1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 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 43 44 45 46
Effects of Accessibility to the Transit Stations on Intercity Travel Mode Choices in Contexts of High Speed Rail (HRS) in the Windsor-Quebec Corridor in Canada
Billy Wong, M.A.Sc Department of Civil Engineering University of Toronto Email:
[email protected]
Professor Khandker Nurul Habib Department of Civil Engineering University of Toronto Email:
[email protected]
Wong & Habib 2014 1
47
Abstract
48
Main objective of this paper is investigating the role of transit station accessibility on intercity
49
travel mode choices in contexts of a proposed High Speed Rail. The study area is the Quebec-
50
Windsor Corridor, which is the most important corridor in Canada and one of the most important
51
corridors in North America. A web-based joint Revealed Preference (RP)-Stated Preference (SP)
52
survey is used to collect data for empirical investigation. To contribute further to travel survey
53
methods, an innovative social media based data collection approach is taken. As opposed to
54
explicit sample frame-based sample selection approach, it applies a reverse procedure of open
55
sample frame-based data collection. The web-based survey is spreaded through social media
56
groups (that are open in sense that information of all individuals are not known explicitly) and
57
the collected responses are screened to match with population distributions. Results prove the
58
potential of such data collection approach in extracting representative samples of the population
59
of concern. The collected dataset, which has close representation of the population, is used to
60
estimate discrete mode choice model (Nested Logit model) of intercity mode choices. Empirical
61
model reveals that intercity travellers are more concerned about access to and egress from transit
62
stations than the main in-vehicle travel while selecting intercity travel modes. The result of this
63
investigate imply that transit station accessibility should be given careful consideration for the
64
success of any innovative travel mode, e.g. High Speed Rail.
65 66 67
Key Words: Intercity travel, transit accessibility, high speed rail, nested logit model
Wong & Habib 2014 2
68
INTRODUCTION
69
The continued economic growth and development of urban areas within Canada have created
70
corridors of population, economic activity, and movement of individuals. Quebec-Windsor
71
corridor is one of such corridor. This corridor serves as the passage of trades and comers
72
between Canada and US. This is the busiest corridor in Canada and one of the busiest corridors
73
in North America. Movement of individuals within such corridor has become the subject of
74
numerous studies from economics, social science, and transportation research groups. Within
75
federal, provincial, and municipal transportation development sectors, considerable efforts are
76
dedicated to assessing the different modes of passenger and freight movement within these
77
corridors (IBI 2002). With increased passenger volumes travelling within these corridors,
78
alternative transportation modes have been proposed and researched in the past without
79
implementation. One such is a high speed rail system (HSR), which has been studied since the
80
early 90’s (Langan 2011) However, increasing oil price in recent years has generated renewed
81
interest in HSR in Quebec-Windsor corridor (Miller 2004). This also generated renewed need
82
for improved understanding on our intercity travel bahaviour. In assessing intercity travel
83
demand, there are issues with data availability on intracity travel. Current publicly available
84
datasets on Canadian intercity travel lack local access and egress information, which is predicted
85
to be a significant factor in intercity mode choice. When travel data is aggregated to the
86
metropolitan level, interpolation of local travel activity is required, which may negatively
87
influence the validity of resulting demand models (Wilson et al 1990). With the lack of local
88
accessibility aspects, it is necessary to create a new survey framework to collect travel data for
89
demand modeling. This paper contributes to two aspects of intercity travel research: travel
Wong & Habib 2014 3
90
survey to collect data on intercity travel and evaluating the influences of transit station
91
accessibility on intercity passenger travel mode choice.
92
Contribution to travel survey for intercity travel includes designing a web-based hybrid
93
survey and application of social media based survey data collection. The hybrid survey combines
94
revealed preference (RP) questions on intercity travel experience and stated preference (SP)
95
experiment on mode choices in context of HSR in the Quebec-Windsor corridor. In terms of data
96
collection, we considered an innovative approach of recruitment process. Recognizing the
97
difficulty of defining the sample frame for a large corridor that crosses multiple provinces, we
98
investigated non-incentivized online recruitment. We investigate the potential of using social
99
media based recruitment process for web-based travel survey. Unlike conventional approach of
100
defined sample frame based recruitment, the social media based approach is an open form
101
approach. The open form refers to the uncontrolled or unknown sample frame, which also
102
resembles a snowball sampling approach. Post data collection, the dataset is matched with
103
overall population characteristics and, if necessary, re-sampled from the collected dataset to
104
match the population characteristics. This approach of data collection for intercity travel is
105
validated in the field for the Quebec-Windsor corridor and it is proven to be effective in
106
collecting representative dataset within a short period of time. The collected dataset is then used
107
to develop discrete choice model of intercity passenger travel mode choice. The estimated model
108
reveals that people put higher value on access to and egress from transit stations than the main
109
mode’s travel time/cost while choosing intercity mode choice. It becomes clear that
110
competitiveness of various transit options (including HSR) depends largely on accessibility to
111
local station locations.
Wong & Habib 2014 4
112
The paper is organized into a number of sections. The next section presents a brief
113
literature review on intercity mode choice investigation. This section is followed by the section
114
on survey design, data collection and empirical investigation. The paper concludes with key
115
findings and policy recommendations.
116 117
LITERATURE REVIEW
118
Intercity travel is defined as a trip that passes through the boundaries of an urban center.
119
Typically an intercity trip would start and end within an urban center. Intercity travel demand is
120
the frequency of trips made between urban centers as well as the modes of travel used. While
121
intercity travel demand research have been published since the 1960s, these models has not been
122
improved as rapidly as the more urban center specific models. Some hypothesized reasons for
123
the slow development of intercity models are due to fewer intercity travel corridors of interest to
124
policy makers in comparison to urban regions, unclear jurisdiction of intercity corridors, larger
125
private sector stake resulting in proprietary information, open-ended definition of study area,
126
limited existing data available for research purposes, and unwilling commitment to invest in
127
long-term research by governments and public agencies (Langan 2011).
128
Application of models such as the logit model to assess intercity travel demand has not
129
been as successful as the urban center counterparts. Aside from the lack of suitable data, there
130
are other fundamental reasons. Compared to daily local journeys (trips to work, school,
131
shopping), intercity trips are made less frequently than urban trips. Due to the lower frequency of
132
intercity travel, there is difficulty modeling a trip maker’s decision to visit another city with logit
133
models using existing data. Unless daily intercity trips are made, a longer time period of trip
134
information collection is required (Sonesson 2011).
Wong & Habib 2014 5
135
An empirical study in using disaggregate choice model of intercity travel demand in the
136
Quebec City – Windsor Corridor, published in 1988 used the 1969 Canadian Transport
137
Commission (CTC) survey to estimate mode choice in the Quebec-Windsor Corridor (Ridout
138
and Miller 1982). The major weaknesses in the dataset was the lack of automobile choice, which
139
limited the accuracy of the demand forecasting logit model as automobile comprises a large
140
market share in intercity travel. While the 1980 Canadian Travel Survey (CTS) was available as
141
a dataset alternative, the aggregation of trip origins and destinations to the Census Metropolitan
142
Area (CMA) would not provide relevant trip access and egress information. The resulting models
143
were all calibrated using a standard maximum likelihood logit estimation procedure. Three
144
models were estimated with each model representing different market segmentation (business,
145
pleasure, or personal). For the business market, it was observed that access was a generic
146
attribute while egress was alternative specific. For non-business and non-personal intercity trips,
147
access was estimated as an alternative specific attribute and egress was generic. For personal
148
intercity trips, only local access, as a generic attribute, was estimated to be statistically
149
significant to mode choice. There was difficulty in estimating the access and egress term, which
150
may be attributed to the use of access/egress distance instead of cost and time. Socioeconomic
151
variables were found to have minimal explanatory function in intercity mode choice.
152
Wilson et al (1990) used the 1985 Canadian Travel Survey (CTS) to estimate MNL
153
models for intercity passenger travel. They developed models for both business and non-business
154
trips in the eastern and western Canadian regions. They proved that the CTS can be effectively
155
used for intercity mode choice modelling when datasets for such investigations are scarce.
156
Forinash and Koppelman (1993) applied and compared a Nested Logit (NL) model
157
against a MNL model using an RP dataset generated from the VIA Rail’s (a national carrier
Wong & Habib 2014 6
158
serving along the Windsor-Quebec corridor) 1989 Passenger Review dataset. They investigated
159
the effect of rail service improvements for weekday business travel in the Toronto-Montreal
160
corridor (which is a part of Windsor-Quebec corridor). They justified the advantage of the NL
161
over the MNL by relaxing the Independent and Irrelevant Alternatives (IIA) property of the
162
MNL to capture correlation among unobserved attributes of similar modes. Bhat (1995) used the
163
same dataset to examine the impact of improved rail service on intercity business travel in the
164
Toronto–Montreal corridor. The focus of this paper was mostly to test an advanced formulation
165
of discrete choice models in terms of performance over the conventional logit model. He found
166
that, compared to the MNL, the Heteroskedastic Extreme Value (HEV) model predicts smaller
167
increases in rail shares and smaller decreases in non-rail shares. The HEV model should allow
168
for cross-elasticity among alternatives compared to a nested logit model and require less
169
computational complexity compared to the multinomial probit model. To test the application of
170
the HEV model, the 1989 Rail Passenger Review from VIA Rail was used. The main focus was
171
on paid business travel in the Quebec-Windsor Corridor and confined between the air, rail, and
172
auto travel modes as travel for non-personal business purposes had less than 1% market share.
173
Five different models were estimated; a multinomial logit model, three nested logit models, and
174
the heteroskedastic extreme value model. From model estimation, the nested logit structure was
175
not significantly better than the multinomial logit models. Compared to both MNL and NL, the
176
HEV model was able to predict smaller changes in level-of-service changes, which may point to
177
an improvement of the HEV model over the commonly used MNL and NL model formulations.
178
Mandel et al (1997) present an intercity mode choice model for Germany. They use a logit model
179
with nonlinear Box-Cox transformations of key level of service variables. Their efforts were
180
mostly focused on capturing non-linear responses with respect to travel time and travel cost for
Wong & Habib 2014 7
181
high-speed rail. Hensher et al. (1999) applied the HEV model to estimate an intercity travel
182
mode choice model using a combination of RP and SP datasets with the objective of identifying
183
the market for a proposed high-speed rail service in the Sydney–Canberra corridor. They found
184
that much more uncertainty exists in the evaluation of non-car modes than of the car mode for
185
that corridor.
186
It is often hypothesized that an individual’s responsiveness to level-of-service variables
187
affects that individual’s mode choice (Bhat 1998). In this paper, Bhat accommodates variations
188
in this responsiveness within a multinomial logit based model. Monte Carlo simulation
189
techniques were also incorporated to approximate the choice probabilities, which are a technique
190
that has been used in empirical applications in the economics field and relatively new to
191
transportation researchers. The 1989 Rail Passenger Review dataset from VIA Rail is used once
192
again to develop the travel demand models. Weekday business-based market segment was
193
selected and recorded trips by bus were omitted from the dataset due to a small percentage of
194
market shares. Three models were estimated in this paper; a multinomial logit model, a fixed-
195
coefficient logit model, and a random coefficients logit model. Level-of-service variables
196
included service frequency, total travel cost, in-vehicle travel time and out-of-vehicle travel time
197
along with socioeconomic variables; income, sex, travel group size, and large city indicator.
198
From model estimation, both the fixed-coefficient logit and random coefficient logit models
199
showed differences in sensitivity of level-of-service variables based on the socioeconomic
200
characteristics of the trip maker, rejecting the response homogeneity assumption of the MNL
201
model. Bhat (1998) shows that not accounting for variations in responsiveness across individuals
202
may lead to inappropriate estimation of mode choice, which may affect policy decisions.
Wong & Habib 2014 8
203
Wen and Koppelman (2001) proposed a general model structure which they refer to as the
204
general nested logit model (GNL), and also applied it to the same 1989 VIA Rail dataset to
205
analyze the effect of rail service improvements. However, the study’s main focus was the general
206
nature of the GNL model formulation and the derivation of the other GEV model structures as
207
special restrictive cases of the GNL model or as approximations to restricted versions of the
208
GNL model. The GNL model is conceptually appealing because of its very general structure and
209
flexibility, while maintaining closed form expressions. However, in practice, informed
210
restrictions that are customized to the application context must be imposed on the GNL model
211
formulation, and thus the flexibility of the GNL model can be realized only if a large number of
212
dissimilarity and allocation parameters are estimated. Grisolía and Ortúzar (2010) recently
213
applied mixed logit approach to investigate the role of travel time perception of on inter-island
214
mode choice model for a limited number of modes (plane and ferries).
215
However, majority of recent studied on intercity passenger travel related studies focus mostly
216
specific aspects of specific intercity mode choices. Among many examples of such
217
investigations, Ahsan et al (2003) investigated market segmentations and role of socio-economic
218
variables on preferences variations among the passengers of intercity bus travellers. Rojo et al
219
(2008) investigated satisfactions of passengers of intercity bus traveller and Rojo et al (2011)
220
investigated how gender plays role in defining passenger satisfactions of intercity bus services.
221
Debrezion et al (2009) applied discrete choice model to investigate relationship between access
222
mode and station choice for intercity trains. Hsio and Hensen (2011) presented models for
223
forecasting demands of intercity air transportation. Rojo et al (2012) investigated the role of
224
service qualities in influencing choice of bus transit for intercity bus only. They did not conduct
225
any systematic modelling exercise of intercity mode choice modelling.
Wong & Habib 2014 9
226
The typical barrier to developing comprehensive understanding on intercity mode choice
227
behaviour is the lack of sufficient data for estimating empirical models. Since intercity trips are
228
not often part of household travel surveys, there are no systematic data collection programs
229
normally available for intercity trips. National travel survey data used for intercity mode choice
230
investigations lack significant details of intercity trip components. Datasets used for majority of
231
research investigations (as reviewed in this section) are collected as parts of specific projects by
232
different government or private organization. Such dataset are mostly designed to collect
233
information of specific aspects of intercity trips than to comprehensively understand intercity
234
travel behavior. To fill this gap in intercity travel research, we designed a comprehensive data
235
collection program considering the Quebec City to Windsor Corridor (QWC) of Canada as the
236
study area.
237 238
SURVEY DESIGN
239
The first task is to create a survey that can collect the required travel information while
240
addressing existing research and dataset gaps. The survey used for this paper collected
241
information on a respondent’s previous experiences of intercity trip, travel behaviour related to
242
non-intercity trips, socioeconomic characteristics in addition to a stated preference (SP) choice
243
experiment. The following sections detail each aspect of the survey design.
244 245
Revealed Preference (RP) Travel Information of Intercity Trips
246
The respondent is asked to provide travel details of his or her last intercity trip within the QWC
247
including basic travel information such as; the destination city, travel mode, travel cost, and trip
248
purpose. Aside from correlating the collected revealed intercity travel data with existing national
Wong & Habib 2014 10
249
travel datasets, the collection of this data may reveal any mode switching tendencies given the
250
introduction of a new travel mode in the QWC.
251 252
Revealed Preference Information on Non-Intercity Travel
253
A set of local travel revealed preference questions was also included in the survey. The
254
collection of this data may be used to correlate daily travel patterns with intercity modal choice.
255
However, the main purpose of the local revealed preference data is to cross reference the
256
demographics of the respondents from this survey with demographics from other travel datasets;
257
such as the ones from Transportation Tomorrow Survey (TTS) from University of Toronto’s
258
Data Management Group.
259 260
Socioeconomics Information of the Respondent
261
This section of the survey inquires about the respondent’s individual and household
262
socioeconomic information. The range of questions included is historically shown to influence
263
transportation-related choices including; age, gender, marital status, employment, household
264
size, household auto ownership, and household income.
265 266
Stated Preference Survey of Intercity Mode Choice
267
The section most useful to modeling mode choice is the intercity mode choice stated preference
268
section. While revealed preference information indicates current travel patterns, the potential
269
modal shift due to the introduction of a currently non-existent high speed rail (HSR) mode
270
cannot be modeled using RP data. The basis of the stated preference survey is presenting the
Wong & Habib 2014 11
271
respondent with a hypothetical intercity travel scenario and inquiring about the respondent’s
272
mode choice.
273
For this paper, a hypothetical intercity trip between an individual’s specified home
274
location in the Greater Toronto Area (GTA) to a location in Montreal is tested. Montreal was
275
chosen as the destination location because existing travel surveys indicate that the greatest
276
percentage of intercity trips within the Quebec City/Windsor corridor and originating from the
277
GTA has a destination in Montreal (Statistics Canada 2010).
278 279
Addressing for previous research gaps, this paper requires greater geographical disaggregation
280
(than previous research) to specifically focus on the effects of local accessibility on intercity
281
mode choice. Unlike existing travel surveys that record origin and destination locations at the
282
cities level, a higher degree of geographical disaggregation is required to obtain local
283
accessibility characteristics. In this paper, the respondent’s origin location is recorded at Forward
284
Sortation Area (FSA) levels, which are the first three digits of a postal code. Utilizing FSA’s is
285
the optimal compromise between stated and perceived local travel times, where further
286
geographical disaggregation may not yield significantly improved estimation results. Similarly,
287
the destination location in Montreal is aggregated to its 19 boroughs. A higher degree of
288
aggregation was chosen for the destination to reduce the necessity to search out a FSA code.
289 290
Attributes and levels in the SP Experiment:
291
A set of attributes was generated to provide the respondent relevant information regarding each
292
travel mode. An initial list of attributes was based on a TSRB study on a high speed rail line in
293
London, UK (Burge et al 2011) and modified to better relate to the Canadian region and existing
Wong & Habib 2014 12
294
Canadian travel mode alternatives. The attribute levels are the various values that can be
295
associated for each attribute. Levels can either be continuous values associated with the attribute
296
(such as a dollar amount with travel cost) or categorical values that associated with a description
297
(such as 1 for booked seating and 2 for representing arranged seating). Table 1 presents the list of
298
the attributes used in the stated preference survey.
299 300 301
TABLE 1: Attributes Considered in Stated Preference Survey Attribute Available Modes Levels Access Mode B/R/A/HSR Car/Transit/Taxi/Non-motorized Access Time B/R/A/HSR Current Travel Time * [0.75/1.00/1.25] Egress Mode B/R/A/HSR Pick Up/Transit/Taxi/Car Rental/Non-motorized Egress Time B/R/A/HSR Current Travel Time * [0.75/1.00/1.25] % On Time C/B/R/A/HSR 70%/80%/90% Departures per Day B/R/A/HSR Current Stated Frequency * [0.5/1.0/1.5] Number of Transfers B/R/A/HSR Direct/1 Transfer/2 Transfers Seat Choice B/R/A/HSR Pre-booked/Assigned/First come first serve Trip Information B/R/A/HSR Mobile schedule/Real-time/Pre-posted Main Travel Time C/B/R/A/HSR Current Travel Time * [0.75/1.00/1.25] Main Travel Cost C/B/R/A/HSR Current Travel Costs * [0.75/1.00/1.25/1.50] Premium Travel Cost R/A/HSR Travel Cost * [1.25/1.50/1.75/2.00] C = car, B = bus, R = rail, A = air, HSR = high speed rail
302 303
With the list of attributes and its respective levels established, an orthogonal stated preference
304
design was completed. A current lack of empirical models that also incorporate all the proposed
305
attributes hindered the use of an efficient SP design. Figure 1 below is one of six scenarios
306
shown to the respondent in the web-based survey. The values populating the SP table is based on
307
the respondent’s stated origin and destination locations and outputs the appropriate travel time
308
and costs while accounting for the different attribute levels. All elements shown in Figure 1 are
309
automatically generated and stored prior to deployment of the survey. Time and distance based
310
elements (access/egress time, travel time, and travel costs) are aggregated at the FSA level and
311
associated times and costs were used during the survey design stage via the use of mapping tools
Wong & Habib 2014 13
312
such as Google Maps. Travel time and costs were determined using publically available tools
313
such as Google Maps. As FSA level geographical aggregation was used, the same base-values
314
would be presented to all respondents in the same FSA. This data was collected and stored as
315
static backend elements that were tied to a specific web URL that referenced the FSA. When the
316
survey asks for the respondents’ FSA origin, the appropriate subset of survey scenarios would be
317
pulled from the survey backend. Total travel time and total travel costs values were also
318
displayed at the bottom of the table to provide assistance to the respondent and reduce survey
319
fatigue.
320
321 322 323 324
FIGURE 1: Single SP Scenario from Survey Design Access/egress time for the SP experiment:
325
Access time is defined as the time required traveling from a respondent’s home location
326
(aggregated to the centroid of the home location’s FSA address) to the specified intercity mode
327
departure station, using the corresponding access mode. The purpose of the access time attribute
328
is to assess if local accessibility has any influence on intercity modal choice. For example, it is
329
relatively easy to travel to all three departure stations from a Toronto downtown location;
Wong & Habib 2014 14
330
however, a similar trip may take longer if travelling from an Ajax home. The baseline access
331
times were determined by using Google Maps to query directions between a FSA centroid
332
location and the three intercity departure locations (Toronto Coach Terminal, Union Station, and
333
Lester B. Pearson Airport). Google Maps was used as the service provided travel times by
334
automobile, transit, as well as non-motorized transport.
335 336
Access/Egress Mode:
337
The access mode attribute is defined as the local mode choice option that the respondent would
338
use to travel to the departure station of an intercity travel mode alternative. This access trip
339
should originate from the respondent’s home location (aggregated to the geographical center of
340
the respondent’s home FSA address) and end at the location of the intercity mode’s departure
341
station. In conjunction with access time the inclusion of the access mode is to assess whether or
342
not the local accessibility of transportation alternative has an effect on a respondent’s intercity
343
mode choice. The possible access mode alternatives are typical travel methods currently used to
344
access the designated intercity departure stations. The cost of utilizing various access modes is
345
also presented to the respondent and is a stratified price based on the access mode. The decision
346
to include a generalized range of travel cost was done to simplify the survey design. In addition,
347
it was assumed that travel cost in each access mode alternative were mutually exclusive. For
348
example, the cost of transit would range between $3 and $6 whereas a trip by taxi with the same
349
origin and destination would be over $10. It is assumed that a respondent would only require one
350
access mode to get from their home location to the intercity mode departure station. However,
351
incorporating multiple access modes would introduce more complexity into the SP design
352
Wong & Habib 2014 15
353
SURVEY IMPLEMENTATION AND DATA
354
The survey is designed as a web-based survey that utilizes online social media sources as the
355
primary method of data collection. With a heavy focus on online social media, a data collection
356
strategy was designed to assess the benefits of web-based data collection.
357 358
Data Collection Method
359
The survey invitations were distributed among different social media sources. Each social media
360
source contains different number of members and has different online activity patterns. So,
361
sample recruitment through these multiple social media sources can be analogized to a pyramid,
362
where each increasing level up the pyramid adds additional survey responses to the total amount;
363
however, the number of expected respondents at each increasing level is decreased due to a
364
reduced social network reach.
365
Under this pyramid analogy, the first step would be to distribute the web-based survey
366
using a social network platform that would produce a large number of respondents knowing that
367
a large percentage of these respondents would fit a narrow socioeconomic demographic. Existing
368
social platforms fitting this would be Facebook and Twitter. The next step would be to distribute
369
the survey onto other networks that may expand to different socioeconomic and demographic
370
groups. For example, university mailing lists may include people outside of the author’s personal
371
Facebook social networks; however, there are fewer people on the listserv and the lack of
372
recognition to the author may lower probability that potential respondents would complete the
373
distributed survey. This sampling technique has an inherent bias of oversampling some
374
respondents, resulting in a biased socioeconomic profile, while under-sampling responses, with a
375
more representative socioeconomic profile, from limited social network reach. Given a large set
Wong & Habib 2014 16
376
of collected data, it is possible to draw a subset from the oversampled demographics; where
377
multiple draws from the same larger subset can be used to validate the model.
378 379
Data Collection Results
380
Overall, the data collection procedure was satisfactory in obtaining complete survey responses.
381
While the number of total responses did not meet initial expectations, the costs to obtain the
382
responses were kept at a minimal. Table 2 below is a summary for each collector source.
383 384 385 386 387
TABLE 2: Summary of Data Collection Program Collector Source: Social Media Groups Facebook – Profile Facebook – Skule Group Facebook – MBA Group Reddit – U of T Reddit – Toronto Skule Nite Intercept UTEK Skulebook EngSoc Digest Transport Email Listserv Metrolinx Employees External Source SurveyMonkey Panel Total
Group Type Close social network Distant social network Personal Favors
Email listings Survey panel
Approximate Size 774 2,745 109 2,314 26,248 27 unknown unknown ~2,000 ~50 unknown unknown unknown
Surveys Started 36 47 10 22 4 8 11 26 122 24 25 13 213 561
Completed Surveys 24 37 10 15 3 7 8 17 81 20 19 10 179 430
388
There are several metrics that can be used to gauge the relative success of each collector source.
389
Each metric has its own validity and has other explanatory variables that are hard to explicitly
390
measure, such as an individual’s social reach. Some possible metrics are; ratio of completed
391
surveys to collector size, ratio of completed surveys to surveys started, ratio of surveys started to
392
approximate size, completed surveys accounting for social reach and collector size. It is evident
Wong & Habib 2014 17
393
that the relationship between the approximate size of the collector source and the number of
394
completed surveys is not linear. This observation is in line with initial expectations that posting
395
survey requests in social media networks outside of an individual’s own social network yields a
396
lower number or responses due to the reduced familiarity between members in the social network
397
with the researcher and/or research project. Especially with a lack of monetary or prize incentive,
398
there is less willingness for individuals to dedicate time on something with little immediate
399
payout.
400
To validate the geographical distribution of collected responses, population and dwelling
401
counts from Canadian Census data was used as comparison baselines. While there are other
402
potential metrics to validate the collected responses, the emphasis on local accessibility is the
403
main motivator to use household population and dwelling counts. One of the main concerns
404
during the data collection process was a possibility where outlying cities in the Greater Toronto
405
Area are underrepresented in comparison with the City of Toronto. With the relatively small
406
number of collected responses compared to the Census population and dwelling counts, the count
407
of responses, population, and dwellings were calculated as percentages relative to the specified
408
list of cities in the survey design. By changing counts into percentages, it is possible to compare
409
the relative distribution of values across the GTA. Figure 2 below compares the distribution of
410
origin locations between survey respondents and Statistic Canada’s 2011 Census figures
411
(Statistics Canada 2013) of population and household totals.
Wong & Habib 2014 18 80% 70% 60%
% Share
50% 40% 30% 20% 10%
2011 Census Population
412 413 414 415
2011 Census Household
Whitchurch-Stouffville
Whitby
Vaughan
Uxbridge
Toronto
Scugog
Richmond Hill
Pickering
Oshawa
Oakville
Newmarket
Mississauga
Milton
Markham
King
Halton Hills
Georgina
East Gwillimbury
Clarington
Caledon
Burlington
Brampton
Aurora
Ajax
0%
Collected Data from Current Survey
FIGURE 2: Comparison between Household Density and Density of Respondents Looking at the distribution of respondents, the overall trend of respondent distribution
416
does follow the distribution of household of 2011 Census; however, the major areas of variance
417
are the outlying areas away from downtown Toronto. This variance is most apparent at the FSA
418
level but is less pronounced when aggregated to the cities level. It is also clear that the spatial
419
coverage of the sample is smaller than the population coverage as represented in census in the
420
study area. However, it should be noted that Toronto is over-represented than other parts in the
421
collected dataset.
422
To validate the socioeconomic distribution of collected responses, age and income
423
attributes were examined based on empirical results. Table 3.1 and 3.2 below are the descriptive
Wong & Habib 2014 19
424
statistics table that summarizes the distribution of respondent age and income and compares the
425
median collected values with both TTS and Statistics Canada data.
426
TABLE 3.1: Descriptive Statistics - Age Age Group Between 18 and 24 Between 25 and 30 Between 31 and 40 Between 41 and 50 Between 51 and 65 Over 65 Undisclosed Mean Age of Respondents Mean Age of Respondents (2011 TTS)
Number of Survey Respondents 226 63 67 35 31 4 4 29.45 38.8
TABLE 3.2: Descriptive Statistics - Income Household Income Group Under $20,000 $20,000 to $29,999 $30,000 to $39,999 $40,000 to $49,999 $50,000 to $59,999 $60,000 to $69,999 $70,000 to $79,999 $80,000 to $89,999 $90,000 to $99,999 Over $100,000* Undisclosed Mean Household Income of Respondents Mean Household Income (Statistics Canada 2013)
Number of Survey Respondents 42 20 24 35 21 21 31 34 33 95 74 $79,929.47 $71,210.00
427 428
429 430 431
*maximum household income was assumed at $200,000 at expected average income of $150,000 for that income group.
432
The mean age from survey respondents was considerably lower than that of the 2011 TTS
433
data; however, this variance may be explained by the larger bias of younger respondents from the
434
data collection program. When looking at mean income, the values between the survey
435
respondents and the 2012 Statistics Canada were similar. The higher mean income of survey
436
respondents may be due to the sampled group.
437
EMPIRICAL INVESTIGATION
438
Validating the data as representative of GTA residents, a discrete choice model of intercity mode
439
choice is developed by using the collected dataset. It is assumed that the utility of a choice
Wong & Habib 2014 20
440
alternative “i” in the choice set “t” (
441
(
442
Type I Extreme Value distribution of the random element, the multinomial logit model (MNL) is
443
developed, which is often used for intercity mode choice analysis (Ashiabor et al 2007). In this
444
paper, we are clear that no all alternatives are independent. They may be number of alternatives
445
having common features that may have influences on mode choices. So, we used nested logit
446
model instead. The nested logit model the limitations of the multinomial logit model, considering
447
that there are multiple levels of conditional choices included in the model in the form of “nests”
448
(Ben-Akiva and Lerman 1985). In the nested logit model, the probability of choosing an
449
alternative “i” will equal to the probability of choosing the alternative “i” conditional to choosing
450
some subset “k”. The full equation to determine the probability of choosing alternative “i” is:
) and a random component (
) is based on the combination of a systematic component ). Considering the Independent and Irrelevant Distributed
451
(1)
452
453 454 455 456
Where, k,
p
defines the scale parameters of nest k and p
defines the root scale parameter
457 458
The model is estimated by STATA/IC 13.0 with intercity mode choice as the dependent
459
variable. The nested logit structure was chosen to separate the automobile mode with common-
460
carrier modes. The model is estimated by using full information maximum-likelihood estimation
461
technique. Based on the possible explanatory attributes collected from the stated preference
Wong & Habib 2014 21
462
survey, the common-carrier modes were grouped by the relative location of the primary
463
departure stations in the GTA. This nesting structure was based on the assumption that travel
464
time is a generic utility while the choice of travel modes is alternative specific and based on the
465
departure station location. In this paper, bus was placed in its own group (Toronto Coach
466
Terminal), rail and high speed rail were placed together (Union Station), and airplane was placed
467
in its own group (Pearson International Airport). Figure 3 presents the diagram of the double
468
nested structure used in this paper and. The NL model is structured to account for the potential
469
effect of local accessibility on intercity mode choice. These nests are organized to separate the
470
automobile mode away from the transit modes.
Choice
Accessibility
No Access
Station Location
None
Toronto Coach Terminal
Generic & Socioeconomic
Auto
Bus
Access
Pearson International Airport
Union Station
Rail
HSR
Air
471 472 473 474
FIGURE 3: Nested Logit Model Structure
475
remaining 86 entries were used as model validation. The entries used for model estimation where
476
randomly chosen and represents all data collector sources. Table 3 is a summary of the estimated
477
model parameters.
478 479
TABLE 4: Parameter Estimates of the Empirical Model
The resulting NL model was estimated using 344 of the total 430 respondent entries. The
Log likelihood = -3557.5799 Wald chi2(44) = 297.08 Prob > chi2 = 0.000
Wong & Habib 2014 22 Description Generic intercity travel attributes Generic local travel attributes within the access nesting group
Attribute Travel Time (in hours) Travel Cost (in units of 10 dollars) Access Time Egress Time
Access Mode - Drop Off
Access Mode - Taxi Alternative specific local travel attributes within the station location nesting group Egress Mode - Local Transit
Egress Mode - Taxi
Household Income Between $30,000 and $39,999
Household-based socioeconomic attribute based on household income
Household Income Between $80,000 and $89,999
Household Income Between $90,000 and $99,999
Household Income Over $100,000
Respondent-based socioeconomic attribute based on age of respondent
Age Between 18 to 24
Age Between 25-30
Mode all
Coeff. -0.21218
P>|z| 0
all
-0.086
0
auto bus/rail/air/HSR auto bus/rail/air/HSR auto bus rail air hsr auto bus rail air hsr auto bus rail air hsr auto bus rail air hsr auto bus rail air hsr auto bus rail air hsr auto bus rail air hsr auto bus rail air hsr auto bus rail air hsr auto
fixed -0.34862 fixed -0.38708 fixed -0.75807 0.298099 -0.02588 0.298099 fixed -0.50081 -0.0333 0.237111 -0.0333 fixed 0.431189 0.151976 0.645616 0.151976 fixed -0.90458 -0.18468 0.116984 -0.18468 fixed -0.06576 -0.12731 -0.08913 -0.22086 fixed -0.38465 -0.32831 -0.76153 -0.38296 fixed -0.0504 -0.29872 -0.17411 -0.11473 fixed -0.24664 -0.30767 -0.01822 -0.07006 fixed 0.729444 0.127588 -0.14689 0.111051 fixed
fixed 0 fixed 0 fixed 0 0 0.874 0 fixed 0.026 0.774 0.059 0.774 fixed 0.011 0.101 0 0.101 fixed 0 0.019 0.538 0.019 fixed 0.778 0.571 0.723 0.319 fixed 0.059 0.079 0.002 0.037 fixed 0.804 0.142 0.427 0.538 fixed 0.068 0.017 0.894 0.548 fixed 0 0.19 0.303 0.244 fixed
Wong & Habib 2014 23
Age Between 31-40
Dissimilarity Parameter for Access Nest Dissimilarity Parameter for Station Location Nest
No Access Nest Access Nest None Nest Toronto Coach Terminal Nest Union Station Nest Pearson International Airport Nest
bus rail air hsr auto bus rail air hsr auto bus/rail/air/HSR auto bus rail/HSR
0.599358 0.25969 0.322517 0.469648 fixed 0.594222 0.756194 0.531022 0.738164 1 0.717817 1 1 0.520145
air
1
0 0.101 0.07 0.001 fixed 0.002 0 0.003 0 0
0
LR test for IIA chi2(2) = 26.72 Prob > chi2 = 0.000
480
In the NL model, the primary explanatory variables (intercity travel time, intercity travel
481
cost, access time, and egress time) were all statistically significant with the expected signs. The
482
variable specifications in the model were developed based in understandings from similar studies
483
in similar contexts. For example, investigations of Ridout and Miller (1982) provided guidance
484
for specifying transit service related attributes, e.g. access, egress etc. In the final model the sign
485
and magnitude of intercity travel time and cost coefficients are also found consistent with Bhat’s
486
(1995) HEV and MMNL (1998) models of travel demand within the same corridor. This proves
487
the justification of nested logit model as opposed to multinomial logit or mixed logit model.
488
Multinomial logit is clearly not suitable as some alternative modes are found to have correlated
489
random utility components and hence the nested logit model is estimated. Similarly, unlike other
490
studies, a mixed logit model was not needed in our context. Possible explanation is that the
491
nested logit formulation is sufficient enough to capture correlated random utilities of and data
492
evidence does not support any further significant overlap between the alternatives as well as any
493
random variations of attribute effects across the population.
Wong & Habib 2014 24
494
The magnitude of local access and egress time coefficients were higher than that of
495
intercity travel time, indicating the importance of local accessibility and transit station locations
496
on intercity travel demand. While these results are inconsistent with prior studies, the local travel
497
time information was shown to the respondent and not implied post-data collection and thus
498
should be more accurate than existing datasets used in prior studies.
499
Aside from local travel time, local access and egress mode parameters were also
500
statistically significant and had intuitive signs. The NL model revealed a preference for being
501
dropped off at Union Station, which is considered a disutility when dropped off at Toronto
502
Coach Terminal or Pearson International Airport. There is also a preference for taking the taxi to
503
the airport over the bus or rail stations, which may be due to the higher travel cost associated
504
with air travel compared to the other intercity travel alternatives. When arriving at Montreal,
505
respondents prefer local transit for an egress mode for all alternatives. Only air travelers consider
506
taxi as a beneficial egress mode choice, which may be correlated with intercity travel costs. The
507
magnitude of these local travel mode coefficients are relatively high when compared that of
508
intercity travel time and cost, which another indication to the importance of local accessibility on
509
intercity travel demand.
510
Income and age socioeconomic attributes were found as statistically significant attributes
511
to this intercity mode choice model. Household income did not appear to have a direct
512
relationship on intercity mode choice as individuals from different household incomes still
513
preferred automobile over transit-based intercity travel alternatives. On the contrary, an increase
514
in age increases the utility of rail, air, and high speed rail travel while decreasing the utility of
515
bus travel. These findings imply that the age has a predictable influence on intercity mode choice
516
but may also be affected by the non-linear effect of household income.
Wong & Habib 2014 25
517
It is apparent that local accessibility is a significant contributing factor to intercity mode
518
choice. The effect of local accessibility is most apparent when long local access or egress times
519
are required for a given intercity travel mode. The use of a SP survey design is instrumental to
520
understanding how changes in local accessibility level-of-service influence intercity modal
521
choice. While socioeconomics does factor into the NL model, it is a smaller degree compared to
522
local accessibility attributes.
523
Waiting time is an important factors of transit related modes and air transportation. In
524
case of transit modes for inter-city travel time, it is expected that passengers know about the
525
schedules and arrive at the stations accordingly. So, waiting time becomes irrelevant in such
526
context. This is the reason, waiting time for transit modes become insignificant and are left out of
527
the model. In case of air transportation, waiting time (boarding time) is also systematic and more
528
or less same for any types of trips. In case of such a systematic factor, the effects are constant
529
and hence difficult to capture in the model. However, we believe that waiting time have effects
530
and even though are difficult to capture in the model those are indirectly captured in the mode-
531
specific socio-economic variable coefficients.
532 533
Model Validation
534
To validate the NL model, the model coefficients were applied to the 86 entries not used for
535
estimation. Comparing between the frequency of stated mode choice and predicted mode choice
536
of both sets of respondents, the resulting modal shares for all five alternatives did not vary
537
beyond 8%. This validation effort is a confirmation that the NL model is an accurate predictor of
538
intercity mode choice. Figure 4 below is a graphical representation of stated and predicted modal
539
shares.
Wong & Habib 2014 26
540 100% 80% 60% 40% 20%
19%
19%
18%
13%
14%
14%
HSR
15%
18%
17%
Air
22%
22%
28%
29%
Model Estimated
Verified Estimated
17% 36%
Rail Bus Auto
0%
541 542 543 544
Stated by Respondent
FIGURE 4: Stated Modal Share and Verification of Nested Logit Model Value of Travel Time Savings
545
The coefficient estimates from nested logit can be used in other forms of econometric analysis.
546
One such form is the calculation of value of time travel savings (VTTS), which is a cost-benefit
547
analysis that assesses the trade-offs between travel cost and travel time. The concept of VTTS
548
arises from the idea that the necessity of travel derives from demand for activities and the idea
549
that travel time has a negative demand. Depending on the importance of the activity, individuals
550
place a certain importance to reduce the travel time required to reach the activity and may be
551
willing to pay a higher travel cost [15]. Some reasons why individuals may want to reduce his or
552
her travel time are to use the time saved to yield a monetary benefit, spend the saved time in
553
recreation or other activities, and to reduce any undesired attribute of travel such as discomfort or
554
fatigue. VTTS is calculated by taking the ratio of the coefficient of travel time to coefficient of
555
travel cost. Alternatively, the access and egress VTTS values measure the additional intercity
556
travel cost that an individual would pay for a departure station to be located closer to his or her
557
origin location. VTTS values obtained from the model estimation coefficients are listed below.
558
Main intercity trip - $24.67/hour
559
Local access - $40.54/hour
Wong & Habib 2014 27
560
561
The VTTS for the main intercity trip is similar to the values of existing literature (EcoTrain
562
2011). For access and egress VTTS, the respective local travel times were divided by the
563
intercity travel cost rather than the cost of local travel. The difference between the intercity
564
VTTS and access/egress VTTS values may be attributed by the relatively lower local travel
565
times. While there are inherent weaknesses to this approach, as the intercity travel cost is
566
independent of local travel time, the survey design would have been too complex to include the
567
extra dimension of local travel cost.
Local egress - $45.01/hour
568 569
CONCLUSIONS
570
There were two primary contributions of this research: assessing the effectiveness of collecting
571
responses using web-based methods (including social network based sampling) and evaluating
572
effects of transit station accessibility on intercity mode choice.
573
With the design of a web-based survey, the procedure to sample from multiple online
574
social networks was used. It became clear that the use of a single individual’s immediate social
575
network was not enough to capture a representative sample of the target population. The process
576
of sampling from multiple online social networks expanded the overall socioeconomic and
577
geographic profiles of respondents at the cost of low response rates when sampling from larger
578
social networks. Another alternative was to utilize the immediate social network of a number of
579
individuals, which yielded improved response rates. The main limitation encountered during
580
data collection was the low ratio between completed surveys and potential audience. The online
581
social networks that have the largest potential audience were typically the ones where one
582
individual has the least impact and audience. This relationship was expected and was
Wong & Habib 2014 28
583
supplemented by utilizing an online survey panel for additional responses. Overall, online social
584
networks were observed to be a viable survey recruitment source when appropriate sampling
585
techniques and communication tools are used to generate interest.
586
The survey used in this paper compiled trip information and travel preference datasets
587
with greater geographical disaggregation compared to currently available public data. From the
588
nested logit model, it is evident that individuals have greater sensitivity to local access and egress
589
travel time compared to intercity travel time on a per unit time level. Additionally, the NL model
590
also found a significant relationship between access/egress modes and intercity stations. These
591
findings validate initial prediction that local accessibility has a relatively large influence on
592
intercity modal choice. Policy changes related to of local accessibility in intercity travel plans
593
may include; the effect of station locations on travel demand, partnerships with local transit or
594
transportation services, and induced intercity travel due to improved local accessibility.
595
Additionally, the lack of local accessibility attributes for the automobile mode may be one of the
596
primary reasons why automobile has a higher modal share compared to existing intercity travel
597
modes. From this model, the introduction of a HSR line in the QWC may result in a large modal
598
shift towards HSR; however, long local access or egress times or a lack of local travel integration
599
may greatly reduce the appeal of HSR.
600
The paper presents an investigation and empirical result on inter-city travel mode choices
601
while collecting data through an innovative and new approach of data collection. Specially, the
602
potential of using social media for travel survey sampling frame is tested. However, this also
603
results in some limitations in the empirical study. Online social network based sampling methods
604
resulted in a skewed sample set towards the socioeconomic characteristics of the individual
605
distributing the survey. The sample collected using social networks did have a bias towards the
Wong & Habib 2014 29
606
university population; however additional responses were collected via an online survey panel,
607
which has more socioeconomically representative respondents compared with census data.
608
Overall, the university population bias is somewhat subsided. Such, bias may not be as evident in
609
other forms of data collection like call centers, street intercept, or mail outs; however the efforts
610
of data collection was greatly reduced. In addition the distribution of revealed travel modes were
611
similar between sampling from online social networks versus a more demographically
612
representative survey panel for this project. Further research should be conducted to assess
613
whether there are any links between internet use and travel preferences. However, it becomes
614
clear from this study that there is a potential that social media can be used as a supplementary
615
survey collection method to reduce the necessity to rely on more traditional sampling methods.
616 617
References Ashiabor, S., H. Baik and A. Trani (2007) "Logit Models for Forecasting Nationwide Intercity Travel Demand in the United States," Transportation Research Record: Journal of the Transportation Research Board, pp. 1-12. Ben-Akiva, M. and S. Lerman (1985) "Discrete Choice Analysis," The MIT Press, Cambridge Massachusetts. Bhat , C. R. (1995) "A Heteroscedastic Extreme Value Model of Intercity Travel Mode Choice," Transportation Research B, vol. 29, no. 6, pp. 417-483. Bhat, C.R. (1998) "Accommodating Variations in Responsiveness to Level-of-Service Measures in Travel Mode Choice Modeling," Transportation Research A, vol. 32, no. 7, pp. 495-507. Burge, P. , C. W. Kim and C. Rohr (2011) "Modelling Demand for Long-Distance
Wong & Habib 2014 30
Travel in Great Britain," Rand Corporation, Santa Monica, CA. EcoTrain (2011) "Updated Feasibility Study of a High Speed Rail Service in the Quebec City - Windsor Corridor," Transport Canada, Montreal. Grisolía, J.M. and J.D. Ortúzar (2010) “Forecasting versus Observed Outturn: Studying Choice in Faster Inter-Island Connection”. Transportation Research Part A 44: 159-168 Ahsan, H.M. , M.M. Rahman and K.M.N. Habib (2003) “Socio-economic conditions and travel behaviour of inter-city bus passengers: Bangladesh perspective”. Journal of Civil Engineering, Institution of Engineers of Bangladesh 30(2): 91100 Rojo, M. , L. dell’Olio, J.L. Moura and H. Gonzalo (2008). User satisfaction in Interurban Bus Transit: Modelling Relevant Variables and their Influence. In: Proceedings of the 10th International Conference on Applications of Advanced Technologies in Transportation, Athens, Greece. Rojo, M. , H. Gonzalo, L. dell’Olio, A. Ibeas (2011). “Modelling Gender Perception of Quality in Interurban Bus Services”. Proceedings of the Institution of Civil Engineers, Transport 164 (1), 43–53. Rojo, M., H. Gonzalo-Orden, L. dell’Olio and A. Ibeas (2012). “Relationship between Service Quality and Demand for Inter-Urban Busses”. Transportation Research Part A 46: 1716-1729. Debrezion, G., E. Pels and P. Rietvelt (2009) “Modelling the joint Access Mode and Rail Station Choice”. Transportation Research Part E 45: 270-283 Hsio, C-Y. and M. Hansen (2011) “A Passenger Demand Model for Air Transportation
Wong & Habib 2014 31
in a Hub-and-Spoke Network”. Transportation Research Part E 47: 111211125. Mandel, B., M. Guadry and W. Rothengatter (1997) “A Disaggregate Box-Cox Logit Mode Choice Model of Intercity Passenger Travel in Germany and its Implications for High-Speed Rail Demand Forecasts”. Annals of Regional Science 31, 99-120. Hensher, D.A. , J. Louviere and J. Swait (1999) “Combining Sources of Preference Data.” Journal of Econometrics 89,197–221 IBI Group (2002) "Making Transportation Sustainable: A Case Study of the Quebec City-Windsor Corridor," Environment Canada, Ottawa. Langan,
P.
(2011)
"All
Canadian
HSR
Studies,"
[Online].
Available:
http://www.highspeedrailcanada.com/p/all-canadian-hsr-studies.html. [Accessed 11 04 2013]. Miller, E.J. (2004) "The Trouble with Intercity Travel Demand Models," Journal of the Transportation Research Board, pp. 91-101 Ridout, R. and E. J. Miller (1982) "A disaggregate logit model of intercity common carrier passenger modal choice," in Proceedings of the Roads and Transportation Association of Canada Annual Conference, Halifax Sonesson, T (2011) "Inter-Urban Travel Demand Elasticities with Emphasis on Trip Generation and Destination Substitution," Journal of Transport Economics and Policy, vol. 35, pp. 301-326 Statistics Canada (2010) "Travel Survey of Residents of Canada 2006 to 2010," Ottawa
Wong & Habib 2014 32
Statistics Canada (2013) "Census Profile - Age, Sex, Marital Status, Families Households, Dwellings and Language for Canada and Forward Sortation Areas, 2011 Census," Ottawa Data Management Group. (2013). Transportation Tomorrow Survey 2011 – Version 1.0 Data Expansion & Validation. Toronto: Transportation Information Steering Committee. Statistics Canada. (2014, 07 23). Median total income, by family type, by province and territory (All census families). Retrieved 20 03, 2015 from Statistics Canada: http://www.statcan.gc.ca/tables-tableaux/sum-som/l01/cst01/famil108aeng.htm Wen, C-H. and F.S. Koppelman (2001) “The Generalized Nested Logit Model”. Transportation Research B 35, 627–641. Wilson, F. R., S. Damodaran and D. J. Innes (1990) "Disaggregate mode choice models for intercity passenger travel in Canada," Canadian Journal of Civil Engineering, vol. 17, pp. 184-191.
618
