Intrazonal or Interzonal?: Improving Intrazonal Travel Forecast in a Four-Step Travel Demand Model
Keunhyun Park, Utah State UniversityShow Abstract
Sadegh Sabouri, University of Utah
Torrey Lyons, University of Utah
Guang Tian, University of New Orleans
Chad Worthen, Wasatch Front Regional Council
Reid H. Ewing, University of Utah
Conventional four-step travel demand models, used by virtually all metropolitan planning organizations (MPOs), state departments of transportation, and local planning agencies, are the basis for long-range transportation planning in the United States. Trip distribution – whether the trip is intrazonal (internal) or interzonal (external) – is one of the essential steps in travel demand forecasting. However, the current intrazonal forecasts based on a gravity model involve flawed assumptions, primarily due to a lack of consideration of differences in land use and street network patterns. In this study, we first survey 25 MPOs about how they model intrazonal travel and find the state of the practice to be dominated by the gravity model. Using travel data from 31 diverse regions in the U.S., we then develop an approach to enhance the conventional model by including more built environment D variables and by using multilevel logistic regression. The models’ predictive capability is confirmed using k-fold cross-validation. The study results provide practical implications for state and local planning and transportation agencies with better accuracy and generalizability.
Deriving and Assessing Activity Hierarchies from Relative Distances in Tours
Florian Schneider, Delft University of TechnologyShow Abstract
Winnie Daamen, Delft University of Technology
Sascha Hoogendoorn-Lanser, KiM Netherlands Institute for Transport Policy Analysis
Serge Hoogendoorn, Delft University of Technology
Knowledge on hierarchies between activities is a prerequisite to understand and predict tour formation. Due to data constraints, fixed hierarchy schemes between activity types are commonly used despite their imprecision in many cases. In this paper, we propose a new method to derive hierarchies between activity types in tours and assess their validity using longitudinal survey data. Assuming that activities for which people travel further are more important and thus have a higher priority in the tour formation, we calculated relative distances of activities in home-based tours that include two different out-of-home activity types. The resulting distributions of relative distances are analyzed for all available combinations of activity types regarding prevailing hierarchy patterns. The outcomes are put into perspective by comparing them to activity durations, which are used in literature as an indicator for an activity’s planning horizon. Subsequently, the method is exemplarily applied to investigate the effect of active modes compared to motorized travelling on activity hierarchies. The results indicate that priorities between pairs of activity types could successfully be derived using distance positions. Both the hierarchies as well as their strengths are in line with intuitive expectations. Moreover, most priorities were confirmed by the outcomes of the activity duration analysis or, conversely, raised interesting questions regarding current hierarchy schemes.
A Three-Level Model for Updating the Urban O-D Demand Matrix Using Vehicle Counts and Vehicle Identification Sensor Data
Mojtaba Rostami Nasab, Sharif University of TechnologyShow Abstract
Yousef Shafahi, Sharif University of Technology
Updating the urban origin-destination demand matrix using data from different sources in networks is essential in transportation planning, traffic management, and operations. This study presented a three-level model to update the origin-destination matrix of large-scale networks using combined data from vehicle counts, such as loop detectors on links as well as automated vehicle identification sensors, such as license plate recognition cameras on partial paths. The first level minimized the distance between the estimated and observed flows of partial paths, whereas the second level minimized the distance between vehicle counts and estimated flows of links. The lowest level found user equilibrium flow patterns for demand tables. To solve the model, an innovative method was offered. The basic ideas of the method included expanding a primary origin-destination matrix with a travel growth factor obtained from comparing estimated and observed link flows, and then adjusting path flows based on percentage differences in the estimated versus observed link or partial path flows. The estimated origin-destination matrices were reassigned under a user equilibrium principle, and the path flows were readjusted. The model was tested to estimate the origin-destination matrix of the large-scale network of Mashhad City with 2 million population, 163 traffic zones, and 2093 links. The estimation quality was evaluated using different error indicator measures. In conclusion, the results showed a significant accuracy to reconstruct the observed flows.
Modeling the Dynamics Between Tour-Based Mode Choices and Tour-Timing Choices in Daily Activity Scheduling
Khandker Nurul Habib, University of TorontoShow Abstract
MD Sami Hasnine, University of Toronto
The paper presents a dynamic discrete-continuous modelling approach to capture individuals’ tour-based mode choices and continuous time expenditure choices tradeoffs in a 24-hour time frame. The analysis of traditional activity-based models (ABM) are typically limited to activity-type, location and time expenditure choices. Besides, mode choice is often simplified to fit in a pre-defined activity schedule. However, decisions of tour departure time, tour mode choice and time expenditure choice for out-of-home activities are intricately inter-related, and common unobserved attributes influence these choices. This paper proposes a random utility maximization (RUM) based dynamic discrete-continuous model for joint tour based mode and tour timing choices. Tour timing choice is modelled as continuous time allocation/consumption choice under 24-hour time-budget. For the tour-based mode choice component, it uses a classical dynamic discrete choice modelling approach. A cross-sectional household travel survey dataset collected in the Greater Toronto and Hamilton Area (GTHA) in 2016 is employed for the empirical investigation in this study. Empirical model shows the capability of handling all possible mode combinations within a tour including ride-hailing services (e.g., Uber, Lyft). Empirical results reveal that individuals variations in time expenditure choice are defined by activity type, employment status, and vehicle ownership. In terms of mode choice, it is clear the emerging transportation service users have different travel pattern than conventional mode users. This practice-ready modelling framework has the potential to test a wide range of travel demand management (TDM) strategies.
Mode Choice Modeling for Hailable Rides: Investigating the Competition of Uber Using a Non-Compensatory Choice Model with Probabilistic Choice Set
Khandker Nurul Habib, University of TorontoShow Abstract
The paper presents an empirical investigation on the demand for TNC services, e.g. Uber in the Greater Toronto and Hamilton Areas (GTHA). It used a dataset of trip mode choices that suitable to be made by any ride-hailing service (e.g Uber). Such trips are named as hailable trips the those were drawn from a large scale household travel survey conducted in the region in the Fall of 2016. In order to have a clear understanding of behavioural processes involved in the choice of travel mode of hailable trips, a new choice model is proposed that jointly models probabilistic choice set formation and conditional semi-compensatory choice. The empirical model did not reveal any clear competition between Uber with private car, public transit, and non-motorized modes. It indicated that urban taxi was its main competitor, but there are notable differences in socio-demographic profiles of taxi and Uber users. For example, Taxi is preferred by older people, but Uber is preferred by younger people and there is no gender difference in such a pattern. In terms of the relationship between considering Uber as a feasible mode and actually choosing it for a trip, Uber sits next to car passenger mode. In such case just accepting it as a feasible option has a large influence on making a final choice of using it. This indicates a potential new segment of the travel market, generated specially for the advent of TNC service, e.g. Uber in Toronto
An Environment–People Interactions Framework for Analyzing Children’s Extra-Curricular Activities and Active Transport
Kevin Y.K. Leung, University of Hong KongShow Abstract
Sebastian Astroza, Universidad de Concepcion
Becky P.Y. Loo, University of Hong Kong
Chandra Bhat, University of Texas, Austin
This paper examines children’s extra-curricular activities in a high density urban environment. The paper offers a framework to understand children’s extra-curricular activities time allocation and active travel participation. Three variables of interest are considered: residential location choice (based on residential density), weekly time spent in four different types of out-of-home after-school activities (academic, sports, arts, and hobbies), and level of active travel. The proposed model takes into account common observed and unobserved effects which may affect all three outcomes simultaneously. The survey data, collected at four Hong Kong primary schools, show that children’s activity and travel behavior within the same city can differ quite substantially based on neighborhood environment (notably residential density) and family socio-demographic background. The empirical findings and analysis provide insights for policy development.
A Semi-Compensatory Choice Model with Probabilistic Choice Set: Combining Implicit Choice Set Within Probabilistic Choice Set Formation
Zohreh Nurul Rashedi, University of TorontoShow Abstract
Khandker Nurul Habib, University of Toronto
The standard random utility model assumes fully rational and compensatory choice behaviour. However, various research studies have proven that non-compensatory /semi-compensatory choice behaviour is more realistic. This paper proposes a semi-compensatory framework for the discrete choice model that combines probabilistic choice set formation along with implicit choice constraints in choice making. It combines the Independent Availability Logit (IAL) with Implicit Constrained Multinomial Logit model (CMNL) to improve the performance of both approaches. The model is applied to investigate mode choice behaviour of the Ottawa-Gatineau regions of Canada’s capital by using the household travel survey data. The explanatory power and elasticity measures of the proposed model are compared with IAL and MNL models. It is found that IAL-CMNL model outperforms both IAL and MNL models and can reproduce choice set formation process more effectively. The empirical investigation shows that IAL-CMNL model results in relatively higher tolerance and softer constraints for cut-off violations compared to the IAL model. Elasticity calculations and outcomes of this research highlight the importance of capturing choice set formation and constrained choice behaviour in the choice modelling.
A Copula-Based Approach to Accommodate Intra-Household Interaction in Activity Stop Generation Modeling
You-Lian Chu, ParsonsShow Abstract
A new stop frequency model was developed to predict the daily number of maintenance activity stops made by individual household heads during a typical weekday. This new model was based on the modification of a multivariate ordered probit (MOP) model by maintaining the probit assumption for the marginal distributions while introducing non-normal dependence among the error terms using copula functions. The copula-based MOP model would not only account for the intra-household interactions in stop-making decisions, but also allow the best functional form to be determined for representing dependencies among household heads; therefore, the new MOP model would relieve the restriction of imposing joint normality on the error terms in the conventional MOP model. Using the New York Metropolitan Transportation Council’s 2010/11 Regional Household Travel Survey data, the copula-based MOP model was employed to examine stop-making behavior for individual household heads residing in New York City and its adjacent counties in Mid-Hudson Valley and New Jersey. Empirical results provide useful insights into the observed effects of socio-demographics, land use density, transportation service, and work schedule together with potential unobserved common effects on the inter-relatedness of spousal stop-making decisions at the household level.
Suitability of Cellular Network Signaling Data for Origin–Destination Matrix Construction: A Case Study of Lyon Region (France)
Mariem Fekih, Hasselt UniversityShow Abstract
Tom Bellemans, Hasselt University
Zbigniew Smoreda, Orange Labs, France
Patrick Bonnel, LET-ENTPE
Angelo Furno, Universite de Lyon
Stéphane Galland, Université de Bourgogne Franche-Comté
Spatiotemporal data, and more specifically origin-destination matrices, are critical inputs to mobility studies for transportation planning and urban management purposes. In this paper, we propose a methodology to infer origin-destination (O-D) matrices based on passively-collected cellular signaling data of millions of anonymized mobile phone users in the Rhône-Alpes region, France. This dataset, which consists of records time-stamped with users’ unique identifier and tower locations, is used to first analyze the cell phone activity degree indicators of each user in order to qualify the mobility information involved in these records. These indicators serve as filtering criteria to identify users whose device transactions are sufficiently distributed over the analyzed period to allow studying their mobility. Trips are then extracted from the spatiotemporal traces of users for whom the home location could be detected. Trips have been derived based on a minimum stationary time assumption that enables to determine activity (stop) zones for each user. As a large, but still partial, fraction of the population is observed, scaling is required to obtain an O-D matrix for the full population. We propose a method to perform this scaling and we show that signaling data-based O-D matrix carries similar estimations as those that can be obtained via travel surveys.
Statistical Properties and Prediction Models of Regional Highway Traffic Distribution
Wei Gao, Northeast Forestry UniversityShow Abstract
Baoyu Hu, Northeast Forestry University
Traffic distribution data are essential basic data for highway network planning, but they are difficult and expensive to obtain. In this study, the statistical properties and prediction models of regional highway traffic distribution are investigated by a practical case of regional highway traffic distribution in Heilongjiang Province, China. We build a topologic graph of the highway traffic distribution and present the degree distribution of the topologic graph and the probability distribution of traffic flow between traffic zones. The two distributions show a high level of heterogeneity and follow Zipf’s law. We also present the distributions of production and attraction of traffic zones. The results show that production and attraction exhibit a positive linear correlation and a high level of heterogeneity, approximately satisfying a power-law decay. We also find that the traffic and the topology are strongly correlated. We present three existing models (gravity model, radiation model, and population-weighted opportunities model) and a proposed multifactor-weighted benefits model considering population, GDP, and area (named as MWB model) for the highway traffic distribution. A comparative study shows that the MWB model is more suitable for predicting the highway traffic distribution than the other three models whether predicting accuracy, cost, or efficiency.
Route Choice Behavior Analysis: An Agent-Based Simulation Perspective
Ali Shamshiripour, University of Illinois, ChicagoShow Abstract
Amir Parsa, University of Illinois, Chicago
homa taghipour, University of Illinois, Chicago
Ahmad Nafakh, University of Illinois, Chicago
Abolfazl (Kouros) Mohammadian, University of Illinois, Chicago
Conventional traffic assignment models assume that people decide on their route merely based upon the expected travel time on the route. In reality, though, diverse drivers would rank a given route differently, based on their preferences towards a range of route characteristics. Instances of factors contributing to this diversity include avoiding tolls, avoiding neighborhoods with high rates of crime, avoiding unsafe roads, lowering fuel costs, willing to drive on scenic roadways, willing to keep the routines, etc. A common practice to account for the variety of route characteristics is to construct generalized cost functions as a combination of various route characteristics. Yet, the literature argues against the use of constant route-characteristic weights in constituting the generalized cost function, suggesting considerable inter-personal heterogeneities in ranking different routes. The current study presents results of a detailed analysis aiming at understanding the major factors affecting travelers’ preferences towards such factors. Furthermore, this study takes into account several contextual variables such as weather conditions, crime prevalence, and the scenic routes. In addition to the route and network attributes, we also found the following variables to be strongly influential: weather characteristics, land use characteristics, and activity characteristics.
Disaggregate Short-Term Location Prediction Based on Recurrent Neural Network and an Agent-Based Platform
Yanjie Dong, Imperial College LondonShow Abstract
John W. Polak, Imperial College London
Aruna Sivakumar, Imperial College London
Fangce Guo, Imperial College London
With growing popularity of mobile and sensory devices, there has been a strong research interest in short term disaggregate level location prediction. Such predictive models have huge application potentials in several sectors to change and improve people’s daily life and experience. Existing methods in this research stream have mainly focused on the sequence of location prediction, with the valuable temporal information overlooked. In addition, data limitations have constrained the development and understanding from different algorithms. In this paper, we have proposed a recurrent neural network-based method (RNN and LSTM, Long Short-Term Memory) for the next and future location prediction. Out model is predicting at the sequence of time hence it can predict both when and where an individual will be in future and duration of the stay at each location. The predictive model is developed based on an agent-based simulation platform which is capable of producing realistic spatial-temporal trajectory data at the individual level. Analysis of the simulated data has shown that RNN and LSTM are capable of predicting future locations with better results than other comparative methods, especially for the agents with high location variability. Online prediction with true location information fed into the model later of the day would greatly improve the predicted results. However, significant variations can be observed at the zonal level, with all methods perform much better on frequently visited locations than less visited locations or irregular visits.
A Framework for Estimating Bikeshare Origin–Destination Flows Using a Multiple Discrete Continuous System
Bibhas Dey, University of Central FloridaShow Abstract
Sabreena Anowar, University of Central Florida
Naveen Eluru, University of Central Florida
This current study identifies two choice dimensions for capturing the bike share system demand: (1) station level demand and (2) how bike flows from an origin station are distributed across the network. A linear mixed model is considered to estimate station level demand while a multiple discrete continuous extreme value (MDCEV) model to analyze flows distribution is employed. The data for our analysis is drawn from New York City bikeshare system (CitiBike) for six months from January through June, 2017. For our analysis, we examine demand and distribution patterns on a weekly basis. A host of exogenous variables including trip attributes, socio-demographic attributes, bicycle infrastructure attributes, land use and built environment, temporal and weather attributes are considered. The model estimation results offer very intuitive results for origin demand and multiple discrete destination choice models. We validated the model by predicting trips to destined stations and found that predicted model performs well for high demand destinations. This analysis will allow bike sharing system planners and operators to better evaluate and improve bikeshare systems.
Tracing the Impact of Macro-Economic Indicators in Micro-Level Households’ Preferences
Milad Ghasrikhouzani, UNSW SydneyShow Abstract
Taha Rashidi, University of New South Wales
John Rose, University of Technology Sydney
This study investigates whether macro-level economic indicators have an impact on the micro-level households’ preferences towards their budget allocation. The hypothesis of this study is that major changes in the macro-economy conditions can cause variations in the way households allocate their budget to different activities (expenditure categories). Typically, households’ tastes are assumed to be unchanged. Under this assumption, changes in the economy can only be reflected in the model input (changes in the explanatory variables). However, this study hypothesizes that in addition to this impact, changes in the economy may change households’ behaviors as well. To investigate this hypothesis, five separate multiple discrete-continuous extreme value models for household expenditure on different commodities are developed using the data of more than 6000 households from 2006 to 2010. The parameters of the models in 2006, 2007, 2009 and 2010 are then statistically tested against the parameters of the model in 2008. This is because the oil market experienced a significant disruption in 2008, where the oil price dropped from $130 to $40 per barrel in six months. According to the results, the estimated coefficients for most of the variables are statistically different in 2008 which suggests that the macro- economic conditions have a significant impact on how people prefer to allocate their budget to different activities.
Direct Demand Models for Non-Motorized Modes: A Comparison of Log-Linear and Generalized Linear Modeling Frameworks
Aditya Medury, Safe Transportation Research and Education CenterShow Abstract
Robert Schneider, University of Wisconsin, Milwaukee
Julia Griswold, University of California, Berkeley
Offer Grembek, University of California, Berkeley
Many non-motorized direct demand models utilize log-linear models, which apply a log-transformed dependent variable within a linear regression framework, to model pedestrian/bicycle counts. The log-transformation ensures that count estimates derived from these models are non-negative, while also accounting for some skewness in the distribution of raw counts. The literature also provides other modeling alternatives for this purpose, especially generalized linear models (GLM), such as negative binomial and Poisson regression models. These models can also address the issues of skewness and non-negative values without requiring a similar transformation of the dependent variable. In this paper, we explore the performance of log-linear and other GLM models using both empirical and simulated datasets, evaluated using comparisons of prediction accuracy and estimated coefficients. We also identify a critical estimation bias that occurs while back-transforming the predictions from a log-linear model, which as commonly applied in direct demand modeling literature, does not correspond to the mean, but rather the median of the underlying lognormal distribution. However, the presence of heteroscedasticity reduces the accuracy of log-linear models overall in comparison to GLM-based methods. Through the use of varied datasets, we provide practitioners with insights into the pros and cons of different model specifications for non-motorized direct demand models.
Factors Affecting the Departure Train Station Choice of Bicyclist in the Western Region of the Netherlands
Manoj Ashvin Jayaraj, Centrum Wiskunde & InformaticaShow Abstract
Jullian van Kampen, Centrum Wiskunde & Informatica
Thomas Koch, Centrum Wiskunde & Informatica
Eric Pauwels, Centrum Wiskunde & Informatica
Rob van der Mei, CWI
Elenna Dugundji, Vrije Universiteit, Amsterdam
With 35,000 km of bicycle pathways, cycling is common among persons of all ages less than 65 years in the Netherlands. Bicycle is often seen as a standalone travel mode but when integrated as part of a multimodal trip with train can be an important solution for long distance journeys, often offering more flexibility and faster access time than other travel modes to a railway station. With 47% of people not preferring to depart from the nearest train station, we investigate in this paper which factors influence departure station choice on combined bicycle-train trips in the western region of the Netherlands, exploring effects of individual socio-economic characteristics of residents, their multi-modal trip attributes, as well as neighborhood characteristics and station attributes. Observations from the Dutch National Travel Survey over years 2015-2017 where a train ride precedes a bicycle ride, or train ride is followed by a bicycle ride are considered for analysis. The choice set consists of four alternatives: the first alternative corresponds to the first closest train station from each departure postcode, and the second, third and fourth alternatives accordingly correspond to the second, third and the fourth closest stations. Estimated results show that the distance to all four stations is negatively significant for departure station choice, whereas in-train travel time, expendable household income and household composition in different categories were not even significant at 10% level. Working hours per week, residential municipality outside a city region and sprinter station type are found to be positively significant.
On the Usefulness of a Combined Mode Choice-Schedule Choice Model: The Case of the Paris-Bordeaux Rail Line (France)
Minghui Chen, Transport, Urban Planning and Economics LaboratoryShow Abstract
Stéphanie Souche Le Corvec, Universite de Lyon
The high-speed rail line (HSR) Ligne à Grande Vitesse Sud Europe Atlantique (LGV SEA) was inaugurated and put into operation on July 2, 2017. Since, we observed a drop in air traffic and a decrease in air service frequency on the Paris-Bordeaux line. In this article, we are interested in the competition between HSR and air transportation services and the influence of this new transport infrastructure on passenger behavior. Using discrete choice models along with data from traveler surveys, we conduct an econometric analysis of traveler demand dealing jointly with mode choice and schedule choice between Paris and Bordeaux. We illustrate that the variables specifically constructed to represent the schedule delay cost are significant, with late arrival generating relatively greater costs compared to the early arrival. This model also allows us to evaluate the quality of transport timetable proposed by the transportation operators with the help of a market share prediction.
Does Security of Neighborhoods Affect Non-Mandatory Trips?: A Copula-Based Joint Multi-Nomial-Ordinal Model of Mode and Trip Distance Choices
Amir Parsa, University of Illinois, ChicagoShow Abstract
Kimia Kamal, Sharif University of Technology
homa taghipour, University of Illinois, Chicago
Abolfazl (Kouros) Mohammadian, University of Illinois, Chicago
The impact of security indicators on travel behavior has been recently investigated in the literature, and researchers more or less found the crime rate of different component of trips, origin, destination, and route, influential on mode choice. This paper, however, presents a copula-based joint multinomial-ordinal model to investigate the impact of security of home location as well as other attributes on joint decision of choosing travel mode and distance. Non-mandatory home-based trips, with home in origins, are selected for this study since this type of trips gives travel makers freedom to choose among different distances and modes for a given travel purpose. The fusion of travel information, weather condition, land use, Google Maps data as well as crime density data sources are used as the primary data for developing the model. Furthermore, employing copula-based approach leads to the model in which unobserved factors affecting both mode and distance choices are considered jointly. The results suggest that crime density of home location neighborhood significantly impacts choosing active modes and travel distance in such a way that increasing crime density of trip origin reduces propensity of choosing active modes and short distances. Statistical significance of copula parameters also confirms that there is intercorrelation between mode and distance choices as a joint decision. Assessing five policy making scenarios reveals that decreasing crime density and increasing percentage of residential, shopping, entertainment, and recreational land use areas can increase propensity of choosing active modes and short distances in non-mandatory trips.
An Integrated Choice and Latent Variable Framework to Incorporate the Influence of Travel Time Variability on Truck Route Choice
Mehek Biswas, Indian Institute of Science (IISc)Show Abstract
Abdul Pinjari, Indian Institute of Science
Subodh Dubey, Tiger Analytics, India
This study proposes a multinomial probit (MNP)-based Integrated Choice and Latent Variable (ICLV) modelling framework that allows simultaneous estimation of route-level travel time variability and incorporation of the influence of such variability on route choice. The travel time on a route is considered as a latent variable and GPS data measurements of route-level travel time are used to identify the parameters of its statistical distribution. Since such measurements are not always available for all routes, the latent variable component of the ICLV framework helps inform the travel time distribution for routes without travel time measurements. The proposed model is applied to an empirical data set on truck route choice using GPS data from Tampa, Florida. The empirical models account for the variability of travel time on a route as a function of the network structure along the route, such as the lengths of different roadway types. The route choice component of the proposed ICLV model demonstrates a superior statistical fit and better predictive ability than traditional choice models that do not consider the influence of travel time variability on route choice – both on the estimation dataset and a validation dataset. A reduced form mixed probit model was found to have a superior data fit than the route choice component of the ICLV model, albeit it fails to offer the insights that the proposed ICLV model offers on travel time variability its influence on route choice.
Household Trip Generation and the Built Environment: Does More Density Mean More Trips?
Qin Zhang, Technical University of MunichShow Abstract
Kelly Clifton, Portland State University
Rolf Moeckel, Technical University of Munich
Jaime Orrego Onate, Portland State University
Trip generation is the first step in the traditional four-step trip-based transportation model and an important transport outcome used in evaluating the impacts of new development. There has been a long debate on the association between trip generation and the built environment with mixed results. This paper contributes to this debate and approaches the problem with two hypotheses: 1) built environment variables have significant impacts on household total trip generation; and 2) built environment variables have different impacts on trip generation by purpose. This study relied on data from the Portland, OR metropolitan area to estimate negative binomial regression models of household trip generation rates across all modes. Results show that the built environment does have significant and positive influences on trip generation, especially for total number of trips, total number of tours, and home-based shopping related trips. Moreover, loglikelihood ratio tests implied that adding built environment to the base model contributed significantly to improving model explanatory and predictability. These findings suggest that transportation demand models should be more sensitive to the effects of the built environment to better reflect the variations in trip making across regions.
Understanding the Relationships Between Demand for Shared Ride Modes: A Case Study Using Open Data from New York City
Raymond Gerte, NoblisShow Abstract
Karthik Konduri, Amazon, Inc.
Nalini Ravishanker, University of Connecticut
Amit Ranjan Mondal, University of Connecticut
Naveen Eluru, University of Central Florida
Karthik Konduri, University of Connecticut
The concept of shared travel, or 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 modes in the form of transportation network companies (TNC), to substantially impact travel behavior. The current body of research on how these modal offerings impact the demand for existing shared modes (e.g. bikeshare, transit) is growing. However, a comprehensive investigation of the temporal evolution of TNC demand and its relationship to other shared modes, is lacking. This research tackles this important limitation by analyzing ridership data for TNC, Taxi, Subway and Citi Bike in New York City using daily ridership data from January 2015 through June 2017. The primary objective was to understand the relationship between TNC and other shared modal offerings while accounting for the influence of temporal trends and other exogenous factors. A Dynamic Linear Modeling (DLM) framework was formulated to accommodate time dependent trends, periodicity, and time varying exogenous factors on the demand for TNC. As a preliminary work, the findings of this study reinforce the observed substitution relationship between Taxis and TNC. The results may also indicate a substitutional relationship between TNC and Citi Bike, and a complementary relationship with Subway, however these results still need to be explored further. With potentially impactful findings for planning and policymakers, the predictive model developed in the study can be used to carry out forecasting in support of short- and long-term operations and planning applications.
Inferring Latent Activity Patterns from Household Travel Activity Diaries
Roger Chen, Rochester Institute of Technology (RIT)Show Abstract
In this paper, we investigate the use of Latent Dirichlet Allocation (LDA) towards identifying latent activity patterns with activity pegs from conventional household travel surveys. The activity-based approach to travel analysis requires forecasting individual activity patterns under future technological and management scenarios, relying on travel-activity diaries as the primary data source. One perspective on the scheduling process considers two stages, where travelers start from a “skeletal” pattern with activity pegs, such as mandatory work activities, and then fill in the remaining time budget with discretionary activities, such as spurious unplanned social activities. We investigate the application of LDA to travel-activity schedules from conventional household travel surveys and its ability to identify activity pegs. Our results indicate eight distinct latent activity patterns that differed primarily in the activities occurring in the late afternoons and early evenings. The results suggest that pre-processing of travel-activity schedule data may help with pattern identification, analogous to removing “stop-words” from text (i.e. and, but, or, the, etc.) that have low informational content. Additionally, based on the mixing distribution of latent patterns from our LDA, we conduct an experiment to recovery travelers’ personal and household characteristics, given observed travel-activity schedules. The results indicate that work related characteristics, such as employment status, have a high accuracy (about 70% correct classification) when predicted by conventional classifiers. Other person and household attributes also did better than a purely random (coin-flip or dice-rolling) classification; however, only attributes with strong associations with activity scheduling and time budgets produced high accuracy in recovery.
Modeling Departure Time Choices Using Mobile Phone Data
Andrew Bwamable, University of LeedsShow Abstract
Charisma Farheen Choudhury, University of Leeds
Stephane Hess, University of Leeds
The rapid growth in passive mobility tracking technologies has led to a departure time choice studies based on GPS data in recent years. GPS data is however still expensive to collect and affected by technical issues like signal losses and battery depletion which create gaps in the data. On the other hand, the rapid growth in mobile phone penetration rates has led to the emergence of alternative passive mobility datasets such as Global System for Mobile communication (GSM) data, which covers wider population and can be used to derive departure time information. This motivates this research where we rigorously compare the strengths and weaknesses of real-world GSM and GPS data to investigate their potential use for modelling departure time choice. We describe practical approaches to extract relevant information from the passive datasets and propose a modelling framework that accounts for the fact that the desired departure times are unobserved. We assume that the preferred departure times randomly vary across the users and apply the mixed logit framework to jointly estimate the distribution parameters of the preferred departure times and the sensitivities to schedule delay. We find that fewer time gaps in the GSM data lead to more reliable model results when compared against those based on GPS data, despite the higher location accuracy of the latter. This is also supported by the comparison of the valuation metrics derived from both models, where those obtained from GSM data are found to be closer to those based on traditional data sources.
A Neighborhood-Based Collaborative Filtering Algorithm for Secondary Activity Location Choice Prediction Using Smart Card Data
Yihong Wang, Delft University of TechnologyShow Abstract
Goncalo Correia, Delft University of Technology
Bart van Arem, Delft University of Technology
Harry Timmermans, Eindhoven University of Technology
Collaborative filtering is a method of predicting the interests of a single person by collecting preference information from many people. Collaborative filtering algorithms have commonly been used to predict the preference of a consumer for a movie or a song in a recommendation system. This data-driven approach only relies on empirical observations and does not require imposing theory-based prior assumptions about behavior, resulting in a more flexible way to capture preferences and potentially a better prediction. In addition, one of the assumptions underlying travel behavior modeling is that different personal attributes (e.g., socioeconomic status) cause the heterogeneity of travel preferences, which is always difficult to model using big data due to the anonymity. Collaborative filtering seems promising for tackling this issue. This work specifically focuses on the problem of predicting one’s secondary activity location (other than work or living). A tailored collaborative filtering algorithm is applied to the three-month metro smart card data from Shanghai, China. Results show that the collaborative filtering algorithm outperforms the other prediction methods, including an estimated multinomial logit model, which shows the relevance of exploring further such method.
Enriching Travel Demand Forecasting Models with a Household Typology
Léa Fabre, Ecole Polytechnique de MontrealShow Abstract
Catherine Morency, Ecole Polytechnique de Montreal
In the Québec province, in Canada, travel demand forecasting relies mainly on individual characteristics, whereas it appears that individual mobility behaviors also significantly depend on the attributes of relatives and people who live in the same household. This paper aims to better understand the interactions between the household structure and individual mobility behavior, as well as the effects of age and life cycle of individuals and their relatives on individual trips. Considering that personal mobility is dependant on household attributes, a household typology is proposed to be included in the travel demand forecasting approach which is initiated with large-scale household travel surveys. An allocation model of the people to a household type completes the process, so that in the future, transport demand could be predicted considering the person herself as well as her household type.
A Latent Class Joint Mode and Departure Time Choice Model for the Greater Toronto and Hamilton Area
Sanjana Hossain, University of TorontoShow Abstract
MD Sami Hasnine, University of Toronto
Khandker Nurul Habib, University of Toronto
This paper presents a closed-form Latent Class Model (LCM) of joint mode and departure time choices. The proposed LCM offers compound substitution patterns between the two choices. The class-specific choice models are of two opposing nesting structures, each of which provides expected maximum utility feedback to the corresponding class membership choice. Such feedback allows switching class membership in response to the changes in choice contexts. The model is used for an empirical investigation of commuting mode and departure time choices in the Greater Toronto and Hamilton Area (GTHA) by using a large sample household travel survey dataset. The empirical model reveals that overall 38 percent of the commuters in the GTHA are more likely to switch modes than departure times and 62 percent of them are more likely to do the reverse. The empirical model also reveals that the average Subjective Value of Travel Time Savings (SVTTS) of the commuters in the GTHA can be as low as 3 dollars if a single choice pattern of departure time choices nested within mode choices is considered. It can also be as high as 67 dollars if the opposite nesting structure is assumed. However, the LCM estimates the average SVTTS to be around 27 dollars in the GTHA. An empirical scenario analysis by using the estimated model indicates that a 50 percent increase in morning peak period car travel time does not sway more than 4 percent of commuters from the morning peak period.
Examining Potential Correlation Among Departure Time, Destination, and Transportation Mode for Discretionary Trips: A Case Study of Shopping and Entertainment Trips
Mahmoud Elmorssy, Istanbul Technical UniversityShow Abstract
H. Onur Tezcan, Istanbul Technical University
This paper aims to examine the inter-relationship (dependency) between departure time, destination and transportation mode alternatives of discretionary trips under a single unifying framework. This can be achieved through developing a number of two-level Nested Logit models that can incorporate different correlation structures. The proposed models have been estimated and tested by using shopping and entertainment trips’ data obtained from the 2015 household survey that was conducted in Eskisehir city, Turkey. In order to reach best models in terms of log likelihood value at convergence, overall goodness of fit and other tests related to significance and signs of coefficient estimates, some specifications which restrict parameters of utility function’s variables with one or more travel dimension have to be assumed and examined. In the light of estimation results, individuals are more likely to jointly decide on “at which departure time”, “to which destination” and “by which mode” rather than separately. Moreover, neglecting the potential correlation among alternatives of departure time, destination and mode leads to inaccurate estimates which results finally in incorrect and improper policy decisions.
A Comparative Study of Methods for Capturing Spatial Correlations in Location Choice Through an Empirical Application on School Location Modeling
Adam Weiss, University of TorontoShow Abstract
MD Sami Hasnine, University of Toronto
Khandker Nurul Habib, University of Toronto
This paper presents a comparison of methods for capturing spatial correlation between location alternatives. Three different methods are compared and analyzed based on their statistical performances. To test their performances, the choice of school location for high school students is examined. This analysis shows that the Spatially Weighted Error Correlation (SWEC) Model outperforms the alternative approaches (spatial autocorrelation and GEV models) with respect to its ability to capture the intricacies of spatial choices. The reasons for this are twofold. First, the model can capture patterns of spatial autocorrelation between alternatives. Secondly, the model captures spatial heteroscedasticity in decision making based on the relative distance of an alternative from the location of the choice maker. This is a substantial improvement over the existing models commonly used models within the literature.
Interlinkage Between Trip Chaining and Mode Choice
Sijia Wang, WSPShow Abstract
Gaurav Vyas, WSP
Peter Vovsha, WSP
Trip destination and mode represent the most fundamental dimensions of travel. Even with a simple trip-based model, sequencing the choice of destination and mode is a non-trivial issue that gives rise to two different model structures. For advanced Activity-Based tour-based Models (ABMs) the whole discussion on the causal relationship between trip destination and mode becomes a more complicated one on how trip chaining and corresponding mode combinations affect each other. The main intention of the current research is to statistically explore various two-way linkages between trip chaining patterns and mode combinations with the ultimate purpose of incorporating these factors in an operational travel model. Joint distributions of tour structure and tour mode combinations were explored with the 2010 Regional Household Travel Survey (RHTS) in the New York Metropolitan Region. The current research contributes to the conceptual side of trip chaining and tour mode modeling by recognizing the special role of joint activities and travel (such as escorting) and a comprehensive analysis and ranking of two-way causal linkages between trip chaining patterns and tour mode combinations. Advanced machine learning methods were employed for this purpose. The current research contributes to the operational modeling side of trip chaining and tour mode by exploring the details of tour formation and combinatorial mode choice models and their possible integration through feedback. To illustrate this, a logit choice model was estimated for tour mode choice combination as a function of tour structure and other variables.
Could a New Mode Alternative Modify Psycho-Attitudinal Factors and Travel Behavior?
Eleonora Sottile, University of CagliariShow Abstract
Francesco Piras, University of Cagliari
Italo Meloni, University of Cagliari
There is ample consensus that, besides objective characteristics, psycho-attitudinal factors play a key role in influencing people’s mode choice. Hybrid choice models use these theoretical frameworks so as to include latent constructs for capturing the impact of subjective factors on mode choice. But recent work in transportation research raised the question about the ability of hybrid choice models to derive policy implications that aim to change travel behaviour, given the focus on cross-sectional data. To address this problem we designed a survey for collecting longitudinal data (socioeconomic and psycho-attitudinal) so as to evaluate, on the one hand, the long term effects on travel mode choice of the implementation of a new light rail line in the metropolitan area of Cagliari (Italy), on the other to detect any changes in the psycho-attitudinal factors and/or in socio-economic characteristics after implementation of those measures. In particular, the objective of the study is to analyse whether these hypothetical changes in individual characteristics are able to affect mode choice from a modelling perspective, through the specification and estimation of hybrid models. Our results show that latent variables were not significantly different over waves, showing that the impact of the psychological construct remained stable over time, even after the introduction of the new light rail. Additionally, we found some evidence that the variables that explain the latent variables could change over time.
Spatial Modeling of Origin–Destination Commuting Flows in Switzerland
Thomas Schatzmann, ETH ZurichShow Abstract
Georgios Sarlas, ETHZ - Swiss Federal Institute of Technology
Kay W. Axhausen, Institute for Transport Planning and Systems
We present a direct demand modelling approach for origin-destination (OD) public transportation commuting flows between municipalities in Switzerland. The purpose is to improve the gravity modelling approach for OD flows by applying a spatial autoregressive regression model and testing different spatial weighting schemes. Besides the usual characteristics to explain commuting, we include a variable based on mean income differences to examine interregional demand patterns. In addition, we treat for the endogenous nature of the newly constructed variable and test its ability to serve as the basis for the construction of a spatial weight matrix, thus replacing the commonly used travel time / distance metric. We apply Ordinary Least Squares (OLS), Generalized Method of Moments (GMM) and Intrumental Variable (IV) estimators to obtain unbiased and consistent parameter estimates. We compare in-sample predictions of the models among each other and to the flows of the National transport model. We use data from the 2000 Federal Census and found significant spatial dependence in the residuals of the gravity model and thus the need for spatial regression models. We use a valid set of instruments to account for endogeneity and show that income differences are underestimated in the gravity and spatial models if assumed exogenous. Neighbouring municipalities affect flows under consideration positively at origins and negatively at destinations. Last, the spatial autoregressive models relying on a combination of origin- and destination-centric weight matrix outperform the gravity models in terms of the predictive accuracy when network and economic distance weights are used.
Application of Machine Learning to Two Large Sample Household Travel Surveys: A Characterization of Travel Modes
Robert Chapleau, Ecole Polytechnique de MontrealShow Abstract
Philippe Gaudette, Ecole Polytechnique de Montreal
Timothy Spurr, GIRO Inc.
Even in a context of rapidly evolving transportation and information technologies, household travel surveys remain an essential source of information for transportation planning. Moreover, as planning authorities become increasingly concerned with reducing the use of the private car, traveller’s mode choice patterns, should be re-examined. In this paper, a machine learning algorithm (random forest) is used to characterize the use of eight different travel modes observed in two consecutive household travel surveys undertaken in Montreal, Canada. The analysis incorporates roughly 160,000 observed trips. The random forest algorithm is trained on the 2008 survey data and is applied to the 2013 survey. The usefulness of the algorithm is evaluated using two numerical representations: the confusion matrix and the importance matrix. The results of this evaluation show that the random forest algorithm can generate a detailed and precise characterization of travel submarkets for four of the most commonly observed modes of travel (auto-drive, public transit, school bus and walk) using eleven attributes of households, persons and trips. The auto-passenger mode is difficult to characterize due to its dependence on unobserved intra-household interactions. The algorithm also has difficulty identifying users of rarely-observed modes (park-and-ride, kiss-and-ride, bicycle) but performs better in this regard than a traditional mode choice model. Finally, traveler’s age and the spatial orientation of origin-destination pairs are found to be decisive factors in the use of the auto-drive mode. This finding, combined with the stability of mode choice patterns observed over five years, highlights the difficulty of significantly reducing automobile use.
Adapting Old-Fashioned Travel Models to Forecast Potential Energy Impacts of New Driverless Vehicles
Di Kang, Virginia Transportation Research CouncilShow Abstract
John Miller, Virginia Department of Transportation
Driverless vehicles (DVs) may alter fuel consumption, depending on unknown operational effects (rising or falling roadway capacity); behavioral responses (longer trips or redevelopment of parking lots), or future policies (discouragement of private DV ownership). Despite the societal importance of impacts on energy, there is not a framework for incorporating the effect of DVs on a region’s fuel consumption into long-range planning. This study developed such a framework, considering the potential DV impacts on capacity, trip length, transit use, parking behavior, land development, and induced travel. Application of the framework to a Virginia case study showed that without any changes to vehicle energy sources but with DVs’ projected ability to eliminate excess acceleration, potential fuel increases are 2.2% (if people who do not have access to a vehicle can then use a DV); 17.1% (if capacity drops); and 23.3% (if trip lengths increase) and the potential decrease is 23.5% (if capacity increases). If the entire fleet adapts hybrid technology, however, net equivalent fuel consumption mostly decreases. The framework enables detailed scenarios to inform policies considered by localities. For instance, commuters seeking to send a vehicle back home to avoid parking charges could increase fuel consumption between 10.7% and 15.8% depending on mode share impacts. Incentives to share DVs, rather than have them privately owned, could yield lesser decreases for this particular region of 1.6% to 8.2%, depending on the degree of matching between a leading traveler’s destination and a following traveler’s origin. A next step is to test the framework elsewhere.