. The SP methodology allows for an assessment of respondents’ behavioral changes relative to changes in several different areas of the transportation systems. In addition, surveying information with similar SP-instruments in the years 2010 und 2015 allows an analysis of changes in peoples’ preferences and willingness to pay between the two years. The goals were: 1)design and carry out an SP survey that would closely resemble the MCMT in regards to the trip characteristics (mode, purpose, length) as well as the spatial and socio-economic properties of the respondents. This is of importance as the MCMT-survey population itself is with around 60´000 respondents representative for the Swiss residential population. 2)offer at the same time enough variation in all attributes in order to estimate significant parameters in the ensuing choice models and thus synchronize two important data sources commonly used for constructing and calibrating transport models. For the latter reason, priority was given to longer trips and trip purposes that have lower shares in the RP data. 3)design an SP experiment, which includes familiar and realistic situations for the respondents to collect reliable information. Filling out a survey based on their own RP-trips was expected to increase respondents´ interest, encourage them to imagine the presented alternatives, and reduce fatigue effects. As SP surveys induce substantial amounts of response burden by asking participants to imagine fictive situations and report their decisions, these issues are of high importance [Axhausen et al., 2014; Schmid et al., 2015]. 4)allow an analysis of the evolution of behavioral preferences as the SP experiments from the years 2010 and 2015 are comparable [ARE, 2012].

2. Methodology The SP survey was formulated as a Stated Choice (SC) experiment, in which the choice situations for each respondent were individually tailored to the trip that the person had reported in the MZMV interview. For that purpose, the real-world attributes of those trips were first determined, using network models for the car alternative and timetable data for the public transport alternative. Based on those values, variations to be used in each choice experiment were determined using an experimental design optimized with NGENE [Rose et al., 2008]. The respondents were then asked to fill out mode and route choice experiments. The attributes that were varied in those experiments for their respective modes were: •walk: travel time; •bike: travel time; •car: in-vehicle travel time, parking search time, fuel costs, tolls, parking costs, reliability; •public transport: in-vehicle travel time, access and egress times, waiting time, ticket costs, number of transfers, service frequency, capacity utilization, reliability. These attributes were varied in realistic ranges to obtain credible responses. Defining these ranges was done in expert workshops, including participants from science, transport suppliers and administration. In addition, capacity utilization was operationalized with the help of schematic images to foster a homogeneous interpretation of low, medium, high and overloaded capacities. In order to ensure better manageability of the relatively complex choice experiments, two subsets of the mentioned attributes were selected (maintaining a base set of variables common to both subsets), only one of which was displayed to each respondent. The combination of the availability of different modes (walk/bike according to distance, car according to the possession of a driving license), the mode of the reported trip chosen for the experiment, and the (randomly assigned) attribute subset lead to a total of 22 different questionnaire types, one of which was given to each respondent. The questionnaires contained eight choice experiments for the mode choice and/or eight for the (car or public transport) route choice, depending on the abovementioned circumstances. As has been mentioned above, prospective respondents to the SP survey were recruited at the end of the MZMV interview by the survey firm. The recruitment period lasted for a total of 16 weeks (2 weeks for a pre-test in April 2015, then 14 weeks between July and October 2015), during which a total of 6’099 respondents were chosen and agreed to participate in the SP. After the respondents had been recruited by the market research firm that carried out the CATI interviews, the corresponding data was delivered to the research team, who then calculated the trip attributes (see above), assigned the variations to be used in the SP experiments and constructed the questionnaire documents. Printing, dispatch and handling of those questionnaires was then again carried out by the survey firm. These steps were conducted within one week after the MCMT-interview to help people to remember the RP trip. A longer time period between RP-survey and SP-study would potentially have resulted in recall issues and resulted in less reliable decisions. After three weeks, reminder letters were sent to respondents who had not yet returned their questionnaires, thus improving response rates.

3. Results Out of the 6’099 recruited respondents, 4’693 effectively returned their questionnaires, resulting in a response rate of roughly 77%. The target of 4’000 respondents was easily achieved. The resulting data was coded and tested for representativeness and consistency. All analyses indicate a high data quality. The SP-study clearly reached its aim to represent the MCMT-survey population in terms of socio-ecomomics and travel behaviour. Issues like item non-response occurred rarely and only few respondents showed non-trading behaviour. These findings indicate a high reliability of the data and can be interpreted as a success of including familiar and realistic situations, the RP-trips, as the foundation for the SP-experiment. Prioritizing longer trips and trip purposes with lower shares in the RP data helped to obtain enough variation in the attributes to estimate the parameters in the ensuing choice models. To analyse the respondents’ travel behavior and their reactions to changing conditions, preliminary discrete choice models based on the Multinomial Logit (MNL) approach were estimated. Those models were based on the entire sample of RP as well as mode and route choice SP data. Non-linear interaction terms were used in the corresponding utility functions, reproducing the effects that had been observed in previous Swiss and other studies [Axhausen et al., 2007; Axhausen et al., 2008; Hess et al., 2007; Hess et al., 2008; Weis et al., 2010; Weis et al. 2012a; Weis et al., 2012b; Fröhlich et al., 2014; Widmer et al., 2017]. Thus, the study reached its aim to match RP-data from the MCMT and the SP-data. Considering the results from the first model estimations, the data may be considered a good foundation for the estimation of more differentiated (e.g. by trip purpose) and complex (e.g. Mixed Logit or Latent Class) models, to be carried out in a follow-up project that has started in the fall of 2016. For applications in the transportation modelling field, the combined RP and SP datasets will allow advancements and the gain of new information regarding the explanatory power of the choice behavior and the trade-offs between the different attributes.

4. Contribution of the paper The paper will include detailed information on the combination of the MCMT and SP-surveys and datasets, on the survey work and the necessary exchange processes to design and run the survey. Thus it will provide example of a practical application of SP experiments based on real-life behavior, while also discussing the strengths and weaknesses of the approach. Furthermore, it will contain a thorough analysis of the data, including descriptive as well as model based statistics. Here, the evolution of key values (such as the value of time) from 2010 to 2015 will be of special interest. To our knowledge, this is the first such comparison that is feasible on a national level and using large-scale data sets which were collected with quite similar survey instruments.

5. Literature ARE (2016) SP-Befragung 2015 zum Verkehrsverhalten, Bundesamt für Raumentwicklung, Berne. ARE (2012) SP-Befragung 2010 zum Verkehrsverhalten im Personenverkehr, Bundesamt für Raumentwicklung, Berne Axhausen, K. W., A. König, G. Abay, J. J. Bates und M. Bierlaire (2007) State of the art estimates of the Swiss value of travel time savings, Vortrag, 68th Annual Meeting of the Transportation Research Board, Washington D.C., Januar 2007. Axhausen, K. W., S. Hess, A. König, G. Abay, J. J. Bates und M. Bierlaire (2008) Income and distance elasticities of values of travel time savings: New Swiss results, Transport Policy, 15 (3) 173–185. Axhausen, K.W., I. Ehreke, A. Glemser, S. Hess, C. Jödden, K. Nagel, A. Sauer und C. Weis (2014) Ermittlung von Bewertungsansätzen für Reisezeiten und Zuverlässigkeit auf der Basis eines Modells für modale Verlagerungen im nicht-gewerblichen und gewerblichen Personenverkehr für die Bundesverkehrswegeplanung, FE-Projekt-Nr. 96.996/2011, BMVBS, Berlin. BFS and ARE (2017) Verkehrsverhalten der Bevölkerung: Ergebnisse des Mikrozensus Mobilität und Verkehr 2015, Bundesamt für Statistik and Bundesamt für raumentwicklung,. Neuchâtel and Berne. Fröhlich, P., C. Weis, M. Vrtic, P. Widmer und P. Aemisegger (2014) Einfluss der Verlässlichkeit der Verkehrssysteme auf das Verkehrsverhalten, Schlussbericht SVI 2012/003, Schriftenreihe, 1472, UVEK, Ittigen. Hess, S., A. Erath und K. W. Axhausen (2007) Reducing bias in value of time estimates by joint estimation on multiple datasets: A case study in Switzerland, Vortrag, European Transport Conference, Noordwijkerhout, Oktober 2007. Hess, S., A. Erath und K. W. Axhausen (2008) Estimated value of savings in travel time in Switzerland: Analysis of pooled data, Transportation Research Record, 2082, 43-55. Rose, J.M., M.C.J. Bliemer, D.A. Hensher und A.T. Collins (2008) Designing Efficient Stated Choice Experiments in the Presence of Reference Alternatives, Transportation Research B, 42 (4) 395-406. Weis, C., K. W. Axhausen, R. Schlich und R. Zbinden (2010) Models of mode choice and mobility tool ownership beyond 2008 fuel prices, Transportation Research Record, 2157, 86-94. Weis, C., M. Vrtic und P. Fröhlich (2012a) Schätzung der Modellparameter für das Gesamtverkehrsmodell Bern und das Gesamtverkehrsmodell Solothurn, BVE, Bern und BJD, Solothurn. Weis, C., M. Vrtic und P. Fröhlich (2012b) Schätzung der Modellparameter für das Gesamtverkehrsmodell Zürich und das Kantonale Verkehrsmodell Zug, AfV, Zürich und AfR, Zug. Widmer, P., T. Buhl, M. Vrtic, C. Weis, L. Montini und K.W. Axhausen (2017) Einfluss des Parkierungsangebots auf das Verkehrsverhalten und den Energieverbrauch, Schlussbericht SVI 2008/002, Schriftenreihe, 1596, UVEK, Ittigen.

Mathijs de Haas (KiM Netherlands Institute for Transport Policy Analysis, Netherlands)
Sascha Hoogendoorn-Lanser (KiM, Netherlands)
Distinguishing different types of observed immobility within longitudinal travel surveys: soft refusal vs. true immobility

ABSTRACT. In every travel survey there are respondents who did not report any trips on a given day. It is important to distinguish true immobile respondents from respondents who simply stated that they were immobile (soft refusers). In a previous paper three methods to identify soft refusers were proposed and tested. They seem to be succesful in identifying soft refusers. In this extended abstract we propose extensions to these methods in order to improve the reliability and make full use of the possibilities that longitudinal travel surveys offer. To test the methods, data from the Netherlands Mobility Panel (MPN) is used.

Hirohisa Kawaguchi (Oriental Consultants Global Co., Ltd., Japan)
Sadayuki Yagi (Almec Corporation, Japan)
Sahrial (PT. SAKA ADHI PRADA, Indonesia)
Donny Cleo Patra Pakpahan (University of Indonesia, Indonesia)
Sample Size of Household Travel Survey in Metropolitan Regions in Developing Countries – A Case Study in Jakarta Metropolitan Region

ABSTRACT. Introduction – Why HTS sample size matters in developing countries? Urban transportation problems are evident, critical and complex in developing countries, especially in emerging economies, due to rapid economic growth together with motorization and urban sprawl in spite of limited resources to solve the problems. For urban transportation studies, aggregate model is dominantly utilized in practice in developing countries apart from developed countries except for a few examples such as Beijing, China. This means that sample size requirement is significant due to aggregation process compared with disaggregate model. In addition, sampling size is sometimes determined mainly by budget constraint and past “examples” or “beliefs” such as a few thousand households without considering required accuracy for a study. When it comes to sampling theories of aggregate model, extensive studies have been done in ‘70s ‘and ‘80s in developed countries1) such as Smith2) and Purvis3). However, it should be noted that these studies are under several assumptions in developed countries which is not always applicable in developing countries such as automobile as a major mode of transportation. Other issue of previous theoretical studies is that impact of sample size reduction is not examined considering cumulative errors of four-step model using actual data in developing countries. The same as other cities in developing world, the Jakarta metropolitan area (Jabodetabek) also faced the issue of survey design for updating urban transportation master plan under various constraints. Fortunately, large-scale surveys, HTS of 167,000 households in 2002 and Commuter Survey with 180,000 households in 2010, have been conducted. However, there is funding issue for this kind of large-scale surveys. While the HTS implementation is a bottleneck of updating the master plan, there is no clear technically sound method on how to design survey and to develop travel demand model for the region. Therefore, objective of this paper is to examine methodology of HTS survey design mainly focusing on sample size determination in developing countries taking Jabodetabek as an example through 1) identifying issues of sample size determination with review of theories and practical examples, 2) examining impact of reduction of sample size through four steps of modeling and 3) discussing direction of survey design of urban transportation studies. Review of Theories and Practical Examples In terms of estimation of an origin-destination (OD) table and development of an aggregate model, Smith2) is often referenced. As over 4% sampling rate is required to estimate 1,000 trips with allowable error of 250 trips and confidence level of 90%, it was concluded that estimating an accurate current OD table is not realistic. It also mentioned that data of 1,000 household or less might be enough to estimate trip generation, distribution and mode choice model by examining each step of the model independently under the assumption that distribution of personal and household attribute is random. In case of studies in developing countries conducted by Japan International Cooperation Agency (JICA), further simplified equation of sample size determination assuming that all trips are equally distributed to each OD pairs has been widely utilized4). In short, theoretical examination is generally based on several unrealistic assumptions in developing countries. With regard to a disaggregate model, Rose and Bliemer5) concluded that sample size of disaggregate model is depending on a critical parameter of a model. de Bekker-Grob et al.6) proposed method of sample size determination for discrete choice model while it requires to develop a model based on a pilot study before implementation of a main survey. According to practical studies in various countries, a few hundred to 1,000 samples are considered as sufficient for estimating discrete choice model for transport analysis7). 12 practical examples of urban transportation studies in developing countries conducted by JICA8)-18) are also reviewed. It highlighted several issues in HTS implementation and sampling in developing countries. In many cities, high income households, which are usually car-owing household, are difficult to be surveyed due to high security level and rejection by housemaid. Base data for sampling such as census data is also sometimes not available. In addition, some studies faced issues in quality control of a number of HTS surveyors such as cheating, use of backup sample without visiting main sample and so forth. It also noted that all models are aggregate models. Jabodetabek is also facing these issues. Case Study of Aggregate Model in Jabodetabek In response to the request of the government of Indonesia, JICA conducted “The Study on Integrated Transportation Master Plan (SITRAMP)” 19) in Jabodetabek from 2001 to 2004. Among 11 transport surveys, HTS was a large-scale home interview survey of household daily travel with a sample size equal to 3% of Jabodetabek population of 21.6 million in 2002. This case study examines the processes of transport modeling and demand forecasting based on the two datasets that were created by grouping odd and even household sequential serial numbers (HHSENOs). It was an efficient way of splitting the dataset into two with locational and socioeconomic information as well as other characteristics equally distributed. Travel demand models and their outputs based on the two datasets were compared with each other and with outputs of the full dataset. In this way, nearly the same models and their synthesized results were expected. Trip generation is an important step since it can generate a control value for the number of trips generated in the study area. Trip purposes are divided into eight home-based trip purposes, following the method in SITRAMP. The samples are also broken down into area determined by the population density (i.e., urban and rural areas) and the household income group (i.e., high-, middle-, and low-income groups). Variables such as population, number of employees and students at residential place, number of employees at working place, and number of students at school place are used for trip production and attraction models. The production/attraction models by trip purposes for urban and rural areas developed from the trip data of high-income households shows lower values of coefficients of determination (R2) because of the limitation of the data compared to other household income groups. Finally, synthesized number of trips by purpose from the models developed with a dataset with odd or even HHSENOs can be cross-validated with the observations from the other dataset. The ratios of synthesized data to the observations ranged from 0.91 to 1.22. Trip distribution for the urban trips and rural trips is developed separately with a gravity model. Network skim for intrazonal trips for each zone is made by a set to half of that to the nearest neighboring zone. Trip purposes are classified into combinations of five basic purposes and three household income groups. Comparison block OD matrices (on a city/regency level) between synthesized trips from the models developed with a dataset with odd or even HHSENOs and the observations from the other dataset revealed some discrepancy. Since the largest share (45%) of the trips are intra-Jakarta trips and inter-city/regency trips are relatively small in number compared to the intra-city/regency trip, this comparison of ratios of synthesized data to the observations could cause more discrepancies, which may not be useful for transportation planning if OD movements (such as desire lines) are focused on. Therefore, impact on the transportation network should also be analyzed in the network assignment. In fact, if compared on a TAZ (transportation analysis zone) level, ratios of synthesized trips from the models developed with one dataset to the observations from the other dataset greatly fluctuate from OD pair to OD pair; whereas, block OD pairs with many trips are not significantly affected. For modal split model, observed person-trip data are stratified into several binary choices together with the corresponding zone-to-zone impedance measures: namely, choice between non-motorized and motorized modes of transport (whereby the share of non-motorized trips varies depending on the distance), choice between public (transit) and private modes of transport (whereby the share of transit trips varies depending on the ratio of generalized time), and choice between motorcycle and automobile mode of transport (whereby the share of motorcycle trips varies depending on the distance). If the modal split summary of synthesized trips from the models developed with one dataset is compared with the observations from the other dataset for each combination of the five basic purposes and the three income classes, the shares are quite close to each other with differences mostly within 1.5%. Even though some differences are observed in the diversion curves, they stand out in longer-distance trips, which comprise only small percentages of the total trips in Jabodetabek. The methods used for network assignment is “multi-user class incremental assignment” on the generalized cost of travel. The user classes for highway assignment are: motorcycle trips by three income classes, auto trips by three income classes, and public transport (i.e., bus volumes as pre-loads). Person trips are also assigned on the transit network by three income classes. Since observed traffic count locations were limited, network assignment results are examined by comparing outputs from the assignment of the synthesized OD matrices from the models developed with a dataset with odd or even HHSENOs. The comparison result shows some differences in the traffic and passenger volumes from the two datasets, but they are mostly within 3% and hence relatively small (See the attached figure). However, if traffic flow around local intersections is focused on and the inflow volumes are compared, the differences become more striking. In conclusion, if the sampling ratio is reduced from original 3% to 1.5% and the sample size becomes half, aggregated results including generated trips, modal shares, and total traffic and passenger volumes on roads and transit lines would bring about the results that are close to those from the original dataset. Meanwhile, significance of OD matrices may fall especially when ODs between TAZs are focused on for transportation planning. Thus, it is essential to take this conclusion carefully when conducting a more in-depth analysis such as traffic flow on a local network and operation planning for transit lines including station design. Discussions and Conclusions While it is hard to maintain accuracy of OD table in TAZ level, odd-even dataset of block OD tables are more or less similar according to the case study. It is sometimes argued that there are cumulative errors of four-step model. On the contrary, it is noteworthy that link traffic volumes of aggregate assignment result of odd and even data set are almost equal. This might be explained by road capacity of each link as complemental information and averaging and offset of various errors in data and models during assignment process. Further studies are required for in-depth analysis of combination of errors in OD tables, road network and assignment process. In Jabodetabek, sample size has been determined to estimate current trip generation/attraction and OD table in the previous studies, however, estimating OD table with resolution of TAZ level requires significant number of samples while complementary information and offset effect of errors are expected during assignment process. In addition, future travel demand forecast is not dependent on current OD table but demand forecast model. If the objective of sample size determination is developing models, disaggregating model might be appropriate considering budget and time constraint, quality control of large-scale survey management and information loss due to aggregation process. The sampling strategy of Jabodetabek assuming disaggregate model is being studied taking characteristics of developing countries into consideration toward updating the master plan. One of biggest issues of survey design is collecting high income households who are dependent on private mode of transport and avoiding quality control issue of a large-scale survey. The critical point of determining a sample size for developing disaggregate models is high income households which are approximately 7% of entire population according to the study in 201015). As a few hundreds to 1,000 persons are required to develop person-based disaggregate model, 4,000 households sample might be enough assuming average household size of 4 persons per household and reserve sample of around 10%. Complementary survey targeting high income households such as survey at shopping mall and work place might work while further studies are awaited.

Parvathy Vinod Sheela (University of South Florida, United States)
Suryaprasanna Kumar Balusu (University of South Florida, United States)
Michael Maness (University of South Florida, United States)
Abdul Pinjari (Indian Institute of Science, India)
When Neutral Responses on a Likert Scale Do Not Mean Opinion Neutrality: Accounting for Unsure Responses in a Hybrid Choice Modeling Framework

ABSTRACT. Introduction Integrated choice and latent variable (ICLV) models have proven to be a remarkable enhancement to discrete choice models with its explicit consideration of socio-psychological factors that can lead to improvements in analysts’ ability to predict outcomes in choice data. The use of Likert scale questions to psychometrically measure attitudes and perceptions is the predominant data collection method used in transportation. Often, these Likert scales are bipolar with a neutral option in the middle of the scale. Modelers have predominantly modeled these Likert scale indicators as ordered responses (continuous or ordered discrete variables). The responses are modeled on a continuum from one extreme (e.g. unlikely) to another extreme (e.g. likely), and thus the middle/neutral response acts as a transition point between the two polar options.

But psychometric research has found that the neutral group of respondents who choose the middle option in a Likert scale is not homogeneous. These respondents are not all truly opinion neutral and thus do not act as a transition group between these extremes (Sturgis et al. 2014, Kalton et al. 1980). This respondents who choose the neutral/middle option often fall into two groups: (1) those individuals who possess true opinion neutrality on the issue and select the neutral option and (2) those individuals without adequate knowledge who choose the neutral option as a way of saying that they do not know or have no opinion (Sturgis et al. 2014). The latter group is not considered in existing implementations of Likert scale indicators in ICLV models. This is due to how existing models treat all neutral responses are opinion neutrality since the response is on a continuum.

Thus, there is a gap in the modeling methodology for distinguishing between opinion neutrality and lack of knowledge and opinion. Hence, there is a need for indicators to be modelled as unordered responses instead of just treating them as ordered responses. In this study, we aim to develop a framework to distinguish opinion neutrality from lack of knowledge/opinion in ICLV models by incorporating unordered response models for the Likert scale indicators. A case study on intended autonomous vehicle use is used to explore the framework’s properties.

Case Study: Intended Mode of Autonomous Vehicle Usage A tremendous increase in the share of autonomous and connected vehicles expected in the foreseeable future can change travel patterns and traffic conditions. Such changes on the transportation network will be greatly influenced by the adoption rate and usage mode (e.g. own, rent, lease, personal or share ride) of autonomous vehicles (AVs). The acceptance of AVs and its different modes depends upon the public perception towards such advanced technologies. These perceptions, directed by people’s attitudes and beliefs, will affect the rate of integration of this technology into the transportation network. Hence, a clear understanding of how these perceptions affect AV adoption is important in accurately modeling consumers’ choice behavior. In this context, modelling the perceptions of benefit and concerns as latent variables can shed some light on its impact on individuals’ decision making process. In this study, ICLV models are used to model the intended mode of AV use. It should also be noted that familiarity with AVs is not high among the general public due to the novelty of the concept. Thus, we should expect to see a significant number of people who would express a lack of knowledge and opinion (“don’t know”) and thus a framework to correct for this is needed to account for possible bias in assuming opinion neutrality.

This study uses a survey designed to collect data on consumers’ perception and intended adoption of autonomous vehicles (Menon, 2015). The survey was distributed to two different samples. The first sample consisted of 1157 respondents from the University of South Florida. The respondents included university lecturers, students, researchers, and staffs who were given a web-based survey. The second sample consisted of 2338 responses from AAA members distributed among the members of AAA South. In the survey, the perceptions of various attributes of AV technology and consumers’ perceptions of the benefits and concerns with AVs are obtained through a five-point Likert scale. Figure 1 shows the observed perceptions of benefits and concerns for each stated category.

As can be seen in the Concern chart (Figure 1), the descriptive statistics show that the proportion of neutral/middle responses is similar between all the concern categories. The Benefits chart (Figure 1) shows a similar relationship although not as strong. The top three categories show similar neutral/middle response proportions, whereas the remaining categories show a smaller proportion of neutral responses (but similar proportion between these categories). The stability of the neutrality proportions motivates the use of this dataset to explore differentiating opinion neutrality and the lack of knowledge/opinions among respondents.

Methodology Two approaches for modelling the indicators are used and compared in this study: (1) assumed homogeneous opinion neutrality and (2) heterogeneous opinion neutrality and lack of knowledge/opinion. In the first approach, the traditional method for modeling the measurement equations as an ordered response is used. Here the indicators of benefit and concern assume homogeneous opinion neutrality. They are measured as ordered responses on a Likert scale – e.g. extremely likely, likely, neutral, unlikely and extremely unlikely – and the choice model is modelled using a Multinomial Logit model (Ben-akiva et al. 2002; Yáñez et al. 2010). In the second approach, we aim to develop a framework to model the indicators as a combination of ordered/continuous and unordered responses. By using data on their stated prior knowledge of autonomous vehicle together with their responses to the Likert scale questions on benefits and concerns, the framework will be able to disentangle opinion neutrality from knowledge/opinion level. We are currently exploring three techniques to achieve this disentanglement.

Assumed Homogeneous Opinion Neutrality ICLV: Method and Results In this study, a latent variable approach is used to model a decision-making process incorporating socio-psychological factors such as attitudes and perceptions which are vital cognitive elements in the decision process. Here, the most preferred way to use fully autonomous AVs is modelled using an ICLV approach. The hypothesis is that an individual faced with a choice of preferred ways of using AVs will assess the impact of the benefits offered by them, the issues while using them on the roadway, and convenience offered by them to decide the mode of AV usage. The structure of the traditional ICLV model with assumed opinion neutrality is given in the figure shown below.

Results from models using the USF and AAA data are discussed briefly here. The latent variables considered, benefits and concerns, have the expected impacts on future AV adoption. The concern latent variable is negatively associated with all modes of using the AVs, being insignificant in owning and renting, but significant in sharing the owned AV. Respondents who embrace AV benefits have a higher likelihood of embracing AV technology for personal use, rather than AV sharing or AV renting. While education does not have a statistically significant impact (though education does play a role through the latent variable structural equations) in owning, it does have a positive impact in renting an AV. Women (compared to men) were more likely to express concerns about the technology and less likely to consider the benefits of the technology in their preferences. Men are more inclined (than women) to be interested in both AV ownership and personal use. Younger adults aged less than 40 years old appear to be more inclined towards owning AV and sharing it than older people, which is consistent with the findings reported by Levire et al. (2017) who note that Generation X and millennials are likely to be the early adopters of advanced technologies. Involvement in a crash does not have a statistically significant impact on owning an AV and using it for personal use even though it plays a role through the latent constructs.

Heterogeneity in Opinion Neutrality and Lack of Knowledge: Expected Method and Results The main motivation behind this approach is to analyze the choice behavior of not only the bipolar groups, but also the neutral and lack knowledge/opinion groups in the sample. The hypothesis is that the neutral/don’t know/can’t say group in the middle of Likert scale cannot be considered as solely a transition group from likeliness to unlikeliness / concerned to not concerned. But, respondents denoting middle responses may also exhibit a lack of knowledge/opinion. The three frameworks we will explore to disentangle opinion neutrality from lack of knowledge/opinions are described briefly as follows:

1. Adding a knowledge level latent variable. This approach follows most closely to existing research. The knowledge level latent variable would include the continuous/ordered indicator from the AV familiarity (knowledge) question. Then unordered indicators (e.g. binary probit models) of the choice of the middle option will be added for each benefit and concern category. 2. Latent Class Model with Indicators. In this approach, four latent classes are implemented representing opinion and knowledge types. The classes will denote those who: (1) perceive high benefits for the AV, (2) are concerned about AV technology, (3) have a neutral opinion of AV benefits and concerns, and (4) lack knowledge or lack an opinion of AVs. As in a traditional latent class model, the class membership model acts as a structural equation for the latent classes (i.e. a discrete latent variable). But, we can use the psychometric indicators to gain additional information on the likely classes that individuals would fall into. 3. Combined Latent Variable Model with Latent Classes. In this approach, an intermediary step in the above Framework #1 between the latent variable and the choice utilities is introduced. The latent variables will feed into the class membership model of a latent class structure (likely with a similar four class structure as denoted in Framework #2). In this way, the latent variable will induce taste variation between respondents of differing opinion/knowledge types.

These approaches will increase the dimensionality of the estimation task, so Multinomial Probit kernel-based (Bhat and Dubey (2014)) ICLV formulations with used. This will provide a common platform to compare model estimation and model results. Additionally, the MNP kernel also allows the use of unordered responses (Bhat et al, (2016)) for indicators – beyond simple binary responses. This could allow us to test representing the indicators as only unordered responses rather than ordered/continuous. We intend to explore this avenue of incorporating indicators into ICLV models as well. We expect to present the estimation results of these models and to discuss their interpretation and practical usage.

Frameworks such as this will allow modelers to account for the proportion of individuals without knowledge of AVs. This enables these frameworks to assist policy makers in designing information campaigns for the public to increase AV awareness. The structural equations for the latent variables will enable modelers to determine which segments of the public are least informed about AV technology. This will allow for more efficient allocation of awareness campaign funding.

References Ben-Akiva, M., McFadden, D., Train, K., Walker, J., Bhat, C., Bierlaire, M. & Daly, A. (2002). Hybrid choice models: progress and challenges. Marketing Letters, 13(3), 163-175.

Bhat, C. R., & Dubey, S. K. (2014). A new estimation approach to integrate latent psychological constructs in choice modeling. Transportation Research Part B: Methodological, 67, 68-85.

Bhat, C. R., Pinjari, A. R., Dubey, S. K., & Hamdi, A. S. (2016). On accommodating spatial interactions in a generalized heterogeneous data model (GHDM) of mixed types of dependent variables. Transportation Research Part B: Methodological, 94, 240-263.

Kalton, G., Roberts, J., & Holt, D. (1980). The effects of offering a middle response option with opinion questions. The Statistician, 65-78.

Lavieri, P. S., Garikapati, V. M., Bhat, C. R., Pendyala, R. M., Astroza, S., & Dias, F. F. (2017). Modeling Individual Preferences for Ownership and Sharing of Autonomous Vehicle Technologies (No. 17-05843).

Menon, Nikhil, "Consumer Perception and Anticipated Adoption of Autonomous Vehicle Technology: Results from Multi-Population Surveys" (2015). Graduate Theses and Dissertations.

Menon, N., Pinjari, A. R., Zhang, Y., & Zou, L. (2016). Consumer Perception and Intended Adoption of Autonomous-Vehicle Technology: Findings from a University Population Survey. In Transportation Research Board 95th Annual Meeting (No. 16-5998).

Sturgis, P., Roberts, C., & Smith, P. (2014). Middle alternatives revisited: how the neither/nor response acts as a way of saying “i don’t know”?. Sociological Methods & Research, 43(1), 15-38.

Yáñez, M. F., Raveau, S., & Ortúzar, J. D. D. (2010). Inclusion of latent variables in mixed logit models: modelling and forecasting. Transportation Research Part A: Policy and Practice, 44(9), 744-753

Chandra Bhat (The University of Texas at Austin, United States)
Gopindra Nair (The University of Texas at Austin, United States)
Ram Pendyala (Arizona State University, United States)
Alternative Decision Mechanisms to Model Ranking Data
SPEAKER: Chandra Bhat

ABSTRACT. The econometric and survey literature has suggested multiple times that ranked data are not as reliable as data obtained from eliciting only the most preferred alternative. But, as we demonstrate in this paper, the issue is more likely to be associated with the rank-ordered logit model used in earlier studies for analyzing ranked data. We then apply a rank-ordered probit model to three travel-related ranking contexts to highlight the fact that ranked data can actually provide reliable and very useful information in surveys.

13:30-15:30 Session 11H: Emergency and Event Transportation
Junji Urata (Kobe University, Japan)
Location: UCEN Lobero
Junji Urata (Kobe University, Japan)
Eiji Hato (The University of Tokyo, Japan)
Incorporating a Dynamical Heterogeneity into Expected Utility of Dynamic Discrete Choice Model - With a Case Study of Tsunami Evacuation
SPEAKER: Junji Urata

ABSTRACT. This paper proposes a dynamic discrete choice model of tsunami evacuation behavior that accounts for the heterogeneity dynamics of expected utility. Individuals decide to evacuate in order to avoid future risk but they only have an estimate of their decision’s expected utility based on perception, which is different from the true expected utility. Our proposed algorithm can estimate parameters by taking into account the heterogeneity of expected utility. The results of our case study validate the existence of heterogeneity dynamics. It is important to evaluate the heterogeneity of expected utility because the behavior of individuals, even in normal times, is often affected by heterogeneity.

Nima Golshani (University of Illinois at Chicago, United States)
Ramin Shabanpour (University of Illinois at Chicago, United States)
Joshua Auld (Argonne National Laboratory, United States)
Kouros Mohammadian (University of Illinois at Chicago, United States)
Hubert Ley (Argonne National Laboratory, United States)
Analysis of Evacuation Tour Formation for No-Notice Emergency Events
SPEAKER: Nima Golshani

ABSTRACT. Disasters, either natural or man-made, occur when an extreme event exceeds people’s ability to cope with. They can result in severe infrastructure damage and loss of life. These events can be categorized into two groups based on their predictability as advance-notice and no-notice emergency events. Advance-notice events refer to disasters which can be predicted days before their occurrence, such as hurricanes and tornados. In such situations, authorities warn the public so that people can plan for the evacuation, if necessary. On the other hand, no-notice emergency events such as terrorist attacks or earthquakes are not predictable and thus, there is no time to develop evacuation plans. Therefore, pre-disaster planning gains significant importance in these situations. To investigate evacuation behavior during no-notice emergency events, we recently proposed an evacuation demand module to be added to the Agent-based Dynamic Activity Planning and Travel Scheduling (ADAPTS) activity-based framework (Auld and Mohammadian, 2012). ADAPTS is an agent-based microsimulation travel demand model, which simulates the process of planning, scheduling, and execution of activities for individuals over a specified timeframe. The proposed behavioral evacuation demand model is only called if a disaster has happened in that simulation time step. Once a disaster happens the evacuation model is called to estimate the new demand and update each individual’s activity schedule. The framework comprises four steps: (i) evacuation decision, (ii) evacuation planning, (iii) tour formation, and (iv) scheduling. The first step deals with the decision to evacuate; individuals can decide to ignore the emergency and thereby continue their previously planned schedule, or they can choose to shelter in place where their schedule is updated with an impulsive indoor activity. Individuals can also choose to evacuate and therefore the corresponding attributes (departure time, destination, and mode) of the new activity need to be determined by the models designed for the evacuation procedures; this is done in the evacuation planning phase of the model. The third phase, which is called tour formation, corresponds to identifying the total number of stops, type of stops, and the total distance/travel time of individuals’ evacuation tours. The outcomes of the evacuation planning phase should be able to influence the models of the tour formation phase and vice versa. Finally, the estimated evacuation demand is put into people’s schedule to be executed by ADAPTS. In this study, we specifically focus on the tour formation phase of the proposed model. To that end, we first estimate a joint ordered-continuous-continuous model of total number of stops, total distance, and total travel time of the evacuation tours. In addition to the interrelation between these three variables, the joint structure is proposed because the number of intermediate stops endogenously affects evacuees’ total travel time and distance. The second step corresponds to estimating a rank ordered logit model to analyze the type of the intermediate stops (e.g., pick-up family members, shop for supplies) in individuals’ evacuation tours. The rank ordered logit model takes the estimated number of intermediate stops for each individual from the first step and estimates the type of each stop.

Junji Urata (Kobe University, Japan)
Eiji Hato (The University of Tokyo, Japan)
Identification of Local Interaction and Asymmetric Impact: Application to Evacuation Departure Behavior
SPEAKER: Junji Urata

ABSTRACT. This study focuses on local interaction in a one-to-one network and its asymmetric impact. Formulae are described for the specification of asymmetric impacts and the importance of asymmetric impacts is validated through an analysis of evacuation departure behaviors during a heavy rain disaster. The introduction of attribution of one-to-one link and different parameters between choices can clarify the influence of local interaction. The empirical analysis show that it is not possible to specify the existence of local interaction if these effects are neglected.

Xiaoyang Wei (College of Architecture and Urban Planning Tongji University, China)
Wei Zhu (Tongji University, China)
Using the time utility model to identify the spatio-temporal behavior patterns of large-scale exposition visitors: A case study of 2014 Qingdao International Horticultural Exposition

ABSTRACT. Planning of large-sclae exposition (LSE) park is a complicated task. Because the visitors’ experiences and safety are important concerns of the exposition administration, it is necessary to estimate the planning’s influence on the visitors’ spatio-temporal behavior with reliable understandings of the visitors’ behavioral patterns and their relationships with the park environment.

Studying the patterns and regularities of the visitors’ spatio-temporal trajectories has been challenging due to their highly heterogeneous nature. At present, there are two kinds of research paradigms about the regularities of the visitors’ spatio-temporal behavior in LSEs or similar environments, such as shopping street. One is to reveal the preferences underlying the visitors’ spatial choices(e.g., Zhu et al. 2008), and the other is to induce the patterns of the spatio-temporal behavior(e.g., Kurose et al. 2001). The main advantage of the spatial preference approaches is the capability to predict the spatio-temporal behavior, while their main disadvantage is that the perspective of behavior interpretation is local, lacking strategic elements. The main advantage of the pattern induction approaches is the capability to reveal the visit strategies, but they also have a main disadvantage of being limited at predicting the behavior.

The purpose of this paper is to propose and validate a new method that further excavates the mechanisms underlying the induced spatio-temporal behavior patterns, trying to combine the two approaches and take their advantages.

Based on a multinomial logit model incorporating time-varying utilities (the time utility model, TUM, Zhu et al. 2006), this method is validated and applied using a data set of 624 visitors’ behavior collected through a questionnaire survey in the 2014 Qingdao International Horticultural Exposition. The expo was held from April 25 to October 25, 2014. The data were obtained by interviewing the visitors when they finished their trips and exited the park. The respondents reported their entrances and entry time, the exits and exit time, the visited places in sequence, the activities and their transportation modes.

For the validation, a trajectory similarity measure is constructed first based on a relative spatio-temporal position alignment algorithm, taking into account the number, visiting order and positions of the visited places. Affinity propagation clustering algorithm is used to classify the trajectories. Five types of the spatio-temporal patterns are induced, which differ in the number of stops, length of paths, spatial extent, and most importantly, the visit strategy.

The TUM is then used to explain the patterns. Five models are estimated using the sub-samples corresponding to each pattern group, with place attractiveness, distance to the place, times already to the place, and choice of going home as explanatory variables, also incorporating the elapsed time since the entry and the current time to capture the parameter variations. The time use of each visitor during the whole trip is estimated based on the visitor’ entry time, exit time, and his/her trajectory, assuming a walking speed of 1m/s and that the time spent at each place is proportional to the attractiveness.

The model results show that the visitors’ spatio-temporal preferences, represented by the model parameters, differ significantly between the five samples, which explains the strategic characteristics of each pattern quite well and proves the validity of the TUM. For example for the first pattern, the positive utility of place attractiveness is minimum and the negative utility of distance is maximum, which explains the characteristics that these visitors move in short distances, small ranges and have lower probabilities of visiting the large pavilions far away. In addition, the relatively high initial utility of going home explains their least visit time and the shortest path lengths on average. The validity suggests that the TUM can directly be used to identify the visitors’ spatio-temporal behavior patterns without inducing the patterns first. It is also a more flexible and generalized approach to capture more complex patterns.

For the identification, an algorithm for classifying the trajectories according the visitors’ preferences is developed. To avoid the problem of model estimation failure due to insufficient data, the algorithm first decomposes the whole sample into undividable sub-samples, each of which includes some trajectories that are similar in the preferences, then clusters them using a hierarchical clustering algorithm with a specifically devised distance measure taking into account the parameter differences and the credibility of the differences. Optimal number of clusters is determined using model selection criteria, specifically the Bayesian Information Criterion.

Using the same data set, six types of the visitors’ spatio-temporal preferences are derived and analyzed. For example, the first pattern resembles the nearest-destination-oriented (NDO) strategy (Kurose et al., 2000), by which the visitor moves in shorter distances in the beginning; the visitors of the second pattern most likely visit attractive places in the beginning, because the attractiveness parameter is the maximum, resembling the attractive-street-oriented (ASO) strategy; the distance parameter of the third pattern is positive in the beginning, which suggests that the visitors prefer farther places, a good correspondence with the farthest-destination-oriented (FDO) strategy.

Overall, the results reveal the complexity of the visitors’ spatio-temporal patterns and preferences in LSE. First, time has significant impacts on the visitors’ preferences, therefore using a single model to explain all the visitors’ behavior is too sketchy. Second, the regularities of the preferences’ temporal variation are heterogeneous. For a same factor, the utility may increase, decrease or remain stable over time, and the initial utilities could also differ, which is a more essential cause of the diversity of the patterns. Third, it shows the defect of inducing the patterns from the trajectories only, because the behavior is constrained by the specific environment, such as different entry points. There may be a same preference pattern underlying different behavior patterns, and in turn, a same behavior pattern may originate from different preference patterns. For example, the visitors of the ASO preference pattern and the visitors of the FDO preference pattern both behave as heading straight to the Theme Pavilion after entering the park, due to the fact that the Theme Pavilion is located in the farthest location from the main entrances.

To summarize, the study finds that the visitors’ spatio-temporal preferences determine their spatio-temporal behaivor patterns, and more general regularities of the visitors’ spatio-temporal behavior can be obtained through the TUM, which has the potential to be used to understand individual spatio-temporal behavior in other situations more deeply and meticulously.

Yuhan Gao (Kyoto University, Japan)
Jan Dirk Schmoecker (Kyoto University, Japan)
Ticket Booking Behavior under Severe Capacity Problems: Sequential Choices during Chinese New Year
SPEAKER: Yuhan Gao

ABSTRACT. Transport systems in China face extreme capacity shortages during the Spring Festival travel season. This study therefore explores traveler’s decision making behavior when booking tickets during this season. With the purpose to establish a sequential choice model that reflects people’s decision making behavior, a mixed RP/SP survey is conduct that reflects the ticketing policy in China. Loop questions are programmed inside the questionnaire to investigate changes of the behavior after people experienced failure in ticket booking. Our sample consists of 452 respondents from different age levels and occupations. The ticket booking process is assumed as a chain of sequential choice. A sequential discrete choice model is built based on the survey data. This study models the first two levels of individual choice sequence. In the second level of the sequential choice model, a generalized choice set that is able to simplify the complex joint choice set is introduced. Two simulations are conducted based on the estimation results of the sequential choice model. The result of first simulation suggests that under fixed total railway capacity, the over-expansion of HSR has a negative impact on low-income people, and this negative impact may decrease the total social utility. The second simulation shows the possibility of creating more capacity for the low-income group by guiding others to book their less preferred tickets. We suggest the methodology introduced in this study can support the development of socially optimal ticketing policies under severe capacity shortages.

15:30-16:00Coffee Break
16:00-18:30 Session 12B: WORKSHOP Time Use and Travel
Chandra Bhat (The University of Texas at Austin, United States)
Location: MCC Theater
16:00-18:30 Session 12C: WORKSHOP Transport for Healthy, Happy, and Holistic Living
Patricia Mokhtarian (Georgia Institute of Technology, United States)
Ram Pendyala (Arizona State University, United States)
Location: Corwin East
16:00-18:30 Session 12D: WORKSHOP Automation and Self-Driving
Yoram Shiftan (Technion, Israel)
Amanda Stathopoulos (Northwestern University, United States)
Location: UCEN SB Harbor
16:00-18:30 Session 12G: WORKSHOP Big Data and Travel
Constantinos Antoniou (Technical University of Munich, Germany)
Location: UCEN Flying A