The Journal of Choice Modelling 6 (2013) 1733

Contents lists available at SciVerse ScienceDirect

The Journal of Choice Modelling

journal homepage: www.elsevier.com/locate/jocm

Modelling travellers' heterogeneous route choice behaviour

as prospect maximizers
Giselle de Moraes Ramos n, Winnie Daamen 1, Serge Hoogendoorn 2
Department of Transport & Planning, Faculty of Civil Engineering and Geosciences, Delft University of Technology, P.O. Box 5048, 2600 GA Delft, The Netherlands

a r t i c l e i n f o

abstract

Available online 28 May 2013

The use of Prospect Theory to model route choice has increased in the past decades. The
main application issue is how to define the reference point, i.e., the value that travellers
use as a reference to distinguish gains and losses in the experienced travel times. Moreover,
the question can be asked whether all travellers have the same reference point or whether
heterogeneity in their behaviour plays an important role.
This paper aims to (i) provide a behavioural interpretation of the reference point,
(ii) investigate the role of heterogeneity in the reference point and (iii) discuss how to take
heterogeneity into account. These aspects are discussed with the aid of an empirical route
choice experiment and a model specification in which travel time is the main variable. Two
model frameworks are proposed, one accounting for heterogeneity and another considering
no heterogeneity in travellers' behaviour, and their outcomes are compared.
Results show improvements in the ability of Prospect Theory to predict route choice
behaviour by accounting for heterogeneity in the reference point. This is particularly the
case when the reference point reflects travelers' route preferences. Statistical analyses show
the significance of accounting for heterogeneity in travellers' behaviour. Thus, we cannot
reject the hypothesis that heterogeneity leads to improvements in the prediction ability of
Prospect Theory.
& 2013 Elsevier Ltd. All rights reserved.

Keywords:
Prospect Theory
Reference point
Travel behaviour and heterogeneity

1. Introduction
Travel is the result of individual choice behaviour regarding (i) whether to leave home to engage in an activity, i.e.,
activity choice, (ii) where to perform the activity, i.e., destination choice, (iii) how to reach the destination, i.e., mode choice,
(iv) when to depart, i.e., departure time choice and (v) which route to take, i.e., route choice (Bovy et al., 2006). Altogether
these travel related decisions directly affect the performance of the transportation network, especially in the scenario of
increased mobility observed in most major city centres. As a result, travel characteristics such as travel times and congestion
patterns are even more severely impacted leading to more travel time uncertainty, which is one of the main impacting
factors on travellers' behaviour (Noland and Small, 1995; Avineri and Prashker, 2003; De Palma and Picard, 2005; Henn and
Ottomanelli, 2006). A proper understanding of travellers' behaviour, therefore, is of fundamental importance to predict
travellers' decisions and to forecast future traffic conditions on the network.
The literature on behavioural theories shows that travellers employ different criteria in the process of evaluating what
the best choice is. The majority of existing route choice models, however, is based on the utility maximization assumption
which assumes that people act rationally in order to get the maximum utility (benefit) from the decision made. In the field
n

Corresponding author. Tel.: +31 15 278 49 14; fax: +31 15 278 31 79.
E-mail addresses: g.m.ramos@tudelft.nl (G. de Moraes Ramos), w.daamen@tudelft.nl (W. Daamen), s.p.hoogendoorn@tudelft.nl (S. Hoogendoorn).
1
Tel.: +31 15 278 59 27; fax: +31 15 278 31 79.
2
Tel.: +31 15 278 54 75; fax: +31 15 278 31 79.

1755-5345/$ - see front matter & 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.jocm.2013.04.002

18

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

of route choice under uncertainty Expected Utility Theory (Von Neumann and Morgenstern, 1947), in particular, is the most
widely used theory (De Palma and Picard, 2005). Situations dealing with routes' travel times definitely involve uncertainty.
For instance, irrespective of the experience someone has gathered due to frequently driving a specific route, the travel times
may reasonably vary depending on the traffic conditions.
Despite the widespread use of Expected Utility Theory, experiments in behavioural studies have found deviations from
its axioms leading to the development of Non-Expected Utility Theories of which Prospect Theory (Kahneman and Tversky,
1979) is the most discussed (Starmer, 2000). Prospect Theory argues that choices are based on gains and losses measured
against a reference point, i.e., values above it are perceived as gains and values below it as losses. What matters, therefore, is
the relative gain and not the final state of wealth or welfare as argued by Expected Utility Theory.
Prospect Theory has been widely used in the field of economics, but applications in the field of transport are relatively
recent (Sumalee et al., 2005; Avineri, 2006; Connors and Sumalee, 2009; Gao et al., 2010). Its application has been facing
two main issues: (i) definition of meaningful reference points within the travel behaviour context and (ii) estimation of
appropriate parameters for the value and weighting functions.
The lack of consensus about the meaning of the reference point in a route choice context has often surfaced in the
literature. For instance, De Palma et al. (2008) suggest that the determination of reference points is one of the major
obstacles for the application of Prospect Theory and that it is likely that the reference point varies depending on individuals
and choice contexts. While for situations dealing with monetary outcomes, zero is the usual reference point (meaning
neither gains nor losses), for situations dealing with travel times this value may vary, for instance, with respect to the
decision maker, to the distances travelled, to the level of stress and to constraints regarding arrival time (Schwanen and
Ettema, 2009; Senbil and Kitamura, 2004). Therefore, questions concerning the meaning of the reference point in situations
involving route choice as well as its pattern over time are raised. In other words, it still has to be made clear what values
travellers use as a reference to distinguish experienced travel times into gains and losses. Do all travellers have the same
reference point or does heterogeneity in their behaviour play an important role? How do repetitive route choices influence
the reference point over time?
As both intuition and literature suggest, attitudes towards risk vary across the population. As a result, travellers may have
different perceptions on how to characterise outcomes into gains or losses and thus different reference points. In order to
capture travellers' perceptions of travel time variability and how this influences their route choices over time, we propose to
investigate travellers' heterogeneity in relation to their reference point. We hypothesise the following:
(i) The reference point varies among travellers and over time, i.e., in case travellers' route preferences change over time,
such as switching to a more reliable route instead of a fast but unpredictable route, the reference point will follow
travellers' new behaviour.
(ii) The reference point reflects travellers' (risk) preferences when making route choice decisions, i.e., the reference point is
aligned with the travel time distribution of the preferred route.
(iii) In case pre-route information is provided, such as travel time, travellers might use that value as a reference point.

Different from what has been observed in the literature, the contributions of this paper lie on the investigation of the reference
point in light of its behavioural appeal and the on the role of heterogeneity in the reference point. As a result, we try to grasp
the rationale behind risky and risk-averse choices throughout the reference point and the relationship between travellers'
heterogeneous preferences and the reference point. This is done by directly applying Prospect Theory in two model specifications,
one accounting for heterogeneity and another considering no heterogeneity. Afterwards, their results are compared.
The model comparison is based on data from an empirical route choice experiment (Bogers, 2009) in which travellers
were asked to make choices among three possible routes: route 1 consisting mainly of highways, route 2 consisting mainly
of rural roads and route 3 consisting partly of highway and partly of urban roads. Two conditions of information provision
and two travel purposes were investigated which resulted in four scenarios.
Rather than claiming a higher importance of heterogeneity in the reference point to other types of heterogeneity related to
risk, we intend to demonstrate, based on empirical data, the benefits of taking the heterogeneity in the reference point into
account. Despite the existence of a great variety of models, (mostly) based on the utility maximization assumption, that are
able to describe route choice behaviour quite well, we propose to investigate the suitability of Prospect Theory due to its
potential to better capture travellers' behaviour. Results from previous research conducted by the authors suggest the
suitability of Prospect Theory to model route choice behaviour and that depending on the reference point Prospect Theory can
perform better than Utility Theory (Ramos et al., 2011). This motivated further investigation into the role of heterogeneity in
the reference point, which is presented here.
Another well-known factor to be observed is that provision of travel information is likely to change the level of
uncertainty of travel related decisions. Modelling its impact is hardly a simple task. Additional information together with
advanced technologies, such as GPS-based path-finders, are likely to contribute to reduce travel time uncertainty, to enable
travellers to choose efficiently among available routes, to save travel time and to reduce congestion (European Commission,
2008). The impact of information, however, is likely to be sensitive to travellers' behavioural and cognitive response to
information that is much less understood. In addition, effects such as experience and learning also seem to play an
important role in the decision making process (Ben-Elia and Shiftan, 2010).

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

19

Although this paper focuses on the role of heterogeneity in the reference point rather than on the impact of information
on travellers' behaviour, given that two conditions of information provision are considered, we expect that travellers show
different route choice behaviour. For instance, in the scenarios in which no information was provided it is expected that
travellers are more willing to explore the routes in order to get familiar with their characteristics. As a result, higher
switching/exploration rates should be observed. As time evolves and travellers get familiar with the travel times of the
routes, routes' switching rates should decrease. In the scenarios in which information was provided, determining a route
exploration pattern is not straightforward because it also depends on the compliance rate with the information. We expect,
however, that exploration rates of travellers that in general do not comply with information to be similar to the situation in
which no information was provided. For travellers that in general comply with the information, as the information provided
was very accurate, we expect that as time evolves and travellers become more confident about the reliability of the
information, compliance rates with the information should gradually increase.
The next section presents main distinctions between the fundamentals of Prospect Theory in relation to Expected Utility
Theory and a discussion about the role of heterogeneity and the reference point. In Section 3, the proposed case study is
presented and in Section 4 two model specifications are introduced. Comparisons between the model specifications are
presented in Section 5. Finally, conclusions and recommendations for future research are provided in Section 6.
2. Behavioural foundations of Prospect Theory
Prospect Theory, a descriptive model of decision making under risk, was developed in 1979 as a critique to Expected
Utility Theory based on the main assumption that choices are based on gains and losses measured against a reference point
(Kahneman and Tversky, 1979). The point of the critique is that due to the reference point, changing the ways in which
options are presented generates predictable shifts in preferences and systematically violates the axioms of Expected Utility
Theory, in particular the independence axiom. The predictable shifts in preferences are not necessarily categorized as good or
bad, but instead reflect how people behave when making choices.
2.1. Initial developments
Considering that a prospect is a consequence associated to a probability of occurrence, the independence axiom states that for
all prospects x, y, z: if xy then (x, p; z, (1p))(y, p; z, (1p)). This implies that the preference between two prospects is
independent of the common components. Consequently, in the evaluation of a prospect their common components are perceived
as if they could be cancelled (Weber and Camerer, 1987). Experiments such as the Allais Paradox (Allais, 1953) presented below have
shown that the independence axiom is not necessarily applicable and that due to both the framing of options and the reference
point, shifts in preferences can be observed. Contrary to the expected utility assumption that people behave as utility maximizers
and have well defined preferences, Prospect Theory argues that preferences vary depending on how alternatives are presented.
The Allais Paradox is the first famous counter example to Expected Utility Theory to show shifts in preferences. Allais proposed
two experiments in which people had to choose between two lotteries in each of them (Table 1). A person with expected utility
preferences would choose option b in both experiments because the independence axiom requires that the difference between the
pairs of lotteries to be ignored. Thus, the common consequence of 0.89 chance of winning $1,000,000 in the lotteries a1 and b1, or
a 0.89 chance of (winning) $0 in the lotteries a2 and b2 would be cancelled and in both experiments the utility of the choice (u)
would be 0.1 u($5M)40.11 u($1M). Allais, however, expected that when faced with these types of choices people would opt for a1
in the first situation (win for certainty) and b2 in the second (very different rewards for slightly different winning probabilities). The
experiments performed confirmed this intuition, and consequently violated the independence axiom.
Following Allais, Prospect Theory was further developed by explaining choice behaviour as a two-step process: an initial
phase of editing and a subsequent phase of evaluation. In the editing phase the options are organized and reformulated by
the application of heuristics to simplify the evaluation. In the evaluation phase the prospect is subjectively valued. The
evaluation phase is divided into two scales: a weighting function (p) and a value function v(x). The probability-weighting
function (p) associates to each probability of occurrence p a decision weight that reflects the impact of p on the overall
value of the prospect. The value function v(x) reflects the subjective value of the outcome, thus measuring the deviations
from the reference point. Thus, neither v(x) is perceived as valuing x nor (p) as valuing p.
Table 1
The Allais Paradox.
Experiment 1

Experiment 2

Lottery a1 (to win)

$1,000,000 with certainty

Lottery a2 (to win)

$1,000,000 with probability 0.11
$0 with probability 0.89

Lottery b1 (to win)

$5,000,000 with probability 0.1
$1,000,000 with probability 0.89
$0 with probability 0.01

Lottery b2 (to win)

$5,000,000 with probability 0.1
$0 with probability 0.9

20

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

Value

Outcome
Gains

Losses

Reference Point
Fig. 1. Illustration of the shape of the value function.

Weighted probability
1
(p)
(p)

1
p (probability)

Fig. 2. Illustration of the shape of the weighting function.

2.2. Cumulative Prospect Theory

In 1992 advances were made to extend Prospect Theory to uncertain outcomes (Tversky and Kahneman, 1992) which led to the
development of Cumulative Prospect Theory. This extension was necessary to allow evaluation of situations involving uncertainty in
which some of the outcomes or probabilities are unknown. This is, for instance, the case for situations involving travel times.
Cumulative Prospect Theory employs cumulative rather than separable decision weights. It is applicable to uncertain and
risky prospects and allows different weighting and value functions for gains and losses due to people's different perception
of gains and losses. Distinct functions associated to positive and negative outcomes were proposed and defined in terms of
specific parameters that enable capturing choice behaviour (Figs. 1 and 2).
The value function is defined by three parameters, , and which are responsible for respectively reflecting (i) the degree of
diminishing sensitivity for gains, (ii) the degree of diminishing sensitivity for losses, i.e., the decrease on the marginal value of gains
and losses with their magnitude, and (iii) the degree of loss aversion, i.e., the aggravation one experiences when incurring losses.
The weighting function, on the other hand, is defined in terms of two parameters, and , that capture the distortion
in the perception of probabilities for gains and losses. Its reversed s-shape is derived from (i) the overweight of small
probabilities, meaning that people are risk-seekers when offered low-probability high-reward prizes and (ii) the underweight of moderate and high probabilities, implying that (a) people are relatively insensitive to probability difference in the
middle of the range and (b) the prevalence of risk aversion in choices between probable gains and certainty; and risk
seeking in choices between probable and sure losses. In addition, the curve of the weighting function starts in zero and
finishes in one ((0)0 and (1) 1) and it is asymmetrical with the inflection point at about 0.3 (Prelec, 1998). The
aforementioned parameter specifications for the value and weighting functions are based on their original formulation
(Tversky and Kahneman, 1992). For the weighting function in particular, alternative formulations such as the power function
(Prelec, 1998) often appears in the literature (De Palma et al., 2008).
Given the distinct functions for positive and negative outcomes, V+ and V respectively, the overall value of a prospect is
defined by V V++V:
n

V
i pvxi
i0

and

i pvxi

i m

where min and m and n are the number of negative and positive outcomes respectively.
Considering that a prospect f (xi, Ai) is given by a probability distribution p(Ai)pi, the probabilistic or risky prospect
can be viewed as (xi, pi), and the decision weights are defined by

n pn ;

m pm

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

for 0in1
i pi pn pi1 pn
i pm pi pm pi1 for 1mi0

21

In order to contribute to a better understanding and application of the above, consider a gamble with eight playing cards
from ace (one) to eight. You have to pick one card from the pack x 1,, 8, check the result, put the card back and then do it
again. If x is odd, you pay $x, otherwise you receive $x. Considering equivalently probable outcomes, f yields the outcomes
(7, 5, 3, 1, 2, 4, 6, 8), each with probability 1/8. Therefore, f+ (2, 1/8; 4, 1/8; 6, 1/8; 8, 1/8) and f (7, 1/8; 5, 1/8; 3,
1/8; 1, 1/8). Using Eq. (2),
V V V
v2 1=2 3=8 v4 3=8 1=4 v6 1=4 1=8
v8 1=8 0 v7 1=8 0 v5 1=4 1=8
v3 3=8 1=4 v1 1=2 3=8
The weighting function is given by
p

p
p 1p 1=

and

p
p 1p 1=

where and define the curvature of the weighting function.

The value function is given by
(
x if x0
vx
if x o 0

where x is the attribute to be measured against the reference point, reflects the degree of loss aversion and and
measure the degree of diminishing sensitivity.
The values of these parameters as originally estimated for situations involving monetary outcomes are 0.88,
2.25, 0.61 and 0.69 (Tversky and Kahneman, 1992). These parameters have also been estimated in a travel
behaviour context of employed parents coping with unreliable transport networks when picking up their children from
childcare at the end of the workday (Schwanen and Ettema, 2009). It was assumed that both the value and weighting
functions were similar for gains and losses, i.e., and . Results show that 1.09, 1.10, 1.271.37 and 0.82,
0.84 revealing some attenuation of the properties of Prospect Theory, in particular the loss aversion effect () and the
distortion in the perception of probabilities ( and ). For a comprehensive review of Prospect Theory, we refer to Kahneman
and Tversky (1979) and Tversky and Kahneman (1992).
2.3. Evaluating travellers' route choice behaviour as prospect maximizers and the role of the reference point
The use of Prospect Theory to model route choice behaviour is relatively recent and studies have been focusing on
the effects of travel time uncertainty on equilibrium conditions (Sumalee et al., 2005, 2009; Avineri, 2006; Connors and
Sumalee, 2009; Gao et al., 2010). In general, comparisons between Prospect Theory and Expected Utility Theory based
equilibrium conditions are shown with the aid of illustrative examples and, despite the emphasis put on the role/
importance of the reference point and other parameters, results are usually preliminary insights into the differences
between these equilibrium conditions. This is mainly due to the lack of consensus about the behavioural meaning of the
reference point.
In situations involving route choices the reference points that are usually adopted are based on the travel times of the
proposed case studies, on average travel times to work as reported by transport studies (30 min) or free-flow travel time.
However, as discussed in Van de Kaa (2010), when commuters are confronted with alternative routes with different travel
time distributions, the reference point might encompass the mean travel time of the preferred route, the distribution over
time and possibly also a notion of an acceptable bandwidth of arrival times and a preferred departure time. In addition, as
many studies suggest, the adopted reference point is influenced by implicit information, explicit information, or even
irrelevant information (White et al., 1994; Kristensen and Grling, 1997).
The behavioural appeal of the reference point is presented in some studies, such as De Borger and Fosgerau (2008),
in which the reference travellers might have in mind when making route choices is discussed. Outcomes suggest the
importance of recent experiences in the definition of the reference point. Connection between the reference point and
travellers' risk preferences, however, has not been found in the literature.
2.4. Role of heterogeneity in the reference point
Another aspect that has not been tackled in the literature is the heterogeneity in the reference point. Experiments in
psychological and social studies have shown that due to people's different perceptions of gains and losses, the perception of
uncertainty and probabilities systematically vary (Wu and Gonzalez, 1996). A similar effect is observed in repeated choice
tasks in which some people may keep on updating their reference while others might systematically choose the most
rewarding probabilistic alternative after a learning period (Schul and Mayo, 2003). Despite the importance of heterogeneity

22

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

as discussed in the literature and shown in travellers' behaviour analysis involving discrete choice models (Polak et al., 2008;
Gopinath, 1995) and car following models (Ossen and Hoogendoorn, 2011), it is usually argued that a single reference point
is used (Zhang et al., 2010; Avineri, 2009) which implies that all travellers value gains and losses similarly. In reality,
however, there might be two or more reference points (Avineri and Bovy, 2008; Van de Kaa, 2008) which may vary not only
among travellers, but also over time.
Over the past decade considerable progress has been made in the characterisation of unobserved taste heterogeneity in
travel choice behaviour (Hess, 2007; Koppelman and Sethi, 2005; Srinivasan and Mahmassani, 2003; Arnott et al., 1992). The
term taste heterogeneity refers to situations in which different decision makers take into account either different factors or
the same factors in different ways than other factors. Two main approaches have been used to represent the variations in
preferences among individuals (Ben-Akiva and Lerman, 1985). The most common is the partition of the data into a fixed
number of segments usually using demographic characteristics (exogenous segmentation). Alternatively, behaviourally
homogeneous segments are directly identified from the data (endogenous segmentation).
Route choice behaviour in particular has been widely investigated in the presence and absence of information and it
appears that travel time is one of the most relevant factors influencing decisions. Due to traffic congestion, however, travel
times are uncertain and travellers are likely to have different perceptions of the benefits of choosing a specific route in
relation to another. As a result, it is natural that they also have different reference points regarding perceptions for gains or a
losses. For instance, while for a more conservative traveller a 40 min travel time between A and B may be considered good,
thus perceived as a gain, for a more aggressive traveller, or someone running late, the same travel time may be perceived
negatively and thus as a loss.
Despite evidence that route choice behaviour varies with drivers' familiarity and prior experience, availability and
perception of alternative routes, travel times, delays, incidents and availability of information (Srinivasan and Mahmassani,
2003), empirical exploration of variations in reference points across respondents are scarce (Hess et al., 2011). In addition, as
discussed in Van de Kaa (2010), the large number of studies investigating interpersonal heterogeneity in application of
choice behaviour strategies have shown a better descriptive ability than homogeneous compensatory attribute processes.
Therefore, in a spatially dynamic framework, such as the one proposed in this paper (in which travellers make daily route
choices among three alternative routes), it becomes particularly important to account for heterogeneity as this helps
capturing individual preferences that persist over time.
3. Case study: investigating travellers' route choice behaviour
This case study is based on the data from the experiment performed by Bogers (2009), in which a travel simulator was
used to investigate travellers' route choice behaviour. The experiment simulated daily departures at 8:00 a.m. and arrivals
within 1 h at the destination for either a meeting with colleagues or a job interview. Travellers were recruited from the
database of the Dutch Organization of Road Users (ANWB) and as an incentive to join the experiment one of the travellers
would be awarded a navigation system. The prize was randomly drawn among the participants without specifying a goal
such as minimizing travel times or arriving on time (which would potentially influence their behaviour). The participants
made 40 consecutive choices among three routes with an approximate length of 30 km. These routes had the following
characteristics: (i) route 1 consisted mainly of highways, it was the fastest route, (ii) route 2 consisted mainly of rural roads,
it was the most reliable route and (iii) route 3 consisted partly of highways and partly of urban roads (going through the city
centre), it was the route of intermediate performance (Fig. 3).
Two conditions of information provision were investigated in this case study: (i) no information provision and
(ii) provision of travel time in minutes. Information was provided at the beginning of every new trip and only after
reviewing this information travellers could decide which route to take. Changing routes after the journey had started was
not possible. The travel information was provided for all three routes and based on the true travel times, thus it was very
accurate for all routes. The information deviated at most 2 min in 85% of the days and approximately between 10 and 30 min
in the remaining 15% of the days. In addition, the information provided usually underestimated the true travel times.
Feedback information with respect to the true travel time of the chosen route was provided at the end of every journey.
Combination of the two travel purposes, i.e., meeting with colleagues and job interview, and the two conditions of
information provision, i.e., no information and information in minutes, resulted in four scenarios to be investigated in this
case study.
An important aspect to be observed is that no prescriptive advice to choose a specific route was provided. Instead
information regarding the travel times of each route was provided. However, as the information was very accurate, we
Route 1: highways
Destination

Origin
Route 2: rural roads
Route 3: highways + urban roads

Fig. 3. Illustration of the characteristics of the routes in the experiment.

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

23

Table 2
Travel time distributions for each route.
Route Description

30 Draws: Gumbel distribution (35, 1)a 10 draws: Gumbel

distribution (70, 1)
All draws: Gumbel distribution (53, 1.25)
All draws: normal distribution (47, 12)

1
2
3
a

Mean travel
time (min)

Variance of travel
time (min)

10th percentile
(min)

90th percentile
(min)

44

233

34

70

53
47

1
146

52
33

54
56

The resulting travel times were randomly distributed over the 40 days.

Table 3
Scenarios investigated and number of travellers in each of them.
Travel purpose

Condition of information provision

No information

Travel time in minutes

Meeting with colleagues

Scenario 1
(20 travellers)

Scenario 3
(42 travellers)

Job interview

Scenario 2
(25 travellers)

Scenario 4
(46 travellers)

assumed that travellers would choose the fastest route, i.e., the route recommended to be the fastest. The travel time
distributions of each route and characteristics of the scenarios investigated are respectively depicted in Tables 2 and 3.
We assumed that during the initial 10 route choices travellers were getting experience about the routes' travel times. As
of the 11th day, we applied Prospect Theory to predict travellers' route choice behaviour. The participants, however, were
not informed that the initial 10 route choices counted as experience period. The idea behind the considered experience
period was to enable travellers to get to know what to expect from the travel times of each route. Then, as of the 11th day,
(in theory) travellers would start making choices following a specific route choice behaviour instead of random route
choices.
4. Model specication
We focus our investigation on the behavioural meaning of the reference point and on its influence on the ability of Prospect
Theory to capture travellers' behaviour. The influence of other parameters was not taken into account. As a result, the
parameters of the value and weighting function are set to the originally estimated values, i.e., 0.88, 2.25, 0.61 and
0.69 (Tversky and Kahneman, 1992). Nevertheless, as discussed in Van de Kaa (2010), despite the attempts to estimate
parameters for the travel behaviour context, the originally estimated parameters might still offer the best functional
description of travel choice under uncertainty. Therefore, in spite of not using specific parameters for the travel behaviour
context, the values adopted are so far considered to be the best estimates.
We propose a simple model specification in which travel time is the only attribute when no information was provided,
otherwise the travel information is the only attribute of the model. By a direct application of Prospect Theory, the predicted
route was determined based on the maximum prospect theoretical value (MaxPT) of travel times or information provided
and then compared with the route chosen by the travellers. The prediction ability of Prospect Theory was determined by the
number of correct predictions for each traveller in the experiment.
4.1. Considering no heterogeneity
Under the no heterogeneity model specification, we assume that all travellers have the same reference point and this value
is updated as travellers get more experience. During the experiment participants made daily route choices among three
routes and we assume that after each route choice travellers' expectations, and consequently their reference point, were
updated. In addition, after each trip the travel time distributions were updated taking the new travel times into account. As a
result, for choices made on the 11th day, data from the initial 10 days was taken into account. For choices made on the 12th
day, data from the initial 11 days was taken into account and so on.
We propose the following reference points for the situations in which no information was provided: (i) the mode of travel
times of the fastest route, (ii) the mode of travel times of the most reliable route, (iii) the average travel time of all routes and
(iv) the minimum travel time of the fastest route. In all cases, the reference point is based on the actual travel times, thus on
the draws. The first reference point implies that travellers derive satisfaction from getting the best probable result; the
second assumes that travellers are more conservative in their choices and tend to avoid losses, therefore behaving risk

24

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

averse; the third reference point does not have a clear behavioural meaning and aims to explore the type of choice predicted
under this circumstance; the fourth reference point is a specific group within the first and in behavioural terms mean that
travellers are (very) aggressive in their choices and only want the best possible route.
For the situations in which travel time information was provided, similar reference points were defined. However, instead
of referring to the travel times of the routes, they refer to the travel information. The consequence of this is that the
weighting function reflects probabilities of occurrence of specific travel information instead of probabilities of occurrence of
travel times. Similarly to the situations in which no information was provided, the reference point is based on the actual
information provided. In addition, a fifth reference point, the travel information itself, i.e., the predicted travel time of the
fastest route, is also proposed in order to investigate whether travellers update their reference based on the information.
Besides this, it is also investigated whether people base their decisions on the real travel times and use the information
solely as a reference, thus in the value function only. Under this condition the MaxPT is determined based on the real travel
times, and not on the travel information and a sixth reference point, travel information_2, is evaluated. Under the proposed
sixth reference point, when evaluating the weighting function we refer to probabilities of occurrence of travel times and
when evaluating the value function we refer to the travel information.
The MaxPT is determined based on Eqs. (3) and (4). In the value function, x measures the deviation from the reference
point (RP) and values x RPTT or RP-info depending on whether information is provided. TT is the real travel time and info
the travel information of a specific route. In the weighting function, p is the probability of occurrence of travel times or travel
information.
4.2. Accounting for heterogeneity
Under the model specification in which heterogeneity is taken into account, we argue that travellers' route choice
behaviour varies across the population. As a result, each traveller has its own reference point depending on its behavioural
characteristics. In addition, similar to the situation in which no heterogeneity is taken into account, we assume that
travellers update expectations and consequently also update their reference point after each route choice.
In order to reflect travellers' usual behaviour, and thus capture risky and conservative behaviours, when no information
was provided we propose that the reference point on day d is defined as: (i) the travel time of the most frequently chosen
route up to the previous day and (ii) the travel time of the most frequently chosen route in the previous 5 days. When
information was provided, the reference point on day d is defined as (i) the travel information of the most frequently chosen
route up to that day and (ii) the travel information of the most frequently chosen route in the previous five days.
For situations involving provision of information, the specifications of the value and weighting functions follow both
aforementioned definitions for the travel information and for the travel information_2, i.e., the weighting function accounts
for both the probabilities of occurrence of travel times and for travel information.
The main difference between the condition in which travellers' heterogeneity is considered and the condition in which it
is not considered is that in the first case each traveller is considered as a unique individual. As a result, each traveller has its
own behavioural characteristics. The consequence of this is that we are able to model travellers' preferences individually,
considering their preferences in all days of the experiment as well as 5 days prior to the choice made (and other situations
that could have been defined). The reference point is the element responsible to capture travellers' preferences in all
situations. Conversely, when no heterogeneity is considered, all travellers are assumed to behave similarly. Therefore, only
their general behaviour can be captured, i.e., the behaviour of the majority. As a result, it does not make sense to propose a
reference point that reflects short term switches in preferences such as condition (ii) above.
5. Results and discussion
The results presented in this section focus on two main aspects: travellers' behaviour and the role of the reference point
in the prediction ability of Prospect Theory. With respect to travellers' behaviour, we discuss their route choice behaviour
with respect to different sources of information. On the role of the reference point in the prediction ability of Prospect
Theory, we discuss the added value of accounting for heterogeneous reference points and aligning it to travellers' behaviour.
These aspects are discussed in Sections 5.1 and 5.2 respectively.
5.1. Travellers' behaviour in relation to information
Travellers' route choice behaviour in scenario 1 (meeting with colleagues+no information) show that during the experience
period, i.e., during the initial 10 days, routes 1, 2 and 3 were respectively chosen 62%, 18.5% and 19.5% of the times. After this
period, i.e., between the 11th and the 40th day, route 1 was chosen 40.5% of the times, route 2 was chosen 48.8% and route 3 was
chosen 10.7%. This indicates a sharp decrease in the preference of both routes 1 and 3 and a corresponding increase in the
preference of route 2. Implication of this is that after getting experience about the routes' travel times, travellers tend to be more
conservative (risk averse) in their choices, thus preferring the most reliable route.
In scenario 2 (job interview+no information), the percentage of travellers choosing the most reliable route in the experience
period was higher than in scenario 1. During the experience period, route 1 was preferred 50.9% of the times, route 2 was preferred
28.8% and route 3 was preferred 20.4%. During the remaining days, route 1 was chosen 34.5% of the times, route 2 was chosen

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

25

52.0% of the times and route 3 was chosen 13.5% of the times. The observed behaviour implies that the more serious the travel
purpose is the more conservative people tend to be in their choices. Therefore, when the travel purpose changed from meeting
with colleagues to job interview, the preference for route 2 increased along the experience period. In addition, it is possible to
observe that irrespective of the travel purpose, the route of intermediate performance was the least preferred route.
Travellers' route preferences in scenarios 1 and 2 appear to be aligned with findings about route reliability reported in
the literature, i.e., travellers value travel time reliability higher than travel time savings (Small et al., 2005; Liu et al., 2004).
Instead of choosing a route that is potentially faster, travellers prefer slower but reliable routes. Figs. 4 and 5 depict
travellers' route choice behaviour in scenarios 1 and 2.
Regarding travellers' reaction and adaptation to information, the behaviour under no information matched our expectations.
At first travellers explore the routes to get familiar with their travel times, but, as time evolves and travellers become more aware of
the route characteristics switching rates decrease (Fig. 6). For the sake of clarity, the reader should consider that a route switch is
Route 1

No. of times each route was

chosen

20

Route 2

Route 3

15

10

0
1

11

16

21

26

31

36

Day
Fig. 4. Travellers' daily route choices in scenario 1 (meeting with colleagues+no information).

No. of times each route was

chosen

20

Route 1

Route 2

Route 3

15

10

0
1

11

16

21

26

31

36

Day
Fig. 5. Travellers' daily route choices in scenario 2 (job interview+no information).

Scenario 1

100%

Scenario 2

Exploration rate

90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
2

10

14

18

22

26

30

34

38

Day
Fig. 6. Travellers' exploration rate in scenarios 1 and 2 (all travellers).

26

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

observed only if the traveller did not choose the same route as the day before (exploring routes) or as 2 days before (back to usual
behaviour).
When information was provided, three groups regarding compliance rate with information were identified: group 1
corresponding to around 45% of the travellers (20 travellers) for which compliance rates with the information were above
60%, group 2 corresponding to around 35% of the travellers ( 15 travellers) for which compliance rates with the information were
between 30% and 60% and finally group 3 corresponding to about 20% of the travellers (10 travellers) for which compliance rates
with the information were lower than 30%. We hypothesized that the switching rates of travellers that in general do not comply
with the advice (group 3) should be similar to the group without information. Fig. 7 shows a tendency of decrease in the
exploration rate up to 2/3 of the experiment and then a tendency to increase again. This is not entirely aligned with our
expectations as it seems that travellers restart exploring the routes. A possible explanation is that although participants in group 3
did not tend to comply with the information, they might have felt intrigued about the accuracy of the information and decided to
explore the routes in order to confirm their expectations. As a result a mixed behaviour arose at the end of the experiment, i.e., in
some occasions travellers complied with the advice and in others they trusted their own experience.
For groups 1 (high compliance rate) and 2 (medium compliance rate) we argued that as time evolves and travellers become
more confident about the reliability of the information, switching rates from the information should gradually decrease.
Compliance rates with the advice should therefore increase. Our expectations did not resemble the observed behaviour. Average
compliance rate with the advice for group 1 was around 85%, but very low compliance rates were also observed (Fig. 8). For
instance, although the lowest compliance rate occurred the day after a very wrong advice, i.e., actual travel times were
underestimated by more than 15 min, lower compliance rates were in general associated to situations in which the information
favoured route 3 (the least preferred route). The reason why route 3 was the least preferred may be related to the fact that it was
neither the fastest nor the most reliable route. For group 2, no pattern regarding compliance with information was identified as it
seems to resemble neither the travel information nor their own experience (Fig. 9). We believe, however, that this may be a
consequence of the design of the experiment in which the information was neither always correct nor always precise. By correct we
meant that the route indicated to be the fastest was actually the fastest and by precise that the indicated travel times were equal to
the experienced travel times. For instance, at times the information deviated between 10 and 30 min from the true travel times.
Nevertheless, a tendency to follow the information was observed (Fig. 10). In scenario 3, route 1 was the most often
recommended, followed by route 3 and finally route 2, and this was the observed pattern for the routes chosen (Fig. 10a).
The average compliance rate with the information during the whole experiment was 63%. In scenario 4, although the information
provided was almost the same as in scenario 3, compliance rates were around 55%, thus considerably lower. In addition, the

Scenario 3 (Group 3)

100%
90%

Scenario 4 (Group 3)

Exploration rate

80%
70%
60%
50%
40%
30%
20%
10%
0%
2

10

14

18

22

26

30

34

38

Day

Compliance rate with the advice

Fig. 7. Travellers' exploration rate in scenarios 3 and 4 (group 3).

Scenario 3: Group 1

Scenario 4: Group 1

100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
2

10

14

18

22

26

30

34

38

Day
Fig. 8. Compliance rate with information of group 1 in scenarios 3 and 4.

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

Compliance rate with the advice

Scenario 3: Group 2

27

Scenario 4: Group 2

100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
2

10

14

18

22

26

30

34

38

Day
Fig. 9. Compliance rate with information of group 2 in scenarios 3 and 4.

Percentage each route was

chosen

Advice 3
2
given
1
100%
80%
60%
40%
20%
0%
1

11

16

21

26

21

26

31

36

Day

Percentage each route was

chosen

Advice 3
2
given
1
100%
80%
60%
40%
20%
0%
1

11

16

31

36

Day
Fig. 10. Travellers' daily route choice in reaction to information in scenarios 3 (Fig 10a) and 4 (Fig 10b). Top part of each figure: the route which the advice
favoured (1, 2 or 3) day by day. Bottom part of each figure: travellers' route choice. Route 1 in black (bottom), route 2 in lilac (middle) and route 3 in dark
grey (top). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

observed pattern for the routes chosen did not fully resemble the pattern of the information provided as route 2 was chosen
more often than route 3 (Fig. 10b).
Similar to the scenarios in which no information was provided, there is a tendency to be more conservative in the choices
when the travel purpose is more serious. Moreover, when the information favoured travellers' preferred routes, the
compliance rate was much higher. For instance, as route 3 was the least preferred route, whenever the advice favoured
this route, compliance rates were (relatively) smaller than in other situations (Table 4). The fact that travel times on route 1
were usually smaller than on route 3 might have influenced the preference of route 1 over route 3. As a result, when higher
travel times on route 1 occurred these were perceived as exceptions. Conversely, higher travel times on route 3 might have
been perceived as a confirmation that this route was not such a good choice. Therefore, even when the information favoured

28

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

Table 4
Route choice preferences in scenarios 14.
Scenario

Travellers' route choice

Route 1 (%)

Compliance rate with advice (%)

Route 2 (%)

Route 3 (%)

40.5

48.8

10.7

34.5

52.0

13.5

3. Info in minutes+meeting colleagues

On average
Information favoured route 1
Information favoured route 2
Information favoured route 3

54.7
67.1
3.6
32.0

22.0
17.1
82.1
19.4

23.3
15.8
14.3
48.6

64.0
67.1
82.1
48.6

4. Info in minutes+job interview

On average
Information favoured route 1
Information favoured route 2
Information favoured route 3

45.4
55.5
1.1
31.3

28.4
24.3
87.0
23.9

26.2
20.1
12.0
44.8

55.0
55.5
87.0
44.8

Percentage of
correct predictions

1. No info+meeting colleagues
2. No info+job interview

100%

Scenario 1

Scenario 2

80%
60%
40%
20%
0%
Mode_Fastest

Mode_Reliable

Average

Min_Fastest

Percenatge of
correct predictions

Reference point

100%

Scenario 3

Scenario 4

80%
60%
40%
20%
0%
Mode_Fastest Mode_Reliable

Average

Min_Fastest

Info

Info_2

Reference point
Fig. 11. Role of the reference point under no heterogeneity. (a) Prediction ability of prospect theory in relation to the reference point in scenarios 1 and 2,
and (b) Prediction ability of prospect theory in relation to the reference point in scenarios 3 and 4.

route 3, there was some distrust with respect to its reliability. On the other hand, when the information favoured route 2, its
reliability was accentuated by the fact that route 2 by itself was already reliable. This explains the (very) high compliance rates.
5.2. Role of the reference point in the prediction ability of Prospect Theory
Results regarding the prediction ability of Prospect Theory are aligned with what has been observed in the literature and
reinforce the importance of establishing meaningful reference points. The need to align the reference point to travellers'
observed behaviour, thus accounting for heterogeneity, can be observed from the data analysis. Discussion regarding the
role of the reference point in situations accounting and not accounting for heterogeneity in travellers' behaviour is
respectively presented in Sections 5.2.1 and 5.2.2.
5.2.1. Considering no heterogeneity
In scenarios 1 and 2, in which no information was provided, travellers were more conservative in their choices and
tended to choose route 2 the most. As a result, when the reference point was set equal to the mode of travel times of the
most reliable route, the prediction ability of Prospect Theory was the highest. On the other hand, the lowest performance
was observed when the reference point was equal to the average travel time of all routes (Fig. 11a). As under this condition
route 3 was predicted the most, it is possible to infer that when the reference point is not based on a specific type of
behaviour, but on average values, routes of intermediate performance benefit from this. In addition, no differences between

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

29

setting the reference point as the mode or minimum travel time of the fastest route are observed. Despite the fact that each
of these reference points has its behavioural appeal, as their values were quite similar, differences in travellers' risky
behaviour could not be captured.
In scenarios 3 and 4, in which information was provided, the role of the reference point is more subtle because its
variability is smaller. Prospect Theory performed best when the reference point was set equal to travel information_2. This
allows inferring that people tend to rely on actual travel times and use the information solely as a reference of how good
their choice is (Fig. 11b). In addition, because route 2 was not often recommended, the worst performance occurred when
the reference point was equal to the mode of the most reliable route. The results of scenario 4 are somewhat disappointing
because the prediction ability of Prospect Theory was only slightly better than in scenario 2, in which no information was
provided, and inferior to that observed in scenario 3, in which only the travel purpose was different. Nevertheless, these
results were somehow expected because the route choices in scenario 4 were more homogeneous and as the experiment
progressed, route 1 was the only route predicted. For instance, as of the 20th day in 75% of the times the advice favoured
route 1, but travellers chose it only 49% of the times.
The Chi Square test was performed to test the hypothesis that improvements in the prediction ability of Prospect Theory
are due to the values adopted for the reference point. In the scenarios in which no information was provided, we tested
the significance of aligning the reference point to travellers' risk preferences. On the other hand, when information was
provided, we tested the significance of aligning the reference point to the information provided. These are the null
hypotheses tested.
The Chi Square test was performed taking into account the number of correct predictions of Prospect Theory associated
to each of these null hypotheses and some base values. The base values associated to different reference points are
depicted in the third and fourth columns of Table 5. The behavioural appeal of each reference point shown in the second
column of Table 5 reflect (i) the behaviour of the majority, i.e., the route chosen the most in each scenario, (ii) the behaviour
of the second most chosen route (Not aligned with the behaviour of the majority (1)), (iii) the behaviour of the route
chosen the least (Not aligned with the behaviour of the majority (2)) and (iv) the information provided. Items (i) and (iv)
respectively refer to the null hypothesis for the situations in which no information was provided and for the situations in
which information was provided. It is possible to observe that the Chi Square statistic considering one degree of freedom is
very significant and the significance level is inferior to 0.01 in all cases. As a result, from the outcomes of scenarios 1 and 2
we cannot reject the hypothesis that aligning the reference point to travellers' behaviour leads to significant improvement in
the prediction ability of Prospect Theory. This suggests that the reference point is aligned with travellers' behaviour. Based
on the outcomes of scenarios 3 and 4 we cannot reject the hypothesis aligning the reference point to the information
provided significantly improves the prediction ability of prospect theory. This suggests that travellers appear to use the
information as a reference point.
5.2.2. Accounting for heterogeneity
By considering each traveller as an unique individual, and consequently adopting distinct reference points for each of
them, the prediction ability of Prospect Theory (substantially) increases. This is particularly observed when no information
was provided (scenarios 1 and 2). Fig. 12 depicts comparisons between the best performance of Prospect Theory under no
heterogeneity with the performance of the proposed reference points when accounting for heterogeneity. It is possible to
observe that setting the reference point as defined in rp ii, i.e., the travel times or travel information of the most chosen
route in the 5 days prior to the route choice, definitely leads to major improvements in the prediction ability of Prospect
Table 5
Chi Square and significance levels for different reference points under no heterogeneity.
Scenario

Reference point

Correct
predictions

Wrong
predictions

Chi Square
a  b ()a

Chi Square
a  c ()a

Chi Square
a  d ()a

(a) Aligned with the behaviour of majority

(b) Not aligned with the behaviour of the majority (1)
(c) Not aligned with behaviour of majority (2)

286
239
113

314
361
487

7.48
(o 0.01)

112.37
( o 0.01)

(a) RP aligned behaviour majority

(b) RP not aligned with the behaviour of majority (1)
(c) Not aligned with the behaviour of majority (2)

383
259
141

367
491
609

41.87
(o 0.01)

171.76
( o 0.01)

(a) Equal to the information

(b) Aligned with the behaviour of majority
(c) Not aligned with the behaviour of majority (1)
(d) Not aligned with the behaviour of majority (2)

772
700
698
623

488
560
562
637

8.46
(o 0.01)

8.94
( o 0.01)

35.64
( o0.01)

(a) Equal to the information

(b) Aligned with the behaviour of majority
(c) Not aligned with the behaviour of majority (1)
(d) Not aligned with the behaviour of majority (2)

733
630
641
658

647
750
739
722

15.37
(o 0.001)

12.26
( o 0.001)

8.15
( o0.01)

Significance level.

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

Percentage of
correct predictions

30

100%

Without heterogeneity

Accounting for heterogeneity - rp ii

Accounting for heterogeneity - rp i

80%
60%
40%
20%
0%
1

Scenarios

Fig. 12. Prediction ability of Prospect Theory with and without accounting for heterogeneity.

Table 6
Chi Square and significance levels for different reference points accounting for heterogeneity.
Scenario

Reference point

Correct
predictions

Wrong
predictions

Chi Square
a  b ()a

Chi Square
a  c ()a

Chi Square
b  c ()a

(a) Accounting for heterogeneity: rp i

(b) Accounting for heterogeneity: rp ii
(c) Aligned with the behaviour of majority

329
398
286

271
202
314

16.61
( o0.001)

6.16
(o 0.01)

42.64
(o 0.001)

(a) Accounting for heterogeneity: rp i

(b) Accounting for heterogeneity: rp ii
(c) Aligned with the behaviour of majority

457
486
383

293
264
367

2.40
( 40.10)b

14.81
(o 0.001)

29.02
(o 0.001)

(a) Accounting for heterogeneity: rp i

(b) Accounting for heterogeneity: rp ii
(c) Equal to the information

773
841
772

487
419
488

7.96
( o0.01)

0.002
(4 0.50)b

8.20
(o 0.01)

(a) Accounting for heterogeneity: rp i

(b) Accounting for heterogeneity: rp ii
(c) Equal to the information

765
833
733

615
547
647

6.87
( o0.01)

1.49
(4 0.10)b

14.76
(o 0.001)

a
b

Significance level.
These values are not statistically significant.

Theory. The real benefit of accounting for heterogeneity may be questioned given that under the reference point as defined
in rp i (travel information of the most chosen route up to day d-1), improvements in scenarios 3 and 4 were not so
substantial. It was observed, however, that despite the fact that on average the number of correct predictions remained
about the same, an increase in the minimum amount of correct predictions per traveller from 2 to 8 was ascertained. This
contributes to the robustness of the model.
The underlying reason for a better performance of the reference point as defined in rp ii rather than as defined in rp i is
that the former allows a quicker adaptation of the reference point to a new route preference and thus to changes in
behaviour. For instance, if a traveller would have chosen route 1 for 10 consecutive days and as of the 11th decided to choose
route 2, the reference point as defined in rp ii would be aligned with the characteristics of the new route in 3 days, thus on
the 14th day. For the reference point as defined in rp i, however, this would be the case only as of the 21st day when the
traveller had driven on the new route for a period longer than on the previous preferred route.
In addition, the data analysis allows inferring that for travellers who appear to have a clear route choice behaviour rather
than random route choices, i.e., route switches are observed, but not such that every other day a different route is chosen,
accounting for heterogeneity in their behaviour is without doubt the reason of substantial improvements. The
characteristics of the routes, however, play an important role and should be very well defined in order to let the reference
point be able to capture travellers' risky preferences. For travellers that usually preferred route 3, which does not have
characteristics so well defined, very low prediction rates were observed. This suggests that not only the reference point
should be aligned with the observed behaviour, but also that the observed behaviour should be a good reflection of the
routes' characteristics.
The Chi Square test tested the hypothesis that improvements in the prediction ability of Prospect Theory are due to the
fact that the heterogeneity of the reference point is taken into account (Table 6). This is the null hypothesis tested for
reference point values as defined for rp i and rp ii. The behavioural appeal of each reference point shown in the second
column of Table 6 reflect (a) heterogeneity as defined for rp i (null hypothesis), (b) heterogeneity as defined for rp ii (null
hypothesis) and (c) behaviour of the majority, i.e., not taking heterogeneity or information provided into account. Reference
points aligned with the behaviour of the majority or to the information are the ones that under no heterogeneity led to the
highest prediction ability of Prospect Theory.
It is possible to observe that the Chi Square statistic considering one degree of freedom is very significant and the significance
level is inferior to 0.01 in the majority of the cases, but not in all of them (Table 6). For instance, we can reject the hypothesis that
improvements in the prediction ability of Prospect Theory in scenario 2 when considering the reference point as defined in rp ii is

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

31

significantly different from considering the reference point as rp i. The same holds for heterogeneity as defined in rp i and the
information provided in scenarios 3 and 4. These outcomes suggest that other factors, such as travel purpose, might be
influencing the better performance of Prospect Theory rather than the reference point.
6. Conclusions and future research
The behavioural meaning of the reference point of Prospect Theory and the role of heterogeneity in travellers' route
choice behaviour have been discussed in this paper. This was done by providing a behavioural interpretation to the
reference point and, with the support of a case study, comparing the results regarding the prediction ability of Prospect
Theory for different reference points and showing the added value of making use of heterogeneous reference points.
The use of Prospect Theory to model traveller behaviour is relatively new. The literature has made clear that there is a
lack of empirical experiments to validate its use for route choice behaviour, in particular due to the need to better
understand the reference point. The presented case study is one of the first attempts to propose a behavioural interpretation
to the reference point in a route choice context and to further investigate the role of heterogeneity in the reference point.
6.1. Conclusions
In line with literature, results show that the reference point considerably influences the prediction ability of Prospect
Theory, thus reinforcing the need of establishing meaningful values. Moreover, by accounting for heterogeneity in travellers'
behaviour, thereby considering that each individual traveller has its own reference point, considerable improvements in the
prediction ability of Prospect Theory are observed. The characteristics of the routes also appear to play an important role and
results suggest that not only the reference point should be aligned with traveller's observed behaviour, but also that the
observed behaviour should well reflect the routes' characteristics.
The Chi Square tests performed to test whether improvements in the prediction ability of Prospect Theory were due to aligning
the reference point to some behavioural premises provided significant results. For situations in which no heterogeneity was
considered, the null hypotheses tested were that aligning the reference point to travellers' risk preferences and to the information
provided lead to significant improvements in the prediction ability of Prospect Theory. These null hypotheses respectively refer to
the scenarios in which no information was provided and to the scenarios in which information was provided. When heterogeneity
was taken into account, we tested the null hypothesis that taking into account travellers' heterogeneous behaviour in the reference
point leads to significant improvements in the prediction ability of Prospect Theory.
The results indicate that we cannot reject the null hypotheses under no heterogeneity, but that one should be careful
with respect to heterogeneity. For instance, if we are able to easily/quickly identify travellers' preference changes and reflect
these changes in the values of the reference point (for example the switch of preferences from a fast and unpredictable route
to a slower and reliable route), we cannot reject the null hypothesis regarding the importance of taking heterogeneous
behaviour into account. Conversely, heterogeneity in travellers' behaviour does not seem to play an important role.
These outcomes suggest that application of Prospect Theory can highly benefit from the use of meaningful reference
points that reflect (i) travellers' heterogeneity, (ii) quicker adaptation of preferences over time, and (iii) travellers' (risk)
preferences. In addition, it has to be observed that a main drawback of not taking heterogeneity into account is that
irrespective of travellers' preferences, either for reliable and slower or unreliable and faster routes, only the behaviour of the
majority can be captured.
Regarding the use of information, results show that despite its potential to influence travellers' behaviour, compliance
rates vary a lot among travellers. Nevertheless, travellers' route choices are more clustered when information was provided,
in particular when the information was aligned with travellers' preferences for a specific route.
6.2. Implications
The main issue underlying the compliance rate with information is determining the extent to which this impacts
network dynamics. For instance, given the high compliance rates in some situations, information provision might influence
travellers to completely switch to alternative routes which would cause a considerable impact in network dynamics.
With respect to the role of heterogeneity in the reference point the results indicate the importance of properly relating travellers'
risky-preferences to the reference point. Given the characteristics of our case study, we were able to properly translate the travel
times of routes 1 and 2 in reference points which were coherent with travellers' observed behaviour/risky-preferences. This,
however, was not possible for route 3 due to its mixed characteristics. The consequence of this is that establishing meaningful
reference points for routes with mixed characteristics, such as urban roads, is not straightforward and makes application of Prospect
Theory more difficult.
6.3. Future research
Despite the potential impact of information provision on network dynamics, further conclusions should be drawn
carefully. This is due to aspects such as travellers' familiarity with the routes and distances travelled as these are factors that

32

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

might also influence compliance with information. These were not part of the analysis presented in this paper, but should be
observed in future research.
Due to the important role of the reference point, both its behavioural appeal and the role of heterogeneity should be
further investigated with the support of more empirical experiments. The same holds with respect to the other parameters
of Prospect Theory. This will help validating the suitability of Prospect Theory to model travellers' behaviour.
Another aspect to be observed is that although we considered a 10 day experience period in which travellers would be
able to explore the routes and get familiar with their characteristics, this paper does not discuss learning. This is because we
focused on investigating the role of heterogeneity in the reference point rather than discussing the role of information on
learning mechanisms. Further analysis regarding the role of learning mechanisms is a topic that should be addressed.
Regarding compliance with information, we argued that non-compliers should behave similar to non-informed travellers.
It is not entirely clear, however, why this behaviour was only partially observed (up to 2/3 of the experiment). Future
research should investigate reasons for disgruntlement of non-compliers, i.e., if this is due to incorrect advices at the
beginning of the experiment or whatever else might be influencing this type of behaviour.

Acknowledgements
This research is financed by the Dutch Organization of Science (NWO) under the Programme TRISTAM: Traveller
Response and Information Service Technology: Analysis and Modelling. We would like to thank Dr. Enide Bogers for making
available the data used in the case study of this paper and the anonymous reviewers for their valuable comments which
helped improving the quality of our paper.
References
Allais, M., 1953. Le comportement de lhomme rationnel devant le risque: critique des postulats et axiomes de lEcole Americaine. Econometrica 21 (4),
503546.
Arnott, R., de Palma, A., Lindsey, R., 1992. Route choice with heterogeneous drivers and group-specific congestion costs. Regional Science and Urban Economics 22,
71102.
Avineri, E., 2006. The effect of reference point on stochastic equilibrium. Transportation Science 40 (4), 409420.
Avineri, E., 2009. The fuzzy meaning of reference-based perceptions in travel choice modeling. In: Proceedings of the 88th Transportation Research Board
Annual Meeting. Washington, DC.
Avineri, E., Bovy, P.H.L., 2008. Identification of parameters for a Prospect Theory model for travel choice analysis. Transportation Research Record 2082,
141147.
Avineri, E., Prashker, J.N., 2003. Sensitivity to uncertainty. Need for a paradigm shift. Transportation Research Record 1854, 9098.
Ben-Akiva, M., Lerman, S.R., 1985. Discrete Choice Analysis: Theory and Application to Travel Demand. MIT Press, Cambridge, MA.
Ben-Elia, E., Shiftan, Y., 2010. Which road do I take? A learning-based model of route-choice behaviour with real-time information. Transportation Research
Part A 44, 249264.
Bogers, E.A.I., 2009. Traffic Information and Learning in A Day-to-Day Route Choice. TRAIL Thesis Series. T2009/5, June 2009. The Netherlands.
Bovy, P.H.L., Bliemer, M.C.J., van Nes., R., 2006. Transportation Modeling: Lecture Notes. CT4801. Delft University of Technology, location is Delft, the
Netherlands.
Connors, R.D., Sumalee, A., 2009. A network equilibrium model with travellers' perception of stochastic travel times. Transportation Research Part B 43,
614624.
De Palma, A., Picard, N., 2005. Route choice decision under travel time uncertainty. Transportation Research Part A 39, 295324.
De Palma, A., Ben-Akiva, M., Brownstone, D., Holt, C., Magnac, T., McFadden, D., Moffatt, P., Picard, N., Train, K., Wakker, P., Walker, J., 2008. Risk, uncertainty
and discrete choice models. Marketing Letters 19, 269285.
De Borger, B., Fosgerau, M., 2008. The trade-off between money and time: a test of the theory of reference-dependent preferences. Journal of Urban Economics 64
(1), 101115.
European Comission, 2008. Communication from the Commission: The Action Plan for the Deployment of Intelligent Transport Systems in Europe. COM
(2008), p. 0886.
Gao, S., Freijinger, E., Ben-Akiva, M., 2010. Adaptive route choices in risky traffic networks: a prospect theory approach. Transportation Research Part C 18, 727740.
Gopinath, D.A., 1995. Modeling Heterogeneity in DISCRETE CHOICE Processes: Application to Travel Demand. PhD Thesis. Massachusetts Institute of Technology.
Henn, V., Ottomanelli, M., 2006. Handling uncertainty in route choice models: From probabilistic to possibilistic approaches. European Journal of Operational
Research 175, 15261538.
Hess, S., 2007. Posterior analysis of random taste coefficients in air travel behaviour modelling. Journal of Air Transport Management 13, 203212.
Hess, S., A., Stathopoulos, A., Daly, 2011. Allowing for heterogeneous decision rules in discrete choice models: an approach and four case studies. In:
Proceedings of the 2nd International Choice Modelling Conference. July 4th6th, 2011, Leeds, UK.
Kahneman, D., Tversky, A., 1979. Prospect Theory: an analysis of decision under risk. Econometrica 47 (2), 263291.
Koppelman, F.S., Sethi, V., 2005. Incorporating variance and covariance heterogeneity in the Generalized Nested Logit Model: an application to modeling
long distance travel choice behavior. Transportation Research Part B 39, 825853.
Kristensen, H., Grling, T., 1997. Adoption of cognitive reference points in negotiations. Acta Psychologica 97, 27288.
Liu, H.X., Recker, W., Chen, A., 2004. Uncovering the contribution of travel time reliability to dynamic route choice using real-time loop data. Transportation
Research Part A 38, 435453.
Noland, R.B., Small, K.A., 1995. Travel time uncertainty, departure time choice, and the cost of the morning commute. Transportation Research Record 1493, 150158.
Ossen, S., Hoogendoorn, S., 2011. Heterogeneity in car-following behavior: theory and empirics. Transportation Research Part C 19 (2), 182195.
Polak, J.W., S., Hess, X., Liu, 2008. Characterizing heterogeneity in attitudes to risk in Expected Utility models of mode and departure time choice.
In: Presented at the 87th Transportation Research Board Annual Meeting. Washington, DC.
Prelec, D., 1998. The probability weighting function. Econometrica 66 (3), 497527.
Ramos, G.M., Daamen, W., Hoogendoorn, S., 2011. Comparison of Expected Utility Theory, Prospect Theory, and Regret Theory for prediction of route choice
behavior. Transportation Research Record 2230, 1928.
Schul, Y., Mayo, R., 2003. Searching for certainty in an uncertain world: the difficulty of giving up the experiential for the rational mode of thinking. Journal
of Behavioral Decision Making 16, 93106.

G. de M. Ramos et al. / Journal of Choice Modelling 6 (2013) 1733

33

Schwanen, T., Ettema, D., 2009. Coping with unreliable transportation when collecting children: examining parent's behavior with Cumulative Prospect
Theory. Transportation Research Part A 43 (5), 511525.
Senbil, M., Kitamura, R., 2004. Reference points in commuter departure time choice: a prospect theoretic test of alternative decision frames. Journal of
Intelligent Transportation Systems 8, 1931.
Small, K.A., Winston, C., Yan, J., 2005. Uncovering the distribution of motorists' preferences for travel time and reliability. Econometrica 73 (4), 13671382.
Srinivasan, K.K., Mahmassani, H.S., 2003. Analyzing heterogeneity and unobserved structural effects in route-switching behavior under ATIS: a dynamic
Kernel Logit Formulation. Transportation Research Part B 37, 793814.
Starmer, C., 2000. Developments in Non-Expected Utility Theory: the hunt for a descriptive theory of choice under risk. Journal of Economic Literature 38 (2),
332382.
Sumalee, A., R., Connors, R., Batley, 2005. The applicability of Prospect Theory to the analysis of transport networks. In: Proceedings of the European
Transport Conference (ETC). Strasbourg, France.
Sumalee, A., Connors, R.D., Luathep, P., 2009. Network equilibrium under Cumulative Prospect Theory and endogenous stochastic demand and supply,
Transportation and Traffic TheorySpringer Science+Business Media.
Tversky, A., Kahneman, D., 1992. Advances in Prospect Theory: cumulative representation of uncertainty. Journal or Risk and Uncertainty 5 (4), 297323.
Van de Kaa, E.J., 2008. Extended Prospect Theory. Findings on Choice Behaviour from Economics and the Behavioural Sciences and Their Relevance for
Travel Behaviour, Ph.D., TU Delft.
Van de Kaa, E.J., 2010. Prospect Theory and choice behaviour strategies: review and synthesis of concepts from social and transport sciences. European
Journal of Transport and Infrastructure Research 10 (4), 299329.
Von Neumann, J., Morgenstern, O., 1947. Theory of Games and Economic Behavior. Princeton University Press, Princeton.
Weber, M., Camerer, C., 1987. Recent developments in modelling preferences under risk. OR Specktrum 9, 129151.
White, S.B., Valley, K.L., Bazerman, M.H., Neale, M.A., Peck, S.R., 1994. Alternative models of price behavior in dyadic negotiations: market prices, reservation
prices, and negotiator assumptions. Organizational Behavior and Human Decision Processes 57, 430447.
Wu, G., Gonzalez, R., 1996. Curvature of the probability weighting function. Management Science 42 (12), 16761690.
Zhang, J., Z., Wang, H., Timmermans, A., Fujiwara., 2010. Integrating the concepts of relative utility and Prospect Theory curvature to represent context
dependence in travel choice Behavior. In: Proceedings of the 12th World Conference on Transport Research. Lisbon, Portugal.