An Econometric Analysis of Public Transportation in Montreal Vincent Chakour Department of Civil Engineering and Applied Mechani

An econometric analysis of public transportation in Montreal

Vincent Chakour

Department of Civil Engineering and Applied Mechanics

McGill University, Montreal

Submitted Electronically August 14, 2013

A thesis submitted to McGill University in partial fulfillment of the requirements of the degree of Master of Civil Engineering

Acknowledgments

I would first and foremost like to express my gratitude to my supervisor, Professor Naveen Eluru, whose inspiring guidance, supervision, and constant support throughout my Master’s program made the completion of this thesis possible. I would also like to thank Guillaume Barreau for modeling the transit trips to McGill in Google Maps, Alex Burkley for his help with land use variable generation, and the transportation team in the Department of Civil Engineering as well as the TRAM research group in the School of Urban Planning for their help and support. I would like to give a special thanks to Ahmed El-Geneidy from the School or Urban Planning, Daniel Bergeron from the Agence Métropolitaine de Transport (AMT), and Jocelyn Grondines from the Société de Transport de Montréal (STM) for providing the necessary data to carry out the studies. I would also like to acknowledge financial support from Natural Sciences and Engineering Research Council (NSERC) of Canada under the Discovery Grants program and from the McGill Sustainability Projects Fund.

Contributions of Authors

Chapters 2, 3 and 4 of this report are based on three separate papers. The paper from Chapter 2 has been accepted for publication in the Journal of Public Transport on September 8th 2012 and presented at the Transportation Research Board 91st Annual Meeting, whereas the paper from Chapter 3 has been presented at the Transportation Research Board 91st Annual Meeting and submitted for publication in the journal of Transportation. The paper from Chapter 4 will be prepared in the summer of 2013 and will be submitted for publication in Journal of Transport Geography. The first paper, titled Travel mode choice and transit route choice behavior in Montreal: insights from McGill University members commute patterns is co-authored with my supervisor, Professor Naveen Eluru and Professor Ahmed El-Geneidy from the School of Urban Planning in McGill University. The second paper, titled Analyzing Commuter Train User Behavior: A Decision Framework for Access Mode and Station Choice and the third paper, titled An Examination of the Influence of Urban Form and Land Use on Bus Ridership, are also co-authored by my supervisor Professor Naveen Eluru. I have received supervision and guidance for all three papers by Professor Naveen Eluru. Specifically, he has provided me with the appropriate knowledge and modeling techniques to carry the studies. Finally, Professor Ahmed El-Geneidy has provided the data necessary to write the first paper, with the 2011 McGill Transportation Survey.

Abstract

The last decade has seen a strong push towards improving the sustainability of transportation systems in urban regions. In this regard, a recurring issue is that of increased car dependence in major North American cities. Policy makers are challenged to find new and innovative solutions to counter the negative externalities of this personal vehicle dependence. For instance, the air pollution and greenhouse gas emissions resulting from private vehicular travel is of a particular concern for the health and safety of future generations. Moreover, the prevalence of sub-urban life in North American cities in the recent years has resulted in increased private vehicle usage while reducing public transportation systems usage. A well planned and efficient public transportation system can provide equitable service and accessibility to the population as well as contributing to the reduction of air pollution and GHG emissions. An effective solution lies in transit agencies and government implementing policies that maximize transit use and minimize car dependence. Not surprisingly, many urban regions are enhancing public transportation infrastructure to address the private vehicle use challenge. A number of research efforts have been focussing on understanding individual behavioral challenges in using transit while several other studies have examined the factors affecting transit operations. These studies provide important information to local agencies and transit agencies to enhance public transit services and operations. This thesis is a collection of three distinct studies, each relating to public transportation issues from different perspectives. The first study examines individual home to work/school commute patterns in Montreal, Canada with an emphasis on the transit mode of travel. The overarching theme of this research is to examine the effect of the performance of the public transportation system on commuter travel mode and transit route choice (for transit riders) in Montreal. We investigate two specific aspects of commute mode choice: (1) the factors that dissuade individuals from commuting by public transit and (2) the attributes that influence transit route choice decisions (for those individuals who commute by public transit).

The second study is an effort to develop a framework for a better understanding of commuter train users’ mode and station choice behavior. Typically, mode and station choice for commuter train users is modeled as a hierarchical choice with mode being considered as the first choice in the sequence. This research proposes a latent segmentation based approach to relax the hierarchy. In particular, this innovative approach simultaneously considers two segments of station and access mode choice behavior: Segment 1 - station first and mode second and Segment 2 – mode first and station second. The allocation to the two segments is achieved through a latent segmentation approach that determines the probability of assigning the individual to either of these segments as a function of socio- demographic variables, level of service (LOS) parameters, trip characteristics, land- use and built environment factors, and station characteristics. Finally, the third study draws attention to the spatial characteristics affecting transit ridership. An analysis of bus stop level boarding and alighting is undertaken by developing ordered response models of the bus stop specific boarding and alighting by time of day. The analysis quantifies the influence of various exogenous factors including public transit accessibility indices (number of bus/metro/train stops around each stop, length of bus/metro/train lines, length of exclusive bus lanes), infrastructure attributes (road length by functional classification, bike lane lengths, distance to central business district, CBD), and land use measures (number of parks and their areas, residential area, number of commerces and their area, government and institutional area, resource and industrial area, and population density). The results obtained from the various studies provide interesting insight to transit agencies. It can assist these agencies in developing strategies to increase transit patronage, thus improving the sustainability of the overall transportation system.

Résumé

Au cours de la dernière décennie, une vague cherchant à améliorer le développement durable dans le secteur du transport urbain s’est manifestée. Dans cette veine, un obstacle récurrent est celui de la dépendance automobile. Les dirigeants ont le défi de trouver de nouvelles solutions et innovation afin de contrer les effets négatifs de cette dépendance. La pollution atmosphérique ainsi que les gaz à effet de serre (GES) provenant du haut taux d’utilisation automobile sont des enjeux importants pour la santé et sécurité des générations à suivre. De plus, l’expansion rapide des banlieues dans les villes nord-américaines encourage la surutilisation de l’automobile tout en diminuant l’usage du transport en commun.

Un réseau efficace et une bonne planification d’un système de transport collectif peut offrir un service équitable et accessible à la population tout en contribuant à la diminution de la pollution atmosphérique et émissions de GES. Une piste intéressante en guise de solution est d’implanter des politiques vertes visant à augmenter l’achalandage du transport en commun et diminuer la dépendance automobile. Sans surprise, plusieurs régions métropolitaines cherchent à améliorer leur service et l’infrastructure du transport collectif afin d’offrir une alternative viable au transport privé. Une multitude d’étude cherchent à mieux comprendre les comportements individuels portant à l’utilisation du transport en commun tandis que d’autres se concentrent plutôt sur le côté opérationnel. Ces études sont d’une importance primordiale, puisque l’information obtenue peut être utilisée par les agences de transport pour optimiser leur service. Cette thèse est un recueil de trois études distinctes, dans laquelle le thème du transport en commun est abordé en différentes perspectives. La première étude examine les déplacements domicile-travail et domicile- école à Montréal, avec une emphase particulière sur le transport collectif. Le thème global de cette recherche est d’examiner l’effet de la performance du transport en commun sur le choix modal et choix de route (pour utilisateurs de transport en commun) d’un individu. Deux aspects du choix modal sont examinés : (1) les facteurs qui dissuadent un individu d’opter pour le transport collectif et (2) les

5 caractéristiques qui influences le choix de la route empruntée (pour les déplacements en transport en commun). Ensuite, la deuxième étude vise à développer une structure ayant pour but de mieux comprendre le choix modal et le choix de gare pour les utilisateurs de trains de banlieues. Ces choix sont typiquement modélisés de manière hiérarchique, dans laquelle le mode d’accès se trouve à être le premier choix dans la séquence. Cette étude propose une approche basée sur la segmentation latente permettant une relaxation de cette hiérarchie. En fait, cette approche innovatrice prend en considération simultanément deux segments des choix de gare et mode d’accès : Segment 1- La gare est choisie avant le mode d’accès et Segment 2 – l’inverse. L’approche de segmentation latente permet détermine la probabilité d’appartenir à l’un ou l’autre des segments pour chaque individu en fonction de données sociodémographiques, du niveau de service du réseau, des caractéristiques de déplacements, des facteurs de l’aménagement de l’espace urbain ainsi que les particularités des gares. La troisième et dernière étude porte une attention particulière au lien entre l’espace urbain et l’achalandage. Une analyse des montées et des descentes à chaque arrêt d’autobus sur l’île de Montréal est effectuée à l’aide d’un modèle à réponse ordonnée pour différentes périodes de la journée. Cette analyse quantifie l’influence d’une multitude de facteurs exogènes autour de chaque arrêt, dont des indices d’accessibilité du transport collectif (nombre d’arrêt d’autobus/métro/train, présence de lignes d’autobus/métro/train, présence de voies réservées pour autobus), l’infrastructure (présence d’autoroutes, artères et de piste cyclable ainsi que la distance au centre-ville) et l’aménagement urbain (nombre de parcs et leurs superficies, nombre de commerces et leurs superficies, les zones résidentielles, les zone industrielles et la densité de la population). Les résultats obtenus de ces études peuvent offrir de l’information intéressante aux agences de transport. En fait, elles peuvent s’en servir pour développer de nouvelles stratégies afin d’augmenter l’achalandage et contribuer au développement durable du système de transport.

Table of Contents

Acknowledgments ...... 1 Contributions of Authors ...... 2 Abstract ...... 3 Résumé ...... 5 CHAPTER 1. INTRODUCTION ...... 9 1.1 Public Transportation in the Montreal Metropolitan Region ...... 9 1.2 Literature Review ...... 10 a. Transit Mode and Route Choice ...... 11 b. Access Mode and Station Choice ...... 13 c. Transit Ridership ...... 14 1.3 Research Questions and Objectives ...... 16 1.4 Structure ...... 17 CHAPTER 2. TRAVEL MODE CHOICE AND TRANSIT ROUTE CHOICE ...... 19 2.1 Introduction ...... 19 2.2 Data and Methodology ...... 21 a. Data ...... 21 b. Modeling Framework ...... 24 2.3 Results ...... 26 a. Empirical Analysis ...... 26 b. Sensitivity Analysis ...... 32 2.4 Conclusion ...... 34 2.5 Tables and Figures ...... 37 CHAPTER 3. ACCESS MODE CHOICE AND STATION CHOICE ...... 43 3.1 Introduction ...... 43 3.2 Data and Methodology ...... 46 a. Data ...... 46 b. Modeling Framework ...... 49 3.3 Result ...... 51 a. Model Estimation ...... 51 b. Model Validation ...... 57

c. Policy Analysis ...... 58 3.4 Conclusions ...... 59 3.5 Tables and Figures ...... 62 CHAPTER 4. URBAN ENVIRONMENT AND BUS STOP RIDERSHIP ...... 66 4.1 Introduction ...... 66 4.2 Data and Methodology ...... 68 a. Data ...... 68 b. Variables Considered ...... 68 c. Summary Statistics ...... 69 d. Visual Analysis ...... 70 e. Methodology ...... 71 4.3 Results ...... 73 a. High Ridership Stops ...... 74 b. Medium Ridership Stops ...... 75 c. Low Ridership Stops ...... 76 d. Correlation Matrices ...... 77 e. Elasticity Analysis ...... 78 4.4 Conclusion ...... 79 4.5 Tables and Figure ...... 81 CHAPTER 5. CONCLUSION ...... 88 5.1 Significant Contributions ...... 88 5.2 Concluding Comments ...... 92 References ...... 93 Appendix ...... 100

CHAPTER 1. INTRODUCTION

1.1 Public Transportation in the Montreal Metropolitan Region

In developed countries such as Canada and United States, a significant number of individuals depend on the automobile as the main mode of transportation. The high auto dependency, in turn, results in high auto travel demand on all roads. At the same time, the ability to build additional infrastructure is limited by high capital costs, real-estate constraints and environmental considerations. The net result has been that traffic congestion levels in metropolitan areas of Canada and the United States have risen substantially over the past decade (Schrank et al., 2011). The increase in traffic congestion levels not only impacts travel delays and stress levels of drivers, but also adversely affects the environment as a result of rising air pollution and greenhouse gas (GHG) emissions. An effective means of reducing the over reliance on the auto mode and ensuing negative externalities is to encourage public transportation ridership (Hodges, 2009). Regrettably, the ease of travel by personal vehicle also often means the hardship of travelling by other modes, most notably by public transportation. A well-designed transit system can provide equitable access to employment and recreational opportunities for the entire urban population, while simultaneously offering significant environmental benefits by offsetting emissions from personal vehicles (FHWA, 2002).

This thesis will concentrate on public transportation in the region of Montreal, Canada. Montreal is the second most populous metropolitan region in Canada with 3.7 million residents. According to the 2008 Montreal origin-destination (OD) survey (AMT, 2008), 67.8% of trips are undertaken by car, 21.4% by public transit, and 10.8% by active transportation (walking and bicycling). Its relatively high share of transit ridership (for a North American city) can be attributed to its multimodal transit system, including bus, metro, and commuter train. There are 4 metro lines, 5 commuter train lines, and over 200 bus lines, managed by different travel agencies. The Société de Transport de Montreal (STM), which serves bus and metro on the

Island of Montreal, has reached a high record of transit ridership in 2011 with 405 million trips, exceeding the previous record of the year 1945 (STM, 2011). Transit ridership has been in a constant decline from the aforementioned year up until the mid-70s due to urban sprawl, increase in vehicular highway capacity, and increase in automobile ownership. The STM’s ridership has then experienced high growth spikes from the Olympics in 1976 as well as the creation of the monthly pass (CAM) in the mid-80s. Finally, the ridership has been growing rapidly in the past decade, mostly due to an increase in government programs and subsidies. In the last 15 years, the transit patronage (bus, metro, train) has increased by over 25% for the Montreal Metropolitan Region.

The recent progress in relation with growing interest on reducing private vehicle usage has led to substantial interest within the travel behavior community on examining the key determinants of transit ridership. These studies have focused on understanding the different factors affecting transit ridership along two dimensions: (1) Individual level or (2) Transit operations. The current research effort contributes to public transportation literature by contributing to both these dimensions. Specifically, the emphasis of this research is on identifying the impact of individual and household socio-demographics, household residential neighborhood characteristics, transportation network attributes, transit service characteristics, and spatial and temporal transit accessibility on transit usage.

1.2 Literature Review

This section presents a detailed review of literature on the main issues addressed in this research. The literature section is organized as three sub-sections with the first sub-section focusing on travel mode and transit mode choice, the second sub-section highlighting earlier work on access mode and station choice decisions while the final sub-section discussing research on transit ridership.

10 a. Transit Mode and Route Choice

Mode choice has received wide attention within the transportation research community in general and travel behavior research community in particular. Transportation researchers have made giant strides in formulating advanced behavior-oriented frameworks and developing enhanced data collection strategies to accurately model travel mode choice decisions. A comprehensive review of earlier literature examining mode choice decisions is beyond the scope of the current paper. We present a brief summary of the most important characteristics of earlier research efforts investigating travel mode choice decisions.

(1) Earlier research has clearly shown that individual and household socio- demographics exert a strong influence on travel mode choice decisions. Specifically, gender, income, car ownership, employment status affect travel mode decisions (Bhat, 1997; Bhat and Sardesai, 2006). (2) Researchers have identified that tour complexity influences mode choice substantially (Stratham and Dueker 1995; Ye et al., 2007). Individuals with more complex commute tours (possibly with multiple stops) prefer to employ the auto mode of transportation. (3) Residential location, neighborhood type and urban form play a prominent role in determining the favored travel mode for commute (Van Wee and Holwerda, 2003; Pinjari et al., 2007; Frank et al., 2008). At the same time, individuals with inclination to commute to work by public transportation locate themselves in neighborhoods with adequate access to transit. (4) There has also been extensive focus on evaluation of the willingness to pay (i.e. amount of money travellers are willing to pay to reduce their travel time by unit time) for reducing travel time (Bhat 1997; Bhat and Sardesai, 2006). In more recent research studies, reliability of travel time is also incorporated within the framework to compute the value of travel time (Noland and Polak, 2002; Small et al., 2005; Bhat and Sardesai, 2006; Li et al., 2010).

(5) Other attributes that influence travel mode choice include travel distance (Scheiner, 2010), and household constraints such as picking up or dropping a child. (6) Advanced modelling frameworks including the mixed multinomial logit model and the generalized extreme value (GEV) models (see Bhat et al., 2008 and Koppelman and Sethi, 2008 for an exhaustive list) have been adopted to investigate travel model choice behavior.

On the other hand transit route choice has received very little attention. There has been very little empirical work within the public transportation community to examine transit route choice behavior from an individual perspective. To be sure, there have been research efforts examining transit route choice within the traffic assignment context. Liu et al., 2010 conduct an extensive review of literature on transit route choice. The paper classifies transit choice literature into three groups: (1) studies that employ shortest-path heuristics, random utility maximization frameworks of route choice within a user equilibrium based assignment (for example Marguier and Ceder, 1984; Lam and Xie 2002; Cepeda et al. 2006), (2) studies that consider intra-day dynamics within transit route choice, and dynamic traffic assignment (for example Nuzzolo and Crisalli, 2004; Hamdouch and Lawphongpanich, 2008), and (3) emerging studies that incorporate day-to-day dynamics, and real-time dynamics in transit route choice behavior (Coppola and Rosati, 2009; Wahba and Shalaby 2009).

The above approaches focus on transit route choice behavior from the system perspective i.e. the focus is on routing transit users based on transit network system pricing, level of service (LOS) measures and network congestion attributes. The individual user behavior is incorporated into the model indirectly. However, there has been little research that examines transit route choice from the individual’s perspective. Bovy and Hoogendoorn-Lanser (2005) is the only study that has investigated transit route choice decisions at the individual level. However, the focus of the study was on examining the influence of route choice with train as the

12 primary mode of transportation with a combination of walking, bicycling and car modes. The study conducted in Rotterdam–Dordrecht region in Netherlands examined the influence of travel time, waiting time, number of transfers (between trains) and walking time on individual route choice. The study developed a hierarchical generalized extreme value model to examine the choice of combination of transit route choice and choice of railway station types. The study was conducted using a small sample of records (235 observations) and considers only one public transportation mode (train). b. Access Mode and Station Choice

The access mode choice to commuter train stations as well as the stations choice for train users has received wide attention within the transportation community. A large proportion of these studies focused on access mode choice. The findings from studies investigating mode choice to train stations are analogous to those obtained from studies on general mode choice. Givoni and Rietveld (2007) show that the availability of a car does not have a strong effect on the choice of access mode to the station. Further, the authors find that improving accessibility to stations by adding newer stations will only result in a mode shift from transit to active transportation (walking and cycling), leaving the car mode share unchanged. Keijer and Rietveld (2000) found that mode choice behavior depends strongly on distance to station. Specifically, active modes of transportation are preferred for shorter distances, whereas driving and transit are favored for longer distances. Finally, Krygsman et al. (2004) found that if the distance to the station exceeds a certain threshold, users will not consider transit alternatives.

On the other hand, research on boarding station choice has found that frequency of trains at the station, parking availability, station facilities, and travel time to station (always considered along with mode choice) plays a major role in the decision process (Debrezion et al., 2007, 2009; Fan et al., 1993; Wardman & Whelan, 1999). The most common approach employed when modeling mode and station choice simultaneously is the nested logit model with mode as the choice in the

13 upper level. It is important to note that only Fan et al. (1993) and Wardman and Whelan (1999) employ disaggregate individual level models. The other studies (Debrezion et al., 2007, 2009) develop aggregate models at the postal code level (not individual level). The aggregate studies employ socio-demographic information at a postal code level and individual level information is not considered. Moreover, most of the access mode and station choice research has been undertaken in the European context where car mode share to train station (drive alone or shared ride) is lower than 15% (Givoni & Rietveld, 2007). The behavioral processes under consideration might be different in the North American context, especially given that the car mode share to station is greater than 60% (much higher for most urban regions). c. Transit Ridership

Several studies examine transit ridership in an attempt to link ridership with socioeconomic characteristics, built environment, and transit attributes across different contexts. Earlier research has focused on understanding the different factors that affect transit ridership at a macro-level (region or country). Taylor et al. (2009), for example, have undertaken a country-wide study for 265 U.S. urbanized areas and concluded that transit ridership is influenced by the regional geography, the metropolitan economy, the population characteristics, and the auto/highway system characteristics. The authors have classified the factors that affect transit ridership as internal (fare, level of service) or external (income, parking policies, development, employment, fuel prices, car ownership, and density levels) variables. They observed that external factors generally have a greater impact on ridership than internal factors.

Some studies examined the effect of trip costs, such as fares, fuel price, and parking price. The elasticity of transit ridership with respect to the fare is negative and inelastic for all transit, and even more so for bus ridership compared to other public transportation modes (Hickey, 2005; Wang and Skinner, 1984). There is also a general consensus that the elasticity of transit ridership with respect to gasoline price is positive and inelastic, especially in medium sized cities (Matson, 2008; Currie and Phung, 2007).

The price of parking also affects transit ridership; imposing a daily parking fee for commuters will significantly increase transit patronage (Hess, 2001).

On the other hand, a distinctive body of literature focuses on the effect of transit attributes and built environment on its patronage. Most of these studies examine the station or stop features affecting ridership or station choice for the rail mode (Brown and Thompson, 2008; Debrezion et al., 2007, 2009; Fan et al., 1993; Frank and Pivo, 1994; Sung & Oh, 2011; Wardman & Whelan, 1999; Weizhou et al., 2009). Debrezion et al. (2009) found that the availability of parking spaces and bicycle standing areas have a positive effect on the choice of the railway station. Brown and Thompson (2008) observed that rail ridership decline in Atlanta could be explained by the employment decentralisation, while Shoup (2008) observed that Transit Oriented Development (TOD) comprised of high commercial intensity positively affects transit ridership at the rail station. In fact, Sung & Oh (2011) also recognized that some TOD factors have a positive effect on transit ridership. They found that important factors affecting ridership at rail stations are land use mix, street network, urban design, and an overall pedestrian friendly area around the stations. To a lesser extent, the ridership has also been analyzed at metro stations (Chan & Miranda-Moreno, 2013; Gutiérrez et al., 2001; Lin & Shin, 2008). Chan & Miranda-Moreno (2013) found that commercial and governmental land use, bus connectivity, and transfer stations are all associated with attracting ridership during morning peak hours. Lin & Shin (2008) observed that transfer stations affect ridership positively. Moreover, the authors found that retail and service area and walkability around the stations (sidewalk length, 4-way intersection) have positive impacts on ridership.

The relationship between bus ridership and built environment has been explored with the use of routes or route segments as the unit of analysis (Stropher, 1992, Peng et al., 1997). This type of analysis has certain limitations such as unequal route lengths, representation of transit service variables, and inter-route relationship analysis. Peng et al. (1997) suggest that although a stop level model requires more detailed data that may not have been available at the time, it may prove to be more appropriate. Few papers have analyzed ridership as a function of the urban environment at a stop level for the bus mode. Ryan and Frank (2009) have studied the influence of pedestrian

15 environments on bus ridership. The authors found that the built environment, specifically the walkability of an area, is a useful tool for predicting transit ridership at a bus stop level. However, they examined total ridership (no distinction between boarding and alighting) and only consider a limited amount of built environment variables. Johnson (2003) also examined ridership at a bus stop level using an ordinary least squares regression, finding that land-use and density have the most important impacts on ridership. More specifically, it was found that multifamily residence, mixed-use, and retail-commercial land uses affect bus boardings. This study focuses its analysis solely on boardings at bus stops, neglecting any possible interactions with the alightings. Moreover, Chu (2004) noted that the presence of bus or trolley stops around a particular bus stop exerts a positive effect on ridership using a standard poisson regression. Finally, Estupiñán and Rodríguez (2008) explored the effect of the built environment on boardings at Bus Rapid Transit (BRT) stations in Bogotá while accounting for the simultaneity of transit demand and supply. The authors highlight the importance of urban environmental interventions to support transit use.

1.3 Research Questions and Objectives

Although the city of Montreal has relatively good public transport service compared to other North-American cities, agencies constantly need to optimize the service to attract more transit riders. As the urban population grows and travel demand increases, comprehensive transportation planning becomes essential to compete with the automobile mode. Adequate planning of transit infrastructure requires a comprehensive understanding of several issues and challenges that public transportation faces.

The entire scope of research on public transportation far exceeds the research agenda of this thesis. For the purpose of this thesis, we will concentrate our efforts on the following questions.

 What is the decision process of individual travel in the context of transit use?  What are the factors influencing the modal choice of a given commute trip?

 How do transit trip characteristics and individual level attributes affect the attractiveness of transit trips?  How can we increase transit mode share to train stations?  What makes a commuter train station more or less attractive for a commuter train user?  How do the built environment and urban design affect ridership?

A series of objectives is established to respond to the preceding research questions.

1. To examine the travel choices for individuals travelling to McGill University  Model mode choice and transit route choice for these commuters  Perform a sensitivity analysis to provide the magnitude of the impact of variables on the choice process at work 2. To investigate access mode choice and station choice for commuter train users  Propose an innovative approach to model these two choices simultaneously  Quantify the effects of the elements influencing these choices 3. To explore the built environment and land use factors influencing bus ridership at the stop level  Present a visual representation of the bus boarding and alighting for several time periods  Perform an ordered regression to quantify the effect of urban form on ridership

1.4 Structure

The remainder of the thesis is comprised of four chapters. Chapter 2 examines individual home to work/school commute patterns in Montreal, Canada with an emphasis on the transit mode of travel. Chapter 3 seeks to develop a framework for a better understanding of commuter train user mode and station choice behavior. Chapter 4 draws attention to the spatial characteristics affecting ridership, for which an analysis of bus stop level boarding and alighting is undertaken. Finally, Chapter 5 completes the thesis by identifying the main findings and contributions of the studies.

CHAPTER 2. TRAVEL MODE CHOICE AND TRANSIT ROUTE CHOICE

2.1 Introduction

In this study, with the objective of enhancing our understanding of public transit usage behavior, we examine individual home to work/school commute patterns in Montreal, Canada. The research is focussed on identifying how the performance of existing transit infrastructure affects transit choice vis-à-vis automobile choice and transit route choice (with multiple transit options available to transit riders). To achieve these objectives the current study employs a two pronged approach. First, we examine the individual decision making process in the context of travel mode choice (automobile versus transit). To elaborate, we identify the factors that dissuade individuals from commuting to work/school by transit. The analysis will enable us to draw insights on the mode choice decision process thus allowing us to make recommendations to enhance the attractiveness of the transit mode to commuters. Second, we study how the performance of the different transit modes in Montreal affect route choice decisions for transit riders. Montreal with its unique multimodal public transportation system consisting of bus, metro and commuter train offers multiple transit route alternatives to individuals commuting to downtown. The examination of individual transit route choice behavior will enable us to identify important attributes that influence route choice decisions. In both phases, the analysis evaluates the impact of various exogenous factors on the choice process including (1) individual and household demographics, (2) level of service measures of the transportation system (auto and public transit), and (3) accessibility to public transportation facilities. The results will be employed to provide recommendations to transit agencies on enhancing transit services in the urban region.

This research study employs a unique survey conducted by researchers as part of the McGill University Sustainability project. The survey collected information on commuting patterns of students, faculty and staff from McGill University. McGill

University, located in downtown Montreal, with its workforce of about 50,000 individuals offers a unique opportunity to examine travel behavior of a large sample of individuals commuting to the downtown. The analysis is undertaken using a multinomial logit model for the travel mode choice component and a mixed multinomial logit model for the transit route choice component. The estimation results are employed to undertake policy sensitivity analysis to evaluate how potential changes to public transportation performance affect travel mode choice and transit route choice.

In this context, the current study offers an opportunity to examine the public transit usage choices of a large sample of commuters travelling to downtown Montreal. It is not surprising that commuters travelling to downtown Montreal have multiple transit alternatives to choose from. It would be possible for an individual to (1) Walk – Bus – Metro – Walk, (2) Walk – Metro – Bus – Walk, (3) Walk – Train – Walk, and (4) Walk – Train – Bus – Walk are all feasible alternatives. These transit alternatives differ in terms of travel time, travel cost, transfers, walking times, and waiting times. It is important to recognize that individuals residing in urban regions with multiple transit route alternatives face an important decision. Understanding this decision framework will allow public transportation agencies to target improved coordination across their services to deliver enhanced transit service to urban residents. There has been very little work undertaken to behaviorally examine how transit users choose among such multiple alternatives (except Bovy and Hoogendoorn-Lanser, 2005). The current study extends Bovy and Hoogendoorn- Lanser (2005) research by considering multiple modes of public transportation (bus, metro and train) and estimating the model for a larger sample of transit road users.

Further, a mixed multinomial logit modelling framework is employed to examine transit route choice model. There are two reasons for adopting the more complex mixed logit model for our analysis. First, the impact of exogenous variables (such as travel time, waiting time, and walking time) might vary across different individuals. In the traditional multinomial logit model framework these intrinsic unobserved taste preferences are not accounted for (Bhat et al., 2008). The mixed

20 logit model allows us to estimate individual level parameters through distributional assumptions on the nature of the parameter. Second, it is possible that there is a host of unobserved attributes that are common to various alternatives an individual faces in the route choice decision. To elaborate, within the multiple alternatives available to different transit riders, it is possible that there are overlapping attributes (observed and unobserved) in the choice set for each individual. The occurrence of such overlap across the alternatives inherently violates the independent and identically distributed error term assumption of the traditional multinomial logit model. Neglecting the presence of such potential dependence across alternatives will result in incorrect estimates of the attribute influence on decision process.

In summary, the current study estimates a multinomial logit model of travel mode choice and a mixed logit model of transit route choice behavior on a large sample of data. The results from the analysis will offer insights that are particularly useful for public transit agencies in Montreal and Canada.

2.2 Data and Methodology a. Data

Study Region

A very good reason for the lack of empirical work on transit route choice behavior is the lack of well-connected multimodal public transportation systems in North America. Montreal, with its unique multimodal system, provides us with a test bed to examine transit route choice behavior.

Data Source

The data employed in the current study is drawn from a web-based survey of the McGill community members (students, staff and faculty) conducted during the months of April and May 2011. The survey collected information on the community

21 members’ socio-demographic information (age, gender, vehicle ownership), and McGill University experience (in years). Further, the survey gathered details on community members’ regular commuting patterns. In particular, the respondents were requested to provide the sequence of their regular commute to McGill with information on their start time to work, arrival time to work, transportation mode, and detailed transit route information for transit users. A screenshot of the web- based survey requesting the commuting pattern information is provided in Figure 1. The figure provides the sequence of questions for a respondent who has walked to the metro station, travelled by metro and then walked to reach campus. Information on the exact metro line is also collected. In addition to the above information, origin and destination postal codes were obtained for all respondents through a McGill internal employee and student database.

The web-survey was hosted and administered internally within McGill University in the months of April and May 2011. A total of 19,662 surveys were distributed among the McGill community members. The survey administered elicited 5,016 responses prior to the closing date. The data thus collected was thoroughly examined for consistency and erroneous reporting and the inconsistent records were eliminated from the database. The resulting sample consisted of 4,698 entries. Of these records 2,616 respondents (56%) are McGill employees (which includes both faculty and staff), and 2,032 respondents (43%) are McGill students, and the remaining 50 respondents (1%) included exchange students, and visiting professors. The reader would note that the web-based survey intentionally oversampled the employee community relative to the student community. For our analysis, we limited ourselves to community members commuting to the downtown campus.

Data Set Assembly for Analysis

The dataset preparation involved two distinct components. The initial part of the data assembly process focussed on compiling the travel mode choice dataset for the car versus transit model. The subsequent part of the data assembly was targeted

22 at generating all transit alternatives for the individuals choosing to commute by transit. The following discussion provides more details of the data assembly process for each component individually.

In our empirical case, we are interested in examining why automobile users are not commuting to work by transit. So, we select only those commuters that employ either the car mode or the transit mode in our analysis. The sample consists of 1845 records. Of these 1291 (70%) respondents commute using transit while 554 (30%) respondents commute by car. For these respondents we need to generate the LOS attributes for modes under consideration. The analyst is only aware of the LOS attributes for the chosen alternative. We need to generate LOS attributes for the competing modes. The research team employed two sources for generating this information. First, car in-vehicle travel times for all individuals (irrespective of their choice) were generated using LOS matrices for postal code origin and destinations. Second, Google Maps were employed to generate the best transit alternative available to the individuals using car at the time of his/her departure to work. The respondent provided transit route information was compiled for respondents who chose transit.

The second component of the data assembly process generated alternative transit routes for the transit commuters. The alternative generation was achieved using a Google Maps procedure that identifies unique alternative transit routes between the respondent’s origin and destination (see Figure 2 for an example). The routes obtained are compared with the respondent’s transit commute route and the chosen alternative is tagged. The transit alternatives for respondents varied from one to six in the following proportions: 5.8%, 33.2%, 32.0%, 22.8%, 5.6% and 0.6%. Clearly, a larger proportion of transit users (88%) have between two to four alternatives to commute to work. This statistic clearly highlights that transit commuters to Montreal downtown region have multiple alternatives to choose from.

Sample Statistics

Descriptive statistics for the samples for travel mode choice and transit route choice are presented in Table 1. The sample statistics for travel mode choice dataset are presented in the top part of the table followed by the statistics for transit route choice dataset.

Travel Mode Choice

The average travel time values for transit and car modes are substantially different. It is not surprising that travel times by transit are superior especially given the large share of proportion of transit users. The sample consists of a larger share of females compared to men. The majority of the respondents are in the age groups of 25-45 and 45-65. A majority of the respondents are full-time McGill community members. The vehicle ownership analysis indicates a large proportion of 0 vehicle and 1 vehicle households in the sample. The number of transfers for transit varies from 0 through 4. The proportion of 0 and 1 transfers (~83%) highlights the well- connected public transportation system in Montreal for the destination of McGill University.

Transit route choice

The average travel time is about 24 minutes for transit alternatives which is higher than the 18 minutes reported earlier because this dataset includes the chosen as well as the not chosen transit alternatives. The average walking time for transit alternatives is about 17 minutes, while the average waiting time is only 3.7 minutes b. Modeling Framework i. Travel mode choice model

A classical Multinomial Logit (MNL) model is employed to examine travel mode choice. The modeling framework is briefly presented in this section. Let q be the index for commuters (q = 1, 2, ..., Q) and i be the index for travel mode alternatives (i = 1, 2,… I). With this notation, the random utility formulation takes the following familiar form:

(1)

In the above equation, represents the utility obtained by the qth commuter in choosing the ith alternative. is a column vector of attributes influencing the choice framework. α is a corresponding coefficient column vector of parameters to be estimated, and is an idiosyncratic error term assumed to be standard type-1 extreme value distributed. Then, in the usual spirit of utility maximization, commuter q will choose the alternative that offers the highest utility. The probability expression for choosing alternative i is given by:

(2) ∑

The log-likelihood function is constructed based on the above probability expression, and maximum likelihood estimation is employed to estimate the α parameter. The reader would note that the travel mode choice model with two alternatives collapses to the conventional binary logit model. ii. Transit route choice model

The mixed logit modelling framework employed to study transit route choice behavior is briefly presented in this section. Let q be the index for commuters (q = 1, 2, ..., Q) and i be the index for transit route alternatives (i = 1, 2,… I). With this notation, the random utility formulation takes the following familiar form:

( ) (3)

In the above equation, represents the utility obtained by the qth commuter in choosing the ith alternative. is a column vector of attributes influencing the choice framework. and are column vector of parameters to be

estimated, where represents the mean effect and represents individual level disturbance of the coefficient. is an idiosyncratic error term assumed to be standard type-1 extreme value distributed. In the current paper we assume that the

elements of are independent realizations from normal population distribution:

~ .

Then, in the usual spirit of utility maximization, commuter q will choose the alternative that offers the highest utility. The probability expression for choosing alternative i is given by:

(( ) )

∫ (4) ∑ ( )

In the usual mixed logit form, the dimension of the integral is same as the

number of elements in vector (see Bhat et al., 2008). The log-likelihood function is constructed based on the above probability expression, and maximum simulated

likelihood estimation is employed to estimate the and parameters. In this paper, quasi-monte carlo (QMC) approach with 400 Halton draws is employed for the MSL estimation (see Bhat 2001 and Bhat et al., 2008 for more details on estimating mixed logit models with Halton draws). The reader would note that in the transit route choice model, alternative specific variables cannot be introduced; hence appropriate interactions with LOS attributes are computed to incorporate the effect of individual socio-demographics on route choice preferences.

2.3 Results a. Empirical Analysis

The empirical analysis in the paper involves the estimation of the travel mode choice model (binary logit model) and the transit route choice model (mixed multinomial logit model). Several variables were considered in the empirical analysis, including individual and household socio-demographics - age, gender, driving license, employment status, vehicle ownership, and LOS attributes - travel

26 time, travel time by mode, walking time, waiting time, number of transfers, and time of day. We also considered several interaction effects among the variables in both the mode choice and transit route choice model. The specification process was guided by prior research and intuitiveness/parsimony considerations. The final specification was based on a systematic process of removing statistically insignificant variables. We should also note here that, for the continuous variables in the data (such as age, travel time, walk and waiting times), we tested alternative functional forms that included a linear form, and non-linear forms such as square terms. In the subsequent discussion, we present the results from model estimations. i. Travel mode choice

In this model we examine the influence of factor influencing respondents’ inclination to use the Transit mode. The mode choice component offers intuitive results. Travel mode choice binary logit model estimation results are presented in Table 2. The car mode of transportation is considered to be the base alternative for all variables except for the travel time variable where we estimate a generic travel time coefficient.

Model fit

The log-likelihood value at convergence for the binary logit model is -718.6. The log-likelihood value at constants is – 1127.5. The hypothesis that the variables in the model do not offer any statistically significant improvement in model fit is re ected at any level of significance. The Mc adden’s ad usted rho-square value for the model is computed. It is defined as 2=1 - (L(β)-M)/(L(C)) where L(β) represents log-likelihood at convergence for the model, represents log-likelihood at sample shares and M is the number of parameters in the model (Windmeijer, 1995). The travel mode choice model has a rho-square value of 0.35 denoting that the model explains travel behavior adequately.

Model parameters

The constant corresponding to the transit mode is significantly positive indicating the inherent preference of transit mode among respondents. Individual and household socio-demographics attributes influence the choice process. Age exerts a significantly negative influence on choosing the transit mode. This is expected because younger individuals of the McGill community (students and younger employees) are more likely to use the public transportation mode compared to older members of the McGill community. The result is further supported based on the influence of the role of the respondent. The adoption of transit is the highest among students followed by staff members compared to faculty members. Among the employees, full-time employees and students are more likely to commute by transit compared to part time employees and students. The full-time members have a more definite work schedule, making it easier for them to commute to work by transit. The license status of the individual significantly affects the choice between transit and car. Within the student community it is possible a number of individuals do not have driver licenses and are captive to the public transportation mode. Household car ownership also has a strong negative effect on the choice of transit mode. Households with more cars are least likely to commute to work by transit.

LOS attributes including travel time, walking time, number of transfers and departure time, significantly influences the choice between auto and transit modes. Specifically, increasing travel time reduces the likelihood of choosing the alternative (see Pinjari and Bhat, 2006, Bhat and Sardesai, 2006 for similar results). The increase in the amount of walking within the transit alternative significantly reduces the likelihood of the respondent using transit for commuting. In fact, the relationship between transit usage and walking time is non-linear with statistically significant linear and square terms. Further, increase in the number of transfers for travelling by transit reduces the likelihood of using transit substantially. The departure time period influences the mode choice. Individuals travelling during the AM peak period are more likely to commute by transit mode. It is heartening for transportation

28 agencies that an increased frequency of transit service during the peak periods contributes to increased transit usage.

ii. Transit route choice model

The mixed multinomial logit model of transit route choice evaluates the propensity for choosing the transit route alternatives based on route LOS attributes and their interactions with a host of individual and household socio-demographics,. The results also support our hypothesis of considering the mixed multinomial logit model as opposed to the traditional multinomial logit model. The results of the estimation are presented in Table 3.

Model fit

The log-likelihood value at convergence for the mixed multinomial logit model with 17 parameters is -767.03. The log-likelihood value for the multinomial logit model with 14 parameters is -779.1. The hypothesis that the additional variables from the mixed logit model do not offer any statistically significant improvement in model fit is re ected at any level of significance. The Mc adden’s adjusted rho-square value for the model is 0.40. The adjusted rho-square denotes that the model describes the route choice behavior satisfactorily.

Model parameters

The transit route alternatives in the choice set are a combination of bus, metro and train alternatives. Hence, it is possible to evaluate the intrinsic preferences of respondents towards commuting by each public transportation alternative. The results indicate a clear preference order for transit alternatives: metro, bus and train. The result is along expected lines given the winter weather conditions in Montreal. Metro service is underground and usually protects commuters from weather. The intrinsic disinclination for the train mode accounts

29 for the expensive monthly pass required for train usage compared with other public transportation modes. The reader should note here that unobserved intrinsic preferences towards the transit modes were insignificant.

In this model, we evaluate the influence of two overall route characteristics on route choice: (a) shortest travel time route and (b) route that allows the respondent to arrive earliest. Individuals are likely to evaluate routes based on such characteristics and hence are considered in the model. These variables are essentially dummy variables that are set to 1 for the route alternatives that satisfy the criterion of interest. The results indicate that commuters are likely to choose alternatives that allow them to arrive at the earliest travel time and are not really influenced whether the alternative is the shortest or not.

The travel time coefficients clearly indicate the negative propensity towards travel for respondents. A closer examination of the travel time results leads to interesting insights. In the model, we introduced travel time by mode. The coefficient on each of these modes provides the sensitivity to travel time for respondents by that mode. The results indicate that individuals find travel time on the bus mode the most onerous while the sensitivity to travel time on metro and train are quite similar on average. Public transportation agencies should investigate the reasons for this apparent discomfort and propose remedial measures to alter this. Further, the results indicate that there is substantial variability across the population on how individuals perceive travel time on the train as indicated by the significant standard deviation (0.043). A plausible explanation for the variability in the effect of travel time is probably related to weather conditions in Montreal. During snow storms trains schedules are often affected thus making commuters place a higher premium on travel time. There is a need for future research to examine this aspect in detail. The reader should note that in spite of the statistically significant variation, the likelihood that train travel time is more onerous than bus travel time is very small (<1%).

The influence of walking time is along expected lines. Specifically, transit route alternatives with smaller walk times are preferred. The model results indicate the presence of a non-linear relationship (linear and square terms). Further, the results indicate a substantial variation on the mean effect of the walking time variable. The result is quite intuitive, because, different individuals are likely to be differentially sensitive to walking time. There are individuals who will consider walking time to transit as an opportunity to relax or exercise while others might consider it a burden. The overall effect at the individual level for walking time results in a downward parabola with a shifting peak (based on the mean value).

The alternatives considered in our analysis involve a significant share of alternatives with transfers. Further, there is a potential waiting time associated with each of these transfer points. We attempted to incorporate their influence on transit route choice in multiple ways. The following specification offered the most intuitive results: (1) number of transfers and (2) waiting time per transfer. As expected, alternatives with fewer transfers were preferred. At the same time, individuals exhibited higher likelihood of choosing alternatives with smaller waiting time per transfer. The reader should note that the impact of number of transfers varied significantly across the population as indicated by the standard deviation coefficient. The variation is expected because it is possible that some individuals are less averse to transfers compared to other individuals. Further, the convenience of a transfer varies substantially depending on where they board and where they make the transfer. In some cases, the transfer points are within the same transfer center while for others, commuters need to walk to farther locations.

In a route choice model, it is not possible to evaluate the effect of socio- demographics directly. Hence, we evaluate their influence by estimating interactions terms with LOS attributes. In the model we consider interactions of gender, age, employment status with total travel time (sum of travel time by all modes in a route). The results offer interesting findings. Travel time interacted with female gender results in a positive coefficient indicating that females are less sensitive to travel time compared to males. To be sure, the overall sensitivity to travel time for

31 females is still negative. However, it is lower than the sensitivity of travel time for males. The results corresponding to the interaction variable involving age and total travel time indicate that with increasing age of the respondent, there is a marginal reduction in the sensitivity of travel time. The result might seem counter-intuitive and requires more detailed future analysis. The interaction of total travel time variable with the role of McGill community members provides intuitive effects. Faculty members are more sensitive to travel time compared to the students and staff members. b. Sensitivity Analysis

The exogenous variable effects presented in Table 2 andTable 3 do not directly provide the magnitude of the impact of variables on the choice process at work. To do so, we conduct a sensitivity analysis of attribute effects on travel mode choice and transit route choice models. i. Travel mode choice

The objective of the policy sensitivity analysis is to investigate the influence of exogenous variables on transit usage. The aggregate “elasticity effects” computation involves the following steps: (a) binary logit model results at convergence presented in Table 2 are used to compute the base probabilities for all respondents in the dataset using the attribute levels as reported. (b) The attribute of interest is chosen and new attribute levels for all respondents are computed in a pre-defined manner. (c) The new attribute computed is employed in the place of the base attribute along with the other base attributes and new probability measures are generated, and (d) percentage change in probabilities relative to the sum of base aggregate shares is computed.

The scenarios considered for analysis include: (a) reduced travel time by transit - five and ten minutes, (b) increased travel time by car– five and ten minutes, (c) reduce walking time for transit – five and ten minutes, (d) reduce transit transfers

32 by 1, and (e) reduce vehicle ownership by 1. The percentage change in mode share for transit and car for the above scenarios are provided in Table 4.

The following observations can be made based on the results. First, the results clearly indicate that travel mode shares are very sensitive to the level of service attributes i.e. by enhancing the public transportation modes we can encourage more travellers to use the transit mode. The changes in travel times by mode provide intuitive results. Second, we see that a change in transit (reduction) or car (increase) travel time lead to similar percentage changes in the overall aggregate share. Third, the influence of walking time on travel mode is marginally lower than the effect of travel time on mode choice. Public transportation agencies must recognize that reducing walking time by increasing accessibility of public transportation mode is less expensive than reducing transit travel time financially. Hence, adequate resources need to be allocated to identify urban pockets that have inadequate transit accessibility (bus, metro or train) and improve accessibility in these urban pockets either by increasing the number of stations or improving feeder services to metro and train stations. Fourth, the reduction in transit number of transfers by 1 would increase transit share by 9.17%. The results indicate that each transfer that individuals are faced with has an effect similar to that of a reduction in travel time by 10 minutes. In other words, individuals consider every transfer that they have to make along their route to be as burdensome as an additional travel time of approximately 10 minutes. The result clearly highlights the need for public transportation agencies to investigate the possibility of developing more direct services between downtown and rest of Montreal. Finally, the effect of vehicle ownership is also staggering on the travel model choice. Even a reduction of household vehicle ownership by 1 might change the share of transit ridership by about 16%. Policy makers need to consider incentives to residents in Montreal towards altering vehicle ownership because it might lead to a significant increase in transit ridership. ii. Transit route choice

The approach employed to undertake sensitivity analysis for the transit route choice model is very similar to the approach described for the travel mode choice except for one small change. In the route choice context, however there are no alternative specific coefficients as the case was in the travel mode choice model. Hence changes to attribute levels do not capture the change in probability adequately. Instead, we focus on changes to attributes based on the presence of different transit modes within the alternative. For instance, for alternatives with bus mode we reduce the travel time by bus by five minutes while the alternatives that do not have bus are not altered.

The scenarios considered for analysis include: (a) reduced travel time by bus, metro and train - five and ten minutes, and (b) reduced walking time for alternatives involving bus, metro and train - five and ten minutes. The change in transit route choice probabilities for all the scenarios is provided in Table 5.

The following observations can be made based on the results. First, change in travel time by bus has the most positive effect, i.e. if alternatives involving bus mode can be improved to reduce travel times the likelihood of individuals choosing that alternative increases substantially. The public transportation agencies and metropolitan organization for Montreal city need to coordinate and develop a dedicated bus priority signalization and/or exclusive bus lanes in order to improve bus travel times. Second, reduction in travel time by train has the least influence indicating that respondents using trains are relatively satisfied with current train travel times. Finally, changes to walking time are likely to affect alternatives with bus and metro substantially, whereas alternatives with trains are only marginally affected by improving accessibility to trains.

2.4 Conclusion

As sustainability in transportation is becoming an increasingly important issue, there is need for us to investigate the automobile dependence and suggest

34 recommendations to encourage more people to employ transit for their travel. Towards this end, we examine two specific aspects of commute mode choice. First, we study the factors that dissuade individuals from commuting to work/school by transit. Second, for individuals commuting to work/school by transit we analyze their transit route choice decision. The data employed in the current study is drawn from a web-based survey of the McGill community members (students, staff and faculty) conducted during the months of April and May 2011. The survey collected information on the community members’ socio-demographic information (age, gender, vehicle ownership), and McGill University experience (in years). Further, the survey gathered details on community members’ regular commuting patterns. The analysis in the research is undertaken using multinomial logit model for travel mode choice component and mixed multinomial logit model for the transit route choice component.

The travel mode choice results clearly highlight the role of travel time, walking time, number of transfers on the propensity to choose transit. Further, the results also indicate that faculty members are least likely to choose the transit mode for commuting compared to staff and students. The policy sensitivity analysis conducted using the convergence results for travel mode choice indicate that changes to travel times by transit mode will result in increase in the proportion of riders using transit. Hence, public transportation agencies must consider the possibility of exclusive bus lanes or bus prioritized signals to improve transit times within the Montreal region. The results also highlight the role of walking time while choosing commute mode. Longer walking times act as deterrents to choosing transit mode. Hence, it is necessary for public transportation agencies to increase bus accessibility as well as provide better feeder access (through bus) to metro and train stations.

The transit route choice results provide interesting insights. The results indicate that individuals find travel time on the bus mode the most onerous while they are similarly sensitive to travel time on metro and train. Public transportation agencies should investigate the reasons for this apparent discomfort and propose

35 remedial measures to alter this. The results also clearly highlight the variability in sensitivity to various exogenous factors across the population supporting our hypothesis of employing a mixed multinomial logit model. The influence of gender on route choice indicates that women are less sensitive to travel time compared to men. Within the McGill context, faculty are likely to be more sensitive to travel time compared to staff and students. The policy analysis conducted indicates that reducing travel time by bus increases the likelihood of such alternatives being chosen substantially. So, public transportation agencies need to enhance bus travel times either through bus priority signalization or exclusive bus lanes. The policy results also indicate that routes with bus and metro alternatives are more sensitive to walking time.

This research is not without limitations. We recognize that the survey is conducted for a single work place. However, the large size of McGill University provides us with a relatively large sample to eliminate any intrinsic biases. The current study does not explore the impact of residential location choice on travel decisions adequately. In future research, impact of individual’s residential neighborhood characteristics on transit mode choice and route choice need to be explored.

2.5 Tables and Figures

Table 1: Summary Statistics

Travel mode choice dataset Mean travel time by transit (min) 17.8 Mean in-vehicle travel time by car (min) 36.1 Gender

Males 39.0 Females 61.0 Age

<25 20.5 25-45 42.7 45-65 33.9 >65 2.9 Employment Type

Part-Time 12.0 Full-Time 88.0 Vehicles Ownership

0 25.9 1 42.9 2 25.7 3 3.9 4+ 1.6 Number of transfers for the transit alternative

0 50.9 1 32.7 2 14.6 3 1.7 4 0.1 Transit route choice dataset

Mean Travel Time 23.9 Mean Total Walking Time 17.4 Mean Total Waiting Time 3.7 Transit route alternatives comprising Bus 69.3 Metro 49.0 Train 15.6 Average travel time by mode (min) Bus 21.6 Metro 10.3 Train 24.9

Table 2: Binary logit model results for Home-Work commute mode choice

Attributes Parameter t-stats

(Car alternative is the base)

Constant 9.4983 8.975

Age -0.2461 -6.261

Age squared 0.0023 5.657

Respondent status

Staff member 0.5099 3.332

Student 0.8231 3.054

Full time member of the 0.3494 1.791 community

Driver license status -1.3592 -3.900

Household car ownership -1.0847 -12.127

In-vehicle Travel time -0.0540 -6.584

Transfers -0.8906 -10.031

Walk time -0.0652 -2.537

Walk time square 0.0012 2.524

Departing time period is AM 0.2689 1.754 peak

Log-likelihood at Convergence -718.644

Log-likelihood at constants -1127.47

McFadden rho-square 0.35

Table 3: Multinomial logit model results for transit route choice

Attribute Parameter t-stats

Transit alternative has bus -0.0379 -0.178

Transit alternative has metro 0.5822 2.041

Transit alternative has train -1.8298 -2.587

The alternative with the earliest arrival time 0.3606 3.505

Travel time in bus -0.2447 -5.712

Travel time in metro -0.1358 -2.817

Travel time in train -0.1430 -2.967

Standard Deviation 0.0430 1.683

Total Walking time -0.3573 -8.376

Total Walking time squared 0.0029 4.134

Standard Deviation 0.1494 5.167

Number of transfers -2.4658 -8.192

Standard Deviation 0.9107 2.165

Waiting Time per transfer -0.0579 -1.811

Total travel time interactions with Socio- demographics

Female 0.0592 2.654

Age 0.0013 1.653

Faculty -0.0475 -1.802

Log-likelihood at Convergence -767.0

Log-likelihood at Equal shares -1310.1

McFadden rho-square 0.40

Table 4: Policy sensitivity analysis of the travel mode choice model

Attribute Car Transit

Travel time by Transit reduced by 5 minutes -10.95 4.67

Travel time by Transit reduced by 10 minutes -20.65 8.80

Travel time by Car increased by 5 minutes -11.04 4.70

Travel time by Car increased by 10 minutes -21.45 9.14

Walking time to transit reduced by 5 minutes -5.94 2.53

Walking time to transit reduced by 10 minutes -17.98 7.66

No. of transfers (for transit) reduced by 1 -21.53 `9.17

Household vehicle ownership reduced by 1 -37.44 15.95

Table 5: Policy sensitivity analysis of the transit route choice model

Bus Metro Train

Reduction in 5 minutes 10 minutes 5 minutes 10 minutes 5 minutes 10 minutes Attribute by

Travel Time 17.87 29.31 8.37 16.37 6.24 17.49

Walking Time 20.36 46.52 17.54 40.35 8.91 14.64

Figure 1: Screenshots of commuter sequence questions in the Web-based Survey

Figure 2: Transit route choice alternatives generated by Google Maps for a sample origin to McGill University

CHAPTER 3. ACCESS MODE CHOICE AND STATION CHOICE

3.1 Introduction

This study examines transit choice behavior, particularly the travel decision framework of commuter train users in Montreal. Montreal, with its unique multimodal transit system consisting of bus, metro and commuter train, offers a rich array of public transit alternatives to individuals travelling to and from different parts of the city. The commuter train provides access to the urban population from the suburbs to the central business district of Montreal. Improving the access to railway stations by transit and non-motorized modes can contribute to a car free life style (Martens, 2004). In this research, we examine the behavior of the commuter train riders in terms of their commuter train station and travel mode to commuter train choices (access mode). The focus of the analysis is on developing a behaviorally representative framework for understanding the decision processes involved in the station and access mode choice.

Towards this end, we propose an innovative latent segmentation approach that simultaneously considers two segments of station and access mode choice behavior: Segment 1 - station first and mode second and Segment 2 – mode first and station second. The allocation to the two segments is achieved through a latent segmentation approach that determines the probability of assigning the individual to either of these segments as a function of socio-demographic variables, level of service (LOS) parameters, trip characteristics, land-use and built environment factors, and station characteristics. Within each segment, the sequence structure imposed is followed to examine the choice processes. To elaborate, in the first segment, mode choice is modeled first and the station decision is modeled conditional on the mode choice decision. In the second segment, the choices are reversed. The latent segmentation based framework will allow us to identify important factors that affect the choice sequence decision while simultaneously modeling the mode and station choices. In fact, through this approach, we allow for

43 two distinct choice hierarchies (mode first and station second (MS) and station first and mode second (SM)) to be simultaneously considered in the analysis as two segments for individuals.

Earlier literature has examined the two causal structures separately and has chosen the one that offered better statistical fit (for example see Newman & Bernardin, 2010). However, these studies fail to recognize that because one causal structure offers improved fit, it does not rule out the presence of the other causal structure (even if for a small proportion of the population). The fit comparison only establishes that on average, one causal structure better represents the dataset. Our approach recognizes this limitation and provides an attractive approach where the two causal structures are effectively weighted to generate a more plausible decision process.

The proposed formulation is a unique structure for examining decision frameworks that are likely to be interconnected. The formulated model is empirically estimated using data from an onboard survey conducted by Agence Métropolitaine de Transport (AMT) for commuter train users. The data consisted of about 24,000 completed questionnaires. The model results clearly highlight the applicability of the proposed approach and provide interesting insights on individuals’ segmentation behavior (in terms of MS and SM) while simultaneously quantifying the influence of individual and household socio-demographics, household residential neighborhood characteristics, transportation network attributes, and transit service characteristics on station and mode choice components.

Previous studies examining station choice consider a very small sample of stations (2 or 3) in the choice set. We observed from the Montréal commuter train data that people exhibit a great variability in terms of the station choice in the database. Residents from the same neighborhood are observed to have boarded the commuter trains at varying locations, indicating that the station choice is not merely a decision to arrive at the nearest commuter train station (not even the nearest 3

44 stations). For a variety of reasons such as seat availability, parking, fare, or better transit coverage, some respondents travel to stations farther from home to board the commuter train. The current research, in addition to examining the access mode choice (drive alone, shared ride, transit, and active transportation), will also investigate the heterogeneity among individuals in choosing the commuter train stations (50 stations in the Montréal metropolitan region).

The decision framework of determining the station at which to board the commuter train and the corresponding travel access mode are interconnected. There is reason to believe that these are potentially simultaneous decisions. There are two possible approaches that have been employed to study these choices. The first approach employs a discrete choice model that has composite alternatives of station and travel mode combination (i.e. every combination of travel mode and station is considered as an alternative). In this approach, it is important to recognize the potential correlations between sets of alternatives. Towards accommodating such correlations, some studies have considered nested logit version of the composite alternative models where one of the decision is placed in the upper level and the other in the lower level (Debrezion, et al., 2007, 2009). The approach, though plausible, imposes a hierarchy that is very hard to validate in the dataset. Further, the number of alternatives explodes very quickly in this approach. For instance, the number of possible combination alternatives might go as high as 200 (4 modes and 50 stations). an alternate approach to account for the simultaneity involved in the decision process is to develop a simultaneous equation model that explicitly accounts for common unobserved heterogeneity across the two decisions (see Pinjari et al., 2011 for a description of such approaches). These approaches are simulation intensive and focus predominantly on the unobserved correlation across the choice processes.

We propose to employ a new latent segmentation based approach that allows us to incorporate simultaneously the two possible sequences (MS and SM). To elaborate, we hypothesize that individuals are likely to consider joint choices or interconnected decisions in a sequence, even if the time difference between these

45 decisions is infinitesimally small. Now, if there was a way to determine the hierarchy (i.e. whether individuals decide first on station or access mode), we can develop a sequential approach to modeling the decision process. Unfortunately, the true sequence is latent to the analyst. Hence, we propose a latent segmentation approach where the first segment follows the station first and mode second sequence and the second segment follows the mode first and station second sequence. The individuals are then allocated to these two segments based on a host of exogenous variables, including socio-demographic variables, LOS parameters, trip characteristics, land-use and built environment factors, and station characteristics. For instance, workers have primary access to automobile in the household and are probably more likely to decide on their mode (automobile) while subsequently depending on the perception of parking availability to decide on the station. Similarly, individuals residing close to the station might decide on the station first and then either walk or take transit (in inclement weather) to arrive at the station.

The approach proposed here improves on a similar method employed earlier in the context of residential and work location choices, which considered one of the two decisions as made first (Waddell et al., 2007). The proposed approach offers many advantages compared to the traditional alternatives. First, we gain a better understanding of the decision processes by examining who are the individuals who choose the station (or mode) first. Second, the approach proposed is free from simulation and easy to implement. Third, the results from our analysis will provide insights to transit agencies on how to improve transit service to reduce the automobile travel to commuter train stations. The regional transit agency AMT is particularly interested in enhancing the transit service to reduce the auto mode share for travel to commuter train stations.

3.2 Data and Methodology a. Data

The primary source of data for the research was based on an onboard survey conducted by the AMT for commuter train users in the month of September 2010. The survey consists of about 24,000 self-compiled responses to the onboard survey questionnaire. The information compiled includes individual and household socio- demographics such as age, gender, vehicle ownership, and occupational status. Also included are residential location, boarding and alighting commuter stations, final destination location, travel mode to the boarding commuter train station and from the alighting commuter train station, and travel departure times. The exhaustive database on the commuter train travel is compiled for analysis by eliminating missing records and inconsistent information.

Level of Service Variable Generation

To undertake travel access mode choice analysis, assembly of LOS attributes for all available alternative modes under consideration is required. In our study, we are faced with the challenge of generating these measures for all the alternatives as well as for all stations possible. A Google Maps based algorithm was used to generate the walk, cycle, drive, and transit time for all viable stations (more details on the process of compiling viable stations is described below). Further, transit alternatives available to the chosen station based on the departure times provided in the survey were also generated using a Google Maps based algorithm. The information on a transit trip was compiled only for those individuals for whom a transit alternative was available. Transit can be unavailable if the station is very close to the individual’s residence or if there are no transit services within 37 minutes of walking for the individual (a threshold implicitly established in Google Maps). For our model analysis, we randomly sample about 4,346 individuals from the 24,000 responses. The reason for sampling was to reduce the computational burden of generating level of service attributes. The survey database is appropriately augmented with the LOS attribute database generated. Also, parking data for each commuter train station was obtained from the AMT.

Station Choice Set Formation

To generate a behaviorally representative station choice set, we focussed on individual level choice set preparation. Considering all the station alternatives in the region as part of the choice set would not truly represent individual behavior.

The objective was to identify the maximum distance commuters are willing to travel vis-a-vis the closest station. Towards this end, we compute a distance ratio measure for every respondent. The measure is defined as:

Once we compute D for all the respondents in the sample, the 95th percentile was used to determine the threshold value. Based on the threshold value, viable alternatives can be generated. However, generating a single ratio measure would be inadequate. For example, a person living 500 m from the closest station might be willing to travel further – 3 km (high D, say 5) whereas the person living 15 km from the closest station might not consider travelling 90 km (assuming same D = 5). So we adopted a distance based D measure that reduces as the distance to closest station increases. We considered D values for the following intervals – 0-0.5 km, 0.5 – 1 km, 1 – 2 km, 2 – 4 km and > 4 km. The maximum viable distance thus computed was used to generate the set of alternatives that are feasible for each respondent. The number of alternatives varies from 1 to 18, with 91% of respondents having between 1 and 5 alternatives.

However, creating only one set of viable stations means that irrespective of chosen mode, the station set is generated in the same fashion for all individuals. This may lead to potential inaccuracy in the analysis, as the station set should also depend on mode (for the Mode first and Station second segment). For instance, an individual who walks to the train station should have a station set for which all stations are at a walkable distance. If this same individual would have driven to train station, it is strongly probable that his or her station set would expand, as more stations are accessible by driving than walking. To address this, a new set of viable

48 stations based on the chosen mode (driver/shared ride transit, active mode (walk/bike)) is generated. The same methodology previously stated is employed for each of the sub-samples to create mode specific station sets. This station set is exclusively employed when modeling station as second choice. In station as first choice model, mode to station is unknown to the analyst and is therefore not incorporated when generating station choice set.

Sample Statistics

Table 6 contains summary statistics for the sample compiled. The following observations can be made based on the summary statistics. The sample consists of 3,902 commuters of whom more than half reported driving to the commuter rail station. Females represent a slightly higher proportion of the sample. The majority of the respondents are between 26 and 54 years of age. Vehicle ownership (recorded as a binary variable) is constant in the different age groups at about 90%, with the exception of individuals 25 and under. The sample predominantly consists of workers, and almost half of the respondents claim to have departed between 6:30am and 7:30am. Further, we can see that individuals do not always board the nearest station – average distance to the nearest station is lower than the average distance to chosen station. Figure 3 illustrates the distribution of the geographic locations of the sample throughout the region. Evidently, there is a strong concentration of respondents around the train lines, notably in the western part of the Island of Montreal and Laval, as well as on the North Shore and the South Shore. b. Modeling Framework

The modeling approach proposed consists of three components: (1) latent segmentation component, (2) Mode choice component for each segment and (3) Station choice component for each segment. The first component represents a binary logit model while the latter two components are two multinomial logit models.

Let q be the index for commuters (q = 1, 2, ...,Q) and i be the index for segment (i = 1 or 2), m be the index for mode choice alternative (m = 1, 2…M), and s be the index for station alternative (s = 1, 2…S). With this notation, the random utility formulation takes the following form:

(1)

(2)

(3)

th where represents the utility obtained by the q commuter in selecting th the i segment, represents the utility obtained by choosing mode alternative m th in the i segment, and represents the utility obtained by choosing station th alternative s in the i segment. , , are column vector of attributes influencing the choice framework. , and are assumed to follow Type 1 Gumbel distribution. The commuter q will choose the alternative that offers the highest utility. are corresponding coefficient column vector of parameters to be estimated.1

The probability expression for each model component takes the usual multinomial logit form given by:

(4) ∑

(5) ∑

(6) ∑

With these preliminaries, the latent segmentation based probability for joint choice of mode m and station s with two segments can be formulated as follows:

1 The reader will note that the second model in each segment is conditional on the first model in the segment. , incorporate the information available to the commuter at that instant in the choice process. For example, if the mode choice is the first alternative, level of service attributes to the chosen station are unknown to the commuter.

(7)

The first term in Equation (7) reflects the first sequence - mode first and station second while the second term reflects the second sequence - station first and mode second. The reader would note that the exogenous variables in the second choice are generated while recognizing the chosen alternative attributes from the first choice process in the segment. The log-likelihood at the individual q is defined as:

Lq = *ln( ) (8) where = 1 if the mode and station combination is the chosen alternative and 0 otherwise.

L = ∑ (9)

The log-likelihood function is constructed based on the above probability expression, and maximum likelihood estimation is employed to estimate the

parameters. The model is programmed in GAUSS matrix programming language.

3.3 Result a. Model Estimation

The universal set of variable considered in our analysis for the latent segmentation sequence choice, mode choice, and station choice components of the proposed model includes individual and household socio-demographics, trip specific attributes, level of service attributes, station attributes, residential and station level land-use and built environment variables.

The individual and household variables considered include age, gender, and vehicle ownership (characterized as Yes/No). The LOS attributes considered include travel time by different modes, average travel times to viable stations, and travel time to closest and chosen stations. The trip specific attributes considered include

51 egress mode and departure time. Station level attributes considered include parking and fare information for the station.

To generate residential and station level land-use and built environment information, a large database of land-use, socio-demographic, transportation network, and vehicle ownership information was generated at the Traffic Analysis Zone (TAZ) level for the Montreal metropolitan region. Naturally, using all these variables simultaneously in model estimation was not possibly due to potential collinearity effects. Hence, a principal component analysis (PCA) based factor analysis was conducted (see Pinjari et al., 2008 for example of PCA based analysis). The reader is encouraged to review Sider et al. (2012) for factor loading and other technical information of the PCA analysis. For this study, the results from the factor analysis are directly employed. The variables from the large dataset were classified into two categories: (1) demand-side variables and (2) supply-side variables. For the demand-based category, three orthogonal factors were derived: (1) zones with high median income and high proportion of newer vehicles, (2) zones with high vehicle ownership and high proportion of larger vehicles, and (3) zones with large proportion of older vehicles. The supply-based variables provided three orthogonal factors: (1) zones with high density, high walkability, and transit oriented developments (TOD), (2) zones with commercial land-use, and (3) zones with government and institutional land-use. All of these six factors were considered in our model estimation.

The universal set of variables was considered in the estimation of the three components – latent segmentation sequence choice, mode choice and station choice. The variables were carefully chosen so as to match the sequence under consideration for each segment. The specification process was based on a systematic process of removing statistically insignificant variables. The process was guided by intuition and findings from earlier literature. The final estimation results are presented in Table 7.

Model Fit Measures

In order to adequately assess the explanatory power of our latent model, its log-likelihood was compared to the log-likelihood obtained from the sequential models run separately. The log-likelihood for separate MS, separate SM, and our latent segmentation model are -6,525.33, -6,198.65, and -5,603.00 respectively. The models are not nested within one another and hence we use the Bayesian Information Criterion (BIC) measure to examine statistical significance. The BIC is defined as - 2ln(L) + K ln(Q), where ln(L) is the log-likelihood value at convergence, K is the number of parameters, and Q is the number of observations. The model with the lower BIC value is the preferred model. The corresponding values for the separate MS, separate SM, and latent segmentation models are 13,094.65, 12,437.51, and 11,288.90 respectively. The BIC values clearly illustrate the statistical superiority offered by the latent segmentation model. At this juncture, it is important to note that the model which imposes SM sequence offers better fit compared to the model with MS sequence. Moreover, the improvement in the combined model indicates that the population share for SM sequence is far from negligible and requires a careful consideration for policy analysis. The results provide evidence to the presence of two hierarchies in the choice process and the benefit of considering these frameworks within the proposed latent segmentation model.

Model Parameters

The model is divided into three distinct components: the sequence choice component, the mode choice first and station choice second component, and the station choice first and mode choice second component. Please note that there are two versions of mode choice and station choice components. i. Sequence Component

The latent segmentation component examines the sequence of the decision process (i.e. whether the individual considers mode first station second sequence (MS) or the station first and the mode second sequence (SM)). The parameter results offer quite interesting insights on the decision process.

Prior to examining the individual parameters, we generated the overall aggregate share of the two segments in the population. We find that about 36% of the respondents’ share is allocated to the MS segment, while the remainder is allocated to the SM segment. The segment shares indicate that the population share for MS and SM sequences is substantial and requires a careful consideration for policy analysis. We find that for the MS segment, the mode shares are 42% (drive alone), 12% (shared ride), 12% (transit), and 34% (active) while for the SM segment, the mode shares are 62% (drive alone), 15% (shared ride), 12% (transit), and 11% (active). These results indicate that at the aggregate level there are significant differences in mode share across the two segment choices. The MS segment has a large share of respondents choosing active transportation while in the SM segment large share of respondents rely on automobile (drive or shared ride). Quite interestingly, the share of transit remains constant across the two segments.

The constant term indicates that given everything else remains the same the MS segment is more preferred relative to the SM segment (of course, the nature of the relationship might be skewed by the values of independent variables as is the case in our context where SM segment has 64% share). The inclination for workers to opt for MS segment indicates that workers are likely to decide on their mode decision first. This is probably a manifestation of the impact of a rigid schedule. .

Individuals with longer walk times to the nearest station prefer to make station decisions first. This attribute, used as a surrogate for proximity to station, suggests that the further you live to a station, the more likely you are to choose the station first. The result appears counter intuitive; however, the individuals that live far from any station are likely to be willing to consider more options for station. Hence, these respondents decide their station first and then choose their mode to

54 arrive there. Individuals leaving before 7.30 am are likely to be assigned to the SM segment. These respondents leaving before 7.30 might make their decision based on seat and/or parking availability and opting for the station first might allow them flexibility in this regard. ii. Sequence 1: Mode First Station Second

The MS segment considers mode choice as the first decision, followed by station choice. The mode choice and station choice parameters discussed below.

The constants of the mode choice model do not have any behavior interpretation because the alternative utility value is determined by the value of the independent variables in the model. Men are less likely to use the car alternatives for traveling to the station. This may be due to the fact that females are usually responsible for dropping off/picking up their children from day care or school and require a car. Household car ownership coefficient indicates that individuals from households with cars are unlikely to choose the transit alternative. The finding is contrary to earlier research (Debrezion, et al., 2009; Givoni & Rietveld, 2007). However, we believe it is more intuitive that individuals with car exhibit preference to drive (as indicated by the overall modal share).

The travel time variables offer interesting results. Two variables are significant in our analysis – average travel time to viable stations and travel time to the closest station. The former coefficient has a negative value while the latter coefficient has a positive value. The reader would note that the net effect is clearly negative because the numerical value of the travel time to closest station is always lower than the average travel time to viable stations. Hence, as expected in any mode choice decision, increasing travel times exerts an overall negative effect on the utilities of the modes (see Debrezion, et al., 2009; Givoni & Rietveld, 2007; Keijer & Rietveld, 2000 for similar results). The direction of the train trip significantly impacts mode choice. Specifically, individuals whose trips are in the direction of the CBD are more likely to choose a car mode to get to the station. Also, we see that individuals residing in areas with high factor values for government and institution

55 area are likely to opt for transit or active transportation alternatives. It is not surprising, as government and institutional land-use zones are usually well connected by transit and often have bicycling facilities. Finally, individuals residing in zones with high score for the factor high car ownership and high percentage of larger vehicles are less likely to use automobile alternatives (drive alone or shared ride). The result though counter intuitive at first glance might be a manifestation of higher active transportation usage in the MS segment – possibly because households while owning larger vehicle fleets are physically active and environmentally conscious.

In terms of the station choice parameters for the MS segment, the variable effects are intuitive. Travel time by the chosen mode has a significant negative impact on station choice. Further, car owners find travel time even more burdensome. As is expected, we observe that train frequency impacts station choice positively. The parking inventory variable has the expected result; higher parking inventory encourages use of the stations. The higher parking inventory serves as a surrogate for station attractiveness. The results are very similar to findings from earlier research (Debrezion, et al., 2007, 2009; Fan, et al., 1993; Wardman & Whelan, 1999). Of course, there is a clear case of endogeneity that we are not considering for this variable. It is entirely possible that these stations have high parking because of the high demand at these stations.

iii. Sequence 2: Station First Mode Second

The SM segment considers station choice as the first decision, followed by mode choice. The station choice and mode choice parameters are discussed below.

In this segment, the surrogate for travel time – considered as drive time to station – impacts the station choice negatively. It is observed that individuals prefer to choose stations that have higher factor values for commercial zones. These zones usually reflect central business districts, and the result is not surprising given the

56 higher accessibility by different modes to stations in these areas. The distance from station to CBD variable presents a negative value, highlighting a preference for stations that are in the direction of the final destination. The frequency and parking inventory variables offer very similar results compared to the station as second choice. The negative coefficient for parking between 1 and 200 can be explained by a better accessibility to some stations with 0 parking, as they can be located in denser and more accessible areas.

The mode second parameter estimates offer similar results as mode first, specifically males, travel time, and train direction. Younger users (25 and younger) have a higher likelihood of choosing the drive alone and shared ride modes in the decision processes. To understand this effect, it is important to note that the drive alone alternative was made available in the modeling procedure to individuals owning a car. As seen in the summary statistics in the above section, the proportion of younger individuals who owns a car is lower than that of all other age groups. Hence, younger individuals that own a personal vehicle are more inclined to use it. Further, most often these individuals do not share their car with others, as opposed to older individuals. Moreover, individuals who have indicated their egress mode as transit are more favorably inclined to use the latter in the access trip. Individuals choosing transit in the alighting trip indicates that they are not averse to using transit. In this segment, individuals residing in zones with high score for the factor high car ownership and high percentage of larger vehicles are more likely to use automobile alternatives. Finally, if the chosen station is located in an area with high factor values for government and institution area, the transit and active modes are less likely to be chosen. The result could be a manifestation of an overall preference for driving (about 77%) in this segment. b. Model Validation

To confirm the validity of the proposed model, we undertake a validation exercise using a hold-out sample not considered in the estimation process with 434 records. We compute the predictive log-likelihood measure for the MS separate

57 model, SM separate model, and latent segment model. To illustrate the strength of our approach, we also compute the predictive log-likelihood measure for various segments of the hold-out sample including gender, occupational status, age categories, mode choice, and LOS attributes. The results from the validation exercise are presented in Table 8. The table provides information on the sub-sample, sample size for the sub-sample under consideration and the predictive log-likelihood for the three models. The results clearly illustrate the increased predictive power offered by the proposed latent segment model. In fact, even for various segments, the latent segment model performs substantial better than the other two sequential approaches (one exception for drivers). Overall, the validation exercise confirms the improved predictive power of the proposed model. c. Policy Analysis

The model estimates provide useful insight on the impact of various parameters on mode and station choice. However, to highlight the utility of the proposed framework in evaluating the impact of various policy alterations, we undertake a series of sensitivity analysis. Specifically, we examine how mode and station choice decisions alter with changes to a set of the exogenous variables such as individual characteristics, LOS measures, land use attributes, and train frequency. The results of the computation are presented in Table 9.

We observe that being male strongly increases the likelihood of transit and active transportation use while being employed has a positive influence on active transportation usage. Car ownership has a very strong negative effect on choosing the transit mode. Reducing the drive time by 5 minutes has a significant negative impact on transit and active transportation. Reducing drive time by 5 minutes to a station significantly increases the odds of choosing that station. It is quite interesting to observe that improving transit service to reduce transit times by 5 minutes has a substantially large positive increase in transit usage. Reducing transit time by 5 minutes is relatively inelastic for station choice and for all modes, except transit share which would experience an increase of over 8%. The effect of a 5 minute

58 reduction in walk time impacts the automobile and transit modes negatively and active transportation very positively. Station choice remains inelastic for walking time. All results from land use changes are inelastic. Particularly, 15% increases in government and institutional area, high vehicle ownership/ high LTD share area, and commercial area almost all present changes of less than 1% in mode share and station choice. Increasing train frequency, as expected, provides increased station level usage with about an 8.9% increase per unit increase.

The policy implications of these findings are quite clear and provide straightforward interpretations. Car ownership plays a major role in transit use, but there can be important difficulties when creating and implementing a policy to reduce car ownership. From our results, the most effective way to increase transit mode share is to increase public transportation service and accessibility. To this end, reducing transit travel time to stations by expanding the feeder bus network or increasing bus frequency can be effective policies. On the other hand, mode and station choice do not seem to react to land use changes. Therefore, the priority for transit agencies in the upcoming years should be to provide stronger transit accessibility to train stations all the while increasing train frequency to attract more users. Also, agencies should consider potential policies that penalize car ownership to reduce vehicle ownership levels to increase transit usage.

3.4 Conclusions

This chapter contributes to literature on transit choice behavior by examining the travel decision framework of commuter train users in Montreal. Its commuter train system provides access to the urban population from the suburbs to the central business district of Montreal. In this research, we examine the behavior of the commuter train riders in terms of their commuter train station and travel mode to commuter train choices (access mode). The focus of the analysis is on developing a behaviorally representative framework for understanding the decision processes involved in the station and access mode choice.

Typically, mode and station choice for commuter train users is modeled as a hierarchical choice with mode being the first choice in the hierarchy. The current study proposes a latent segmentation based approach to relax the hierarchy. In our approach, we simultaneously consider the two possible sequences while also determining the individual allocation each segment as a function of socio- demographic variables, LOS parameters, trip characteristics, land-use and built environment factors, and station characteristics. The proposed model is estimated using an onboard survey conducted by the AMT for commuter train users. The information compiled includes individual and household socio-demographics such as age, gender, vehicle ownership, and occupational status. Also included are residential location, boarding and alighting commuter stations, final destination location, travel mode to the boarding commuter train station and from the alighting commuter train station, and travel departure times. The survey database was appropriately augmented with the LOS attribute database generated as well as parking data for each commuter train station.

A host of variables were considered in our analysis. In the sequence model, the results indicate that as the distance from the station by active forms of transportation increases, individuals are less likely to select a station first. This suggests a form of stickiness to station for people living very close to the station. For the mode choice model, an inclination exists for workers and individuals leaving before 7:30 am to opt for mode first. Young persons, females, car owners, and individuals leaving before 7:30 am have an increased propensity to drive to the commuter train station. Travel time by mode has a negative impact on the mode choice. The station model indicates that travel time by the chosen mode has a significant negative impact on station choice. Finally, the presence of parking encourages use of the stations; not surprising because parking spots at a station is a function of parking demand. Frequency also has a positive impact on station choice. The model developed was used to undertake a validation exercise on a hold-out sample. The latent segment model’s performance was compared with the two sequential models. The results clearly highlight how the latent segment model

60 outperforms the two sequential models. The model estimates are employed to undertake policy analysis highlighting the role of gender, car ownership, train frequency and level of service in mode and station choice decisions.

This paper is not without limitations. The study does not account for various station specific attributes such as percentage of seat availability and parking availability (different from inventory). Individuals might consider driving longer for these reasons. However, data collection for such attributes is rarely considered. Transit agencies need to enhance their database inventory to better understand transit choice behavior. Further, a joint choice of mode and station choice with a large alternative set within the latent segmentation context is a possible direction for future research.

3.5 Tables and Figures

Table 6: Summary Statistics

Summary Statistics (N=4,346) (%)

Drive 51.1 Passenger 12.5 Transit 11.3 Active (Walk/Cycle) 25.1 Gender

Males 43.6 Females 56.4 Age

≤25 19.8 26-39 34 40-54 37.2 ≥55 9

Car ownership (Yes/No) * Age groups ≤25 57.5 26-39 92.6 40-54 92.2 ≥55 88.2 Status

Worker 81.8 Student 17.6 Other 0.6 Time left home

Before 6:30 am 18.2 Between 6:30 am and 7:30 am 47.6 After 7:30 am 34.2 Station Characteristics

Distance to stations (km) Average distance to nearest station 3.6 Average distance to chosen station 4.9 Number of stations with:

No parking at station 14 Between 0 and 200 spots 12 Between 201 and 500 spots 12 Between 501 and 700 spots 7 Between 701 and 1000 spots 4 More than 1000 spots 3

Table 7: Latent segmentation based mode and station choice model results

Variables Sequence Choice

MS SM Coefficient t-stat Coefficient t-stat Constant 0.788 4.30 - - Individual Demographics Worker 0.407 2.80 - - LOS Measures (in hours) Walk time to closest station -2.311 -10.11 - - Trip Characteristics

Time left home before 7:30 am -0.415 -3.47 - - Mode First Mode Second Coefficient t-stat Coefficient t-stat Drive alone constant -7.421 -4.24 -1.329 -1.53 Shared ride constant -9.039 -5.15 -3.431 -3.94 Transit constant -1.437 -4.44 0.060 0.37 Individual Demographics

25 years old and younger

-Drive alone - - 0.809 5.43 -Shared ride - - 0.809 5.43 Male

-Drive alone -0.618 -1.89 -0.719 -5.50 -Shared ride -0.618 -1.89 -0.719 -5.50 Car ownership

-Transit -2.268 -5.27 - - LOS Measures (in hours)

Average travel times to viable stations -20.927 -10.46 - - Travel time to closest station 16.158 8.08 Total travel time to chosen station - - -0.342 -2.81 Trip Characteristics

Egress mode is transit

-Transit - - 0.633 4.11 Trip is in the direction of CBD

-Drive alone 3.692 2.18 3.149 3.68 -Shared ride 3.692 2.18 3.149 3.68 Land-use and Built Environment Factors

Government and institutional areas (at origin) - Transit 0.889 2.00 - - -Active transportation 0.889 2.00 - - Government and institutional areas (at station) -Transit - - -0.103 -3.54 -Active transportation - - -0.103 -3.54 High vehicle ownership/High LTD share

-Drive alone -0.919 -2.75 0.655 5.43 -Shared ride -0.919 -2.75 0.655 5.43 Station Second Station First Coefficient t-stat Coefficient t-stat LOS Measures (in hours)

Drive time to station - - -7.171 -14.58 Travel time with chosen mode to station -6.448 -3.75 - - Travel time for car owners -2.339 -6.14 - - Station Service

Train frequency 1.573 6.74 0.163 6.40 Distance from station to CBD - - -0.037 -4.20 Land-use and Built Environment Factors

Commercial area - - 0.314 2.66 Parking Inventory (0 parking is base)

Between 1 and 200 parking spots 3.697 3.79 -1.300 -5.14 Between 201 and 500 parking spots 1.361 1.54 1.082 5.26 Between 501 and 700 parking spots 3.261 3.34 0.935 4.36 Between 701 and 1000 parking spots 3.815 3.88 1.204 5.40 More than 1000 parking spots 4.051 4.16 2.207 9.71

Table 8: Predictive log-likelihoods for the three models

Segments N MS separate SM separate Latent Full validation sample 433 -1,441.1 -719.7 -690.6 Male 172 -306.2 -305.3 -239.1 Workers 364 -593.6 -600.9 -505.4 Age≤25 114 -170.4 -176.1 -150.2 Drivers 218 -259.5 -259.7 -210.5 Transit Users 51 -135.8 -148.3 -132.5 Active transportation users 109 -159.8 -166.9 -118.0 Drive time to station <10 298 -407.3 -424.7 -365.5 Drive time to station 10 - 20 107 -168.7 -165.9 -168.2 Drive time to station > 20 28 -126.5 -129.1 -64.1 Walk time <10 minutes 85 -30.8 -35.1 -26.7 Walk time <15 minutes 46 -82.1 -88.8 -72.0 *The underlined number indicates the best fit log-likelihood value within the three models

Table 9: Elasticity Analysis

Mode Choice Station Choice Drive Alone Shared Ride Transit Active

Individual Characteristics - Male -10.37 -19.28 35.57 15.94 - Car Ownership 4.49 7.12 -50.71 10.45 - Worker -3.46 -3.22 -0.43 9.41 -0.26

Travel Time Drive time (-5min) 1.74 2.73 -4.51 -3.09 15.35 Transit time (-5min) -0.74 -1.36 8.37 -1.63 0.61 Walk time (-5min) -2.71 -2.84 -3.50 9.02 0.21

Land Use Gov Area (+15%) 0.10 0.18 -0.26 -0.18 - High LTD/Veh (+15%) 0.38 1.02 -1.76 -0.53 - Commercial Area (+15%) - - - - -0.28

Frequency +1 - - - - 8.87 +2 - - - - 16.14

Figure 3: Home location of survey sample

CHAPTER 4. URBAN ENVIRONMENT AND BUS STOP RIDERSHIP

4.1 Introduction

In the previous chapters, we focused on travelers’ behavioral characteristics associated with transit ridership - individual level analysis. The present chapter will draw attention to the spatial characteristics affecting ridership – at a transit system level. The emphasis is on a systems perspective where transit ridership is studied from the perspective of the transit provider. Specifically, we explore the influence of urban form and land use factors on bus ridership at the stop level in Montreal.

The current study’s objective is quantifying the influence of transit system operational attributes (such as headway, transit accessibility indices), transportation system infrastructure attributes (such as road network characteristics, bike lanes) and built environment attributes (such as presence of parks, residential area) on the disaggregate stop level boardings and alightings by time of day for the bus network in the Montreal region. To be precise, the emphasis is on the quantification of the influence of various attributes on boardings and aligthings using econometric models. The results will provide transit agencies a mechanism to study the influence of transit accessibility, transit connectivity, transit schedule alterations (to increase/reduce headway), and land-use pattern changes on ridership. .The framework developed can be applied to predict ridership at potential new stop locations. Moreover, the ridership information at stop level by time of day provides the transit agency an effective mechanism to predict transit bus occupancy - an important measure for vehicle fleet allotment for various lines.

There is emerging recognition on quantifying the influence of built environment, transit and transport infrastructure on transit usage. However, a majority of the studies exploring this relationship focus on the rail or metro mode. The analysis for the rail and metro modes are computationally less intensive as the number of stations rarely amount to more than 50 in an urban region. The focus of this paper is to examine this relationship specifically for the bus mode, at the bus stop level. In our empirical context, we focus our evaluation on a transit stop system of about 8000 stops.

As discussed in the literature review section, a number of studies explored the association between built environment and bus ridership, but have either considered daily ridership as a sum of boardings and alightings or analyzed boardings only (Chu, 2004; Estupiñán and Rodríguez, 2008; Johnson, 2003; Ryan and Frank, 2009). The analysis is adequate for an overall picture of transit ridership in the region but is inadequate to comprehensively examine the influence of various attributes highlighted earlier. Moreover, to draw any conclusions on vehicle fleet decisions a daily ridership measure is inadequate.

Of course, incorporating the stop level boardings and alightings along various time periods provides us with unique challenges of its own. For instance, the consideration of four time periods for boardings and alightings result in eight dependent variables for each stop. It is important not only to consider different time periods in the analysis, but to assess the possible unobserved interactions between them as well. The dependent variables are all reported for the same stop and hence are likely to be affected by common unobserved factors. Our analysis quantifies the dependencies between the eight dependent variables using an innovative Composite Marginal Likelihood (CML) method that has recently been employed in transportation literature (Ferdous et al., 2010, 2011; Seraj et al., 2012; Sidharthan et al., 2011).

The current study is implemented in two stages. In the first stage we undertake an exploratory analysis of the boarding and alighting data in the Montreal urban region. As a part of this exercise, we generate a visual representation of the bus ridership for different time periods of the day. The second stage analysis contributes to existing literature by quantifying the effect of the land use, transit infrastructure and transportation infrastructure attributes on bus ridership using CML ordered probit models.

4.2 Data and Methodology a. Data

The data employed in this study is drawn from data collected by STM. Approximately 15% of STM bus fleet is equipped with infrastructure that counts boardings and alightings with specific information, such as the location, time of day, and bus number. The sampling procedure allows us to obtain an accurate average of ridership for each bus stop across the Island for a typical weekday. STM has also provided data on bus frequency for each bus stop for all time periods.

The original data has been processed in order to generate total ridership for each bus stop by time period. The dependent variable data compiled for the purpose of this analysis consists of bus boarding and alighting for different time periods for about 8000 bus stations across the Island of Montreal. The time periods considered in our analysis (as provided in the data compiled) are the am peak (6:30 – 9:30), pm peak (15:30 – 18:30), off peak day (9:30 – 15:30), and off peak night/morning (18:30 – 6:30). The average boarding and alighting numbers per bus stop for the entire day amount to 110. The corresponding values for various time periods are: (1) am peak period – 28, (2) off peak day – 35, (3) pm peak period – 28 and (4) off peak night – 20. Across the 8000 stops the ridership varies significantly from 0 to 8000 riders. To accommodate for the large variability in boarding and alighting at the stops, we categorize the various stops into three groups – low, medium, and high ridership. The demand profile for the stops in the three groups is expected to be very different and warrant group specific analysis. The thresholds for the low, medium, and high groups are less than 50, 50 – 250, and more than 250 respectively (boarding + alighting for 24 hours). As you would expect, the finalized groups have the largest sample of stops in the low category (3574), and the lowest sample of stops in the high category (1813). b. Variables Considered

The analysis quantifies the influence of various exogenous factors including stop level transit operational variables (average headway for time period, number of lines passing through the stop, night bus passes through stop), public transit accessibility indices (number of bus/metro/train stops around each stop, length of bus/metro/train lines,

68 length of exclusive bus lanes), transportation infrastructure attributes (road length by functional classification, bike lane lengths, distance to central business district, CBD), and stop level land use measures (number of parks and their areas, residential area, number of commercial enterprises and their area, government and institutional area, resource and industrial area, employment density, walkscore). The attributes highlighted are computed for various buffer sizes (200m, 400m, 600m, 800m, 1000m) and for Traffic Analysis Zones (TAZ) drawn around the bus stop using Geographic Information Systems (GIS). c. Summary Statistics

Table 10 presents some summary statistics for all variables used in the models. All lengths and areas are in kilometers and kilometers squared respectively. A large variance exists between average boarding and alighting for the different stop categories and time periods, confirming the necessity to analyze them separately. Naturally, more lines pass through high ridership stops than medium or low, on average. Moreover, average headway for a time period varies from 10 minutes to 100 minutes; peak periods and high ridership stops have lower headway. The number of bus stops and metro stations in different buffer sizes consistently decrease from higher ridership to lower ridership stops. Unsurprisingly, the total length of bus routes in a 600 meter buffer around the stops decreases from the higher to the lower ridership categories, but the opposite can be observed for the same variable at a TAZ level. The explanation is quite simple: TAZ size varies throughout the Island of Montreal, where larger TAZs are generally located far from the city center, where higher ridership exists. The bus route length will be higher in the larger TAZs not because of actual service length, but rather because of the area analyzed. The nature of TAZ variables – with impact of land area - necessitates the adoption of buffer level variables wherever possible. The same logic applies to train line length in TAZ, while metro line length in TAZ decreases since metros are only present close to the city center. Finally, on average, high ridership stops are located in areas with more reserved bus lanes.

The length of major roads and bicycle paths around the stops decreases for lower categories, whereas length of highway remains relatively constant. The number of parks

69 and commercial enterprises and their respective areas decrease for lower categories, while government and institutional area, residential area, park and recreation area, and resources and industrial area all increase for lower ridership stops. Once again, the size of the TAZs has a large role to play in these values. d. Visual Analysis

The visual representation of the bus ridership is generated for 4 categories, namely for boardings and alightings for AM and PM peaks. To easily represent the transit ridership origin and destination in the urban region, the hourly ridership was illustrated using the kriging function in GIS, an interpolation technique in which the surrounding measured ridership values are weighted to derive a predicted value for an unmeasured location.

Prior to analyzing individual maps, some general trends are discussed briefly. First, we notice that for all time periods, some areas always have a high ridership. This is explained by the presence of a bus terminal or a metro station in that area - transfer points that attract higher demand particularly because of high number of bus lines and bus stops. In fact, we notice a consistently greater ridership along the metro lines. Second, some areas and neighborhoods in Montreal have generally lower ridership. This is especially true for the West Island (the left-most part of the Montreal island in Figure 1), an area in which public transportation services are generally lower than that of the rest of the city.

Figure 4 presents a visual depiction of bus boarding and alighting in Montreal for the 2 time periods. These maps clearly show similar ridership patterns for AM Boardings and PM Alightings, as well as for AM Alightings and PM boardings. These trends can be simply explained with individuals boarding buses in residential areas and alighting in the city center or near the workplace in the morning and the opposite occurring in the afternoon. On one hand, the AM Boardings/PM Alightings are characterized with high ridership in areas further from the center of the city, which are mostly considered as residential areas. On the other hand, the AM Alightings/PM Boardings present high ridership around transit infrastructure, such as along metro lines or near train stations.

70 e. Methodology

The Composite Marginal Likelihood (CML) for ordered response probit (ORP) model is employed to examine the effect of exogenous variables on ridership at bus stops. This model allows observing possible correlations between boardings and alightings for the multiple time periods. For instance, we might observe that boardings in the AM peak are positively correlated with alightings in the PM peak.

Let q (q = 1, 2, …, Q) be an index to represent bus stops, i (i = 1, 2, 3, …, I) be an index to represent boarding/alighting – time period combinations, where I=8. Then, let the ridership interval value for combination i be Ki + 1 (i.e., the discrete levelsbelong in {0, 1, 2, …, Ki} for category i). The index k takes value of ridership intervals such as “Alighting per hour between 0 and 10” (k=1), “Alighting per hour between 10 and 20” (k=2), etc. The intervals vary for each group of models, namely for each combination of ridership (alighting, boarding) and ridership level (high, medium, low). The equation system for the standard ordered response model is:

k * k 1  i  y qi   i (1)

where corresponds to the latent ridership propensity for a stop q. xq is an (L × 1)- column vector of built environment attributes: stop level variables, public transportation accessibility indices, infrastructure attributes, and land use measures

k for a stop q. is the corresponding (L × 1)-column vector of variable effects. θi is the lower bound threshold for ridership category k of combination i (

0 1 2 K i 1 0 K i 1  i   i   i ...   i ;  i   ,  i   for each category i). The model structure requires for the θ thresholds to be strictly ordered in order to adequately distribute the latent ridership propensity in the observed ridership categories.

Finally, ɛq is an idiosyncratic random error term that impacts ridership propensity, which may include the presence of a bus shelter at stop q. The ɛqi terms are assumed independent and identical across stops (for each and all i). For identification reasons, the variance of each ɛqi term is normalized to 1. However, we allow correlation in the ɛqi terms across combinations i for each stop q. Specifically,

. Then, ɛq is multivariate normal distributed with a mean vector of zeros and a correlation matrix as follows:

0 1 1 2  1 3  1 I  0 2 11  2 3  2 I  ~N  , , o r q   0    1 III1 2 3 (2)

 N  0, Σ  q ~

The off-diagonal terms of Σ capture the error covariance across the underlying latent continuous variables of the different combinations; that is, they capture the effect of common unobserved factors influencing the propensity of

ridership at bus stops. For example, if  12 is positive, it implies that boardings in the AM peak period for a stop q will likely be positively correlated with boardings in the PM peak. Of course, if all the correlation parameters (i.e., off-diagonal elements of Σ), which we will stack into a vertical vector Ω, are equal to zero, the model system in Equation (1) collapses to independent ordered response probit models for each ridership category.

Given the preliminaries above, we employ a pairwise marginal likelihood estimation approach, which corresponds to a composite marginal approach based on bivariate margins (see Ferdous et al., 2010, Varin and Czado, 2008; Apanasovich et al., 2008; Varin and Vidoni, 2008; and Bhat et al., 2009 for the use of the pairwise likelihood approach in the past). The pairwise marginal likelihood function for station q may be written as follows:

II1 L( ) P r( y  m , y  m ) CMLq,  qi qiqgi qg i11 g  i 

mq g1 m q g  1 m q i  1 m q g II1   x ,     x ,        x ,     x ,  22 i iqig gqgig  i iqig gqgig    , (3) m m1 m m i11 g  i   q i  x ,  q g    x ,     q i    x ,  q g    x ,  22 i iqig gqgig  i iqig gqgig 

L ( )  L ( ) and CML  CML ,q (4) q

The above expression just requires evaluation of Bivariate normal probabilities and can be computed at a high level of precision. The estimates obtained by maximizing the logarithm of the above function are consistent and asymptotically normal distributed (see Ferdous et al., for more details on inference metrics).

4.3 Results

The empirical analysis in the study involves estimating the effect of the built environment and urban design on ridership at a stop level using an ordered regression model. We are examining the ridership in different dimensions. First, we are analyzing three categories of stops; high, medium, and low ridership. For each of these, hourly boardings and alightings are modeled separately. Further, each time period (am peak, pm peak, off peak day, off peak night) is considered. Therefore, we are estimating an 8 dimensional multivariate ordered probit model using the CML approach for each category (low, medium and high). The final specification was based on a systematic process of removing statistically insignificant variables. Table 18 located in the appendix, presents the categories (in hourly rates) employed for the different models

Prior to examining each category in more detail, some general trends that appear throughout the models will be discussed. First, we notice that in each category, the AM Boarding and PM Alighting models have similar specifications. The same applies for PM Boarding and AM Alighting. In each case, both models present similar significant variables with the comparable effects. Evidently, they capture the morning and afternoon commute impacts. This is along expected lines because an individual boarding at stop A near his residence in the morning is likely to alight at that same stop A in the afternoon.

73 a. High Ridership Stops

Table 11 provides the final model specification for the “high” category. The model results presented include a column for each time period. Each row represents the impact of an exogenous variable (“-“ indicates no significant variable impact).

The headway (in minutes) has a negative and very significant effect for all ridership models. In other words, stops with higher frequency have higher ridership. The presence of public transportation around the stop has a positive and significant effect on ridership. This holds true especially for presence of bus stops and metro stations in a 200 meter buffer, effectively showing that most high ridership stops are located in an area with substantial public transportation facilities. The number of surrounding train stations has an effect only on AM Boarding, suggesting that individuals’ board at high ridership stops after traveling by train in their morning commute. Specifically in the context of Montreal, this most likely represents individuals boarding buses at stops near the central station, where the largest train station is located. Overall, these high ridership stops seem to be transfer points, close to metros and located in areas with extensive public transportation facilities.

The presence of major roads around the stop exerts a positive effect on ridership and is significant only for Off Peak Night Boarding and AM Alighting. This may be because of the location of transit on major roads. The length of highways in an 800 meter buffer exerts the opposite effect, indicating that stops in the vicinity of highways are more likely to have fewer riders. Again, this effect is only significant for Off Peak night Boarding. Finally, the further the stop is to the CBD, the fewer alightings are likely to occur for the Off Peak Night period.

The variables capturing the presence of parks offer interesting results. The area of the parks around the stop has a significantly negative effect, whereas the number of parks exerts an opposite effect. This suggests that ridership is likely to be higher in an area with several parks of small dimensions, as the walkability of the area would benefit from the presence of parks without constraining road areas for transit to operate. Nevertheless, the net effect is positive overall. To demonstrate this overall positive effect, the average park area in a 600 meter buffer for the “high” category is 0.086 km2,

74 and the average number of parks for the same buffer size is 8.41. Therefore, in the AM Boarding, the overall park effect can be calculated as -0.632*0.086 + 0.014*8.41 = 0.0633. There is a similar equilibrium effect between the number of commercial enterprises and their area. In fact, their interaction results in an overall positive manner, effectively demonstrating that stops in these areas are more likely to have high ridership. Government and institutional area near the stop is likely to increase the ridership, notably for the AM Alighting time period. The presence of residential area exhibits expected trends. Specifically, higher residential area implies lower PM boarding and lower AM alighting illustrating the presence of the commuting pattern - individuals alight buses in the morning and board them in the afternoon near their workplace. We observe that the employment density at the TAZ level exerts a negative effect on boardings and the opposite effect on alightings. Finally, the resources and industrial area exerts a negative effect on ridership, particularly on boarding. b. Medium Ridership Stops

Table 12 is the final model specification for the “medium” category, for which all stops have a total ridership (boarding + alighting) of between 50 and 250 for a given day.

The headway variable has the same effect for the medium category as the high one. However, the number of lines affects the ridership negative, most notably for AM Boarding and PM Alighting. Although this may seem counterintuitive, it actually illustrates the competition between different bus lines passing through the same stop. To understand this, we must remember that the headway was obtained by adding the frequency of all bus lines for a stop. For equal frequency, stops with multiple lines will effectively have a lower frequency per bus line. For example, stop A has a 10 minute headway and 2 bus lines passing through it; the effective headway, assuming equal frequency for all lines, will be 10*2 = 20 minutes for each bus line.

The effect of transit for medium ridership stops is not as straight forward as the high ridership stops. The presence of bus stops around the stops (600m and 800m radii) impacts the ridership in a positive manner for AM Boarding and PM Alighting, while total bus line length in the TAZ exerts the opposite effect for these same time periods. Similar to the contrasting park effects in the high category, the presence of buses (line length

75 and stops) has an overall positive effect on ridership. The presence of bus stops stays insignificant, however, for the PM Boarding and AM Alighting, for which the total bus line length in the TAZ affects ridership positively, resulting once again in an overall positive effect for presence of bus. Train line length at the TAZ level has a negative effect on ridership, principally on AM Boarding and PM Alighting, while the presence of train stations in the vicinity of these stops are likely to increase ridership for the PM Boarding and AM alighting. This indicates that these stops serve as transfer points for commuter trains. Overall, the medium ridership stops seem to be transfer stops for trains as well as residential stops in somewhat transit accessible areas.

Presence of major roads around a stop is likely to increase ridership for PM Boarding and AM Alighting, whereas the distance to CBD impacts ridership in a negative manner for these same models. Highway length in an 800 meter radius exerts a negative impact on patronage for the PM and Off Peak Day time periods. Finally, an increased presence of bicycle paths has a positive effect for AM Boarding and PM Alighting. Again, all these results are intuitive, when considering the effects of commuting patterns.

The land use variables also clearly demonstrate these commuting patterns. The ridership for AM Boarding and PM Alighting are positively affected by the number of parks and their area as well as the residential area, and negatively affected by the number of commercial enterprises near the stop. The opposite is also true for PM Boarding and AM Alighting, for which stops located in residential areas are less likely to have high ridership. c. Low Ridership Stops

Table 13 is the final model specification for the “low” category, for which all stops have a total ridership (boarding + alighting) of less than 50 for a given day.

Once again, bus headway at stops impacts ridership negatively. It is important to note that a stop through which a night bus passes has no significant effect on the ridership, most notably for the Off Peak Night periods.

The public transportation infrastructure for low ridership stops has similar effect to the previous ridership models. For instance, the number of bus stop in the vicinity has

76 a positive and significant effect on ridership, which indicates that there is higher ridership in more transit accessible area.

Generally, the presence of major roads impacts the ridership negatively, whereas the presence of highways has the opposite impact. Moreover, the presence of bicycle paths is likely to increase the ridership. It is however not significant for PM Boarding and negative for AM Alighting. The ridership for these two categories is also negatively affected by the distance to CBD.

The presence of parks (number and area) has the same overall positive effect from the previous models. The residential area mostly has a positive effect on ridership, except for PM Boarding and AM Alighting models, demonstrating once again that these stops are mostly situated away from areas in which housing predominates. This is also attested with commercial areas, resource and industrial, job density, as well as government and institutional areas, exerting a negative effect on ridership, except for these two same models. d. Correlation Matrices

Table 14 provides the correlation matrix for the eight dimensions of the high ridership stop models, where values of 0 represent an insignificant correlation effect. All the non-zero elements in the Table are statistically significant at the 95% level. We notice that boardings for all time periods are positively correlated to each other (top left corner of Table 14), as are the alightings (bottom right corner of Table 14). The AM Boardings have a negative correlation with alightings for the same time period, whereas the PM Boardings and AM Alightings have the opposite relationship indicating that unobserved factors that result in an increase in boardings are likely to contribute to a reduction in alightings. Finally, the results indicate that ridership in Off Peak Day and Off Peak Night time periods also exhibit significant dependencies. These results clearly highlight the presence of unobserved dependencies across the eight dependent variables for each stop. Ignoring the presence of such unobserved dependencies would result in incorrect estimates for the observed variables.

Table 15, which presents the correlation matrix for medium ridership stops, offers similar results to the high stops. In fact, boardings for all time periods have a positive relationship to each other, just like the alighings. However, all correlations between boardings and alightings are significantly negative, suggesting that medium ridership stops serve either as a boarding stops or an alighting stops.

Table 16 presents the results of the correlation matrix for the low ridership stops. Once again, boardings and alightings are positively correlated to each other. However, the correlation between boardings and alightings are either positive or insignificant, with the exception of Boarding PM and Alighting OPN.

e. Elasticity Analysis

In order to highlight the effect of various attributes, an elasticity analysis for high ridership stops was conducted for both boardings and alightings for the peak periods and presented in Table 17. Specifically, we are calculating the change in ridership for changes in transit and land use attributes. To present the elasticity in an easier to understand format, we present the change in ridership levels by computing the product of change in probability of ordered alternative with the median value of the ordered category. To provide a sense of the resulting changes based on the proposed elasticity scenarios, the average ridership per hour for high stops is included in the Table. Several observations can be made from the results presented in Table 17. First, we notice that the transit accessibility and service attributes (headway and number of bus and metro stops) have a stronger influence on boardings compared with land use attributes (job density, residential area, and commercial area). Second, increasing headway, which translates into a decline in service, will result in a decrease in ridership as expected. However, the effect of the change on Boarding AM and Alighting PM (typically the destination end of the trip) is more pronounced compared with the Boarding PM and Alighting AM. The results indicate that ridership is more sensitive to headway change in the direction of commute. Third, the addition of a metro station has much larger influence on ridership relative to the addition of bus stops. This is not

78 surprising as the cost of adding a metro stop is substantially larger than the cost of adding a bus stop.

The policy implications of these findings are quite clear and provide straightforward interpretations. From our results, it is clear that the most effective way to increase is to increase public transport service and accessibility. Since ridership does not seem to react to land use changes, the main priority for these agencies should be to expand their network. One of the priorities for the STM in the upcoming years is to extend the metro network to the east. Our study findings provide evidence that expanding the network is likely to increase bus ridership. Moreover, our approach can be applied to calculate expected ridership with new stops.

4.4 Conclusion

In this chapter, we examine the influence of the urban form and land use factors affecting bus ridership at the stop level by time of day in Montreal. The data employed in our study was drawn from data collected by the STM consisting of counts of boardings and alightings at each bus stop in the public transit network of Montreal. The time periods considered in our analysis were the am peak (6:30 – 9:30), pm peak (15:30 – 18:30), off peak day (9:30 – 15:30), and off peak night/morning (18:30 – 6:30). The various stops were categorized into three groups – low, medium, and high ridership – to accommodate for the large variability in ridership for different stops. The exploratory analysis through visual representation allowed us to observe the following ridership characteristics. Similar ridership patterns exist between AM Boardings and PM Alightings, as well as between AM Alightings and PM boardings. These trends can be simply explained with individuals boarding buses in residential areas and alighting in the city center or near the workplace in the morning and the opposite occurring in the afternoon. On one hand, the AM Boardings/PM Alightings are characterized with high ridership in areas further from the center of the city, which are mostly considered as residential areas. On the other hand, the AM Alightings/PM Boardings present high ridership around transit infrastructure, such as along metro lines or near train stations.

The empirical analysis in the study involves quantifying the effect of the built environment and urban design on ridership at a stop level using a CML ordered probit model. The analysis considers a host of exogenous factors including public transit infrastructure and accessibility indices, infrastructure attributes, and land use factors. We analyzed boardings and alightings for three categories of stops - high, medium, and low ridership stops - for four time periods -am peak, pm peak, off peak day, off peak night, estimating a total of 3*2*4 = 24 models.

Transit facilities (such as presence of metro stations, bus stops, and reserved bus lanes) and the presence of parks have a positive impact on ridership, while presence of highway has a negative impact. The impact of certain land use indices (commercial area, government and institutional areas, and residential areas) is temporally dependent. The results from the correlation estimates highlight the intricate nature of unobserved factors affecting boarding and alighting across various time periods. The elasticity analysis undertaken provides useful insight. Specifically, we observe that the most effective way to increase ridership is to increase public transport service and accessibility, whereas changes in land-use are very inelastic.

The research is not without limitations. We recognize that capturing the effects of the urban design is a delicate process, which can occasionally provide results which are difficult to explain. Further research can be carried out to develop more comprehensive set of land use variables in order to model ridership adequately.

4.5 Tables and Figure

Table 10: Summary Statistic

Average High Medium Low N 1813 3350 3574

Boardings per hour AM Peak 31.62 6.15 0.94 PM Peak 36.59 4.36 0.79 Off Peak Day 21.48 3.06 0.44 Off Peak Night 9.18 1.18 0.18

Alightings per hour AM Peak 34.74 4.54 0.89 PM Peak 32.95 6.15 0.99 Off Peak Day 21.40 3.12 0.50 Off Peak Night 8.49 1.47 0.26

- Stop level variables Number of lines passing through stop 2.19 1.75 1.37 Night bus passes through stop 0.46 0.26 0.16

- Transit accessibility indices

Number of bus stops in a 200m buffer 6.67 5.33 4.49 400m buffer 17.17 14.97 13.12 600m buffer 33.80 30.24 25.88 800m buffer 56.59 51.23 43.01 1000m buffer 85.03 77.11 63.91

Number of metro stops in a 200m buffer 0.17 0.05 0.03 400m buffer 0.25 0.16 0.15 600m buffer 0.40 0.33 0.30

Number of train stations in a 200m buffer 0.03 0.01 0.01 400m buffer 0.05 0.03 0.04 1000m buffer 0.18 0.17 0.19 Number of train stations in the TAZ 0.24 0.21 0.31

Bus line length in a 600m buffer 15.92 13.52 12.39 Bus line length in the TAZ 6.72 7.49 10.09 Metro line length in the TAZ 0.11 0.08 0.08 Train line length in the TAZ 0.80 1.69 2.79

Reserved bus lane length in a 200m buffer 0.14 0.05 0.03

- Infrastructure attributes

Major roads length in a 400m buffer 2.25 1.84 1.80 600m buffer 4.65 3.99 3.91 800m buffer 7.85 7.02 6.82

Highway length in a 800m buffer 2.48 2.35 2.74

Bicycle length in a 400m buffer 1.17 1.13 0.98 600m buffer 2.65 2.45 2.05 1000m buffer 7.26 6.61 5.40

- Land use measures

Park area in a 200m buffer 0.007 0.007 0.006 400m buffer 0.035 0.033 0.028 600m buffer 0.086 0.079 0.064

Number of parks in a 200m buffer 1.31 1.19 0.97 400m buffer 3.99 3.69 2.84 600m buffer 8.41 7.65 5.68 800m buffer 14.50 12.85 9.68 1000m buffer 22.04 19.52 14.91

Number of commerces in a 200m buffer 49.93 33.04 20.17 400m buffer 150.45 105.26 79.73 600m buffer 306.80 222.09 170.59 800m buffer 507.17 377.38 293.70 1000m buffer 760.67 570.63 444.89 Commercial area in the TAZ 0.027 0.021 0.022 Governmental and institutional area in 0.044 0.046 0.073 the TAZ

Residential area in the TAZ 0.295 0.396 0.510 Park and recreational area in the TAZ 0.056 0.065 0.090 Resources and industrial area in the 0.075 0.149 0.329 TAZ *For the sake of brevity, only the attributes significant in the empirical analysis are shown in the summary statistics.

Table 11: Ordered probit models for the High ridership category

Boarding Alighting Am peak Pm peak Off peak day Off peak night Am peak Pm peak Off peak day Off peak night B (t-stat) B (t-stat) B (t-stat) B (t-stat) B (t-stat) B (t-stat) B (t-stat) B (t-stat) - Stop level variables Lines through stop 0.139 (3.73) 0.17 (4.73) 0.118 (3.16) Night bus through stop 0.387 (4.34) 0.575 (6.45) - Transit indices Bus stops in a 200m buffer 0.083 (5.58) 0.112 (7.54) 0.08 (5.43) 0.109 (7.2) 0.043 (2.99) 0.033 (2.25) 0.047 (3.17) 0.051 (3.52) Metro stations in a 200m buffer 0.548 (4.28) 0.851 (6.16) 0.412 (2.53) 1.345 (9.41) 0.738 (5.88) 0.298 (2.15) 0.861 (6.61) 0.675 (5.36) Train stations in a 200m buffer 0.928 (3.32) Reserved bus lane length in 200m buffer 0.724 (4.58) 0.653 (4.19) 0.384 (2.37) 0.796 (5.06) 0.372 (2.29) 0.394 (2.51) Metro line length TAZ 0.474 (2.16) Train stations TAZ 0.327 (3.51) - Infrastructure Major roads length in a 400m buffer 0.079 (2.63) 0.105 (2.99) 0.128 (4.33) 0.072 (1.99) 0.098 (2.8) Hway length 800m buffer -0.038 (-2.3) -0.037 (-2.31) -0.064 (-3.48) -0.034 (-2.11) -0.076 (-4.02) -0.049 (-2.73) Straight line distance to CBD -0.019 (-2.24) - L-U measures Park area in a 200m buffer -8.222 (-2.16) 600m buffer -1.378 (-2.27) -1.279 (-2.1) Parks in a 200m buffer 0.066 (2.45) 600m buffer 0.021 (2.79) 0.022 (2.88) 0.012 (2.21) Commerces in a 200m buffer 0.003 (2.67) 600m buffer -0.001 (-5.43) -0.001 (-2.87) 800m buffer -0.001 (-3.49) -0.001 (-3.97) Comm. area TAZ -1.24 (-1.94) 1.496 (2.47) 1.388 (2.23) 2.239 (3.55) 1.661 (2.67) 3.462 (5.53) Gov&Inst area TAZ 1.32 (2.57) 2.612 (4.11) 0.952 (2.35) Residential area in the TAZ -0.715 (-4.56) -0.838 (-5.22) Resources & Industry TAZ -0.812 (-2.64) -0.643 (-2.09) -1.318 (-3.9) -0.905 (-2.98) Threshold 1 -0.292 (-2.47) 0.349 (2.39) -0.006 (-0.06) 0.331 (2.78) 0.554 (3.75) -0.644 (-5.44) 0.603 (4.11) -0.293 (-1.93) Threshold 2 0.886 (7.43) 1.951 (12.72) 1.287 (11.76) 1.508 (12.24) 1.713 (11.22) 0.656 (5.58) 1.911 (12.52) 0.608 (3.99) Threshold 3 2.135 (16.45) 3.055 (18.52) 2.04 (17.43) 2.911 (21.03) 2.707 (16.69) 1.952 (15.4) 2.727 (16.91) 1.774 (11.25)

Table 12: Ordered probit models for the Medium ridership category

Boarding Alighting Am peak Pm peak Off peak day Off peak night Am peak Pm peak Off peak day Off peak night B (t-stat) B (t-stat) B (t-stat) B (t-stat) B (t-stat) B (t-stat) B (t-stat) B (t-stat) - Stop level variables Lines through stop 0.219 (6.45) 0.22 (6.37) 0.171 (5.22) 0.104 (2.54) 0.108 (3.09) 0.22 (6.42) 0.121 (3.64) Night bus through stop 0.452 (5.04) 0.55 (7.51) - Transit indices Bus stops in a 200m buffer 0.029 (2.18) 0.025 (1.91) 600m buffer 0.012 (3.54) 0.017 (5.24) 0.017 (5.1) 0.009 (2.6) 800m buffer 0.005 (2.27) Metro stations in a 400m buffer -0.497 (-5.98) -0.222 (-2.83) 600m buffer -0.168 (-3.33) -0.11 (-2.16) -0.148 (-2.87) Train stations in a 200m buffer -0.713 (-2.15) 400m buffer 0.418 (2.34) 0.457 (2.5) 1000m buffer -0.172 (-2.31) Bus line length 600m buffer 0.021 (3.69) Bus line length TAZ -0.442 (-7.27) 0.014 (2.03) 0.025 (3.84) -0.04 (-6.34) Train line length TAZ -0.013 (-2.83) -0.014 (-3.09) 0.013 (3.25) - Infrastructure Major roads length in a 400m buffer -0.677 (-3.17) 0.06 (3.34) 600m buffer 0.069 (5.33) HWay length 800m -0.049 (-2.72) -0.037 (-3.11) -0.041 (-3.29) -0.064 (-4.89) Bicycle path length in a 400m buffer 0.108 (2.58) 600m buffer -0.064 (-2.7) 0.05 (2.14) Distance to CBD -0.015 (-2.41) -0.018 (-2.8) - L-U measures Park area 400m buffer -2.911 (-3.82) Parks in a 400m buffer 0.026 (3.54) 0.023 (3.1) 600m buffer 1.11 (3.39) 0.013 (3.25) 1000m buffer 0.009 (3.95) Commerces in a 400m buffer -0.001 (-3.01) 0.001 (3.04) 0.001 (2.38) -0.001 (-2.61) 600m buffer Comm. area TAZ 2.58 (3.91) Gov&Inst. Area TAZ 0.85 (3.58) Residential area TAZ 0.045 (4.56) -0.439 (-4.04) -0.54 (-4.89) 0.37 (3.76) -0.222 (-2.69) Threshold 1 -0.172 (-1.23) 0.028 (0.23) 0.032 (0.26) 0.092 (0.77) 0.123 (0.88) -0.416 (-3.01) -0.444 (-4.67) -0.505 (-3.98) Threshold 2 1.128 (7.97) 1.811 (14.21) 1.152 (9.33) 1.023 (8.49) 1.58 (11.01) 1.069 (7.69) 0.71 (7.46) 0.297 (2.35) Threshold 3 1.9 (13.23) 2.895 (21.25) 1.99 (15.73) 1.691 (13.78) 2.41 (16.23) 1.953 (13.79) 1.601 (16.22) 0.885 (6.95)

Table 13: Ordered probit models for the Low ridership category

Boarding Alighting Am peak Pm peak Off peak day Off peak night Am peak Pm peak Off peak day Off peak night B (t-stat) B (t-stat) B (t-stat) B (t-stat) B (t-stat) B (t-stat) B (t-stat) B (t-stat) - Stop level variables Lines through stop 0.322 (6.07) 0.282 (5.77) 0.383 (7.83) 0.417 (7.37) 0.253 (5.17) 0.31 (6.11) 0.361 (7.27) 0.427 (8.03) - Transit indices Bus stops in a 400m buffer 0.048 (6.59) 600m buffer 0.027 (7.04) 0.025 (6.81) 0.021 (5.6) 1000m buffer 0.01 (6.47) 0.012 (7.34) Metro stations in a 200m buffer -0.725 (-3.5) -0.483 (-2.62) Train stations TAZ 0.11 (2.17) - Infrastructure Major roads length in a 400m buffer -0.14 (-5.88) 600m buffer -0.087 (-4.54) 800m buffer 0.018 (2.09) Hway Length 800m buffer 0.037 (2.04) Bicycle length in a 600m buffer -0.116 (-4.98) 1000m buffer 0.07 (5.35) 0.053 (4.42) 0.1 (7.41) 0.059 (4.54) 0.049 (4.04) 0.049 (4.01) Distance to CBD -0.029 (-5.95) -0.028 (-5.12) - L-U measures Park area in 400m buffer Park I n 600m buffer 0.028 (5.07) 0.04 (7.32) 0.021 (3.49) 0.019 (3.44) 0.042 (7.28) Commerces in a 600m buffer -0.001 (-4.37) -0.001 (-4.21) 800m buffer -0.001 (-4.88) 1000m buffer -0.001 (-6.52) -0.001 (-4.34) -0.001 (-4.36) Comm. area TAZ 2.069 (3.01) Gov&Inst areaTAZ -0.804 (-3.05) -1.077 (-4) -0.462 (-2.07) Residential area in the TAZ 0.289 (4.41) -0.215 (-3.11) 0.323 (4.73) 0.233 (2.76) -0.293 (-4.12) 0.378 (5.71) 0.3 (5.03) P&R TAZ -0.642 (-3.87) -0.387 (-2.48) -0.334 (-2.29) -0.664 (-2.81) -0.377 (-2.8) -0.329 (-2.35) -0.451 (-2.59) Reso&Ind TAZ -1.111 (-9.45) -1.163 (-10.47) Threshold 1 1.184 (7.54) 0.235 (2.61) 1.663 (13.03) 2.912 (20.89) 0.112 (0.82) 0.924 (6.01) 1.707 (13.54) 1.868 (14.78) Threshold 2 1.85 (11.63) 1.048 (11.37) 2.437 (18.59) 0.826 (6.04) 1.684 (10.82) 2.373 (18.39)

Table 14: Correlation matrix for high ridership stops

Boarding Alighting AM PM OPD OPN AM PM OPD OPN

AM 1 0.5974 0.7104 0.7359 -0.1602 0 -0.0949 -0.1494 PM 1 0.8369 0.7862 0.1439 0 0 -0.0797

OPD 1 0.7974 0.1368 -0.0838 0.0643 -0.1264 Boarding OPN 1 -0.0915 -0.0711 0 -0.0663

AM 1 0.5052 0.7046 0.5104 PM 1 0.8549 0.8789

OPD 1 0.8191 Alighting OPN 1

Table 15: Correlation matrix for medium ridership stops

Boarding Alighting AM PM OPD OPN AM PM OPD OPN

AM 1 0.4275 0.6783 0.6964 -0.3698 -0.2111 -0.3018 -0.3604 PM 1 0.756 0.674 -0.0696 -0.28 -0.216 -0.322

OPD 1 0.7976 -0.17 -0.302 -0.1717 -0.3605 Boarding OPN 1 -0.2728 -0.3112 -0.243 -0.3416

AM 1 0.2788 0.5005 0.4007 PM 1 0.741 0.7596

OPD 1 0.714 Alighting OPN 1

Table 16: Correlation matrix for low ridership stops

Boarding Alighting AM PM OPD OPN AM PM OPD OPN

AM 1 0.4799 0.6748 0.6317 0 0.146 0.0681 0 PM 1 0.6912 0.5631 0 0 0 -0.0854

OPD 1 0.6766 0 0.0774 0.1488 0 Boarding OPN 1 0 0.0951 0.0813 0.0978

AM 1 0.3857 0.5479 0.498 PM 1 0.7195 0.6546

OPD 1 0.7192 Alighting OPN 1

Table 17: Elasticities for High Ridership Stops

Boarding AM Boarding PM Alighting AM Alighting PM Average Ridership 31.5 36.6 34.7 33.0 Headway + 1 min -9.2 -5.1 -5.8 -9.8 + 2 min -18.1 -10.2 -11.5 -19.2 + 5 min -42.2 -24.5 -27.5 -44.3 Bus stops in 200m buffer + 1 7.5 8.8 3.5 4.7 + 2 15.3 18.0 7.1 9.6 Metro stops in 200m buffer + 1 83.4 81.7 85.2 53.2 Job Density in TAZ + 15% -0.3 -0.4 - 0.3 Residential Area in TAZ + 15% - -2.0 -2.4 - Commercial Area in TAZ + 15% - 0.5 - 0.3

Figure 4: Ridership for Different Time Periods

CHAPTER 5. CONCLUSION

The last decade has seen a strong push towards improving the sustainability of transportation systems in urban regions. In this regard, a recurring issue is that of increased car dependence in major North American cities. A well planned and efficient public transportation system can provide equitable service and accessibility to the population as well as contributing to the reduction of air pollution and GHG emissions. A number of research efforts have been focussing on understanding individual behavioral challenges in using transit while several other studies have examined the factors affecting transit operations. These studies provide important information to local agencies and transit agencies to enhance public transit services and operations. This report has added to existing literature by examining aspects of the public transportation system in Montreal, Canada.

Montreal has seen its transit service demand rise radically in the last years. In fact, it has reached a high record of transit ridership in 2011 with 405 million trips, exceeding the previous record of the year 1945. The increasing effort for sustainable transportation as well as the gas price hike in 2008 could explain this rising demand. The recent progress in relation with growing interest on reducing private vehicle usage has led to substantial interest within the travel behavior community on examining the key determinants of transit ridership. This thesis is a collection of three distinct studies examining several aspects of public transportation.

5.1 Significant Contributions

The first study examined individual home to work/school commute patterns for McGill University with an emphasis on the transit mode of travel. The overarching theme of this research was to examine the effect of the performance of the public transportation system on commuter travel mode and transit route choice (for transit riders) in Montreal. We investigated two specific aspects of commute mode choice:

(1) the factors that dissuade individuals from commuting by public transit and (2) the attributes that influence transit route choice decisions (for those individuals who commute by public transit). The contribution and findings of this report on this subject can be reported as follows:

 The travel mode choice results clearly highlight the role of travel time, walking time, number of transfers on the propensity to choose transit  Within the McGill context,, faculty members are least likely to choose the transit mode for commuting compared to staff and students  The policy sensitivity analysis conducted indicate that changes to travel times by transit mode will result in increase in the proportion of riders using transit  The results also highlight the role of walking time while choosing commute mode. Longer walking times act as deterrents to choosing transit mode. Hence, it is necessary for public transportation agencies to increase bus accessibility as well as provide better feeder access (through bus) to metro and train stations.  Individuals find travel time on the bus mode the most onerous while they are similarly sensitive to travel time on metro and train.  The influence of gender on route choice indicates that women are less sensitive to travel time compared to men  Faculty are likely to be more sensitive to travel time compared to staff and students  The policy analysis conducted indicates that reducing travel time by bus increases the likelihood of such alternatives being chosen substantially  It also indicates that routes with bus and metro alternatives are more sensitive to walking time, making it imperative that public transit agencies consider mechanism to reduce passenger walk times to metro and bus

 An innovative latent segmentation based approach is proposed and implemented  The results indicate that as the distance from the station by active forms of transportation increases, individuals are less likely to select a station first suggesting a form of stickiness to station for people living very close to the station  Also, workers and individuals leaving before 7:30 am are more likely to opt for mode first  Young persons, females, car owners, and individuals leaving before 7:30 am have an increased propensity to drive to the commuter train station  The station model indicates that travel time by the chosen mode has a significant negative impact on station as second choice  Travel time has a negative and significant impact on station choice  Presence of parking encourages use of the stations; not surprising because parking spots at a station is a function of parking demand  The latent segment model’s performance was compared with the two sequential models, clearly highlighting how the latent segment model outperforms the two sequential models

Finally, the third study draws attention to the spatial characteristics affecting ridership. An analysis of bus stop level boarding and alighting is undertaken by developing ordered response models of the bus stop specific boarding and alighting by time of day. The analysis quantifies the influence of various exogenous factors including land use attributes (road length by functional classification, bike lane lengths), built environment measures (number of parks and their areas, residential area, commercial area, government and institutional area, population density) and public transit infrastructure and accessibility indices (number of bus, metro and train stops around each stop, length of bus lanes, length of exclusive bus lanes). Regarding this research, here are the contribution and findings of this report:

 A distinction is made for different time periods as well as boarding vs. alighting when undertaking the analysis  A visual representation of the transit ridership in the Montreal region  It is observed that some areas always have a high ridership, particularly those along metro lines or those comprising a bus terminal  Several areas and neighborhoods in Montreal have generally lower ridership, which is the case for the West Island, an area where public transportation services are generally lower than that of the rest of the city  Boardings in morning and alighting in afternoon are generally away from city center  Boarding in afternoon and alighting in morning are generally closer to city center  An estimating the effect of the built environment and urban design on ridership at a stop level using ordered regression models was performed (3 different categories)  For all models, AM Boarding and PM Alighting models have similar specifications and comparable variable effects  High ridership stops can serve as transfer points for metros  Medium ridership stop can serve as transfer points for train

 Low ridership stops generally located in residential areas  Stops located in areas with more transit facilities are more likely to have higher ridership  Parks have an overall positive effect on ridership  The effect of other land use variables vary temporally  An elasticity analysis has proven that to increase ridership effectively, transit agencies must increase service and accessibility. Land-use changes do not provide significant changes in transit ridership.

5.2 Concluding Comments

There is still much research to be carried out in the public transportation field. In fact, one of the greatest challenges in the coming decades is the need for sustainable development and smart growth. For government and policy makers to implement the appropriate strategies, an adequate knowledge of the issues must be acquired, and this can be reached with thorough research. The future of public transportation is one of great potential with the gradual integration of new technologies, namely Intelligent Transportation System (ITS). By implementing ITS in an effective manner, public transportation service improve, thus increasing ridership and contributing to a more sustainable system.

References

AMT 2008, Mobilité des personnes – Enquête O-D, obtained from Agence métropolitaine de transport (accessed on 19th July 2011, http://www.enquete- od.qc.ca/docs/EnqOD08_Mobilite.pdf)

Bhat, C.R. (1997), Work Travel Mode Choice and Number of Nonwork Commute Stops. Transportation Research Part B, Vol. 31, No. 1, pp. 41-54

Bhat, C.R. (2001). Quasi-Random Maximum Simulated Likelihood Estimation of the Mixed Multinomial Logit Model. Transportation Research Part B, Vol. 35, No. 7, pp. 677-693

Bhat, C.R., and R. Sardesai (2006). The Impact of Stop-Making and Travel Time Reliability on Commute Mode Choice. Transportation Research Part B, Vol. 40, No. 9, pp. 709-730

Bhat, C.R., N. Eluru, and R.B. Copperman (2008). Flexible Model Structures for Discrete Choice Analysis. Handbook of Transport Modelling, 2nd edition, Chapter 5, pp. 75-104, edited by D.A. Hensher and K. J. Button, Elsevier Science

Bovy, P. H. L. and S. Hoogendoorn-Lanser (2005). Modelling route choice behaviour in multi-modal transport networks. Transportation 32(4): 341-368.

Calthrop, E. and S. Proost (2006). Regulating on street parking. Regional Science and Urban Economics 36, 29–48.

Cepeda, M., R. Cominetti, and M. Florian (2006). A frequency-based assignment model for congested transit networks with strict capacity constraints: characterization and computation of equilibria. Transportation Research B 40(6): 437-459.

Cervero, R., and R. Gorham (1995). Commuting in Transit Versus Automobile Neighborhoods. Journal of the American Planning Association, 61(2), 210-225.

Coppola, P. and L. Rosati (2009). Simulation-based evaluation of advanced public transportation information systems (APTIS), in: N. H. M Wilson and A. Nuzzolo (Eds) Schedule-Based Modeling of Transportation Networks: Theory and Applications, pp. 195–216 (New York: Springer).

Currie, G. and J. Phung (2007). Transit ridership, auto gas prices, and world events: new drivers of change? Transportation Research Record 1992, 3–10.

Debrezion, G., E. Pels and P. Rietveld (2007). Choice of departure station by railway users. European Transport, XIII(37), 78-92.

Debrezion, G., E. Pels and P. Rietveld (2009). Modelling the joint access mode and railway station choice. Transportation Research Part E: Logistics and Transportation Review, 45(1), 270-283.

Eluru, N. (2010). Developing advanced econometric frameworks for modeling multidimensional choices : An application to integrated land-use activity based model framework. Dissertation, University of Texas at Austin (182)

Fan, K., E. Miller and D. Badoe (1993). Modeling Rail Access Mode and Station Choice. Transportation Research Record, 1413, 49-59.

FHWA. (2002). Status of the Nation's Highways, Bridges, and Transit.

Frank, L., M. Bradley, et al. (2008). Urban form, travel time, and cost relationships with tour complexity and mode choice. Transportation 35(1): 37-54.

Frank, L. D., and G. Pivo (1994). Impacts of mixed use and density on utilization of three modes of travel: single-occupant vehicle, transit, and walking. Transportation research record, 44-44.

Givoni, M., and P. Rietveld (2007). The access journey to the railway station and its role in passengers’ satisfaction with rail travel. Transport Policy, 14(5), 357-365.

Guo, Z., N. Wilson and A. Rahbee (2007). The Impact of Weather on Transit Ridership in Chicago. Transportation Research Record 2034, 3–10.

Hamdouch, Y. and S. Lawphongpanich (2008). Schedule-based transit assignment model with travel strategies and capacity constraints. Transportation Research B 42(7-8): 663-684.

Hess, D. B. (2001). Effect of free parking on commuter mode choice: Evidence from travel diary data. Transportation Research Record: Journal of the Transportation Research Board, 1753(1), 35-42.

Hickey, R. L. (2005). Impact of Transit Fare Increase on Ridership and Revenue: Metropolitan Transportation Authority, New York City. In Transportation Research Record: Journal if the Transportation Research Board, No. 1927, Transportation Research Board of the National Academies, Washington, D.C., pp. 239–248.

Hodges, T. (2009) Public Transportation’s Role in Responding to Climate Change, Prepared for February 3rd 2012 at: www.fta.dot.gov/documents/PublicTransportationsRoleInRespondingToClimateCha nge.pdf

Keijer, M. J. N. and P. Rietveld (2000). How do people get to the railway station? The dutch experience. Transportation Planning and Technology, 23(3), 215-235.

Koppelman F.S, and V. Sethi (2008). Closed-form discrete choice models. D.A. Hensher, K.J. Button (Eds.), Handbook of Transport Modelling (second ed.), Elsevier Science, pp. 211–225

Krygsman, S., M. Dijst and T. Arentze (2004). Multimodal public transport: an analysis of travel time elements and the interconnectivity ratio. Transport Policy, 11(3), 265-275.

Lam, S.-H. and F. Xie (1981). Transit Path-Choice Models That Use Revealed Preference and Stated Preference Data. Transportation Research Record 1799: 58- 65.

Li, Z., D. A. Hensher, and J.M. Rose (2010). Willingness to pay for travel time reliability in passenger transport: A review and some new empirical evidence. Transportation Research Part E: Logistics and Transportation Review 46(3): 384-403.

Liu, Y., J. Bunker, and L. Ferreira (2010). Transit Users’ Route-Choice Modelling in Transit Assignment: A Review. Transport Review 30(6): 753-769.

Mattson, J., (2008). Effects of Rising Gas Prices on Bus Ridership for Small Urban and Rural Transit Systems. North Dakota State University.

Marguier, P. H. J. and A. Ceder (1984). Passenger waiting strategies for overlapping bus routes. Transportation Science 18(3): 207-230.

Martens, K. (2004). The bicycle as a feedering mode: experiences from three European countries. Transportation Research Part D: Transport and Environment, 9(4), 281-294.

Newman, J. and V. Bernardin (2010). Hierarchical ordering of nests in a joint mode and destination choice model. Transportation, 37(4), 677-688.

Noland, R. B. and J. W. Polak (2002). Travel time variability: a review of theoretical and empirical issues. Transport Reviews 22(1): 39-54.

Nuzzolo, A. and U. Crisalli (2004). The schedule-based approach in dynamic transit modeling: a general review, in: A. Nuzzolo and N. H. M. Wilson (Ed) Schedule-Based Dynamic Transit Modeling: Theory and Applications, pp. 1–22 (Boston, MA: Kluwer).

Pinjari, A.R., R.A. Pendyala, C.R. Bhat and P. Waddell (2011). Modeling the Choice Continuum: an Integrated Model of Residential Location, Auto Ownership, Bicycle Ownership, and Commute Tour Mode Choice Decisions. Transportation, 38(6), 933- 958.

Pinjari, A.R., N. Eluru, C.R. Bhat, R.A. Pendyala and E. Spissu (2008). Joint Model of Choice of Residential Neighborhood and Bicycle Ownership: Accounting for Self-

Selection and Unobserved Heterogeneity. Transportation Research Record: Journal of the Transportation Research Board, 2082(-1), 17-26.

Pinjari, A. R., R. M. Pendyala, C.R. Bhat and P.A. Waddell (2007).Modeling residential sorting effects to understand the impact of the built environment on commute mode choice. Transportation 34(5): 557-583.

Pinjari, A.R., and C.R. Bhat (2006). On the Nonlinearity of Response to Level of Service Variables in Travel Mode Choice Models. Transportation Research Record, Vol. 1977, pp. 67-74

Rajamani, J., C.R. Bhat, S. Handy, G. Knaap and Y. Song, (2003). Assessing Impact of Urban Form Measures on Nonwork Trip Mode Choice After Controlling for Demographic and Level-of-Service Effects. Transportation Research Record: Journal of the Transportation Research Board, 1831(-1), 158-165.

Ryan, S. and L.D. Frank (2009). Pedestrian environments and transit ridership. Journal of Public Transportation, 12(1), 39–57.

Scheiner, J. (2010). Interrelations between travel mode choice and trip distance: trends in Germany 1976-2002. Journal of Transport Geography 18(1): 75-84.

Schrank, D., T. Lomax and B. Eisele (2005). The 2011 Urban mobility report: Texas Transportation Institute.

Shoup, D. (2005). The High Cost of Free Parking. Chicago, IL: Planner’s Press.

Shoup, D., 2006. Cruising for parking. Transport Policy 13, 479–486.

Sider, T., A. Alam, M. Zukari, H. Dugum, N. Goldstein, N. Eluru et al. (2012). Quantifying The Effects Of Land-Use And Socio-Economics On The Generation Of Traffic Emissions And Individual Exposure To Air Pollution At A Metropolitan Scale. , Technical Paper, Department of Civil Engineering and Applied Mechanics, McGill University.

Small, K. A., C. Winston, et al. (2005). Uncovering the Distribution of Motorists' Preferences for Travel Time and Reliability. Econometrica 73(4): 1367-1382.

STM. (2012). Plan Stratégique 2020 [online]. Available from http://stm.info [Accessed 16 February 2013].

Strathman, J. and K. Dueker (1995). Understanding trip chaining. Special Reports on Trip and Vehicle Attributes, US Department of Transportation.

Sung, H., and J.-T. Oh (2011). Transit oriented development in a high-density city: identifying its association with transit ridership in Seoul, Korea. Cities, 28(1): 70–82.

Taylor, B., D. Miller, H. Iseki and C. Fink (2009). Nature and/or nuture? Analyzing the determinants of transit ridership across US urbanized areas. Transportation Research Part A 43, 60–77.

Van Wee, B., H. Holwerda, et al. (2003). Preferences for modes, residential location and travel behaviour: the relevance of land-use impacts on mobility. European Journal of Transport and Infrastructure Research 2 2(3-4): 305-315.

Van Wee, B., and R. Van Baren (2002). Preferences for modes, residential location and travel behaviour: the relevance of land-use impacts on mobility. European Journal of Transport and Infrastructure Research, 2(2), 305-315.

Waddell, P., C.R. Bhat, N. Eluru, L. Wang and R.A. Pendyala (2007). Modeling Interdependence in Household Residence and Workplace Choices. Transportation Research Record: Journal of the Transportation Research Board(2003), 84-92.

Wahba, M., Shalaby, A.: MILATRAS: a new modelling framework for the transit assignment problem. In: Wilson, N., Nuzzolo, A. (eds.) Schedule-Based Modelling of Transportation Network. Springer, New York (2009a)

Wang, G. K. H. and D. Skinner (1984). The Impact of Fare and Gasoline Price Changes on Monthly Transit Ridership. Transportation Research B 18, 1: 29-41.

Wardman, M. and G.A. Wheelan (1999). Using geographical information systems to improve rail demand models. Final Report to Engineering and Physical Science Research Council.

Windmeijer, F. A. G. (1995). Goodness-of-fit measures in binary choice models. Econometric Reviews 14, (101-116).

Ye, X., R. Pendyala, G. Giovanni (2007). An exploration of the relationship between mode choice and complexity of trip chaining patterns. Transportation Research B 41: 96-112.

Zhang, M. (2004). The role of land use in travel mode choice. Journal of the American Planning Association 70(3): 344-360.

Appendix

Table 18: Ridership intervals for different stop categories

AM PM Off Peak Off Peak Peak Peak Day Night k=1 0-10 0-10 0-2.5 k=2 10-25 10-20 2.5-5

High k=3 25-50 20-30 5-10 k=4 50 + 30 + 10 +

k=1 0-2 0-1.5 0-0.5 k=2 2-6 1.5-3 0.5-1

k=3 6-10 3-4.5 1-1.5 Medium k=4 10 + 4.5 + 1.5 +

k=1 0-0.5 0-0.25 0-0.25

k=2 0.5-1 0.25-0.5 0.25 + Low k=3 1 + 0.5 + -

100