- The document discusses how increasing numbers of tolled roads in a network like Sydney could lead to "toll saturation", where commuters reach a limit on how much they are willing to pay in tolls. - It presents models to test the hypothesis that as toll costs accumulate across multiple roads, the value that commuters place on travel time savings (VTTS) will decrease due to this toll budget constraint. - The study uses data from Sydney, which has 9 existing tolled roads and 5 more planned, to estimate models that incorporate a toll saturation effect and examine how it impacts VTTS for commuters.

1. How much is too much for tolled road users: Toll saturation and
the implications for car commuting value of travel time
savings?
David A. Hensher ⇑
, Chinh Q. Ho, Wen Liu
Institute of Transport and Logistics Studies, The University of Sydney Business School, The University of Sydney, NSW 2006, Australia
a r t i c l e i n f o
Article history:
Received 10 November 2015
Received in revised form 15 August 2016
Accepted 16 October 2016
Available online 5 November 2016
Keywords:
Toll saturation
Budget constraint
Multiple toll roads
Value of travel time savings
New empirical evidence
a b s t r a c t
The current practice of forecasting the demand for new tolled roads typically assumes that
car users are prepared to pay a higher toll for a shorter journey, and they will keep doing so
as long as the toll cost is not higher than their current value of travel time savings. Practice
ignores the possibility that there could be a point when motorists stop driving on toll roads
due to a toll budget constraint. The unconstrained toll budget assumption may be valid in
networks where the addition of a new toll road does not result in a binding budget con-
straint that car users may have for using toll roads (although it could also be invoked for
existing tolled routes through a reduction in use of a tolled route). In a road network like
Sydney which offers a growing number of (linked) tolled roads, the binding budget con-
straint may be invoked, and hence including additional toll links might in turn reduce
the car users’ willingness to pay for toll roads to save the same amount of travel time.
When this occurs, car users are said to reach a toll saturation point (or threshold) and begin
to consider avoiding one or more toll roads. Whilst toll saturation has important implica-
tions for demand forecasting and planning of toll roads, this type of behaviour has not been
explored in the literature. We investigate the influence that increasing toll outlays has on
preferences of car commuters to use one or more tolled roads as the number of tolled roads
increases. The Sydney metropolitan area offers a unique laboratory to test this phe-
nomenon, with nine tolled roads currently in place and another five in planning. The evi-
dence supports the hypothesis that the value of travel time savings decreases as a
consequence of toll saturation.
Ó 2016 Elsevier Ltd. All rights reserved.
1. Introduction
It is often suggested that when a new toll road is introduced into a network that already has a number of tolled roads, the
accumulation of daily or weekly tolls should be taken into account in determining the probability of a traveller using one or
more tolled facilities. There may, however, be a limit on how much individuals are willing to outlay to save travel time, given
personal budgets for specific expenditures and competing demands on their income. Although traffic assignment methods
have the ability to take into account the accumulating toll outlay through a generalised cost or time expression, the binding
toll budget constraint, which is likely to impact on how much an individual is willing to pay as more tolled options arise, is
typically not taken into account and reflected in the value of travel time savings (VTTS) applicable to the accumulating set of
http://dx.doi.org/10.1016/j.tra.2016.10.012
0965-8564/Ó 2016 Elsevier Ltd. All rights reserved.
⇑ Corresponding author.
E-mail addresses: David.Hensher@sydney.edu.au (D.A. Hensher), Chinh.Ho@sydney.edu.au (C.Q. Ho), Wen.Liu@sydney.edu.au (W. Liu).
Transportation Research Part A 94 (2016) 604–621
Contents lists available at ScienceDirect
Transportation Research Part A
journal homepage: www.elsevier.com/locate/tra

2. tolled situations. Simply applying the same VTTS to all tolled links may be questionable in the presence of a toll budget
threshold.
Whilst this position has intuitive practical merit, there are few jurisdictions where the growing number of tolled links
may result in consideration of potential toll budget constraints on travel preferences. The Sydney metropolitan area is
one example of a real world laboratory where we currently have nine tolled roads, totalling 135 one-way kilometres, of
which two have more than one fixed tolled entry or exit location, and one that is distance-based with a financial cap at
20 km for a single continuous trip. There are five more tolled links currently being planned to be in place over the next eight
years, an additional total of nearly 50 km, with a mix of fixed and distance-based charging regimes. Sydney will then have
over 185 one-way (multi-lane) kilometres of tolled routes.1
Depending on the geographical spread of travel activity, it is not
uncommon for Sydney residents to spend between $2000 and $5000 per annum on tolls (often exceeding outlays on fuel) for the
journey to and from work. This is especially evident for travel between the north-west/west and the central business district
(CBD), as shown in the empirical setting section below.
If there is a real possibility that the growth in toll road capacity will start to bite into travel budgets, we can no longer
assume that there is an open-ended commitment to toll routes used for a given level of travel time savings. There should
be a correction of the VTTS to reflect this budget threshold, which is likely to vary across the population of travellers. Such
an upper limit is referred to as the toll saturation effect.2
The idea of saturation effects linked to income is not new in general in
demand studies; however it has not been investigated, as far as we are aware, as a phenomenon that needs to be included in
potential adjustments in VTTS as increased toll opportunities are introduced to a road network. We have been unable to find any
explicit assessment of the role of a ceiling figure on the amount paid on tolls.
If the VTTS is significantly revised downwards after accounting for toll saturation (i.e., toll saturation effect results in a
downward prediction of toll route patronage), we may have identified one basis for explaining the errors in forecasts of toll
road patronage (see Bain, 2009, 2011; Flyvbjerg et al., 2006; TRB, 2006; Li and Hensher, 2010; Welde and Odeck, 2011).3
For
example, The Clem7 toll road forecasts in Brisbane have been shown to be affected by four key factors, with the VTTS, labelled as
the willingness to pay in Fig. 1, accounting for nearly 30% of the forecast outcome. Thus, it is reasonable to assume that the VTTS
estimate plays a major role in the determination of patronage forecasts for tolled roads. A lower VTTS will inevitably bring the
forecast and actual patronage estimates closer together.
The paper is organised as follows. We begin with the presentation of a model that can explicitly test for the potential role
of a toll saturation effect on the VTTS. We then set out the design of a choice experiment to capture the data necessary to
establish preferences for travel as mixtures of tolled and non-tolled links under varying budget thresholds, followed by a
summary of the empirical setting and sampling strategy required to ensure relevant exposure to the existing tolled network
and potential access to the new tolled routes. We then estimate a nonlinear logit model to obtain parameter estimates asso-
ciated with time-cost trade-offs under toll budget constraints, and establish the role that toll saturation plays in the VTTS.
The paper concludes with the important implications the evidence has on the demand for toll road travel and project
appraisal.
2. Modelling approach
Consider a spatial setting in which an individual has a series of tolled routes available to travel to and from work. Each
tolled route has an associated cost with the total toll costs on route j 2 J defined by Eq. (1).
Tollcostj ¼
XL
l¼1
dl  Tolll; dl ¼ 1 if route j involves tolled link l; 0 otherwise ð1Þ
Taking a two-week period of commuting on route j (2wkij) as the travel context, we define a personal budget that an indi-
vidual i has allocated for toll roads for commuting purpose, as budgeti. The individual’s preference for a particular route con-
figuration (as a mix of tolled and non-tolled links) can be defined initially as a linear additive utility expression of the form
given in Eq. (2), where we include the toll outlay in the last two weeks of commuting, the budget threshold, trip time, socioe-
conomic influences, and other unobserved effects (eij).
Uði; jÞ ¼ a þ bsat
Toll2wkij
budgeti
þ btimeTimej þ bsocio  Socioi þ eij
¼ a þ bsatpsati þ btimeTimej þ bsocio  Socioi þ eij
ð2Þ
1
See http://www.smh.com.au/nsw/how-much-is-too-much-for-sydney-toll-roads-20150410-1mif9l.html.
2
Rob Bain refers to this as the affordability overlay (personal communication 9 November 2015).
3
A major global bank has told us (verbally) that they discount patronage by 60% when selling the project to potential private equity investors. This is a
startling response but one which aligns well with the growing evidence globally of the high forecasts that typically equate to over 50% error at least in the first
2–10 years of a toll road project’s life.
D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621 605

3. The toll saturation rate is defined as the total toll costs over a two week period relative to an individual’s total toll budget
for the same period; hence psati is the current level of toll saturation of commuter i, who has a 2-week budget of budgeti to
spend on toll roads, and has spent a total of Toll2wkij ¼ Workdays  2  Toll
oneway
j on commuting in the last 2 weeks, with the
number of days travelling to and from work in the last two weeks being Workdays.
The marginal (dis)utility of travel time and the marginal (dis)utility of trip cost given the toll budget threshold are given
in Eq. (3), and the implied value of travel time savings ($/person per minute or hour) as Eq. (4), noting that VTTS is condi-
tioned on the toll budget assigned.
MUtime ¼ btime; MUcost ¼
bsat  2  Workdays
budgeti
ð3Þ
VTTS ¼
MUtime
MUcost
¼
btime
bsatð2 Â WorkdaysÞ
 budgeti ð4Þ
As the personal budget of an individual worker for toll road usage increases, VTTS is expected to increase as shown in Eq.
(4). However, each person will have an initial budget which is likely to remain the same for a given level of time savings, and
this renders the functional form given in (2) uninformative as the only possible way to explore the variation of personal VTTS
as shown in Eq. (4) through varying a personal toll budget.
Given that our objective is to investigate the effect of toll saturation on VTTS, we need to structure the functional form of
the preference expression in such a way to be able to identify how VTTS might change (i.e., decrease) as we approach and go
beyond the toll budget threshold that represents what an individual is prepared to spend on tolls.
In summary, we want to know what happens when a new toll road (and/or an increase in tolls on existing tolled roads)
triggers the personal toll outlay budget constraint for each commuter. Do car users adjust their personal budget to use toll
roads (when they see the value of the time savings) or does the personal budget come into play leading to avoidance of one
or more toll roads in the short term as they decide that they can no longer afford the toll costs? The answers to these ques-
tions can be revealed by investigating, in a choice experiment, the effect of increasing toll costs on the share of routes with a
varying number of tolled links (e.g., current route, current plus 1, current plus 2, or free road).
Eq. (4) is therefore too restrictive since it does not enable us to obtain the relationship between VTTS and toll saturation.
An alternative preference expression is one that recognises the impacts of toll budget constraint on the trade-off between
travel time and toll outlay, and hence it should condition the entire observed component of the utility expression. One way
of showing this is to modify Eq. (2) as Eq. (5) or Eq. (6), which are distinguished by the impact of a toll budget threshold.
That is, whether the toll saturation effect is associated with the accumulated experience over a fixed period of time (Eq.
(6)) or it is associated with a specific trip (i.e., the one way trip travel time and toll outlay in Eq. (5)). We investigated both
possibilities,4
but Eq. (6) is behaviourally more meaningful in that it is the accumulating toll expenditure (TotTolls) over a period
of travel (TotTime) that results in consideration of the toll budget threshold. Note that in Eqs. (5) and (6), we add 1 in the con-
ditioning function to assist the interpretation of the toll saturation effect. That is, when the toll saturation effect is not signif-
icant (identified by bsat being not statistically different from zero), the budget constraint function receives the value of one, and
hence Eqs. (5) and (6) collapses to the standard utility expression (i.e., without the impact of budget constraint). Including 1 in
the conditioning function also helps solving the issue of some commuters having a zero level of toll saturation (i.e., toll non-
users and toll avoiders).
Uði; jÞ ¼ ð1 þ bsatpsatiÞ Â ðaj þ btollTollj þ btimeTimejÞ þ eij ð5Þ
0%
10%
20%
30%
40%
50%
Day/Year expansion Lack of growth Willingness to pay Network
Fig. 1. Factors affecting the CLEM7 (Brisbane) toll road forecasts. Source: The RiverCity Motorway (2010).
4
We have estimated Eq. (5), but do not report any findings even though the model had almost the same overall statistical fit, but a key attribute, btoll, was
only marginally significant. It is available on request. The VTTS in this model are summarised in footnote 11.
606 D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621

4. Uði; jÞ ¼ ð1 þ bsatpsatiÞ Â ðaj þ btollTotTollsj þ btimeTotTimejÞ þ eij ð6Þ
The logic behind the model specification is that the amount an individual is willing to pay for the benefits of a tolled route
(all explanatory variables) is conditioned on how much someone is prepared to outlay on tolls (the toll budget), and that the
way this can be incorporated into a utility expression is through treating the toll budget as information that is otherwise
contained in the random component or unobserved effects. Given that the toll budget is likely to vary across a sample,
we can relate it to the error variance. We use this interpretation to obtain the form used in model estimation.
This approach to incorporate toll saturation is analogous to the approach developed by Swait and Adamowicz (2001a,b) to
accommodate complexity as a constraint, in which the theoretical context is aligned with information theory in order to pro-
vide a measure of information content or uncertainty. Information theory refers to an approach taken to quantify the amount
of information contained in an experiment or phenomenon (e.g., Soofi, 1994). Toll saturation is a source of information quan-
tity. Analogously to Swait and Adamowicz, we assume that toll saturation affects the utilities only through the stochastic
component and assume that differences in toll saturation generate differential consistency levels in preferences across indi-
viduals, which will be reflected in the standard utility expression Vjq + ej by affecting the variances of the assumed distribu-
tion for the disturbances. As shown in Swait and Adamowicz (2001b), under the usual distributional assumptions associated
with logit model form, the conditioning expression is the scale function l(E), where l is inversely related to the variance of
the errors. Importantly, so long as the conditioning expression is a function of object attributes X, and decision maker char-
acteristics, the resulting model does not have the Independence of Irrelevant Alternatives property (unlike the standard MNL
model). This is referred to as the Heteroscedastic MNL model, similar to the idea presented in Hensher and Rose (2012) and
in Hensher et al. (1999) as a parametrized heteroscedastic MNL (PHMNL) model. Although there may be other ways of incor-
porating toll budgets and toll saturation, the approach proposed has great appeal in that it conditions all of the observable
sources of influence on the relative utility associated with each alternative.5
Thus, this is a way of recognising that each alter-
native is processed conditioned on the amount a sampled commuter is willing to outlay on tolled roads to gain time benefits.
More specifically, Eqs. (5) and (6) can be rewritten in a compressed form, similar to the one derived by Swait and
Adamowicz (2001b) and Hensher et al. (1999):
Uði; jÞ ¼ lðEijÞ Â ðVij þ eijÞ ð7aÞ
Uði; j
0
Þ ¼ lðEij0 Þ Â ðVij0 þ eij0 Þ ð7bÞ
where j0
is an alternative route to route j; l(Eij) is the conditioning function that conditions the standard utility expression
(Vij + eij) on the level of toll saturation associated with an alternative. This conditioning is a form of heteroscedasticity. Eij
recognises that an individual level of toll saturation, proxied by the ratio of the toll budget to toll outlay, conditions the mar-
ginal (dis)utility of each and every attribute, observed and unobserved, associated with the jth alternative in a pre-defined
choice set.
In Eqs. (7a) and (7b), the random variables l(Eij)eij are IID Gumbel but with unit scale factors. Thus, multiplying the stan-
dard utility by a non-negative6
lij, the probability expression remains unchanged, as shown in (8).
Pr½Uij P Uij0 Š ¼ Pr½Vij À Vij0 P eij À eij0 Š
¼ Pr½lijðVij À Vij0 Þ P lijðeij À eij0 ÞŠ
ð8Þ
Given the IID property of the error difference, it follows that the probability of choosing an alternative is an MNL-like
model with the observed sources of utility l(Eij)Vij as given in Eq. (9).
Prij ¼
exp½lðEijjkÞ Á VijðXijjbÞŠ
P
j0
2Ji
exp½lðEij0 jkÞ Á Vij0 ðXij0 jbÞŠ
ð9Þ
where k and b are parameters to be estimated; E and X are the observed variables associated with each alternative and each
individual.
With the conditioning utility function specified in Eqs. (5) and (6), VTTS is now non-linear, given that psati is a function of
the level of the toll cost outlay over a cumulative period of time; here it is two weeks (i.e., psati = Toll2wkj/budgeti). Using Eq.
(6), the marginal (dis)utility of total toll cost and total travel time can be expressed as Eqs. (10) and (11), with VTTS being
defined as a non-linear expression in Eq. (12).
MUtoll ¼
@U
@Toll
¼ ð1 þ bsatpsatiÞ Â ðbtoll  Workdays  2Þ þ bsat
Workdays  2
budgeti
 ðaj þ btollTotTollsj þ btimeTotTimejÞ ð10Þ
5
We investigated conditioning only some of the attributes in the linear component and that was found to be a statistically poor model and indeed one
inconsistent with the way that error variance should scale the entire utility expression.
6
There are many ways to specify the conditioning function such that its value is non-negative, one of which is to use an exponential form (see Hensher and
Ho, 2016); however, in this paper we let the empirical data speak for itself and we verify that the conditioning function is indeed positive for all sampled
individuals.
D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621 607

5. MUtime ¼
@U
@Time
¼ ð1 þ bsatpsatiÞ Â ðbtime  Workdays  2Þ ð11Þ
VTTS ¼
MUtime
MUcos t
¼
btimeð1 þ bsatpsatiÞ
btollð1 þ bsatpsatiÞ þ bsat
budgeti
ðaj þ btollTotTollsj þ btimeTotTimejÞ
ð12Þ
The interpretation of the effect of toll saturation, or the toll budget constraint, on the VTTS is conditioned on the sign of
the toll saturation parameter (bsat) and other parameters entering Eq. (12). As is in the standard utility function, travel time
and cost are both expected to have a negative parameter (i.e., btime 0, btoll 0), given that an increased travel time and travel
cost results in an increase in the marginal disutility. As all commuters have the non-negative level of toll saturation (i.e.,
psati P 0), when its associated parameter bsat is significantly positive, the conditioning function 1 + bsatpsati will have a value
larger than 1. This means that the toll budget constraint will scale up the standard disutility (see Eq. (6)). In addition, the
scaling effect of the toll budget constraint on the marginal disutility is expected to be greater for toll cost (Eq. (10)) than
for travel time (Eq. (11)), given that the VTTS for commuting is usually smaller than $60 per hour (i.e., btime btoll). Thus,
a positive parameter bsat suggests that the VTTS will reduce when people approach and go beyond their toll budget. Con-
versely, a negative parameter bsat suggests the reverse: the VTTS will increase as the level of toll saturation increases, whilst
a non-significance parameter bsat suggests that the VTTS is not influenced by the toll budget constraint (i.e., people will will-
ing to pay the same amount to save a unit of travel time as they currently do, regardless of how much they have already
outlaid on tolled roads). We hypothesise a positive parameter estimate for the toll budget constraint and test this hypothesis
with the empirical data (see footnote 11).
With these expected signs, Eq. (12) suggests that as the budget an individual prepared for the use of toll roads increases,
VTTS increases since the second term of the denominator becomes less negative (marginal cost approaches zero when the
toll budget approaches infinity). It also shows that as the proportion of toll cost to total toll budget, psati, increases (i.e., peo-
ple approach or go beyond their toll saturation point), VTTS decreases since the first term of the denominator becomes more
negative. Thus, the denominator of the VTTS function expressed in Eq. (12) has two parts: the first part reflecting the cumu-
lative effect of tolls, and the second part reflecting the current level of toll saturation on future route choice and hence, the
willingness to pay for higher toll costs to save more travel time. The VTTS as expressed in Eq. (9) will vary across the alter-
natives even if all parameters are specified as generic because the marginal disutility of cost depends on the level of accu-
mulated toll outlay (TotTollsj) and time (TotTimej) associated with each alternative. This approach provides an appropriate
formulation for identifying the effect of toll saturation on VTTS, using the concept of a demand curve as illustrated in Fig. 2.
Other variables can also be included in the conditioning function to recognise the residual heterogeneity effect after indi-
vidual toll budget constraint has been accounted for. Of particular interest is the impact of toll costs being reimbursed on
worker’s choice of tolled roads. This effect is explored by modifying the conditioning function to include a dummy variable,
EmpPay, indicating whether the cost of using tolls for commuting are reimbursed by the worker’s employer.7
Uði; jÞ ¼ ð1 þ bsatpsati þ bempEmpPayiÞ Â ðaj þ btollTotTollsj þ btimeTotTimejÞ þ eij ð13Þ
Commuters are expected to be less sensitive to toll costs if the cost of using tolls for commuting are reimbursed by the
worker’s employer, and hence bemp is expected to be negative such that its effect will offset the effect of the toll budget con-
straint (i.e., workers become less sensitive to toll costs when their employer pays). Other socio-economic variables such as
age and income can also be included, either in the conditioning function (the term in the first parenthesis of Eq. (13)) or in
the standard utility function (the term in the second parentheses of Eq. (13)). The theory for including socio-economic vari-
ables in the conditioning function is that commuters in different age and income groups may respond to travel time and
travel cost differently. Conversely, if the same variables are found to be better included in the standard utility function, this
suggests that workers with the same level of income and age group response to the addition of more tolled links differently,
depending on their current level of toll saturation. As both specifications are theoretically relevant, the decision of where to
include socio-economic variables is based on the statistical ground. That is, a specification that produces the highest model
fit to the empirical data is selected for the final model specification.
3. Empirical setting
The number of toll roads in Sydney is set to increase substantially over the next few years with a number of new toll roads
being added to a network which already has nine toll links and a total of 135 km of tollways. The existing toll roads are the
M2, M5, M7, Cross City Tunnel (CCT), Lane Cove Tunnel (LCT), the Military Road E-Ramp, the Sydney Harbour Bridge (SHB),
Sydney Harbour Tunnel (SHT) and the Eastern Distributor (ED). Fig. 2 shows the geographical location of these toll roads,
together with the ones that are to be constructed in Sydney.
Of the tollways under construction, the WestConnex project has been identified by the New South Wales (NSW) Govern-
ment as a key infrastructure project, which aims to ease congestion and facilitate growth of Sydney. The scheme is currently
the largest integrated transport and urban revitalisation project in Australia. The 33 km route will be built in three stages,
7
We considered separate psat parameters for whether the employer paid the tolls or not but this model produced singularity in the estimated variance
matrix of estimates.
608 D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621

6. with the final stage opened to traffic in 2023. Stage 1 will widen (2015–2017) and extend (2016–2019) the M4 motorway.
Stage 2 (2015–2019) will build new twin tunnels with more than double capacity along the M5 East, and Stage 3 (2019–
2024) will join the M4 and M5 with two new 9-km tunnels, each with three lanes. Upon completion, the WestConnex will
form a continuous high speed motorway to link between the west of Sydney, the Central Business District, Sydney Airport
and Port Botany (see Fig. 3). These are areas of significant importance to the Sydney and national economy.
Tolling has been identified as necessary to fund the WestConnex project, which is expected to deliver significant travel
time savings across Sydney for toll road users. However, the addition of this new tollway, together with the 9-km under-
construction NorthConnex tollway, means that the toll burden will spread further and become greater as two existing
motorways which are free to motorists – the M4 and the M5 East – will start to charge toll fees once the WestConnex
has added extra lane capacity. As tolls increase, the toll outlay may start to bite into personal budgets. With a large number
of tolled motorways, it is likely that a toll budget constraint will soon become a phenomenon in Sydney, if it has not already
happened. When this occurs, motorists are said to reach their toll saturation point and are expected to change (route) beha-
viour. They may adjust their personal budget to use toll roads (when they see the value of the time savings) or they may
consider avoiding one or more toll roads (when they are no longer willing to spend that much money saving travel time).
This phenomenon is investigated in this study using a stated preference (SP) experiment. The next section describes the
experiment in detail.
4. The experiment
The centrepiece of this project is an SP experiment designed to understand how an increase in toll cost influences a
motorist’s choice of route for commuting, and whether the impact of the same increase in toll on their choice of commuting
route will vary across different levels of toll saturation. Commuting is selected given that it is the travel segment where
motorists are more likely to pay to use tolled routes. To replicate the real world situation of more tolled roads to be added
to the Sydney’s toll network described in the ‘empirical setting’ section above, this study designs the SP experiment by add-
ing one or two more tolled links into the current commuting route offerings. Thus, information about the current commuting
pattern is firstly sought for the purpose of designing the SP experiment.
The survey instrument has five major parts. The first part asks respondents to report their commuting patterns over the
last two weeks with questions relating to their usual commuting mode, home postcode and workplace suburb, number of
days of travel to work, and whether their commuting route involved any toll road. Depending on the answers to these ques-
tions, the survey instrument determines whether the respondent is eligible to proceed to the second part of the survey,
which is designed to collect information on the journey to work (JTW) and the journey from work (JFW). In this second part,
respondents are asked to report their usual departure and arrival times, the number of times they used each of the nine toll
roads for commuting in the last two weeks, the approximate travel time and toll cost on each tolled link if they used them for
commuting, the typical, slowest and quickest commuting times on the current route, as well as their estimation of travel
time on an alternative route (i.e., free route for current toll road commuters and tolled roads for non-toll commuters). These
sources of information are useful for establishing the current level of toll saturation, and the travel time and toll cost on the
current route, which are used in the design of the SP experiment described below.
Market price
$0
$5
$10
$15
$20
$25
$30
$35
$40
$45
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 110%
Proportion of toll costs to toll budget
WTP ($/hour) for tolled roads in relation to toll budget
Fig. 2. The potential role of toll saturation on VTTS.
D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621 609

7. The third part of the survey asks respondents to consider a situation when a 2-week toll outlay increases to the point that
they would (i) seriously consider avoiding, and (ii) definitely stop using one or more toll roads for commuting. The former
value is considered as the lower bound and the latter the upper bound of a personal toll budget.8
Combining these budgets
Fig. 3. Sydney’s toll road network: existing and new tollways. Source: Saulwick (2015).
8
The levels reported are in a sense perceived estimates; however they have real meaning to each individual regardless of whether there may be some
amount of reporting error and are a best estimate of a likely toll budget. This is akin to the perception of levels of travel time associated with chosen and non-
chosen alternatives.
610 D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621

8. with the toll outlay in the last two weeks (collected in the second part), we can determine how far each commuter is away from
their saturation point, and hence when they are not prepared to spend more money on tolls. This information is used for the
assignment of the hypothetical scenarios in which motorists face their toll budget constraints when they trade-off travel time
and toll costs. Apart from the toll budget, this section also asks respondents questions relating to who actually pays the toll costs
for commuting and whether commuters receive any financial help from other people in case their toll budget is smaller than the
amount that they have outlaid on tolls.
The fourth part is the SP experiment which offers a total of four alternative routes for commuting trips. These alternatives
are the current route, the current route plus one more tolled link, the current route plus two more tolled links, and a free
route. Current toll road users face all four alternatives, whilst commuters who do not use toll roads for commuting (i.e.,
non-toll commuters) are limited to choosing amongst the first three alternatives (i.e., the free route is blocked out since
the current route is already a free route). Each alternative route is described by the number of tolled segments, the travel
time on the tolled and free segment, the total one way travel time and one way travel cost, as well as an accumulating toll
costs and time saved over a 2-week commuting period. Fig. 4 provides an illustrative choice screen for a respondent whose
current commuting route includes three toll segments and costs $10 per one way.
Sitting behind the SP experiment are 14 D-efficient designs (see Hensher et al., 2015) which were customised for each
respondent by pivoting the travel time and travel cost on the current route that respondents revealed in the second and third
parts of the survey. That is, the current route is used as the reference alternative, and as the survey covers an entire
metropolitan area, it is necessary to use many designs to cover different reference points and to make choice tasks mean-
ingful at the individual level. Attributes to pivot are the total travel time on tolled roads, the total travel time on free roads,
and the total toll costs. Table 1 shows the pivot levels for each attribute and the rules employed to assign scenarios to
respondents. The choice experiment was designed using NGene (Choice Metrics, 2012).
Priors for the SP designs were obtained from Rose and Hensher (2014). In addition, pivot designs require reference levels
(see second column of Table 1) which we selected to cover all commuting patterns, including medium to long commute with
large/small/no time component on toll roads (D1–D6, and D13) and short commute with large/small/no time component on
toll roads (D7–D12, and D14). The SP experiments were designed with conditions such that the toll cost always increases and
total travel time always decreases as more tolled links are added to the network. The former condition is quite easy to meet
with a selection of positive pivot levels for toll costs (see Table 1) but the latter condition requires a careful selection of pivot
levels for the two components of travel time (i.e., time on free segments and time on tolled segments). Nonetheless, a
removal of choice tasks that do not satisfy this condition is necessary. To have enough degrees of freedom, each SP exper-
iment was designed with 12 choice tasks but each respondent was asked to review only four of the 12 choice tasks that sat-
isfy both design conditions. Respondents were requested to indicate their preferences in terms of the acceptability of each
route (yellow9
line in Fig. 4), and the number of times per fortnight they would choose each route for commuting (green line in
Fig. 4). Information was grouped in such a way that allows respondents to base their answers on a one-way commuting time
and toll cost (blue block) or the cumulative time and toll costs over a typical two weeks of commuting (green block in Fig. 4).
The fifth and final part of the survey asks respondents to describe themselves and their households. Standard questions
relating to gender, age, personal income, household structure, and vehicle ownership are used to collect contextual informa-
tion for consideration in the choice model.
We undertook extensive pre-testing before finalising this instrument. It is important to note that we are studying a topic
in which there is extensive experience with using tolled routes (or avoiding them) in Sydney, and hence the topic is very
meaningful to respondents and of immense importance and interest in the Sydney context where the construction of
new tollroads is always in the news, and results in much discussion about how expensive it is getting. Individuals get a state-
ment each month on toll usage and indeed see the amount taken out of their bank accounts on a regular basis.
5. Sampling and sample profile
The survey was conducted using the Lightspeed GMI panel which has many thousands of panellists who are commuters
in the study area. Ethics approval (Project No.: 2015/393) was obtained for the experiment and each respondent received a
small incentive (as either cash, points redeemable for a gift card, or equivalent money donated to a charity depending on
their preference) for a completed survey. The main survey was conducted from 30th July to 11th August 2015 after a pilot
survey of 45 workers was carried out from 16 to 18 July 2015. In both the pilot and main surveys, respondents were recruited
via an e-mail directing them to a customised online survey. A sample of 500 valid responses was contracted and sampling
quotas were applied to obtain 400 tollway users. To account for data quality removals, however, the quotas were set at 480
tollway users and 120 non-tollway users (i.e., we over-recruited by 20%). All car drivers who travelled to work a minimum of
four days over the last two weeks were recruited. No other screening criteria or quotas were applied.
A total of 5651 workers were invited to undertake the survey and a sample of 600 completed respondents was obtained,
resulting in an incidence rate (IR) of 10.6%. This IR is very close to the rate (10%) derived from the annual Sydney Household
Travel Survey that a randomly selected worker who lives in Sydney will commute by car and use at least one toll road (i.e.,
toll road commuter). An extensive process of cleaning and validating the data reduced the sample to 410 usable respondents
9
For interpretation of color in Figs. 4 and 11, the reader is referred to the web version of this article.
D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621 611

9. (311 toll road commuters). Apart from 30 respondents with very short commuting trips and 46 speedsters (i.e., respondents
who completed the experiment too quickly (as our subjective assessment)), an additional 114 observations were removed
based on one or multiple cleaning rules10
which consider the consistency of information across different attributes. Fig. 5
shows the distribution of respondents by their home and workplace, together with the alignment of tollways and freeways.
As can be seen in Fig. 5, most respondents live in the northwest and southwest of Sydney, where the use of tollways for com-
muting to/from the employment hubs (the CBD, Lower North, Parramatta and Macquarie Park) is expected to shorten travel
time significantly, compared to an alternative free route.
Fig. 4. An illustrative choice set screen.
10
The rules are available on request.
612 D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621

10. Fig. 6 shows the number of toll roads used for the journey to work (JTW) of the sampled workers. The Journey from work
(JFW) is very similar. Of the commuters whose travel involved toll roads, the majority use one toll link with the most popular
toll roads being the M5, followed by the SHB, M7, M2 and the Eastern Distributor (ED). However, it is not uncommon for the
JTW to involve more than one tolled link. The most popular combination of toll roads are the M5 and M7 ($4723 per annum),
the SHB and LCT ($2462 per annum), the ED and CCT ($4046 per annum), M7 and M2 ($6739 per annum), and SHB, LCT and
M2 ($5539 per annum) with the number in parentheses being the annual toll outlay on commuting, assuming a 5-day work-
ing week and a 48-week working year (4 weeks vacation). The sample average annual gross personal income is $93,000 per
annum (Table 2), which after tax is around $68,000. The range of toll outlays associated with the toll activity summarised
above are from 2 to 9% of the after tax income for toll users (although there are a number of users in excess of 9%). As indi-
cated, the toll outlay for toll road commuters is substantial, and an addition of more tolled links may result in an increasing
number of commuters not prepared to pay tolls to save travel time. Fig. 7 shows the current level of toll saturation amongst
toll road commuters. One in five toll road commuters (65 out of 311 workers) have reached their saturation point, with an
average level of toll saturation amongst toll road commuters around 60%. Thus, some commuters can still sustain increasing
toll costs; but a substantial proportion appear to be no longer prepared ‘to pay to save’.
Table 2 completes the commuters’ profile. On average, the JTW or JFW of a sampled car commuter takes close to an hour,
with one-third of the commuting time being on toll roads.11
Over the last two weeks, commuters have outlaid, on average, $50
on toll roads with the maximum amount of toll outlay of $374. The toll outlay is currently smaller than the budget commuters
have for commuting on toll roads, with an average gap between toll outlay and toll budget of $37 ($87 À $50 = $37) for 2-week
commuting or $3.70 per day if commuters travel to and from work five days per week. The average age of sampled workers is
43 years and a vast majority (80%) work fulltime. Five percent of the workers have their commuting tolls covered by employers,
and another 4% of workers pay commuting tolls through their own business. In terms of gender and occupation, the sampled
workers spread quite evenly across both sexes and cover all occupations.
Table 1
Pivot levels of the SP experimental designs and assignment rules.
Attribute Reference level Pivot level Design Apply for current route which has. . .a
Total travel time on tolled segments 35 min 0, 5, 10 D1–D3 TOTtime P 40 min and
Total travel time on free segments 25 min À10, À15, À20 D1–D3 Ftime P 25 min and
Travel time on free route 90 min À10, 0, 10, 20 D1–D3 Tcost 0 and
Total one-way toll costs $14 2, 3, 5 D1 Budgleft 5 or
5, 7, 10 D2 Budgleft = 5–10 or
7, 10, 12 D3 Budgleft 10 or
Total travel time on tolled segments 30 min À10, À15 D4–D6 TOTtime P 40 min and
Total travel time on free segments 10 min À5, 0 D4–D6 Ftime 25 min and
Travel time on free route 60 min À10, 0, 10, 20 D4–D6 Tcost 0 and
Total one-way toll costs $10 2, 3, 5 D4 Budgleft 5 or
5, 7, 10 D5 Budgleft = 5–10 or
7, 10, 12 D6 Budgleft 10 or
Total travel time on tolled segments 15 min 0, 5, 10 D7–D9 TOTtime 40 min and
Total travel time on free segments 20 min À10, À15, À20 D7–D9 Ftime P 25 min and
Travel time on free route 50 min À10, 0, 10, 20 D7–D9 Tcost 0 and
Total one-way toll costs $7 2, 3, 5 D7 Budgleft 5 or
5, 7, 10 D8 Budgleft = 5–10 or
7, 10, 12 D9 Budgleft 10 or
Total travel time on tolled segments 15 min À10, À5 D10–D12 TOTtime 40 min and
Total travel time on free segments 20 min À5, 0 D10–D12 Ftime 25 min and
Travel time on free route 50 min À10, 0, 10, 20 D10–D12 Tcost 0 and
Total one-way toll costs $7 2, 3, 5 D10 Budgleft 5 or
5, 7, 10 D11 Budgleft = 5–10 or
7, 10, 12 D12 Budgleft 10 or
Total travel time on tolled segments 0 min 5, 10, 15 D13 TOTtime P 40 min and
Total travel time on free segments 60 min À20, À30, À40 D13 Ftime = 0
Total one-way toll costs $0 2, 3, 5, 7 D13
Total travel time on tolled segments 0 min 3, 5, 10 D14 TOTtime 40 min and
Total travel time on free segments 30 min À10, À15 D14 Ftime = 0
Total one-way toll costs $0 2, 3, 5, 7 D14
a
TOTtime = Total travel time one way; Ftime = Time on free segments; Budgleft = toll budget left for each commuting trip, calculated as the difference
between upper bound budget less the toll outlay in the last 2 weeks of commute and adjusted for the number of commuting trips over 2 weeks.
11
A number of commuters live in the Central Coast, which is over 90 km from the CBD. In addition, commuters coming from the far Outer West spent
significant time on connected toll roads (i.e., M7, M2, Lane Cove Tunnel and Harbour Bridge).
D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621 613

11. 6. Results
A number of alternative models were estimated, with the final model (M2) summarised in Table 3 together with a base
model (M1) that does not take into account each individual’s toll budget threshold. The representative component of the
estimated utility expression for the threshold model is given in Eq. (10). The observed sources of influence on utility asso-
ciated with the jth tolled alternative are heteroscedastic-conditioned (Hensher et al., 2015) on the percent of the toll budget
expended on tolls, and a dummy variable for whether an employer paid the tolls. We also investigated the role of income, but
did not find it to be significant as a replacement for the toll budget. We suggest that this is because the toll budget is some-
thing that cannot be proxied by income (given low correlation of À0.052 with psat and 0.161 with the toll budget) and is
related to the way that individuals partition their overall budget (as in the Strotz utility tree model (Strotz, 1957)) which
Fig. 5. Distribution of respondents by home and work postcodes.
614 D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621

12. does not assure that the amount allocated to a specific expenditure (in the current context, this is the amount outlaid on
tolls) is strictly proportional to income. An analysis of the empirical data suggests that there are many high income workers
living in areas where no toll roads are available for commuting and hence, they do not prepare to pay much on tolls to save
travel time. Fig. 8 shows a weak relationship between the annual personal income and toll budget. Thus, we chose to focus
on the toll budget which was statistically significant compared to replacing it with personal income.
An extensive set of explanatory variables were investigated, both as attributes describing the alternatives in the choice
experiment (Fig. 4) and socioeconomic and contextual characteristics. The selected influences summarised in Table 3 include
the age of the traveller and a series of dummy variables (1/0) for the location of the workplace and the availability of toll
roads for commuters who were observed to commute on a free road (i.e., non-toll commuters). These commuters are clas-
sified into two groups: one consciously avoids tolls and one does not have toll road options for commuting. These parameters
are significantly negative, suggesting that non-toll commuters are more sensitive to toll costs than toll commuters. We did
not find the commuter’s income and household characteristics except the age of the commuter to be statistically significant
Fig. 6. Number of toll roads involved on journey to work.
Table 2
Descriptive profile of sample.
Mean Std. Dev. Minimum Maximum
Journey to and from work travel time (min) 56 23 22 150
Travel time on toll roads to and from work (min) 20 20 0 140
Total toll outlay in last 2-week commuting ($)*
50 59 0 374
Toll budget for 2-week commuting ($) 87 88 0 500
Respondent age (year) 43 14 20 70
Personal income ($1000) 93 48 10.4 260
Worker pays tolls (1/0, base = other arrangement)*
57% n/a 0 100
Own-business pays tolls (1/0, base = other arrangement)*
4% n/a 0 100
Employer pays tolls (1/0, base = other arrangement)*
5% n/a 0 100
Male worker (1/0, base = female worker) 53% n/a 0 100
Fulltime worker (1/0, base = Casual/Volunteer) 80% n/a 0 100
Part-time worker (1/0, base = Casual/Volunteer) 14% n/a 0 100
Professional worker (1/0, base = labourer) 30% n/a 0 100
Admin worker (1/0, base = labourer) 27% n/a 0 100
Clerical worker (1/0, base = labourer) 14% n/a 0 100
Self-employed (1/0, base = labourer) 8% n/a 0 100
Sales worker (1/0, base = labourer) 7% n/a 0 100
Trading worker (1/0, base = labourer) 4% n/a 0 100
Workers with other occupations (1/0, base = labourer) 6% n/a 0 100
Note: n/a = standard deviation is not meaningful for dummy variables.
*
Statistics are based on the sub-sample of toll road commuters.
D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621 615

13. influences, and we surmise that the role of income may in part be taken into account through its influence on the toll budget
threshold.
Vði; jÞ ¼ ð1 þ bsatpsati þ bempEmpPayiÞ Â ðaj þ btollTotTollsj þ btimeTotTimej þ bsocioSocioiÞ ð14Þ
The overall goodness of fit of the models resulted in a pseudo-R2
of 0.421 for the threshold model and 0.419 for the base
model. There is an improvement in behavioural power of the threshold model given the AIC values of the two models. We
investigated a random parameter version of the model; however, a number of the critical parameters of interest were not
statistically significant with respect to the standard deviation parameter estimate. We have stayed with the nonlinear
non-random parameters version in this paper.
Fig. 7. Current level of toll saturation amongst toll roads commuters.
Table 3
Toll Road choice model with and without toll budget thresholds.
Attribute, variable Alta
Model w/o budget constraint (M1) Model w/budget constraint (M2)
Conditioning expression
Level of toll saturation, psat (%) 1–3 n/a 0.1949 (3.43)
Tolls paid by employer, EmpPay (1/0) 1–3 n/a À0.3688 (À5.45)
Linear utility inputs
2-week toll outlay, TotTolls ($) All À0.0034 (À3.62) À0.0026 (À3.85)
2-week commuting time, TotTime ($) All À0.0014 (À5.48) À0.0019 (À6.75)
Age of worker, Age (year) 2 À0.0199 (À2.74) À0.0191 (À3.41)
Age of worker, Age (year) 3 À0.0230 (À2.76) À0.0242 (À3.51)
Work in CBD (1/0) 1–3 0.1178 (0.64) 0.1792 (1.66)
Work in Lower North shore (1/0) 1–3 À0.5866 (À2.49) À0.4478 (À3.18)
Work in Eastern Suburbs (1/0) 1–3 À0.2315 (À0.96) À0.2044 (À1.28)
Non-toll commuters, avoiding tolls (1/0) 2–3 À0.4073 (À1.50) À0.3546 (À2.25)
Non-toll commuters, no toll options (1/0) 2–3 À0.7848 (À3.13) À0.6671 (À3.80)
Alternative specific constant, a 2 À1.3662 (À4.47) À1.1794 (À4.85)
Alternative specific constant, a 3 À1.5466 (À4.33) À1.3027 (À4.53)
Alternative specific constant, a 4 À1.2632 (À7.05) À1.3025 (À10.37)
Tolls paid by employer, EmpPay (1/0) 1–3 0.9466 (3.80) n/a
Model fit
Log-likelihood at zero À2273.52 À2273.52
Log-likelihood at convergence À1320.30 À1317.58
Pseudo-R2
0.419 0.421
AIC (sample adjusted) 1.626 1.624
Note: Model is based on 1640 observations and adjusted for multiple observations per respondent.
a
Alternative: 1 = current travel using one or more tolled routes; 2 = plus one more tolled segment; 3 = plus two more tolled segments, 4 = free route.
616 D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621

14. The percentage of the toll budget threshold expended on tolls is the variable of most interest, with a t-value of 3.53, which
is significant at the 99% level of confidence. The positive sign is in line with the rationale set out in the modelling approach
above. It conditions the marginal disutility of total travel time and total toll outlays over a two week period and influences
the marginal disutility in different ways (see Eq. (12)).
As hypothesised, when the accumulated toll outlay relative to the offered total travel time savings reaches the toll budget
threshold, individuals are likely to opt out of choosing to use additional tolled routes (or use existing routes less frequently)
that offer time savings, with the result that the implied VTTS is lower than what would be obtained if the toll budget was
ignored in model estimation and derivation of VTTS. Fig. 9, based on the choice experiment levels of offered total travel times
and tolls, shows the extent of the toll budget being exceeded under the time-cost trade-offs presented in the toll alternatives
in the choice experiment. The value of 1.0 on the horizontal axis is the level at which toll outlays coincide with the toll bud-
get threshold. Fig. 9 shows clearly that under the scenarios when tolls increased, toll costs start to bite into the personal toll
budget, with more people approaching and exceeding their toll saturation point (i.e., toll saturation level P 1).
What does this mean for the VTTS? In the presence of the imposed toll budget threshold, additional tolled roads which
offer potential travel time savings are less likely to be chosen as part of the commuter’s travel itinerary because the accu-
mulated outlay of tolls for all toll road usage over a period of time (i.e., two weeks), is such that individuals cannot justify
paying the increasing amount of tolls to gain an additional travel time benefit. Consequently, the implied amount of money
they are willing to outlay to save a unit of time is capped, resulting in an observed lower VTTS than would otherwise be
obtained under an unconstrained toll budget threshold.
Table 4 shows the effects of budget constraints on VTTS on commuting trips.12
The VTTS finding associated with Model M1
which does not explicitly take into account the toll budget threshold is $24.24 per person hour with a standard error of 7.09. The
estimated VTTS is statistically significant at the 99% level using a Wald test to implement the Delta method to compute the vari-
ances of nonlinear functions to obtain the standard error estimates (see Hensher et al., 2015, Chapter 7). The mean VTTS in M1 is
51.5% of sample average weekly income. When we account for the toll budget constraint (Model M2), the VTTS estimate asso-
ciated with the current trip experience is on average $12.04 per person hour (statistically significant with a standard error of
$2.94), or 25.6% of sample average weekly income. When we investigate the time-toll trade-off associated with introducing an
additional toll segment, we find that the mean VTTS decreases to $6.09 per person hour (statistically significant with a standard
error of $1.60), and adding a further toll segment decreases the mean estimate to $5.70 per person hour (statistically significant
with a standard error of $1.52). That is, the more toll roads are added to the network such that more people reach and go beyond
their saturation point, the average VTTS becomes smaller, and so does its standard error. However, we observe a flattening out
of the VTTS for current plus 1 and current plus 2. The evidence suggests that many commuters in Sydney are already close to
their toll budget threshold, which when reached will flatten the commuter’s willingness to pay to save extra commuting time.
Fig. 10 shows the distribution of the VTTS under the toll budget threshold constraint of each commuter associated with
the choice scenarios assessed in the choice experiments. The vertical axis is density and the horizontal axis is VTTS. Taking a
Fig. 8. The relationship between personal income and the toll budget.
12
The predicted utility associated with the conditioning expression is positive, as expected, for each and every sampled respondent. The mean is 1.051, with a
standard deviation of 0.103 and a range from 0.672 to 1.147.
D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621 617

15. VTTS of $10 per person hour as an example, under the current level of toll saturation, a relatively large number of sampled
commuters are willing to pay more than $10 to save an hour of commuting time (i.e., the area to the right of $10). However,
once additional tolled links are added to the network (the second and third kernel density distributions to the right in Fig. 10)
such that toll costs start to bite into personal toll budgets, which limits the total amount commuters are prepared to outlay
on using toll roads over a given period of time, an hour of savings in commuting time is no longer valued at $10 or more by a
relatively larger number of commuters, with the proportion of commuters with a VTTS lower than $10 per person hour being
a far greater percent of the sample.
Fig. 11 presents the distribution of VTTS across workers with different toll budget thresholds. As expected, workers with a
higher budget for tolls have a higher VTTS; however, the VTTS increases at a diminishing rate. A majority of the sampled
commuters have a 2-week toll budget between $20 and $200, and Fig. 11 shows that under the current route, a commuter
with a $20 budget has an average VTTS of around $8 per hour. If the toll budget increases to $100, the VTTS increases to $14
per hour (i.e., a $6 increase); however, with the same $80 increase in a toll budget from $100 to $180, we only see a $2
increase in VTTS ($14–$16 per hour). This direction and rate of change is consistent across levels of tolls associated with
existing tolled roads as well as future new tolled routes (in our study relating to one and two additional tolled segments).
What is of particular interest, however, is to see how the VTTS of the same commuter (i.e., same toll budget) changes
when more tolled links are added to the network. This is revealed by comparing the VTTS for the same budget across the
three profiles shown in Fig. 11. Take a toll budget of $200 as an example, under the current level of toll saturation (green
circle); a worker is willing to pay, on average (the green line), $16 to save one hour of commuting time. When more tolled
links are added to the network (as shown the blue triangle and red cross series) such that the level of toll saturation of work-
ers with a $200 budget increases, they will value an additional hour of savings of commuting time at only $8 (i.e., $8 less).
This means that a higher level of toll saturation results in a decrease in the VTTS and the decrease is at a diminishing rate.
That is, there is not much change to the VTTS when we go from adding one to adding two more tolled links. This is because a
majority of the sample are already at the saturation point when adding one more toll roads (see Fig. 9), and at this point extra
savings in commuting time are valued at the ‘floor price’.
7. Conclusions
This study has shown that the current practice of assuming that the time-cost trade-off associated with a new tolled link,
which ignores the accumulated time and cost outlaid in using existing toll routes, is very likely to result in an upward biased
Current route
2 more tolled
links
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0 1 2 3 4 5
CummulaƟvepercentageofworkers
Level of toll saturation
Level of toll saturation under current and future toll scenarios
Fig. 9. Relationship between toll outlays associated with the choice experiments and respondent toll budget thresholds.
Table 4
Effect of toll saturation on commuter’s VTTS ($/person hour).a
VTTS Model w/o budget constraint (M1) Model with budget constraint (M2)
Current route Current plus 1 Current plus 2
Mean $24.24 $12.04 $6.09 $5.70
Standard Error $7.09 $2.94 $1.60 $1.52
a
The equivalent VTTS for a single trip are respectively for current route, current route plus 1 and current route plus 2: $18.89 (StErr $5.98), $9.55 (StErr
$2.98), and $8.98 (StErr $2.78). These mean and standard error estimates are higher than shown for the total time and total toll models in the table. The ratio
of the means for the accumulated and single trip times and costs VTTS are 0.637, indicating that when we ignore the accumulation of travel time and tolls in
the representation of travel (focussing only on a single trip), the VTTS is 57% higher (1/0.637).
618 D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621

16. Kernel density of VTTS under current road
VTTS
.010
.020
.030
.040
.050
.060
.070
.080
.090
.000
10 20 300
DENSITY
DENSITY
Kernel density of VTTS under scenario with one more tolled link
VTTS
.0250
.0500
.0750
.1000
.1250
.0000
0 5 10 15 20-5
DENSITY
DENSITY
Kernel density of VTTS under scenario with two more tolled links
VTTS
.0250
.0500
.0750
.1000
.1250
.0000
0 5 10 15 20-5
DENSITY
DENSITY
Fig. 10. Kernel density of VTTS for reference alternative, add 1 and add 2 tolled segments.
D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621 619

17. estimate of the mean VTTS. This is a very important finding, raising serious concerns about the estimates of VTTS associated
with all current studies used to forecast patronage demand for new tolled links when there are existing tolled links in place.
It may also have implications for the frequency of use of the existing tolled links.
The use of a VTTS distribution that recognises the distribution of toll saturation in a population appears to be important in
practice. For example, if we work with the findings in Fig. 11 for the context of additional tolled roads, the range of VTTS is
from $0.1 to $20 per person hour. We know from many previous toll road patronage studies undertaken in Australia by var-
ious consortia bidding for the concession that there are many travellers who are not prepared to pay a toll to save time,
which is shown very clearly in Fig. 11; there are however travellers who are willing to pay substantially higher amounts
to save time, well above the toll prices paid. A frequency distribution can be used to represent the incidence of each VTTS
on the assumption that the toll saturation distribution in the sample applies to a population. In applications, the analyst
can then work with this distribution of VTTS in the exact same way that they might use a distribution of VTTS associated
with a random parameter formulation. The approach we have developed has relevance even if there are no tolled roads
in place, since toll budget thresholds are real and can be applicable even for assessment of the VTTS associated with an initial
tolled route.
A question of importance is to decide which estimate of VTTS to use when additional tolled roads are introduced into an
existing tolled network. Should we use the VTTS associated with a proposed (i.e., additional) road or some average across all
tolled roads, current and prospective? Our preferred position is to recognise that all tolled roads are most likely to be
reviewed in the presence of a new tolled route, and hence the VTTS of relevance is the one that is aligned with the toll budget
threshold in the presence of all tolled roads on offer. In our empirical inquiry for Sydney, this would be no greater than $12,
and possibly closer to $6 per person hour (Table 4). It is notable that a typical estimate used by bidding consortia in many
previous tolled road studies is close to $18.23 per person hour ($2005) for commuters (see Hensher, 2011, Table 7.1), or
updated to close to $22.23 per person hour in $2015 (using an average inflation rate of 2% per annum), similar to the uncon-
strained toll budget estimate herein of $24.24 per person hour.
Although we have focussed on commuting travel and commuting trip-related toll budget thresholds, the approach can be
implemented for all travel purposes. The approach developed and the resulting evidence suggests that all VTTS estimates are
likely to be lower than those obtained by all previous studies, with the obvious implications on practice for both travel
demand forecasts and the time benefits used in appraisal of projects.
In ongoing research, additional data will be collected in face to face interviews to assess the extent to which the findings
in this paper are robust. This is a necessary task given the implications that the findings have on the over-estimation of the
mean VTTS in practice. Subsequent evidence, however, does not detract from the value of raising a concern about the
absence of any consideration of toll saturation in VTTS estimation, an issue that has often been mentioned in passing in
meetings on toll road traffic forecasts13
but which has never been seriously investigated.
Acknowledgments
This research was partially funded by the University of Sydney Business School Research Scheme. We also thank a num-
ber of banks and construction companies who have expressed curiosity about the likely limits to toll road use as the accu-
mulated cost increases. We also thank Rob Bain, Gillian Akers and John Rose for the many discussions we have had on this
topic, as well as the detailed comments and suggestions of two referees.
0
5
10
15
20
25
30
0 100 200 300 400 500
VTTS($/HOUR)
TOLL BUDGET ($)
Current route Current plus 1 toll Current plus 2 tolls
Fig. 11. Plot of VTTS against toll budget threshold for reference, add 1 and add 2 tolled segment alternatives.
13
The lead author has been involved in numerous toll road studies for private sector consortia where this matter has been mentioned.
620 D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621

18. References
Bain, R., 2009. Error and optimism bias in toll road traffic forecasts. Transportation 36 (5), 469–482.
Bain, R., 2011. On the Reasonableness of Traffic Forecasts, TEC, May, pp. 213–217. http://www.trafficresearch.co.uk/.
Choice Metrics, 2012. NGene. Choice Metrics, Sydney.
Flyvbjerg, B., Holm, M.K.S., Buhl, S., 2006. Inaccuracy in traffic forecasts. Transp. Rev. 26 (1), 1–24.
Hensher, D.A., 2011. Value of Travel Time Savings Handbook in Transport Economics. In: de Palma, André, Lindsey, Robin, Quinet, Emile, Vickerman, Roger
(Eds.), Edward Elgar Publisher, UK, pp. 135–158 (Chapter 7).
Hensher, D.A., Rose, J.M., 2012. The influence of alternative acceptability, attribute thresholds and choice response certainty on automobile purchase
preferences. J. Transp. Econ. Policy 46 (3), 451–468.
Hensher, D.A., Ho, C.Q., 2016. Experience conditioning in commuter modal choice modelling – does it make a difference? Transp. Res. Part E 95, 164–176.
Hensher, D.A., Louviere, J.J., Swait, J., 1999. Combining sources of preference data. J. Econom. 89, 197–221.
Hensher, D.A., Rose, J.M., Greene, W.H., 2015. Applied Choice Analysis. Cambridge University Press, Cambridge.
Li, Z., Hensher, D.A., 2010. Toll roads in Australia. Transp. Rev. 30 (5), 541–569.
Rose, J.M., Hensher, D.A., 2014. Tollroads are only part of the overall trip: the error of our ways in past willingness to pay studies. Transportation 41 (4), 819–
837.
Saulwick, J., 2015. How much is too much for Sydney toll roads? Sydney Morning Herald. http://www.smh.com.au/nsw/how-much-is-too-much-for-
sydney-toll-roads-20150412-1mif9l.html (accessed on 11 May 2015).
Soofi, E.S., 1994. Capturing the intangible concept of information. J. Am. Stat. Assoc. 89, 1243–1254.
Strotz, R.H., 1957. The empirical implications of a utility tree. Econometrica 25 (2), 269–280 https://www.jstor.org/stable/1910254?seq=1#page_scan_
tab_contents.
Swait, J., Adamowicz, W., 2001a. The influence of task complexity on consumer choice: a latent class model of decision strategy switching. J. Consum. Res.
28, 135–148.
Swait, J., Adamowicz, W., 2001b. Choice environment, market complexity, and consumer behavior: a theoretical and empirical approach for incorporating
decision complexity into models of consumer choice. Organ. Behav. Hum. Decis. Process. 49, 1–27.
The RiverCity Motorway, 2010. CLEM7 Tolled Road. Leighton Industries, Brisbane.
Transportation Research Board (TRB), 2006. Estimating Toll Road Demand and Revenue, NCHRP Synthesis 364. Transportation Research Board, Washington,
DC.
Welde, M., Odeck, J., 2011. Do planners get it right? The accuracy of travel demand forecasting in Norway. Eur. J. Transp. Infrastruct. Res. 11 (1), 80–95.
D.A. Hensher et al. / Transportation Research Part A 94 (2016) 604–621 621