See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/284452240

An Application of RP-SP Data for Joint Estimation of Mode Choice Models

Conference Paper · January 2016

CITATIONS READS
0 1,030

3 authors:

Mohamed Salah Mahmoud Khandker Nurul Habib

University of Toronto University of Toronto
21 PUBLICATIONS 581 CITATIONS 331 PUBLICATIONS 3,889 CITATIONS

SEE PROFILE SEE PROFILE

Amer Said Shalaby

University of Toronto
184 PUBLICATIONS 3,839 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

A Study of Car and House Ownership in the face of Increasing Commuting Expenses (CHOICE) View project

Passenger Rail and Urban Transit Reliability View project

All content following this page was uploaded by Khandker Nurul Habib on 19 February 2020.

The user has requested enhancement of the downloaded file.

 1 Application of RP/SP Data to the Joint Estimation of Mode Choice Models:
2 Lessons Learned from an Empirical Investigation into Cross-Regional
3 Commuting Trips in the Greater Toronto and Hamilton Area
4
5
6
7
8
9 Mohamed Salah Mahmoud*, Ph.D.
10 Department of Civil Engineering
11 University of Toronto
12 35 St. George Street, Toronto, Ontario, M5S 1A4, Canada
13 Tel: 416-946-8299; Email: mohamed.mahmoud@utoronto.ca
14
15
16
17 Khandker Nurul Habib, Ph.D., P.Eng.
18 Assistant Professor
19 Department of Civil Engineering
20 University of Toronto
21 35 St. George Street, Toronto, Ontario, M5S 1A4, Canada
22 Tel: 416-946-8027; Email: khandker.nurulhabib@utoronto.ca
23
24
25
26 Amer Shalaby, Ph.D., P.Eng.
27 Professor
28 Department of Civil Engineering
29 University of Toronto
30 35 St. George Street, Toronto, Ontario, M5S 1A4, Canada
31 Tel: 416-978-5907; Email: amer@ecf.utoronto.ca
32
33

34
35

36

37
38
39 Submission Date: August 1, 2015
40 Word Count: Abstract (193) +Paper (5,199) + (6) figures/tables (1,500) + References (593) = 7,485
41 * Corresponding Author
 1 ABSTRACT
2 This study presents an investigation on the mode choice behaviour of cross-regional commuters
3 in the Greater Toronto and Hamilton Area (GTHA). Cross-regional trips are defined as those
4 crossing the boundaries of two or more municipal or regional jurisdictions, each served by a
5 local transit operator. The GTHA has nine local transit and one regional transit systems with
6 little or no coordination between them. The complexity of cross-regional trips stems from the
7 multimodal nature of long distance travel across multiple regional/local municipalities. Using
8 Revealed Preference (RP) and Stated Preference (SP) data of the Survey of Cross-Regional
9 Intermodal Passenger Travel (SCRIPT), a set of econometric joint RP/SP mode choice models
10 are developed. This paper presents a comparison between a conventional Multinomial Logit
11 (MNL) mode choice model and two models that relax the independent and irrelevant alternative
12 (IIA) assumption, namely Nested Logit (NL) and Parameterized Logit Captivity (PLC) models.
13 The joint RP/SP models reveal meaningful insights into cross-regional commuters’ mode choice
14 behaviour. The developed models now constitute the core of a policy analysis tool, called
15 Interactive Model for Policy Analysis of Cross-Regional Travel (IMPACT), which predicts
16 changes in aggregate modal shares in response to new policies.
 1 INTRODUCTION
2 We live in the era of “mobility,” an era in which individuals have become more mobile in terms
3 of their ability to plan and undertake trips using various travel modes within multimodal
4 transportation systems. Such systems offer multiple travel options of various modes (walking,
5 cycling, automobile, public transit, etc.) and connections among them. The development of
6 effective multimodal transportation systems cannot be achieved without proper planning of the
7 levels of integration between the various travel modes; transit modal integration is the
8 cornerstone of this development (1). Transit modal integration refers to schemes that facilitate
9 combinations of transit services and other motorized or non-motorized modes such that
10 passengers can seamlessly interchange between different modes of travel. As such, more
11 opportunities are created from the efficient integration of transit services with other modes by
12 increasing and diversifying users’ travel options. As a result, improving transit modal integration
13 has been considered by many regional transit authorities as a promising strategy to boost
14 effectiveness of transit systems (1, 2).
15 Due to the unabated sprawl of metropolitan regions, the number of cross-regional commuter trips
16 (i.e. trips that cross boundaries of two or more municipal or regional jurisdictions, each served by
17 a local transit operator) has been growing continuously (3, 4). Consequently, cross-regional
18 commuters (especially those who use transit modes) have been experiencing longer than typical
19 commuting travel times. The peak-period commuting trips in many North American regions are
20 dominated by private automobile drivers. Despite the availability of alternative and more
21 sustainable travel modes (such as transit, carpooling, car-sharing, walking and biking), driving
22 alone is often more attractive since it provides shorter travel times and ubiquitous accessibility
23 while still being financially competitive.
24 The phenomenon of private automobile domination becomes more prominent in the case of
25 cross-regional commuting trips. Unlike intra-regional commuting travel (trips originating from
26 and destined to the same municipal jurisdiction), cross-regional commuters have a unique set of
27 possible travel modes since non-motorized modes are mostly infeasible. In addition, cross-
28 regional trips often originate in suburban areas where transit accessibility is relatively inadequate
29 because of low population density and sparse land use. That is, non-automobile cross-regional
30 trips may involve the use of multiple transit services or the interaction between two travel modes,
31 which often results in delays caused by the typical lack of service coordination, not to mention
32 the absence of fare integration. Therefore, in such cases, transit options do not provide
33 competitive travel times and/or costs compared to automobile options, which explains the latter’s
34 dominance. This highlights the importance of dedicating special attention to the enhancement of
35 multimodal transportation systems’ current planning practices by improving transit modal
36 integration. Evaluating the effectiveness of such initiatives and supporting policies requires a
37 proper understanding of individuals’ intermodal travel behaviour. This paper is focused on
38 acquiring a thorough understanding of the mode choice behaviour of cross-regional travellers,
39 many of whom have viable intermodal options, by quantifying the effects of the key factors that
40 influence their travel decisions. That is, the primary objective of this study is to develop a policy-
41 sensitive mode choice model that constitute the core of a policy analysis tool to facilitate testing

1
 1 the effect of new policies on the aggregate mode share of cross-regional commuting trips. To
2 fulfill this objective, different decision structures of mode choice models are empirically
3 investigated and a preferred decision structure is recommended to be used for policy analysis.
4 LITERATURE REVIEW
5 In order to develop policy-sensitive travel demand models capable of capturing the details of
6 cross-regional trip legs such as access, transfer, and egress times, exhaustive data on the trip
7 patterns of cross-regional travellers’ behaviour are required. Typical commuting travel surveys
8 do not provide sufficient data to conduct this type of analysis. This is due to several reasons:
9 cross-regional trips are often underrepresented in survey samples, the collected data do not
10 provide the necessary level of detail on inter- and intra-modal trips, and the majority of typical
11 travel surveys rely predominantly on Revealed (observed) Preference (RP) trip data. RP data are
12 limited to the current conditions that the individuals may encounter in existing travel scenarios.
13 In order to investigate individuals’ preferences towards services or policies that do not exist at
14 present, Stated Preference (SP) surveys and data are more appropriate (5-7). However, previous
15 empirical analyses have shown that SP data suffer from systematic bias due to the specific nature
16 of the SP experimental design as well as the inconsistency of individuals’ stated preferences with
17 their actual choices (8). Alternatively, using joint RP/SP data allows for scale adjustment of
18 parameter estimates to correct the systematic bias of the SP data (9, 10). That is, data from
19 RP/SP surveys provide comprehensive datasets that enable the development of joint RP/SP
20 models that can explain the probabilistic response to changes in transportation level-of-service
21 attributes as a result of introducing new policies. In a recent study, data from an RP/SP
22 commuting survey were used to investigate the influence of transit service attributes on mode-
23 switching behaviour (11). The study showed that joint RP/SP models outperformed (in terms of
24 the goodness of fit and explanatory power) corresponding models developed using SP data only.
25 Several studies have used pooled RP/SP data to develop discrete choice models (12). To estimate
26 RP/SP models jointly, each data record should include RP information and all repeated SP
27 choice scenario information for the same individual across the survey population. As such, the
28 estimation routine of these models inherently assumes that the data come from two different
29 populations or datasets. Brownstone, Bunch and Train (13) developed joint RP/SP MNL and
30 simple mixed logit models for vehicle type choice. The study emphasizes the importance of
31 combining RP/SP data to exploit the strengths and mitigate the weaknesses of each type. In
32 addition, the model estimation results showed that advanced model structures are statistically
33 better than the standard logit model. Similarly, Hensher, Rose and Greene (9) used pooled RP/SP
34 data to develop nested and mixed logit mode choice models.
35 The development of joint RP/SP models may involve the estimation of coefficients that are
36 uniquely determined by the RP or the SP data, coefficients that are equally identified within the
37 two datasets (before applying the scale effect), and coefficients that are identified differently
38 within the two datasets. Despite the recent advances in joint estimation of RP/SP models, there
39 are no clear guidelines in the literature for classifying data-specific or pooled coefficients
40 especially for advanced discrete choice models (13). In a few cases where all attributes are
41 considered generic (pooled) across both datasets, using a joint RP/SP model for prediction can be

2
 1 less problematic (14). However, in most cases, data-specific coefficients must be estimated. For
2 instance, SP attributes are often not available in the RP context and therefore cannot be estimated
3 using the RP data. Similarly, different coefficients of the same attribute can be estimated across
4 the two datasets, which can be tested empirically. Cherchi and Ortúzar (14) presented a
5 discussion of the development and application of joint RP/SP models when the RP and SP data
6 show different systematic or random taste heterogeneity. Despite the wide range of applications
7 of joint RP/SP models, most of the attention has been given to advances in model estimation
8 without giving proper attention to the use of such models for forecasting and policy analysis
9 which limits the practical aspects of jointly estimated RP/SP models.
10 This paper presents an empirical investigation into the development of joint RP/SP mode choice
11 models and provides practical implications on the use of such models for policy analysis. The
12 developed models reveal meaningful insights into cross-regional commuters’ mode choice
13 behaviour. In addition, the models constitute the core of a policy analysis tool which predicts
14 changes in aggregate modal shares in response to new policies. That is, this paper investigates
15 mode choice and/or captivity behavioural patterns in the context of cross regional trips and
16 addresses some of the limitations of applying advanced joint RP/SP models in practice.
17 STUDY AREA AND DATA
18 The Greater Toronto and Hamilton Area (GTHA), as shown in Figure 1, forms Canada’s largest
19 urban region. The GTHA has nine local transit (LT) and one regional transit (RT) operator, each
20 having its own transit network with little or no coordination between them. The complexity of
21 cross-regional trips stems from the multimodal nature of long distance travel across multiple
22 regional/local municipalities. Such trips may involve the use of multiple transit services or the
23 interaction between two travel modes. In 2011, approximately one in three trips in the GTHA
24 crossed a regional boundary (4). Data from a recent survey on cross-regional trips in the region
25 showed that more than 70% of automobile users and 95% of transit users have commuting times
26 longer than 30 minutes. In addition, the average travel time of cross-regional commuting trips by
27 local transit modes (with walk access) is 75 minutes. Further, the trip patterns and demographics
28 of cross-regional commuters in the GTHA differ considerably from regular commuters (15).
29 In this paper, RP and SP data of the Survey of Cross-Regional Intermodal Passenger Travel
30 (SCRIPT) are utilized to develop a set of econometric mode choice models to improve our
31 understanding of cross-regional commuters’ travel behaviour. Details of the survey design and
32 data collection procedure can be found in (15). SCRIPT provides detailed information about
33 respondents’ current travel options and their stated preferences concerning different travel
34 alternatives in response to changes in the current level-of-service attributes.
35 The collected RP, SP, and household and personal information was used to construct a complete
36 database for empirical model development. The joint RP/SP dataset encompasses seven data
37 points for each individual: one RP and six SP records. Level-of-service attributes of the
38 considered travel modes were generated using a multimodal trip planner tool (15). The tool
39 provides detailed disaggregate travel time components and travel costs (including parking costs
40 at work locations and park-and-ride stations) for each trip record. The RP data are joined with

3
 1 the prevailing level-of-service attributes. However, the SP data are joined with the modified
2 level-of-service attributes based on changes in attributes’ levels according to the experimental
3 design of the survey. That is, respondents’ SP mode choices are associated with the modified
4 level-of-service attributes shown to them during the SP experiment.
5

6
7 Figure 1 The Greater Toronto and Hamilton Area1
8 A total of 704 trip records are used in the analysis. The variables used in empirical investigation
9 are presented in Table 1 along with the corresponding polices tied to each variable (if any).
10 These variables are carefully selected in accordance with the policies investigated in the SP
11 experiment. The modelling framework considers nine alternatives; it explicitly distinguishes
12 between auto driver and auto passenger modes, regional and local transit modes, and different
13 access modes.

1
http://findtheway.ca/en/

4
 1 Table 1 Definitions of Variables
Variable Name Description Corresponding Policy
Travel Cost/Fare Travel cost including gas and parking cost at Increase driving cost/reduce
the work location or transit fares transit fares
P&R Cost at LT Parking cost at local transit park-and-ride Increase parking cost at LT
Stations locations per day park-and-ride stations
P&R Cost at RT Parking cost at regional transit park-and-ride Introduce paid parking at RT
Stations locations per day park-and-ride stations
In-vehicle Travel In-vehicle travel time if Wi-Fi is not available Reduce transit travel time
Time (no Wi-Fi) on regional transit trains
In-vehicle Travel In-vehicle travel time if Wi-Fi is available on Provide new transit features
Time (Wi-Fi) regional transit trains (Wi-Fi on board)
Out-of-vehicle Travel Sum of access, waiting, transfer, and egress Improve transit accessibility
Time travel times
Next LT Vehicle 1 if the next transit vehicles’ arrival time Provide accurate transit
Information information is available; 0 otherwise information
Provision
Need for a 2nd 1 if the individual needs to make more than Reduce number of
Transfer one transfer between regional transit and other intermodal transfers
travel modes; 0 otherwise
Number of Vehicles Number of vehicles per household N/A
per Household
Number of persons Number of persons per household N/A
per Household
Transit Pass 1 if the individual owns a transit pass; 0 N/A
Possession otherwise
Trip O/D: City of 1 if the trip origin/destination is from/to the N/A
Toronto City of Toronto; 0 otherwise
Gender (Male) 1 if male; 0 otherwise N/A
2

3 DECISION STRUCTURES
4 The modelling frameworks can be classified into two groups: conventional mode choice models,
5 which suffer from the independence and irrelevant alternatives (IIA) property, and advanced
6 modelling frameworks, which relax this assumption. Conventional mode choice models include
7 the multinomial logit (MNL) model, which is the most common discrete choice model
8 formulation. To relax the IIA property, two advanced choice model structures are adopted: the
9 nested logit (NL) and the parameterized logit captivity (PLC) models. Details on the econometric
10 model formulations and probability calculations can be found in (16-20).
11 Figure 2 shows the joint RP/SP decision structures of the three models. An artificial tree
12 structure is assumed to identify the differences between the two datasets: RP and SP data. This is
13 captured through a scale parameter that is estimated for the SP data relative to a normalized
14 (fixed to 1) scale parameter for the RP data. In addition, the scale parameter allows for capturing
15 the heteroscedasticity in individuals’ responses. As such, the scale parameter was parameterized
16 as an exponential function of the respondents’ attributes that can explain scale variation (e.g.

5
 1 socio-demographic or land use attributes) (21). The repeated RP/SP observations are treated as
2 panel data to allow for the correction of the group behaviour (i.e. each observation is represented
3 by each individual’s actual choice of the RP context and six choices of the SP hypothetical
4 scenarios) (14). For a sample of Z individuals with multiple responses, the joint likelihood
5 function can be expressed as follows:
! ! !

� � = (Pr(�)!" )!!"# , [1]

! ! !

6 where � is a vector of variable coefficients that maximizes the likelihood function � � , d

7 represents either the RP data or one of the six choice situations of the SP data, and yndz =1 if
8 person z chooses alternative n from a variable choice set C in dataset d and zero otherwise. As
9 shown in equation [1], the joint likelihood function is defined in a closed form which facilitates
10 the use of the developed models for policy analysis. The scale parameter was parameterized
11 using individual-specific attributes in order to account for the correlation across repeated
12 observations of the same individuals. In addition, it is worth mentioning that SP choice scenarios
13 where low confidence levels were reported by the respondents are not considered in the
14 estimation process.
15 Unlike the MNL model structure, where equal competition between all pairs of alternatives is
16 assumed, the NL model structure categorizes common alternatives in “nests” and as such
17 considers a partially common error term component for within-nest alternatives. It is important to
18 note that the choice probabilities do not imply that individuals make their decisions based on the
19 assumed nesting order. The nest-specific (upper level) scale parameter µj is identified to capture
20 the difference in within-nest variation across the nests. The value of the nest-specific scale
21 parameter µj is constrained between the dataset-specific scale parameter µ and positive infinity.
22 This indicates a non-zero correlation between within-nest alternatives. A higher value of µj
23 relative to µ indicates lower variation across nest alternatives (i.e. higher substitution between
24 within-nest alternatives). Such constraints are adopted in the models’ estimation routines to test
25 different nesting structures empirically, and thereby, they are either accepted or rejected.

6
 1
2 Figure 2 The Artificial Nested Structure of Joint RP/SP Models
3 On the other hand, the parameterized logit captivity (PLC) model structure is a modified version
4 of the dogit model structure originally proposed by Gaundry and Dagenais (16). As a latent class
5 model, according to an inclusion probability, individuals are either classified as rational
6 (strategic) or captive (myopic) users. Rational users are assumed to make a strategic choice
7 according to the level-of-service attributes, while in contrast, captive users are assumed to rely
8 on a specific alternative under all circumstances, irrespective of the level-of-service attributes of
9 other options. It is important to note that the use of “captive” or “myopic” terms in this context
10 does not imply that individuals are fundamentally irrational or that they explicitly do not have

7
 1 other options in their choice sets. These individuals are best described as “captive-by-choice”
2 users. Such individuals may have no desire to consider some of the alternatives that may be
3 readily available to them and therefore are implicitly dependent on other modes of travel. This
4 implicit dependency may be a result of individuals’ preconceptions of specific alternatives or
5 their travel inertia towards these modes. For instance, perceived higher mobility with a car or
6 viewing bicycling to work as an efficient way to exercise may implicitly cause a commuter to be
7 attracted to one mode over another. Therefore, even if a mode may become more attractive from
8 a level-of-service perspective, the decision maker might not consider the strategic trade-off
9 between modes (i.e. based on their level-of-service attributes) due to unobserved reasons. In
10 other words, these individuals are not necessarily “irrational” users; they may be making a
11 rational choice according to factors unobserved to the modeller. However, it is difficult to
12 capture directly such implicit decisions regarding which alternatives to be considered within
13 individuals’ choice sets. Therefore, this perception of modal dependency is treated as a latent
14 choice. The PLC model structure allows to capture, in part, the unobserved mode dependency by
15 parameterizing two utility functions: the systematic utility functions and the parameterized
16 captivity function. That is, typical level-of-service attributes (travel times and travel costs/fares)
17 are included in the systematic utility functions of the model, which represents the
18 rational/strategic choice component. Such attributes contribute to the trade-offs between
19 alternatives that are made by rational individuals prior to making a decision. In contrast, socio-
20 demographic and mode-specific attributes are added to the parameterized captivity functions.
21 These factors are more likely to influence individuals’ dependency on one specific alternative.
22 EMPIRICAL MODELS
23 Various model specifications were tested and compared to one another until the reported final
24 model specifications with the highest explanatory power were reached. Table 2 shows the
25 estimation results of the joint RP/SP MNL, NL, and PLC models. In general, all the parameters
26 are estimated with the expected signs and relative values, and they are statistically significant at
27 the 95% confidence interval except for the provision of information on local transit vehicles’
28 arrival times, which is statistically significant at the 90% confidence interval. The reported rho-
29 squared value of the MNL model is 0.158, which is better than the reported value for a
30 corresponding SP-only model (22). Clearly, the use of the combined RP/SP data enhanced the
31 goodness of fit and explanatory power of the joint model. Similarly, the reported rho-squared
32 values of the NL and PLC models are 0.161 and 0.160, respectively. The log likelihood ratio
33 tests of the three models showed that the reported models fit the data significantly better than the
34 corresponding constant-only models.
35 The estimated parameters of the joint RP/SP models are classified into three groups: parameters
36 exclusively estimated by either the RP or the SP datasets, parameters estimated with different
37 coefficients in each dataset, and parameters estimated with the same coefficient (before taking
38 the scale parameter effect into consideration) in the RP and SP datasets. Typically, dataset-
39 specific (RP/SP) alternative specific constants (ASC) coefficients are estimated, while variables
40 that belong to one dataset and scale parameter factors are uniquely estimated by one dataset
41 (usually the SP data) (13). Other level-of-service attributes that appear in both datasets as well as
42 socioeconomic attributes are estimated with the same coefficients. Since the specification of the
43 utility functions of the choice alternatives may vary across the RP and SP datasets, the associated

8
 1 data-specific ASC may have different values or signs. In fact, the ASC captures the average
2 impact of the unobserved factors in the model specification on the utility function of each
3 alternative. Therefore, the relative values and/or signs of one data-specific ASC compared to the
4 corresponding estimates of the other dataset’s ASC do not have a particular implication on the
5 interpretation of the estimated model’s results.
6 Assuming a unit scale parameter for the RP data, the SP scale parameter is relatively estimated
7 as a parameterized exponential function of a constant and individuals’ gender. The estimated
8 parameters of the SP scale factor are statistically significant, confirming the assumed tree
9 structure of the two datasets. The SP scale parameter is estimated to be lower than the RP scale
10 parameter, which indicates that the variance within the SP data is higher than in the RP data.
11 This typical finding explains that the SP data encompass an induced variation according to the
12 introduced changes in some attributes of current transportation systems’ elements. As explained
13 earlier, the SP scale is parameterized using socio-demographic attributes to capture the
14 heteroscedasticity across the sample. For instance, the parameter estimates of the SP scale factor
15 suggest that males have a lower scale factor than females, indicating that the latter group has a
16 lower variance across their SP choices, and vice versa.
17 The effects of the introduction of paid parking at regional transit park-and-ride stations, the
18 provision of information on local transit vehicles’ arrival time (for agencies that currently do not
19 provide this service), and the introduction of Wi-Fi service on regional transit trains on the
20 probability of commuters’ mode choice are captured through the SP environment of the model.
21 The results showed that introducing pay parking at park-and-ride stations has a negative impact
22 on the probability of choosing park-and-ride as a travel mode. On the other hand, providing
23 individuals with real-time information on transit vehicles’ arrival time would increase the
24 likelihood that they will choose local transit as their travel mode. Similarly, introducing Wi-Fi
25 service for regional transit users is expected to increase the modal share of regional transit.
26 Several model specifications were tested to quantify the effect of the availability of Wi-Fi on
27 commuters’ mode choice. The results of preliminary models showed that the introduction of Wi-
28 Fi on regional transit modes is only statistically significant for individuals who spend 40 minutes
29 or more on regional transit vehicles (i.e. individuals whose in-vehicle travel time is greater than
30 or equal to the average in-vehicle travel time for regional transit users within the sample). This
31 finding triggered further investigation of in-vehicle travel time interaction with the availability of
32 Wi-Fi on regional transit trains. The final model specification shows two in-vehicle travel time
33 parameters, one if Wi-Fi is available and the other if Wi-Fi is not. The two coefficients are
34 estimated with the expected negative sign and are statistically significant. The coefficients are
35 statistically distinct from each other. The test statistics value is found to be -2.19 which passes t-
36 statistics of -1.96 at the 95% confidence interval. The estimated coefficient of in-vehicle travel
37 time if Wi-Fi is available has a smaller negative effect on the probability of choosing regional
38 transit modes than the estimated coefficient of in-vehicle travel time if Wi-Fi is not available.
39 This indicates that individuals are more likely to choose regional transit if Wi-Fi service is
40 available on regional transit trains. In contrast, the need for a second transfer between regional
41 transit and other travel modes has a significant negative effect on the probability of choosing
42 regional transit modes.

9
 1 The effects of personal and household attributes on individuals’ travel mode choices were
2 captured in different ways across the three estimated models. Such attributes are included in the
3 systematic utility functions of the MNL and NL models. The model results showed that the
4 number of vehicles per household has a positive impact on the probability of choosing car-
5 dependent modes such as auto driver and park-and-ride. Similarly, transit pass possession
6 increases the probability of the selection of transit as a travel mode. Further, individuals who
7 commute from/to the City of Toronto are more likely to use transit. This is likely driven by the
8 city’s unique multimodal transit system, high-density land use, and supportive transit policies.
9 The NL model is estimated with a lower-nest scale parameter for all transit mode alternatives.
10 The scale parameter of the transit nest is estimated to be greater than the corresponding upper-
11 level scale parameters (i.e. RP/SP data scale parameters). In other words, the variance across the
12 nested alternatives is lower than the variance across all mode alternatives, which indicates that
13 the correlation between the nested alternatives is higher than the correlation between all
14 alternatives. This is consistent with the underlying assumption of the NL model formulation;
15 therefore, the empirical model is accepted.
16 Despite the dissimilarities between the PLC and the previously presented model formulations,
17 the empirical results are consistent with the aforementioned findings of the MNL and NL
18 models. The estimated parameters of the parameterized captivity function provide further
19 insights into individuals’ modal reliance/captivity. Based on the estimated parameters of the
20 parameterized captivity function, the model suggests that approximately 75% of the individuals
21 in the sample make their choices based on a rational trade-off between all feasible mode
22 alternatives that are available to them (i.e. by comparing alternatives’ level-of-service attributes).
23 However, the remaining 25% of individuals in the sample are classified as captive users who
24 may depend on one specific mode alternative and inattentive to the rational trade-off between
25 other alternatives. Out of which, 17% are found to be car-dependent users which explains the
26 high modal share of the auto driving mode in the data. The parameter estimates of the
27 parameterized captivity function show that an increase in the number of vehicles per household
28 results in an increase in the probability of auto driver reliance. Similarly, individuals with a
29 higher number of vehicles compared to the number of persons per household are less likely to
30 consider auto passenger as their sole travel mode. In contrast, possessing a transit pass and
31 having one end of the trip in the City of Toronto increases the probability of transit dependency.

10
1 Table 2 – Estimation Results of Joint RP/SP MNL, NL, and PLC models
RP/SP – Joint Estimation MNL NL PLC
Log Likelihood of Full Model -4,497.462 -4,156.844 -4,491.302
Log Likelihood of Constant-only Model -5,341.4035 -4,954.521 -5,346.789
Rho-squared Value 0.158 0.161 0.160
Number of Observations 704
Variable Mode Parameter (t-Statistics)
Systematic Utility Function: RP Coeff. SP Coeff. RP Coeff. SP Coeff. RP Coeff. SP Coeff.
Alternative Specific Auto driver 2.6427 0.6422 4.8019 1.6428 2.9507 2.0285
Constant (11.011) (4.442) (6.544) (2.286) (4.206) (4.99)
Alternative Specific Auto passenger 3.2428 -0.4277
0 (fixed) 0 (fixed)
Constant (4.483) (-0.66)
Alternative Specific Local transit with walk access 1.0894 -1.3721 1.873 0.5527 6.4198 3.0635
Constant (2.711) (-4.141) (2.471) (0.799) (3.688) (3.46)
Alternative Specific Local transit with auto driver access (park- -0.5885 -1.0607 5.1629 2.8293
0 (fixed)
Constant and-ride) (-0.572) (-1.389) (2.673) (1.857)
Alternative Specific Local transit with auto passenger access -1.7627 -2.3267 0.6285 -1.2924 2.973 1.4805
Constant (kiss-and-ride) (-3.246) (-6.399) (0.858) (-1.993) (2.46) (1.968)
Alternative Specific Regional transit with walk access 0.723 -1.2086 2.7149 -0.178 2.2227 3.126
Constant (1.466) (-3.53) (3.529) (-0.272) (0.616) (3.026)
Alternative Specific Regional transit with auto driver access 0.8518 -0.9091 3.0243 0.1223 2.7899 3.0047
Constant (park-and-ride) (2.351) (-3.114) (4.305) (0.187) (2.178) (3.435)
Alternative Specific Regional transit with auto passenger access -0.4608 -1.4898 1.8666 -0.4047 -3.1359 1.8054
Constant (kiss-and-ride) (-1.118) (-5.209) (2.63) (-0.625) (-0.247) (2.493)
Alternative Specific Regional transit with local transit access 0.8244 -1.2861 2.7889 -0.1934 -0.1359 3.4659
Constant (1.661) (-3.469) (3.611) (-0.293) (-0.023) (3.063)
Travel Cost/Fare All mode alternatives -0.0762 -0.0802 -0.2203
(-8.291) (-8.264) (-5.536)
P&R Cost at LT Local transit with auto driver access (park- -0.4633 -0.4494 -0.6239
Stations and-ride) (-2.97) (-3.088) (-2.141)
P&R Cost at RT Regional transit with auto driver access -0.0756 -0.0715 -0.19
0 (fixed) 0 (fixed) 0 (fixed)
Stations (park-and-ride) (-2.158) (-2.23) (-2.111)
In-vehicle Travel Time All mode alternatives -0.0355 -0.0363 -0.1093
(No Wi-Fi) (-9.206) (-8.972) (-4.759)
In-vehicle Travel Time All regional transit alternatives -0.0306 -0.0314 -0.0972
0 (fixed) 0 (fixed) 0 (fixed)
(Wi-Fi) (-7.252) (-7.434) (-4.374)

11
Out-of-vehicle Travel All transit alternatives -0.0616 -0.059 -0.2032
Time (-9.785) (-8.679) (-5.14)
Next Local Transit Local transit with walk access, local transit
Vehicle Information with driving access, local transit with 0.2395 0.2072 0.3795
Provision passenger access, and regional transit with (1.793) (1.703) (1.697)
local transit access
Need for a 2nd Transfer All regional transit alternatives -0.4614 -0.478 -0.7511
(-3.26) (-3.504) (-2.019)
Number of Vehicles per Auto driver 0.1934 0.2348
0 (fixed)
Household (3.133) (3.732)
Ratio of Number of Auto passenger
-0.7046
Vehicles to Number of 0 (fixed) 0 (fixed)
(-4.11)
Persons
Transit Pass Possession All transit alternatives 2.4008 2.2591
0 (fixed)
(10.823) (10.137)
Trip O/D: City of All transit alternatives 0.6106 0.7167
0 (fixed)
Toronto (3.692) (3.824)
Exponential Function of Scale Parameter:
Constant SP scale factor -0.1042 -0.1198 0 (fixed) -0.3186
0 (fixed) 0 (fixed)
(-1.145) (-1.263) (-2.597)
Gender SP scale factor -0.1781 -0.1878 0 (fixed) -0.207
0 (fixed) 0 (fixed)
(-4.101) (-3.986) (-4.034)
Exponential Function of Nesting Scale Parameter*:
Constant All transit alternatives nest scale factor -2.0926
0 (fixed) 0 (fixed)
(-3.125)
Parameterized Captivity Function:
Alternative Specific Auto driver 0.1348 -2.2695
0 (fixed) 0 (fixed)
Constant (0.181) (-4.623)
Alternative Specific Auto passenger -1.4374 -1.7724
0 (fixed) 0 (fixed)
Constant (-1.702) (-3.246)
Alternative Specific Local transit with walk access -5.6489 -8.2565
0 (fixed) 0 (fixed)
Constant (-7.617) (-9.593)
Alternative Specific Local transit with auto driver access (park- -15.6694 -10.0521
0 (fixed) 0 (fixed)
Constant and-ride) (-0.589) (-9.158)
Alternative Specific Local transit with auto passenger access -8.5856 -9.1894
0 (fixed) 0 (fixed)
Constant (kiss-and-ride) (-7.624) (-9.474)
Alternative Specific Regional transit with walk access 0 (fixed) 0 (fixed) -6.0311 -8.4754

12
 Constant (-7.596) (-9.068)
Alternative Specific Regional transit with auto driver access -5.1156 -7.7924
0 (fixed) 0 (fixed)
Constant (park-and-ride) (-6.866) (-9.474)
Alternative Specific Regional transit with auto passenger access -6.333 -8.0629
0 (fixed) 0 (fixed)
Constant (kiss-and-ride) (-8.522) (-9.532)
Alternative Specific Regional transit with local transit access -6.1803 -8.9118
0 (fixed) 0 (fixed)
Constant (-7.997) (-9.34)
Number of Vehicles per Auto driver 0.3749
0 (fixed) 0 (fixed)
Household (2.47)
Ratio of Number of Auto passenger
-2.9066
Vehicles to Number of 0 (fixed) 0 (fixed)
(-2.808)
Persons
Transit Pass Possession All transit alternatives 4.7976
0 (fixed) 0 (fixed)
(10.301)
Trip O/D: City of All transit alternatives 1.8845
0 (fixed) 0 (fixed)
Toronto (5.015)
1 *The scale parameter of the lower nest is estimated as an exponential function in addition to the scale parameter of the upper nest to ensure that the nesting
2 structure is consistent with the theoretical derivation of the nested logit model (i.e. the lower nest’s scale parameter is higher than the upper nest’s scale
3 parameter; both scale parameters are positive).

13
 1 In general, the findings of the three models are consistent. As such, the relative values of the
2 estimated parameters are analogous. For instance, out-of-vehicle travel times (including access,
3 egress, wait, and transfer times) are perceived to be 1.7, 1.6, and 1.9 times higher than in-vehicle
4 travel time according to the MNL, NL, and PLC model results, respectively. Similarly, the value
5 of travel time savings (VOT) is calculated for the different travel time components according to
6 the estimated travel time and cost parameters of the three developed models. The VOT of the in-
7 vehicle travel time in case Wi-Fi is provided/or not is estimated to be $27.95/$24.09,
8 $27.16/$23.49, and $29.77/$26.47 according to the MNL, NL, and PLC models, respectively.
9 The estimated VOT are considered reasonable values given the average wage rates within the
10 sample data (more than 50% of the sample’s household yearly income is $100,000 and above).
11 The VOT for regional transit modes if Wi-Fi is available is lower than the VOT if Wi-Fi is not
12 available. This result indicates that individuals are willing to pay up to six cents per minute to
13 use Wi-Fi on regional transit trains.
14 In terms of the estimated models’ goodness of fit, both the NL and PLC models showed higher
15 rho-squared values than the MNL model. In particular, the NL model showed a slightly higher
16 rho-squared value than the PLC model. The application of joint RP-SP PLC models for policy
17 analysis is an open area of research. However thus far, the model provides valuable insights in
18 terms of the latent segmentation of individuals within the sample as it identifies a probabilistic
19 weight of individuals’ choices to account for their level of dependency on specific modes. The
20 limitation of applying the PLC model for forecasting stems from the complexity associated with
21 the model calibration. Therefore, the NL is recommended to be used for policy analysis as
22 discussed in the following section.
23 POLICY ANALYSIS
24 The developed joint RP/SP NL model is used to predict corresponding changes in aggregate
25 modal shares of cross-regional commuting trips in response to new transportation policies. To
26 effectively use the developed model for predicting market shares, the RP environment is used
27 (because it represents the actual behaviour) and the SP specific variables (i.e. variables that are
28 identified using the SP data which reflect the newly introduced policies) are moved to the RP
29 utilities (without adjusting for the scale effect) to develop a “hybrid” model that can be used for
30 forecasting. The common practice is to adjust the ASC to reproduce market shares that are
31 sufficiently close to those of the sample data (14, 20). After calibrating the ASC, the model can
32 be used in predicting demand changes in response to changes in the explanatory variables.
33 The NL model was calibrated using the full sample RP data and used to test the effectiveness of
34 various policy initiatives. The Interactive Model for Policy Analysis of Cross-Regional Travel
35 (IMPACT) was developed to facilitate testing the policies under investigation. Figure 3 shows a
36 snapshot of the Web-based policy analysis tool with the base-case aggregate modal shares of the
37 full sample RP data. To showcase the model prediction functionality, four independent policies
38 (one policy at a time) are investigated. In each policy analysis, corresponding explanatory
39 variables are adjusted and the predicted model shares are compared to the base-case aggregate
40 modal shares.

14
 1 Policy 1: Introducing Wi-Fi on Regional Transit
2 In this policy analysis, it was assumed that all individuals in the dataset will have access to Wi-Fi
3 service on all regional transit modes. Based on the model’s predicted modal shares, the modal
4 share of regional transit modes has increased from 13.0% to 14.2%. This increase in the regional
5 transit modal share is associated with a 0.9% decrease in the driving modal shares.
6 Policy 2: Introducing Paid Parking at Regional Transit Park-and-Ride Stations
7 The effect of introducing paid parking at regional transit park-and-ride stations (which is
8 currently free) is investigated. In general, the predicted modal shares show a decrease in the
9 regional transit modal share and, in particular, a decrease in the regional transit park-and-ride
10 with auto driver access modal share. Table 3 shows the predicted modal shares of regional transit
11 modes in response to the introduction of different parking costs at regional transit park-and-ride
12 stations.
13 Table 3 Predicted Regional Transit Mode Shares
Parking Cost at Regional Transit Regional Transit Regional Transit Park-and-Ride with
Park-and-Ride Stations Modes* Auto Driver Access Mode**

$1 12.8% 7.4%
$3 12.3% 6.7%
$5 11.9% 6.0 %
14 *Base-case mode share is 13.0%
15 ** Base-case mode share is 7.8%
16
17 Policy 3: Reducing Transit Co-Fares to/from Regional Transit
18 In this policy analysis, regional transit users are exempted from paying the co-fare when local
19 transit is used for access/egress. Accordingly, the model predictions show an increase in regional
20 transit modal shares of approximately 0.76%.
21 Policy 4: Increasing Driving Cost
22 Similarly, the current driving and parking costs at work locations for the driving mode
23 alternatives (i.e. auto driver and auto passenger) are increased by 50%. Based on the new
24 estimated modal shares, the modal shares of the driving mode alternatives decrease by 2.8%,
25 and, interestingly, a corresponding increase in regional transit with driving access (park-and-ride
26 and kiss-and-ride) modal shares of 1.7% is captured.

15
1
2 Figure 3 Snapshot of IMPACT

16
 1 CONCLUSIONS
2 This paper presents an empirical investigation of the travel behaviour of cross-regional
3 commuting trips in the Greater Toronto and Hamilton Area (GTHA). Three mode choice
4 decision structures are examined using joint RP/SP data; namely, Multinomial Logit (MNL)
5 model, Nested Logit (NL) model, and Parametrized Logit Captivity (PLC) model. To the
6 authors’ knowledge, this application of the PLC model using joint RP/SP data is unique.
7 The utilization of RP/SP data stems from the need to test new policies that target improving
8 transit services and transit modal integration while allowing for the scale adjustment of
9 parameter estimates to correct the systematic bias of the SP data. Joint RP/SP data provide
10 comprehensive datasets that enable the development of joint RP/SP models that can explain the
11 probabilistic response in accordance with changes in transportation level-of-service attributes as
12 a result of the introduction of new policies. For instance, the effects of the introduction of paid
13 parking at regional transit park-and-ride stations, the provision of information on local transit
14 vehicles’ arrival time, and the introduction of Wi-Fi service on regional transit trains on the
15 probability of commuters’ mode choice are quantified. The results of the three developed models
16 are presented in a comparative fashion. In general, the findings of the three models are
17 consistent, however, the NL model showed higher rho-squared values than the MNL and PLC
18 models. Therefore, the NL was recommended to be used for policy analysis.
19 The NL model was calibrated and used for the prediction of changes in aggregate modal shares
20 in response to the introduction of new transportation policies. To facilitate this analysis, the
21 Interactive Model for Policy Analysis of Cross-Regional Travel (IMPACT) was developed as a
22 Web-based policy analysis tool. The results of sample policies are presented to showcase the
23 model prediction functionality. As such, this paper presents a unique empirical investigation into
24 the development of joint RP-SP mode choice models utilizing advanced decision structures
25 which provides practical implications on the use of such models for policy analysis.
26 The next step of this research is to investigate the use of the PLC model for policy analysis. In
27 addition, the RP/SP data included information on individuals’ expected departure time in
28 response to the selected travel mode after each SP scenario. Investigating the effect of mode
29 choice on adjusting the departure time choice can provide behavioural insights into the
30 interaction between travel mode and departure choices.

17
 1 REFERENCES
2 [1] Goldman, J., G. Murray, C. Sullivan, B. Whitaker, M. Chase, A. Reisman, N. Whelan, and T.
3 Spencer. Improving Transit Integration Among Multiple Providers, Volume I: Transit Integration
4 Manual. 2015.
5 [2] Mertrolinx. The Regional Transportation Plan for the Greater Toronto and Hamilton Area
6 (GTHA): “The Big Move”, 2008.
7 [3] Data Management Group. 2006 Transportation tomorrow survey. Joint Program in
8 Transportation, University of Toronto, 2008.
9 [4] Data Management Group. 2011 Transportation Tomorrow Survey. Joint Program in
10 Transportation, University of Toronto, 2013.
11 [5] Hensher, D. A. Stated preference analysis of travel choices: the state of practice.
12 Transportation, Vol. 21, No. 2, 1994, pp. 107-133.
13 [6] Hensher, D. A., P. O. Barnard, and T. P. Truong. The role of stated preference methods in
14 studies of travel choice. Journal of Transport Economics and Policy, 1988, pp. 45-58.
15 [7] Louviere, J. J., and D. A. Hensher. Using discrete choice models with experimental design
16 data to forecast consumer demand for a unique cultural event. Journal of Consumer research,
17 1983, pp. 348-361.
18 [8] Wardman, M. A comparison of revealed preference and stated preference models of travel
19 behaviour. Journal of Transport Economics and Policy, 1988, pp. 71-91.
20 [9] Hensher, D. A., J. M. Rose, and W. H. Greene. Combining RP and SP data: biases in using
21 the nested logit 'trick' - contrasts with flexible mixed logit incorporating panel and scale effects.
22 Journal of Transport Geography, Vol. 16, No. 2, 2008, pp. 126-133.
23 [10] Idris, A. O., K. M. N. Habib, and A. S. Shalaby. Modal shift forecasting model for transit
24 service planning: survey instrument design.In The 12th International Conference on Advanced
25 Systems for Public Transport (CASPT), 2012.
26 [11] Idris, A. O. Modal shift forecasting models for transit service planning.In, University of
27 Toronto, 2013.
28 [12] Haener, M. K., P. C. Boxall, and W. L. Adamowicz. Modeling recreation site choice: Do
29 hypothetical choices reflect actual behavior? American Journal of Agricultural Economics, Vol.
30 83, No. 3, 2001, pp. 629-642.
31 [13] Brownstone, D., D. S. Bunch, and K. Train. Joint mixed logit models of stated and revealed
32 preferences for alternative-fuel vehicles. Transportation Research Part B: Methodological, Vol.
33 34, No. 5, 2000, pp. 315-338.
34 [14] Cherchi, E., and J. D. D. Ortúzar. Use of mixed revealed-preference and stated-preference
35 models with nonlinear effects in forecasting.In Transportation Research Record, 2006. pp. 27-34.

18
 1 [15] Mahmoud, M. S., K. M. N. Habib, and A. Shalaby. Survey of Cross-Regional Intermodal
2 Passenger Travel (SCRIPT): A Joint RP/SP Survey Integrated with a Multimodal Trip Planner
3 Tool.In Transportation Research Board 94th Annual Meeting, 2015.
4 [16] Gaundry, M. J., and M. G. Dagenais. The dogit model. Transportation Research Part B:
5 Methodological, Vol. 13, No. 2, 1979, pp. 105-111.
6 [17] Koppelman, F. S., and C. Bhat. A self instructing course in mode choice modeling:
7 multinomial and nested logit models. US Department of Transportation, Federal Transit
8 Administration, Vol. 31, 2006.
9 [18] Mahmoud, M. S., A. Weiss, and K. N. Habib. Latent Captivation or Mode Culture?
10 Investigation into Mode Choice Preference Structures in Competitive Modal Arrangements.In
11 Transportation Research Board 94th Annual Meeting, 2015.
12 [19] Stopher, P. R. Captivity and choice in travel-behavior models. Transportation engineering
13 journal of the American Society of Civil Engineers, Vol. 106, No. 4, 1980, pp. 427-435.
14 [20] Train, K. E. Discrete choice methods with simulation. Cambridge university press, 2009.
15 [21] Habib, K. M. N., M. S. Mahmoud, and J. Coleman. Effect of parking charges at transit
16 stations on park-and-ride mode choice: Lessons learned from stated preference survey in Greater
17 Vancouver, Canada.In Transportation Research Record, 2014. pp. 163-170.
18 [22] Mahmoud, M. S. Demand Modelling of Cross-Regional Commuting Trips in Multimodal
19 Networks.In, University of Toronto, 2015.

19

View publication stats