Investigation of commuting trip timing and mode choice in Greater Toronto Area

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/46490356
An investigation of commuting trip timing and mode choice in the Greater

Toronto Area: Application of a joint discrete-continuous model
Article in Transportation Research Part A Policy and Practice · August 2009

DOI: 10.1016/j.tra.2009.05.001 · Source: RePEc
CITATIONS READS
75 220
3 authors, including:
Khandker Nurul Habib Eric Miller

University of Toronto University of Toronto
266 PUBLICATIONS 2,097 CITATIONS 228 PUBLICATIONS 5,955 CITATIONS
SEE PROFILE SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Sustainable Systems View project
DEVELOPING AN EVALUATION TOOL FOR TRANSIT ORIENTED DEVELOPMENT (TOD) AROUND THE STATIONS OF GO TRANSIT View project
All content following this page was uploaded by Khandker Nurul Habib on 24 February 2020.
The user has requested enhancement of the downloaded file.

Habib, Day and Miller 1
An Investigation of Commuting Trip Timing and Mode Choice in the Greater Toronto Area:
Application of a Joint Discrete-Continuous Model
Khandker M. Nurul Habib, Ph.D., P.Eng.

Department of Civil & Environmental Engineering
School of Mining and Petroleum Engineering (SMPE)
3-004 Markin/CNRL Natural Resource Engineering Facility
University of Alberta
Edmonton, Alberta, Canada T6G 2W2
E-mail: khandker.nurulhabib@ualberta.ca
Phone: 1-780-492-9564
Fax: 1-780-492-0249
Nicholas Day
IBI Group
77 Franklin Street, 7th Floor
Boston, Massachusetts, USA 06830
E-mail: nday@ibigroup.com
Phone: 617-450-0701
Fax: 616-450-0702
Eric J. Miller, Ph.D.

Director and Professor
Cities Centre
University of Toronto
455 Spadina
Toronto, Ontario M5S 2G8, Canada
Phone: 416 978 4076
Fax: 416 978 7162
miller@ecf.utoronto.ca
ABSTRACT
The trip timing and mode choice are two critical decisions of individual commuters mostly define peak
period traffic congestion in urban areas. Due to the increasing evidence in many North American cities
that the duration of the congested peak travelling periods is expanding (peak spreading), it becomes
necessary and natural to investigate these two commuting decisions jointly. In addition to being
considered jointly with mode choice decisions, trip timing must be also be modelled as a continuous
variable in order to precisely capture peak spreading trends in a policy sensitive transportation demand
model. However, in the literature to date, these two fundamental decisions have largely been treated
separately or in some cases as integrated discrete decisions for joint investigation. In this paper, a
discrete-continuous econometric model is used to investigate the joint decisions of trip timing and mode
choice for work commuting trips in the Greater Toronto Area (GTA). The joint model, with a
multinomial logit model for mode choice and a continuous time hazard model for trip timing, allows for
unrestricted correlation between the unobserved factors influencing these two decisions. Models are
estimated by occupation using 2001 travel survey data for the GTA. Across all occupation groups,
strong correlations between factors influencing mode choice and trip timing are found. Furthermore, the
estimated model proves that it sufficiently captures the peak spreading phenomenon and is capable of
being applied within the activity-based travel demand model framework.
INTRODUCTION
The trip timing and mode choice decisions of commuters are critical considerations for transportation
demand modellers. These two decisions along with route choice directly determine the temporal
distribution of demand experienced on any given piece of transportation infrastructure in an urban area
(Sushilo and Kitamura, 2007). Certainly, increasing congestion and transportation demand management
policies in modelled future year scenarios can be expected to influence both temporal and modal
decision making. Despite the importance of the impacts of trip timing decisions on the distribution of
transportation demand, in practice most transportation demand models used at the regional planning
level do not explicitly model the temporal dimension of travel. For example, the latest version of the
Southern California Association of Governments (SCAG) four stage Regional Transportation Model
(SCAG, 2003) for the Greater Los Angeles Area and the Greater Toronto Area’s GTA Model (Miller,
2007) use peaking factors derived from observed origin-destination survey data to subdivide generated
trips/journeys into discrete time periods for trip assignment. These fixed peaking factors calculated from
base year data are inherently insensitive to the changing temporal conditions and policies of future years.
Activity-based models, on the other hand, handle trip timing decisions more comprehensively by
considering the detailed impacts of scheduling constraints on trip timing and mode choice at the
disaggregate level (Davidson et al, 2007). However, even such activity-based models have difficulties
with introducing policy sensitivity to the distributions of the activity frequencies and start times that
serve as key inputs to their activity schedulers. Activity-based models such as TASHA, FAMOS, and
ALBATROS, use base year distributions cross-classified by different variables to simulate activity-
travel demand (generation) (Roorda et al., 2008; Pendyala et al., 2005; Arentze and Timmermans, 2004).
This approach is insufficient when considering policies and scenarios that have the potential to
significantly shift travel trends away from the base year distribution of activity start times. Activity-
based models by Vovsha and Bradley (2004) and Bowman and Ben-Akiva (2000), on the other hand, are
examples of models that include explicit tour-based time-of-day discrete choice models to schedule the
travel tours of individuals. Although this approach represents a significant improvement over the use of
base year distributions, as will be discussed in the next section, limitations remain with the
representation of time as a discrete variable. Overall, a better understanding and representation of the
distribution of activity start times/trip departure times is required.
In reality, it is reasonable to expect significant differences in observed home-work departure time
distributions by mode of travel. Each mode offers very different levels of service throughout the day.
For example, transit modes are considerably less attractive during off-peak periods due to lower service
levels, while the auto mode is more attractive during off-peak periods due to lower traffic congestion. In
the extreme, some transit modes (e.g. commuter rail) may be completely un-available during certain
time periods due to limited periods of operation. As a result, an integrated treatment of mode and time of
day choice that explicitly considers their interdependence or endogeneity is important. In this paper, 24
hour joint models of Home-Work (H-W) trip timing and mode choice are estimated using revealed
preference survey data from the Greater Toronto Area (GTA). A joint discrete-continuous model
specification, with a continuous time hazard model for home-to-work (H-W) trip timing and a
multinomial logit model for mode choice, is used to capture the interconnected nature of mode choice
and trip timing decisions.
The paper is arranged as follows: first a review of pertinent literature on trip timing models is
presented; this is followed by a description of the theory behind the joint model structure used, a
discussion of the detailed model estimation, model sensitivity analysis and model application. The paper
concludes with a summary of the investigation and recommendations for future studies.
MODELLING TRIP TIMING
Discrete Choice Models

The vast majority of trip timing models in the literature employ a discrete choice utility maximization
framework, which requires the definition of finitely many contiguous time period choices. Small (1982)
constructs a multinomial logit (MNL) time allocation model for home-work morning departure time
choice that incorporates scheduling considerations. In Small’s model, each commuter is presented with
the choice between twelve discrete work arrival times periods grouped into 5-minute arrival intervals.
Small argues that during work trip scheduling, individuals consider their expected travel time and make
a trade-off between arriving early, which involves some time wasted, and arriving late, which can have
more severe consequences that depend on one’s level of work hour flexibility.
A significant issue encountered when applying discrete choice models to time of day choice is that
neighbouring time period choices are likely to be correlated, especially if the discrete time periods are
short. The appropriate discretization of time in travel demand modelling is always a significant
challenge because of the natural correlations that exist among different time periods (Russo et al, 2009).
However, this correlation is a violation of the Independence of Irrelevant Alternatives (IIA) property of
the multinomial logit (MNL) model, which requires that the unobserved components of the utility of
each alternative be uncorrelated. As such, a significant proportion of more recent time of day choice
research has focused on applying more complicated discrete choice models with less restrictive
assumptions on correlation. For example, Small (1987) and others have derived models belonging to the
Generalized Extreme Value (GEV) family to allow for the accommodation of correlation among time
period alternatives. Small derived the Ordered Generalized Extreme Value (OGEV) model specifically
for problems with ordered alternatives, such as the naturally ordered time period choices in a time of day
choice model. Logically, one would expect the highest degrees of correlation between adjacent time
period choices, with correlation decreasing as the distance between alternatives increases. The OGEV
model recognizes this fact by introducing a correlation parameter for each pair of alternatives that is a
function of the distance between them. When applying the more theoretically robust OGEV framework
to the work departure time survey dataset previously used in Small (1982), likelihood ratio tests revealed
that the OGEV model did not perform well in comparison to a NL model and the original MNL
framework. Furthermore, the OGEV model’s correlation parameters were only found to be marginally
significantly different from a value of one. A study by Bhat (1998), on the other hand, is an example of a
considerably more successful application of the OGEV model to trip timing decisions. Bhat jointly
models urban shopping trip departure time and mode choice using a nested modelling structure with a
MNL model at the top level for mode choice (drive alone, shared ride, transit) and an OGEV model at
the bottom level for departure time choice between five contiguous discrete time periods that span the
entire day. Further more general GEV formulations with very flexible correlation structures have also
been applied to trip timing problems by various authors. A more prominent example uses the mixed
MNL, which “is no less general than the [Multinomial Probit] MNP model in that it can also estimate a
complete variance-covariance matrix” (de Jong et al., 2003).
Regardless of their relative merits, the discrete choice models discussed above share common
theoretical and practical issues that stem from their framework to represent time. Even an application of
the probit model, which allows for the most general variance-covariance matrix, would not solve all of
the problems associated with the discrete representation of time. As discussed by Bhat and Steed (2002),
even the most complex discrete choice models suffer from the following problems:
• It is difficult to conclusively argue what exact time interval size should be used for
discretization;
• The times near the boundaries of adjacent discrete periods are assumed to be distinct (and
possibly independent) choice alternatives; and
• Any temporally based policies for evaluation must also be applied within the discrete time
periods defined during model estimation.
This is also true even in very recent studies using trip or activity timing models (Ozbay and Yanmaz-
Tuzel, 2008; Ettema et al, 2007). At the most basic level, all of these problems stem from the fact that
time is more naturally (and simply) represented as a continuous variable. Attempting to force time into a
discrete framework is inherently limiting, often requiring unrealistic simplifying assumptions
(independence of adjacent time periods) or highly complex models (OGEV, Mixed Logit, etc.) to
compensate. Furthermore, employer-specified work start times and a close familiarity with en route
travel times and congestion variability can be expected to restrict commuters to relatively narrow home-
work departure windows within which it is most appropriate to represent time as a continuous variable
(Bhat, 1998). Another point to consider is that the discrete choice framework for trip timing is often
incompatible with state of the art activity-based microsimulation models since exact skeletal activity
start times are required as inputs to their activity scheduling components (Roorda et al., 2008). As such,
a continuous representation of time for commuting trip timing is selected for this paper.
Continuous Time Models

An alternative to the application of discrete choice theory is modelling trip timing following the
‘equilibrium scheduling theory’ (EST) proposed by Vickery (1969). EST is a trip timing model that
considers that travellers justify early/late arrival times relative to a desired arrival time. Hyman (1997)
and van Vuren et al (1999), for example, use this approach to capture congestion effects on trip timing
decisions. This approach is often cited as reducing the necessity of a strict discretization of time.
However, this approach deviates from the application of discrete choice theory primarily at the
theoretical level since the specification of the basic trade off in trip time decisions is basically the same
for the EST approach and discrete choice approach (deJong et al, 2003). Individuals are effectively
presented with the discrete choice between arriving at the desired time, and shifting to a later/earlier
time. The discretization of time at an appropriate level still remains as an issue. Indeed, to avoid the
practical difficulties associated with discretizing; theoretically, time should be considered as a
continuous entity (Bates and Polak, 1997). Indeed, Palma and Marchal (2002) prove that the continuous
time assumption is unavoidable if one is to capture transportation system dynamics properly; they
discuss the use of continuous time as an event-based approach.
In a recent study, Bhat and Steed (2002) estimate an event-based model, which is a continuous
time hazard duration model of urban shopping trip timing decisions. Hazard models recognize the
dynamics of event/activity durations by considering the conditional probability of event termination,
usually as a function of covariates (explanatory variables). In this case, the authors specifically apply a
time varying proportional hazard model that accommodates time varying coefficients and time varying
explanatory variables. Furthermore, duration is split into smaller grouped intervals where the baseline
hazard rate is assumed to be constant. The authors argue that the definition of these time intervals can be
motivated by the fact that individuals typically report rounded off times in surveys (e.g. to nearest 5
minutes). The individual’s actual departure time is not necessarily observed, just its interval. However,
the discretization does place strong assumptions on the form of the baseline hazard rate, simplifying
model estimation considerably. Indeed, it should be recognized that a model with a different
discretization scheme will have a different baseline hazard rate and hence different results.
In a continuous time study that departs from the time hazard model, Ettema and Timmerman
(2003) propose that an individual’s departure time for each activity can be determined through time
budget constrained utility maximization, where one’s total utility is derived from activity participation
(positive utility) and travel (negative utility). A critical assumption in their model is the hypothesized
form of the marginal utilities of activity participation (the utility gained per unit time). For each activity,
the authors propose parametric marginal utility functions of the exponential form that exhibit high
utilities at times near the preferred period of activity participation and continuously decreasing utilities
as one moves farther away from the preferred period (on either the late or early side). As a result, the
specification also requires that the estimation dataset include each individual’s planned activity schedule
that includes a list of activities to participate in, in what order, approximate timings, and by which
transportation mode. Although the results of the model are promising, the model applies highly
specialized data that requires the detailed activity schedule of individuals, including complete activity
start and end times. Such specialized data is not available in the estimation dataset used for this paper
and is unlikely to be available in a large dataset suitable for statistically robust model estimation results.
Joint Discrete-Continuous Models

All of the literature discussed in the previous sub-sections has modelled trip timing independently from
mode choice, with the exception of the MNL-OGEV discrete choice model for shopping trips by Bhat
(1998). If a continuous time model is used for trip timing and a discrete choice model is used for mode
choice, a discrete-continuous model is required for joint model estimation. This paper is concerned with
modeling trip timing and mode choice decisions in an effort to better understand the peak spreading
phenomenon. As a result, a discrete-continuous modelling approach, considering time as a continuous
variable and mode choice as a discrete decision, is most appropriate. This section presents a selection of
previous joint discrete-continuous models that are relevant to this study.
Early studies by Krishnamurthy and Raj (1988), and Barnard and Hensher (1992) present similar
joint discrete-continuous models of brand choice and consumption, and shopping location choice and
rate of expenditure respectively. Most hazard model applications to time of day choice have focused on
situations where activities end as a result of a single cause. A study by Bhat (1996), however, extends
the hazard modelling framework by deriving a general “competing risk” model with multiple entry/exit
states (causes) and applies it to the joint discrete-continuous choice of after-work activity participation
and activity duration. Hensher and Ton (2002), on the other hand, discuss a joint mode choice and
departure time modelling framework within the integrated model, TRESIS: a transportation, land use
and environmental strategy impact simulator for urban areas. However, very little information is widely
available in the literature about the complete formulation of the joint model structure in TRESIS at this
point in time. Further, Pendyala and Bhat (2004) use a joint discrete-continuous model to investigate the
relationship between maintenance activity start time and activity duration. In a recently published paper,
Munizaga et al. (2006) derive and apply a discrete-continuous econometric model similar to the one
used in this paper to work time allocation and mode choice. However, their study did not consider trip
timing, which is important for assessing peak period traffic congestion and peak spreading. More
recently Habib et al (2008) applied a discrete-continuous modelling approach to investigate the social
context of activity scheduling in terms of finding the relationship between the start time / duration of
social activities and with-whom an individual socializes with.
MODEL STRUCTURE
This paper presents a joint model for an individual’s continuous choice of Home-Work (H-W) trip
timing along with his or her discrete mode choice. From a practical modelling standpoint, it is equivalent
to model H-W departure time and work arrival time since adding the expected travel time to one’s H-W
departure time obtains the work arrival time. A hazard-based model is employed to model the work
timing decisions, which are represented by the number of minutes elapsed past midnight until the
departure/arrival time, while the discrete mode choice is modelled using a multinomial logit model
(MNL). The joint estimation of these two choices results in a discrete-continuous model; the MNL
effectively “selects” a sample for the hazard model where the alternatives of the MNL model represent
different possible temporal entry/exit states. The structure and features of each individual discrete and
continuous model component and the joint model are discussed in the following three sections.
Multinomial Logit (MNL) Mode Choice Model

The multinomial logit (MNL) model, derived from random utility theory, is a commonly used discrete
choice model in transportation problems. If each alternative mode m has a person-specific utility for
person i (Umi), and given a vector of explanatory variables (xmi), the utility is typically expressed in the
following linear in parameters form:
U mi = Vmi + ε mi = β m xmi + ε mi (1)
Where, Vmi is the systematic utility, which is a function of a set of explanatory variables (xmi) and
corresponding parameters of weighting factors (βm). The random variable εmi in the utility function is the
unobservable individual and alternative specific error term. According to Random Utility Maximization
(RUM) theory, an alternative mode m will be chosen by an individual i, if the utility of that alternative
mode is the maximum of all alternative modes.
max
U mi > U ni
n = 1,2,3......M , n ≠ m
⎧ max ⎫
Vmi > ⎨ U ni ⎬ − ε mi
⎩n = 1,2,3......M , n ≠ m ⎭
So (2)
max ⎛ ⎧ max ⎫ ⎞
Pr(U mi> U ni ) = Pr ⎜⎜Vmi > ⎨ U ni ⎬ − ε mi ⎟⎟
n = 1,2,3......M , n ≠ m ⎝ ⎩n = 1,2,3......M , n ≠ m ⎭ ⎠
= Pr (Vmi > (Vni + ε ni ) − ε mi )
= Pr (Vni ≤ Vmi + (ε mi − ε ni ) )
The above equations reveal that the systematic utility of a chosen alternative is a function of the
difference between two random error terms: the error term of the chosen alternative (εmi) and the error
term of the second best alternative (εni). Here, we assume that the error term of the utility function of
equation (1) is identically and independently distributed (IID) Type I Extreme-Value (Gumbel).
According to the properties of Type I Extreme-Value distribution, the maximum over an IID Extreme-
Value random variable is also extreme value distributed and the difference of two IID Extreme-Value
random terms is logistically distributed (Johnson et al, 1995; Pendyala and Bhat, 2004). Hence, the
implied cumulative distribution of the random error term of the chosen alternative, F(εmi) can be written
as (Ben-Akiva and Lerman, 1985; Train, 2003 ):
exp(Vmi ) exp( β m xmi )
Pr(Vmi > Vni ) = F (ε mi ) = = (3)
exp(Vmi ) + ∑ exp(Vni ) exp( β m xmi ) + ∑ exp( β n xni )
n≠ m n≠ m
Equation (3) is the formula of well known Multinomial Logit (MNL) model.
Continuous Time Hazard Model

Continuous time hazard models are primarily concerned with the time or duration until event
termination. In this paper this is the time until one departs for work, or equally when the “at home”
event terminates with a work trip. The basic formula describing event termination in hazard models is
the hazard rate λ(t)dt, which is the conditional probability of event termination occurring between time t
and t+dt given that the event has not terminated before time t. For unit time step, dt=1, the hazard rate is
expressed simply by λ(t). The mathematical expression of the hazard rate is: λ(t)=f(t)/[1-F(t)]. Where
the cumulative distribution function F(t) describes the probability of event termination before time t,
F(t) = Prob [T < t] and f(t) is the corresponding probability distribution function, f(t) = dF(t)/dt
Similarly, the survivor function S(t), which defines the probability that the event’s duration will
be greater than or equal to time t, is defined as S(t) = Prob [T≥t] = 1 – F(t).Therefore, the hazard
function can also be re-written as: λ(t)=f(t)/S(t). In hazard models, the hazard rate is specified as a
function of an assumed baseline distribution and covariates (Zi). There are two common specifications
for accommodating covariate affects in hazard models: Proportional Hazard and Accelerated Hazard
models. In proportional hazard models it is assumed that the covariates modify the hazard function
directly by having a proportional or multiplicative effect; the hazard rate is effectively decomposed into
one term dependent on time and another dependent only on the covariates (Hensher and Mannering,
1994). In the accelerated hazard model, on the other hand, it is assumed that the covariates rescale or
accelerate time directly in the baseline survivor function. The hazard rate varies over time as it is
accelerated or decelerated by the covariates. The accelerated hazard specification was chosen to model
trip timing decisions in this paper since one can expect considerable dynamics in the hazard rate due to
the behavioural nature of trip timing decisions; the proportional hazard formulation may not be realistic
in such situations (Lee and Timmermans, 2007). Furthermore, the accelerated hazard model has the
added benefit of allowing one to directly interpret the relationship between covariates and event duration
due to their multiplicative effect on time.
Assuming that the covariates act in the form exp(θZ), the accelerated hazard model can be
expressed as: S(t|Z)=S0[t.exp(θZ)] and λ(t|Z)=h0[t.exp(θZ)]exp(θZ). Following the method outlined by
Kiefer (1988), it can be shown that the accelerated hazard model can be expressed in the following
equivalent linear form (for each individual i and mode m ):
ln(t mi ) = θZ mi + α mti (4)
Where, the assumed distribution of the error term αmti (normal with mean zero and variance σmti) defines
the form of the specific accelerated hazard model by defining its probability density function, which in
our case becomes a lognormal accelerated hazard model. So, the PDF, f( . ) and CDF, F( . ) of the error
term of our continuous time model can be written as (Johnson et al, 1995):
1 ⎛ ln(t mi ) − θZ mi ⎞
f (α mi ) = ϕ ⎜⎜ ⎟⎟
t miσ mti ⎝ σ mti ⎠
(5)
⎛ ln(t mi ) − θZ mi ⎞
F (α mi ) = Φ⎜⎜ ⎟⎟
⎝ σ mti ⎠
Joint Model
The joint estimation of an accelerated hazard model for departure time choice and a multinomial logit
model for mode choice requires that the random error terms of both models are correlated. This means
that the joint distribution of εmi and αmti has non-zero correlation coefficients. In our case the random
error terms εmi and αmti have different types of distributions: εmi is IID Extreme-Value Type I distributed
with zero mean and unit variance, while αmti is normally distributed with zero mean and σmti variance.
One way of specifying the correlation between a normal and a non-normal random variable is to
transform both random variables into an equivalent standard normal variable and specify the joint
distribution as equivalent multivariate normal distribution (Lee,1983). In our case, the marginal
distributions of the two random variables (F(εmi) and F(αmti) are described in equation (3) and equation
(5) respectively. Transforming these marginal distributions into equivalent standard normal variables, it
can be shown that (Lee, 1983):
ε mi
*
= J1 (ε mi ) = Φ −1[ F (ε mi )]
(6)
α mi
*
= J 2 (α mti ) = Φ −1[ F (α mti )]
Here, Φ-1 indicates an inverse of the cumulative standard normal variable. The transformed
variables J1(εmi)and J2(αmti) are transformed standard normal variables of the corresponding random
variables εmi and αmi respectively. The joint decision process of mode choice and departure time now can
be described by the considering that the transformed standard normal variables are bivariate normal
(BVN) distributed with correlation ρmti: BVN[J1(εmi ), J2(αmti), ρmti]. Hence, the joint probability of
observing that any individual (just ignoring the individual identifier, i) choosing an alternative mode m
and a corresponding departure time, tm can be expressed as follows(Habib et al., 2008):
Pr(Time = t m ∩ Mode = m) = Pr(Time = t m ∩ ε ≤ J 1 (ε m ))
1 ⎛ ln(t m ) − θ m Z m ⎞ ⎛⎜ J 1 (ε m ) − ρ mt J 2 (α mt )] ⎞⎟ (7)
ϕ ⎜⎜ = ⎟⎟Φ
σ mt t m ⎝ σ mt ⎠ ⎜⎝ 1 − ρ mt ⎟
2
⎠
Here φ( . ) and Φ( . ) indicate the PDF and CDF of the Standard Normal distribution,
respectively. Based on the above formulation, the likelihood (L) and log-likelihood (LL) function of a
sample of observations with sample size N, (i=1, 2, ……..N) can be written as:
Dmi
⎛ 1 ⎞
Dmi ⎛ ⎛ ⎛ ⎞⎞
⎜ φ ⎜ ln(t mi ) − θZ mi ⎞⎟Φ⎜ J1 (ε mi ) − ρ mti ((ln(t mi ) − θZ mi ) / σ mti ) ⎟ ⎟
N M
L = ∏∏ ⎜⎜ ⎟⎟
⎜⎜ ⎜ ⎟ ⎜
i =1 m =1 ⎝ t miσ mti ⎠ σ mti ⎟ ⎟⎟
(1 − ρ mti )
2
⎝ ⎝ ⎠ ⎝ ⎠⎠
N M
LL = ∑ ∑D mi ln( (φ (ln(t mi ) − θZ mi ) / σ mti ) − Dmi ln(t miσ mti ) (8)
i =1 m=1
+ Dmi ln( Φ ( ( J1 (vmi ) − ρ mti ((ln(t mi ) − θZ mi ) / σ mti ))) / (1 − ρ mti ) ) )

2
Here Dmi is the indicator variable for choosing alternative m by individual i. In this paper, parameter
estimates for β, θ, ρ, and σ are obtained by maximizing the above log-likelihood function using code
written in GAUSS, which applies the BFGS optimization algorithm (Aptech, 2006). The case of M
number of modes, leads to a total of (M-1) number of σ and (M-1) number ρ parameters that can be
estimated.
In terms of the interpretation of correlation parameters of the joint mode choice and departure
time choice model, according to equation (7) a negative value of the correlation coefficient implies a
positive correlation between the unobserved factors affecting the discrete choice of alternative m and the
continuous time t. Therefore, a negative value of ρmt implies that choosing alternative mode m influences
longer t values, i.e. later departure/arrival times and vice versa.
DATA PREPARATION AND EXPLANATORY VARIABLE DEFINITIONS

A subset of the 2001 Transportation Tomorrow Survey (TTS), a multimodal travel survey conducted in
the Greater Toronto Area (GTA) every five years, was used to generate a base set of feasible Home-
Work-Home (H-W-H) tours (DMG, 2005). At the completion of data preparation, each individual
observation (102,975 in total) had a set of available mode choices and an observed feasible mode choice,
an observed H-W departure time, an observed H-W arrival time, and observed work duration. For
indirect H-W-H tours, the number of stops on both the individual H-W and W-H journeys were
calculated and recorded. The available modes are: Auto Driver (1), Auto Passenger (2), Transit all way
(3), Transit Park and Ride (4), GO Park and Ride (5), and Walk (6). Here transit indicates local transit
and GO is a regional transit system (commuter rail) in the GTA.
TABLE 1, below, defines the complete set of explanatory variables used in the final specifications of the
joint H-W Departure Time models. Most mode choice level of service variables were calculated on a
tour basis (as noted in the Table), while the trip timing covariates were calculated only for the H-W
journey. Auto travel times were calculated on an hourly basis for all 24 hours of the day by assigning all
observed zone-to-zone trips to an EMME/2 traffic assignment model. All other household and individual
level sociodemographic variables and trip attributes were obtained directly from or generated from TTS,
except where noted otherwise in TABLE 1.
MODEL ESTIMATION AND SPECIFICATION PROCESS

The model was first specified for the professional occupation group through the use of an iterative
“backward selection” procedure. This procedure started by estimating a model that included all possible
explanatory variables with a plausible causal relationship with the utility of each mode and/or H-W
departure time. Statistically insignificant variables were iteratively removed from the model
specification based on the estimated t-Stat and P-values of their parameters. A key measure used during
the professional model specification was the implied value of travel time savings for auto users, which
was obtained by dividing the model’s auto travel time and cost parameter estimates. The magnitude of
the auto value of travel time savings is an indicator of the validity of any model; it must be of a plausible
magnitude (on the order of prevailing wage rates) and should show realistic differences between
occupation groups.
Table 2, summarizes the joint H-W departure time model estimation results for the mode and
departure time choice components of the model, key distributional parameters (correlation coefficients
and hazard model variances) and presents goodness of fit measures for each occupation type. The vast
majority of the estimated parameters in both the discrete mode choice and continuous trip timing
components of both joint models remained statistically significant across all occupation groups,
surpassing critical t-Stat values at the 5% significance level by a wide margin in most cases. The few
parameters which became statistically insignificant when moving from the base professional departure
time model to the other models remain in Table 2 to illustrate the full extent of the differences between
occupation groups..
As per the formulation of the joint discrete-continuous model, the negatively valued correlation
coefficients indicate a positive correlation between the unobserved factors affecting the choice of mode
and departure time. Due to the joint nature of the model, one can equivalently interpret the negative
correlation coefficients as implying that: unobserved factors affecting workers choosing that mode
influence departing during later time periods and unobserved factors affecting workers choosing to
depart during later time periods influence choosing that particular mode. The converse is true for
positively valued coefficients. In general, the estimated correlation coefficients in Table 2 are valued as
expected, with relative magnitudes that reflect the observed differences between occupation groups and
modes. Most importantly, all correlation coefficients (except for one) are statistically significant across
all occupation types and modes, indicating that there are significant amounts of unobserved factors
affecting mode choice and departure time choice that are not captured if independent models are used
for mode choice and departure time separately. Hence, the joint model specification is merited. The
interpretations of the values of the individual correlation coefficients are discussed below. It should be
noted that interpretation of the correlation parameters of the joint models are conditional on the fact that
the observed variables failed to capture these unobserved relationships. The interpretations refer to
affects over and above those explained by the variables used in the model.
For the professional occupation group, GO and transit Park and Ride (P&R) users have the most
positively valued correlation coefficients, which means that they can be expected to depart the earliest
relative to the users of other modes. This is consistent with both the constrained nature of the P&R mode
(limited parking and schedule constraints) and the fact that long distance trips are common among P&R
users. The walk coefficient, on the other hand, is by far the most negative, which reflects the fact that
this mode is only feasible for relatively short trips. Walkers are generally able to depart from home
much later in comparison to the users of other modes. Lastly, the auto and transit coefficients are both
negative and of a similar magnitude, with the auto coefficient being less than the transit one. This is to
be expected since both modes are not highly constrained and compete within similar markets. The auto
coefficient is more negative due to the fact that auto users generally have more flexibility than transit
users due to the lack of any hard mode-related scheduling constraints.
The results for the other occupation groups are to some extent consistent with that of the
professionals. The auto mode correlation coefficient is always among the most negative except for the
manufacturing group, indicating that all auto users except those employed in manufacturing generally
have the flexibility to depart later due to its consistent availability and high level of service during most
periods of the day. Similarly, the walk mode coefficient (with the exception of the manufacturing group)
is generally among the most negative, which is a result of the short length and inherently flexible nature
of walk trip departures that allow individuals to depart comparatively later. The local transit mode
correlation coefficients, on the other hand, are positive for all occupation groups except the professional
group. This indicates that local transit users can be expected to depart later in the day if they are not
professionals. This can be explained by the fact that the professional occupation group usually has more
defined traditional 9-to-5 working hours, requiring local transit users of this group to depart earlier in the
day if they are to arrive on-time. The other occupation groups, however, can generally be expected to
have work start hours that are more spread through the day, especially for retail and manufacturing
workers, allowing for later departures. The transit P&R coefficients, on the other hand, are consistently
among the most positive, reflecting their highly constrained nature. However, once again the results for
manufacturing workers are different from that of the other occupation groups; the GO P&R coefficient
has the most negative value in the manufacturing model.
Although the interpretations of the correlation coefficients are generally consistent across
occupation groups, there are some marked deviations for the manufacturing group in particular. The
reason behind such differences between occupations is likely due to the basic differences in home-work
location distribution patterns, work start time flexibility, and work duration variations. Indeed, unlike
professional and general office workers, which can generally be expected to arrive at their workplaces
during traditional office work start times (i.e. 9:00 am), manufacturing and retail workers generally do
not have a single predominant work arrival time. Their work hours are spread throughout the day into
distinct work shifts (e.g. morning, afternoon, evening, etc.). This makes the interpretation of their
correlation coefficients inherently more difficult; each mode has vastly different levels of attractiveness
that depend on one’s departure time (and therefore work start/shift time), origin, and desired destination.
For example, during early morning and late evening work shifts, the auto mode may be the only feasible
mode for long distance travel. This is especially likely to be the case for manufacturing workplaces
which are typically located in very auto-oriented suburban areas.
Certainly, there are many competing factors at work in each of the different occupation groups,
including vastly different work shift timing policies and geographic considerations that affect the
availability and level of service offered by each mode. This illustrates the fact that one has to be careful
when interpreting the values of individual correlation coefficients since they only capture the correlation
between unobserved factors (not explainable by observed variables) affecting mode choice and
departure time patterns. In such cases, it is nearly impossible to find out a definite reason for differences
between correlation values; we can only discuss likely reasons. The merit of joint estimation is that it
captures such unobserved correlations, which would be completely overlooked if independent models
were used instead (causing the estimated parameters in independent models to be biased). To be sure,
regardless of the interpretation of each correlation coefficient’s absolute value, the most important result
is that mode choice and trip timing decisions are indeed correlated; their very statistical significance
guarantees this important result.
The estimated mode-specific variances in TABLE 2 describe the variability in the observed H-W
departure times for each mode. They directly influence the random spread of the normal error terms in
the hazard model. The mode-specific specification was chosen in favour of a generic one to account for
the highly constrained nature of the P&R modes. Models that used generic variances across all modes
did not reproduce observed P&R cumulative distributions well since they did not sufficiently restrict the
choice of the mode to narrow early morning time periods when parking is available. Indeed, it is evident
from TABLE 2, that the variances of the two P&R modes are consistently much smaller than the other
modes across all occupation groups. The other three modes do show statistically significant differences
in the values of their variances, but they remain of relatively similar magnitude. This is to be expected
since their availability is not restricted to narrow times in the same manner that the P&R modes are since
they are generally available throughout the day. Lastly, it should be noted that although the absolute
magnitudes of the mode specific variances are different between occupation groups, their relative
magnitudes and therefore their interpretations are similar. These differences reflect the observed spread
in times and the model’s level of uncertainty for each given occupation; higher variance magnitudes are
observed for the occupations with lower likelihood ratio values.
TABLE 2 also summarizes the parameter estimation results of the mode choice component of the
joint H-W departure time model by occupation type. The column entitled “Mode” specifies the utility
function(s) that each variable lies in. For professionals, all parameters are statistically significant and of
their expected sign and magnitude. When moving from this base model to the other occupation groups,
few parameters lose their significance and the majority of all remaining statistically significant
parameters are also of their expected sign and relative magnitude. Some key behavioural interpretations
of the parameter estimates and significant differences between occupation groups are discussed below.
In general, across all occupation groups, the mode specific parameter estimates of the full-time
dummy variable indicate that full time workers are more likely to drive and use Park and Ride, but are
less likely to walk. This is a result of the fact that full time workers are more likely to own a personal
vehicle and have more control over their household’s vehicle(s). While the positive parameter for the
total stops variable for all occupation groups indicates that increasing numbers of intermediate stops on
the H-W-H tour increase one’s likelihood of driving. This is to be expected since non-auto modes are
considerably less flexible and convenient when multiple stops are required along one’s journey.
Furthermore, across all occupations, the parameter for auto in-vehicle time, AIVTT, is more negative
than the transit in-vehicle travel time parameter, TIVTT, indicating that each unit of time spent
travelling on transit is perceived to be worse than it is in one’s personal automobile. The parameter for
transit wait time, TWAIT, on the other hand, is an order of magnitude more negative than TIVTT, which
implies that time spent waiting for transit is even more onerous. Such differences in the perceptions of
time by mode and by in-vehicle and out-of-vehicle time are a typical and expected result in most mode
choice models.
Moving on to the parameter estimates for socioeconomic variables, individuals in households
with two or more vehicles are found to be much more likely to drive, significantly less likely to take
local transit or walk, and more likely to take Park and Ride (all with respect to the likelihood of being an
auto passenger). While older individuals are more likely to drive, less likely to walk, and individuals
living in higher income zones are less likely to use transit and less likely to walk. Similar trends are
observed across occupation groups.
Inconsistent results are found for the effects of the urban density of the household zone,. The
parameter estimates for professionals indicate that individuals are less likely to drive if they reside in
higher density areas; while the results for manufacturing workers indicate the opposite effect (the
parameter is insignificant for the other occupations). The first negative effect is the expected one since
higher density areas are typically more transit and walk friendly in nature. The unexpected result for
manufacturing workers, however, is likely due the fact that most manufacturing workplaces are located
in very suburban (low density) locations. In this case, the walkability of the work location is much more
important than that of one’s household location. Indeed, higher urban densities at one’s workplace are
found to have a consistent positive effect on the likelihood of choosing the local transit and walk modes
regardless of one’s occupation. Furthermore, the parameter for the work urban density is an order of
magnitude larger than that of the home urban density for all occupation groups.
To further illustrate the differences between each occupation group, auto value of time (VOT)
savings were calculated from each model’s parameter estimates by dividing the auto travel time
parameter by the total auto cost parameter. As expected, professional workers were found to have the
highest value of travel time saving at $43/hr. For manufacturing, general office, and retail workers, on
the other hand, the values of travel time saving were found to be approximately $13/hr, $32/hr and
$16/hr respectively. Professionals and general office workers are found to have the most similar
characteristics, which is to be expected since both occupation groups are employed in similar
workplaces and travel during similar congested periods of the day. Overall, the absolute and relative
magnitudes of the values of time are plausible, with expected similarities and differences between
occupations.
TABLE 2 also summarizes the estimation results for the continuous time departure time
component of the joint model by occupation type. The vast majority of parameters are statistically
significant and of expected sign and magnitude. Insignificant parameter values are primarily relegated to
mode specific variables in occupation groups with small numbers of observations for a particular mode.
Key parameter values and similarities and differences by occupation group are discussed below.
Firstly, full-time workers were found to depart earlier than part-time workers across all
occupation groups. The work duration parameter, on the other hand, was found to be consistently
negative and of similar magnitude across occupation groups, implying that individuals depart earlier
when they have longer total work hours. This is an intuitive result; individuals must arrive at work
earlier if they are to allow for longer work days.
As expected, the larger the household size, the later one’s departure time; with more individuals
in the household, one is more likely to work during traditional morning work hours and is likely to be
more constrained by facilitating trips for other household members (e.g. dropping children off at school
or daycare) on the way to work. Older individuals and males were found to depart earlier than younger
individuals and females respectively. Furthermore, increasing numbers of stops on one’s Home-Work
journey were found to cause individuals to depart earlier, as expected. All other factors being held equal,
stops on the Home-Work journey require earlier departure times if one is to arrive at work on-time; they
make one’s effective travel time longer.
Higher urban densities in the household and workplace zone were found to have a positive effect,
indicating that individuals can depart later when the urban area along their commute is more conducive
to walking and taking transit. Finally, individuals with access to free parking in planning districts 1 and
4, downtown Toronto and North York Centre respectively, were found to depart later than those that had
to pay for parking. This effect is to be expected since people with free parking have more flexibility, as
they do not have to compete for limited spaces, and do not have to consider early arrival incentives such
as “early bird” discounts. It also should be noted that this variable was not statistically significant when
applied throughout the entire study area since almost all individuals employed outside of the highly
dense areas of planning districts 1 and 4 have free parking.
Now moving on to the mode-specific travel time parameters, increasing travel times have a
consistent negative effect on departure time, causing individuals to depart from home earlier across all
modes and occupations. This result is to be expected since longer travel times force workers to depart
earlier if they are to arrive at work on-time (assuming that work hour flexibility is low). Although the
mode specific travel time coefficients are all of the same order of magnitude, an interesting result is that
the auto travel time parameter has the most negative value of all modes for all occupation groups. The
transit travel time parameter is consistently two times less negative than the auto travel time parameter
across occupations, which implies that transit users depart from home later than auto users, even if their
travel times to work are identical. This is a counter intuitive result. However, it should be recognized
that this interpretation is dependent on the fact that all the other covariates in the model completely
specify one’s arrival time at work, with the travel time simply modifying this arrival time to obtain one’s
H-W departure time. Indeed, it is highly likely that the travel time parameters are not capturing the
effects of travel time on departure time alone. Further complicating this interpretation is the fact that the
subtractive effect of the travel times on one’s eventual H-W departure time is non-linear (all covariates
lie in an exponential function).
Lastly, it should be noted that most covariates in the hazard model have a negative effect, causing
individuals to depart earlier relative to the “base” start time for each occupation group, which is
calculated by taking the exponential of the estimated constant for each occupation. All covariates and
the normal error term modify this “base” start time by increasing or decreasing the value of the
constant’s value in the exponential. The lack of positively forcing covariates, beyond the randomness of
the normal error term, is primarily due to the fact that detailed knowledge of official workplace hours
and degrees of work hour flexibility are unavailable at the individual level in the TTS dataset. Without
such knowledge, it is difficult to identify why one would have the option of departing later or even
earlier relative to the typical departure distribution for the person’s occupation. As previously discussed,
such detailed knowledge is usually only collected in detailed activity diary surveys, which by their very
nature have very small sample sizes. It is difficult to get such detailed knowledge in form of a large
regional travel demand survey such as TTS. However, the advantages of using TTS are plainly evident
from the large number of statistically significant, policy sensitive variables obtained in the parameter
estimation results of this section. Furthermore, even with more detailed data in hand, it must be
remembered that hazard models are not utility-based and do not consider departure time “alternatives”;
covariates can only take on one sign and can therefore only force timing in one direction. For example,
under flexible work hours it is unclear whether an individual would shift their departure time earlier or
later in an effort to avoid peak period congestion charges.
MODEL SENSITIVITY ANALYSIS

For the purposes of the sensitivity analysis, observed H-W departure times were used to retrieve
appropriate EMME/2 assigned auto travel times (by hour of departure) for each individual. This was an
appropriate assumption since detailed forecasts were not desired and the impacts of peak spreading on
overall daily trends were not expected to be major for this paper. These auto times were used in
combination with the other non-time varying variables as inputs to the estimated joint H-W departure
time model presented in the previous section. To test the impacts of increasing congestion and auto
travel times during peak periods on mode and departure time choice, five scenarios were tested,
increasing the observed auto travel times by 10% increments from 10% to 50% in the am and pm peak
periods, 6:00-9:00am and 3:30-6:30pm respectively. It should be noted that the effects of the increasing
auto travel times on Park and Ride auto access times were also accounted for in addition to the auto
drive and passenger modes.
As expected, the sensitivity analysis demonstrated that the model is capable of capturing the
decrease in utility of the auto mode as its travel time increases during peak periods. The daily auto mode
share continually decreases from 80% in the base scenario to 77% in the scenario where peak period
auto travel times were increased by 50%. In reality, the magnitude of the change in mode share as a
result of increasing congestion would be moderated by the travel time savings that would occur on the
auto network as auto users switch modes. This would be captured by model iterations with inputs of new
trip timing trends by mode into EMME/2. This is unnecessary in this sensitivity test since the travel time
increase is not a result of increasing trip making; it is fixed at given level for each scenario.
In the case of investigating peak spreading, moving from the base case to the 50% increase in
auto travel times, the auto PHF (ratio of peak hour to peak period trips) decreases from 35.7% to 34.6%.
As such, the model captures peak spreading effects of increasing auto travel times during peak period,
shifting travel outside of the peak hour. Although this appears to be a very modest decrease, one should
consider the fact that a PHF value of one third corresponds to the most highly spread situation; the base
case situation is already a very spread situation in the GTA. In order to fully understand how the model
is capturing the peak spreading phenomenon, it is important to investigate what time periods the former
peak period auto trips have shifted to. Specifically, it is illustrative to analyze the am shoulder peak
hours on the early and late sides of the am peak period, 5:00-6:00am and 9:00-10:00am respectively.
Through this analysis it was found that auto users travelling during the am peak period in the base case
scenario were being shifted to the early am shoulder peak hour; its proportion of daily auto trips steadily
increases as the auto travel times are increased during the am peak period from 19.0% in the base case to
22.3% in the +50% scenario. This is to be expected, since as auto travel times increase in the am peak
hour, auto users must depart earlier if they are to arrive at work on-time (assuming no work hour
flexibility). Therefore, the model is capable of capturing the reactive dimension of peak spreading. The
proportion of daily auto trips in the late shoulder peak hour predicted by the model however remains
relatively constant, decreasing from 24.9% in the base case to 24.2% in the +50% scenario. Although
undesired, this result should be expected based on the model parameter estimates presented in previous
section; very few covariates have a positive forcing on H-W departure time and auto travel time itself
had a negative effect. It should be noted that the slight decrease in the proportion of auto trips in the late
am shoulder peak hour is not due to the temporal shift of departures. It is a result of a slight decrease in
the auto mode share during this period for workers that depart from work during the pm peak period (the
auto travel time increase was applied to both the am and pm peak periods). In order to capture conscious
peak spreading to later periods, it is likely that more detailed survey data would be required, with work
hours and work flexibility policies defined in addition to one’s actual arrival time at work. Such data
could be easily collected within a survey such as TTS since activity diaries are not required.
MODEL APPLICATION
Although not conceptually difficult, a full application of the joint models estimated in this paper requires
the use of an iterative procedure since key model inputs (i.e. travel times) are dependent on one’s
departure time, which is the output of the model’s continuous trip timing component. As such, a full
forecasting application of the joint trip timing and mode choice models, to analyze policies and/or
changing conditions in both the base case and future year scenarios, was not presented herein. However,
the basic method that one would use to apply the home-work departure time and mode choice model is
sketched out in this section.
During model estimation, auto travel times for each individual were computed by hour of
departure for the entire 24 hour day through the assignment of observed trips to the network in an
EMME/2 traffic assignment model. Each individual’s observed departure time period was used to select
an appropriate auto travel time from the twenty four individual assignment results. This was appropriate
during model estimation since each person only experiences one “real” travel time, based on their actual
departure time, regardless of the model’s eventual prediction of their departure time. During model
application however, especially in the forecasting context, the knowledge of one’s departure time period
is unknown before the model is applied. Therefore, when applying the model, an assumed or
representative starting value is required for each individual’s one-hour departure time period to generate
appropriate auto travel time inputs. An iterative application procedure is required, where each
individual’s assumed one-hour departure period is continually updated by the model’s predicted time
period until they match. This process ensures consistency between the departure time periods used to
generate auto travel times and the model’s predicted departure times for each individual. The iterative
process is outlined below in the following five steps and summarized in Figure 1.
• Step 0: Randomly draw initially assumed Home-Work departure time (T0) for each individual
from observed base year distributions cross-classified by person, household, and scheduling
attributes (or any other relevant attributes).
• Step 1: Obtain auto travel times (AIVTT) from traffic assignment model (using observed base-
year trips) using the one-hour period that each individual’s assumed departure time (T0) lies in.
• Step 2: Use joint model to microsimulate predicted departure time (T*) and mode choice (M*) for
each individual using the appropriate auto times (AIVTT) and other non-time varying variables.
• Step 3: If the assumed and predicted departure times T0 and T*do not lie within the same one-
hour period for any given individual, readjust that person’s assumed departure time by letting
T0=T*. For individuals where modelled and assumed departure time periods are identical, T* and
M* are the final predictions for that individual and no readjustment of T0 is necessary. Return to
Step 1 to re-calculate auto travel times for individuals with newly assumed departure times (T0).
Once every individual’s assumed and predicted departure time periods are identical and thus
each individual has a final prediction, proceed to Step 4.
• Step 4: The final predicted mode and departure time choice for each individual have to be
inputted into the traffic assignment model to update travel times from observed base year
conditions. Since, the AIVTT previously obtained in Step 1 was calculated from the assignment
of observed base year trips to the network, new auto travel times (AIVTT) by hour must be
calculated to reflect the joint model’s newly predicted mode and departure time choices1. Go to
Step 2 with the newly corrected travel times AIVTT. As a result of the changes in AIVTT,
individuals may change predicted departure time periods, causing departure time periods to once
again be re-adjusted as above until assumed and predicted departure time periods are the same.
Eventually this process would equilibrate with no individual switching time periods as a result of
updating level of service variables from new traffic assignments.
In the method outlined above, it should be recognized that the first step of randomly drawing
assumed departure times is identical to the process used to generate the skeletal components of an
individual’s activity schedule in activity based models such as TASHA. In this method, however, the
initially predicted departure time from observed distributions are updated by the model, improving this
initial random guess. Although one could potentially use any starting value for the initial departure time
periods, this method is most likely to converge in the least number of iterations as most policies and
future year scenarios will at the very least be similar to base year conditions.
CONCLUSIONS AND FUTURE WORK

Trip timing and mode choice decisions are two of the most fundamental decisions made by commuters.
It is critical that transportation demand models are capable of modelling these two key choice
dimensions since they directly influence the temporal distribution of travel demand experienced on
every piece of transportation infrastructure in an urban area. From a behavioural point of view,
individuals can be expected to choose their departure times based on expected travel times for available
modes of transport and their preferred/required activity arrival times. Furthermore, one’s anticipated
time of travel directly impacts the level of service and feasibility of each available mode. Thus, both the
temporal and modal decisions are interrelated, requiring commuters to jointly consider both choice
dimensions together when scheduling trips.
As a result of these behavioural considerations, this paper departed from the traditional approach
of modelling modal and temporal decisions independently and postulated that mode choice and work
trip timing decisions are inherently correlated and must be modelled as a joint decision. A joint discrete-
continuous model specification, with a continuous hazard model for H-W trip timing and a multinomial
logit model for mode choice, was used to capture the interconnected nature of mode choice and trip
timing decisions. The estimated model has robustly demonstrated that a significant proportion of factors
influencing these two decisions are unobserved and highly correlated. Moreover, the estimated mode
specific variances for the hazard component of the departure model revealed considerable differences in
the spread of home-work departures between modes. In addition, the joint model has a wide range of
policy sensitive variables that are statistically significant by a wide margin and of sensible sign and
magnitude. From model sensitivity tests, it has been demonstrated the joint model is capable of
capturing peak spreading as a result of increasing congestion during peak period in the GTA.
Commuters were found to shift their departures to earlier shoulder peak periods in response to
increasing travel times in the am and pm peak periods. Such behaviour is to be expected in cases where
work hour flexibility is minimal; individuals must shift their departure times earlier in response to the
congestion if they are to remain on-time for their arrival at work.
The main limitation of the model presented in this paper is the fact that beyond the consideration
of full-time versus part-time work there are no workplace-related covariates to explain departure/arrival
1
This step is especially important if the policies or scenarios under consideration are expected to seriously alter the
temporal and/or modal distribution of travel in comparison to base year observed conditions. If one is simply fitting the
model to base year conditions or slight variations thereof, Step 4 is not necessary.
time differences and different levels of work hour flexibility among individuals employed within the
same occupation group. To adequately capture such trends, knowledge of workplace policies, including
official workplace start times and degrees of arrival flexibility, is required at the individual level.
However, such detailed scheduling related information typically requires small purpose-built activity
diary surveys since this type information is not available in the TTS survey and in most other large
regional surveys of similar size and scope. When designing future surveys to support models concerned
with activity scheduling, a balance must be struck between the richness of the activity information
solicited and the required dataset size for statistically robust parameter estimates. Ideally, a survey
similar in scope and size to TTS would be expanded to include only key work activity information such
as the required/preferred arrival time and the level of flexibility.
As an extension of this work, future research includes developing a trivariate joint model that
considers the full correlation between trip timing, work duration, and mode choice. The trivariate joint
model will be able to capture the complete daily cycle of commuters’ trip making behaviour.
REFERENCES
Aptech Systems (2006), GAUSS User’s Manual. Maple Valley, CA.
Arentze, T.A., and Timmermans, H.J.P. (2004) “A learning-based transportation oriented simulation
system,” Transportation Research B 38, 613-633.
Barnard, P.O. and Hensher, D.A. (1992), "Joint Estimation of a Polychotomous Discrete-Continuous
Choice System: An Analysis of the Spatial Distribution of Retail Expenditures", Journal of
Transport Economics and Policy, XXVI (3), 299-312.
Bates, J. Polak, J.W. (1997) “Time Period Choice Modelling: A Preliminary Review”, Project Report,
Centre for Transport Studies, Department of Civil and Environmental Engineering, Imperial
College, London, UK
Bhat, C. (1998) “Analysis of Travel Mode and Departure Time Choice for Urban Shopping Trips,”
Transportation Research B 32, 361-371.
Ben-Akiva, M., Lerman, S. (1985) “Discrete Choice Analysis: Theory and Application to Travel
Demand” MIT Press
Bhat, C. R. (1996) “A Generalized Multiple Durations Proportional Hazard Model with an Application
to Activity Behavior during the Evening Work-to-Home Commute.” Transportation Research B
30, 465-480.
Bhat, C., Steed, J. (2002) “A Continuous-Time Model of Departure Time Choice for Urban Shopping
Trips.” Transportation Research B 36, 207-224.
Bowman, J.L., and Ben-Akiva, M.E. (2000) “Activity-Based Disaggregate Travel Demand Model
System with Activity Schedules,” Transportation Research A: Policy and Practice 35, 1-28.
Data Management Group (DMG) (2005). Transportation Tomorrow Survey. Joint Program in
Transportation, University of Toronto. http://www.jpint.utoronto.ca/dmg/tts.html [accessed 15
October 2008]
Davidson, W., Donnelly, R., Vovsha, P., Freedman, J., Ruegg, S., Hicks, J., Castiglione, J., Picado, R.
(2007) “Synthesis of first practices and operational research approaches in activity-based travel
demand modeling”. Transportation Research Part A: Policy and Practice 41, 464–488.
deJong, G., Daly, A., Pieters, M., Vellay, C., Bradley, M., Hofman, F. (2003) “A Model for Time of
Day and Mode Choice Using Error Components Logit.” Transportation Research Part E, 39.
Ettema, D., Timmermans, H. (2003) “Modeling Departure Time Choice in the Context of Activity
Scheduling Behavior,” Transportation Research Record 1831, 39-46
Ettema, D., Bastin, F., Polak, J., ,Ashiru, O. (2007) “Modelling the joint choice of activity timing and
duration” Transportation Research A: Policy and Practice 41, 827-841.
Hensher, D.A., and Ton, T. (2002) TRESIS: A transportation, land use and environmental strategy
impact simulator for urban areas, Transportation, 29(4), pp 439-457
Habib, K.N., J.A. Carrasco, and E.J. Miller (2008), "Social context of activity scheduling: Discrete-
continuous model of relationship between ‘with whom’ and episode start time and duration,"
Transportation Research Record, 2076, 81-87.
Hensher, D. A., and Mannering, F. L. (1994) “Hazard-Based Duration Models and their Application to
Transport Analysis.” Transport Reviews 14, 63-82
Hyman, G. (1997) “The Development of Operational Models for Time Period Choice.” Department of
Environment, Transport and Regions, HETA Division, London, UK.
Hess, S., Daly, A., Rohr, C., Hyman, G. (2007) “On the development of time period and mode choice
models for use in large scale modelling forecasting system” Transportation Research A: Policy
and Practice 41, 802-826.
Johnson, N.L., Kotz, S., Balakrisnan, N. (1995). “Continuous Univariate Distributions, Volume 2.”,
Willey Inter. Science
Kiefer, N. M. (1988) “Economic Duration Data and Hazard Functions.” Journal of Economic Literature
26, 646-679.
Krsihnamurthi, L., Raj, S.P. (1988) “A Model of Brand Choice and Purchase Quantity Price
Sensitivity.” Marketing Science (7), 1-20.
Lee, B. and Timmermans, H. J.P. (2007) “A Latent Class Accelerated Hazard Model of Activity
Episode Durations.” Transportation Research Part B 41, 426-447.
Lee, L-F. (1983) “Notes and Comments Generalized Econometric Models with Selectivity.”
Econometrica 51, 507-512.
Miller, E.J. (2007) “A Travel Demand Modelling System for the Greater Toronto Area, Version 3.0:
Volume II GTAMODEL Version 3.0 Model System Documentation,” Joint Program in
Transportation, University of Toronto, Ontario, Canada.
Muizaga, M. A., Jara-Diaz, R.C.S.R, Ortuzar, J. (2006) “Valuing Time with a Joint Mode Choice-
Activity Model.” International Journal of Transport Economics 33(2), June.
Ozbay, K., Yanmaz-Tuzel, O. (2008) “Valuation of travel time and departure time choice in the
presence of time-of-day pricing”. Transportation Research Part A: Policy and Practice 42, 577–
590.
Palma, A., Marchal, F. (2002) “Real Case Application of Fully Dynamic METROPOLIS Tool-Box: An
Advocacy for Large Scale Mesoscopic Transportation System.” Network and Spatial
Econometrics 2, 347-369
Pendyala, R., Bhat, C.R. (2004) “An Exploration of the Relationship between Timing and Duration of
Maintenance Activities” Transportation 31, 429-456
Pendyala, R.M., Kitamura, R., Kikuchi, A., Yamamoto, S., Fuji, S. (2005) “Florida Activity Mobility
Simulator: Overview and Preliminary Validation Results.” Transportation Research Record
1921, 123-130.
Roorda, M.J., Miller, E.J., Habib, K.M.N (2008) “Validation of TASHA: A 24-h Activity Scheduling
Microsimulation Model.” Transportation Research Part A: Policy and Practice 42, 360-375.
Russo, F., Vitetta, A., Comi, A. (2009) “Estimation of Target Time Distribution for Agri Food Products
by Road Transportation” In Schedule-Based Modeling of Transportation Network, Chapter 14,
267-283, publisher Springer US.
Small, K.A. (1982) “The Scheduling of Consumer Activities: Work Trips.” American Economic Review
72 (June), 467-479.
Small, KA (1987) “A Discrete Model for Ordered Alternatives,” Econometrica, No. 55(2), 409-424.
Southern California Association of Governments (SCAG) (2003) “Year 2000 Model Validation &
Summary: Regional Transportation Model.” http://www.scag.ca.gov/modeling/index.htm
Accessed in October, 2008
Susilo, Y.O., Kitamura, R. (2007) “Structural changes in commuters’ daily travel: the case of auto and
transit commuters in the Osaka metropolitan area of Japan, 1980–2000”. Transportation
Research Part A: Policy and Practice 42, 95–115.
Train, K. (2003) Discrete Choice Methods with Simulation. Cambridge University Press, New York,
NY. <http://elsa.berkeley.edu/books/choice2.html>
van Vuren, T., Carmichael, S., Polak, J., Hyman, G., Cross, S. (1999) “Modelling Peak Spreading in
Continuous Time” in Proceedings of European Transport Conference-PTRC, Cambridge, UK
Vovsha, P., and Bradley, M. (2004) “Hybrid Discrete Choice Departure-Time and Duration Model for
Scheduling Travel Tours,” Transportation Research Record, No. 1894, p. 46-56.
Vickery, W.S. (1969) “Congestion Theory and Transport Investment.” American Economic Review 59,
251-260
TABLE 1 Mode Choice and Trip Timing Variable Definitions

Round
Variable Definition
Trip
WDuration n/a Work duration in hours
nHWstops N # of intermediate stops on Home-Work Journey
nWHstops N # of intermediate stops on Work-Home Journey
TotalStops Y nHWstops + nWHstops
med_inc+ n/a Median household income of zone in thousands of dollars
Dwell_House n/a 1 if individual’s dwelling is a house, 0 otherwise
n_person n/a Number of individuals residing in household
Natural logarithm of urban density of household zone (residents plus jobs per
Ln(Hurb_dens) n/a
hectare of built area)
Natural logarithm of urban density of employment zone (residents plus jobs
Ln(Wurb_dens) n/a
per hectare of built area)
Full_Time n/a 1 if worker is employed full time, 0 otherwise
TwoPlusVeh n/a 1 if household has more than one vehicle, 0 otherwise
Ln(Age) n/a Natural logarithm of individual’s age
Male n/a 1 if individual is male, 0 otherwise
FreeParking 1 if individual has free parking at workplace in PD1 (downtown Toronto) or
n/a
(PD1&PD4) PD4 (North York Centre), 0 otherwise
Straightline centroid-to-centroid distance from household to employment
Distance N
zone in Km
AIVTT Y Round trip auto in-vehicle time in minutes
ACost Y AM peak period auto fuel cost in dollars
"All Day" average zonal parking cost in dollars, equal to zero if individual
Pkcst N
has free parking at workplace
Total Auto Cost Y ACost + Pkcst
Round trip transit in-vehicle travel time in minutes
TIVTT Y
(peak period)
TWAIT Y Round trip transit wait time in minutes (peak period)
TWALK Y Round trip transit walk time in minutes (peak period)
TCOST Y Round trip transit fare in dollars (peak period)
FARE/DISTANCE Y TCOST divided by Distance
AIVTTP&R Y Round trip Park & Ride auto access time to closest station in minutes
ACostP&R Y Round trip Park & Ride auto access fuel cost to closest station in dollars
HW_AIVTT N Home-Work auto in-vehicle travel time in hours
Total H-W local transit travel time in hours
Total Transit Time N
(TIVTT + TWAIT + TWALK)/(2*60)
Total Transit P&R Total H-W local transit P&R travel time in hours
N
Time (TAIVTT TIVTT + TWAIT + TWALK)/(2*60)
Total GO P&R Total H-W GO P&R travel time in hours
N
Time (TAIVTT TIVTT + TWAIT + TWALK)/(2*60)
+ Adapted from 2001 Census Canada data on incomes by census tract.
TABLE 2: Estimated Parameters of Joint Model

Professional Manufacturing General Office Retail
Variable Name Mode Para. t-Stat Para. t-Stat Para. t-Stat Para. t-Stat
Correlation Coefficients
ρAuto 1,2 -0.02 -8.00 0.02 5.78 -0.04 -8.48 -0.05 -11.52
ρTransit 3 -0.03 -2.89 0.06 7.60 0.10 11.14 0.01 1.70
ρTransitP&R 4 0.10 5.63 0.30 5.22 0.21 5.77 0.30 7.58
ρGOP&R 5 0.12 9.05 -0.13 -2.10 0.09 3.46 0.28 7.04
ρWalk 6 -0.09 -7.33 0.11 6.34 -0.21 -7.70 -0.01 -0.80
Variances for Hazard Component
σAuto 1,2 0.19 1152.4 0.29 813.10 0.20 577.89 0.27 680.05
σTransit 3 0.16 455.3 0.26 284.68 0.16 299.73 0.26 325.79
σTransitP&R 4 0.11 139.6 0.16 32.78 0.11 76.34 0.15 46.92
σGOP&R 5 0.10 214.3 0.11 44.21 0.10 107.65 0.11 53.61
σWalk 6 0.17 225.9 0.27 124.54 0.20 111.97 0.29 177.30
Goodness of Fit Measures
Mean Loglikelihood
-8.423 -8.332 -8.425 -8.492
(Null Model)a
Mean Loglikelihood
-6.233 -6.613 -6.395 -6.801
(Convergence)
Adjusted Rho-Square 0.260 0.206 0.241 0.199
Logit Mode Choice Model Components
Drive_Const 1 -1.17 -19.32 -3.80 -45.47 -2.03 -21.93 -3.68 -49.89
Transit_Const 3 1.97 68.23 0.78 16.41 2.26 48.36 1.56 34.01
TransitP&R_Const 4 -0.24 -3.10 -1.58 -8.63 -0.64 -6.51 -1.10 -11.12
GOP&R_Const 5 -1.06 -7.36 -1.10 -2.11 -0.92 -3.74 -1.00 -2.77
Walk_Const 6 0.52 4.52 0.18 1.04 -0.04 -0.15 0.03 0.24
Full_TimeDrive 1 0.40 24.59 0.16 7.17 0.26 13.54 0.41 27.86
Full_TimeTransitP&R 4 1.25 16.76 0.14 0.79 1.22 12.96 0.75 8.36
Full_TimeGOP&R 5 0.95 15.25 1.05 3.55 0.64 6.90 0.96 8.22
Full_TimeWalk 6 -0.24 -8.54 -0.27 -6.66 -0.16 -3.91 0.11 4.21
TotalStops 1 0.90 113.31 0.93 54.44 0.82 64.21 0.77 55.84
Total AutoCost 1,2 -0.05 -58.54 -0.06 -30.68 -0.05 -34.57 -0.06 -33.84
Fare/Distance 3,4,5 -0.43 -44.71 -0.43 -26.12 -0.45 -26.36 -0.55 -39.38
ACostP&R 4,5 -0.20 -22.97 -0.76 -11.81 -0.49 -16.25 -0.55 -10.96
AIVTT 1 -0.03 -119.23 -0.01 -19.01 -0.03 -56.48 -0.02 -26.54
AIVTTP&R 4,5 -0.02 -24.24 0.04 7.21 0.001 0.87 0.02 4.42
TIVTT 3,4,5 -0.01 -74.99 -0.01 -30.63 -0.01 -34.61 -0.01 -36.85
TWAIT 3,4,5 -0.15 -155.14 -0.068 -61.34 -0.12 -90.92 -0.07 -62.19
TWALK 3,4,5 -0.03 -88.49 -0.02 -39.24 -0.03 -54.62 -0.02 -48.25
Distance 6 -0.841 -150.70 -0.794 -80.42 - -78.63 - -96.04
1.1496 0.8943
TwoPlusVehDrive 1 0.75 62.53 0.69 47.80 0.68 36.26 0.58 35.69
TwoPlusVehTransit 3 -1.06 -72.69 -1.13 -61.09 -1.12 -52.51 -1.25 -68.82
TwoPlusVehTransitP&R 4 0.32 13.16 0.40 5.04 0.39 9.40 0.68 10.71
TwoPlusVehGOP&R 5 0.25 11.01 0.45 5.77 0.12 3.06 0.17 2.86
TwoPlusVehWalk 6 -0.57 -25.86 -0.49 -15.24 -0.81 -19.85 -0.75 -28.40
ln(Age)Drive 1 0.84 57.67 1.25 61.54 0.95 43.30 1.42 75.85
ln(Age)GOP&R 5 0.67 19.20 0.28 2.33 0.56 9.22 0.61 6.35
ln(Age)Walk 6 -0.19 -6.61 -0.20 -4.48 -0.008 -0.12 0.14 4.00
MaleDrive 1 0.86 90.94 1.55 112.56 1.05 51.27 1.00 65.42
MaleTransit 3 0.39 33.98 0. 50 31.42 0.48 21.17 0.10 5.97
MaleWalk 6 0.95 51.68 1.10 36.52 0.95 22.97 0.63 25.68
-
med_incDrive 1 -0.001 -11.54 -0.0003 -1.84 -0.001 -5.07 -3.08
0.0004
med_incTransit 3 -0.002 -20.32 -0.002 -8.83 -0.003 -13.22 -0.003 -15.98
med_incWalk 6 -0.004 -19.62 -0.005 -11.52 -0.001 -2.82 -0.005 -18.06
ln(Hurb_dens)Drive 1 -0.04 -8.47 0.03 3.76 0.005 0.59 0.0006 0.07
ln(Wurb_dens)Transit 3 0.36 95.77 0.35 46.98 0.27 43.80 0.42 63.41
ln(Wurb_dens)Walk 6 0.61 91.23 0.49 36.13 0.61 44.69 0.42 38.33
Value of Travel Time
1,2 $43.14 $12.99 $32.32 $16.10
Saving ($/hr)b
Continuous Departure Time Model Component
Constant n/a 6.64 2137.98 6.80 961.63 6.84 1214.79 7.19 1169.76
Full_Time n/a -0.09 -107.75 -0.12 -62.58 -0.10 -87.39 -0.15 -113.87
Wduration n/a -0.02 -270.48 -0.03 -135.54 -0.03 -151.81 -0.03 -162.16
HW_aivtt 1,2 -0.16 -171.74 -0.17 -72.42 -0.18 -86.35 -0.23 -89.37
Total Transit Time 3 -0.09 -78.11 -0.09 -36.63 -0.10 -62.93 -0.11 -45.31
Total Transit P&R
4 -0.13 -32.61 -0.08 -2.73 -0.14 -16.28 -0.10 -5.90
Time
Total GO P&R Time 5 -0.12 -134.77 -0.17 -20.11 -0.14 -70.40 -0.16 -29.40
Distance 6 -0.009 -8.95 0.02 5.68 -0.03 -8.84 -0.01 -2.42
Dwell_House n/a -0.01 -21.24 -0.03 -26.19 -0.008 -8.59 -0.002 -1.71
n_person n/a 0.003 21.01 0.01 34.77 0.007 23.12 0.006 17.21
ln(age) n/a -0.05 -77.04 -0.09 -58.15 -0.09 -69.14 -0.17 -116.75
Gender=1 n/a -0.003 -7.99 -0.04 -34.79 0.02 20.89 0.02 22.83
nHWstops n/a -0.01 -21.55 -0.008 -5.45 -0.04 -33.46 -0.08 -54.50
ln(Hurb_dens) n/a 0.006 23.95 0.005 7.93 0.06 11.70 0.001 1.84
ln(Wurb_dens) n/a 0.005 31.98 0.01 30.68 0.003 8.86 0.009 20.28
Free Parking
n/a 0.01 11.97 0.002 0.62 0.01 5.53 0.02 7.38
(PD1&PD4)
“Base” Start Timec n/a 12:44 14:57 15:33 22:10
Mode Definitions: Auto Drive (1), Auto Passenger (2), Transit Walk (3), Transit P&R (4), GO
P&R (5), Walk/Bike (6)
a
Independent discrete and continuous models, with no parameters for mode choice (sample
share) and a single constant for the hazard model.
b
VOT is calculated from AIVTT and Total Auto Cost parameter estimates.
c
Calculated by: exp(Constant)/60

Figure 1: Joint H-W Departure Time Model Iterative Application Procedure
View publication stats

Investigation of commuting trip timing and mode choice in Greater Toronto Area

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Investigation of commuting trip timing and mode choice in Greater Toronto Area

Uploaded by

Copyright:

Available Formats

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

An investigation of commuting trip timing and mode choice in the Greater

Article in Transportation Research Part A Policy and Practice · August 2009

Khandker Nurul Habib Eric Miller

SEE PROFILE SEE PROFILE

Sustainable Systems View project

The user has requested enhancement of the downloaded file.

Khandker M. Nurul Habib, Ph.D., P.Eng.

Eric J. Miller, Ph.D.

MODELLING TRIP TIMING

Discrete Choice Models

Continuous Time Models

Joint Discrete-Continuous Models

Multinomial Logit (MNL) Mode Choice Model

Continuous Time Hazard Model

+ Dmi ln( Φ ( ( J1 (vmi ) − ρ mti ((ln(t mi ) − θZ mi ) / σ mti ))) / (1 − ρ mti ) ) )

DATA PREPARATION AND EXPLANATORY VARIABLE DEFINITIONS

MODEL ESTIMATION AND SPECIFICATION PROCESS

MODEL SENSITIVITY ANALYSIS

CONCLUSIONS AND FUTURE WORK

TABLE 1 Mode Choice and Trip Timing Variable Definitions

TABLE 2: Estimated Parameters of Joint Model

View publication stats

You might also like