ABSTRACT. The adoption of technology-enabled transportation services is transforming transportation demand and supply by providing a unique opportunity for the introduction and extensive deployment of a wide range of new shared mobility services. Among various types of new shared mobility services, the popularity and availability of ridehailing services (also known as ridesourcing, on-demand ride services or transportation network companies, or TNCs), such as Uber and Lyft in the U.S. market, have grown rapidly. For example, a recent study of ridehailing services in the City of San Francisco showed that the trips made with these services (about 170,000 trips per day) accounts for more than 15% of all trips inside the city of San Francisco on a typical weekday for about 20% of total vehicle miles traveled (VMT) inside the city of San Francisco, and about 6.5% of total VMT including both intra- and inter-city trips (SFCTA 2017). The rapid growth in the adoption and frequency of use of the ridehailing poses significant challenges for both transportation planners and policy makers. In this paper, we plan to address some of these challenges, in particular through improving our understanding of the factors that affect the frequency of use of ridehailing services. The behavioral studies that have focused on ridehailing usually address one of these two main research questions: (1) What factors affect the adoption and/or frequency of use of ridehailing services; and (2) How the adoption of ridehailing affects individuals’ travel patterns and travel behavior, including vehicle ownership, mode choice, and vehicle miles traveled (e.g., Rayle et al. 2014; Alemi et al. 2017; Alemi et al. Under review; PEW research center 2016; Clewlow and Mishra 2017; Shared-Use Mobility Center 2016). Most studies in this area, to date, are based on the analysis of descriptive statistics and self-reported behavioral changes. For example, a recent study by the Pew Research Center (2016) showed that out of the 15% of respondents in their sample who reported that they have used on-demand ride services (N=4,787), only 3% and 12% reported that they have used on-demand ride services on a daily and weekly basis, respectively. This research also confirmed that younger adults tend to use Uber and Lyft services more frequently. In another study, Feigon et al. (2016) showed that the most frequent users of on-demand ride services live in middle-income households (annual incomes of $50 to 75K). This paper builds on our existing research effort investigating the factors that affect the use of ridehailing services. We use data from the 2015 California Millennial Dataset, a rich source of information on attitudes, lifestyles, travel patterns and the characteristics of the built environmental in the place of residence for members of the millennial generation (i.e., young adults, 18-34 by 2015) and Generation X (i.e., middle-age adults, 35-50 by 2015). The dataset was collected with an online survey, which was designed as part of the project and administered to a sample of 2400 California Residents, recruited through an online opinion panel. We used a quota sampling approach to recruit respondents from each of the six major regions of California and three dominant neighborhood types (urban, suburban and rural), while controlling for sociodemographic targets including household income, gender, race and ethnicity, and presence of children in the household. The data collection is part of a longitudinal study of the emerging transportation trends in California, designed with a rotating panel structure, with an additional wave of data collection planned in spring 2018. For additional information on the survey content and data collection, see Circella et al. (2016). In previous related stages of this research, we estimated a Zero-inflated Ordered Probit model and an Ordered Probit model with Sample Selection to simultaneously model the adoption and frequency of use of Uber and Lyft (Alemi et al. forthcoming). Both models were able to correct the self-selection bias associated with the common unobserved factors that affect both the selection (i.e., adoption of ridehailing) and frequency of use the services. Our results show that the factors affecting the adoption of ridehailing services can be different from those affecting the frequency of use of these services. For example, we found that sociodemographic variables cannot explain much of the variation in the Uber/Lyft frequency models, while these variables can explain a large proportion of variation in the adoption model. Among different built environment variables, we find that land use mix and activity density contribute to respectively decreasing and increasing the frequency of use of ride-hailing services. The results of both models also confirm that individuals who live in zero-vehicle households and those with higher shares of long-distance leisure trips made by plane are more likely to be among the frequent users of ridehailing services. Interestingly, both models reveal that there is a competition between carsharing and ridehailing services: users of carsharing services are more likely also to adopt Uber and Lyft, but they tend to use them less frequently. In this paper we expand our analyses of the frequency of use of ridehailing services and investigate the heterogeneity behind the decisions regarding the adoption and frequency of use of on-demand ride services through the estimation of a Latent-Class Zero-inflated Ordered Probit Choice Model and Latent-Class Sample-Selection Model. To our knowledge, this is the first paper that integrates latent-class models into a zero-inflated and sample-selection modeling framework. The application of this model will help us identify the factors affecting the adoption and frequency of use of ride-hailing services while controlling for variation in individuals’ lifestyles and taste heterogeneity. We expect individual’s lifestyles to impact an individual’s travel decisions. In addition, the development of classifications by lifestyle orientation could facilitate the forecasting of the adoption and frequency of use of on-demand ride services. We are currently testing different model specifications to simultaneously estimate the class-membership model based on factors affecting individual’s lifestyle and the choice models including both adoption and frequency of use of ride-hailing services. We expect to have final results ready to be presented by the time of the conference.

REFERENCES Alemi, Farzad, Giovanni Circella, Susan Handy, and Patricia L. Mokhtarian. Under review, 2017. “What Influences Travelers to Use Uber? Exploring the Factors Affecting the Adoption of On-Demand Ride Services in California”. Submitted to Journal of Travel behavior and Society, and presented at the Transportation Research Board 96th Annual Meeting, Washington DC, January 2017, Paper No. 17-05630. Alemi, Farzad, Giovanni Circella, and Susan Handy. Under review. “Exploring the Latent Constructs behind the Use of On-Demand Ride Services in California”. Journal of Choice Modeling, Special Issue of International Choice Modeling Conference (ICMC 2017). Alemi, Farzad, Giovanni Circella, and Susan Handy (forthcoming) “On-demand Ride Services in California: Investigating the Factors Affecting the Frequency of Use of Uber/Lyft”. Accepted for presentation at the Transportation Research Board 97th Annual Meeting, Washington DC, January 2018. Circella, Giovanni, Lew Fulton, Farzad Alemi, Rosaria M. Berliner, Kate Tiedeman, Patricia L. Mokhtarian, and Susan Handy. 2016. “What Affects Millennials’ Mobility? PART I: Investigating the Environmental Concerns, Lifestyles, Mobility-Related Attitudes and Adoption of Technology of Young Adults in California.” Project Report, National Center for Sustainable Transportation. University of California, Davis, May 2016; available at http://ncst.ucdavis.edu/wp-content/uploads/2014/08/05-26-2016-NCST_Report_Millennials_Part_I_2016_May_26_FINAL1.pdf (last accessed on Nov 1, 2017). Clewlow, Regina R., and Gouri Shankar Mishra. 2017. “Disruptive Transportation: The Adoption, Utilization, and Impacts of Ride-Hailing in the United States.” Institute of Transportation Studies, University of California, Davis, Research Report UCD-ITS-RR-17-07. Feigon, Sharon, Colin Murphy, and Shared-Use Mobility Center. "Shared Mobility and the Transformation of Public Transit." TCRP Report 188 (2016). Hallock, Lindsey, and Jeff Inglis. 2015. “The Innovative Transportation Index: The Cities Where New Technologies and Tools Can Reduce Your Need to Own a Car.” http://www.uspirg.org/sites/pirg/files/reports/Innovative_Transportation_Index_USPIRG.pdf (last accessed on July 5, 2017). Pew Research Center. 2016. “Shared, Collaborative and On Demand: The New Digital Economy”. May 2016. http://www.pewinternet.org/files/2016/05/PI_2016.05.19_Sharing-Economy_FINAL.pdf. (last accessed on Nov 1, 2017). Pew Research Center. 2016. “Shared, Collaborative and On Demand: The New Digital Economy”. May 2016. http://www.pewinternet.org/files/2016/05/PI_2016.05.19_Sharing-Economy_FINAL.pdf. (last accessed on Nov 1, 2017). Rayle, Lisa, Susan Shaheen, Nelson Chan, Danielle Dai, and Robert Cervero. 2014. “App-Based, On-Demand Ride Services: Comparing Taxi and Ridesourcing Trips and User Characteristics in San Francisco.” Working Paper. University of California Transportation Center (UCTC). https://www.its.dot.gov/itspac/dec2014/ridesourcingwhitepaper_nov2014.pdf (last accessed on Nov 1, 2017). San Franciscio County Transportation Authority (SFCTA). 2017. “ TNCs Today: A Profile of San Francisco Transportation Network Company Activity”. Draft Report. June 2017. http://www.sfcta.org/sites/default/files/content/Planning/TNCs/TNCs_Today_061317.pdf (last accessed Nov 1, 2017). Shaheen, Susan, Adam Cohen, and Ismail Zohdy. 2016. “Shared Mobility: Current Practices and Guiding Principles.” Report No. FHWA-HOP-16-022 (April 2016). https://ops.fhwa.dot.gov/publications/fhwahop16022/fhwahop16022.pdf (last accessed Nov 1, 2017). Shared-Use Mobility Center. 2016. “Shared Mobility and the Transformation of Public Transit.” http://www.apta.com/resources/reportsandpublications/Documents/APTA-Shared-Mobility.pdf (last accessed on Nov 1, 2017)

ABSTRACT. Mobility-on-Demand (MoD) systems are generally designed and analyzed for a fixed and exogenous demand, but such frameworks fail to answer questions about the impact of these services on the urban transportation system, such as the effect of induced demand and the implications for transit ridership. In this study, we propose a unified framework to design, optimize and analyze MoD operations within a multimodal transportation system where the demand for a travel mode is a function of its level of service. An application of Bayesian optimization (BO) to derive the optimal supply-side MoD parameters (e.g., fleet size and fare) is also illustrated. The proposed framework is calibrated using the taxi demand data in Manhattan, New York. Travel demand is served by public transit and MoD services of varying passenger capacities (1, 4 and 10), and passengers are predicted to choose travel modes according to a mode choice model. This choice model is estimated using stated preference data collected in New York City. The convergence of the multimodal supply-demand system and the superiority of the BO-based optimization method over earlier approaches are established through numerical experiments. We finally consider a policy intervention where the government imposes a tax on the ride-hailing service and illustrate how the proposed framework can quantify the pros and cons of such policies for different stakeholders. 

ABSTRACT. See attachement

ABSTRACT. The significant share of private car use in the Americans’ trips (87% of daily trips and 91% of commute trips, according to the United States Department of Transportation (1)) highlights the key role of vehicle ownership decisions in shaping travel behavior. To better understand individuals’ vehicle decisions, one should also investigate the impacts of the ever-growing smartphone-based mobility-on-demand services and the non-vehicle transportation alternatives (i.e., bike, walk, and transit), in addition to the individuals’ characteristics and the vehicles’ features. This study addresses this need by focusing on the vehicle transaction decisions (i.e., longitudinal effects of historic vehicle decisions), among other dimensions involved in vehicle decision-making process (e.g., ownership, fleet size, body type, and vehicle usage). We contribute to the relevant literature by integrating a discrete choice-based model of vehicle transaction decisions with a latent variable model of individuals’ preferences for mobility-on-demand services and their “green lifestyle”. The latter factor reflects individuals’ attitudes towards non-vehicle transportation alternatives.

ABSTRACT. Quality of life is an important dimension of individual well-being that needs better understanding for making informed policy decisions. Within this context, quality of life is especially critical with changing demographics and lifestyles, as these changes demand greater improvements in services – especially transportation mobility. To this end, transportation mobility is critical and essential to the maintenance of life satisfaction and subjective well-being. Exploring this relationship between activity participation and subjective well-being is of immense importance, especially in the context of activity-based travel analysis. Activity participation is heavily based on life cycle stages of individuals and the role they play in households. This study sheds light on activity time-use and task allocation in different life cycle stages and the resulting travel demand along with assessing and monitoring individuals’ episodic feelings (or emotional well-being) that contribute to their overall quality of life. The paper contributes to existing studies that investigate aspects of traveler attitudes and perceptions, activity time-use and modal choices on their subjective well-being (Devos et al., 2017, Enam 2017, Ravulaparthy et al., 2015). Subjective well-being (SWB) is a major topic of interest in research with emergence of new fields such as hedonic psychology, positive psychology and happiness economics (Kahneman and Krueger 2006; Kahneman et al., 2004; Frey and Stutzer 2005). Travel behavior research is starting to explore the importance of these aspects with a focus on understanding the links between activity engagement, time-use and mode choice as related to global life satisfaction and episodic well-being. However, we note that most of the current studies assume a causal structure in studying SWB and activity engagement choices. In first line of research, studies have determined individuals’ activity and travel patterns by utilizing well-being and quality of life indicators as explanatory variables within the random utility-based models of activity travel patterns (Abou-Zeid and Ben-Akiva 2012; Polydoropoulou et al., 2010). In second line of work, individuals’ SWB is viewed as a portion of derived utility from activity engagement and time-use patterns – that focus on incorporating measures of activity and travel characteristics to predict individuals’ well-being (Ravulaparthy et al., 2013, Archer et al., 2013; Goulias et al., 2013; Ettema et al., 2010). To this end, the role of time is critical in understanding emotional well-being and quality of life for both researchers and policy makers. Essentially, activity engagement and travel entails use of time, which is a truly constrained resource in which 24 hours should be distributed across these episodes, thereby also impacting an individual’s daily satisfaction. Thus, in this study we aim to jointly explore the underlying correlation structure of well-being and activity time-use patterns, thereby addressing a major gap in the literature on imposition of causality in studying these relationships. As noted by Ettema et al., (2010), unraveling this joint relationship between activity engagement and individuals’ well-being poses a major issue in-terms of bias on findings that could be misleading from a policy standpoint. Furthermore, we also analyze the temporal nature of this underlying correlation by conducting an episodic-level dynamic analysis using the sample from 2010-2013 American Time Use Survey (ATUS) well-being module. The data records reported levels of well-being for an episode an individual participates in as measured using indicators related to – happiness, feeling of tiredness, stress, sad, pain and meaningful during activity engagement. For instance, questions are framed as “taking all things together, how happy did you feel during this time of the activity? Please use a scale from 0 to 6, where 0 means not at all happy and 6 means very happy”. In addition, the data also collects information about activity and travel characteristics along with a broad array of sociodemographic information. We also report in Table 1 the sample characteristics of 2010 ATUS data by activity type classified into in-home and out of home activities. To unravel the joint relationship between activity time-use and episodic feelings (or emotional well-being), we approach this in a two-stage process. First, we construct a composite measure of degree of episodic well-being that individual experiences from activity engagement using latent class cluster analysis (LCCA) framework. In the second stage, we jointly analyze and examine individuals’ activity time-use (or duration) and derived degree of episodic happiness using a bivariate ordered probit model structure. This modeling framework also enables us to investigate the underlying correlation structure along with understanding the satiation levels (or thresholds) in activity engagement choices that act as turning points for activity scheduling from reported indicators to episodic feelings. Overall, this two-stage process also further helps understand the dynamics of this underlying correlation structure between activity time-use and episodic happiness. This is achieved by estimating a series of econometric models using the ATUS data from year 2010 to 2013. We believe analyzing this correlation structure dynamically is significant in further understanding the temporal stability of time-use and episodic well-being (Levinson and Wu 2005). For instance, Ravulaparthy et al., (2015) report a strong negative correlation between activity duration and episodic happiness from elderly activity engagement choices based on a cross-sectional study. Furthermore, two separate models are estimated for in-home and out-of-home activities to specifically investigate the differences in correlation structures between time-use and episodic happiness. In this process, we control for a broad array of covariates related to individual and household characteristics, activity and travel attributes and social context among other factors. Additionally, to further generalize results we also control for history-dependency by incorporating activity characteristics before the sampled activity in measuring episodic wellness. We also incorporate individual lifecycle stages as related to household composition, age and marital status into the modeling exercise to study the differences in their time-use patterns and reported levels of well-being. Finally, conclusions are drawn from the model estimation results along with major findings for policy implications and future methodological directions.

ABSTRACT. Activity-based models have been developed in recent years as a response to the assumptions and simplifications of the of the mobility demand in classic aggregated models. With the increasing availability of disaggregated mobility information, large-scale activity-based demand can be generated with high accuracy. In these models, a virtual population represents people from an area of interest, and the objective is to estimate realistic sequences of activities for each of these persons. Hence, the complexity of this task depends on the period of time in which activities are scheduled. Moreover, the complexity grows exponentially with the length of the period of time. Recently, multiday datasets have become more accessible, and some researchers have turned their attention to multiday activity-based models. In this work, a methodology to generate an activity-based weekly mobility demand is proposed. The approach is based on a mental map structure where activities are categorized as fixed or flexible, and three information structures are used to schedule activities: (i) a fixed activity skeleton, and (ii) a flexible activity agenda and (iii) a set of evoked places. To evaluate the model, this proposed methodology is applied to a synthetic population from Singapore. First, fixed activity patterns are extracted from the national household interview travel survey of Singapore (HITS). These results are used to detect primary activities of frequent public transport users from public transport smart card data (CEPAS), and subsequently to recognize weekly patterns of fixed activities. Then, random forest classifiers are trained to assign weekly fixed activity skeletons to agents of the Singapore scenario. Subsequently, flexible activity type and place type models are estimated using HITS and these results are employed to assign a flexible activity agenda and a set of evoked places to each agent of the synthetic population. Finally, agents are put together in a large-scale simulation platform: MATSim. Agents start the simulation with incomplete plans containing fixed activities only. During the mobility simulation, agents use a scheduling algorithm to plan flexible activities and generate trips between them. The evolutionary algorithm of MATSim is then used to optimize weekly activity sequences. Results show accurate spatio-temporal distributions of activity performing by day of the week, accurate travel time distributions, and accurate public transport rideship distributions. The process is also tractable computationally; in 2 days and 3 hours, 100 iterations were executed of a 10% scenario of Singapore.

ABSTRACT. Do people decide how to travel based on expectations or experiences of time use or feelings during a trip? This research investigates associations between two major components of the Positive Utility of Travel (PUT) concept—travel-based multitasking and subjective well-being (SWB)—and mode choice. A novel revealed preference survey of around 550 Portland, Oregon, commuters—in which PUT attributes were measured for chosen modes as well as for alternative non-chosen modes—was used to test associations of multitasking and SWB with mode choice using an integrated choice and latent variable (ICLV) model. The SWB was a 9-item adapted version of the Satisfaction with Travel Scale (STS). Surprisingly, many instances of activity participation were not or even negatively associated with mode choice. Alternatively, the STS measure of SWB from the travel experience was positively and significantly associated with commute mode choice. Results suggest that travelers may consider experiences and expectations regarding their well-being and happiness when choosing a commute mode, but also that travel-based multitasking may be more about “killing time” than making productive use of it. These findings have implications for improving transport-related SWB and understanding future mobility.

ABSTRACT. This study identifies the interrelationship between individual's mobility and perceived Quality of Life (QoL).We consider two latent variables for mobility regarding its subjective and objective dimensions. The first one is perceived level of mobility (subjective) which is obtained by survey data. The second is the level of mobility observed from actual bus use records (objective mobility) obtained from smart card data. This study targeted Shizuoka city, a mid-size city in Japan and the analysis was conducted by dividing the data samples into three groups. i.e non-elderly (less than 65 year), young elderly (65-74 years), and old elderly (over 75 years). Finally we identify whether subjective mobility and objective mobility has more influence on mobility satisfaction as well as QOL trough a hybrid model estimation and discuss how through the activation of bus use the overall QOL and social welfare can be increased. In addition, the result would help to introduce the new transit system, in order to provide more flexible and more appropriate services for an aging society.

ABSTRACT. Bicycling is an environmentally friendly, healthy, and affordable mode of transportation that is viable for short distance trips. Urban planners, public health advocates, and others are therefore looking for strategies to promote more bicycling, including improvements to the built environment that make bicycling more attractive. This study presents an analysis of how key built environment characteristics relate to bicycling frequency based on a large sample from the 2012 California Household Travel Survey and detailed built environment data. The built environment characteristics we explore include residential and intersection density at anchor locations (home, work, school), green space, job access, land use mix, and bicycle infrastructure availability. Analyses are conducted separately for three distinct demographic groups: school-age children, employed adults, and adults who are not employed. The key conclusion from this work is that the relationship between bicycling and some built environment characteristics varies between types of people – most dramatically between adults and children. To develop targeted policies with scarce resources, local policymakers need specific guidance as to which investments and policy changes will be most effective for creating “bikeable” neighborhoods. Our work indicates that the answer depends – at least in part – on who these bikeable neighborhoods are meant to serve.

ABSTRACT. ***Extended Abstract is attached***

Abstract: Shared economy is a social-economic phenomenon that prioritizes utilization over ownership. Shared mobility is defined as the "shared use of a vehicle," allowing users to get access to a private transport mode on an "as-needed basis". Bike sharing, which is based on a short-term rental of a bicycle, has found in the pertinent literature to have benefits related to the decrease of private cars ownership. Given their benefits the identification of the exogenous factors that affect the travel behaviour with regards to bike-sharing systems is needed, in order to expand such concepts to new cities and to increase their performance and reliability. Several studies have focused on the identification of these factors on a city level or a station level. However, these studies are restrictive on both a spatial (only focusing on individual cities or stations, or considering cities as a whole) and temporal level (varying from 1 week to a few months year) limiting in essence, both their predictive and explanatory power. Nowadays, the existence of open data platforms has enabled the possibility of collecting large datasets both in the transport and geography. Utilizing open data, this study has created an automated data collection and analysis framework to identify the relationships between arrivals and departure rates from bike sharing transportation systems with exogenous factors in multiple cities on the scale of influence zones. An automated procedure has been developed that consists of (see Figure~\ref{fig:methodology}): 1) data collection from open-source data sources, 2) data analysis and processing and 3) model building and selection using three methods: multiple linear regression with a stepwise regression selection technique, generalized linear models (GLM) with a lasso technique and gradient boosting machine (GBM).

ABSTRACT. Introduction Encouraging the use of active travel modes such as walking and cycling is vital for ensuring a sustainable urban development. In many European metropolitan areas, cycling is becoming increasingly popular within the recent decades. The Stockholm travel survey shows that the cycling share in the Greater Stockholm Area has reached 16% in 2015 (Petersen et al., 2015; Liu et al., 2017). On the other hand, large-scale transport models, which serve as the main tools for policy evaluation and cost-benefit analysis, are often designed for modelling motorised travel modes such as private car and public transport. In Sweden, the national transport model Sampers (Muriel and Algers, 2002) uses a simplified specification for cycling as a mode alternative, where the travel distance calculated from the routing on a car network is used as a level-of-service variable in the utility function of cycling. In that sense, a change in bicycle infrastructure will not result in changes in the utility function of cycling (because changes in bicycle infrastructure are not reflected in changes in the car network) and therefore will not affect model predicted cycling demand. Another issue is that the zoning system used in most large-scale transport models is too coarse to model cycling demand, as cycling trips are often short-distance trips and therefore many have their original-destination pairs within a given zone. Within-zone cycling trips are difficult to model properly as the detailed level-of-service variables for those trips are unknown. This study presents a tour-based transport model that addresses the above-mentioned issues in modelling cycling demand. First, it uses a detailed bicycle network containing more than 200,000 links, covering the whole Greater Stockholm Area. A generalised cost calculated by a route choice model from Halldórsdóttir (2015) is included as one attribute in the utility function of cycling. It considers detailed types of bike lanes and they are categorised into motorised road without any bicycle facilities; separated bike lane; painted bike lane. It also takes into account link cumulative elevation gain and built environment. Other variables such as bicycle parking and population information are also included in the utility function of cycling. Therefore, the model predicted cycling demand is responsive to a change in bicycle infrastructure, which makes the model suitable for evaluating such changes. Second, the model utilizes a refined zoning system with totally 5808 zones each of the size of 250 m×250m covering the entire Greater Stockholm Area, whereas there are 717 zones in the Sampers’s zoning system for the same area. Third, the model is calibrated using the newest Stockholm travel survey 2015, and therefore the model can represent the travel behaviour that is up-to-date. However, it is worth noting that the travel survey only includes observations in September 2015. Therefore, the seasonal variation cannot be captured in this data. Apart from the improvements mentioned above, the model also considers cycling as an access travel mode for a trip by public transport. Therefore, the model treats cycling and public transport as complementary and the model is capable of evaluating the impact of an improvement in bicycle infrastructure on cycling as well as cycling to public transport stations. Model formulation (please see the attached file as the full abstract. since symbols are not allowed here, I remove all the description of model)

Preliminary results The travel survey 2015 was used to calibrate the model. Given that the observed destination choice of each trip is only available at the traffic zone level that is used in Sampers, the Sampers zoning system is used in the estimation stage. Currently only the model for morning commute trips are calibrated. The descriptive statistics show that around 16% of the morning commute trips (6:00-9:00) observed in the data are cycling trips. Among trips by public transport (mode share 50.4%), 6% of these trips have cycling as access mode. The initial estimation results indicate that the model can capture the positive effect of bicycle parking on the propensity of cycling although the effect is weak (t-value 1.31). The coefficient of generalised cost for cycling is higher for men, indicating that men are willing to cycle longer distances than women. The model also successfully captures the effects of different time components for public transport with different access travel modes. The results indicate a substitution rate 1:2 between in-vehicle time and transfer walk time/egress walk time. In other words, one-minute in transfer walk time/egress walk time is equivalent to 2 minutes in in-vehicle time. Similarly, the substitution rate between in-vehicle time and initial waiting time/transfer waiting time is 1:1.5. The substitution rate between in-vehicle time and access walk time is 1:1.5 but 1:1 for car access time. For cycling as access mode, the generalised cost is used instead of travel time, and the substitution rate between in-vehicle time and cycling generalised cost for the access part is 1:1. Elasticities and marginal effects are derived to compare different variable effects. The preliminary results show that the elasticity of generalised cost on cycling mode share is -1.45, indicating that 1% increase in the generalised costs for all OD pairs corresponds to 1.45% (0.22 percentage points) decrease in the cycling share. On the other hand, the effect of bicycle parking is marginal with the elasticity 0.034% (0.006 percentage points). The calibrated model is then used to generate cycling demand on the much more refined zoning system (5808 zones). Generated cycling demand is assigned back to the cycling network using an All-or-Nothing assignment. A synthetic population is also generated by sampling observed individuals from the national travel survey 2011-2013 and 2015. The model predicted cycling flow will be compared with observed count data available at several critical links in the network. Further study will focus on conducting scenario analysis and evaluating possible bicycle infrastructure changes. Models for other trip purposes will also be developed. Possible model development towards cost-benefit analysis of bicycle infrastructure will be discussed.

ABSTRACT. Many communities of Alaska rely on less conventional modes of transportation to fulfill their basic mobility needs. Here, we present an a analysis of place connectivity, statewide trauma data, and the results from the 2016 Pacific Northwest Transportation Survey to better understand issues related to travel on “mixed-use” facilities. Specifically, this research documents the use of non-motorized and off-highway vehicle modes of travel in rural and small-urban communities of Alaska.

ABSTRACT. Text mining of essays written by Gen Z about self-driving cars

ABSTRACT. As the increased convenience of driving and enabling various activities while driving results in reducing the value of travel time, the increased utility of private vehicles affects transportation behavior, such as a modal shift to an (autonomous) owner-driven car, an increase in long-distance traffic, and change in urban land-use by the improved accessibility in the future. Thus, this research aims to study the change in future transportation behavior by implementing stated preference survey for an autonomous vehicle and applying the modal choice model (estimated based on the survey) to activity-based model. Despite some limitations, this study will serve as useful reference for transportation planning regarding AVs and provides a quantitative method for evaluating autonomous cars' impact on transportation behavior.

ABSTRACT. A prominent focal point for a bundle of emerging new technologies is the autonomous vehicle (AV, a.k.a. self-driving/automated car/vehicle, etc.). Many people agree that the introduction of AVs will bring massive changes in nearly everything related to transportation, and hence the changes in transportation systems or human behavior under full automation are of keen interest to researchers, planners, and policymakers alike. Because fully automated AVs have not been realized yet, those topics are of dominant interest. However, in the behavioral research realm, the ultimate question about AVs would be “How will people change their travel-related behavior, if personal vehicles were fully automated and replaced all conventional vehicles?” Although any such reactions offered now must be considered rather volatile, it is not too soon to begin assessing responses to an all-AV future. To address this question (i.e. to envision future travel behavior in the era of fully automated self-driving cars), we are conducting a survey throughout the state of Georgia in the US. This study has a number of merits: Geographical coverage, representativeness of the sample, focus on behavioral change, and rich information (e.g. attitudes, use of technologies, current behavior, socio-demographics). We expect to model AV-use propensity models. Thus, the focal models of this study will be identifying the personal characteristics associated with higher versus lower propensities to use AVs when those are the only personal motorized vehicles available.

ABSTRACT. Please see the Attachment!

ABSTRACT. With the increasing availability of volunteered geographic information (VGI), the study of human travel behaviors in a small scale becomes plausible. However, the progressively increased coverage inevitablely sacrifices the quality of the volunteered geographic information. Therefore, approaches to quantify the uncertainty, or accuracy, of VGI emerges as a research challenge. Conventional research concentrates on investigating the uncertainty of VGI from perspectives such as the place positions, attributes and topological relations. There is few work on assessing the uncertainty of VGI from a semantic perspective, which plays a vital role in understanding human travel behaviors. To address this gap, our work proposes an approach to take into account the semantic uncertainties of places to assess the quality of VGI. The semantics are specifically extracted from the feature type of geographic information (e.g., place types such as Restaurant, Hotel and Police Office) as it is semantically rich in terms of characterizing the functionality, popularity, and human perception of places. In order to assess the semantic uncertainty of feature types, we design a set of statistics, named as semantic signatures, to quantitatively characterize feature types. Semantic signatures are furthure grouped into three categories: spatial, temporal and thematic, giving the respective perspectives they are extracted from. Finally, we illustrate some initial experiments showing the feasibility of applying semantic uncertainty analysis to understand human travel behaviors.

ABSTRACT. Cycling has been recognized as one of important travel modes in cities due to its societal and environmental benefits of alleviating traffic congestion, improving air quality, decreasing fuel consumption, increasing public health, and providing an affordable mode of transport. Although efforts and investments to increase bike use, the low growth rate in cycling has been a subject for investigation in the last decade. Researchers, policy makers and transportation agencies have invested extensive resources to identify factors that influence cycling as cycling mode share takes up only 1.3% of the commuters in Canada and 1% of all trips in the U.S. in 2001 (e.g., National Household Traffic Survey of America 2013). In order to understand the underlying reasons for such low cycling levels and to see a substantial increase in mode shift, an understanding of cyclist needs and perceptions is required.

Discrete route choice modeling has been paid attention to in recent years to investigate travel behavior of cyclists. Route-based models such as C-Logit and Path-Size Logit (PSL) were widely used for bicycle route choice analysis (Ben-Akiva & Bierlaire 1999) to evaluate the effects of attributes related to the whole trip traces. These route choice models only addressed the similarities between the considered (sampled) set of routes. In order to consider sampled and non-sampled alternatives (routes), Expanded Path Size Logit (EPSL) model has been suggested that made use of a sampling approach to compare the chosen path to a set of alternative paths available to the cyclist. All above efforts lead to better understand traveler behavior in selecting choices from generated alternative choice sets. However, in the presence of intensive applications of discrete route modeling, limited studies have applied machine learning techniques to identify factors explaining travel decisions and to uncover the underlying decision-rules. Machine learning is the practice of bringing quantitative data, analyze and visualize them in ways to bear on decision making and predicting futures by finding patterns from existing data. Machine Learning (ML) techniques use different algorithms to extract knowledge/information from large datasets. Decision tree, and random forest are two popular, strong and non-parametric ML methods, which are able to predict future responses (predicting cyclist route choice) within a black box framework. As an advantage of ML analytics, the introduced methods are able to handle complex data collected from different resources such as videos, pictures, surveys, text in efficient ways. The application of ML in different transportation domains has become popular recently (Wong et al., 2017). However, the application of ML in travel behavior research field is still limited mostly to analyzing observed movement patterns and to make short-term travel demand predictions. As addressed by transportation scholars, it is essential to use ML methods as data oriented techniques in identifying factors justifying travel decisions.

With respect to the above arguments, this paper has been motivated to analyze cyclist route choice by applying Expanded Path Size Logit (EPSL) model along with the Metropolis-Hastings (MH) sampling algorithm. The results are compared with findings achieved from ML techniques, which to best of our knowledge, has not been used together for cycling route choice analysis. Our study makes use of data from a large-scale GPS-based travel survey, as well as Toronto’s geographic information system (GIS) road network databases to model Torontonian’s bicycle route choice. The GPS bicycle trajectory data include valuable information such as trip purpose, date, time and season. Decision tree and random forest complete non-parametric analysis on data collected from GPS based travel survey to predict cyclist route choice according to given attributes. MH path-sampling algorithm is applied on a road network to generate the choice set, and a multivariate route choice framework while EPSL considers the effects of various attributes on cyclist route choices by including correlations between the sampled and non-sampled alternatives.

ABSTRACT. Efficient and robust machine learning frameworks for explaining latent behaviour are becoming more important with the increased use of very large scale multiple discrete-continuous (MDC) choice data in travel behaviour modelling. In such type of data, there are various complex inter-dependencies because of the multiple unobserved heterogeneous population behaviour types. Moreover, advanced utility based modelling techniques may not be able to capture the different facets of high-dimensional information and the estimation techniques used may not be able to manage such large data. This paper propose a novel generative modelling methodology for MDC data based on a Restricted Boltzmann Machine framework. We consider a large scale regional travel survey which contains information on 624,845 individual trips. Trip mode, purpose and distance measures are specifically considered in our case study. Our preliminary studies have demonstrated the applicability of the proposed method and could significantly improve the understanding of latent behaviour in MDC choice data.

ABSTRACT. Despite having been known for a long time (e.g., McCulloch & Pitts, 1943; Rosenblatt, 1958), and despite having been occasionally used for the analysis of travel behavior since more than a decade ago (Hensher & Ton, 2000; Mohammadian & Miller, 2002), Artificial Neural Networks (ANNs) have only lately become – by a distance – the most prominent and promising Artificial Intelligence (AI) model for the analysis of travel behavior in the context of large, emerging data sources (e.g. Karlaftis & Vlahogianni, 2011; Chen et al., 2016; van Cranenburgh & Alwosheel, 2017). This sharp increase in the popularity of ANNs as a tool for travel behavior analysis has resulted from a range of improvements in ANNs’ capabilities, increases in computational power, and the rapidly increasing size and diversity of data which are at the disposal of choice modelers. This paper aims to help pave the way for further and effective deployment of ANNs for travel behavior analysis. It does so by highlighting and articulating an easily overlooked aspect of the ANN-methodology, which is of crucial importance for its successful use in a travel choice modeling context. More specifically, we study the relation between i) the assumed characteristics of the Data Generating Process (DGP; in this case the assumed model of travel choice behavior or decision rule), and ii) the size of the data that is required for meaningful, reliable travel choice analysis using ANNs. The core idea behind this relation is intuitive: if the DGP is relatively complex – e.g. highly non-linear – then a given ANN needs more data to be able to generate a reliable representation of the DGP, leading to accurate predictions. Despite or perhaps because of this straightforward intuition, choice modelers employing ANNs so far seem to have ignored important results from the AI literature which rigorously define this relation between the complexity of the DGP and resulting data-requirements, in the context of empirical analysis using ANNs. Such concepts as the Universal Approximation Theorem (Cybenko, 1989; Hornik et al., 1989), the notion of Probably Approximately Correct (Valiant, 1984) and the so-called V-C dimension (Vapnik & Chervonenkis, 1971) have helped AI-researchers in various fields of application determine the required size of their dataset as a function of the assumed characteristics of the DGP. This paper aims to introduce these theoretical concepts and notions from the AI-literature to the travel behavior research community, and moreover to translate them in a way that they can be readily used by travel choice modelers. By doing so, we aim to help travel behavior researchers who wish to use ANNs for discrete choice analysis, in the process of selecting data sets or collecting data. To focus our attention, we limit our discussion to the context of two particular travel choice models as DGPs: one is the well-known linear-additive MNL model based on utility maximization premises, which is the workhorse of discrete choice analysis and in many ways the least complex choice model available (Ben-Akiva & Lerman, 1985; Train, 2009). The other is the Random Regret Minimization model (in MNL form), which is one the most used behavioral alternatives to the canonical linear in parameters utility based MNL model (Chorus et al., 2008; van Cranenburgh et al., 2015). The regret function embedded in most RRM models is highly non-linear and includes attributes of all alternatives in the choice set. As such it is a considerably more ‘complex’ choice model than its utility based counterpart, something which for example shows in considerably higher runtimes (Guevara et al., 2016). As such, the comparison between these two models (i.e., DGPs) serves well to highlight how, in the context of discrete choice analysis based on ANNs, data-requirements follow from the characteristics – i.e., level of complexity – of the DGP. Our study thus consists of two parts. In Part 1 we will introduce all relevant concepts, notions and theorems that have been developed in the ANN literature to determine minimum sample sizes as a function of model complexity. We will make sure to present these ideas in a notation and framework that connects directly with conventional modeling practice in the travel behavior research community. In Part 2 we will use these ideas in a concrete example, for illustration purposes and to establish face validity. More specifically, we will show how Random Utility and Random Regret DGPs differ in terms of their data requirements, in the context of model estimation with ANNs. We conclude our study with the derivation and discussion of implications for researchers and practitioners in the field of travel behavior analysis. To get a flavor of the analyses which we performed in Part 2, we here present some first results. Our ‘empirical’ setting is a simple travel mode choice between three alternatives (car, bus, train) based on two attributes (travel time, travel cost). We generate two synthetic datasets containing mode choices: one dataset uses a Random Utility DGP (in MNL-form) and the other one uses a Random Regret DGP (also in MNL-form). Subsequently, we derive – using the introduced concepts from the ANN-literature – the theoretically expected minimum (training) sample size needed to achieve a reliable representation of the DGP by an appropriately specified ANN. We do this for the RUM and RRM DGPs, and show how – in line with expectations – the theoretically required minimum (training) sample size is larger for the latter. Finally, we verify this theoretical result by training ANNs, for each DGP, using increasingly large subsets of the synthetic data. As Figure 1 (RUM) and Figure 2 (RRM) show, the out of sample predictive ability – measured in terms of out of sample LogLikelihood – of the corresponding ANNs is found to increase sharply up to the theoretically identified minimum (training) sample size, after which marginal increments in model fit become notably smaller. This suggests that the theoretically established minimum sample size provides a reasonable indication of practical (training) sample size requirements for the two different DGPs.

ABSTRACT. Many existing studies have identified residential self-selection to be associated with travel behavior. Actually, any omitted factors may cause the occurrence of self-selection. However, the existing studies have assumed that residential self-selection stems from two sources: attitudes and sociodemographic traits. As evidenced by research on the life-oriented approach, residential behavior can be associated with other life choices such as job location, school location, household expenditure, and time use. Thus, it is not unrealistic to assume the existence of self-selections associated with various life choices (i.e., multiple life-oriented self-selection effects). This study makes an initial effort to capture the multiple life-oriented self-selection effects on travel behavior, by implementing a questionnaire survey to 900 residents living across the whole Japan in 2014. The survey contains information about eight life domains, in addition to ownership of different types of vehicles, major daily travel mode, and travel behavior with respect to each of the eight domains. Furthermore, the survey investigated attitude-based self-selection with respect to all the above eight life domains and travel behavior, i.e., attitude toward life. This study empirically confirms that attitudes toward life affect both travel behavior and other life choices, where the corresponding variables of attitude toward life are further used to explain travel behavior. The findings from this study suggest that the life-oriented approach could pave a new way for exploring the issues of self-selection in the context of travel behavior research in a systematic way.

ABSTRACT. Please see attached PDF document.

ABSTRACT. Motivation Households’ longer-term decisions such as residential location and vehicle transaction have a dynamic nature, as these are inter-dependent decision processes. For instance, decision of where to live interacts with the decision of how many vehicles to own (Rashidi and Mohammadian 2011). Moreover, such longer-term decisions evolve over the life-course and interact with life-cycle events. For example, birth of a child is found to effect the choice of residential location (Strom 2010), as well as vehicle transaction (Oakil et al. 2014). Moreover, longer-term decisions have an inherent process orientation. For instance, residential location is a process of decision to move (i.e. mobility), and location choice (Habib 2009). Although considerable progress has been made in developing theories, modeling methodologies, and simulation frameworks for longer-term decisions; limited studies have addressed the evolution of multi-domain decision interactions along the life-course of the households as well as the process orientation of decisions within the empirical and computation procedures of the models. Particularly, the following two research questions demand further investigation: 1) How to accommodate life-trajectory dynamics and process orientation during modeling longer-term decisions of residential location and vehicle transaction? 2) How to advance the microsimulation of longer-term decisions by taking a life-course perspective and predict the micro-level evolution of urban regions? This study attempts to address the above-mentioned research questions by proposing the development of an agent-based life-oriented longer-term decision simulator that includes the following components: population synthesis, life-stage transition, residential location, and vehicle transaction.

Theoretical Context This study adopts a life-oriented approach to develop a longer-term household-level decision simulator. Life-oriented approach focuses on the inter-dependencies among the decisions and changes occurring at different life-domains of people (Zhang 2017, Zhang 2015). A life-oriented approach takes a life-course perspective, as temporal variation of the interactions over the life-course of households is addressed. Life-course perspective emphasizes on the effects of changes at different life-domains in shaping individuals’ or households’ behavior (Chatterjee and Scheiner 2015). The proposed life-oriented model microsimulates households’ longer-term decisions longitudinally along their whole life-time or a segment of the life-time. Individuals enter the simulator through birth or in-migration. They grow older within the system, and exit through death or out-migration. Along their life-course, changes at different life-domains occurs, such as marriage, child birth, job change, residential location change, and vehicle transaction, among others. These decisions and changes interact with each other. These decision interactions generate discrepancies between the desired and current situation of a household and induces stress. Hence, households’ decision in one domain might trigger a decision in another domain that holds the potential to minimize the stress. Such multi-way interactions also have a temporal dimension. For example, households require an adjustment period to adapt prior or after a change in life-stage, due to the limitations in time and money budget. The mechanism utilized to accommodate the interaction among the changes is through introducing lead and lag events. A lead event refers to an event on occurrence, and a lag event refers to an event in anticipation. The simulator also addresses the underlying process orientation of the decisions. For example, residential relocation is modeled as a process of mobility, and residential location. In the first stage of mobility, households decide to move or stay at a location. Households deciding to move, enter the second stage of residential location. This stage is assumed as a two-tier process of location search and choice. In the first-tier of search, households undertake a search process to generate a pool of location alternatives. Finally, in the second-tier of location choice, households move to one of the locations from the pool of alternative locations. The process of relocation follows the theory of residential stress (Rossi 1955), thus it is conceptualized to accommodate the effects of continual stress at different life-domains. Similarly, vehicle transaction is assumed as a process of: first vehicle purchase, acquisition, disposal, and trade.

Methods Innovative modeling methods are developed to address the process orientation and life-trajectory dynamics. For example, a novel fuzzy logic-based method and latent segmentation-based logit (LSL) modeling technique are adopted to develop the two-tier search and location choice model. The search process is assumed as a stress releasing mechanism. The fuzzy logic method accommodates the inter-dependencies between the stress-driven push and pull factors. The push factors include: to live in proximity to work/key activity locations, to live in desirable neighborhood/dwelling, and due to life-cycle events. The pull factors refer to: distance to work location, percentages of non-movers in the neighborhood, and distance to CBD. Since households’ choices of residential locations are strongly influenced by their affordability, this study assumes that each push factor is constrained by the following two parameters: household income and average value of the property. The output from the search model is a pool of household-specific alternative location, which feeds the location choice model in the second tier as the choice set. In addition, demographic elements in the life-stage transition component follows a heuristic modeling approach. Advanced econometric models are developed to accommodate the effects of repeated choices during the life-course of the households, as well as capture unobserved heterogeneity. For instance, vehicle transaction component adopts a LSL modeling technique. In the case of microsimulation, the proposed simulator is an agent-based model, assuming individuals and households as the agents. Urban form is represented at the most micro geographic unit of parcel, since residential location component is conceptualized to be modeled considering parcels as the spatial unit of analysis. It is a discrete time model, which moves forward by simulating agents decisions at each simulation time-step. The simulation starts with a sample of baseline population, and the relationships among the agents in the population are maintained throughout the simulation period. The system state does not hold the equilibrium assumption, rather it is always in a dynamic dis-equilibrium state. It is designed as a modular-based modeling system, which allows the application of each module and subsequent micro-models in isolation, and offers the opportunity to improve any component without affecting the whole simulation framework. All urban form elements are considered as exogenous in the current version of the software.

Data Sources This study utilizes data from a number of sources. The primary data source is a retrospective Household Mobility and Travel Survey (HMTS), administered from September 2012 to April 2013 in Halifax. The HMTS collected life-history information across the life-domains of the households, including housing history, vehicle ownership history, compositional change in household size and employment size, and employment records, among others. The secondary data sources include, Public Use Microdata File (PUMF), Census information, Nova Scotia Property Database, location of different activity points from the Desktop Mapping Technologies Inc. (DMTI), and road network and land use information from the Halifax Regional Municipality (HRM).

Model Estimation Results The model tests the interactions among multi-domain decisions and life-cycle events as lead and lag events. In the case of vehicle transaction, model results suggest that considerable heterogeneity exists in the two latent segments. For instance, birth of a child or member move in is found to trigger vehicle acquisition in one segment and deter in another segment. The effect of historical deposition is also confirmed on vehicle transaction decisions. For example, birth of a child confirms a two-year lagged effect on vehicle acquisition. Interestingly, the first time vehicle purchase behavior is found to be considerably different than vehicle acquisition decisions (addition of a vehicle to the existing vehicle fleet). For instance, addition of a job reveals significant heterogeneity across the two segments in the case of first time vehicle purchase. The model confirms that households require three years of adjustment period after getting a job to purchase their first vehicle. The same variable exhibits a higher probability for vehicle acquisition in both segments, and confirms a smaller adjustment period of one year. Similar findings are observed for the mobility, and residential location models. Another interesting finding is that the accommodation of process orientation within the models offer improvement in the empirical estimation technique. For example, the fuzzy-based two-tier location choice model improves the model fit compared to the traditional random sampling-based location choice model.

Microsimulation Results The proposed simulator is currently operationalized for Halifax, Canada, and known as the integrated Transport, Land Use, and Energy (iTLE) model. The iTLE model implements following components: population synthesis, life-stage transition, mobility, location choice, first time vehicle purchase, and vehicle transaction. The model generates baseline synthetic information for the year 2006, and runs simulation for a 15-year period from 2007 to 2021 at a yearly time-step. A full-scale validation of the model suggests that the iTLE generates reasonably satisfactory estimates of the population. This study also offers microsimulation results regarding the spatio-temporal evolution of Halifax. In terms of the demographic events, the iTLE predicts the rates for birth, death, and marriage to be 11.73, 6.72, and 5.97 per 1000 individual, respectively in 2021. The predicted rates for in-migration, out-migration, and residential movers are 40.71, 24.33, and 152.04 per 1000 households respectively in 2021. Overall, an increase of 14.08% population is predicted in 2021 compared to 2006. In the case of mobility, kernel density plots suggest that younger head households are predicted to be more frequent movers than their older counterpart. The mean duration of stay of the population with age <40, 40-54, 55-64, and 65 and above are predicted to be 3.53, 6.05, 6.60, and 7.24 years, respectively. The predicted housing pattern suggests a higher density of the households in the locations within 25km from the CBD over the simulation years of 2007-2021. The proportion of total households is predicted to increase from 68% in 2007 to 71% in 2021 in these high density neighborhoods. In terms of the household compositional configuration of the high density neighborhoods, kernel density plot suggests that a higher density of single person households are predicted in the urban core (Figure 1-a). As household composition changes through marriage and having child, the density is predicted to be more variable and skewed towards suburban neighborhoods (Figure 1-b). In the case of vehicle ownership level, higher proportion of the households with zero vehicle ownership is predicted in the Halifax urban core. The density is predicted to be more variable and distributed in the suburban areas with the increase in vehicle ownership level. The average vehicle per household member is predicted to be 0.72 in 2021 and around 75% of the households are predicted to own at least one vehicle per household member in 2021. In the case of vehicle transaction, 2-D kernel plot suggests that the first vehicle purchase is predicted to involve a higher proportion of younger head (average age 39 years) lower income (average household income $33,000) households in 2021. The vehicle ownership level composition analysis of the neighborhoods suggests that the higher income DAs in the urban core and surrounding suburban areas are predicted to have higher average vehicle ownership per household, which is as high as 2.2 in 2021.

Conclusions In summary, the contributions of this study are two-fold: 1) to develop methods for process-oriented life-course modeling; for example, fuzzy logic-based method is developed for process-oriented location modeling that accounts for households’ continual stress at different life-domains, and 2) implement a new longer-term household-level decision simulator that assists in predicting the housing pattern, neighborhood composition, vehicle ownership, and vehicle transaction pattern of an urban region. Finally, this study is a significant step forward towards adding the capacity in longer-term decision simulator to test the response of population at different life-stages under alternative land use and transport scenarios.

ABSTRACT. As individuals pass through different stages of their life their experiences can influence future behaviours. As well, those experiences may or may not play a role related to future expectations. In this study, we examined whether past experiences with different modes would explain current behaviour and anticipated behaviour.

ABSTRACT. We compare the ability of three types of positioning data, namely, CDR, MPS and GPS to provide common mobility metrics, such as activity patterns, home and work locations, and travel distances. Daily probability distributions of activity locations and durations are determined using kernel density computation with automatic bandwidth detection. These distributions are then used to calculate various mobility indicators. With the use of higher resolution passive position technologies becoming increasingly common, this research will further the discussions on the accuracy and potential of mobile data for travel behavior research.

ABSTRACT. Mobile phone telco data represents very valuable information for transport planners. Its large-scale coverage together with its spatiotemporal resolution makes it compatible with agent-based simulations for transport planning and a promise to improve travel demand models. However, such data is particularly vulnerable to breaches of privacy. Even if anonymized, there is a risk that users could be re-identified. In this work, we propose a model capable of generating individual space and time traces without looking at any individual telco trace. We define a set of aggregated histograms needed from the telco company to generate a population of individual mobility patterns using a Dynamic Bayesian Network. We present the validation results against the original telco data, showing that the model proposed is a viable tool to exploit telco data for transport planning without compromising users’ privacy.

ABSTRACT. Advances in smartphone technology have enabled researchers to collect travel data based on GPS-based smartphone apps. Researchers have investigated several methods to infer trip characteristics from data collected with these apps, such as origin/destination or mode. However, automatically predicting complete trip information from GPS traces has received less attention in the literature. This research develops a machine learning-based framework to identify complete trip information based on GPS traces as well as online data from public transit agencies, and location-based social network services (specifically, General Transit Feed Specification (GTFS) data and Foursquare API). The framework has the potential to be integrated with smartphone apps to produce all trip characteristics traditionally collected through household travel surveys. We use data from a recent, large-scale pilot smartphone travel survey from in Montreal. The collected GPS traces augmented by GTFS and Foursquare data are used to train and validate three random forest models to predict mode of transport, transit itinerary as well as trip purpose. According to cross-validation analysis, the random forest models show prediction accuracy of 86%, 81% and 71% for mode, transit itinerary and purpose of trip, respectively. The results are comparable with models previously developed in the literature. Furthermore, the validation results show that the machine learning-based framework is an effective and automated tool to support trip information extraction for large-scale GPS-based household travel surveys, which have the potential to be a reliable and efficient (in terms of costs and human resources) data collection technique.

ABSTRACT. The goal of this study is to investigate the relationships between wayfinding styles and mobility patterns and navigational preferences. Urban wayfinding behavior is defined by the strategies that people use to decide how to move from one place to another within a city (Montello 1995). It relates to the preferences, selection and application of navigation strategies, the attitude towards travelling, and ability to reach the intended destination. The research question is to what extent do wayfinding styles differ for groups of travellers and their mobility patterns and navigation preferences? The hypotheses are that more active mobility patterns correlate to more wayfinding abilities, and that with more wayfinding abilities a stronger preference occurs for taking short cuts, while the preferences for time, distance and number of turns may depend on the travel mode and urban environment. First a theoretical framework has been developed for the identification of wayfinding styles based on literature and a factor analysis derived from 23 self-reported preferences towards wayfinding and navigation. Furthermore, the results illustrate fourteen variables (relating to socio-demographic, mobility patterns and navigational preferences) that exercise significant differences among the clusters of wayfinding styles, while the built and urban environment did not yield any significant differences. The contribution of relating wayfinding behavior to revealed mobility patterns and navigational preferences could provide new insights into the decision-making process of people while travelling and improve the content of travel information.

ABSTRACT. See attached file.

ABSTRACT. Route guidance provided through mobile applications are increasing in popularity with the advancement of telecommunication technologies and, these have the potential to effectively handle demand in congested systems. The success of this strategy, however, lies in understanding what factors may influence how the user reacts to the route guidance provided by such application.  The aim of this paper is to analyze data generated by the “Metropia” application, and investigate which factors contribute to increased user compliance with the suggested routes. Data from over 31,000 trips covering peak and off-peak periods and weekdays and weekends were analyzed. The results indicate that route compliance is greater during off-peak periods and along less congested routes. The app’s incentive system (credits) positively affects the compliance with the suggested route. Past compliance with the route guidance was the strongest predictor of the extant of compliance for any trip. Identifying ways to increase initial compliance with route guidance and uncovering ways to encourage users to comply with guidance under congested conditions are critical to successfully using route guidance for transportation demand management.

ABSTRACT. Introduction In container terminals, hinterland workload forecasting is of essential importance for storage planning, daily and hourly equipment allocation and human resources management. It has been observed that one of the main factors increasing the inefficiency of a container terminal is unproductive moves (UPM), i.e. relocations of containers for any purpose beside inspection and customs. Several studies have illustrated that accurate information provisions on drayage truck arrivals can result in an important reduction of Unproductive Moves (UPMs) (Goodchild and Noronha, 2010). The number of UPMs can be applied as Key Performance Indicator (KPI) to measure terminal efficiency. Therefore, based on the prediction of the Dwell Time (DT) of a container, a daily pick-up probability can be assigned to each container depending on its arrival day. This information would permit stowage officers to stack the containers in such a manner that the containers with the higher pick-up probabilities could be retrieved easily without requiring extra UPMs. Literature review revealed a limited research on the factors that affect the DT of containers (Rodrigue, 2008; Moini et al., 2012). Furthermore, research on freight behavioural modelling literature pointed out reliability as one of the most important factors that influence the choices related to the transportation of the different products (Ben-Akiva et al., 2015; Feo et al., 2011; Fries, 2009). In the context of this research we considered Freight Forwarders (FFs) as the key decision makers that determine the decision of when to pick-up an import container from a port container terminal and developed a Hybrid Choice Model (HCM) where we inserted the “importance of reliability” as a latent variable. Modelling Framework We developed a questionnaire based survey that addressed factors that may influence their decisions on when to pick-up a container from a terminal. Specifically, we requested general information on the main characteristics of the FF and the description of a typical import pick-up from the container terminal. In addition we asked FFs to reply on statements about how they perceive their firm’s reliability. In the last part of the questionnaire we designed an SP experiment were respondents were asked to state how many days, after getting customs clearance, they would leave import containers in the terminal before picking them up. We collected data from 34 FFs in the Middle East during August 2015. Each FF was presented with 8 different scenarios and our total sample consists of 264 observations. The proposed HCM includes an explanatory variable that cannot be directly measured; this is the latent variable which describes the importance that FFs give to the reliability of the services their company offers to clients. For the development of the latent variable model two types of equations are necessary: the measurement equations that link the indicators to the latent variable and the structural equation that quantifies the influence of the company’s socioeconomic characteristics to the latent variable (Ben-Akiva et al., 2002; Walker and Ben-Akiva; 2002; Tsirimpa et al., 2009; Kamargianni and Polydoropoulou, 2014; Kourounioti and Polydoropoulou, 2015). The structure of the HCM is shown in Figure 1 in which the complete set of structural and measurement equations is sketched depicting the relationships between explanatory variables and each partial model.

FIGURE 1 HCM Model Structure For the development of the choice model the continuous DT was divided into the following discrete time intervals: 1. Interval 1: Duration 0-1 days 2. Interval 2: Duration 2-6 days 3. Interval 3: Duration 7 days 4. Interval 4: Duration 8-9 days 5. Interval 5: Duration over 9 days For the development of the HCM we made the assumption that the importance a FF gives to providing reliable services influences the decision related to DT. We expect that the higher the importance of reliability for the operations of the company the sooner the container will be picked up from the terminal. Apart from the latent variable in the choice model we inserted: • the container characteristics: o Container type: dummy variable equal to if it is a 20’ft container or not o Royal Client: dummy variable equal to 1 if the client to whom the container belongs is a royal client of the company, meaning that uses only the specific FF to execute transportation of his/hers containers. • seasonality o Spring: dummy variable equal to 1if the pick-up is realised in spring. o Monday: equals to 1 when the container is discharged from the customs inspection on Monday. o Warehouses: dummy variable equal to 1 when the FF owns warehouses. • Rel= latent variable “importance of reliability”. • A disturbance effect (η) term was inserted to account for the panel effect. • The error term ε.

We assumed that the “importance of reliability” depends on the following characteristics of the FF which we inserted in the structural equation of the latent variable model: • Less than 10 employees: dummy variable equal to 1 when the FF company has less than 10 employees. • More than 30 employees: dummy variable equal to 1 when the FF company has more than 30 employees. • Delayed deliveries: dummy variable equal to 1 when the FF company delivers more than once per week delayed deliveries. • Scheduled service: dummy variable equal to 1 when the FF executes a scheduled service to the container terminal. • ω= random distribution of errors.

Finally, respondents were asked to state the level of their agreement with the statements in Table 2. The FFs of the sample disagreed that they ensure on-time deliveries only when monetary fines are imposed by the clients. They disagree that delayed shipments only to their good clients may harm the reliability of their companies. In addition, they state that they are willing to accept additional measures from the container terminals in order to guarantee the on-time and undamaged delivery of containers. These statements were inserted as indicators in the equations of the latent variable measurement model. Table 2. Indicators of the Latent Variable “Importance of Reliability” Latent Variable=Importance of Reliability State your level of agreement using a scale (1= totally disagree….7= totally agree) Mean Std. Dev. I would be willing to comply with additional measures to increase a container's safety 6,24 1,044 I would be willing to comply with additional measures to decrease delays. 6,05 0,590 I try to avoid delays only when there are monetary fines imposed by the client 3,43 3,187 It is important to be able to inform my client on time when a container will be delivered. 6,15 0,498

Model Results Model estimations were conducted using Python Biogeme 2.4. The results of the choice model are presented in the table below. Table 3. HCM results Parameters t-stat βo βo1 -1,09 -2,82 βo2 -2,93 -2,25 βo3 -5,52 -3,38 βo4 -5,73 -3,00 βtwenty Βtwenty1 -0,876 -1,39 Βtwenty2 -0,699 -1,55 Βtwenty3 -0,915 -1,84 Βtwenty4 -1,52 -2,13 βroyal_client Βroyal_client1 0,561 1.97 Βroyal_client2 0,352 0,77 Βroyal_client3 0,481 1,99 Βroyal_client4 -0,484 -1,86 βMonday βMonday1 0,442 0,83 βMonday2 0,538 0,98 βMonday3 1,57 2,83 βMonday4 -0,6 -1,96 βspring βspring1 0,437 1,12 βspring2 0,06 0,94 βspring3 0,576 1,98 βspring4 -1,15 -2,34 β<100 β<1001 0,0468 0,08 β<1002 -1,31 -2,85 β<1003 -0,763 -1,56 β<1004 -1,11 -2,00 Βwarehouses βwarehouses1 0,870 2,04 βwarehouses2 0,153 0,66 βwarehouses3 0,199 0,44 βwarehouses4 -0,701 -1,95 γrel γrel1 0,854 4,37 γrel2 0,646 3,16 γrel3 0,123 2,15 γrel4 0,460 2,85 γrel5 0,490 2,08 Η 1 3,39 Sample size 254 R2 0,357 Model results showed a negative correlation between the 20’ft containers and the utility in all time intervals. Containers that belonged to a royal client of a company are picked up faster and presented higher utility in the lower time intervals. In addition, when the container was discharged from customs early in the week the utility of earlier time intervals increased. Pick-ups during spring are conducted faster. FFs without warehouses did not use the terminal as a storage and presented negative correlation in the last time intervals Companies that conducted less than 100 pick-ups per month tended to leave the containers for less time in the terminal. This variable has a negative sign only in the last time intervals. As we can see for the value of tstat, the latent variable “importance of reliability” influences the DT model. The βς of the latent variable decreased as the DT intervals increased. The model results agreed with our initial assumption that the more important reliability is for a FF the faster the pick-up will be conducted. In the structural equation of the latent variable “importance of reliability” we insert the characteristics of the FF (Table 3). The results of the model showed that the companies with less than 10 employees desired to be more reliable. This can be explained either by the close relationships they develop with their clients or by their need to attract more clients. On the contrary, larger companies with more than 30 employees seem to give less importance to their reliability. The FFs that admitted to frequently face delayed deliveries seem to be less sensitive to providing reliable and trustworthy services. Finally, FFs that believed that reliability is very important to their clients operate a scheduled service for pick-ups to the terminal in order to be able to serve their clients better. Table 3. Results of the structural and measurement model of the latent variable Structural Model Parameters t-stat β_rel 5,03 30,69 θ<10_employees. 0,141 6,60 θ>30_ employees. -0,763 -6,77 θdaily_deliveries -0,368 -6,75 θsceduled_service 0,0247 -4,35 σrel 0,821 6,77 Measurement Model Parameters t-stat a1 0,00 a2 -4,58 -3,11 a3 -2,93 2,39 a4 6,667 5,309 λ1 1,00 λ2 2,07 6,77 λ3 0,806 10,150 λ4 1,296 20,923 υ1 2,48 22,29 υ2 0,0606 22,39 υ3 2,38 31,69 υ4 -0,214 -2,726

Research Implications Terminal operators face a lack of information from the landside transportation parties on how many and which containers will be picked-up every day. Undoubtedly, the accurate prediction of the next day’s tasks would also lead to the optimal allocation of equipment and human resources to avoid overproviding or underproviding equipment and personnel. Excess resources could lead to higher and highly unproductive operational costs. Lack of resources could cause delays on service, congestion inside the terminal and the surrounding road network, which ultimately leads to unsatisfied customers. In addition, being able to understand the determinants of DT can be useful when designing efficient policies to control the amount of time containers spend in the terminal before being picked up. Because many shippers do not own their own storage facilities, as well as low demurrage fees, many shippers choose to keep their cargoes in the terminal’s yard; however, this practice impacts terminal capacity. Therefore, the ever-growing volume of transported cargoes, in combination with the lack of available space for terminal expansion, is expected to force terminal operators to enforce monetary policies or various operational restrictions such as delivery or pick-ups after appointment or higher demurrage fees.

References Ben-Akiva, M., Bolduc, D. and J. Park , (2013). “Discrete Choice Analysis of Shippers’ Preferences.”Freight Transport Modeling, ed. Ben-Akiva, M., H. Meersman, and E., van de Voorde, Emerald Group Publishing Limited, United Kingdom. Ben-Akiva, M., Walker, J., Bernardino, A., Gopinath, D., Morikawa, T. and A. Polydoropoulou, (2002). “Integration of Choice and Latent Variable Models.” Perpetual motion Travel behavior research opportunities and application challenges, In H. Mahmassani (Ed.), Elsevier, Oxford, United Kingdom. Bierlaire, M. (2003). “BIOGEME: A free package for the estimation of discrete choice models”. Presented at the 3rd Swiss Transport Research Conference, Ascona. Bierlaire, M. (2015). “BisonBiogeme: estimating a first model.” Technical report TRANSP-OR 150720. Transport and Mobility Laboratory, ENAC, EPFL. Feo, M., Espino, R., and L. Garcia, (2011). “A stated preference analysis of Spanish freight forwarders modal choice on the south-west European Motorway of the Sea.” Transport Policy, vol. 18, is. 1, pp. 60-67. Fries, H., (2009). “Market potential and value of sustainable freight transport chains.” Zurich: ETH Zurich. Goodchild, M. and P. Val Noronha, (2010). “MeTrIS: Metropolitan Transportation Information System: Applying Space Based Technologies for Freight Congestion Mitigation”, U.S. Department of Transportation, Final Report.

Kamargianni, M., and A. Polydoropoulou, (2014). “Generation’s Y Travel Behavior and Perceptions Towards Walkability Constraints among Three Distinct Geographical Areas.” Presented in the 93rd Annual Meeting of the Transport Research Board of the National Academies, Washington D.C.. Kourounioti, I. and A. Polydoropoulou, (2015). “Understanding Freight Forwarders Time-of-Day Choice Decision Making Framework- A Greek Case Study.” Paper presented at Transportation Research Board, January 2015. Moini, N.,M. Boile, S. Theofanis and W. Levanthal, (2008). “Estimating the determinant factors of container dwell times at seaports”, Maritime Economy and Logistics, vol. 14, pp. 162-177

ABSTRACT. In order to approach a system optimal network state, some travellers need to be directed to routes with higher travel times (i.e. they need to take a detour). Our study aims to explore impacts on road network performance and equity resulting from the provision of social route advice. We discuss our findings in light of observed individual compliance behaviour obtained from a small-scale field experiment. As such, we do not only directly link individual compliance behaviour – specifically in response to social routing advice – with network effects, but also show the potential of information-based demand measures in improving road network efficiency.

ABSTRACT. This paper investigates the relationship between the use of ridehailing services, such as Uber and Lyft in the U.S. market, and the use of other means of transportation and other components of travel behavior. The availability and popularity of ridehailing services are quickly growing. So do their impacts on transportation demand and traffic congestion in cities. For example, a recent study of ridehailing services in the City of San Francisco showed that the share of total trips made with these services (approximately 170,000 trips per day) exceeds 15% of all trips inside the city of San Francisco on a typical weekday (SFCTA 2017), which is equivalent to 20% of total vehicle miles traveled (VMT) inside the city of San Francisco, and 6.5% of total VMT including both intra- and inter-city trips. The impacts of ridehailing services on different components of travel behavior is not clear yet and vary depending on the local context in which the services are provided, the characteristics of users and the availability of other alternatives. Similar to other new shared mobility services, Uber and Lyft can expand the set of mobility options available to most users, separating access to transportation (and automobility) from the fixed cost of auto ownership, increasing the number of reliable, comfortable and affordable options available for a trip (Taylor et al. 2015). As a result, ridehailing services can potentially increase the attractiveness and feasibility of living in a zero-/lower vehicle household. Overall, the behavioral studies that have investigated the adoption of ridehailing services can be grouped in two main areas, depending on their main focus: (1) studies that investigate the factors associated with the adoption and frequency of use of ridehailing, and (2) studies that discuss the potential impacts of the use of ridehailing on travel patterns and other components of travel behavior, such as vehicle ownership, mode choice and vehicle miles traveled (Rayle et al. 2014; Shared-Use Mobility Center 2016; PEW research center 2016; Clewlow and Mishra 2017). To date, the debate about the factors affecting the use, frequency of use and also the potential impacts of ridehailing services is still dominated by descriptive analyses and self-reported behavioral changes. Further, none of the discussed studies confirms the causal relationships among the use of on-demand ride services and different components of travel behavior, largely because of the lack of high-quality data (in particular, longitudinal data) that can shed light on this topic, and help explain the causal relationships associated with it (which would be difficult to evaluate using cross-sectional data). Additional difficulties include the eventual maturation of the impacts of ridehailing use over time and the heterogeneity in the impacts of these services associated with the local context and the characteristics of users (Taylor et al. 2015; Circella et al. 2017). This means that it is not yet clear the extent to which the adoption of ride-hailing services causes any changes in vehicle ownership, mode choice, and vehicle miles traveled as opposed to being those conditions caused by other variables. This study builds on a larger research effort undertaken to investigate the relationships among residential location, individual attitudes, lifestyles, travel behavior and vehicle ownership, the adoption of shared mobility services, and the aspiration to purchase and use a vehicle vs. use other means of transportation in California. A rich dataset was collected in fall 2015 with a comprehensive online survey administered to a sample of 2400 California residents, including millennials (i.e., young adults, 18-34) and members of the preceding Generation X (i.e., middle-age adults, 35-50). The data collection is part of a longitudinal study of the emerging transportation trends in California, designed with a rotating panel structure, with additional waves of data collection planned in spring 2018. We used a quota sampling approach to recruit respondents from each of the six major regions of California and three dominant neighborhood types (urban, suburban and rural), while controlling for sociodemographic targets including household income, gender, race and ethnicity, and presence of children in the household. For additional information on the survey content and data collection, see Circella et al. (2016). Using the California Millennials Dataset, our research team has started to investigate the factors affecting the adoption and frequency of use of ridehailing. Some of these results were summarized in previous publications (Alemi et al. 2017; Alemi et al. under review), in which we explored the factors affecting the use of ridehailing services through the estimation respectively of a binary logit model and of a latent-class adoption model. In this paper, we explore the relationships between the adoption of ridehailing and other components of travel behavior. In particular, we first perform a latent-class analysis of the self-reported behavioral changes induced by the adoption of Uber/Lyft on the use of other travel modes. We then explore the relationships between the frequency of use of ridehailing and the average weekly Vehicle Miles Driven (VMD) reported by individuals. We are currently testing the use of latent-class with a continuous distal outcome to simultaneously classify individuals based on their commute and non-commute travel patterns and to predict their average weekly VMD from the latent-class membership while controlling for other confounders such as built environment variables, attitudes, and socio-demographics. The work described in this extended abstract is in progress. Preliminary results are available from the first part of the analysis. The latent-class analysis of self-reported behavioral changes can provide more meaningful and scientifically interesting results against the noisy background compared to other approaches: Three rather well defined latent-classes were identified in our preliminary latent-class analysis of self-reported Uber/Lyft impacts. We find that the use of Uber/Lyft is more likely to reduce the use of public transportation and walking/biking for urban dwellers, i.e. those who live in a neighborhood with better transit access and quality (transit access and quality were measured based on the information provided by Alltransit website using the geocoded home address) and who are more likely to live in zero-/ very low vehicle-owning households. In contrast, the use of Uber/Lyft is more likely to increase the use of public transportation among suburban dwellers and those who live in low-vehicle households; however, the members of this class tend to be among the least frequent users of ride-hailing services. The last class consists of those who live in lowest quality and access transit areas. The members of this class tend to live in a household with one or more vehicle per household drivers (i.e. vehicle sufficient household) and tend to use Uber/Lyft 1-3 times a month. We find that the use of Uber/Lyft is more likely to reduce the amount of driving among the member of this class. The results from the complete analysis will be available by the IATBR conference in July 2018.

REFERENCES

Alemi, Farzad, Giovanni Circella, Susan Handy, and Patricia L. Mokhtarian. Under review. “What Influences Travelers to Use Uber? Exploring the Factors Affecting the Adoption of On-Demand Ride Services in California”. Submitted for publication in Travel Behavior and Society, and presented at the Transportation Research Board 96th Annual Meeting, Washington DC, January 2017, Paper No. 17-05630. Alemi, Farzad, Giovanni Circella, and Susan Handy. Under review. “Exploring the Latent Constructs behind the Use of On-Demand Ride Services in California”. Submitted for publication in the Journal of Choice Modeling, Special Issue of International Choice Modeling Conference (ICMC 2017). Circella, Giovanni, Lew Fulton, Farzad Alemi, Rosaria M. Berliner, Kate Tiedeman, Patricia L. Mokhtarian, and Susan Handy. 2016. “What Affects Millennials’ Mobility? PART I: Investigating the Environmental Concerns, Lifestyles, Mobility-Related Attitudes and Adoption of Technology of Young Adults in California.” Project Report, National Center for Sustainable Transportation. University of California, Davis, May 2016; available at http://ncst.ucdavis.edu/wp-content/uploads/2014/08/05-26-2016-NCST_Report_Millennials_Part_I_2016_May_26_FINAL1.pdf (last accessed on Nov 1, 2017). Circella, Giovanni, Farzad Alemi, Kate Tiedeman, Rosaria M. Berliner, Yongsung Lee, Lew Fulton, Patricia L. Mokhtarian, and Susan Handy. 2017. “What Affects Millennials’ Mobility? PART II: The Impact of Residential Location, Individual Preferences and Lifestyles on Young Adults' Travel Behavior in California.” Project Report, National Center for Sustainable Transportation. University of California, Davis, October 2016. available at https://ncst.ucdavis.edu/wp-content/uploads/2015/09/NCST_Report_Millennials_Part_II_2017_March_31_FINAL.pdf (last accessed on Nov 1, 2017) Clewlow, Regina R., and Gouri Shankar Mishra. 2017. “Disruptive Transportation: The Adoption, Utilization, and Impacts of Ride-Hailing in the United States.” Institute of Transportation Studies, University of California, Davis, Research Report UCD-ITS-RR-17-07. Pew Research Center. 2016. “Shared, Collaborative and On Demand: The New Digital Economy”. May 2016. http://www.pewinternet.org/files/2016/05/PI_2016.05.19_Sharing-Economy_FINAL.pdf. (last accessed on Nov 1, 2017). Rayle, Lisa, Susan Shaheen, Nelson Chan, Danielle Dai, and Robert Cervero. 2014. “App-Based, On-Demand Ride Services: Comparing Taxi and Ridesourcing Trips and User Characteristics in San Francisco.” Working Paper. University of California Transportation Center (UCTC). https://www.its.dot.gov/itspac/dec2014/ridesourcingwhitepaper_nov2014.pdf (last accessed on Nov 1, 2017). San Franciscio County Transportation Authority (SFCTA). 2017. “TNCs Today: A Profile of San Francisco Transportation Network Company Activity”. Draft Report. June 2017. http://www.sfcta.org/sites/default/files/content/Planning/TNCs/TNCs_Today_061317.pdf (last accessed Nov 1, 2017). Shared-Use Mobility Center. 2016. “Shared Mobility and the Transformation of Public Transit.” http://www.apta.com/resources/reportsandpublications/Documents/APTA-Shared-Mobility.pdf (last accessed on Nov 1, 2017) Taylor, Brian D., Ryan Chin, Melanie Crotty, Jennifer Dill, Lester A. Hoel, Michael Manville, Steve Polzin, et al. 2015. “Between Public and Private Mobility: Examining the Rise of Technology-Enabled Transportation Services.” Special report 319. Transportation Research Board: Committee for Review of Innovative Urban Mobility Services.

ABSTRACT. Car-sharing could have substantial benefits. However, there is a lack of evidence about if more people choosing car-sharing would reduce private car usage or, instead, shrink public transport demand. This work aims to bridge the gap by first studying carsharing choice behavior via a mode choice analysis and then revealing the pattern of modal split changes via a scenario analysis. Policy implications are subsequently obtained on the possible measures that could effectively bring down private car usage. The case study is Taiyuan-China; stated and revealed preference data are both collected. Mixed nested logit models are developed to study the combined SP/RP data. The analysis is conducted separately for a shorter trip case (2km to 5km) and a longer trip case (more than 5km) to examine if results would differ by distance. It is found that raising the cost of private car usage (travel cost, parking cost) should be prioritized for shorter trips since car is more difficult to be substituted when trip distance increases. Shorter trips also need such direct measures to help suppress the demand for private car when promoting a car-sharing service; otherwise car-sharing would attract more bus users instead. Longer trips need a more effective solution to bring down private car usage and that is discovered as making car-sharing service more appealing so that it can serve as a practical substitute to private car. A number of informative indicators (e.g. value of travel time savings, direct and cross point elasticity) are also derived to enrich the findings.

ABSTRACT. Abstract The concept of shared travel, making trips with other users via a common vehicle, is far from novel. However, a changing technological climate has laid the tracks for new dynamically shared ride modes to substantially impact travel behaviors. These new options are known as dynamic ridesharing, ridesourcing, or transportation networking and the two most prominent providers are Uber and Lyft. Current research is beginning to investigate demand for these new systems and how they impact existing shared modes. However, a comprehensive investigation of all shared ride modes in a particular region, and the temporal evolution of the demand for a mode and its relationship to other shared modes, is missing. The proposed research tackles this important limitation in literature by analyzing ridership data for all major shared ride mode offerings from New York City including subways, taxis, other dynamic rideshare providers, and Citi Bike. This is done to test the primary hypothesis whether new dynamic rideshare modes are drawing ridership from other shared modes, as has been proposed in existing literature. The data is examined at the neighborhood level using a hierarchical dynamic linear state space modeling framework to explore the changes in demand over time and to analyze the potential interactions across modes. In addition to the spatial and temporal factors, information on the land use, socioeconomic, demographic, and environmental factors will also be accounted for in the model. The findings of this work will offer substantial insights into not only different shared ride mode offerings but also potential interrelationships across the modes. These findings in turn can be used to analyze the impacts of shifting demand due to emerging transportation technologies within the context of a major metropolitan area. It also provides valuable tools for policy insight on how future shared technologies, like that associated with fleet based autonomous vehicles, can integrate into existing multimodal networks.

INTRODUCTION Modern transportation systems, like the ecosystems of daily life they support, both influence and are influenced by shifts in technology. A prime example is the emergence of the smartphone. Every year, the functionality and capabilities of smartphones continue to grow on the supplier side, and on the demand side, there is an increasingly pervasive adoption by consumers (1). Spurred by the same technological and consumer advances, transportation networks, and the modes offered, are experiencing significant changes. One major change that has already influenced how people travel today is dynamic ridesharing. Dynamic ridesharing, also known as ridesourcing or transportation networking, is the concept of using location aware and data enabled mobile devices to actively link riders with drivers through the use of an app-based exchange. To put the adoption of these systems in perspective, Uber, the industry leader, has made over 5 billion trips worldwide since 2010 (2). Lyft, announced in July 2017 that it was completing over 1 million trips per day in the US (3). According to Second Measure, these two ridesharing giants encompass almost 98 percent of the worldwide market, with the other 2 percent coming from smaller companies like Gett, Juno, Sidecar, and Via (4). The growth of dynamic ridesharing is irrefutable, but there has been substantial debate regarding the impact of this growth on other shared ride modes. In a 2017 report by Brooklyn based Schaller Consulting, the sustainability of dynamic ridesharing in New York City was questioned. The report highlights increases in vehicle-miles traveled (VMT) and an undermining of demand for established transit networks as some of the negative impacts due to the growing demand for dynamic ridesharing systems (5). Additionally, recent work by Gerte et al. explored trends in demand for dynamic ridesharing, specifically Uber in New York City in 2014 and 2015 using a panel regression model (6). Their findings indicated that while demand for Uber is growing at the macroscopic scale, the rate of that growth stagnates in heavily residential neighborhoods over time. While these findings do provide some insight into how neighborhood demographic, economic, and built environment variables influence demand for Uber, the research lacks sufficient data to understand how Uber is integrating into the modal landscape of Manhattan. The goal of the research proposed in the paper is to leverage the vast amount of openly available data from New York City to fully explore the temporal evolution of shared mode offerings and also understand the relationships across modes. In particular, the research will bring in ridership data from the full range of shared ride mode offerings including dynamic ridesharing such as Uber, Lyft, Gett, Via, and Juno, and also existing ridesharing services such as the city’s taxi, subway, and bikeshare networks. In an effort to explore the trends over time, demand data for all modes from 2014 through 2017 will be explored. The rest of the extended abstract is organized as follows. In the following sections, relevant literature is discussed, the data sources and some pre-processing elements are identified, the planned modeling framework is presented, and finally expected findings and potential impacts are offered.

RELEVANT LITERATURE Existing literature regarding dynamic ridesharing has been growing over the past few years as data becomes more openly available and the impact of these systems continues to grow. Schaller (2017) explored the impact of this growth in the context of New York City at a high level and concluded that dynamic rideshare growth has detrimental impact on the sustainability goals of New York City (5). In particular, the author notes that by increasing the number of trips made by auto mode dynamic ridesharing does not lead to reduction in VMT and greenhouse gas emissions. This work however does not provide a quantitative framework for understanding the trends and also relationship between dynamic ridesharing and existing shared ride options. Thus it is not usable for forecasting and policy analysis (5). Gerte et al. explored Uber demand over the same study area (6). However, in their work they did not consider any other shared ride modes. There are research efforts that have sought to compare and contrast Uber and Taxi Modes, a traditional shared ride offering, in NYC. Correa et al. (2017) used linear and spatial models in their work to quantify the demand for both Uber and taxis across the entire city from 2014 to 2015. They found that transit access time, roadway length, vehicle ownership, education, employment, and income significantly influence the demand (7). In a comparison of green cabs and Uber outside of Manhattan, Poulsen et al. (2016) found increasing demand for both modes, with Uber growing at a much faster rate (8). Zhao et al. (2016) explored modeling taxi demand at a high spatial resolution using various algorithmic techniques and compared findings for taxi demand predictability to that of Uber. Using their framework for taxi demand, they were able to also predict Uber demand at a higher maximum predictability (9). McKenzie and Baez (2016) explored using Uber and taxi demand to identify events around the city (10). All of these works look at the interaction between a dynamic rideshare mode and taxi, its most similar competitor. However, these works still exclude the majority of additional traditional and new shared mode offerings. Hoffmann et al. (2017) is one of the very few papers that attempted to explore the interplay between different ridesharing options including dynamic ridesharing and public transportation. The authors aimed to investigate the fluctuations between subways and ridesharing while including data on yellow cabs, green cabs, and Citi Bike. They found that subway disruptions positively influence demand for Uber and Yellow cabs suggesting that these modes can serve as a substitute to transit (11). However, their focus on the relationship between shared ride modes was limited to planned or unplanned disruption events. After reviewing the literature, it can be seen that there is a need to understand the temporal evolution of these modes and also to simultaneously analyze the tradeoffs across all shared modes within a single cohesive framework. The proposed research tackles this important limitation in literature by analyzing ridership data for all major shared ride mode offerings from New York City including subways, taxis, other dynamic rideshare providers, and the Citi Bike.

DATA Prior to any comprehensive investigation, a considerable amount of data needs to be collected and compiled for the proposed research. The list of shared services that will be considered is as follows: dynamic ridesharing (Uber, Lyft, Gett, Via, and Juno), taxi (yellow and green), subway, and bikeshare (Citi Bike). Ridership data for Uber is available as pickup trip ends from April-September 2014 and for each month from January 2015 through June 2017. The 2014 dataset comes from a freedom of information (FOIL) request by researchers at FiveThirtyEight, and the latter data is published openly by the city’s for-hire vehicle (FHV) regulators, the Taxi and Limousine Commission (TLC) (12, 13). Along with information on Uber, trip pickup information for the other dynamic ridesharing providers listed above is also included in the FHV files. Ridership information for NYC Taxis (yellow and green) is available through the TLC from 2009 through June 2017 (13). Information on NYC Subway ridership is available through turnstile count data published by Metropolitan Transit Authority (MTA). An additional FOIL request was made by the authors to obtain bus ridership data from the MTA (14). Bikeshare ridership data is made available by Citi Bike at the disaggregate trip level from 2013 through 2017 (15). The built environment (16, 17), environmental (18), and demographic (19) variables have been collected (from Census and other open data sources from NYC) and aggregated in a previous work by Gerte et al. (2017) (6). Due to the limitations in how taxi and dynamic rideshare data are spatially presented all other sources are aggregated to the Taxi Zone level, a spatial system used by the TLC. For reference, Taxi Zones coincide well with existing census tracts and can provide insight at the neighborhood level. In addition to spatial aggregation, the data will also be temporally aggregated to a week based system for tractability.

METHODOLOGY To investigate the influences each shared mode has on one another over both space and time, a hierarchical dynamic linear modeling framework proposed by Gamerman and Migon (1993) will be used in this research (20). The methodology is appropriate for the study time series of cross-sectional data, as is the case with weekly ridership data at neighborhood level for different shared ride modes from 2014 through 2017. The modeling framework proposed Gamerman and Migon can be used to not only explore time dependent and multi-level spatial effects for each mode but the approach can also be used analyze the relationship across different shared ride modes. Through the specification of the observation equation, structural equation(s), and the system equation, different spatial, temporal, and inter-modal effects of interest in this study can be explored. The model will be estimated using Bayesian estimation approaches in R programming language.

EXPECTED FINDINGS AND CONCLUSIONS A discussion of the expected results is presented here. Using the hierarchical dynamic linear modeling framework, the proposed research will attempt quantify the influence new modes such as Uber and Lyft have on other existing and new shared ride services. Additionally the model will not only allow the study of evolution in the demand for individual services over time but also allow the study of how these relationships between modes is shifting over time. Based on findings of other researchers (7-8, 11), there is reason to believe that dynamic rideshare providers do draw demand away from taxis and in some instances away from bus and subway networks as well. Hoffmann et al. (2017) highlight this in their investigation of the influence of subway disruption on dynamic rideshare, taxi, and bikeshare demand. At a high level, this can potentially be explained by the heightened accessibility of these modes through their apps and their ability to provide door-to-door type service. In regard to bikeshare, there is potentially a symbiotic relationship wherein certain group of travelers use bikeshare for short trips, however, they use dynamic ridesharing to complete longer trips or as an alternative to bikeshare during poor weather. Also, by controlling for zonal level attributes like built area and demographics, the study will also offer insights on the relationship between demand for these services and the socioeconomic and demographic factors (e.g. how ethnic background, income, and education influence the demand for these shared services). The limitations of this work are in some ways linked to data availability. As rideshare data is available as only pickup trip ends a fully complete picture of how these modes are used capturing the destination choices and trip lengths cannot be drawn. A second limitation is that this data only represent met demand, and riders that were for some reason not able to access transit, bikeshare, taxi, or rideshare are inherently excluded. With this noted, this proposed research fills an important research gap by analyzing the full slate of shared modes while simultaneously looking at how they interact with one another. This can provide powerful insight for both planners and policy makers in deciding how to address the growth of these new systems in a systematic and sustainable way.

RELEVANCE TO THE CONFERENCE The proposed research addressed two thematic tracks of interest to the conference namely, modeling and simulation, and new data collection/surveys. Related to the former, the proposed research uses a hierarchical dynamical modeling approach to study temporal and multi-level spatial effects in demand for different shared ride modes. Through specification of the model components, the study also attempts to capture the relationship between different shared ride modes. Related to the later theme, the study brings together a variety of data at different spatial and temporal resolutions to analyze an important research question about the growth and evolution of existing and new shared ride mode offerings.   REFERENCES 1. Poushter, J. (2016). Smartphone ownership and internet usage continues to climb in emerging economies. Pew Research Center, 22. 2. Holt, R., Macdonald, A., & Gore-Coty, D. (2017, June 29). 5 Billion Trips. Retrieved 2017, from https://newsroom.uber.com/5billion/ 3. Dickey, M. R. (2017, July 05). Lyft is now completing one million rides a day. Retrieved 2017, from https://techcrunch.com/2017/07/05/lyft-is-now-completing-one-million-rides-a-day/ 4. Molla, R. (2017, August 31). Uber's market share has taken a big hit. Retrieved 2017, from https://www.recode.net/2017/8/31/16227670/uber-lyft-market-share-deleteuber-decline-users 5. Schaller, B. (2017). Unsustainable? The Growth of App-Based Ride Services and Traffic, Travel and the Future of New York City. Schaller Consulting, Brooklyn, NY. 6. Gerte, R., Konduri, K., & Eluru, N. (2017). Is There a Limit to Adoption of Dynamic Ridesharing Systems? Evidence from Analysis of Uber Demand Data from New York City. 7. Correa, D., Xie, K., & Ozbay, K. (2017). Exploring the Taxi and Uber Demand in New York City: An Empirical Analysis and Spatial Modeling (No. 17-05660). 8. Poulsen, L. K., Dekkers, D., Wagenaar, N., Snijders, W., Lewinsky, B., Mukkamala, R. R., & Vatrapu, R. (2016, June). Green Cabs vs. Uber in New York City. In Big Data (BigData Congress), 2016 IEEE International Congress on (pp. 222-229). IEEE. 9. Zhao, K., Khryashchev, D., Freire, J., Silva, C., & Vo, H. (2016, December). Predicting taxi demand at high spatial resolution: Approaching the limit of predictability. In Big Data (Big Data), 2016 IEEE International Conference on (pp. 833-842). IEEE. 10. McKenzie, G., & Baez, C. (2016, January). Uber vs. Taxis: Event detection and differentiation in New York City. In International Conference on GIScience Short Paper Proceedings (Vol. 1, No. 1). 11. Hoffmann, K., Ipeirotis, P., & Sundararajan, A. (2016). Ridesharing and the Use of Public Transportation. from http://aisel.aisnet.org/icis2016/DataScience/Presentations/14/ 12. FiveThirtyEight, (2015) Uber Trip Data. [Data Files]. Retrieved July, 2017, from https://github.com/fivethirtyeight/uber-tlc-foil-response/tree/master/uber-trip-data 13. NYC Taxi & Limousine Commission, (2017) TLC Trip Data. [Data Files]. Retrieved October, 2017. From http://www.nyc.gov/html/tlc/html/about/trip_record_data.shtml 14. Metropolitan Transit Authority. (2017). Turnstile Data. [Data Files]. Retrieved September, 2017. From http://web.mta.info/developers/turnstile.html 15. Citi Bike. (2017). Citi Bike Trip History Data. [Data Files]. Retrieved 2016. From https://www.citibikenyc.com/system-data 16. NYC Planning (2016). PLUTO. [Data File]. Retrieved July, 2017, from https://www1.nyc.gov/site/planning/data-maps/open-data/dwn-pluto-mappluto.page 17. NYC Open Data (2016) [Data File]. Retrieved July, 2017, From https://opendata.cityofnewyork.us/ 18. National Centers for Environmental Information. NOAA. [Data File]. Retrieved July, 2017, from https://www.ncdc.noaa.gov/data-access 19. United States Census Bureau. (2015). [Data Tables]. Retrieved 2017. From https://factfinder.census.gov/faces/nav/jsf/pages/index.xhtml 20. Gamerman, D., & Migon, H. S. (1993). Dynamic hierarchical models. Journal of the Royal Statistical Society. Series B (Methodological), 629-642.

ABSTRACT. Abstract: While municipal governments promote carsharing in their toolkit of policies for reducing private vehicle ownership, traffic engineers loath to allow carshare operators (CSOs) access to on-street parking. CSOs argue that visibility is key to more effective member recruitment and higher utilization of their vehicles. This paper reports on a quantitative analysis of the impact of visibility on recruitment of new members and on vehicle utilization rates. Data from 20 years of member recruitment and vehicle deployment, along with two years of vehicle utilization were analyzed for a station-based CSO in Vancouver. The findings shed light on details that contribute to vehicle utilization rates and recruitment effects associated with the visibility of new vehicles introduced to a given neighbourhood. The findings provide new evidence-based information for CSOs and decision-makers to fine-tune municipal government policies aimed at reducing car ownership through increasing visibility and utilization of CSO vehicles.

ABSTRACT. The focus of this paper is on investigating the potential impacts of ride hailing through a better understanding of (1) the socio-demographic characteristics of ride hailing users, and (2) the spatio-temporal (trip distance and time-of-day) usage patterns of ride hailing services. A unique multi-month trip-level data set, provided by an Austin-based ride hailing company, along with several publicly available data sources, are employed to undertake the analysis.

ABSTRACT. 1. Introduction The advances in simulation techniques involving Quasi Monte-Carlo approaches (for example Halton draws proposed by Bhat, 2001) and increasing computational power have revolutionized the application of maximum simulated likelihood based discrete choice (DC) modeling frameworks. Researchers have formulated and estimated advanced DC model structures involving single dependent variable (such as random parameters and/or error components based multinomial logit models (see Bhat et al., 2007), involving multiple dependent variables (such as simultaneous equation models and/or self-selection based models with endogeneity treatments (see Abay et al., 2013; Eluru and Bhat, 2007). In recent years, the application of these advanced models have become common practice in transportation and other related fields (Cao et al., 2009; Golob, 2003; Guevara et al., 2004). While the model structures have advanced significantly there is one aspect of these frameworks that has received less attention. In the simulation based approaches, traditionally, the emphasis is on capturing the orthogonal error variances with minimal consideration for covariance terms. For instance, consider a random parameters based multinomial logit model. While it is reasonable to assume that, the impact of observed variables varies across decision makers. To accommodate for this, the analyst traditionally adopts a univariate normal distribution assumption on the parameter. However, when multiple random parameters are estimated, it is possible that there is inherent correlation between the various random parameters. Ignoring the presence of such correlation when it is present could lead to biased estimates. Accommodating for this correlation requires developing estimates where the different elements of the variance-covariance matrix are estimated. In recent years, several researchers have recognized this and have proposed approaches to address this challenge (see Abay et al., 2013; Guevara et al., 2009). Bhat (2011) recently proposed the Maximum Composite Marginal Likelihood (MACML) approach that provides an elegant solution. The MACML approach obviates the need for simulation and allows for a relatively reasonable mechanism for estimating the error variance-covariance matrix discussed above. The approach has been applied in several research studies where the covariance parameters were estimated. However, the approach implicitly characterizes the relationship between the various parameters in the form of multivariate normal distribution. Further, extending the approach to more general error-covariance distributional assumptions would not be straight forward. To be sure, most studies that consider error covariance structure in simulation or approximation frameworks have considered multivariate normal distributions only. The current study is proposed to relax this normality assumption by incorporating a copula relationship based random simulation draw generation for estimating variance-covariance matrices. Specifically, we update the current maximum simulated estimation approaches to consider copula based simulation draws. The research will consider six different copula structures: 1) Gaussian, 2) Farlie-Gumbel-Morgenstern (FGM), 3) Clayton, 4) Gumbel, 5) Frank and 6) Joe (a detailed discussion of these copulas is available in (Bhat & Eluru, 2009)). Among these copulas; Gaussian, FGM and Frank copulas represent symmetric dependency structures that ensure higher dependency for unobserved variables around the mean of the distribution. Clayton copula allows for stronger dependency between the unobserved variables for the lower tails of the distribution. Gumbel and Joe distributions offer the mirror image to Clayton copula by allowing for stronger dependency toward the positive tails of the distribution. Between Joe and Gumbel copula, Joe copula allows for a stronger positive tail dependency. By allowing for various copula relationships, the research relaxes the restrictive assumption that the error variance-covariance structure is multivariate normal distributed. The copula generated random draws can account for distinct correlation structures and are likely to enhance the dependency coverage. The proposed approach requires very minimal changes to the existing infrastructure of estimating advanced discrete choice models. Further, the proposed approach can serve as an alternative to any current approach where simulation based maximum likelihood estimation is feasible.

2. Econometric Approach See pdf file

3. Data Source and Empirical analysis The proposed econometric model is estimated using a travel mode choice data sourced from 2008 origin-destination (O-D) travel survey of the greater Montreal area. The analysis is focused on trip level mode choice for home based work trips. The modes considered in our analysis are: driver, passenger, transit, walk, bike, park/kiss and ride, and other mode. Other mode includes taxi, school bus, paratransit and intercity travel. In the final estimation sample, the distributions of these modes are as follows: driver 63.76%, passenger 5.43%, transit 20.58%, walk 4.79%, bike 1.53%, park/kiss and ride 3.60%, and other mode 0.32%. In our analysis, we selected variables from five broad categories: Level-of-Service (LOS) Measures (including travel time and travel cost), Household Sociodemographics (including number of household vehicles and number of household members), Trip maker’s Characteristics (including trip maker’s age, gender and possession of driving license), Urban Form Measures (such as distance of origin from central business district (CBD), distance of destination from CBD, length of highway at origin) and Trip Characteristics (including trip start time). The empirical analysis of the proposed research will follow three steps: model estimation, comparison of data fit measures, and validation analysis. In terms of model estimation, we will estimate four different models: (1) multinomial logit (MNL) model, (2) mixed MNL model without random variable correlations, (3) mixed MNL model with random variable correlations with a multivariate normal distributional assumption and (4) mixed MNL model with random variable correlations presented by different copula functions. Then, in the second step, we will evaluate the performance of the estimated models by using different information criterions (such as Bayesian Information Criterion (BIC)) to identify the model with best data-fit measures. In the final step, we will perform a validation experiment to ensure that the statistical results obtained above are not a manifestation of over fitting to data. We will evaluate both the aggregate and disaggregate measure of predicted fit by using the hold-out validation sample. At the disaggregate level we will compute predictive log-likelihood, Akaike Information Criterion (AIC), Akaike Information Criterion corrected (AICc) and BIC. At the aggregate level, root mean square error (RMSE) and mean absolute percentage error (MAPE) will be computed by comparing the predicted and actual (observed) shares of modes. The predictive fit measures will be computed for two best fitted models identified from step two for the comparison purposes.

REFERENCES Abay, K. A., Paleti, R., & Bhat, C. R. (2013). The joint analysis of injury severity of drivers in two-vehicle crashes accommodating seat belt use endogeneity. Transportation Research Part B: Methodological, 50, 74-89. Bhat, C. R. (2001). Quasi-random maximum simulated likelihood estimation of the mixed multinomial logit model. Transportation Research Part B: Methodological, 35(7), 677-693. Bhat, C. R. (2011). The MACML estimation of the normally-mixed multinomial logit model. Retrieved from Bhat, C. R., & Eluru, N. (2009). A copula-based approach to accommodate residential self-selection effects in travel behavior modeling. Transportation Research Part B: Methodological, 43(7), 749-765. Bhat, C. R., Eluru, N., & Copperman, R. B. (2007). Flexible model structures for discrete choice analysis Handbook of Transport Modelling: 2nd Edition (pp. 75-104): Emerald Group Publishing Limited. 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. Eluru, N., & Bhat, C. R. (2007). A joint econometric analysis of seat belt use and crash-related injury severity. Accident Analysis & Prevention, 39(5), 1037-1049. Golob, T. F. (2003). Structural equation modeling for travel behavior research. Transportation Research Part B: Methodological, 37(1), 1-25. Guevara, C., Cherchi, E., & Moreno, M. (2009). Estimating random coefficient logit models with full covariance matrix: comparing performance of mixed logit and Laplace approximation methods. Transportation Research Record: Journal of the Transportation Research Board(2132), 87-95. Ladron de Guevara, F., Washington, S., & Oh, J. (2004). Forecasting crashes at the planning level: simultaneous negative binomial crash model applied in Tucson, Arizona. Transportation Research Record: Journal of the Transportation Research Board(1897), 191-199.

ABSTRACT. The growing complexity in people’s daily activity-travel behavior along with the overwhelming traffic congestion in major urban areas have motivated development of microsimulation activity-based models (ABMs) as policy sensitive tools to analyze travel demand. Several ABMs have been developed over the past decade. Accounting for the dynamics of activity planning and scheduling is a pivotal feature of such models. However, due to the high complexity of this process, most ABMs rely on predetermined assumptions on the order and timing of planning for activities and their attributes. In fact, such assumptions impose significant restrictions on the behavioral realism and precision of these models. Introducing the notion of activity planning horizon, ADAPTS (Agent-Based Dynamic Activity Planning and Travel Scheduling) and POLARIS (Planning and Operations Language for Agent-based Regional Integrated Simulation) have been proposed to open up opportunities of relaxing assumptions of this kind. An important module of each of these models is the activity conflict resolution, which intends to resolve cases where two or more generated activities conflict in time. This module plays a pivotal role in overall accuracy of the models, as errors in its logic would cause significant levels of imprecision at the aggregate level. The present study intends to propose a new activity scheduling framework, to be implemented in both PLARIS and ADAPTS, that (1) can be applied for any conflicting situation regardless of the number and configuration of conflicting activities, (2) enhances the behavioral robustness of the model, and (3) guarantees feasibility of its resolving strategies.

ABSTRACT. The increase in data sources, computational power and the incorporation of several psychological approaches have prompted the analysis of multiple choice heuristics in a single discrete choice model. Although these models are subject to identifiability issues, there is no formal analysis of this problem and how it relates to the various heuristics used. In this paper, the identifiability of multiple choice heuristic is analysed, showing how the differences among them are crucial in identification. By simulating multiple choice experiments in a real transport mode choice context, we also show how the behavioural differences affect the recovery of individual choice heuristics. We conclude by discussing how different choice heuristics are inter-related in their capacity to be identifiable simultaneously.

ABSTRACT. The Value of Travel Time (VTT) is fundamental in transport economics. The past decades the state-of-the-art practice for VTT estimation has been to use Stated Choice (SC) data. However, there is now plenty of evidence of reference dependence and gain-loss asymmetry in SC data, implying that such data do not reveal long-term stable preferences. This is a serious problem since the value of time is often applied in welfare analyses, where long-term stability of the preferences is a key assumption. A potential reason for the strong reference dependence found in SC data is the emphasis on a short-term reference point often used in SC data to reduce hypothetical bias. In the long-run there is no stable reference point.

An alternative to SC data is to use revealed preferences data and a mode choice model to estimate the VTT. Observed behaviour has adapted to the (stable) travel conditions and should thus be ruled by the long-term stable preferences. Many countries collect NTS (national travel survey) data and spend considerable resources on making them representative, which is an argument for using them for VTT estimation. However, a key problem in the use of NTS data for VTT estimation is the measurement errors in the travel time and travel cost variables. Time and cost in NTS data is either self-reported or derived from a network assignment model.

In this paper we explore the errors in the self-reported and model computed time and cost variables by treating travel time and travel cost as latent variables in the estimation of a mode choice model. We use Swedish NTS data, and a Transcad network to simulate travel time and cost with the state-of-practice method in large-scale modelling. We show how the magnitude of the errors in the input variables can be quantified, and explore the possibility of controlling for these errors to estimate the VTT on NTS data. We also explore the errors in the time and cost variables in a descriptive analysis, for instance with regard to rounding errors and driving costs.

We admit that we face a potential identification problem, i.e. that the random error in the choice model cannot be separated from error in the latent time and cost variables. In this case the assumption of the error structure in the choice model influences the estimated errors in the time and cost variables. We explore this issue by making sensitivity tests in the model formulation. However, given that we use a state-of-practice mode choice model, we argue that it is also relevant to explore the errors in the time and cost variables given this model. We use maximum likelihood measurement-error models, focussing on their application to mode choice models. To our knowledge, no previous study on large-scale transport models has explored the impacts of different model assumptions in error quantification in this way.

Previous studies have shown that regression models - including discrete choice models - are sensitive to errors in variables if not accounting for them, hence parameter estimates are biased towards zero, an effect known as regression dilution. Perception, reporting (e.g. rounding) and modelling errors are just a few of the errors transport models face; therefore, models that can account for these errors are of paramount importance to prevent biased estimates.

Based on previous work, we start by estimating a nested logit model, assuming no error in the input variables. Subsequent models use latent variables to quantify errors in the time and cost variables. We estimate different model specifications, some of which exploit specific model assumptions of the distributions for the latent variables, and show how this influences results. Moreover, we test how the results depend on the assumed error distribution of the time and cost variables.

We apply the 2005/06 Swedish National Survey for the Greater Stockholm Region. This provides 3485 observations, 1556 used PT, of which 46% used bus, 41% used metro and 13% used train. The remaining 1929 observations are divided between walk, bicycle, car driver and car passenger.

Results indicate that time and cost variables normally used in mode choice models, whether reported or derived from networks, carry errors with them; hence, parameter estimates are diluted, and therefore biased. We also find that assumptions regarding the latent variable prior distribution affect parameter estimates, and that skewed distributions for time and cost variables, outperform the normal distribution. The goodness of fit of the assumed error distributions for time and cost variables was measured through the analysis of the measurement equation residuals, and we find that, ceteris paribus, multiplicative measurement-error models outperform additive ones.

The error quantification shows that residuals for cost variables exhibit much larger variance than do time variables. This suggests that cost parameters incur larger errors than time parameters. This results shed some light on the common belief that cost indicators are more error-prone than time indicators because the modeller typically lack cost data based on individual characteristics, such as the type of fuel, car and driving behaviour or fare for public transport. Furthermore, there are reasons to expect that these errors attenuate the cost parameters in transport models and lead to under estimation of the response to pricing measures in appraisal.

In our data, an advantage for analysing time over cost variables is that two time indicators are available for the chosen alternatives – calculated travel time by the assignment software and self-reported time. Results show that whilst the time error variance for public transport modes is similar between the two indicators, we can observe large discrepancies for the car alternatives, the variance of the residuals for reported travel times being larger.

The resulting VTT from the different models are reported. Models not accounting for measurement errors yield higher values of time – between two and four times the values currently used in appraisal – than the models with latent variables, primarily due to higher cost parameter estimates. Results thus suggest that when the model do not account for errors in the travel cost variables, cost parameter estimates are diluted resulting in a too high VTT. Furthermore, VTT estimates from the final specification with latent values yield lower estimates than current VTT from SC data used in appraisal.

ABSTRACT. Motivated by the promising performance of alternative estimation methods for mixed logit models, in this paper we derive, implement, and test expectation-maximization (EM) and minorization-maximization (MM) algorithms to estimate the semiparametric logit-mixed logit (LML) and mixture-of-normals multinomial logit (MON-MNL) models. In particular, we show that the reported computational efficiency of the MM algorithm is actually lost for large choice sets. Because the logit link that represents the parameter space in LML is intrinsically treated as a large choice set, the MM algorithm for LML actually becomes unfeasible to use in practice. We thus propose a faster MM algorithm that revisits a simple step-size correction. In a Monte Carlo study, we compare the maximum simulated likelihood estimator (MSLE) with the algorithms that we derive to estimate LML and MON-MNL models. Whereas in LML estimation alternative algorithms are computationally uncompetitive with MSLE, the faster-MM algorithm appears emulous in MON-MNL estimation. Both algorithms -- faster-MM and MSLE -- could recover parameters as well as standard errors at a similar precision in both models. We further show that parallel computation could reduce estimation time of faster-MM by 45% to  80%. Even though faster-MM could not surpass MSLE with analytical gradient (because MSLE also leveraged similar computational gains), parallel faster-MM is a competitive replacement to MSLE for MON-MNL that obviates computation of complex analytical gradients, which is a very attractive feature to integrate it into a flexible estimation software. We also compare different algorithms in an empirical application to estimate consumer's willingness to adopt electric motorcycles in Solo, Indonesia. 

ABSTRACT. Multi-stage behavior change theory shows promise in the realm of 'soft' transport policy for creating personalized mobility management campaigns. We use the results of an online survey, designed around the conceptual framework posited by the Transtheoretical Model, to construct and compare (a) cumulative link and (b) continuation ratio ordinal regression models regarding walking and cycling adoption. These two model formulations hold distinct insights into the process of behavior change. Furthermore, we investigate the implications of varying the number of stages in our analyses for tailoring active travel policy, mainly through differential variable effects across the thresholds demarcating stage boundaries. We find that various manifestations of identity and revealed multimodal travel behavior are critical for promoting the uptake of walking and cycling.

ABSTRACT. Road transport contributes to about 25% of the EU's total emissions of CO2, and car driving is also associated with unhealthy behavior. Active modes of transport (i.e., cycling and walking) are not only environmentally friendly, but also seen as healthy alternatives (Park, Rink, and Wallace 2006, Hamer and Chida 2008), in particular for relatively short trips within cities. Positive interventions or “soft measures”, such as travel planning, subsidies, marketing, rewards, and PT discounts could stimulate the use of sustainable transport options. Unlike Fiscal regulations to discourage car use, Positive intervention is promoted for behavior change studies since it provides a way to prevent socio-economic inequity (i.e. poor people cannot afford to use the car anymore, whereas rich people are less affected or not at all).

To measure the actual changes of behavior by positive interventions, we will use smartphones to automatically detect travel behavior, offer real-time information about the traffic and provide rewards accordingly, which overcome the stated preference survey that can only measure the intention for changes. Additionally, subjects are fully aware of the fact that they are participating in an experiment is the problem in current effective positive incentive programs (e.g., (Hu, Chiu, and Zhu 2015), (Sanjust, Meloni, and Spissu 2014), (Ben-Elia and Ettema 2011), (Usui et al. 2008) , (Zhu et al. 2015)). This may induce them to behave in a more positive way in order to pleasure the experimenter (List and Levitt 2006). Furthermore, direct measurements of the effects are often lacking and behavioral change is often measured indirectly. So to provide a real-world context is one precondition of this study to gain rich insights into the complex interactions of the user with the environment, or the user context. Last but not least, aforementioned scientific studies were not specifically focusing on cycling. However, commercial apps such as Strava, CycleMaps, BetterPoints, Fietstelweek, CommuteGreener are providing reward to promote cycling but lack rigorous scientific analysis to evaluate their effects. Therefore, this study uses positive interventions in real-world Living labs to promote sustainable travel behavior (change), such as cycling. the SMART (jointly developed by Mobidot and the municipality of Enschede) app has been used to promote cycling. The objective of this paper is to test the effectiveness of positive interventions in a living lab environment for cycling promotion.

First, we use persuasive technology (Cialdini 2001) to design the interventions into SMART app to carry out the interventions. The positive interventions are: (1) Task/challenges for cycling promotion; (2) Feedback on historical behavior; (3) Providing event and traffic information messages; (4) Rewards that provided upon completion of the challenge; (5) A regularly updated and maintained feature, and third party redeem stores to provide users with credible and authentic feeling. The SMART app is show in Figure 1, which depicts the SMART app dashboard. Users can explore the whole functions of the app from this page. In the front end, users can select and see the challenges provided by the operator, as well as their observed trips (including route and mode), this is done by continuously tracking (using GPS, accelerometer data, etc.), and using advanced algorithms that combine travel speeds and travel routes to determine which transportation mode the traveler is using. When users complete a challenge, we provide rewards (points) upon completion of the challenge. The earned points can then be redeemed for various discounted products and services.

Second, we created a real-life context—a living lab to truly analyze the traveler’ behaviors. Participants who used the SMART app were recruited via different municipal communication channels in which the main objective was to promote cycling in general in a challenge and reward campaign. Particularly, participants were not told that they were participating in an experiment and could join at any time within the duration of the campaign through the SMART app. As a result, we created a real-life context. However, as the participants could immediately use all functionalities of the SMART app, we are not able to do a before measurement, which is an important drawback of this study. This living lab was carried out in the city of Enschede, a midsized city in the Netherlands with approximately 158,000 inhabitants. There are currently two campaigns in this living lab, one fixed choice challenge campaign, and one multiple choices challenge campaign. Each experiment includes two parts: first the implementation, monitoring and evaluation of the challenge and rewarding scheme, and secondly post-challenge survey to evaluate the participants’ view on their behavior. The whole case study was designed and carried out through the SMART app. The fixed choice challenge campaign, named the Boswinkel challenge, participants were challenged to cycle at least 10 times along a newly paved cycling road (in the Boswinkel neighborhood). The campaign period was from the beginning of October to the beginning of December 2016 (week 41-49 in 2016). During this time period, there were no other challenges. Users who downloaded the SMART app needed to select and join the challenge. After finishing the challenge, a voucher was awarded to the user who could redeem it in the local shops. In total, 139 SMART users joined the Boswinkel challenge. However, 59 users did not have more than 4 weeks continuously GPS travelling data, and were therefore excluded from the analysis. Of the remaining 70 users, 32 completed the challenge. The other 38 users who did not complete the challenge. And 43 out of 139 participants took part in the post-challenge survey. On the other hand, in the multiple choices challenge campaign, developer provided multiple challenges per month, participants can choose the challenge they like. The campaign started in March 2017, and is still on going. Every month, a cycling related challenge is created, within a period of two weeks. For example, in April 2017, participants can choose to cycle 2, 5, 10, 25, or 50 times, or 5, 20, 50, 100, or 200 kilometers in two weeks. Currently, campaigns for three months were analyzed, which is distance or frequency multiple choice challenge in April and July, and only frequency multiple choice challenge in May. Note that April and July were provided the same challenge campaign. It total, 568 users joined the multiple choices challenge campaign, however, only 43 users joined all challenge in April, May and July. 145 out of 568 users participated in the post-challenge survey.

We found that the interventions resulted in new bike trips, on the newly paved road, made by users who completed the challenge, even after the campaign had ended. According to the survey, 44% of participants responded that their behavior had changed due to the campaign, and that they would keep on cycling more often after the campaign. Most users who did not complete the challenge said they do not need to use this road, but 20% of them indicated that they would cycle more if they got challenges and rewards more suited to their personal situation. This implies that challenges are more effective when they are customized based on individuals’ historical travel patterns. However, the extra bike trips on the Boswinkel route were mostly created by change in route choice, since users who completed the challenge mostly live nearby the Boswinkel route and already cycled often. The effectiveness of the interventions for modal shift is not significant, the data shows a slight decrease and increase respectively of the average numbers of car and bike trips, these findings are not statistically significant. In addition, as we used a living lab setting, it is hard to do (controlled) experiments with before and after measurements. To overcome these drawbacks, we are in process to use traffic count data to validate the external effects (such as weather). The fixed choice challenge campaign implies that challenges are more effective when they are customized based on individuals’ historical travel patterns. Therefore, we run the multiple choices challenge campaign. We found 41% of participants who choose same change in April and July, spent less days to complete the challenge, which proves that the multiple choices challenge campaign is more effective than fixed choice challenge campaign which shows no significant results on cycling trip distance change. Additionally, participants tend to choose more difficult distance related challenges than cycling frequency challenges, which might result from that more than 20% participants cycled long distance per trip in average (more than 7 km). Currently, we are in the process of modal shift analysis, and to find out how continues challenges impact participants’ choice each month.

Overall, our study examined the potential impact of interventions to promote cycling by using the SMART app in a living lab environment. Moreover, it indicates that challenges are more effective when they are customized based on individuals’ historical travel patterns, as a result, to increase cycling distance. However, the interventions effectiveness analysis on modal shift is in process.

ABSTRACT. Walking has emerged as a specific research area within travel behavior because of renewed interest in promoting environments that support active and sustainable modes. The benefits of increased walk trips have motivated planning agencies to pursue development of new tools to assess how policy and environmental changes would affect the amount of walking in their communities (Singleton, et al., 2017). These tools need a sound behavioral basis and more precise relationships in order to be an effective tool to guide policy. Although the vast research on travel behavior has established that the built environment has an effect on walking and other transportation outcomes, there are still many questions. Some of the most compelling questions are related to the functional forms and assumed linearities of these relationships. A better understanding of these relationships not only improves model predictive capacity, it also has implications for the ability to transfer findings across different metropolitan areas and areas within one region.

To this end, we explore the statistical and structural relationships between walking activity and various built environment measures. Specifically, this paper will build upon previous work on pedestrian behaviors and examine the following: ● The nonlinearities in the pedestrian behavior and built environment relationship, ● The transferability of these relationships across several urban regimes (intra-regional and inter-regional environments), and ● Multivariate specification of the behavioral choice model to avoid sources of endogeneity by isolating variables.

This research paper is organized as follows. First, we identify some key findings and gaps in the literature to guide our work. Then, we will discuss our methodology to examine the above questions and makes use of spatially-explicit travel behavior data from cities from the Americas. Finally, we discuss our findings and their relevance for both methodological approaches to researching these behaviors and for policies aiming to achieve higher rates of walking.

ABSTRACT. 1 Introduction Bicycle sharing programs are a new non-motorized transport solution which is of great importance to mitigating traffic problems in cities. Though in many countries, a public bicycle system and a bike sharing system are similar, differences between them may be evident in areas where such programs are rapidly developing. Public bicycle schemes (PBS) are often run or subsidized by government and involve massive docking station constructions. It is a not-for-profit scheme and often requires a membership card to enjoy the service. Bike sharing schemes (BSS) on the other hand are for-profit use the Internet for linking users and bicycles. With the help of a smart phone application, a bike can be rented and returned everywhere. In order to cover the last mile problem of other modes of transit, China first introduced PBS in 2005 and has been expanding these systems ever since. In the meantime, bike sharing companies attempted to operate BSS systems in large cities like Beijing, but were not successful (Cui 2016). Not until 2016, when China’s smart phone users hit one billion and smart phone network payment represented 70% of the all payments, did bike sharing companies find their opportunity for success (Chinese bike sharing market research report 2016). Hangzhou is the first city in China to include PBS as a part of the public transportation system (Fishman 2013). As of March, 2017, there were 3,772 docking stations with 85.8 thousand public bikes in total (Cao 2016). A BSS started at the beginning of 2017, and is continue to growing. As of May 1st, 2017, nine companies had put their bikes in the city, bringing about the trend of using shared bikes, although with some problems (Yang 2017). This paper initiates a case study on PBS and BSS in Hangzhou focusing on how user characteristics differ for both systems, and to compare the two schemes from a user perspective.

2 Related works Research on different aspects of PBS or BSS is growing. Focusing on the potential impacts of a shared bike scheme, Javier discussed the possibility for the scheme to be put forward in urban areas (Molina-García et al. 2015), while Ricci drew attention to the influences created by BSSs in Europe and the U.S. on health care, economy and the environment (Ricci 2015). With respect to user characteristics, Buehler proposed the method to reduce private car use by examining city features, citizen travel characteristic, PBS and policy implementation in five European cities (Buehler et al. 2017). Zhao et al. (2014) suggest a positive correlation of the ridership and turnover rate in Chinese cities. Campbell et al. (2016) modeled the factors influencing user choice to change from existing mode to shared bike or shared e-bike and discovered that e-bike tends to attract users as a bus replacement. In terms of bicycle operation, Wang and Wei (2013) gave advice following a comparison of three kinds of operation modes. DeMaio (2014) pointed out the strengths and shortcomings of operators. Zhang et al. (2015) summed up the features of operations based on case study on five cities in China. Advice on policy to PBS development was given by Midgley (2011) according to analysis on bike facilities, investment, operation, benefit and problems concerne. In the field of modeling, Etinne et al. (2014) built a model on analyzing the travel distance to provide information for the location selection of docking stations. Vogel (2010) established a spatial-temporal model and modeled the bike rent in peak hours to optimize docking station location planning. These observations and studies touch on policy, operations, and docking location planning which are essential to early development. However, none of them have clearly examined PBS and BSS by user characteristics or their perception of differences, which is important for market development and the durability of the scheme. To fill this research gap, this paper will examine the factors that influence the use frequency and to what degree they relate to PBS and BSS. The results can help to explore the relationship between the two bike schemes from the user perspectives and to discover the key features of each scheme for users.

3 Survey method To detect differences between the users of PBS and BSS services, a questionnaire was developed focusing on the users’ attributes and their experiences on public bikes and shared bikes. Respondents were also asked about their latest riding experiences on both services. Based on the technique of random sampling, the data was collected by a joint on-line and on-site survey using the same questionnaire. 266 questionnaires in total were used as the sample data of the research.

4 Results To figure out the structure of the use frequency, users are classified into three categories: “frequently use”, “occasionally use” or “hardly use” public bikes or shared bikes. These three categories are determined by the question set in the questionnaire asking for the users’ cycling frequency by public bike and shared bike as “several times a week”, “several times a month” or “almost don’t use”, respectively. 4.1 Who are using these two bicycle services? Descriptive analysis (Table 1) pointed out that the user structures for both service are quite similar. For both services, male users are more common than females. Users are mostly younger than 35, with high cellphone data flow of more than 500M purchased monthly. Most of them do not own a car or e-bike. Commuting, going to school and for leisure were the top 3 travel purposes. As the travel distance increases, the bicycle use decreases. The majority of respondents had at least a bachelor’s degree. For the PBS system, having a graduate level degree was associated with greater use, but not for the BSS system.

Table 1 Descriptive statistics of users for Public Bicycle and Shared Bicycle Schemes Variable Levels Use Frequency of Bicycle Schemes Public Bicycle Shared Bicycle Fa (%) Ob (%) Hc (%) Fa (%) Ob (%) Hc(%) Gender Male 10.9 12.4 32.7 26.7 13.9 15.4 Female 5.3 9.0 29.7 14.3 9.4 20.3 Age 18-24 6.0 9.4 34.2 22.9 13.2 13.5 25-29 4.9 6.4 13.2 12.0 4.9 7.5 30-34 2.6 2.6 5.6 4.5 2.3 4.1 35-39 0.0 2.3 2.3 0.4 1.1 3.0 40-44 1.5 0.4 3.4 0.8 1.1 3.4 ≥45 1.1 0.4 3.8 0.4 0.8 4.1 Education Background High school or below 1.1 0.4 1.9 0.4 0.8 2.3 Bachelor 9.8 11.7 47.0 28.2 16.5 23.7 Master/ Ph.D. 5.3 9.4 13.5 12.4 6.0 9.8 Monthly Cell Phone Data Purchased 0-100 M 0.4 0.4 6.0 1.1 1.5 4.1 101-300 M 0.4 1.9 2.6 1.1 1.1 2.6 301-500 M 0.8 1.5 4.1 1.1 1.1 4.1 ≥500 M 14.7 17.7 49.6 37.6 19.5 24.8 Car Ownership No car 8.7 12.5 44.1 31.6 14.8 19.0 1 car 6.1 6.1 10.6 7.2 4.6 11.0 ≥2 cars 1.1 3.0 7.6 2.3 3.8 5.7 E-bike Ownership No e-bike 9.5 12.2 45.2 30.8 14.4 21.7 1 e-bike 6.1 7.2 14.8 9.9 6.8 11.4 ≥2 e-bikes 0.4 2.3 2.3 0.4 1.9 2.7 Travel Purpose Commute 4.7 3.2 7.5 8.3 2.4 4.7 Go to school 2.8 3.2 18.2 12.6 3.2 8.3 Go home 0.4 2.4 2.8 3.2 1.2 1.2 Out for business 1.6 1.6 2.0 2.0 0.8 2.4 Back to work/school 0.4 1.6 3.6 2.8 2.4 0.4 Pick up kids 0.0 0.00 0.8 0.0 0.0 0.8 Out for meal 1.2 0.8 0.4 1.2 1.2 0.0 See a doctor 0.4 0.00 0.4 0.4 0.0 0.4 Shopping 0.4 2.0 3.2 1.6 2.8 1.2 Leisure 4.0 5.1 13.8 8.7 8.3 5.9 Visit friends 0.4 0.8 1.2 0.4 1.2 0.8 Others 0.4 2.0 7.1 2.0 0.8 6.7 Travel Distance 0-2.9km 8.3 11.6 32.4 24.1 11.6 16.6 3-5.9km 6.2 9.1 16.2 12.9 8.7 10.0 6-8.9km 2.1 2.5 5.4 4.6 2.9 2.5 ≥9km 0.4 0.00 5.8 2.5 2.1 1.7 Note: Sample size = 266 Fa. Frequently use; Ob. Occasionally use; Hc. Hardly use

4.2 What are the factors influencing the use frequency of both bicycle systems? To further detect the factors that influence use frequency, “education background”, “family car ownership”, “travel purpose” and “travel distance” are the four important ones for PBS, whereas “gender” and “monthly cell phone data purchased” are the significant factors for BSS. 4.3 Which specific level of the factors influencing the use frequency of both bicycle schemes? Ordinal logistic regression analysis suggests for public users, the education background of “master/Ph.D.”, travel purposes of “commute”, “out for business” and “out for meal” influence the use frequency, when the BSS counterparts are monthly cell data purchased “more than 500M” and various travel purposes of “commute”, “go to school”, “out for business”, “back to work/school”, “out for meal”, “shopping” and “leisure”. Thus, the BSS system may relate more to an internet-based sharing system related to very flexible use, whereas the PBS system relates to high frequency trips such as commuting likely due to the hard infrastructure requirements.

5 Discussions The outcomes have given some clues to examine the existence of the two systems. In Hangzhou, the public bicycle service has been operating for 12 years with its system being updated and specialized maintenance takes rather good care of the bicycles. It has already become rooted in the city as one of the public transportation modes and attracted a certain group of users. Because its stable performance and the city’s features fit for cycling, the habit of riding bikes has been awakened by it in Hangzhou, which has provided an ideal environment for shared bikes to come into use. From the study, the characteristics of users are alike in both systems since to users, they are both bike for renting. Admittedly, to the service system, the sudden emergence of massive amount of shared bikes operated by various companies has created some impact on public bicycle use. Though BSS has developed a large group of users in a short time, PBS still keeps its clients. To some extent, the previous problem of hard to rent and return bikes especially in peak hours has been mitigated. Furthermore, the two services have their own strengths. Managed by the government without the purpose of gaining profits, PBS provides service with good quality at a low cost. The rent and return procedures is easy as long as the users bring their citizen card. Evidence shown by the study that senior users tend to choose public bikes more because they are easy to ride and do not require a cellphone application download. For BSS, the biggest benefit of it is to park the bicycle freely so long as there are bikes near your location and you have a smart phone. The concept that the bike can take people to the door of their destination is quite attractive. Unlike the public bicycles, the problem of “no available docking in the nearest station” no longer exists. From this point of view, shared bikes are more flexible and convenient to users. Despite of these strengths, there is still space for improvement. For public bicycles, it should keep pace with technology development to open the access for mobile renting system to attract more young users while at the same time allowing for those without such technology to continue to use and benefit from the system. With the subsidy provided by the government, aided bikes (i.e. electric motor assist bikes) can be further developed and put into use to maintain its strength at attracting senior clients. The most severe problem for BSS shared bicycles is that they stay unattended after being put into use. Aimed at earning profits and getting user data, operators do not care if the users put the bicycle in appropriate places that do not interfere with others (i.e. sidewalk crowding or such). Bikes blocking the door way or occupying sidewalks can be found everywhere. Thus, government should take action and work together with operators to add parking area restrictions into cell phone application. To help the BSS develop sustainably, government should supervise the operators, creating an on-line platform which includes all the shared bicycle service so that users only need to download one application and get all the information they want. This can be a win-win cooperation.

References Buehler R, Pucher J, Gerike R, et al. Reducing car dependence in the heart of Europe: lessons from Germany, Austria, and Switzerland. Transport Reviews, 2017, pp. 1-25. Cao Y. Investigation and analysis of Hangzhou public bicycle service satisfaction. Statistical Theory and Practice, Vol. 11, 2016, pp. 44-47. Chen M., Wang D., Sun Y., et al. Service evaluation of public bicycle scheme from a user perspective: a case study in Hangzhou, China. Transportation Research Record, Vol. 2643, 2017, pp. 28-34. Chinese bike sharing market research report 2016. Big Data Research. http://www.bigdata-research.cn/content/201702/383.html. Accessed July 12, 2017. Cui L. Mobike: bicycle version of Didi in the car-hailing world. Traffic Construction and Management, 2016, pp. 74-77. Demaio P. Bike-sharing: History, Impacts, Models of Provision, and Future. Journal of Public Transportation, Vol.12, No.2, 2014, pp. 41-56. Etienne, C., and Latifa O. Model-based count series clustering for bike sharing system usage mining: a case study with the Vélib’ system of Paris. ACM Transportations on Intelligent Systems and Technology, Vol. 5, No.3, 2014, pp. 1-21. Fishman, E., S. Washington, and N. Haworth. Bike share: a synthesis of the literarue. Transport Reviews, Vol. 33, No. 2, 2013, pp. 148-165. Li, Z.-C., M.-Z. Yao, W. H. K. Lam, A. Sumalee, and K. Choi. Modeling the Effects of Public Bicycle Schemes in a Congested Multi-Modal Road Network. International Journal of Sustainable Transportation, Vol. 9, No. 4, 2015, pp. 282–297. Midgley P. Bicycle-sharing schemes: enhancing sustainable mobility in urban areas. Department of Economic and Social Affairs, United Nations, New York, 2011. Molina-García J, Castillo I, Queralt A, et al. Bicycling to university: evaluation of a bicycle-sharing program in Spain. Health Promot Int, Vol.30, No. 2, 2015, pp. 350-358. Ricci M. Bike sharing: A review of evidence on impacts and processes of implementation and operation. Research in Transportation Business & Management, Vol. 15, 2015, pp. 28-38. Vogel P., D. C. Mattfeld. Modeling of repositioning activities in bike-sharing systems. Bruges: World Conference on Transport Research, 2010. Wang W., and Wei W. Comparative study of commercial operation mode of public bicycle systems worldwide: based on perspective of institutional economics. UPI, Vol. 28, No. 3, 2013, pp. 64-69. Yang W., A comparison between public bikes & sharing bikes of Hangzhou residents’ characteristics. Zhejiang University, 2017. Zhang L., Zhang L., Duan Z. et al. Sustainable bike-sharing systems: characteristics and commonalities across cases in urban China. Journal of Cleaner Production, Vol. 97, 2015, pp. 124-133. Zhao, J., Deng, W., and Song,Y. Ridership and effectiveness of bikesharing: the effects of urban features and system characteristics on daily use and turnover rate of public bikes in China. Transport policy, Vol. 35, 2014, pp. 253-264.

ABSTRACT. Pedestrian time-use tendency as well as spatial choices provides a good indication of people’s preference for spatial domains, and it is important for evaluating urban-planning policies in the central areas of cities. However, the existing models for such time-use analysis during walking-related activities are insufficient. 
In this study, we propose a modeling approach for a sequential time-space choice made by pedestrians in urban spaces. We focus on the utility of spending time for a certain spatial domain, and use a dynamic discrete-continuous choice framework to describe the sequential spatial choice and time allocation to each domain while considering the corresponding expected utility of future options. Dynamic-activity domains are introduced for analyzing spatial attributes without predefined links or zones. Such domains represent the gross plans made by pedestrians before making a detailed decision for their immediate area. 
The size and shape of each activity domain is latent. We focus on the situation under time-budget constraints and assume that each domain is limited by the time-space prism between the pedestrian’s origin and destination. Decision timing is another factor used to determine the size of a domain. We assume some latent classes of the unit distance for decision making and estimate the class shares by using the EM algorithm. The utility of each domain can be assumed to be an aggregate utility of each point inside it, with the weight of each point being unknown. We define these weights in a data-oriented manner by using GPS-based observational data. 
The parameters of the model are estimated using GPS-based observational data, and reasonable results are obtained.

ABSTRACT. Please see the attached pdf-file.

ABSTRACT. Autonomous driving might soon become reality due to the rapid development of vehicle technologies in the past years and due to digitalization trends in all areas of the daily life. Experts expect that this might change personal mobility and mode choices since the driver do not have to concentrate on the driving task and can engage in other activities while travelling in a more comfortable way (Anderson et al., 2014, Fraedrich et al., 2016). One use case which is discussed as most attractive for autonomous driving is using it on long-distance trips since these are often felt as exhausting and tedious tasks (Trommer et al., 2016). Also, highway pilot might come on the market sooner than fully autonomous driving in urban environment since it requires a lower level of automation (ERTRAC, 2017). The aim of the study presented in this paper is to estimate how the value of travel time savings (VTTS) for long-distance trips may change when autonomous vehicles are available on the market.

In general, long-distance passenger trips are special cases of travelling since they are rare events (except of extended commuting) but with a high mileage travelled. The modal split for these trips is dominated by the private car; with increasing distance the share of air travel increases as well. Kuhnimhof et al. (2014) showed that long-distance travel has increased in the last years while everyday mobility almost stagnated. Understanding mode choice decisions for long-distance trips and how these will change following the introduction of autonomous vehicles on the market becomes more and more relevant since long-distance travel has a great impact on the environment (Goeverden et al., 2016). At the same time, determinants of mode choice preferences for long-distance travel, such as VTTS, are addressed in theoretical and empirical works considerably less than mode choices by everyday trips (Abrantes & Warman, 2011). Results from a meta-analysis suggest that the decision-making process by mode choices for long-distance trips differs from everyday mobility decisions. For instance, the user type is suggested to have stronger effect than the real mode-specific effect (Shires & Jong, 2009). Moreover there is still a lack of empirical studies addressing VTTS for autonomous driving for long-distance trips.

In order to provide empirical insights on the impact of autonomous driving on VTTS for long-distance travel, an online survey was conducted in 2017 using a sample which was representative for Germany by age and gender. The design of the survey included a combination of revealed and stated preference methods. In the revealed preference part, the participants had to report a recently made leisure or business trip which was of at least 100 kilometers (one direction). Detailed information on the trip, such as destination, trip duration, trip purpose, mode of transportation used for the trip as well as number of companions and overnight stays had to be reported.

The stated preference part of the survey included two stated choice experiments consisting of eight choice situations each. In each choice situation the participants had to choose one mode of transportation for their trip based on the attributes presented for each of the alternative options. Since we aimed to examine how mode choice decision may change when autonomous driving becomes available, we used the first choice experiment which addresses current mode choices as a base line. Consequentially, in the first choice experiment only the currently available modes of transportation for long-distance trips were presented. This included private car, bus, train, and airplane. The second experiment focused on potentially future available alternatives. Hence, the private car was presented as a vehicle able to drive autonomously.

The attributes of the alternatives included in-vehicle time and cost for all modes of transportation as well as access/ egress time, waiting time, and changes only for the alternatives airplane, train and bus. In order to present more realistic situations to the participants, the trip length reported in the revealed preference part of the survey was used as a base, i.e. reference, for computing individual choice situations. The in-vehicle time and cost were estimated using the reported trip length as well as average speed and cost per mode of transportation drawn from a passenger travel demand model and existing cost analysis for Germany (Winkler and Mocanu, 2017, Winkler et al., 2016, ADAC, 2017). A pivot design was used by reducing or increasing the in-vehicle time and cost in the choice situations. The attributes access/ egress time as well as waiting time had fixed attribute levels. Additionally, the alternatives included an attribute indicating whether people had to change or not vehicle during the journey. In order to enhance the design efficiency of the choice experiments, a Bayesian efficient design (Bliemer and Rose, 2006), using the software Ngene (ChoiceMetrics, 2012), was created. The prior parameter values for creating the efficient design were drawn from a pilot study with 25 participants. Before the second choice experiment, the concept of autonomous driving was introduced using a short animated video.

To analyze the data from the two stated choice experiments, mixed logit models were performed using the software PhytonBiogeme (Bierlaire, 2016). We chose mixed logit models (MMNL) over multinomial logit (MNL) in order to cope with some restriction of the MNL, especially considering heterogeneity within the sample and the panel effect within the data (eight choice situations per person) (Hensher and Greene, 2002; Train, 2002). To obtain the final model specification, an iterative procedure was used. The final models included the alternatives´ attributes in-vehicle time, access/ egress time, waiting time, changes, and cost as well as individual mobility characteristics, such as rail pass and car availability. In-vehicle time and the waiting time as well as the alternative-specific constant were considered as random parameters. The distributions of all random parameters were simulated by using 5,000 MLHS draws (Hess et al., 2006).

The results of the model estimations show plausible parameters´ values and signs. Also, time and cost parameters as well as individual mobility characteristics have a statistically significant effect on mode choices for long-distance trips. A comparison of the parameters of the model 1 (current mode choice preferences) and model 2 (future mode choice preferences) shows similar trends. The in-vehicle time in a train is perceived less negatively than time spent in other modes of transportations. Traveling by inter-urban bus is perceived more negatively compared to using a train or private car. The preferences toward using an airplane depend strongly on trip purpose (business or holiday trip) and distance. A comparison of the car in-vehicle time coefficient from the first model with the autonomous car in-vehicle time coefficient from the second model suggests that riding autonomously is perceived less negatively than driving manually.

Since the autonomous vehicle can, by description presented in the short introduction video, be used as a conventional one (i.e. be driven manually), we additionally performed a model in which we estimated two different in-vehicle time coefficients indicating driving autonomously or manually. In order to do this, two dichotomous variables were created based on the response to a question examining a preference toward autonomous driving. The estimated coefficients confirm the results of the first estimations, suggesting that people who prefer riding autonomously perceive car in-vehicle time less negatively than people who prefer to drive manually. Moreover, people who prefer to use an autonomous vehicle as a conventional one perceive the in-vehicle in the car time consequentially very similar as using a non-automated vehicle. Again, this confirms the effect of autonomous driving on time perception.

Further results analysis suggests that driving is perceived more negatively than using a train on long-distance journeys. At the same time, the time spent driving autonomously seems to be perceived similar as sitting in a train. The cost coefficients were estimated considering differences between people belonging to different income classes. The results confirm that people with high household income perceive costs less negative than people with middle or low income. Consequentially, VTTS depends on mode of transportation and the values differ between people belonging to different income classes. The present paper summarizes the results of the study on the impact of autonomous driving on the VTTS by long-distance travel. The results of the study confirm the theoretical postulated reduction of VTTS when travelling autonomously in a vehicle is compared to driving manually. This study incorporates two stated choice experiments in a single survey. The main advantage of the proposed approach is comparing current and future mode choice prefrences using the same sample. This allows identifying changes after introducing a new concept, such as the autonomous driving, by using the results from the first experiment as a base-line. Furthermore, understanding current decision making processes of long-distance travel is very important against the background of the specific characteristics of this type of trips. At the same time, we acknowledge the limitation of the stated prefrence approach, such as a hypothetical bias occuring especially by alternatives not yet available on the market, such as the autonomous driving in a car. Thus, we focussed with the analysis on higher level constructs, such as the VTTS. The full paper will describe more detailed the methodological approach and study design as well as the model estimations steps. It will summarize the main results discussing them in the context of existing reseach works on VTTS for long-distance trips. Also, it provides estimated values for travel time savings, including values for autonomous driving, presented by mode of transportation and for different income classes. Furthermore, effects of trip purpose and trip distance as well as the benefits and challenges related to the estimation of VTTS for autonomous driving using stated preference approach will be discussed.

References

Abrantes, P.A.L, Wardman, M.R. (2011). Meta-Analysis of UK Values of Travel Time: An Update. In: Transportation Research Part A, 45(1), pp. 1-17. ADAC. ADAC Autokosten Frühjahr/Sommer 2017. 2017. Anderson, J. M., Kalra, N., Stanley, K. D., Sorensen, P., Samaras, C. & Oluwatola, O. 2014. Autonomous Vehicle Technology - A Guide for Policymakers. Bierlaire, M. (2016) PythonBiogeme: a short introduction. Report TRANSP-OR 160706, Series on Biogeme. Transport and Mobility Laboratory, School of Architecture, Civil and Environmental Engineering, Ecole Polytechnique Fédérale de Lausanne, Switzerland. Bliemer, M.C.J. and Rose, J.M.. 2006. Designing Stated Choice Experiments: State-of-the-Art. 11th International Conference on Travel Behaviour Research, Kyoto, Japan. ChoiceMetrics, 2012. Ngene 1.1.1 User Manual & Reference Guide. ERTRAC, 2017. Automated Driving Roadmap. Version 6. Date: 03.04.2017. Working Group „Connectivity and Automated Driving“. Fraedrich, E., Cyganski, R., Wolf, I. & Lenz, B. 2016. User perspectives on autonomous driving. A use-case-driven study in Germany. Geographisches Institut, Humboldt-Universität zu Berlin, Arbeitsbericht 187. Hess, S., Train, K.E., Polak, W., 2006. On the use of Modified Latin Hybercube Sampling (MLHS) method in the estimation of a Mixed Logit Model for vehicle choice. In: Transportation Research Part B, Vol. 40, pp. 147-163. Hensher, D. A., and W. H. Greene. The Mixed Logit Model: The State of Practice. Kluwer Academic Publishers, 2002. Kuhnimhof, T., Frick, R., Grimm, B. and Phleps, P. 2014. Long Distance Mobility in Central Europe: Status Quo and Current Trends. 42th European Transport Conference, Frankfurt, Germany. Shires, J.D., Jing, G. C. d. (2009). An international meta-analysis of value of travel time savings. In: Education and Program Planning, 32(4), pp. 315-325. Train, K. E. Discrete Choice Methods with Simulation. Cambrige University Press, 2002. Trommer, S., Kolarova, V., Fraedrich, E., Kröger, L., Kickhöfer, B., Kuhnimhof, T., Lenz, B., Phleps, P. (2016). Autonomous driving: the impact of vehicle automation on mobility behaviour. Van Goeverden, K., van Arem, B. and van Nes, R. 2016. Volume and GHG emissions of long-distance travelling by Western Europeans. Transportation Research Part D. Winkler, C., Nordenholz, F., Knörr, W. (2016). Analysing the modal shift to rail potential within the long-distance passenger travel market in Germany. In: Proceedings of the 44th European Transport Conference, Barcelona, Spain. Winkler, C., Mocanu, T. (2017): Methodology and Application of a German National Passenger Transport Model for Future Transport Scenarios. In: Proceedings of the 45th European Transport Conference, Barcelona, Spain.

ABSTRACT. Please see the submission

ABSTRACT. This paper presents the initial steps of an empirical investigation into the impacts of widespread autonomous vehicle adoption under a number of different use cases for the largest urban metropolis in Canada: the Greater Golden Horseshoe. In particular, this paper provides an overview of the rational for a number of different decisions made with respect to the design of a data collection instrument. The data collection instruments development was done in order to better understand how end users are expected to use AVs (conventional ownership versus ride hailing services) and how these use cases differ along a number of different lines. In particular, factors such as trip purpose, AV penetration rate on the roadway and cost of AVs  are examined. The rational for how this information is collected and why it is important is reviewed in the paper. Finally the paper presents the results from a preliminary modelling exercise done in the context of the responses to the stated preference portion of the survey. The paper concludes with key findings and questions from the preliminary analysis and planned future work to be done using the collected data.

ABSTRACT. Agencies dedicate a great deal of time and effort developing and using software tools to estimate future travel behavior. It’s not clear how people’s travel choices will change with the emergence of disruptive technologies, such as on-demand ride services, autonomous vehicles (AVs), nor is it clear how predictive tools will need to evolve.

Given differing opinions and plenty of speculation on the potential impacts of these technologies in transportation, the focus of our research is to bracket this uncertainty, translate technology trends into the context of a model, understand model sensitivity, and bolster confidence in investment decisions given an uncertain future.

For this purpose, we examined a few regional travel demand models to determine their sensitivity to the foreseeable effects of AVs, one of the disruptive technologies. The findings are expected to serve to enhance the discussions on how AVs may influence the regional planning process.

1. Approximation of Foreseeable Effects of the AVs in Travel Demand Models There are varying levels of vehicle autonomy, ranging from Level 1 (very limited autonomous functions) to Level 4 (fully autonomous driving with no driver required). Our research focused on the likely effects of Level 4 autonomy (on the NHTSA scale).

• Value of Time Because time spent in an AV is likely less stressful and provides opportunities to engage in other activities while traveling to a destination, people may tolerate higher travel times than they do today. Put differently, the personal cost of lost travel time is likely to be lower with AVs than it is with human-driven cars. It is expected travelers using autonomous vehicles will have lower values of time because the opportunity cost of driving will be reduced.

• Use of Auto Mode Automobile travel opportunities, in general, should be increasingly available. People who are unable to drive today such as teenagers, the elderly, people who lost their license will have the opportunity to make auto trips at their discretion rather than rely on walk/bike/transit modes or another driver. Automation and car-sharing will allow travelers access to an auto trip on-demand and in real-time. Similarly, whether they arrive at work via auto or other mode, people will have ready access to a shared and/or autonomous vehicle from their workplace. It will be substantially easier to make an auto-based trip.

• Roadway Capacity When vehicles are autonomous and connected, they offer the greatest potential to increase roadway capacity especially on freeways. The increase in capacity will come from shorter headways, higher speed, less weaving, and more stable traffic flows. We expect that freeway capacity will increase, first on freeways and expressways, then on major arterials.

• Auto Occupancy Autonomous vehicles have a much higher likelihood of being shared than personally owned vehicles. Shared autonomous vehicles will likely become a hybrid of transit and the private car, offering on-demand point to point service for multiple passengers with an origin or destination in similar areas. The ride-sharing aspect pushes down the cost of the service and increases the likelihood of sharing.

• Vehicle Availability Models generally have variables tied to trip rates and auto availability. AVs may increase trip rates due to their greater convenience and ready availability. Greater convenience could lead to more discretionary vehicle trips for shopping, social, leisure or recreational purposes. Additionally, people not licensed to drive will be able to make vehicle trips. Vehicle availability will increase for many households and at workplace locations – especially those in urban areas.

• Transit Access In an area with a highly developed rail transit network, transit usage may increase with better auto availability, as the increase in auto availability allows for more people to utilize the park-and-ride and kiss-and-ride to access rail systems. During the more congested times of day (AM, MD, and PM), access to the rail system by car provides a faster trip than driving the whole way.

• Parking Cost There would be better use of parking supply and lower parking costs with autonomous vehicles. People pay a premium today to park close to their destination. But with an autonomous vehicles, the car would drop a passenger off, and even if the car needs to park, it would have a lot more options (such as under-utilized lot, warehouse, or even return home). On top of that, parking could become more tightly packed, increasing supply and potentially decreasing cost.

Most models include a variable for parking cost in areas where costs are imposed. AVs have the potential to lower or even eliminate these traditional parking costs. However, cities in the future may impose pick-up and drop-off costs for AV use depending on location to help manage peak period traffic demands.

• Terminal Time Travel models define the time needed to park a car and walk to a destination as “terminal time.” The higher a terminal time, the less likely a person will choose an auto for a particular trip.

Whether owned or shared, a Level 4 AV might pick up and drop off¬ passengers closer to their destination. If so, there will be less search time for a parking space, and less walking to the destination from a parking space. AVs are likely to reduce terminal times by eliminating the need to park. The amount of reduction though will depend on the curb space management policies in cities and how they prioritize curb space use.

2. Results Analysis The results of model testing on foreseeable effects of AVs are included in the table below. As shown in the table, the range of results is generally consistent with professional expectations with some exceptions. There are substantially varying levels of sensitivity to changes in model inputs among models or across regions. From the results, we can also see that current modeling platforms appear sensitive to the likely impacts of AVs, but as we learn more about travel behavior within an AV context, these tools will need significant enhancement.

3. Conclusion Current modeling platforms are capable of roughly estimating the effects of AVs on travel demand and behavior, as long as autonomous travel can be approximated via a small number of scenarios. However, it is noteworthy that the range of predicted outcomes for the same scenario varies appreciably between certain models.

While it is certainly possible that the future effects of AVs could vary by geographic region, it would be much more useful to directly model these variables in a context-specific manner rather than approximating them. Future travel forecasting models will need to consider new tools and methods that more directly capture AV-related variables to more reliably predict future outcomes associated with autonomous technology.

ABSTRACT. Some sorts of travel behavior are often promoted by government or society as a whole in order to attain social objectives. The examples extend from legal behaviors such as parking in the designated parking lot or speeding within the maximum speed limit, to recommended behaviors such as choosing transport modes with less carbon dioxide emissions for reducing climate risk or modes with physical activities for improving public health. Policy measures for promoting these transport modes also vary from economic incentives such as subsidy or fine, to psychological approaches such as mobility management. One of the psychological measures is to appeal to the moral consciousness of the people for promoting legal or socially recommended travel behavior. By activating the moral consciousness, some of the people are assumed to choose legal or socially recommended travel behaviors more than illegal or socially discouraged ones. However, to find the effective psychological measure is not an easy task. First, the travel behavior information from a simple questionnaire survey before/after the implementation of a particular psychological measure might be possibly biased because the respondent often disguise his moral consciousness when he answers the questions. Second, the travel behavioral model for describing the change of travel behavior, assumed as the basis for most psychological measures, is so dynamic and complicated that a simple questionnaire survey is quite difficult to capture the whole picture of the psychological state including moral consciousness inside the respondent’s brain simultaneously. The alternative approach to the questionnaire survey is the application of brain function measurement by functional Magnetic Resonance Imaging (fMRI). The fMRI approach has been developed mostly in neuroscience more than a decade, and it has also been applied to behavioral science recently. If (travel) behavior is assumed to be controlled by brain, each element assumed in the travel behavioral model should be coincided with the activation of a certain brain area or the relationship among certain brain areas. The fMRI can monitor the whole brain areas of activation simultaneously when the respondent receives a psychological or other stimulus. The fMRI can also capture the unbiased picture of the activated brain areas even when the participant tries to disguise his moral consciousness. In this study, I conducted fMRI experiments for examining the relationship between moral consciousness and travel behavior by monitoring the activation of the participant’s brain areas in a situation where the questions in relation to moral consciousness were indicated. This study was approved by a research ethics review committee for Tokyo Institute of Technology (A 16130 "Behavioral Consciousness under Social Dilemma Situation" Approval number No. 2016124), and informed consent for participants, information management and anonymization, etc. were carried out according to the guidelines of the committee.

For the fMRI experiments I used the fMRI scanner (3.0-T General Electric Signal scanner) of Tokyo Institute of Technology fMRI laboratory. This device produces a very powerful magnet, and generates a magnetic field (3 Tesla) around it. The participant is requested to enter in this device in a supine state and to read the sentences of the questions on the screen through the dedicated goggle. The participant responds to the question by pressing either the YES or NO button on the response pad, and I monitor the change of the blood flow in the participant’s brain at the response time visualized by the devise. In this study, all 20 participants were 20 to 25 years old right-handed Japanese men who were students of Tokyo Institute of Technology. The participants for the experiment were recruited from the bicycle parkers at the Tokyo Institute of Technology. The participants were composed of 10 legal bicycle parkers who parked their bicycles in the designated area, and 10 illegal bicycle parkers who parked their bicycles outside of the designated area. The legal bicycle parkers were assumed to have higher moral consciousness than the illegal bicycle parkers in this study. I invited 20 participants enough for performing population analysis as a whole and within each legal/illegal category. The place of recruitment was a legal bicycle parking area beside the university cafeteria, and illegal bicycle parking areas around it. After confirming which participant parked which area, I asked for cooperation. The legal bicycle parking area is crowded with students and staff especially during lunch break, but it is widely known that there is an alternative legal bicycle parking area which is mostly vacant and located within even 1 to 2 minute walking distance from the place of the recruitment. Therefore, I can assume that the participants who illegally parked their bicycles are defined as the participants whose illegal bicycle parking was actually observed on a day of recruitment. I conducted population analysis by using the software SPM8. In the analysis, first, signal intensity data representing the blood flow rate is extracted for each voxel representing each area of the brain from the image subjected to the smoothing operation, and arranged as a data matrix. Second, from the design matrix defining the data matrix and the experimental condition, the parameter matrix is obtained by the general linear model. I extract brain areas where a significant difference in fMRI signal intensity is observed between the experimental condition in question and the baseline by t statistical test. The mean and standard deviation are calculated for each condition assuming that the signal intensities within each condition (10 items per condition) are homogeneous. I prepared 40 visual stimulus sentences used in the experiment, and they were classified into 4 conditions of (o) general as baseline (a) social dilemma (b) moral-impersonal dilemma (c) moral-personal dilemma, 10 sentences per condition. The conditions (o), (b), and (c) corresponded to non-moral, moral-impersonal, and moral-personal dilemmas of Greene et al. (2001) respectively, and many of the sentences used in this study in (o), (b), and (c) were translated texts into Japanese from Greene et al. (2001). For the condition (a), I generated new sentences that are categorizes in social dilemma in relation to travel behavior. The sentences were related to the behavioral choice between illegal or socially discouraged behaviors, and legal or socially recommended ones. For the study objectives in this study, I investigated the difference between (o) and (a) only. Each sentence is divided into three sections, "Background", "Selection" and "Action" for each paragraph, and presented in order. While the "Action" screen is displayed, I ask each participant to answer the question. When the presentation of the sentence is finished, a mark of "+" is displayed on the screen, and the participant is asked to look at it without thinking anything. I repeat this 40 times. Presentation randomly chooses (o) (a) (b) and (c), and puts a break every ten trials. The examples of (a) social dilemma problem and (o) general as baseline visual stimulus sentence are indicated below: (a) An example of social dilemma problem In order to shop, Nick decided to go to the supermarket in the distance of 5 minutes on foot in the rain. Nick is recommended to walk as much as possible for his health by the doctor, and he thinks whether to go to the supermarket on foot or by car. Nick's car has very good fuel efficiency and hardly emits carbon dioxide. If Nick goes to the supermarket on foot, he gets wet by rain and it takes time, but the amount of physical activity increases, the health condition improves, and he can save medical expenses of the country. If Nick goes to the supermarket by car, he does not get wet by rain and saves time, but the health condition worsens, and the medical expenses of the country increase. Nick would not like to get wet by rain and would like to save time, so he decided to go to the supermarket by car. Do you think that Nick's behavior is correct? If you think that it is correct, press the index finger button. If you think it is incorrect, press the middle finger button within 15 seconds. (o) An example of general as baseline John is working on a farm and is going to harvest herbs on a tractor. During the tractor's driving, John approaches a branch point that splits into right and left roads. If John operates the tractor and moves to the right road, he can harvest 20 ton turnip. If he proceeds to the left road, he can harvest a 10 ton turnip. If he does not do anything, the tractor will automatically advance to the left road. John operated the tractor and proceeded to the right road to harvest 20 tons turnip instead of 10 tons. Do you think John's behavior is correct? If you think that it is correct, press the index finger button. If you think it is incorrect, press the middle finger button within 15 seconds.

From the results of the analysis, the brain areas working in the most participants regardless of legal or illegal bicycle parking, were the anterior cingulate cortex and the frontal pole. The anterior cingulate cortex has a role of receiving information from the brain group "limbic system" which is deeply involved in emotion, and it was assumed to be activated in this study when processing information received from "limbic system." Since the information on emotion is collected in the anterior cingulate cortex, it was observed to be activated regardless of legal or illegal bicycle parking. Since the frontal pole is activated when making decisions, and in this study I presented stimulus sentences and asked participants to answer the question as to whether the choice is correct or not, it was considered to be activated in both legal and illegally bicycle parkers. Regarding the brain areas which became significant only for legal bicycle parkers whose moral consciousness is supposed to be higher than illegal bicycle parkers, the insular cortex and the temporal pole are the areas which are related to the emotion of "sympathy" when the people think about others and society. The anterior cingulate cortex is also a brain area which is related to the emotion of "sympathy," and the level of activation of the area by legal bicycle parkers is higher than the level by illegal bicycle parkers. The prefrontal cortex dorsolateral part and hippocampus are the brain areas which are related to "memory," and in particular, the prefrontal cortex dorsolateral part manages the memories concerning the schedule of the future, and is activated when planning for the future. For legal bicycle parkers, there was a possibility of imagining the future predicted from past memories, and selecting actions. Regarding the brain areas for illegal bicycle parkers, the activation of the areas which are related to "sympathy" or "memory" was not observed significantly. In this study, I conducted fMRI experiments for examining the relationship between moral consciousness and travel behavior by monitoring the activation of the participant’s brain areas in a situation where the questions in relation to moral consciousness were indicated. I successfully extracted the brain areas in relation to "sympathy" or "memory" which were activated when legal bicycle parkers with higher moral consciousness answered the questions. The findings can provide useful information for activating moral consciousness when the effective psychological measure is designed to promote legal or socially recommended travel behavior.

ABSTRACT. Please see attached uploaded abstract below.

ABSTRACT. Car-following represents a critical component of driving behaviour models that are applied in microsimulation tools. Existing models address several factors that influence car-following behaviour, where the effects of surrounding traffic conditions have received considerable attention. However, driving behaviour is also affected by several factors e.g. drivers’ characteristics, emotional state, fatigue etc. The current work presents the development of a traditional stimulus-response GM car-following model where the effects of time pressure and drivers’ stress have been included in the model specification. Model estimation is based on data collected with the University of Leeds Driving Simulator. The car-following data have been extracted from a motorway setting while physiological “stress” indicators (i.e. heart rate, blood volume pulse, skin conductance) have been collected with a wristband that participants used during the driving simulator experiments. The raw physiological data are used to extract features to be used as indicators in a latent variable that represents stress. The latent variable is then used in the sensitivity-stimulus component of the model. The results indicate that car-following behaviour is not only influenced by time pressure but also from surrounding traffic conditions. The tests of parameter equivalence, applied for the estimates derived from different simulator scenarios, are rejected and thus indicate that drivers’ sensitivity depends on the behaviour of the other drivers. Moreover, the level of stress influences car-following behaviour. These results indicate that human factors significantly influence driving behaviour and should be consider in modelling approaches.

ABSTRACT. The aim of this study is to investigate the relationship between cyclists’ risk perception and their stated preferences in the same risky situations across two different elicitation methods: (1) computer-based videos and (2) virtual reality simulations of road situations, beyond the static pictures are used as a baseline. The behavioural responses on cycling in presented risky conditions are collected together with the stated responses on propensity to cycle and risk perception. Beyond, biofeedback methodology is employed in the form of the electroencephalography (EEG) to provide insight into the temporal sequence of cortical risk processing which gives a better understanding of mechanisms underlying choices. The main contribution of this paper is to better understand travel choices under risk by using the EEG, which provides a multifaceted analysis of the dynamics of risky decision-making. Secondly, the study provides the validation of virtual reality as a tool of risk preference elicitation. Thirdly, this study will potentially contribute to the improvement of decision-making models where they will account for the difference between automatic and rational processes. In terms of real-life implications, the results will shed light on cycling behaviour and ways to improve cycling experience and safety as multiple previous studies show that risk is one of the main deterrents of cycling.

ABSTRACT. Please see attached uploaded abstract below.

ABSTRACT. 1. Introduction A large number of empirical studies indicate that the activity-travel behavior differs across different age groups in the United States. People within different groups may face certain mobility challenges. For example, the older women are found to make least number of trips per day, and they are more likely to suffer medical conditions that limit their ability to travel (Collia et al., 2003), whereas the school children are facing the issue of declining in active transportation (walk or bike) trips which could lead to a worrisome loss of physical activity (McDonald, 2007). Several studies indicate that studying the activity-travel behavior of certain age group is not adequate since the formative experiences have significant influences on individual’s lifelong beliefs, values and behaviors (Bush 2003, McDonald 2015, Garikapati et al., 2016). Bush (2003) pointed out that although both within the 65 or elder age group, the Baby Boomers (born 1946-1964) behaviors differently from those of the Silence Generation (born 1928-1945), since greater portion of them are licensed drivers with high degree of education compared their prior counterparts. The Millennials (born 1979-2000) also exhibit different activity-travel behavior. They have been found to travel less and own fewer vehicles. They also tend to reside in urban or suburban place and spend more time at home (McDonald, 2015). However, a study found the difference in activity-travel behavior of the Millennials only exists in their early adulthood, and they will adopt the activity-travel pattern similar to their prior generation counterparts as they age (Garikapati et al., 2016). Although more and more studies appear to incorporate the generational cohort effect to study activity-travel behaviors, they only analyze univariate dimensions of activities and travel choices. Also, most of these analyses are conducted at an aggregate level thus ignoring the temporal dimensions of activity-travel behaviors. In terms of the temporal dimensions, not only the timing and amount of time spent for traveling and activities is important, the sequence in which the activities and travel are pursued is also important. However, few studies have considered the sequence property of individual’s activity-travel pattern. In order to better understand the transportation needs of different generational cohorts, it is important to study activity-travel behaviors by considering the full pattern over the course of a day defined as a function of the purpose, timing, duration, mode, and accompaniment. Using data from four different waves of the American Heritage Time Use Survey (AHTUS) (including 1965-66 dataset, 1985 dataset, and two waves of American Time Use Survey (ATUS) from 2005 and 2012) and five waves of the National Household Travel Survey (NHTS) (including 1983 dataset, 1990 dataset, 1995 dataset, 2001 dataset and 2009 dataset), this study will examine the activity-travel patterns of five different twentieth century American generations. In particular, the generations considered in the analysis include: GI Generation (birth year: 1901-1924), Silent generation (birth year: 1925 – 1943), Baby Boomers (birth year: 1944 – 1964), Generation X (birth year: 1965 – 1981) and Millennials (birth year: 1982 -2000). Activity-travel patterns will be characterized as sequences based on multiple attributes including activity type, activity duration, travel mode (where applicable) as well as the accompaniment. As can be seen, the temporal attributes including timing and duration will be embedded in the sequence. Three types of clustering methodologies will be applied to the sequences to identify representative patterns. In an effort to characterize patterns within a generation and across generations, representative patterns will be generated by considering sequences for all generations together and also for each generation separately. The representative patterns will be analyzed to identify the similarities and differences in activity-travel patterns of different generations. The rest of the abstract is organized as follows. An overview of the two types of datasets that will be used in this research is presented in the next section. In the following section, the clustering methodologies that will be used are described. In the fourth section, the anticipated results are described. Lastly, in the fifth section, the relevance of the research work to the conference is discussed.

2. Datasets American Heritage Time Use Survey (ATUS) is a database of cross-sectional time use datasets of US individuals collected over six decades. The surveys were designed to measure the amount of time individuals spend doing various activities, such as work, watching television and socializing among others over a 24 hour period. The data are collected from a representative sample of Americans of age 15 and older. Each respondent was interviewed regarding the activities s/he conducted during 24-hour time period, starting at 4 a.m. the previous day and ending at 4 a.m. on the interview day. For each activity, the respondent was asked to report the duration of the activity, and for most activities, the respondent was also asked about if anyone accompanied the respondent during the activity and where the activity took place. Additionally, sociodemographic information such as age, gender, employment status of respondent was also collected in the AHTUS surveys. As noted earlier, four different AHTUS datasets will be used in this research. While AHTUS provides detailed information regarding time use choices, it doesn’t provide enough details about the travel characteristics. In addition to time spent on activities, time spent by individuals on traveling and other travel characteristics such as travel mode are also important. To also understand the travel patterns, National Household Travel Survey (NHTS) will be used. NHTS is cross-sectional dataset providing information about the daily travel behaviors of a representative set of households from the United States. The survey has been sponsored by the Federal Highway Administration and conducted periodically since 1983. The most recent survey was conducted in 2016, and the data has yet been published. In this analysis, data from five previous waves will be used. The dataset includes four types of information, namely, household, person, trip, and vehicle. The household file contains detailed sociodemographic information of each participant household, such as, household income, household location and household size. The person file contains information of each household member including age, gender, and employment. Detailed trip information of all the trips made by each respondent is included in the trip file, such as, trip origin and destination, trip purpose and travel mode. The vehicle file contains information about all vehicles owned by each surveyed household. By combining datasets from multiple waves, large number of observations about different generations at various lifecycle stages can be obtained. Subsequently, activity-travel patterns of generations can be explored while also controlling for the impact of sociodemographic attributes (e.g. age), life cycle factors (e.g. marriage status) and situational variables (e.g. period effects). 2.1 Attributes Used for Defining Activity-Travel Pattern Researchers who have used clustering method for analyzing activity-travel behaviors have tried to incorporate multiple attributes to describe the activity-travel patterns. The commonly used attributes include number of activities, activity type, locations visited, travel modes and average duration of activities and travel (Wilson, 2001, Jon et al., 2001, Shoval & Isaacson, 2007). Based on previous studies, the following attributes will be used to define activity-travel patterns: 1) activity type, 2) travel mode, and 3) accompaniment. Additionally, timing and duration of different activity and travel episodes is embedded into the pattern by considering activity-travel patterns as a sequence of 1440 events corresponding to the 1440 minutes in a day. Each event is described using the information about the three attributes identified above. The timing and duration of the activity and travel episode were also embedded in the sequence by using the same values of activity type, mode, and accompaniment for each one minute interval from start of the activity/trip episode to the end of the activity/trip episode.

3. Methodology for Clustering Activity-Travel Patterns The section describes the three sequential data clustering methods that will be used to identify representative activity-travel patterns for different generations. The methods for clustering sequential data generally fall into three categories: proximity-based method, feature-based method and model-based method. All three methods will be applied in this study to analyze activity-travel patterns. 3.1 Proximity-based method There are mainly two steps in proximity-based method. First, a distance function which devises the dis/similarity between each pair of sequences must be selected. Second, a distance-based algorithm is then used to obtain the clusters. Several distance functions have been proposed such as Euclidean distance, City-block distance and Pearson correlation (Xu & Wunsch, 2005). However, the limitation of some distance functions is that they are not able to capture the sequential dis/similarity. Levenshtein distance function is widely used in biology field to compare the sequential similarity between gene sequences. This function calculates the dis/similarity between two sequences “a” and “b” by finding the smallest operations (insertion, deletion and substitutions) that are needed to transform “a” into “b”. This distance function will be used to compare the individual’s activity-travel pattern. In order to identify clusters from the combined data across multiple surveys, randomized search approach will be applied – this approach is designed explicitly to tackle large-scale data (Ng & Han, 2002). 3.2 Feature-based method Instead of comparing the raw sequences, feature-based method begins with the extraction of a set of features from each individual sequence that captures the temporal information (Bicego et al., 2003). All the sequences are then transformed to feature space. Feature extraction can simplify the sequence by representing it as point vectors. In our analysis, Walsh-Hadamard transformation will be applied to extract features of sequence. Then a standard point clustering algorithm will be applied to obtain the clusters. This method can significantly reduce computational overhead for large-scaled datasets when compared with proximity-based methods (Xu & Wunsch, 2005). 3.3 Model-based method Model-based approaches attempt to build a statistical model for each cluster, and the clustering solution is then identified by a set of models that best fit the data (Bicego et al., 2003). Li and Biswas proposed a framework for model-based clustering. They assumed that the sequential data are generated by a mixture model which is represented by K component Hidden Markov models and a hidden independent discrete variance C, where each value of C represents a component cluster, and then the clustering solution is identified by selecting the mixture model that best represents the data. This framework has been successfully tested and applied to physiology, ecology and social science domains. This methodology will be applied in this research to identify clusters in activity-travel patterns (Li & Biswas, 2002).

4. Anticipated Results Utilizing the methods identified above, it is anticipated that representative activity-travel patterns for each generation will be identified. Sociodemographic attributes, lifecycle factors, and other situational attributes will be used to characterize and analyze the clusters further. It is expected that, across generations, some representative clusters will be similar while others will be very different. For example, the activity-travel pattern for workers across generations will show some similarity. However, some patterns may differ significantly across generations. For example, activity-travel pattern of younger Millennials may differ from the patterns of the older Millennials and the Boomers. They are expected to spend more time for in-home activities and less time for traveling. The results from the research are envisioned to help us better understand the similarities and differences in activity-travel patterns of different generations.

5. Relevance to the Conference The research proposed addresses the thematic track of “pattern recognition” of interest to the conference. Pattern recognition has been used in different areas such as speech, text recognition, face recognition, DNA sequence analysis. But it is a fairly new method for analyzing individual’s activity-travel pattern. In this paper, three different clustering methods for analyzing patterns will be used to identify representative activity-travel patterns for each generational cohort. Using this approach, we do not predefine any patterns but use the clustering algorithm to help us identify previously unknown patterns of daily activity-travel. In addition to understanding the activity-travel patterns of different generational cohorts in the U.S., it is envisioned that the findings from the research can help answer a number of questions about the activity-travel behaviors of two generations namely Millennials and Baby Boomers.

ABSTRACT. Please see attached PDF document

ABSTRACT. It is said that a declining trend of young people’s car ownership and usage has been observed in Japan, similar to some other developed countries. However, the evidence is limited. Considering that owning and using a car spends a lot of money, choices and factors from other life domains would affect the decisions of young people on car-related expenditure behavior under the constraint of limited budget. So, it may be more effective to analyze car-related behavior from a different angle through understanding the changing expenditure patterns, instead of only considering the changing trend of young people’s car ownership and usage. Hence, in this study, a specific emphasis is placed on the interactions among expenditures for different life domains to obtain a more thorough understanding of young people’s car ownership and usage behavior. To this end, a resource allocation model based on multi-linear function (RAM-MLF) (Yu and Zhang, 2015) is applied by which the interactions between car-related expenditures and other domain expenditures could be reflected. To capture changes in expenditure behaviors over time, the model results are compared by using a household-level long-term and large-scale national survey data: i.e., National Survey of Family Income and Expenditure in Japan collected in 1984, 1989, 1994, 1999, 2004, and 2009. This is the first study in the context of Japan.

ABSTRACT. Transport trends like increasing car ownership, driving more frequently and over longer distances were not questioned for decades. Most industrialised countries have witnessed a steady path of motorisation, driven by increasing wealth and income. But many of these countries nowadays notice a saturation or even a decline in car use, a trend described by the term ‘peak car’ (Goodwin & Van Dender, 2013). The trend of decreasing car use has been noticed in Sweden and Norway already in the late 1990s, but it did not receive scientific attention until the 2010s when a similar trend was documented in Australia, North America, Japan and various West-European countries. Kuhnimhof et al. (2013), for example, reported how car use in terms of car-km travelled (as driver or passenger) per trip maker per day decreased or stagnated in a selection of industrialised countries (-14% in the USA 2001-2008, -1.9% in France 1993-2007, -1.5% in Germany 1997-2007 and +0.9% in the UK 1996-2005). Ever since, a growing interest in the topic emerged. Research so far has revealed remarkable differences among socio-demographic groups. Notably, car use per capita has declined among Millennials or Generation Y (i.e., those born between early 1980s and mid-1990s) (Raimond & Milthorpe, 2010; Kuhnimhof et al., 2011; IFMO, 2013).

After having described this ‘peak car’ trend, many studies now try to understand why young adults are less likely to obtain a driver license, and use cars less often. Various possible explanations have been distinguished (Delbosc & Currie 2013; Van Dender & Clever 2013; van Wee 2015) ranging from situational/contextual changes (e.g., demographic changes, changes in motoring affordability, changes in driver licensing schemes) to attitudinal shifts (e.g., young adults being more climate aware, less car-oriented but more technology-oriented, and having urban preferences). The economic crisis of 2008 is often suggested as an important contextual explanation, but ‘peak car’ already occurred around the turn of the millennium. The size and structure of economic determinants of transport demand might therefore be changing, and other type of determinants like mobility attitudes might become more important. It seems like mobility attitudes of Generation Y ‒ compared to older generations ‒ are in favour of more sustainable modes of transport. Hopkins (2016), for example, found that environmental awareness does not prevent young adults from learning to drive but it influences modal choices for those with a driver license.

Nevertheless, research on Generation Y’s mobility attitudes has two important shortcomings. First of all, evidence is generally based on studies (surveys or interviews) organised among young adults only (e.g., Belgiawan et al., 2014; Delbosc & Currie, 2014). Because mobility attitudes are not compared over generations, it remains unclear if the so-called sustainable mobility attitudes of Generation Y are unique to this birth cohort and if it likely represents a new norm. Second, most studies use cross-sectional data which do not offer insights into changes in mobility attitudes over time. Consequently, it is also uncertain if mobility attitudes of Generation Y remain stable over time as they age. Only a few longitudinal studies over different generations exist (e.g., Schiener & Holz-Rau, 2013; McDonald, 2015), but these studies generally use national travel surveys that offer information on factors like gender, income and employment status but not on attitudes.

This study therefore tries to address both shortcomings by comparing data from two cross-sectional surveys (2007, 2017) organised in Ghent, Belgium. Both surveys have comparable questions on mobility attitudes and behaviour, and allow a comparison over generations. Moreover, the 2017 data also offers information on a new generation of young adults called post-Millennials or Generation Z (i.e., those born after mid-1990s) that has reached young adulthood by now. This generation has received no attention so far in the debate on ‘peak car’. Both data sets are used to test three hypotheses about the so-called uniqueness of young adults’ mobility attitudes: (i) Generation Y has a different attitude towards mobility compared to older generations (i.e., being less car-oriented), and this attitude remains stable over time, (ii) Mobility attitudes of Generation Z resemble those of Generation Y, thus representing a new norm towards mobility among young adults, (iii) The influence of mobility attitudes on car use is much stronger for younger generations compared to older generations.

It is important to note that the two surveys (2007, 2017), although containing similar information on mobility attitudes, were organised for different purposes. The 2007 survey was an Internet survey on lifestyles, attitudes and mobility. Invitations to participate were emailed to staff and students of Ghent University and the University of Antwerp. The sample was overrepresented by especially higher educated respondents. The 2017 survey was also an Internet survey, but more focused on subjective well-being, residential relocations and mobility. Invitations to participate were send out in a much more structured way, by mailing postal invitations to respondents in a selection of urban and suburban neighbourhoods in Ghent. This resulted in a representative sample. In order to construct two more or less comparable datasets for our analyses, we restricted the two samples to only those respondents from Ghent with an age 18 or above and an education of secondary school or higher. This resulted in a sample of 312 respondents for 2007 and 1,576 respondents for 2017. Generations were defined as follows: Boomers, born between 1946 and 1966; Generation X, born between 1967 and 1978; Generation Y, born between 1979 and 1990; Generation Z, born in 1991 or after.

Central in this study are mobility attitudes. In both surveys, respondents were asked to indicate on a five-point Likert scale how much they agreed with five statements on mobility (1 = totally disagree, 5 = totally agree). Statements were: ‘Daily travel is boring’, ‘I love being on the road’, ‘Travel time is wasted time’, ‘Arriving at my destination is the only good thing about daily travel’, and ‘I like to discover new and unfamiliar places’. These statements were inspired by previous research on attitudes and mobility (see, e.g., Bohte et al., 2008, for the Netherlands; Bagley & Mokhtarian, 1999, for the USA). A first step in our analysis included a factor analysis (principal component analysis) of these survey items. For both survey years, we found each time one factor representing a positive mobility attitude (41.9% and 36.8% explained variance for 2007 data, and 2017 data respectively).

A second step in our analysis included for each survey year separately an ANOVA to see how this positive attitude towards mobility differs between generations in 2007 and 2017. At first sight, it seems that younger generations obtain lower mean scores, meaning they are less positive towards mobility compared to older generations. This especially holds for the 2017 data (Gen Z: -0.1607, Gen Y: 0.0059, Gen X: 0.0677, Boomers: 0.1980), but not entirely for the 2007 data when Boomers seemed to be the least positive (Gen Z not available, Gen Y: -0.0139, Gen X: 0.1515, Boomers: -0.1738). An ANOVA-test confirms that differences between generations are significant in both years, although for 2007 this is only at a p-value slightly lower than 10% (2007: F-test is 2.412 with p-value 0.091, 2017: F-test is 8.203 and p-value 0.000). In a next step, post hoc tests were used to determine which specific generations significantly differ from each other. For the 2007 data, significant differences (at p-value below 10%) were only found between Generation X and Boomers, not with Generation Y. For the 2017 data, it was especially Generation Z that significantly differed from all other generations, including Generation Y. Generation Y itself did not significantly differ from Generation X again, but differences with Boomers were now found to be significant at p-value of 5%.

In sum, this ANOVA partly supports the first and second hypothesis of this study. Mobility attitudes of young adults seem to be less positive compared to older generations. This especially holds for Generation Z in 2017, but also Generation Y seems to be one of the least positive generations in 2007 as well as 2017. However, differences between Generation Y and X are less strong than initially assumed.

In a next step, we tested the third hypothesis that car use of younger generations is influenced by mobility attitudes to a larger extent compared to older generations. For that, we used the factor scores on the mobility attitudes as input in a series of hierarchical logistic regressions explaining car use for leisure activities. At the same time though, we must be aware that other variables such as age, gender, or car ownership might be associated with attitudes as well. To make sure that such variables do not strongly impact the entire association between attitudes and car use, we have entered these variables in separate ‘blocks’ into the logistic regression analysis. Socio-economic and demographic variables for which we want to control for are entered first, residential location and car ownership in a second block, and attitudes finally in a third block. For each survey year, we estimated one regression model for the whole sample and additional models for each generation separately (thus 4 models in total for the 2007 data, 5 models for the 2017 data) in SPSS Statistics 24.

Logistic regression does not report the proportion of explained variance (R²) as in OLS regression. However, some pseudo R²-measures are reported for each block of predictors that has been added to the models. When interpreting the changes in Nagelkerke R², it is interesting to see that attitudes do not add much explanatory power compared to the other two blocks of variables. This holds for all regressions models over all generations.

Having a closer look at the influence of individual variables (thus not at the block level), it is interesting to see that a positive attitude towards mobility seems to discourage car use for leisure activities. Respondents who are ‘not bored by daily mobility’ or who ‘love being on the road’ tend to use another transport mode than their cars for leisure activities. But this effect of attitudes was found to be significant for Generation X only (in 2007: odds ratio of 0.241 with p-value of 0.060, in 2017: odds ratio of 0.954 with p-value 0.025) and not for the other generations. On the contrary, car use of younger generations is significantly influenced by car ownership and having a driver license. This holds for Generation Y using the 2007 data and Generation Z using the 2017 data. But as Generation Y ages other variables become important. Using 2017 data, we no longer found a significant effect of having a driver license (because probably by then the majority obtained a driver license) but instead found significant effects of age (with less car use among older Generation Y), employment status (with less car use among part-time employment compared to full-time employment) and residential location (with more car use for those residing in suburban compared to urban neighbourhoods). The Wald statistic however indicates that car ownership remains the most important variable of influence in both survey years and for all generations. Consequently, the third hypothesis that attitudes have a stronger influence on modal choices of younger generations compared to older generations seems not to be true.

In sum, the findings of this study suggest that the ‘peak car’ debate should be cautious when it tries to explain the stagnation and decline in car use among young adults on the basis of changes in attitudes. Attitudes provide only a marginal explanation of modal choices. This is also in line with findings from social practice research in mobility (e.g., Shove et al., 2012; Kent & Dowling, 2013; Shove et al., 2015). This cultural paradigm in mobility research does not specifically focus on peak car, but it is also overtly critical on over-reliance on attitudes to determine sustainable behaviour. However, one must keep in mind that findings of our study are based on two datasets with a different set-up and organisation. Nevertheless, it was useful to obtain some preliminary insights into the so-called uniqueness and stability of younger generations’ mobility attitudes. We hope this might inspire more research to determine whether attitudes from Generation Y (and now also Generation Z) remain stable over time (representing a cohort effect), or change due to contextual changes (period effect) or personal advancements in life stages (age effect).

ABSTRACT. Japanese government recently promotes a driving cessation policy which strongly recommends elderly drivers to voluntarily return their driving license to reduce traffic accidents caused by them. However, the driving cessation may lead to a loss of independence, identity and self-esteem for elderly drivers. Because its decision is made by not only the elderly drivers themselves but also for their family members, the driving cessation would result in (1) the decrease of opportunity participating in out-of-home activities, and/or (2) the increase in family members’ burden such as picking up and dropping off the elderly. Hence, household discussion on driving cessation between older drivers and their family members usually becomes quite sensitive and often postponed until an accident occurs. Given such a situation, some manuals have been made for family members to communicate with older adults about their driving cessation (NHTSA and ASA, 2007), yet effective communication on driving risks (and available alternative travel modes) has not been explored well in the literature.

To fill in the above mentioned gap, this study develops a collective household model of driving cessation of older adults, which explicitly takes into account (1) altruistic attitudes of household members (two different types of altruistic attitudes are considered in this study, as mentioned below in detail), (2) multiple household members’ preferences, collectively forming a household preference, and (3) the effects of communication contents on household preference. The developed model is empirically estimated based on the data collected from households (with at least one elderly member) who had/have been considering driving cessation of older adults.

ABSTRACT. Background: Car-following models represent the longitudinal movement behaviour of drivers in a traffic stream. Most car-following models consider the stimulus from a leader vehicle, such as the relative speed, along with other moderating characteristics (e.g., headway) between the leader and the follower vehicles, to model longitudinal movements of the follower vehicle (or the subject vehicle). An important precursor step to this is the identification of the leader vehicle for a given subject vehicle. In homogeneous traffic conditions, it is straightforward to identify the lead vehicle because vehicular traffic streams typically follow lane discipline with the vehicles traveling one behind the other. In heterogenous traffic conditions, however, the identification of a leader is not a trivial task due to the weak lane discipline arising from a wide variation in the vehicle types, sizes, and maneuvering characteristics as well as driving behaviours uncommon to those in homogenous traffic streams (Choudhury and Islam 2016). Such indiscipline in driving behaviours is common in many counties outside the western world and perhaps not uncommon in large and congested cities of the west. In fact, some studies argue that considering a single leader vehicle may not adequately represent driving behaviours (Hoogendoorn and Ossen 2006). This could be particularly so in mixed traffic conditions, with several vehicles and infrastructural elements around a subject vehicle influencing its microscopic movement characteristics. Building on these arguments, one can posit the presence of an ‘influence zone’ around each subject vehicle, where multiple vehicles and roadway elements within the ‘influence zone’ influence the microscopic movement characteristics of the subject vehicle; as opposed to the subject vehicle following a single leader vehicle.

Another characteristic of car following models is that they model the unidimensional, longitudinal movement behavior of vehicles. Lateral movements are largely treated as a part of lane-changing maneuvers, which are modeled separately. This is again a feature common to traffic streams with strong lane discipline. In heterogenous traffic streams with weak (or without) lane discipline, however, a considerable portion of vehicular movements tend to be two-dimensional, with simultaneity of longitudinal and lateral movements – particularly so for two-wheelers. Therefore, it is likely that a two-dimensional characterization of vehicular movements, where both longitudinal and lateral movements are considered simultaneously, better represents driving behaviours in heterogeneous traffic streams.

Current Research: This study proposes a discrete-continuous choice modeling framework for modeling driving behaviours while considering two-dimensional movements based on the influence of multiple leaders in an influence zone around the subject vehicle. The aspects of driving behaviors analyzed at each instance (or time step) are: (1) the angular direction of movement with respect to the longitudinal direction, (2) the decision to accelerate, decelerate, or remain in a steady state, and (3) the extent of acceleration/deceleration conditional on the vehicle’s decision to accelerate/decelerate. All these decisions at an instance characterize the two-dimensional movement of a vehicle at that instance. The discrete choice component of the model is used for analyzing the decision to accelerate, decelerate, or remain in a steady state as well as the angular direction of movement categorized into discrete intervals. The continuous choice component is used for analyzing the extent of acceleration or deceleration conditional on the above decisions. The influence of common unobserved factors is captured through flexible correlations across choice alternatives within the discrete choice component as well as correlations between the discrete and continuous choice components.

The proposed framework accommodates the role of multiple stimuli from vehicles within an influence zone around the subject vehicles. To do so, stimuli (in the form of relative velocities) from multiple vehicles ahead and abreast of the subject vehicle are considered as explanatory variables in the above models. In addition, the role of other variables describing the microscopic traffic environment around the subject vehicle – longitudinal headways, lateral gaps, and the role of roadway geometry and static obstacles may be considered as potential explanatory variables in determining the above identified driving choices – angular movement, accelerate/decelerate/remain in steady-state, and the extent of acceleration/deceleration. Different sizes and shapes of influences zones may be explored to empirically assess which size and shape provides the best statistical fit to observed (and validation) data and provides most plausible behavioural insights. Such analysis helps in identifying which vehicle(s) influence which decision(s) of the subject vehicle and in ordering the influential vehicles in their extent of influence.

To be sure, discrete choice frameworks have been used in the past to model the combination of angular movement and the decision to accelerate/decelerate/remain in steady state for vehicular movements (Shiomi et al., 2012) and pedestrian movements (Robin et al. 2009). However, these studies do not consider the extent of acceleration or deceleration. Other, car-following based choice models that jointly consider the decisions of acceleration/deceleration along with the extent of acceleration/deceleration (Choudhury and Islam 2016, Koutsopoulos and Farah 2012) consider longitudinal movements only, as opposed to two-dimensional movements. More importantly, as discussed earlier, the vast majority of literature on driving behavior considers the stimulus from a single leader only as opposed to potentially multiple stimuli from multiple vehicles. This paper aims at plugging such gaps in each of the above identified studies. Furthermore, this study accommodates the following important properties of driving behavior as postulated by the theories of traffic flow (Chakroborty 2006): (1) a subject vehicle responds to stimuli after a time lag (or reaction time), which typically varies across drivers and vehicle types (Toledo 2003), (2) the extent of stimulus that one needs to respond (by accelerating or decelerating) might also vary across drivers and vehicle types, (3) the magnitude of accelerations typically do not exceed beyond a safe distance (or time) headway between the subject vehicle and the closest lead vehicle, (4) the magnitude of decelerations typically do not exceed beyond a comfortable limit that tends to vary from one situation to another, (5) most driving behavior attempts to reach a stable state where the subject vehicle does not feel the need to accelerate or decelerate. To accommodate heterogeneity, we treat reaction times and safe distance/time headways as random variables, whose distributions are identified using observed decisions and the extent of acceleration/deceleration. In addition, to impose realistic limits on the magnitude of acceleration and deceleration, we utilize truncated distributions for the corresponding variables, whose truncated values are themselves assumed to be randomly distributed to allow for heterogeneity in the maximum possible magnitude for acceleration and deceleration values based on vehicle type and the microscopic traffic environment. Finally, the estimated models will be subject to stability analysis to assess how well the models reflect the driver’s desire to reach a stable state if traffic conditions allow.

Most of the choice-based studies utilize multivariate extreme value (MEV)-based modeling frameworks to accommodate complex correlation patterns discussed earlier. In this paper, we employ the mixed multinomial probit (mixed MNP) structure that can accommodate complex correlation patterns among the utility functions of choice alternatives, correlations between discrete and continuous elements, and the heterogeneity of reaction time and safe distance/time headway variables relatively easily, thanks to recent advances on estimating MNP models using analytic approximation techniques.

Empirical Data: A mixed traffic video data of 30 minutes from an urban arterial stretch of 250 meters in Chennai city of India is used for empirical analysis in this study. The raw video data was originally processed by Kanagaraj et al. (2015). The processed data includes individual trajectories of 3005 vehicles (26.6% cars, 56.4% two wheelers, and 17% other vehicles), with each vehicle’s trajectory including the spatial position, speed, and acceleration/deceleration values in both the longitudinal and lateral directions at a 0.5 sec resolution. This sums up to a total of 111,629 observations. For each observation of a subject vehicle (i.e., a vehicle whose behaviour is being analyzed), we further processed the data to define different shapes and sizes of influence zones around it and extract the traffic and roadway information within the influence zone. For example, Figure 1 depicts a rectangular influence zone around a subject vehicle, with two vehicles directly ahead of it (middle-front compartment), one in the left-front compartment, one in the right-front compartment, and one each in the left and right sides.

Preliminary Results: Preliminary results are presented in Table 1 for regression models of acceleration of cars (conditional on the decision to accelerate) for the influence zone in Figure 1. As can be observed from the table, including the influence of multiple vehicles improves the model fit substantially. Further, the first lead vehicle in the middle front compartment is the most influential, followed by the second lead vehicle in the middle front compartment, and vehicles in the left and right front compartments, in that order. Furthermore, comparing the last two columns suggests that truncating the distribution of acceleration being modeled leads to a significant improvement in goodness of fit. Finally, the signs of all explanatory variables are consistent with their expected effect on the acceleration behaviour of cars. In summary, the preliminary results corroborate our hypothesis that driving behavior in mixed traffic streams is influenced by multiple vehicles and roadway elements around the subject vehicle.

Expected Contributions: The anticipated contributions of this study are: (a) comprehensive discrete-continuous framework to model the driving behavior in heterogeneous traffic streams, (b) Identification and characterization of the influence zones for determining the potential leaders to be considered in predicting the acceleration/deceleration behaviour of vehicles in mixed traffic conditions, (c) Insights on the influence of different vehicles and roadway elements in the influence zone on the driving behaviour of subject vehicles in mixed traffic conditions, and (d) insights on the differences between the driving behaviors of cars and two-wheeler vehicles.

References

Chakroborty, P.,2006. Models of vehicular traffic: An engineering perspective. Physica A: Statistical Mechanics and its Applications 372(1), 151-161. Choudhury, C., Islam M., 2016. Modelling acceleration decisions in traffic streams with weak lane discipline: A latent leader approach.Transportation Research Part C: Emerging Technologies 67, 214-226. Hoogendoorn, S., Ossen S., 2006. Empirical analysis of two-leader car-following behavior. European journal of transport and infrastructure research 6(3), 229-246. Kanagaraj, V.,Asaithambi, G., Toledo, T., Lee, T., 2015. Trajectory data and flow characteristics of mixed traffic. Transportation Research Record: Journal of the Transportation Research Board(2491), 1-11. Robin, T., Antonini, G., Bierlaire, M., Cruz, J., 2009. Specification, estimation and validation of a pedestrian walking behavior model. Transportation Research Part B: Methodological 43(1), 36-56. Shiomi, Y., Hanamori, T., Eng, M., Nobuhiro, U., Shimamoto, H., 2012. Modeling traffic flow dominated by motorcycles based on discrete choice approach. Proceedings of 1st LATSIS Conference. Toledo, T., 2003. Integrating driver behavior modeling, Ph. D. Dissertation, Department of Civil and Environmental Engineering, Massachusetts Institute of Technology.

ABSTRACT. This paper investigates the opportunities and application challenges of using Floating Car Data (FCD) for understanding and modeling user behavior in parking search at the roadside. The literature typically studies the strategic approach of parking in order to assess policies impacts, while this paper investigates the problem of the reproduction of user behavior and time spent to search a parking lot at the destination, as it can deeply affect car travel time modeling and estimation. For this reason, a methodology to calculate the time spent for on-street parking search by using FCD data is proposed. The parking problem involves different issues, such as parking policies, parking search behavior, parking search time estimation. It has been investigated using different approaches, such as stated preference surveys, field observations, discrete choice models, numerical and simulation models. Differently, this paper investigates the opportunities provided in recent years from the introduction of a huge amount of GPS data, which leads the research to new data-driven models [Kaplan and Bekhor (2011), Montini et al. (2012) and Van der Waerden et al. (2014)]. The literature presents very few papers that investigate the problem of estimating on-street parking time (e.g. Belloche, 2015) but none seems to consider the opportunity offered by FCD data. The analysis of the FCD data challenges leads to a new methodology for computing the parking search time. Accurate travel times are useful both for planning (off-line application) and for user’s information and dynamic routing (real-time application), as they influence modal choice as well as departure time or destination choices. The proposed methodology focuses on understanding and modeling paths that vehicles perform in the final part of their trips before reaching the destination with the support of FCD data. In order to capture the parking search starting nearby destination, which plays a key role in the estimation model, an analysis of route complexity has been carried out. Given a study area in which traffic zones and transport infrastructures are individuated to define the supply model. If the proposed methodology is used for real-time purposes, the supply model is updated in its performances at fixed steps according to available real-time data sources. Once defined the supply model, a vehicle matching procedure is used to associate available FCD data to the transport network allowing to define the set of probe vehicles per destination zone, which are used to calculate the revealed LoS attributes for the considered time period. At the same time, a dynamic shortest path algorithm is used to carry out predicted LoS attributes for the above probe vehicles according to the shortest path in terms of transport generalized cost moving from the generic position of a probe vehicle at a specific time to its destination. The dynamic approach in path search is needed because relevant differences occur between peak and off-peak hours or inside the same peak hours due to congestion, and hence network performances (link costs) differ considerably within the simulation period. By this approach, as paths are identified according to departure time and path travel times are estimated by considering link travel times in the instant in which vehicles cross them, dynamic algorithms are able to define correctly origin-destination routes on the network. Revealed and predicted LoS attributes are computed for all probe vehicles reaching the same destination traffic zone; they are used to estimate a gap function, which allows us to define the park search ceiling distance per destination traffic zone. This value represents the distance from the destination that defines the start of the vehicle search for a parking lot in the considered destination traffic zone, which is modeled by spiral trajectories around the destination, through which the parking search time is computed. Application challenges have been tested in the city of Rome aiming at demonstrating the goodness of the proposed approach to obtain both aggregate indicators about parking time for planning purposes and detailed results for real-time applications. The paper reports an in-depth analysis of application results that are here summarized in Table1. Referring to the whole study area, this table describes the incidence of parking time with respect to the total time through the ratio between the parking search time and the total travel time in the morning peak hour on weekdays in both average value and upper bound of the confidence interval. Results show the average on-street parking search time is about the 4% of the whole travel time and this percentage reaches the maximum value of 10% for destination zones in the city center. These values rise to the 8% and 18%, respectively, if we consider the confidence interval upper bound of parking search time. Application results also allowed to identify future directions of this research, which mainly regard efforts for greater robustness of the methodology, the specification and calibration of different gap functions and application to other cities, by considering also a further validation based on the comparison with survey data. Moreover, if experienced privacy constraints on FCD data will be solved, it will be also interesting to specify and calibrate models for different vehicle types.

ABSTRACT. Route choice behavior has predominantly been analyzed from the angle of a single mode, most often the car. Considering route choice in the broader context of multi-modal networks yet opens the way to more complex policy analysis and wider applications. On many levels, the behavior of travelers in multi-modal networks is more complex to model than that of car drivers. To represent a multi-modal trip as a path, it is necessary to combine the networks of available modes via transfer, waiting and/or access links into a so-called supernetwork (Sheffi, 1985). The main modeling challenges are the limited availability of public transport services, since transit lines are subject to a frequency or a schedule, and the definition of alternatives to the observed path. Not only is it more complex to generate realistic path alternatives in a multi-modal network, but there may be a bias in parameter estimates induced by the selection of a restricted choice set (Frejinger et al., 2009). This paper tackles these challenges by applying the recursive logit to model the choice of transit modes and route in a real network. The model is based on the assumption of a full available schedule. The approach presents numerous advantages. First, route choice preferences can be consistently estimated without generating choice sets of paths. Second, the model can be used to predict fast and accurately path choices in real network by sampling from estimated link choice probabilities. Although the network is much larger than previous applications of the RL model with over 1 million links, we obtain reasonable computational times. Third, the approach allows to include all transit services without restriction in one large-scale network, providing the possibility to estimate realistic rates of substitution between different attributes.

ABSTRACT. The agent-based transport simulation framework MATSim allows for the simulation of dynamic transport scenarios with agents, that adaptively make travel choices. Regarding mode choice, a heavily randomized process is used to date, which allows for very unrealistic mode decisions in the short run to arrive at consistent mode shares after a large number of iterations. The authors show that implementing a discrete mode choice model may drastically increase the convergence speed of the simulation, but point out that considerable future research is necessary to make travel decisions consistent and to back the process with a strong theoretical foundation.

ABSTRACT. 1. INTRODUCTION Logistics operators are currently facing the challenge of addressing evolving service expectations among consumers. Amidst growing volumes of goods delivery operations due to diffusion of ICT and e-commerce, customers are also demanding transparency, reduced cost and shorter shipment times. It is consequently essential for city logistics providers to adapt to the market evolution and satisfy the evolving demands related to e-commerce and same-day delivery. One solution to address these issues is crowd-shipping. Crowd-shipping is built on the notion of matching and connecting people or companies needing to send a package with drivers who are willing to deliver the parcels. A particularity of the system is that drivers are typically not professional shippers but occasional couriers from the crowd seeking to earn extra income by delivering packages while traveling along their planned routes and taking advantage of the empty space in their vehicles. The crowd-shipping market is attracting attention as it offers various advantages to its users. This solution is generally less costly than traditional delivery services, allows more flexibility and offers customized services (Rougès and Montreuil, 2014; Goetting and Handover, 2016). It adapts to individual shipping demand and enables users to have more control over pick-up and delivery conditions using smartphone technology. Further, crowd-shipping has the potential to reduce the environmental impacts by optimizing deliveries and reducing the total number of vehicles on the road (Cohen and Muñoz, 2016; Paloheimo et al., 2016). These features make crowd-shipping a promising solution to disrupt the traditional package delivery industry. Yet, crowd-shipping remains an immature market which has an unknown potential to attract new customers. Recent studies have sought to identify the characteristics of potential crowd-shipping users. Punel and Stathopoulos (2017) show, through stated preference scenario studies, that high income individuals are more likely to try crowd-shipping, as are people with high level of employment and low education level. Modeling the willingness-to-work as a crowd-shipping driver, Miller et al. (2017) finds that more likely adopters are those not holding a graduate degree, respondents travelling for leisure and who enjoy the time spent in their vehicle. Such studies provide initial insights concerning the behavior of crowd-shipping users. However, it is also important to understand and explore individuals’ personal motivations for using crowd-shipping. This paper aims to analyze crowd-shipping user behavior and motivations to begin building a knowledge framework around the public view and use of innovative goods delivery systems. Empirical analysis of the individual and contextual factors that promote crowd-shipping acceptance is essential to improve operations and underlying business models of crowd-shipping as well as designing policies to decrease the negative impacts related to freight movement. Through Structural Equation Modeling (SEM), we define several latent factors which characterize crowd-shipping users. The findings promise a better understanding of users’ expectations of the system. The rest of the abstract is organized as follows. First, we provide a description of the data used in this study. Then, we present the methodology to build the model. Finally, we show the SEM results.

2. DATA The analysis is based on an online survey conducted in June 2016 designed for 800 potential respondents living in four U.S. states namely, California, Florida, Georgia, and Illinois. The questionnaire was divided into five sections. The first one covered questions to determine respondents’ previous experience with the parcel and crowd-shipping industries. The second section included a set of stated preference questions to understand whether respondents would shift from a traditional delivery system to crowd-shipping under various scenarios. Section 3 examined the preferences of individuals toward the overall parcel industry, whereas, respondents were asked to give their opinion about potential advantages and disadvantages of crowd-shipping. The fourth section studied attitudes by presenting statements related to two groups of motivations: sense of community and money saving. Finally, section 5 collected the socioeconomic and demographic information. Respondents were selected from both crowd-shipping users and non-users, as the survey aimed to understand the potential users from the general public and their evaluation of crowdsourced delivery options. In total, 587 individuals completed the survey with a return rate of 73.4%. Following the data cleaning process, 533 observations are retained for further analysis.

3. METHOD AND MODEL To understand individuals’ motivations for using crowd-shipping, a SEM is developed. The model aims at explaining the dependent variable of being a crowd-shipping user through factors defined by attitudinal item statements.

3.1. Statement Variables The SEM is primarily built using respondents’ answers extracted from the third and fourth sections of the questionnaire. These sections asked respondents to rate their level of agreement with various statements using a 7-point Likert scale ranging from “Strongly Disagree” to “Strongly Agree”. Statements were designed using suggestions from both a pilot study run prior to the survey and a literature survey, and represented four categories: • Crowd-Shipping’s Positive features (CSP): assets of crowd-shipping systems over traditional delivery services which would encourage its use. • Crowd-Shipping’s Negative features (CSN): reservations about using crowd-shipping • Sense of Community (SC): statements about the potential of crowd-shipping to build a community, as well as personal belief and motivation for belonging to a community in general. In this study, sense of community is defined as the membership in a group where its members interact and help each other in order to improve their common condition. • Money Saving (MS): statements about the economic advantages of crowd-shipping and their personal motivation for saving money. The latter two categories were inspired by previous work where central motivations for using sharing economy systems were to make social connections and to save money (Bellotti et al. 2015, Tussyadiah & Pesonen 2016, Hamari et al. 2015). The questions were tailored in the current study so that respondents expressed thoughts about their own priorities but also their opinion about how well crowd-shipping corresponded to these values. In total, respondents evaluated 27 statements with 15 being retained to build the final SEM shown below.

3.2. Structural Equation Modeling The construction of the structural equation model follows a 4-step process. The first step consists in running exploratory factor analysis over all 27 statements. To apply a first filter and remove inconsistent statements, we start with an item analysis procedure by computing the correlation of each statement versus the rest, and plotting the Locally Weighted Scatterplot Smoother (LOWESS) plots of items against the rest. We then run Exploratory Factor Analysis (EFA) to determine the structure of the factors of the model, and confirm our findings by running Confirmatory Factor Analysis (CFA). The second step introduces the dependent variable of the model, i.e. ‘Crowd-shipping User’ status. We study how this variable is explained by the factors revealed in the first step. The third step introduces the socio-demographic and built-environment variables in the model, as it is probable that such features have an impact on the likelihood of using crowd-shipping, according to preliminary studies. Finally, the fourth step accounts for covariance between the latent variables.

4. INITIAL RESULTS Figure 1 shows the structure of the final model. Being a crowd-shipping user is explained by 3 factors: Finance, Help, and Social & Innovation. The Finance factor refers to individuals looking to save money. It loads on 5 items related to the economic benefits of crowd-shipping, as well as respondent’ general behavioral trait towards saving money. The Help factor represents the interest in helping others. It loads on 3 items which support the subset of the ‘sense of community’ feature related to crowd-shipping as a platform for mutual help with deliveries. Finally, the Social and Innovation factor represents individuals interested in the novel features of crowd-shipping as well as its social dimension of creating relations with other users. It loads on 8 items related to crowd-shipping advantages as well as a subset of the ‘sense of community’ construct. In addition, a number of causal variables that explain the factors are identified among the socio-demographic and built-environment variables. Table 1 presents the factor analysis results while Table 2 shows the general results and statistics of the structural equation model. Study of the sign shows that the Social & Innovation factor has a positive impact on the use of crowd-shipping contrary to the two others which load negatively. Such findings suggest that current crowd-shipping users are not primarily driven by the economic advantages of the system, nor by the mutual aid dimension. However, they value the innovative features of crowd-shipping, as well as its capabilities to create and encourage interaction between its users.

5. SUMMARY AND FURTHER EXPLORATION These results are crucial for the development of crowd-shipping companies. They provide a first exploration of crowd-shippers’ personal motivations for using the system. The model highlights the characteristics they value and care about and how well those values are reflected by the crowd-shipping system. Such results allow the research community, along with the crowd-shipping industry to have a better understanding of the main motivations and representations of crowd-shipping on behalf of its users. Further work will explore the impact of the broader built environment to influence the latent values and user outcomes. In the current results an overarching latent variable ‘diversity’ that seizes on the level of entropy in land-use features in the area of the respondent is included. Initial results suggest that less uniform the local conditions of employment, the less the three value factors seem to matter for the respondent. Further research is needed to pinpoint the right theoretical framework and measurement variables to determine how values and behavior surrounding crowd-shipping are impacted by local circumstances of respondents.

ABSTRACT. This study explores the relationships among environmental, organizational, technological, and individual difference factors, defined as mechanisms influencing ICT acquisition and decision quality of small and medium third-party logistics (3PL SMEs) in the emerging logistics market environment. 3PL SMEs in Lagos State, Nigeria is used as a case study. Mainly, the study investigates whether ICT acquisition and decision quality of the 3PL SMEs depend on the mediation effects of the perceived usefulness and ease of use of ICT; the moderation effects of ICT experience and education status of the owner/managers. Also, whether it depends on the interactive results of the scope of business, consumer readiness, and the facilitating conditions on the usefulness and ease of use of ICT. We adopted the covariant-based structural equation modelling for the analysis based on its suitability for multivariate study. We leveraged integrated framework for e-commerce adoption for SMEs for the research. The result of the investigation indicates that ICT acquisition and decision quality among the 3PL SMEs is probably the outcome of the interrelationship of the above factors. It also connect to a location that is already challenged by other context-specific factors besides resource scarcity. Individually, the result shows that the likelihood of ICT acquisition and decision quality of the 3PL SMEs depends on the facilitating conditions, the scope of business, and ICT experience/education status of the owner/managers. The study provides a nuanced understanding of how poor facilitating conditions in the study site has led to a limited scope of business, low ICT acquisition and decision quality of the 3PL SMEs and outlines practical implications for the policymakers,  logistics operators, and the ICT-vendors. The research limitations deal with a small sample. Nevertheless, the generalization to other locations with a similar environment is plausible as the study site represents a vital commerce centre in the region.

ABSTRACT. Please see the attached file.

ABSTRACT. Background In the United States, trucks are essential to freight transportation. Trucks carry 1,996 billion ton-miles in 2014, which is 37.7% of total ton-miles transported in that year (BTS, 2017). Researchers have found that autonomous vehicles have impact on both inter- and intra-regional passenger transportation (LaMondia et al., 2016; Perrine et al., 2017). Investment in and use of autonomous trucks (Atrucks) may affect not only national and regional economies, but trade patterns, production levels, and pricing of goods. Self-driving, fully-automated or autonomous vehicles are an emerging transportation technology that may transform both passenger and freight transport decisions. Atrucks can potentially reduce some environmental impacts, reduce crash rates, and increase efficiency in warehousing operations, line-haul transportation, and last-mile deliveries. Commercial trucks consume about 20% of the nation’s transportation fuel, and self-driving technologies are predicted to reduce those diesel fuel bills by 4-7% (Liu and Kockelman 2017; Barth et al., 2004; Shladover et al., 2006). Convoy systems would allow long-distance drives with large quantities of goods, through which Atrucks could reduce fuel use by 10-15% (Clements and Kockelman, 2017). Crash counts may fall by 50% or more (Kockelman and Li, 2016), along with various insurance costs. While there is active investigative interest of self-driving cars’ impacts on travel and traffic, research into the implementation impacts of Atrucks is severely lacking.

Data This study anticipates changes in U.S. highway and rail trade patterns following widespread availability of Atrucks. Data sets used include the boundary and zonal data from the U.S. Commodity Flow Survey (CFS), trade flow data from the U.S. DOT’s Freight Analysis Framework (FAF) version 4, and industry-by-industry transaction table and regional purchase coefficients of year 2008 from IMPLAN. Methodology The random-utility-based multiregional input-output (RUBMRIO) model is used to simulate changes in freight flows across the U.S, via a nested-logit model (Ben-Akiva and Lerman, 1978) for shipment or input origin and mode, including the shipper’s choice between Atrucks and conventional or human-driven trucks (Htrucks). Kim et al. (2002) used such a model for estimating interregional commodity flows and transportation network flows to evaluate the indirect impacts of an unexpected event (an earthquake) on nine U.S. states, represented by 36 zones. Kockelman et al. (2005) explored the RUBMRIO model’s application on Texas’s 254 counties, across 18 social-economic sectors and two modes of transport, meeting foreign export demands at 31 key ports. Huang and Kockelman (2010) developed a dynamic RUBMRIO model to equilibrate production and trade, labor markets and transportation networks simultaneously for Texas’ counties over time (better recognizing starting distributions of labor and employment). An equilibrium trade-flow solution (where all producers obtain the inputs they need, and all export demands are met) can be achieved in RUBMRIO, driven by foreign export demands, domestic demands and trade flows between counties that are distributed on the basis of random utilities of rail, Atruck and Htruck. The total production (outputs) of each zone is multiplied by its’ corresponding technical coefficient in order to estimate the total consumption (set of input) required for purchases from all domestic counties including the zone itself. Flow-weighted averages of shipments’ travel costs then create input costs, which merge together to create output costs, as commodities (and labor) flow through the production and trade system. Once the solutions have stabilities (with domestic flow value changing by less than 1% between iterations), final disutilities of travel and trade provide mode shares by OD (origin-destination) pair and industry sector. The RUBMRIO model’s utility functions for domestic and export trade-flow splits (across shipment origin alternatives) depend on the cost of acquiring input type m from zone i, as well as zone i's population. Since there are three mode alternatives for these shipments, with the two truck modes sub-nested, the competing travel costs can be shown as logsums (the expected maximum utility or minimum cost of acquiring that input from different origin zones). Parameter assumptions for random utility functions are adapted from Du and Kockelman’s (2012) work, which has two levels of random utility structure – for origin and mode choices. Results With the assumption that Atrucks lower trucking costs by 25% per value-mile delivered, domestic truck flow values are predicted to rise 2%, while rail flow values fall 16.1%. Rail flows are predicted to rise at short distances (under 250 miles between counties) and very long distances (over 1550 miles), while trucks see flow increases on longer trips: between 350 miles and 2150 miles. Atrucks are shown to be quite competitive for mid- and long-distance trade. However, when input access distances exceed 2000 miles, railway’s lower costs prove very competitive, for many commodities (e.g., those that are less time-sensitive, low value per ton, and/or perishable). For mode splits between Atrucks and Htrucks in terms of transported freight value across domestic trade-flow distances, Htrucks’ appear more valuable in ranges from 0-300 miles in distance, while Atrucks are clearly superior in any distance greater than 750 miles. Interestingly, truck movements appear to peak at just 150 miles of (inter-county) travel distance, for domestic shipments, while Atruck flow values do not peak until 2,750 miles of travel distance. 2,750 miles is essentially the distance separating the nation’s two largest regions: New York City and Los Angeles (as well as Miami to Los Angeles, for example), making this an important OD pair for many commodities (like finance, insurance and service goods). In terms of ton-mile change, domestic truck flows are forecasted to decrease by 8% per ton-mile with rail flow falling by 12.6% per ton-mile. Machinery, miscellaneous, durable and non-durable manufacturing trade flows (between U.S. counties) are predicted to experience a large decrease (greater than 70%) as a result of Atruck implementation. This boost trend in ton-mile also happens to agriculture, forestry, fishing, hunting, chemicals, plastics and primary metal manufacturing, which showed an increase of greater than 60% instead. Atrucks probably show the advantage which could transport these commodities to further destinations in time. With the availability of Atrucks, food, beverage, tobacco product, computer, electronic product and electrical equipment manufacturing ton-mile by trucks increase by approximately 30%. However, rail flow of these products increases by more than double. Although the advent of Atrucks increases the demand, railway remains to be an effective and efficient way for transporting these commodities. Seven sectors see a decrease in total (domestic) value shipped, while 13 sectors see an increase. In terms of export flows, predicted truck flow ton-miles witness increases arranging from 9.8% to 95.7%, except for durable and non-durable manufacturing, which decreases by 90.9%. This might be the same reason as discussed for domestic flow. Total rail flow of commodities headed for U.S. export zones rises by 75.7% while total truck flow decreases by 1.5%. Interestingly, total export flows see rising trend for all types of commodities. Flow patterns for trucks and railroads will have changed before and after the introduction of Atrucks (where truck flows are the sum of Atruck and Htruck flows). For domestic trade flows, rail trade patterns suggest a dramatic shift toward continuing various trans-continental rail flows, following the introduction of Atrucks, from an end-point concentration in central Colorado. Atrucks will probably replace some of the mid- or long-distance flows from central Colorado so that rail trade goes directly from west locations to east locations. Truck flows are predicted to lose many interactions between the western U.S. and Floridian and northeastern regions, but experience greater interactions among northwestern regions. Export trade flows are much lighter than domestic flows, in general, and their connections among southern, northwestern and northeastern U.S. regions appear enhanced by rail ties, following introduction of Atrucks, while trucks appear to lose overall trade flows between the nation’s northwest and southeast regions. Since great uncertainty still exists about the relative costs of acquiring and deploying Atrucks, multiple scenarios were tested, with different parameter assumptions. Atruck operating costs are expected to be much lower than Htruck costs, overall, thanks to a reduction in operator/attendant burden from the driving task and Atrucks’ greater utilization, as their attendants rest/sleep or perform other duties (and are not subject to strict hours of service regulations, since they cannot cause a fatal crash, for example). Wages and benefits may fall, or simply shift from administrative and service workers that used to be officed (e.g., those managing carrier logistics, customer service calls, or shipper billing) to workers that now travel between states on-board a moving office (and help with pickups and deliveries, as those arise). Interestingly, Atruck splits (either by dollar-miles carried or ton-miles transported) are very stable across these scenarios, at around 90 percent and 85 percent, regardless of the relative price variation. Conclusion Overall, this study is an initial attempt to reflect self-driving trucks in freight systems. Lower-cost trucking operations will impact choice of mode as well as input origins, affecting production and flow decisions for domestic and export trades. 20 commodity types are tracked using the 2012 CFS and FAF4 data sets. Sensitivity analysis allows for variations in predictions, given the great uncertainty that accompanies shippers’ future cost-assessments, adoption rates, and use of Atrucks. Such predictions should prove helpful to counties and regions, buyers and suppliers, investors and carriers, as they prepare for advanced automation in our transportation systems.

ABSTRACT. The increasing contribution of freight transport to global warming and congestion has spurred an interest in various freight transport policy options including a modal shift to less carbon intensive modes. In order to achieve this, there is a need to model and understand freight decision-makers (National Academy of Science, 2010, Samini et al, 2011; Windisch et al., 2010). Contrary to this importance, freight models have not been developed at the same pace as passenger ones (Hensher and Figliozzi, 2007). One of the most highlighted reasons for this is the difficulty of obtaining reliable disaggregated data, so most of the freight models addressing modal choice end up being aggregated models without a solid behavioural foundation (Ellion et al 2017; Pourabdollahi et al, 2012). Nevertheless, the contributions to the behavioural understanding of freight choices have increased in recent years (Chow et al, 2010; de Jong et al, 2004; de Jong et al 2012; de Jong and Ben-Akiva 2007; Fowkes, 2007; Samini et al, 2011; Tavasszy, 1998). Freight transport behaviour differs from passenger transport, not only in terms of the data availability, but also in terms of actors and product diversity and in the complexity of the decisions (Arumotayanun and Polak, 2009; de Jong et al 2012). In freight there are multiple decisions makers whose preferences and attitudes influence the choice making process. The identification and targeting of these key actors is crucial for successfully modelling the phenomena. Besides, the multiplicity of products also makes the behaviour of these agents heterogeneous. So far, the mode choice in freight transport has been dealt as a decision of a purely discrete nature, assuming that the alternatives are perfect substitutes. This may be true for passenger transport, but not necessarily for freight transport. There can be cases where multiple-discreteness can occur. Consider a shipper who has to send some amount of goods over a time period between a month and a year to a receiver. At a tactical level, where some of the planning and commercial decisions are made, there can be room for analysing and choosing multiple alternatives (not always the same mode). Different clients at different locations can be analysed, depending on transport aspects and even for the same location different modes can be chosen. The selection of multiple alternatives can be made in order to reduce risk, obtain smoother cash flows and to have a stable inventory. The possibility of splitting the total amount of goods to be sent in a month or a year over different modal alternatives can be interpreted as one of the particularities of the freight choice process. One decision-maker may have at the same time the choice alternative to ship a certain amount of tonnes in one big shipment or in several smaller ones, and depending on the context, by different routes or different destinations and modes. Discrete-continuous choice modelling had a big breakthrough after the development of the Multiple Discrete-continuous extreme value model (MDCEV) (Bhat, 2005; Bhat, 2008). The MDCEV framework is attractive because of its closed form and relative simple interpretation of the parameters (Pinjari and Bhat, 2011). Mixed and nested applications (Chalastri et al, 2017) of this framework have also been developed, together with forecasting algorithms (Pinjari and Bhat, 2010). There are some applications of a discrete-continuous framework in freight modelling. Some efforts have been made in modelling the joint decision of mode and shipment size (McFadden et al.,1985; Abdelwahab and Sargious, 1992; Holguin-Veras, 2002; de Jong and Ben-Akiva, 2007; Windisch et al. 2012; Johnson and de Jong, 2011), but not in the form of MDCEV models. An application of the MDCEV framework has been found at an urban level, where Khan and Machemehl (2017) modelled tour patterns of urban commercial vehicles. No application of the MDCEV model has been found for freight modelling at a regional level.

CONTEXT The Argentinian soy production, one of the country’s main agricultural products, is currently a modern and innovative sector. It has transformed by the introduction of technology, genetically modified seeds, fertilizers and new ways of managing the supply chain. Figure 1 shows the main configuration, together with their main actors and their interactions. Figura 1: Soy supply chain The soy’s supply chain can be divided into two. The receivers, such as exporters, industry or consumers and the senders, such as producers and consolidators. As seen in figure 1, producers have two options for selling their crops. They can sell to exporters or industry (directly for the large producers, though brokers for the other producers) or they can sell their products to a consolidator. Since 46% of the production is carried by a large number of producers (Regunaga, 2010), they are likely to use the latter option. Besides, some large producers might also need conditioning for their seeds before selling them and this is normally carried out by consolidators. This makes them a key agent in the decision of where and how the inland transport is carried out in Argentina. A previous study carried out by the authors (unpublished) applied to the south of Buenos Aires Province, in Argentina, found out that the main drivers for mode and port choice were: the price paid in the port, the transport cost, the distance to the port and the minimum headway the rail service could provide.

DATA The datasets that will be used consists of one RP data set from 2014 and an SP survey, which was completed recently. Depending on its compatibility, the SP data used in a previous study could be used as well. The SP data come from a survey carried out with the support of the Ministry of Transport of Argentina, orientated to commercial decision-makers in consolidator firms. 12 choice tasks with 4 alternatives each were presented to each respondent. The choice alternatives represent two unlabelled ports (one closer and other further away) with 2 labelled modes each (road and rail). The ports, although unlabelled, were the 2 main ports for each consolidator, according to their location. The variables tested, each with 3 levels, were price at the port, freight price, travel time, transport lead time, reliability and minimum shipment size. An efficient design was used for the creation of the scenarios using as seed values the ones found in the previous study made by the authors. The response variable was the percentage of available cargo they would allocate to each of the four alternatives, similarly as used by Brooks et al (2012).The available cargo was stated in the scenarios and varied only with the stated size of the company. It is expected that this method can capture small trade-offs between alternatives that are not strong enough to induce a shift to an alternative, but can lead to a change in the allocation between alternatives. In total, 54 valid questionnaires were collected from different regions of the country. Although this is a small number compared to surveys in passenger transport, it is an acceptable number of responses considering the context of freight data collection. The RP data consists of a consignment bill database where every grain vehicle movement is registered. This data has no linkage to the shipper, carrier nor receiver due to confidentiality issues. The information contained by this database consists of consignment bill number, transport mode, year of harvest, origin, destination, load weight and date of unloading. The database consists of 4,6726,09 records of interregional grain transport at a national level, where 2,932,686 are from places where rail and road are both present. As the records consist of vehicle movements and not of shipments, there is the need to consolidate them in order to analyse the choice context with more accuracy. This merging of records is made under the assumption that consecutive records with the same product, origin, destination and data consists on the same shipment. After the merging, 1,104,243 records remain. With this RP data we will not be able to investigate the multiple discreteness that might be present in this choice context because of the impossibility to link these shipments with others. Nevertheless, the MDCEV collapses into MNL if only one alternative is chosen it meaning that its framework might still be useful to scale the SP data to create a more comprehensive model (Bradley and Daly, 1994).

MODEL STRUCTURE Considering a non-negative vector x of consumption quantities of the k alternatives available, the utility function U(x) the consumer intends to maximise in the MDCEV model is given by the following expression (Bhat, 2008):

U(x)= ∑_(k=1)^K▒〖γ_k/α_k ψ_k {(x_k/γ_k +1)^(α_k )-1} 〗 (1)

There are 3 key parameters in the formulation according to Eq. (1) (Lu et al, 2016). The parameter ψ_k represents the baseline utility when the consumption of the alternative k is 0. The higher ψ_k is, the more likely is that that a positive quantity of the alternative is consumed. The role of γ_k is to allow the zero consumption for the good k. It works as a translation parameter, and thus explains part of the satiation. The α_k parameter represents the decrease in the marginal utility with the increasing consumption of alternative k. It is normally constrained to a negative value to represent the satiation effect, otherwise, it would imply positive marginal returns. Due to the similar characteristics that γ_k and α_k have, they cannot be estimated simultaneously so ether you fixate α_kfor a gamma-profile or you fixate γ_k for an alpha-profile. The MDCEV can adjust to different price levels of alternatives and a budget constraint. In this case, the budget would 100% of the cargo the consolidators had and the price would be the same for the four alternatives. Additionally, in the SP presented there was no outside good present. The outcomes for the MDCEV model will be compared to those of discrete choice models: Multinomial logit (MNL) model for discrete observations Multinomial logit (MNL) model for market shares for each observation (as in the aggregate logit model).

PROPOSED CONTENT OF THE PAPER The objectives of this paper are twofold. Firstly, to understand the behaviour of grain consolidators in Argentina using a disaggregate modelling approach. Secondly, to investigate the presence of multiple discreteness in freight modelling. To achieve this, an MDCEV model will be estimated and its performance will be compared against other models (discrete choice models) that do not necessarily include a discrete-continuous approach. The paper will be organized as follows: First a literature review to accommodate the current state of research in freight behavioural models. Second, an explanation of the models used. Third a description of the RP data and the SP experiment and data. Fourth, the model estimation results for MDCEV and discrete choice models. In fifth place, a comparison on the models and on the implications in the behaviour of grain consolidators in Argentina. Finally a discussion on the multiple-discreteness in freight transport and the adequacy of the MDCEV application.