Content uploaded by Md Sami Hasnine
Author content
All content in this area was uploaded by Md Sami Hasnine on Nov 30, 2020
Content may be subject to copyright.
Determinants of travel mode choices of post-secondary students in a large metropolitan 1
area: the case of the city of Toronto 2
3
Md Sami Hasnine, MASc 4
PhD Candidate 5
Department of Civil Engineering 6
University of Toronto 7
35 St George Street 8
Toronto, ON, M5S1A4 9
Tel: 647-283-9067; Email: sami.hasnine@mail.utoronto.ca 10
11
12
TianYang Lin 13
MEngCEM Candidate 14
Department of Civil Engineering 15
University of Toronto 16
35 St George Street 17
Toronto, ON, M5S1A4 18
Tel: 905-764-5379; Email: tian.lin@mail.utoronto.ca 19
20
21
Adam Weiss, MASc 22
PhD Candidate 23
Department of Civil Engineering 24
University of Toronto 25
35 St George Street 26
Toronto, ON, M5S1A4 27
Tel: 416-567-4970; Email: adam.weiss@mail.utoronto.ca 28
29
30
Khandker Nurul Habib, Ph.D., PEng 31
Associate Professor 32
Department of Civil Engineering 33
University of Toronto 34
35 St George Street 35
Toronto, ON, M5S1A4 36
Tel: 416-946-8027; Email: khandker.nurulhabib@utoronto.ca 37
38
39
40
41
42
43
44
2
ABSTRACT 1
The paper presents an investigation on the mode choice behaviour of post-secondary students 2
commuting to school in the city of Toronto. It uses a large-scale dataset collected through a web-3
based travel diary survey among all students of four universities (seven campuses) in Toronto. 4
Multinomial logit (MNL), nested logit (NL) and cross-nested logit (CNL) models are used for 5
investigating home to school trips mode choices. In terms of goodness-of-fit, the CNL 6
outperforms the MNL and NL model. Furthermore, the proposed CNL model shows fundamental 7
improvements over the MNL and NL models by capturing non-proportional substitution patterns. 8
Empirical models reveal that the mode choice behaviour of female students who travel to 9
downtown campuses differ significantly from female students who travel to suburban campuses. 10
Female students who travel towards downtown are more transit and active mode oriented than 11
those who travel towards outside of downtown. This study also shows mobility tool ownerships 12
(i.e., transit pass, car and bike ownership) and age groups have distinctive influences on student’s 13
mode choice behaviour. Using the CNL model as a tool for policy scenario analysis, it is found 14
that public transit users are highly sensitive to changes in travel time. In the context of policy 15
implementation, if bike and ride mode is encouraged during peak hour commuting, there is likely 16
a large amount of latent demand for this mode. 17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
3
1. Introduction 1
The years as students in universities and colleges represents an important transitional period in 2
most people's lives. As students move towards adulthood in this life stage, they also form 3
preferences and habits that will likely have impacts on their later life. Activity-travel behaviours 4
are no exception (Kamruzzaman et al. 2011; Khattak et al. 2011; Balsas 2003). As this 5
demographic cohort move into the workplaces and join the larger commuting population, 6
understandings of their travel behaviour, especially their mode choice behaviour, are valuable in 7
the context of demand forecast and long-term planning for urban transportation. Current post-8
secondary students are parts of the millennials and with baby boomers retiring, travel behaviour 9
of millennials are of profound interest to the transportation and urban planners. 10
However, it is until recently that transportation researchers started to look at travel behavior of 11
post-secondary students. Previous studies on this topic mainly followed two approaches: (1) 12
mathematical models (2) descriptive statistics. Studies which relied on mathematical models 13
mainly used Multinomial logit (MNL) or Nested logit (NL) models (Lavery et al., 2013; 14
Grimsrud and El-Geneidy, 2013; Kuhnimhof et al., 2012; Lavery et al., 2013; Rodrı́guez and Joo, 15
2004; Whalen et al., 2013; Zhou, 2012). Over the last three decades, the mode choice models are 16
matured to such extent that it is common to use MNL and NL model. However, issues of 17
preference heterogeneity and imperfect substitution patterns are overlooked in MNL models. 18
While some models have moved towards NL models, a more comprehensive examination of 19
substitution patterns is advisable through the application of other generalized extreme value 20
models such as the cross-nested logit (CNL) (Rodrı́guez and Joo, 2004). The use of advanced 21
generalized extreme value (GEV) models are useful in capturing the non-proportional 22
substitution patterns, which can be informative in revealing and forecasting travel behaviours of 23
particular demographics. 24
25
Alternatively, an array of studies relied on descriptive statistics to capture post-secondary students’ 26
travel patterns (Boyd et al., 2003; Delmelle and Delmelle, 2012; Limanond et al., 2011; Miralles-27
Guasch and Domene, 2010; Shannon et al., 2006; Uttley and Lovelace, 2016; Khattak et al., 2011). 28
Though studies using descriptive statistical analysis reveal a general picture of the travel behaviour 29
of post-secondary students, without adequate modelling tool it is not feasible to forecast future 30
travel demand and analyze the implication of various transport policies. In terms of data collection, 31
two types of surveys are commonly found in the literature which investigated students’ mode 32
choice phenomenon: (1) Household level travel survey and (2) Single university or campus based 33
survey. Though a limited body of literature uses household-level travel surveys in their study, 34
household level travel surveys are often observed to undercount the post-secondary student 35
population (Grimsrud and El-Geneidy, 2013; Khattak et al., 2011). As a result, the importance of 36
single university or campus based surveys is well argued in literature. Many past studies used 37
single university or single campus based surveys (Boyd et al. 2003, Delmelle and Delmelle, 2012; 38
Limanond et al., 2011; Miralles-Guasch and Domene, 2010; Rodrı́guez and Joo, 2004; Shannon et 39
al., 2006; Uttley and Lovelace, 2016; Whalen et al., 2013). However, numerous cities and city 40
regions host multiple universities with multiple campuses. In these cases, conducting surveys on 41
the students from a single campus can be problematic, since the mode choice decision of the 42
students from suburban campuses may vary a lot than the students from the downtown campuses. 43
4
As such, a concurrent study across multiple institutions and campuses would be beneficial. To our 1
knowledge, there have been no applications of cross campus university travel surveys. 2
3
To contribute to the growing literature on post-secondary student’s mode choice behaviour, this 4
study makes use of a comprehensive travel diary survey capturing all universities (four 5
universities and seven campuses) in Toronto, a population of 0.18 million post-secondary 6
students. This travel survey is a unique case which represents a very important population cohort 7
in a metropolitan area where university campuses are scattered both in suburban and downtown 8
areas. Single campus surveys would fail to capture this complete picture for cities and city 9
regions with multiple universities. 10
11
In this paper CNL model is proposed along with MNL and NL models to reveal the influences of 12
various contextual factors alone or in interactions with socio-economic variables in defining 13
trade-offs in home to campus trip mode choices of the students. It is found that proposed CNL 14
model outperforms the MNL and NL model in the context of goodness-of-fit. Furthermore, CNL 15
model has a closed formulation and can be estimated without applying simulation techniques. 16
17
The paper explains the research as follows. The next section presents a brief literature on mode 18
choice investigations of post-secondary students. The following sections present a discussion on 19
the dataset used for empirical investigation; modelling methodology and a discussion on 20
empirical results. The paper concludes with summary of key findings and recommendations for 21
future research. 22
23
2. Literature Review 24
Travel behaviour of university (post-secondary) students has been garnering attention in recent 25
years even though the body of literature is still relatively compact. Few studies adopted 26
econometric models to investigate students’ commuting mode choice. Whereas the remainders 27
relied on descriptive statistics. The literature review in this paper considers three major themes: 28
(1) variables influencing transit mode choice, (2) variables influencing active transportation 29
mode choice, and (3) variables influencing driving mode choice. For each theme, we covered 30
studies which adopted econometric models and then studies which adopted descriptive statistics. 31
2.1 Variables influencing transit mode choice 32
A small number of the past literature exploited econometric model to investigate variables 33
influencing students’ transit mode choice, most of which are limited to MNL, NL or mixed logit 34
model. Zhou (2012, 2014, 2016) has a series of studies where MNL and two-stage least squares 35
non-recursive models are exploited to investigate post-secondary students travel mode choice in 36
Los Angeles. These studies show that subsidized transit pass, higher access to transit services 37
and decent transit services encourage students to choose transit. Rotaris and Danielis (2014) 38
estimate a mixed logit model where they found similar findings as Zhou (2012, 2014, 2016). 39
Danaf et al. (2014) estimated nested logit models to find major determinants that affect students’ 40
mode choice. They found that increasing parking cost, decreasing transit travel time and cost can 41
influence students to choose public transit over car. 42
5
Few studies relied on descriptive statistics to specifically investigate students’ mode choice with 1
respect to transit usage. Among these studies, one study of a fare-free transit program for the 2
students at University of California at Los Angeles found that when cost is taken out of the 3
picture, students would make use of public transit for most of their travels (Boyd et al. 2003). In 4
this same study, the authors compared the modal share before and after introducing a shuttle bus 5
program for university students and employees. The authors found that higher transit service 6
frequency and lower transit access distance both increase transit ridership (Boyd et al. 2003). 7
Zhan et al. (2016) estimated hierarchical tree-based regression models based on survey data 8
collected from eight universities in China. This non-parametric investigation reveals that transit 9
network, access station location, commuting distance, school location significantly influence 10
students’ mode and commuting frequency. Grimsrud and El-Geneidy (2014) in their study of the 11
20-34 age cohorts in Montreal were also able to find transit preference among university 12
students. They show summary statistics of repeated cross-sectional origin–destination survey 13
data of Greater Montreal. In another study, Grimsrud and El-Geneidy (2013) show the general 14
trend of increasing affection toward transit modes among this group of travellers. 15
2.2 Variables influencing active transportation mode choice 16
Few studies are focused on how different factors influence students’ active transportation mode 17
choice such as walk and bike. In terms of mode choice, post-secondary students, and young 18
adults are found to have higher preference for transit and active transportation modes such as 19
walking and biking (Khattak et al., 2011). Akar et al. (2013) investigated how female students 20
travel behaviour is different in the context of bicycling in particular. The estimated MNL model 21
from the study shows that female students are more concerned about safety when they choose to 22
bike. In particular, female bikers are more inclined to bike if there are bike trails or segregated 23
bike lanes relative to biking in mixed automobile traffic. 24
Several other studies show that commuting distance is one of the crucial factors for selecting 25
biking and walking for university students (Lundberg and Weber, 2014; Rybarczyk and 26
Gallagher, 2014). Rybarczyk and Gallagher (2014) rely on binary logit model to investigate how 27
distance from home to campus can affect the likelihood of choosing to bike or walk as a 28
commuting mode. In their study, Rybarczyk and Gallagher (2014) found that a visible bicycle 29
culture would encourage students to choose biking as a commuting mode, while safety is the 30
main hindrance for the students when it comes to choose walking as a commuting mode. 31
Rodriguez and Joo (2004) estimated a series of discrete choice model to investigate the 32
relationship between the use of active transportation modes and the built environment . 33
Rodriguez and Joo estimated three models, including a MNL, NL and heteroscedastic extreme 34
value (HEV) model. Their analysis exploits local topographic information (slope, sidewalk etc.) 35
as explanatory variables and found that these variables have significant influence on walking and 36
biking. However, the estimated NL and HEV models show lower goodness-of-fit than MNL 37
model which is unusual, since majority of the literature show that capturing non-proportional 38
6
substitution pattern improves the model’s goodness of fit significantly (Koppelman and Bhat 1
2006). 2
Uttley & Lovelace (2016) find that the mode share of cycle users did not increase after 3
implementing a cycling promotion scheme at a UK university, suggesting that students’ long-4
term habit influence their mode choice. Duque et al. (2014) demonstrate that conventional 5
transportation demand management (TDM) strategies, which are introduced to attract more 6
university employee and students to active transportation modes, may fail to attract university 7
staff to change their existing commuting mode choice, while car ownership and higher income 8
may encourage staff to still drive to school. 9
Whalen et al. (2013) address some of the shortcomings of previous literature by including 10
variables such as socio-demographic status, attitudes, built environment, and mode and trips 11
specific factors. Interestingly, they highlighted the positive utility of travel time for cycling and 12
driving, a counterintuitive result. This finding requires further investigation to determine if travel 13
time is correlated with the random utility component or if there is a high degree of collinearity 14
between travel time and other variables included in the utility of each alternative. 15
2.3 Variables influencing driving mode choice. 16
In contrast to studies focusing on active transportation modes (i.e., walk and bike) usage, another 17
set of studies focused on the factors that influence students to choose driving mode. Soria-Lara et 18
al. (2016) and Sultana (2015) both try to find the factors that influence students to commute to 19
school by car. Soria-Lara et al. (2017) find that weekly vehicle kilometers travelled are correlated 20
with daily stay at campus, gender, and age. Another study, which is done on the same university 21
students, found that the lack of adequate biking infrastructure influences students to commute by 22
car (Miralles-Guasch and Domene, 2010). In contrast to Soria-Lara et al. (2016), Sultana (2015) 23
find that gender and income have little to no effect in parking pass purchasing decisions. Sultana 24
(2015) also found that parking pass purchase decisions largely depend on students’ habit and car 25
ownership. Davison et al. (2015) found that female students and part time students are more 26
inclined to choose car as commuting mode. This finding essentially shows that there are some 27
existing barriers which are discouraging female students to choose active transportation modes 28
for commuting purpose. 29
Several studies show that students’ long-term habit play a crucial role in choosing driving as a 30
mode, even though the utilities of other modes are higher in some cases. Both Kerr et al. (2010) 31
and Klöcknera and Friedrichsmeie (2011) show that students’ long-term habit, or inertia, 32
encourages them to use car as commuting mode. Few other studies suggest that by putting more 33
emphasize on education and environmental concerns we can reduce the driving mode share (Kerr 34
et al., 2010; Kim et al., 2014). Studies by Balsas (2003) and Shannon et al. (2006) substantiated 35
such findings along with the fact that post-secondary student’s population may have latent 36
demand for the use of non-motorized modes. 37
38
7
From the discussion above, it is found that of the few econometric investigations on this area of 1
research, all have either relied on small range of explanatory variables or used simple discrete 2
choice model overlooking substitutions patterns and/or preference heterogeneities in mode 3
choices. In addition, post-secondary students in general are chronically under-represented in 4
household travel surveys and household travel surveys provide the core database in most 5
regional planning studies. 6
This paper contributes to the literature in two ways. First, it uses a comprehensive travel survey 7
data collected from all universities in Toronto. The dataset represents students living all over the 8
Greater Toronto Area and commuting to campuses either in the downtown and suburban areas. 9
Secondly, the paper explicitly investigates preference heterogeneity through estimation of 10
various forms of discrete choice models. It exploits closed from advanced formulations, e.g. 11
cross-nested logit model to capture clustering and non-proportional substitution patterns of mode 12
choices of post-secondary students. The study also uses wide varieties of personal, household 13
and land use attributes to investigate the key determinants of school trip mode choices of post-14
secondary students in Toronto. 15
16
3. Survey Implementation and Descriptive Statistics 17
The data for this study come from a web-based survey, which is named as “StudentMoveTO”, 18
conducted among the university students in the City of Toronto. Four universities have been 19
chosen based on their higher number of existing student: a) Ontario College of Art and Design 20
(OCAD), b) Ryerson University, c) York University and d) the University of Toronto. Among 21
these four universities, the University of Toronto and York University have multiple campuses 22
across the region. The University of Toronto has three campuses in three locations namely, St. 23
George, Scarborough, and Mississauga. York University has also two campuses: Glendon and 24
Keele. As such the survey sample frame is the students from all seven campuses of these four 25
universities. These four universities have a combined total of around 184,000 students. The time 26
frame for the data collection of the StudentMoveTO is during Fall of 2015. Emails were sent to all 27
students’ university email addresses. Among the entire student body, 15226 students completed 28
the survey, which corresponds to a response rate of 8.0%. This survey asked for respondents’ 29
personal information, household information and travel diaries. For the travel diary, each 30
respondent reported all trips made by him/her during the day prior to the date of the survey which 31
is a standard practice for many household level travel survey data collection schemes in North 32
America (e.g., Transportation Tomorrow Survey 2011). The objective of this study is to develop 33
commuting mode choice models. As such, it is required to retrieve the commuting trips of the 34
students from the database. When breaking down the total number of trips taken by trip destination 35
purposes, commuting trips to school represents just under a quarter of the total. Of those, around 36
70% are made on weekdays. A total of 3208 students’ records are eventually retrieved from those 37
reported a commuting trip to school on a weekday in their travel diary. StudentMoveTo classifies 38
commuting modes into eight distinct classes as follows: 39
Driving 40
Auto Passenger 41
Local Transit with Walk Access 42
Park and Ride 43
Kiss and Ride 44
8
Bike and Ride 1
Walk 2
Bike 3
Table 1 shows the descriptive statistics of the sample dataset used in this study. A preliminary 4
analysis of the sample statistics shows that the average household size is 3.6 and the average 5
number of dependent children in the household is 0.25. The average age is 22.53 with a standard 6
deviation 5.46, which is intuitive since this survey is designed exclusively for university students. 7
The dataset includes different type of household mobility tools such as a car, bike, and transit pass 8
ownership. It is found that only 14% of the students have their own car, while 42% have transit 9
passes, and 32% of them have a smart fare payment card (Presto card) which allows them to pay 10
at all regional transit stations and select local transit stations. The regional planning agency, 11
Metrolinx, commissioned the Presto card and promotes it as an integrated fare payment method 12
across different transit agencies in the region. The local transit agency - the Toronto Transit 13
Commission (TTC), only accepts payment through Presto card at selected locations. The 14
ownership of Presto card influences all four transit related modes. However, it has less influence 15
when compared to the local transit pass. The possible explanation is the limited rollout of Presto 16
card system across the local transit (TTC) network within the City of Toronto. As of July 2016, 17
only 31 subway stations have presto card reading facilities out of a total of 69 subway stations 18
(Toronto Transit Commission 2016). Card readers have been installed on all streetcars, but there 19
are almost none on buses. As such, the practical use of the Presto card is severely limited when 20
compared to a Metropass, the traditional transit pass distributed by the TTC. As such, for cross-21
regional commuters, Presto card is a supplementary mobility tool to access other transit systems 22
in the region. 23
24
The bike ownership percentage is also high (49%). This mobility tool ownership information has 25
inherent relations with the mode share of the sample population. For instance, the high transit pass 26
ownership corresponds to the high mode share of the local transit with walk access (48.57%). Walk 27
mode also has a significant share (22.54%). Many students live near the university. As such, 28
walking is a suitable option for them. Table 1 also shows the home to school level of service (LOS) 29
values for the respondents. Various LOS (e.g., driving time and cost, in-vehicle travel time for 30
transit, access time to transit, and waiting time for transit) values are estimated using calibrated 31
traffic assignment models. The transportation network is coded into the EMME travel demand 32
modelling software platform. A second ordered linear approximation traffic assignment with 33
background transit assignment model is used to estimate a driving travel time origin-destination 34
(O-D) matrices. A link based probabilistic shortest path algorithm is used to estimate driving cost 35
O-D matrices based on per unit distance cost function. This auto cost includes the average vehicle 36
maintenance plus fuel cost per kilometres multiplied with distance. A fare based congested transit 37
assignment model is used to generate transit LOS matrices. These traffic and transit assignment 38
models are calibrated using the 2011 Transportation Tomorrow Survey (TTS) data. These models 39
consistently produce expected travel time for any given pair of traffic analysis zone (TAZ). A 40
detail description of these traffic assignment models can be found in Miller et al. (2015). 41
42
43
44
45
46
9
Table 1. Summary Statistics of the selected variables 1
Variable Mean St. Deviation Maximum Minimum
Household Size 3.60 1.50 16.00 2.00
Number of Dependent Children in the Household 0.25 0.72 5.00 0.00
Number of Cars in the Household 1.11 1.04 9.00 0.00
Age of the Respondents 22.33 5.46 62.00 0.00
Driving License Dummy (Yes=1,0=No) 0.59 0.49 1.00 0.00
Car Ownership Dummy (Yes=1,0=No) 0.14 0.35 1.00 0.00
Rideshare Membership (Yes=1,0=No) 0.05 0.21 1.00 0.00
Transit pass Ownership Dummy (Yes=1,0=No) 0.42 0.49 1.00 0.00
Presto Card Ownership Dummy (Yes=1,0=No) 0.35 0.48 1.00 0.00
Bike Ownership Dummy (Yes=1,0=No) 0.49 0.50 1.00 0.00
Bike share Membership Dummy (Yes=1,0=No) 0.01 0.11 1.00 0.00
Auto Cost ($) 2.17 2.32 17.74 0.00
Auto In-vehicle Travel Time (min) 17.59 17.85 99.29 0.00
Transit Fare ($) 2.35 2.47 13.13 0.00
Transit In-vehicle Travel Time (min) 36.31 35.57 175.80 0.00
Transit Wait Time (min) 5.02 4.73 48.27 0.00
Transit Walk Time (min) 18.00 10.79 263.53 0.00
Drive Access Time (min) 1.44 0.86 21.08 0.00
Bike Access Time (min) 4.80 2.88 70.27 0.00
Home to School Distance (Km) 15.30 15.61 94.39 0.02
The distance in kilometers to the nearest bus stop from the
postal code centroid
0.27 0.35 10.88 0.00
The distance in kilometers to the nearest rail stop from the
postal code centroid
3.54 2.08 21.06 0.05
The distance in kilometers to the nearest streetcar stop from
the postal code centroid
9.64 10.55 64.59 0.00
The distance in kilometers to the nearest subway stop from
the postal code centroid
6.86 9.36 64.50 0.01
The employment density (employees per sq. km) 2011
divided by 1000
9.24 18.39 271.18 0.00
Gender (%)
Female 64.68
Male 35.32
University (%)
University of Toronto 65.07
Ryerson University 20.90
York University 11.59
OCAD University 2.44
Student Status (%)
Undergraduate 77.31
Graduate 21.85
Mode Share (%)
Driving 5.52
Auto Passenger 5.33
Local Transit with Walk Access 48.57
Park and Ride 3.02
Kiss and Ride 7.86
Bike and Ride 0.22
Walk 22.54
Bike 6.95
2
10
In this study, several land use attributes related to the respondent’s “home” zone are incorporated 1
into the analysis, such as transit frequency, employment density and area of sidewalks in 2
kilometers squared. In addition, the distances in kilometers to the nearest bus stop, rail stop, 3
subway station from the postal code area centroid are also fused with the dataset. Figure 1 shows 4
the spatial distribution of the university of Toronto St. George campus students’ home and school 5
location and their commuting mode. As is expected, most of the bike and walk mode users are 6
found in downtown Toronto. This is intuitive, since bike infrastructure facilities and pedestrian 7
friendly environments are not very common outside of the downtown areas (Hasnine et al., 2017). 8
In addition, transit with walk access mode users mostly live near subway stations or in locations 9
where they could easily access the bus. It is also clear from the figures that driving and auto 10
passenger mode users are travelling to school from comparatively further distances. Finally, park 11
and ride, and kiss and ride users mostly live outside of the downtown area and are commuting long 12
distances to school. 13
(a) Driving and Auto Passenger
(b) Local Transit with Walk Access
(c) Park and ride, and Kiss and Ride
(d) Walk and Bike
14
Figure 1: Home locations and mode share of the students’ University of Toronto St. George 15
Campus 16
17
For estimating any mode choice model, it is required to generate the feasible alternative sets by 18
using feasibility rules. The following rules have been set to define the availability of the eight 19
modes under study. 20
Driving - the respondent owns a driver’s license and his household owns a car 21
11
Auto passenger - available to everybody 1
Local transit walk access – depends on transit network assignment model result regarding 2
the availability of local transit and the one-way travel time should be less than 150 min. 3
Park and Ride – depends on transit network assignment model result regarding the 4
availability of park and ride station within a threshold distance, the one-way travel time 5
should be less than 150 min and the similar conditions of driving. 6
Kiss and Ride - depends on transit network assignment model result regarding the 7
availability of transit and the similar conditions of auto-passenger. 8
Bike and Ride – depends on transit network assignment model result regarding the 9
availability of transit and household owns a bike. 10
Walk - commuting distance is no greater than 3 km 11
Bike – commuting distance is not greater than 10 km and household owns a bike 12
13
4. Econometric Modelling Framework 14
15
Three models are estimated in this study and these are multinomial logit (MNL), nested logit 16
(NL), and cross-nested logit (CNL) models as shown in Figure 2. MNL is the most popular 17
modelling structure in the family of discrete choice models (McFadden 1973). The MNL model 18
assumed that the random utility components of modal alternatives are independently and 19
identically distributed extreme values. This assumption leads to identical cross-elasticities for all 20
other alternatives with respect to one specific alternative (Wen & Koppelman 2001; Train 2003). 21
This represents a perfect substitution pattern where all alternatives are perceived to be exact 22
substitutes of each other. In reality, alternatives may not be perceived to be exactly substitute of 23
each other, especially for special demographic groups like post-secondary students. A Nested 24
Logit model can overcome this assumption by allowing nests of alternatives with different 25
substitution within the nest as opposed to alternatives that are out of the nests (Williams 1977). A 26
further advancement of the NL model is the CNL model, where one alternative can be a member 27
of multiple nests (Wen & Koppelman 2001; Train 2003). CNL allows us to capture different 28
cross-elasticities between pairs of alternatives. 29
For an individual the random utility of mode (𝑚) can be written as: 30 𝑈= 𝛽𝑋+𝜀 (1) 31
Where, 𝑈 is total utility of mode m 32 𝑉 are systematic utility of mode m, where 𝑉= 𝛽𝑋 33 𝛽 is the parameter vector 34 𝑋 are attributes to the corresponding mode 35 𝜀 are random utility components with zero mean and µ scale parameter. 36
37
The general MNL formulation of the probability of choosing a particular mode can be written as: 38
𝑃=𝑒µ∗
∑𝑒µ∗
(2) 39
40
In the NL model probability of choosing a mode is equal to the probability of choosing that 41
mode (m) conditional to choosing the same nest (T) which the mode belongs to. In our study, the 42
12
nest T has four alternatives (j). If µ is the root scale parameter and µ is the scale parameter of 1
the transit nest, for NL formulation probability of choosing a particular mode within transit nest 2
can be written as follows: 3
𝑃| () =µ
∑µ∗
∗ µ
µ (∑µ∗)
µ
µ (∑µ∗)
∑µ∗
(3) 4
5
6
7
8
9
10
11
12
13
Figure 2. Various modelling structure 14
15
16
17
18
19
20
21
22
13
In equation (3) µ
µ is the coefficient of expected maximum utility which should be between 0 to 1
1. If a particular mode (l) is not within the nest then the probability of choosing the mode, 2
𝑃=𝑒µ∗
𝑒µ
µ (∑µ∗)
+∑𝑒µ∗
(4) 3
In the CNL model, we have two nests: transit and active mode. No allocation parameter is 4
considered in this case (i.e. equal allocation between nests). In this case, the probability of choosing 5
mode m within a nest can be written as follows: 6
𝑃=
⎣
⎢
⎢
⎡
(𝑒)µ
∑(𝑒)µ
∗ 1
µ𝑒 (∑()µ
)
∑1
µ 𝑒 (∑()µ
)
⎦
⎥
⎥
⎤
(5) 7
8
In a mode choice model for a sample of N individual with each individual having the options of 9
m alternatives the log likelihood function becomes (Ben-Akiva & Lerman 1985, Aptech Systems 10
2016): 11
𝐿𝐿(β) = 𝑦
ln (𝑃) (6) 12
Whereas, 𝑦 = 1 if person 𝑖 choose mode 𝑚 and zero otherwise. 13
14
The marginal effect of MNL, NL and CNL model can be calculated by equation (7), (8) and (9) 15
respectively, 16
17
Marginal Effect (MNL) = (1-𝑃)* 𝛽 (7) 18
19
Marginal Effect (NL) = ((1-𝑃)+µ𝑇− 1∗ (1 − 𝑃𝑚|𝑛𝑒𝑠𝑡)) * 𝛽 (8) 20
21
Marginal Effect (CNL) = ∑ |((1−)(µ)∗(|))
* 𝛽 (9) 22
23
5. Empirical model 24
Three modelling structures are presented in this section: MNL, NL and CNL. Table 2, 3, 4 and 5 25
shows the model estimation results. Variables in the final specifications are selected based on the 26
expected sign, and statistical significance (95% confidence interval) of corresponding 27
parameters. The final specification of MNL has 35 parameters, the NL has 36 parameters and the 28
CNL has 38 parameters. Some of the variables in these three models are not statistically 29
significant at 95% confidence interval, but they are retained in the models because it is felt these 30
variables provide significant insight when comparing the three different modelling structures. 31
32
For all three models, the goodness-of-fit against the equiprobable model is measured. Three 33
models with three different substitution patterns allow us to compare effects of different 34
variables on mode choice preferences of students. For discussion, we used marginal effects of the 35
variables in each model. Figure 3 presents the marginal effect comparison of some selected 36
variables. Marginal effects are estimated by using probability weighted sample enumeration 37
14
(PWSE) technique. 1
2
These model results will be discussed in the context of three categories of variables: LOS, socio-3
economic and land use. In addition, rather than describing the three models’ result separately, a 4
comparison of the three models will be presented here. CNL model shows the best goodness-of-5
fit (adjusted rho-square value against equiprobable model) than the MNL and NL model, which 6
means a significant improvement over the past studies which mainly relied on MNL or NL 7
(Lavery et al., 2013; Grimsrud and El-Geneidy, 2013; Kuhnimhof et al., 2012; Lavery et al., 8
2013; Rodrı́guez and Joo, 2004; Whalen et al., 2013; Zhou, 2012). In addition, as oppose to 9
Rodriguez and Joo (2004), our study finds that GEV model outperform MNL and NL model. 10
11
In regards to the ASCs, most are statistically significant with the exception of a selected few. 12
With regards to the level of service variables, all are found to be statistically significant with the 13
sole exception of travel time in the MNL model. The value of travel time for these three models 14
varies in between 2.48$/hour to 3.83$/hour, which is little bit lower than the previous studies (y 15
($4.60 per hour) (Rodrı́guez and Joo, 2004). This lower value of travel time is intuitive since this 16
survey exclusively sampled students and the majority of them are not yet employed. 17
18
The travel time variables consist of total travel time, access time towards transit station (by walk, 19
bike or car) and transit wait time. The sign of travel time parameter is negative which is different 20
from the finding of Whalen and Paez (2013). This negative travel time parameter is intuitive, 21
since an individual is more likely to choose a mode which has lower travel time (Hasnine, 2015; 22
Hasnine et al., 2016). For CNL model travel cost is normalized by the commuting distance for all 23
transit modes to get statistically significant parameter. 24
25
Various household level mobility tool ownership level and socio-economic attributes are also 26
investigated in this study. From the marginal effect comparisons, it is found that the “number of 27
cars per household member” variable has a strong influence on the park and ride mode than the 28
driving mode (Figure 3). A Higher number of the household car allows the household member to 29
use the car without sharing it with someone, which encourages the park and ride mode. As such, 30
this finding is intuitive. For all three models, this variable shows similar results. Transit pass 31
ownership also influences commuters to choose transit related modes such as transit with walk 32
access, park and ride, kiss and ride, and bike and ride. From the marginal effect analysis, it is 33
found that the marginal effect of transit pass ownership for local transit is higher than for park 34
and ride, and kiss and ride. Similar result is found for the presto card ownership. 35
36
With consideration to gender, the mode choice behaviour of female students is investigated in 37
the context of downtown versus suburban campuses. The female students who commute to 38
downtown campuses and who commute to suburban campuses exhibit very distinct behaviour of 39
mode choice (Figure 3d and 3e). For instance, it is found that female students are more inclined 40
to use park and ride, and kiss and ride. This finding echoes the results of some previous studies 41
(Akar et al. 2013, Davison et al. 2015). Female students who travel to downtown campuses are 42
less inclined to choose walk, bike and bike and ride mode. These results are supported by the 43
female bike mode share of the City of Toronto, where only 35% of the people who bike to work 44
are female (City of Toronto, 2009). In the suburban campuses, it is found that females are more 45
15
inclined to choose driving or auto passenger in comparison to transit mode. The inadequacies of 1
transit services in the outskirts of Toronto likely play a significant role in this behaviour. 2
3
Age is another important variable for understanding the post-secondary students’ commuting 4
behaviour (Soria-Lara et al. 2017). In this study we classified students into two groups: (1) 5
students aged between 18 and 22, (2) students aged between 23 and 25. Empirical model results 6
show that there are significant differences in the mode choice behaviour of these two age groups. 7
For both age groups, utility provided by driving is taken as the reference utility value. The 8
students who are aged between 18 and 22 are less inclined to choose bike and ride as a 9
commuting mode. On the other hand, the student aged between 23 and 25 are less inclined to 10
choose auto-passenger, local transit with walk access and kiss and ride. This finding, in fact, 11
reveals a threshold age of 22 when the older youth change their previous mode preference, and 12
this older age cohort are more inclined to choose driving as a commuting mode. 13
14
Table 2. Model Estimation Result for MNL, NL and CNL 15
16
MNL NL CNL
Number of Observation 3208 3208 3208
The number of Parameter Estimated 35 36 38
Loglikelihood of the full model -2247.73 -2245.88 -2214.52
Loglikelihood of equiprobable model -4446.99 -4446.99 -4446.99
Rho-Square value against equiprobable model 0.4946 0.4950 0.5020
Adjusted Rho-Square value against equiprobable model 0.4867 0.4869 0.4934
17
Table 3. Travel level of service variables and Nesting variables 18
19
MNL NL CNL
Variable Name Mode Estimates t-stat Estimates
t-stat Estimate
s
t-stat
Alternative Specific
Constants
Driving 0.0
(Fixed) ---- 0.0(Fixed) ---- 0.0
(Fixed) ----
Auto passenger 1.147 3.624 0.893 2.532 0.131 0.367
Local transit 2.931 9.252 2.756 8.099 2.392 6.273
Park and ride -1.848 -6.945 -1.361 -4.138 -1.472 -4.085
Kiss and ride 1.132 3.513 1.271 4.320 0.943 2.862
Bike and ride -1.398 -2.793 -0.768 -1.531 -2.084 -4.158
Walk 8.563 16.435 8.367 15.466 7.332 10.777
Bike 1.606 1.318 1.616 1.324 0.966 0.741
Travel Cost All modes -0.075 -1.983 -0.069 -1.829 --- ---
Driving and
auto passenger --- --- --- --- -0.247 -6.363
Travel Cost/Distance in
Kilometers
Local transit,
park and ride,
kiss and ride,
bike and ride
--- --- --- --- -1.263 -3.650
Travel Time (In vehicle
travel time+out of
vehicle travel
time+waiting time)
All modes
-0.0031 -1.429 -0.0044 -2.052 -0.0142 -5.130
16
Distance Walk -1.825 -10.514 -1.836 -10.581 -1.815 -6.710
Bike -0.408 -7.414 -0.418 -7.602 -0.419 -7.175
Coefficient of Expected
Maximum Utility of
Transit Nest
--- --- 0.795 -2.111 0.671 -1.640
Coefficient of Expected
Maximum Utility of
Active Transport Nest
--- --- --- --- 0.953 -0.634
1
Table 4. Sociodemographic variables 2
MNL NL CNL
Variable Name Mode Estimates t-stat Estimates t-stat Estimates t-stat
Number of cars per
household member
Driving, Park
and Ride 4.614 11.623 4.101 8.549 3.697 7.856
Transit pass ownership
dummy (1=yes, 0=no)
Local Transit,
Park and Ride,
Kiss and Ride,
Bike and Ride
1.792 13.323 1.782 13.242 1.830 13.446
Presto Card ownership
dummy (1=yes, 0=no)
Local Transit,
Park and Ride,
Kiss and Ride,
Bike and Ride
0.896 6.685 0.886 6.575 0.657 4.882
Bike ownership dummy
(1=yes, 0=no)
Bike 1.666 1.830 1.587 1.754 1.282 1.428
Female Students
Dummy Who Commute
to Downtown Campus
Local Transit 0.980 4.992 0.946 4.903 0.812 4.178
Park and Ride 1.446 4.973 1.330 5.040 1.201 4.655
Kiss and Ride 1.273 5.460 1.205 5.458 1.026 4.777
Walk 0.772 2.441 0.743 2.357 0.531 1.692
Bike 0.816 2.762 0.781 2.653 0.578 1.961
Female Students
Dummy Who Commute
to Suburban Campus
Auto Passenger 0.742 2.925 0.717 2.849 0.930 3.674
Local Transit -0.705 -3.109 -0.705 -3.189 -0.406 -1.777
Park and Ride -1.168 -2.704 -1.105 -2.948 -0.763 -2.189
Kiss and Ride -0.857 -3.073 -0.804 -3.143 -0.485 -1.909
Walk -2.127 -5.046 -2.089 -4.969 -1.926 -4.521
Bike -1.766 -3.415 -1.829 -3.555 -1.575 -3.094
Age between 18 to 22 Bike and Ride -1.608 -1.479 -1.532 -1.723 -1.646 -1.559
Age between 23 to 25 Auto Passenger -0.828 -2.659 -0.718 -2.331 -0.732 -2.396
Local Transit -0.528 -2.927 -0.467 -2.807 -0.467 -3.014
Kiss and Ride -0.702 -2.714 -0.606 -2.696 -0.572 -2.855
Number of dependent
children per number of
household members
Driving 0.780 1.524 0.643 1.229 0.613 1.168
Bike -1.539 -1.932 -1.359 -1.703 -1.380 -1.736
3
4
5
6
7
17
Table 5. Land Use Variables 1
MNL NL CNL
Variable Name Mode Estimates t-stat Estimates t-stat Estimates
t-stat
Square Km of
commercial land
Driving -3.413 -2.190 -3.035 -1.950 -1.319 -0.756
Auto Passenger
1.230 1.021 1.237 1.025 1.062 0.876
Area of sidewalks in
kilometers squared
Walk 1.330 3.453 1.260 3.280 1.238 3.158
2
This study has also found influence from the number of dependent children on mode choice. The 3
influence of the number of children is captured through the normalized variable called “number 4
of dependent children per number of household members”. Households with a higher number of 5
children are less likely to choose biking as a mode and are more likely to choose driving. This 6
result is intuitive since household with a higher number of children will need to drop off their 7
children at school or the day care. Furthermore, seating for children is not accommodated easily 8
on a bike, and even when it can be, the number of children that can be accommodated is very 9
limited. There is also the safety concerns that most parents will have when bringing their 10
children on their bicycles. As such, it is more likely for a household with many children to 11
choose to use a car. 12
13
We have also incorporated several land use variables in the models. It is also found that if a 14
certain area has higher amount of sidewalks, it increases the likelihood of choosing walk mode. 15
The effect of the commercial land (square kilometer) is also investigated in this study. It is found 16
that students are less inclined to drive to school zones with larger areas of commercial land 17
(square kilometer). 18
19
20
21
18
1 2
Figure 3. Marginal Effect Comparison of Selected Variables. 3
4
Finally, in the CNL model, the bike and ride mode is observed to cutting across the transit and 5
active nests. This suggests that there is a strong inherent correlation between the active mode and 6
transit nests. This intuitively tells us that young people are interested in transit and they want to 7
bike as well. It is worth mentioning that, during peak hour Toronto Transit Commission does not 8
allow taking bikes on board a transit vehicle, a practice that is likely turning away student users 9
as suggested by the nest correlation. This finding tells us if bike and ride is encouraged during 10
peak hour commuting, there is likely a large group of bicyclists out there ready to take 11
advantage. 12
6. Policy Analysis 13
14
On the basis of the estimated models, it is possible to calculate how changes in various policies 15
will influence the mode shift of the sample. Since CNL model exhibits the best goodness-of-fit, 16
we have used CNL model for the policy scenario analysis. Figure 4 reaveals the change in mode 17
shares due to change of level of service variables for all eight modes. For example, Figure 4(a) 18
shows how the base mode share, in this case - driving, changes for each percent change of the 19
level of service values (travel time, cost). For driving, it is clear that students are particularly 20
sensitive travel cost and parking cost. This has signifiant policy implications. If parking cost is 21
reduced to 50%, it will encourage students to drive more (around 0.4% increase in driving mode 22
share), and vice versa. It is found that the effect of travel cost for auto-passenger is less than that 23
for driving. For trasit modes (transit with walk access, park and ride, and kiss and ride), it is 24
found that students are more sensitive to in-vehicle travel time than wait time and access time. 25
For transit with walk access specifically, it is found that a 50% reduction in the transit fare 26
increases the transit with walk access mode share to 2.0%. Therefore, providing some incentives 27
to trasit passes can encourage students to take local transit. This findings support many previous 28
studies (Rotaris and Danielis, 2014; Zhou, 2012, 2014, 2016).For bike and walk, it is found that 29
students are highly sesititve to commuting ditance, especially for walk mode. A 30% reduction in 30
commuting distance would increase the probability of choosing walk mode by 3.0%. 31
19
1 (a) (b) 2
3 (c) (d) 4
5 (e) (f) 6
20
1 (g) (h) 2
3
Figure 4. Change in Mode Shares due to change of Level of Service Variables 4
5
7. Conclusions and Recommendations for future studies 6
7
The objectives of this paper are to better understand: (1) how we can capture preference 8
heterogeneity and non-proportional substitution patterns through an advanced generalized 9
extreme value (cross-nested logit) model which can also provide better goodness-of-fit than 10
conventional models such as MNL and NL, and (2) how mode choice decisions are made by 11
students of a metropolitan area. A cross-campus university travel survey is exploited to 12
investigate post-secondary students’ mode choice behaviour. The models estimated and the 13
policy scenario analysis performed in this study are expected to contribute a better understanding 14
of the different factors (i.e., level if service, socio-economic and land use variables) that affect 15
students’ commuting mode choice. The major findings of this study are summarized as follows: 16
17
Rodrı́guez and Joo (2004) showed that MNL model outperforms NL and HEV models in terms 18
of goodness-of-fit. Our study actually shows that CNL model outperforms MNL and NL model 19
in terms of goodness-of-fit, which is commonly found in literature (Koppelman and Bhat, 2006). 20
The inclusion of the CNL model allows for investigation of a mode’s possible correlations with 21
multiple nests. The results revealed that bike on board mode are correlated with both the active 22
mode and transit mode nest. This suggests that if adequate bike and ride infrastructure is 23
provided, there is likely a large number of students who will take advantage of these facilities. In 24
particular, during morning peak period, bike and ride mode is still not encouraged by the transit 25
agencies in the city of Toronto. As such, if sufficient facilities and conducive policies are 26
provided to encourage this mode (i.e., allowing bikes on board transit vehicle during peak 27
periods), it will encourage a large number of student commuters to shift to bike and ride mode 28
from driving and auto passenger modes. 29
30
Our study shows that various level of service (travel time, cost, and distance) attributes have 31
significant policy implications in the context of public transit mode choice. Whalen et al. (2013) 32
highlighted the positive utility of travel time for students in their model. However, we found the 33
travel time parameter to be negative which echoes various past studies (Akar et al., 2013; 34
Rodrı́guez and Joo, 2004; Koppelman and Bhat, 2006). The marginal effect analysis in our study 35
21
shows that by reducing in-vehicle travel time we can encourage students to choose transit as 1
commuting mode. It is also found that providing some incentives to transit passes will encourage 2
post-secondary students to choose local transit as the commuting mode. The literature 3
summarized in section 2.1 supports this empirical finding (Rotaris and Danielis, 2014; Zhou, 4
2012, 2014, 2016). Similar to previous studies, our study shows that students are very sensitive 5
to commuting distance while choosing biking or walking as a commuting mode (Lundberg and 6
Weber, 2014; Rybarczyk and Gallagher, 2014). 7
8
Our study particularly tried to capture the differences in mode choice between female students’ 9
who commute to downtown campuses and who commute to suburban campuses. It is found that 10
these two groups exhibit very distinct behaviour in mode choice. It is found that female students 11
are more inclined to use park and ride, and kiss and ride when they commute to downtown 12
campuses. Female students who travel to suburban campuses are more likely to choose auto-13
passenger. This finding echoes the results of some previous studies (Akar et al. 2013, Davison et 14
al. 2015). The empirical model results show that students aged between 18 and 22 are less 15
inclined to choose bike and ride. Whereas, the student aged between 23 and 25 are less inclined 16
to choose auto-passenger, local transit with walk access and kiss and ride. The age cohort 23 to 17
25 are particularly interested to choose driving as their commuting mode. 18
19
The proposed modelling frameworks offer a flexible tool to better understand the travel 20
behaviour of a very influential segment of the population. However, the framework in this paper 21
can be extended by including multiple trips (tour based mode choice). In additions, this study can 22
also be extended by investigating the implication of dynamic discrete choice on tour based mode 23
choice modelling framework which will reveal more behavioural insight. 24
25
Acknowledgement 26
27
The study was partially funded by an NSERC Discovery grant. Authors acknowledge the 28
contributions of StudentMoveTO survey team for making the dataset available for this research. 29
Contributions of Zhejiang Wang and Yan Zhuang Tony are also acknowledged for support in the 30
data cleaning stage. However, views and opinions presented in the paper belong only to the 31
authors. 32
33
34
35
36
37
38
39
40
22
References: 1
1. Aptech Systems. (2016). Gauss User's Manual. http://www.aptech.com/. Accessed July 30, 2
2016 3
2. Akar, G., Fischer, N., & Namgung, M. (2013). Bicycling choice and gender case study: The 4
Ohio State University. International Journal of Sustainable Transportation, 7(5), 347-365. 5
3. Balsas, C. J. (2003). Sustainable transportation planning on college campuses. Transport 6
Policy, 10(1), 35-49. 7
4. Ben-Akiva, M. E., & Lerman, S. R. (1985). Discrete choice analysis: theory and application 8
to travel demand (Vol. 9). MIT press 9
5. Boyd, B., Chow, M., Johnson, R., & Smith, A. (2003). Analysis of effects of fare-free transit 10
program on student commuting mode shares: BruinGo at University of California at Los 11
Angeles. Transportation Research Record: Journal of the Transportation Research Board, 12
(1835), 101-110. 13
6. City of Toronto. 2009. City of Toronto Cycling Study. 14
7. Data Management Group. (2011). Transportation Tomorrow Survey 15
8. Davison, L., Ahern, A., & Hine, J. (2015). Travel, transport and energy implications of 16
university-related student travel: A case study approach. Transportation Research Part D: 17
Transport and Environment, 38, 27-40. 18
9. Danaf, M., Abou-Zeid, M., & Kaysi, I. (2014). Modeling travel choices of students at a 19
private, urban university: insights and policy implications. Case Studies on Transport 20
Policy, 2(3), 142-152. 21
10. Delmelle, E. M., & Delmelle, E. C. (2012). Exploring spatio-temporal commuting patterns 22
in a university environment. Transport Policy, 21, 1-9. 23
11. Duque, R. B., Gray, D., Harrison, M., & Davey, E. (2014). Invisible commuters: assessing a 24
university’s eco-friendly transportation policies and commuting behaviours. Journal of 25
Transport Geography, 38, 122-136. 26
12. Grimsrud, M., & El-Geneidy, A. (2014). Transit to eternal youth: lifecycle and generational 27
trends in Greater Montreal public transport mode share. Transportation, 41(1), 1-19. 28
13. Grimsrud, M., & El-Geneidy, A. (2013). Driving transit retention to renaissance: trends in 29
Montreal commute public transport mode share and factors by age group and birth 30
cohort. Public Transport, 5(3), 219-241. 31
14. Hasnine, M. S., Weiss, A., & Nurul Habib, K. (2017). Stated Preference Survey Pivoted on 32
Revealed Preference Survey for Evaluating Employer-Based Travel Demand Management 33
Strategies. Transportation Research Record: Journal of the Transportation Research Board, 34
(2651), 108-117. 35
15. Hasnine, M. S., Weiss, A., & Nurul Habib, K. (2016). Development of an Employer-Based 36
Transportation Demand Management Strategy Evaluation Tool with an Advanced Discrete 37
Choice Model in Its Core. Transportation Research Record: Journal of the Transportation 38
Research Board, (2542), 65-74. 39
16. Hasnine, M. S. (2015). A Comprehensive Study on the Effectiveness of Office based TDM 40
Policies (Masters dissertation, University of Toronto (Canada)). 41
23
17. Kamruzzaman, M., Hine, J., Gunay, B., & Blair, N. (2011). Using GIS to visualise and 1
evaluate student travel behaviour. Journal of Transport Geography, 19(1), 13-32. 2
18. Khattak, A., Wang, X., Son, S., & Agnello, P. (2011). Travel by university students in 3
Virginia: Is this travel different from travel by the general population? Transportation 4
Research Record: Journal of the Transportation Research Board, (2255), 137-145. 5
19. Kerr, A., Lennon, A., & Watson, B. (2010). The call of the road: factors predicting students’ 6
car travelling intentions and behaviour. Transportation, 37(1), 1-13. 7
20. Kim, J., Schmöcker, J. D., & Fujii, S. (2016). Exploring the relationship between 8
undergraduate education and sustainable transport attitudes. International Journal of 9
Sustainable Transportation, 10(4), 385-392. 10
21. Klöckner, C. A., & Friedrichsmeier, T. (2011). A multi-level approach to travel mode 11
choice–How person characteristics and situation specific aspects determine car use in a 12
student sample. Transportation Research Part F: Traffic Psychology and Behaviour, 14(4), 13
261-277. 14
22. Koppelman, F. S., & Bhat, C. (2006). A self instructing course in mode choice modeling: 15
multinomial and nested logit models. 16
23. Kuhnimhof, T., Wirtz, M., & Manz, W. (2012). Decomposing Young Germans' Altered Car 17
Use Patterns: Lower Incomes, More Students, Decrease in Car Travel by Men, and More 18
Multimodality. Transportation Research Record: Journal of the Transportation Research 19
Board, (2320), 64-71. 20
24. Lavery, T. A., Páez, A., & Kanaroglou, P. S. (2013). Driving out of choices: An 21
investigation of transport modality in a university sample. Transportation research part A: 22
policy and practice, 57, 37-46. 23
25. Limanond, T., Butsingkorn, T., & Chermkhunthod, C. (2011). Travel behavior of university 24
students who live on campus: A case study of a rural university in Asia. Transport 25
policy, 18(1), 163-171. 26
26. Lundberg, B., & Weber, J. (2014). Non-motorized transport and university populations: an 27
analysis of connectivity and network perceptions. Journal of Transport Geography, 39, 165-28
178. 29
27. Miller, E.J., Kucirek, P., & Vaughan, J. (2015). A Mode by Any Other Name, presented at 30
the 14th International Conference on Travel Behaviour Research, Windsor, UK, July 19-23 31
28. McFadden, D. (1973). Conditional logit analysis of qualitative choice behavior. 32
29. Miralles-Guasch, C., & Domene, E. (2010). Sustainable transport challenges in a suburban 33
university: The case of the Autonomous University of Barcelona. Transport policy, 17(6), 34
454-463. 35
30. Rodrı́guez, D. A., & Joo, J. (2004). The relationship between non-motorized mode choice 36
and the local physical environment. Transportation Research Part D: Transport and 37
Environment, 9(2), 151-173. 38
31. Rotaris, L., & Danielis, R. (2014). The impact of transportation demand management 39
policies on commuting to college facilities: A case study at the University of Trieste, Italy. 40
Transportation Research Part A: Policy and Practice, 67, 127-140. 41
32. Rybarczyk, G., & Gallagher, L. (2014). Measuring the potential for bicycling and walking at 42
a metropolitan commuter university. Journal of Transport Geography, 39, 1-10. 43
24
33. Sultana, S. (2015). Factors associated with students' parking-pass purchase decisions: 1
Evidence from an American University. Transport Policy, 44, 65-75. 2
34. Shannon, T., Giles-Corti, B., Pikora, T., Bulsara, M., Shilton, T., & Bull, F. (2006). Active 3
commuting in a university setting: assessing commuting habits and potential for modal 4
change. Transport Policy, 13(3), 240-253. 5
35. Soria-Lara, J. A., Marquet, O., & Miralles-Guasch, C. (2017). The influence of location, 6
socioeconomics, and behaviour on travel-demand by car in metropolitan university 7
campuses. Transportation Research Part D: Transport and Environment, 53, 149-160. 8
36. Toronto Transit Commission. (2016). TTC Presto Stations. 9
https://www.ttc.ca/PDF/Maps/Presto_%20Station_Map.pdf. Accessed July 30, 2016 10
37. Train, K. (2003). Discrete choice methods with simulation. Cambridge university press. 11
38. Uttley, J., & Lovelace, R. (2016). Cycling promotion schemes and long-term behavioural 12
change: A case study from the University of Sheffield. Case Studies on Transport 13
Policy, 4(2), 133-142. 14
39. Wen, C. H., & Koppelman, F. S. (2001). The generalized nested logit model. Transportation 15
Research Part B: Methodological, 35(7), 627-641. 16
40. Whalen, K. E., Páez, A., & Carrasco, J. A. (2013). Mode choice of university students 17
commuting to school and the role of active travel. Journal of Transport Geography, 31, 132-18
142. 19
41. Williams, H. C. (1977). On the formation of travel demand models and economic evaluation 20
measures of user benefit. Environment and planning A, 9(3), 285-344. 21
42. Zhan, G., Yan, X., Zhu, S., & Wang, Y. (2016). Using hierarchical tree-based regression 22
model to examine university student travel frequency and mode choice patterns in China. 23
Transport Policy, 45, 55-65. 24
43. Zhou, J. (2016). Proactive sustainable university transportation: Marginal effects, intrinsic 25
values, and university students' mode choice. International journal of sustainable 26
transportation, 10(9), 815-824. 27
44. Zhou, J. (2014). From better understandings to proactive actions: Housing location and 28
commuting mode choices among university students. Transport Policy, 33, 166-175. 29
45. Zhou, J. (2012). Sustainable commute in a car-dominant city: Factors affecting alternative 30
mode choices among university students. Transportation research part A: policy and 31
practice, 46(7), 1013-1029. 32
33
34
35
