Vous êtes sur la page 1sur 18

Theoretical Issues in Ergonomics Science

Vol. 13, No. 5, SeptemberOctober 2012, 528545


Modelling driver behaviour: a rationale for multivariate statistics
J.C.F. de Winter
*
and Riender Happee
Department of BioMechanical Engineering, TU Delft, Delft, The Netherlands
(Received 18 June 2009; final version received 6 December 2010)
This article provides a discussion of driver behaviour models. Based on a
literature survey and using experimental driving simulator data, three types of
driver behaviour models are evaluated: motivational models, adaptive control
models and trait models. It is shown that motivational models are comprehensive
but lack specificity. Many adaptive control models yield precise results but tend
to be overly specific and psychologically implausible. Trait models have not been
successful in the past and have been criticised for providing post hoc explanations
and for the absence of a multifactorial structural approach. We demonstrate that
if the right multivariate statistical techniques are used, trait models are suitable
for the parsimonious modelling of predictive-valid driver characteristics, thereby
taking an intermediate position on a dimension ranging from specific to
unspecific models.
Keywords: driver behaviour models; driver assessment; motivational models;
multivariate statistics; driving simulator
1. Introduction
Sensor systems onboard contemporary vehicles record vast amounts of data, providing
new possibilities for personalised driver assistance. According to Carsten (2007), the
potential of a system that understands the driver would be huge. Based on the real-time
monitoring of driver state and performance, it could give feedback and assistance to the
driver and adapt the operation of the vehicle according to the drivers needs in order to
improve road safety. The challenge faced by the researchers is to translate the raw sensor
data into a driver profile that represents the drivers characteristics, which potentially vary
over time. A driver behaviour model is a prerequisite to assist in this data processing.
During the last decades, an enormous variety of driver behaviour models have been
proposed. This article provides a review using the familiar classification of Michon (1985;
Figure 1), who proposed a two-dimensional arrangement (behaviourally oriented vs.
psychologically oriented, and taxonomic vs. functional) to define four basic types of driver
behaviour models: (1) task analyses, (2) trait models, (3) mechanistic/adaptive control
models and (4) motivational/cognitive models.
1
A taxonomic model is an inventory of
facts; the relationships are those of pertinent structure, for example, subordination,
identification, sequential relations, proportions and likelihood. In contrast, functional
models contain components which interact dynamically. Inputoutput models represent
*Corresponding author. Email: j.c.f.dewinter@tudelft.nl
ISSN 1463922X print/ISSN 1464536X online
2012 Taylor & Francis
http://dx.doi.org/10.1080/1463922X.2010.546437
http://www.tandfonline.com
observable behaviour and mathematical functions, whereas psychological models are
based on presumptions regarding the processes taking place inside the head of the driver.
In this study, we characterise driver behaviour models on a dimension ranging from
specific to unspecific (Figure 2). We show that motivational models tend to be unspecific,
suffering from a vagueness which makes them unfalsifiable. Adaptive control models often
fall into the opposite trap by being overly specific, fitting random patterns, while lacking
predictive power. We argue that driver behaviour models should be parsimonious and that
trait models may well be the most promising for driver behaviour modelling, taking an
intermediate position on the specificunspecific dimension. We illustrate our findings
using the literature and results from driving simulator experiments.
2. Motivational models
Motivational driver models make explicit assumptions about the drivers mental state
(Michon 1985). A chronological overview of motivational models was provided by Vaa
(2007), starting with the safe travel concept of Gibson and Crooks (1938) and ending with
the task difficulty model of Fuller (2005).
Risk homeostasis theory (RHT; Wilde 1988) is perhaps the most familiar motivational
model. RHTposits that a driver acts on the basis of a target level of accident risk. This target
risk level is determined by four utility factors: (1) the expected advantages of risky
behaviours, (2) the expected costs of these, (3) the expected benefits of cautious behaviours
and (4) the expected costs of these. At any moment in time, the road user compares his or
her personal target level of risk with the level of risk experienced or anticipated and attempts
to reduce any difference to zero. The theory further states that whenever a technical,
educational or other intervention is introduced that does not alter the target level
of risk, short-term fluctuations in the traffic accident loss per capita may occur,
but these will ultimately be eliminated after a short time, such that the crash rate
returns to the pre-intervention level (Wilde 1988). RHT has triggered extensive
Taxonomic Functional
Input-output
(Behavioural)
Task analyses
Mechanistic models
Adaptive control models
-Servo-control
-Information flow control
Internal state
(Psychological)
Trait models
Motivational models
Cognitive (process) models
Figure 1. Summary of driver behaviour model types.
Source: Michon (1985, p. 490; Figure 3). Reprinted with the kind permission of Springer Science and
Business Media.
Specific
Quantitative
Covers details of the driving task
Overfitted
Focuses on the vehicle dynamics
Unspecific
Qualitative
Covers the entire driving task
Unfalsifiable
Focuses on the human mind
Figure 2. Dimension of driver behaviour models, ranging from specific to unspecific models.
Theoretical Issues in Ergonomics Science 529
debates (Wilde et al. 2002, Cacciabue 2007) and, although criticised (ONeill and Williams
2004), it has been successful in bringing the concept of behavioural adaptation and the use
of driver incentives to the attention of road safety researchers (Trimpop 1996).
A more recent and well-cited motivational model is Fullers (2005) task-capability
interface (TCI). TCI describes the interaction between the determinants of task demand
and driver capability. The task demand is determined by factors such as the environment,
other road users and speed, with capability being determined by training, education and
experience. Task difficulty homeostasis is proposed as a key subgoal in driving, and the
choice of speed is argued to be the main solution to the problem of keeping task difficulty
within driver-preferred bounds.
Motivational models take the form of a comprehensive theory, often covering the
entire driving task. Critics of motivational models, however, have pointed to an
overreliance on confirmation rather than refutation, lack of specificity (Ranney 1994)
and impreciseness (Rothengatter 2002). An important concern is underdetermination,
which means that two rival motivational models are inconsistent with each other but both
consistent with the available evidence. For example, RHT has been found to compete with
zero-risk theory, which states that subjective risk is zero most of the driving time
(Na a ta nen and Summala 1974). Similarly, Janssen and Tenkink (1988) stated that
phenomena presented as supportive of RHT can also be adequately explained in
any utility-maximisation model (as proposed earlier by e.g., Blomquist (1986) as well as
ONeill (1977)).
Another uncertainty has been whether RHT applies to individual drivers or to a group
of drivers (Michon 1985, Huguenin and Rumar 2001). It is important to accurately
distinguish between individual and collective behaviours. The speedaccuracy relationship,
for instance, is familiar as a trade-off made by an individual. However, between subjects,
a positive correlation has been found. That is, those who react more quickly are also
generally those who make fewer errors (Jensen 2006).
As a response to the limitations of motivational models, hierarchical models have been
introduced, which Ranney (1994) referred to as the second-generation motivational
models (p. 747). The literature contains several hierarchies of driving behaviour, usually
of three levels (Rasmussen 1983, Michon 1985, Van der Molen and Bo tticher 1988) but
also of four levels (Hatakka et al. 2002, Hollnagel et al. 2003, Panou et al. 2007).
Researchers have also proposed the combining of various hierarchies and taxonomies into
a two-dimensional matrix (Hale et al. 1990) or a three-dimensional cube (Summala 1996,
Theeuwes 2001). These hierarchies work well as qualitative concepts, but there is no
quantitative support in the literature.
2.1. Experiment 1. Unfalsifiability
We re-used data from a previous driving simulator study (De Winter et al. 2006) to test
both RHT and TCI. The experiment originally aimed to explore how participants change
their behaviour during a lane-keeping task in terms of speed and accuracy, and to
investigate what happens to their perception in relation to aspects, such as pleasure, risk
and task difficulty. In this context, we investigated whether RHT and TCI hold during the
lane-keeping task. We predicted constant levels of task difficulty and risk because the TCI
states that drivers are motivated to maintain a preferred level of task difficulty (Fuller
2005, p. 467) and to keep task difficulty within selected boundaries (p. 467),
530 J.C.F. de Winter and R. Happee
whereas RHT states that drivers aim to maintain a target constant level of risk (which is
expected to be constant in the short term in a constant environment).
The experiment was conducted in a fixed-base driving simulator (Green Dino BV).
A total of 10 participants (8 men, 2 women; age range 1830 years) without previous driving
experience completed four identical sessions. In each session, participants drove a 7.6-km
lap on a two-lane rural road with 25 bends of radii in the order of 1030 m. The lane width
was 5 m. Participants had to steer, accelerate and brake by themselves; gear-changing was
automatic. The participants were instructed to drive as well as they could, keeping in the
right lane with both hands on the steering wheel, and to adopt a speed of 2030 km/h in
sharp corners. No other vehicles were present and no feedback was provided. Thus, the
experiment was a simple self-paced training session in the elementary task of driving a car
around a track.
The dependent variables were the mean speed and the standard deviation of lateral
position (SDLP; a measure of lane-keeping accuracy). These variables were calculated
separately for each road segment and subsequently averaged.
After each session, participants completed a written questionnaire, consisting of a
series of questions on a 17-point scale. The selected questions under this investigation were
(translated from Dutch): (1) How difficult did you find the task during the preceding
session?, from 0 (very easy) to 16 (very difficult). This question had to be answered with
respect to straights, normal turns, sharp turns and mild turns separately, and was averaged
into one task difficulty level, and (2) According to your own feeling, how high did you
regard your risk level during the preceding session?, from 0 (high) to 16 (low). Afterwards,
these measures were transformed into a scale running from 0% (very easy/low risk) to
100% (very difficult/high risk).
Figure 3 shows the speed-accuracy ratio (mean speed divided by SDLP) of the 10
participants as a function of the session number. The participants improved on this
1 2 3 4
20
40
60
80
100
120
140
160
180
200
Session
S
p
e
e
d
-
a
c
c
u
r
a
c
y

r
a
t
i
o

(
1
/
3
.
6
s
)
Figure 3. Speed-accuracy ratio (mean speed [km/h] divided by the SD of lateral position [m]) of the
10 participants as a function of session number.
Theoretical Issues in Ergonomics Science 531
aggregate indicator (t 4.52, p 0.001, comparing sessions 1 and 4 using a paired t-test).
Figure 4 shows the results of the questionnaire. Task difficulty decreased (t 6.18,
p50.001) as well as participants sense of risk (t 2.90, p 0.020).
The results of the experiment showed that subjective risk and subjective task difficulty
decreased with increased experience, which disagreed with predictions of constant levels
of risk and task difficulty. Colleagues shown these results were rather sceptical. Responses
included: Has anyone ever stated that RHT applies during first learning experiences? I
could imagine that learner drivers are still building a perception of risk and develop a
target risk. Moreover they could pursue a constant risk, but may initially fail to realise the
target due to inability. Some also questioned whether RHT and TCI applied in simulated
environments. Furthermore, it was argued that TCI does not specify how wide the range of
acceptable task difficulty can be, and that RHT does not explicitly state that risk should
always be constant; in fact, the target level of risk allows variations according to a cost
benefit approach.
In other words, it seemed quite easy to save the motivational models from
falsification. Our colleagues responses highlighted the problem of motivational models
being unfalsifiable, violating a critical aspect of scientific modelling (see Huguenin 1988,
Ranney 1994, Elvik and Vaa 2004, for similar comments regarding RHT and other
motivational models). This makes it difficult to establish trust in the value of these
motivational models for driver assessment.
3. Adaptive control models
Adaptive control models range from microscopic engineering models (Brackstone and
McDonald 1999) and manual control models in the frequency or time domain
1 2 3 4
0
20
40
60
80
100
S
c
o
r
e

(
%
)
Feeling of risk
Task difficulty
Session
Figure 4. Mean questionnaire responses as a function of session number. The error bars are shown
at 1 SD from the mean. The results have been offset slightly on the horizontal axis so that overlap
of the error bars can be seen more clearly.
532 J.C.F. de Winter and R. Happee
(McRuer and Jex 1967, McRuer et al. 1977), to complex computational simulations of
traffic situations (Cacciabue 2007). According to Michon (1985), adaptive control models
can be subdivided into classic control models (involving signals that are continuous in
time) and information-flow control models (involving discrete decisions). Adaptive control
models are used for driver assessment through parameter identification techniques (Boer
et al. 2005). Overviews of adaptive control models are provided by Bekey et al. (1977),
Brackstone and McDonald (1999), Guo and Guan (1993), Jagacinski and Flach (2003),
MacAdam (2003), Plo chl and Edelmann (2007) and Reid (1983).
In contrast to motivational models, adaptive control models provide precise quanti-
tative results. However, the engineering-oriented character of adaptive control models is
not supported by all behavioural researchers. This became apparent after a publication on
mathematical car-following models by Brackstone and McDonald (1999) in the journal
Transportation Research Part F: Traffic Psychology and Behaviour. The article generated
four commentaries (Boer 1999, Hancock 1999, Ranney 1999, Van Winsum 1999). Ranney
(1999) criticised Brackstone and McDonald for making detailed assumptions that were not
well-founded from a human perspective. Many factors that are known to influence car-
following, such as weather, road conditions, age, gender, motivations and strategic
aspects, were found not to be covered in the mathematical models. Hancock (1999) was
critical of the mathematical fitting procedure: Deriving equations from physical
descriptions of motion and subsequently trying to fit these to data derived from
behavioral response both literally and figuratively, puts the cart before the horse
(Hancock, p. 198). Boer (1999) considered that the aspect of individual differences was
overlooked: Psychologically mediated situation dependent within and between driver
variability is a likely cause for the lack of agreement between the multitude of car-
following models (Boer 1999, p. 205) and that most driver models assume that drivers
have access to variables that are mathematically convenient but perceptually implausible
(Boer et al. 2000).
A common characteristic of many adaptive control models is that despite being able
to accurately fit the measured data they are not good at making predictions. The problem
of overfitting was also recognised by Boer (1999) when commenting on Brackstone and
McDonalds review (1999): Experience tells us that the adopted modeling abstraction has
been pushed too far as exemplified by the fact that the identified coefficients of the same
model differ considerably between experiments (Boer 1999, p. 202). Indeed, adequate fit
with the observed data is not a sufficient condition for model quality (Roberts and Pashler
2000, Pitt et al. 2002, Ginzburg and Jensen 2004). Researchers are not only interested in
how well a model can describe a process but also in how well the model, with its
parameters constrained to fixed values can predict other (e.g. future) data. Models of
higher complexity (more parameters, functions, connections) are more flexible in
describing data, but at the price of reduced predictive validity. For example, the optimal
control model of human behaviour (Kleinman et al. 1970) contains many free parameters
which hampers its usefulness and predictive ability (Mulder 1999).
Even though the risk of overfitting is disconcerting, many studies have actually
recommended increasing the level of complexity. For example, McRuer et al. (1977)
discussed extension and refinement, such as including model components for driver
decisions and driver judgement factors and interactions. Guo and Guan (1993)
recommended better preview and correction strategies in order to render better tracking
accuracy, while Brackstone and McDonald (1999) suggested extending the car-following
models to include motivational and attitudinal factors. MacAdam (2003) concluded that
the models should include more functionality such as driver sensory input processing
Theoretical Issues in Ergonomics Science 533
elements, and according to Michon (1985), a model that displays a sufficiently broad
spectrum of realistic driving behaviours will be inherently complex and will embody at
least between 5000 and 10,000 elements. We argue that a driver behaviour model should
not aim to predict exactly what a driver is doing at any moment. Because of the numerous
design factors and random influences involved in driving, this is probably a meaningless
or impossible endeavour (Carsten 2007).
A recent convincing application of an adaptive control model was provided by
Sheridan (2004). He found that using a simple adaptive control model led to a
refined conceptual understanding of driver distraction and improved modelling and
prediction of driver performance as a function of motorway and vehicle design. Sheridan
recognised the importance of model simplicity: Unlike several complex qualitative models
of distraction causality and effects on performance, the proposed control model is
sufficiently simple and explicit to have a degree of predictive capability (p. 588). Other
successful applications of adaptive control models have been for the simplest of tasks, such
as car-following, curve negotiation and regulation against wind gusts (Allen et al. 2005,
Boer et al. 2005).
3.1. Experiment 2. Inter- and intra-individual variability
Adaptive control models act on an external disturbance and/or target signal, such as a
vehicle in front driving with variable speed (in a car-following task), or lateral wind gusts
or a desired vehicle path curvature (in a lane-keeping task). Adaptive control models
typically cannot function in the absence of such external input.
Data from a previous, unpublished car-following experiment in the same type of fixed-
base driving simulator as Experiment 1 were reanalysed. Ten licensed drivers participated
(2 women, 8 men; age range 1834 years). The initial aim of the experiment was to
investigate driver attention and performance during a 1 h car-following task. Participants
were instructed to follow a vehicle at a constant comfortable distance. The lead vehicle had
a constant speed of 100 km/h and no other cars were present in the virtual world. The
experiment took place on an endless straight road with constant lead vehicle speed, so the
virtual situation can be considered stationary. Background music played during
the experiment. Half of the experiment per participant involved a secondary task
(rhythmic tapping), but this had only a small effect on driver performance and behaviour.
Figure 5 shows that there were considerable inter- and intra-individual differences in the
following distance. That is, the mean following distance differed between participants, and
the SD of following distance indicates that individual participants did not drive at a
constant speed either; low-frequency variations and drifts were apparent in the following
distance.
The bottom line is that there is a large random variability between and within drivers,
even without external disturbances due to infrastructure and other vehicles. It seems
fruitless to use adaptive control models to explain this type of variability at the vehicle
control level. Such efforts will overfit the data.
3.2. Experiment 3. Longitudinal and lateral control
Many adaptive control models adopt a clearcut distinction between longitudinal and
lateral driving behaviour (MacAdam 2003). From the standpoint of modelling the physical
534 J.C.F. de Winter and R. Happee
vehicular response, such a distinction is justified. However, when attempting to describe
individual differences, this distinction may not be appropriate.
Data were used from a previous car-following experiment in a driving simulator
(De Winter et al. 2008). Fourteen participants (12 men, 2 women, age range 2351 years,
all with car driving experience) followed a preceding car in heavy traffic in four 12 min
sessions on a motorway, while testing various configurations of an accelerator force
feedback system. Analysis of a correlation matrix of performance measures revealed that
participants who were more active with the brake pedal were more active with the steering
wheel as well (r 0.95, p50.001; Figure 6). Note that the particular car model was
programmed for motorway use only and had independent lateral and longitudinal vehicle
dynamics. Hence, although the sample size was limited, this study showed that a
behavioural factor linked longitudinal and lateral control behaviour.
40 50 60 70 80 90 100 110 120 130
10
15
20
25
30
35
40
45
50
55
Mean following distance (m)
S
D

o
f

f
o
l
l
o
w
i
n
g

d
i
s
t
a
n
c
e

(
m
)
Figure 5. SD of following distance vs. mean following distance of 10 participants who completed
a car-following experiment.
0 2 4 6 8 10 12 14
0.100
0.102
0.104
0.106
0.108
0.110
0.112
0.114
0.116
SD of brake pedal position (%)
S
D

o
f

s
t
e
e
r
i
n
g

w
h
e
e
l

a
n
g
l
e

(

)
Figure 6. SD of steering-wheel angle vs. SD of brake-pedal position of 14 participants who followed
a preceding car while maintaining the right lane on a motorway.
Theoretical Issues in Ergonomics Science 535
In summary, attempts to explain driver behaviour using adaptive control models of
vehicle control can lead to an inadequate description of the data. Underlying traits such as
driver aggression could explain the observable behaviours more parsimoniously.
4. Trait models
Trait models describe relationships between driver characteristics without incorporating
dynamic functions. Generally, trait models have historically concerned the statistical
identification of the accident-prone driver, a concept which was proclaimed dead about
50 years ago (Haight 1964). Michon (1985) was critical of trait models and stated that
empirical connections are at best correlative. Ranney (1994) provided a review of the study
of individual differences in road traffic crashes and concluded that the differential crash
involvement paradigm should be abandoned. Ranney was particularly critical about post
hoc explanations and the absence of a multifactorial structural approach: The use of
simple correlational methods without multifactorial structural models raises questions
about the meaning of significant correlations (Ranney, p. 736).
Despite the lack of success of trait models in the past, statistical models and the
identification of accident-prone drivers have continued to receive positive attention
(Af Wa hlberg and Dorn 2009). Elvik and Vaa (2004) explained that predicting car crashes
is inevitably a statistical procedure. According to Carsten (2007), understanding driving
requires a statistical model rather than a deterministic model that attempts to explain
exactly what a driver does at a certain moment. Similarly, Rothengatter (2002) stated that
the study of individual differences can provide a basis for driver training and accident
prevention.
We propose exploratory factor analysis (EFA) as a starting point for the statistical
analysis of driver behaviour. EFA uses the correlations between variables to reduce the
data to a small number of common factors. The proposed use of EFA for studying driver
behaviour is neither new nor surprising. EFA (or principal component analysis, a similar
technique) has been applied in Driver Behaviour Questionnaire responses, resulting in the
invention of violation and error factors (Reason et al. 1990, Blockey and Hartley 1995,
A

berg and Rimmo 1998, Lajunen et al. 2004) that are predictive of accidents (De Winter
and Dodou 2010). EFA has also been applied to study phenomena, such as decision-
making style and driving style (French et al. 1993), driver aggression and anger (Hauber
1980, Iversen and Rundmo 2002, 2004), driver stress (Gulian et al. 1989), driver fatigue
(Matthews and Desmond 1998), perceptual motor skills and safety skills (Lajunen et al.
1998), skills and safety-motives (Lajunen and Summala 1995), problem driving (Hartos
et al. 2000), driver vengeance (Wiesenthal et al. 2000), driver emotion (Mesken 2006),
avoidance of driving (Stewart and St. Peter 2004) and attitudes and habits concerning
speeding (De Pelsmacker and Janssens 2007). The majority of EFA studies are based on
self-reported driving behaviours. However, a smaller number of studies have applied EFA
on objective driving data (Janssen 1994). Following EFA, related multivariate statistical
methods that are directed towards hypothesis testing and causality, such as confirmatory
factor analysis (CFA) and structural equation modelling (SEM), can facilitate a model for
understanding driver behaviour. Several important steps have already been taken.
For example, Groeger (2000a) tested a four-factor cognitive theory of driving behaviour
using CFA. A similar type of study was presented in Grayson et al. (2003). SEM has
been applied to driver behaviour by, for example, Chen (2009), De Pelsmacker
and Janssens (2007), Sato and Akamatsu (2008), Ulleberg and Rundmo (2003),
536 J.C.F. de Winter and R. Happee
Warner and A

berg (2006) and Warner et al. (2010). SEM may also be useful for testing
models which currently operate at the qualitative level only, such as Rudin-Brown and
Parkers (2004) qualitative model of behavioural adaptation. The merit of SEM for travel
behaviour research and driver behaviour studies has also been acknowledged by Golob
(2003) in a review.
Trait models operate at a less global level than motivational models, but do not aim to
predict driver behaviour in all its detail such as the time-varying controls issued to steer
pedals and other interfaces as do adaptive control models. Indeed, there are some
qualitative and unspecific aspects to multivariate statistics, such as the researchers
decision about the number of factors to extract and the subjective interpretation of the
factors. Moreover, factors are not uniquely defined by the data, and factorial invariance
for different populations or different moments in time is another aspect to be considered.
Hence, it can be argued that multivariate driver modelling takes an intermediate position
on a dimension ranging from unspecific motivational models to highly specific adaptive
control models (Figure 2). The focus of multivariate modelling is not on fitting the data as
accurately as possible (as with many adaptive control models), nor is it on covering the
entire driving task (as with many motivational models), rather it aims for the construction
of parsimonious models by finding a trade-off between model fit and model simplicity and
generalisability. In this respect, analogies can be made with psychometric models of
human intelligence and personality: there is considerable disagreement whether there is
one general intelligence factor or a set of partially independent intelligence factors (Neisser
et al. 1996, Gottfredson 1997). It is most likely that there is no single answer, as all models
fit data to different degrees. A similar notion applies to driving: it is to be expected that
a single model of driver behaviour does not exist. However, models should be compared
with respect to their parsimony and their predictive value.
A car is a deterministic system, many aspects of which can be adequately modelled
using mathematics, elementary mechanics and dynamics. Humans are less deterministic,
comprising inter-individual differences (e.g. skill, age and gender), intra-individual
differences (e.g. fatigue, concentration and emotion), physical and mental processes and
limitations, learning and various forms of error. As indicated by Hancock (1999): The
larger question remains. Will the physicist and the psychologist ever meet? Someday they
must. Yet clearly much remains to be achieved if those of a mathematical persuasion are to
change their fundamental perspective to a psychological focus while many in psychology
learn to use the austere scalpel of numbers in their descriptions of behavior. Hopefully,
such a union will be of great value in transportation and many worlds beyond (Hancock,
p. 199). Multivariate statistics may be the solution for modelling car driving, which is
irreducibly a human-machine process.
4.1. Experiment 4. Factor analysis for driver assessment: extracting a speed factor
We conducted a statistical analysis of 520 pre-licence drivers who followed a 7 h driver
training course in a simulator over multiple days (De Winter et al. 2007). It was found that
the correlation matrix of the mean task-completion times of a diverse range of tasks (e.g.
changing gear, completing road stretches, stopping the car and car-following) had one
dominant eigenvalue. Performing EFA reduced the task-completion time to one general
factor, here called the general speed factor.
In later work (De Winter et al. 2009), we replicated the calculation of the speed score in
a sample of 804 pre-licence drivers (366 men, 438 women) who are trained in the simulator
Theoretical Issues in Ergonomics Science 537
prior to their first on-the-road lessons. Figure 7 illustrates that the speed score predicted
violations (e.g. close following, violating traffic rules) in the simulator. Figure 8 shows that
extreme behaviours (i.e. speed score greater than 1) were particularly prominent amongst
young male drivers. This is in line with the well-established fact that young male drivers
are more involved in serious on-the-road accidents. Figure 9 shows that the speed score
predicted the time taken to obtain a drivers licence, further contributing to the predictive
validity of the speed factor (see also De Winter et al. 2009).
The empirical demonstration of a general speed factor is in line with motivational
models in which speed is generally considered the primary variable (Vaa 2007), with
Rothengatter (1988) and Groeger (2000b) finding that speed is a consistent variable over
time and locations, and also with kinematic/dynamic analyses of steering and braking
showing that speed is a crucial variable determining the criticality of a situation. Clearly,
the general speed factor may be a promising predictive-valid construct in future
multivariate trait models of driver behaviour. Future advanced driver assistance systems
(ADAS) could incorporate a dynamic factor analysis, calculating the drivers speed score
in real time as a function of the time taken to complete various tasks during a trip. This
speed score, representing the drivers overall tendency for speed, could be a dominant
variable in the adaptive ADAS interface.
4.2. Limitations
This study is not free from limitations. Inevitably, it does not provide a comprehensive
description of all driver behaviour models. Other overviews of driver models can be found
in Cacciabue (2007), Michon (1985) and Ranney (1994). With respect to the taxonomy of
Michon (Figure 1), task analyses were not evaluated here, nor were mechanistic models,
7 6 5 4 3 2 1 0 1 2 3
3
2
1
0
1
2
3
4
5
6
Speed score
V
i
o
l
a
t
i
o
n

s
c
o
r
e
Men
Women
Figure 7. Violation score (representing the number of violations that occurred in the simulator) vs.
the speed score for each of the 804 participants. The correlation between speed score and violation
score was 0.68.
538 J.C.F. de Winter and R. Happee
2 1.5 1 0.5 0 0.5 1 1.5 2
100
120
140
160
180
200
220
Speed score
L
a
s
t

S
i
m

t
e
s
t

p
a
s
s
e
d

(
d
a
y
s
)
Figure 9. Mean duration between the last simulator driving lesson and the day the learner driver
passed the on-road practical driving test (LastSim-TestPassed). The participants for whom complete
data were available (n 727) were sorted on the basis of their speed score, and 10 groups of equal
size were created. The horizontal axis shows the mean speed score and the vertical axis shows
the mean LastSim-TestPassed. The vertical lines are the 95% confidence intervals of the mean
LastSim-TestPassed. The correlation between the speed score and LastSim-TestPassed was 0.28.
17.79 17.91 17.94 17.98 18.01 18.10 18.39 19.01 20.78 27.93
0
10
20
30
40
36
44
39
42
33
47
40
41
36
44
46
34
35
46
36
44
32
49
33
47
Age
P
e
r
c
e
n
t
a
g
e

o
f

s
t
u
d
e
n
t
s
Men
Women
Figure 8. Percentage of men/women with a speed score greater than 1. This figure was created by
sorting the 804 participants according to their age at the first driving simulator lesson and
subsequently dividing them into 10 groups of 80/81. The horizontal axis represents the mean age
of the 80/81 participants in the groups. The vertical axis represents the percentage of men/women.
The numbers above the markers represent the number of men/women per group.
Theoretical Issues in Ergonomics Science 539
for example, a model that regards a traffic stream mathematically as a continuous
incompressible fluid.
It may be considered that this study oversimplifies and fails to discuss the nuances of
the many different driver behaviour models, which all serve different purposes and focus
on different aspects of the driving task. For instance, some adaptive control models
discussed in the literature operate at a very fine granularity, being merely descriptive or
analytical (such as detailed descriptions of the task of changing gear; Carsten 2007).
Motivational models, on the other hand, attempt to explain which psychological factors
affect driver behaviour and why drivers make certain decisions. Comparing such very
different models is inherently difficult. Even so, the meta-theoretical dimension identified
in driver behaviour models, ranging from unspecific to specific models, is considered an
important result.
We focused mainly on individual differences in car driving behaviour. The study of
individual subject variability, particularly the use of psychophysiological indicators, is
certainly also of relevance. As indicated by Brookhuis (2008): One of the major problems
of an adequate adaptive vehicle control system is to detect and assess inadequate driving
by the driver of the motor vehicle; when and why performance drops below the red line,
where and what exactly is this red line (p. 58). Brookhuis et al. (2003) provided a
framework for detecting driver impairment and showed that the difficulty with
psychophysiology is that it provides an indirect index of driver impairment. Also, one
has to deal with the consequences of signal-detection theory, and it is difficult to avoid
false alarms while at the same time being sensitive enough to detect changes in driver state.
In order to detect driver impairment, it is necessary to consider the separate influences of
variability within and between drivers. We expect that multivariate trait models are also
useful in this respect, such as models incorporating robust individual differences while
simultaneously being able to identify time-varying latent constructs such as driver
impairment in real time.
Finally, a remark with respect to cognitive models (Figure 1); Michon (1985) refers to
cognitive models as adaptive control models of thought. Indeed, our concerns about
limited generalisability and overfitting of adaptive control models also apply to complex
cognitive models (Fum et al. 2007).
5. Conclusion
Motivational models have been shown to be valuable in generating ideas regarding the
mechanisms related to driver psychology. However, they have received extensive criticism
regarding their lack of specificity and it is concluded that they cannot be used to create
a predictive-valid driver profile.
Adaptive control models have been shown to be useful for quantitative driver
parameter identification in basic tracking tasks. However, here it was revealed that latent
correlations, which are not readily captured by adaptive control models operating at the
vehicle dynamics level, can be present in data. Even under stationary conditions, there
were considerable differences between and within drivers. Moreover, psychological
implausibility and the problem of overfitting were discussed. Overfitting is a dangerous
practice because if you overfit you think you know more than you really know
(Gru nwald 2000, p. 148).
Trait models have been criticised in the past for using post hoc explanations without a
multifactorial structural approach. The right statistical techniques should be used in order
540 J.C.F. de Winter and R. Happee
to go beyond establishing correlations between single variables. Multivariate statistical
analyses, in particular EFA and SEM, were proposed as promising future tools for the
parsimonious modelling of driver behaviour.
Acknowledgements
The research of J.C.F. de Winter is supported by the Dutch Technology Foundation STW, Applied
Science Division of NWO (Netherlands Organisation for Scientific Research) and the Technology
Program of the Ministry of Economic Affairs. The research of R. Happee is supported by the
Ministry of Economic Affairs through the projects Driver Observation in Car Simulators (DrivObs,
agentschap.nl HTASI09004/Eureka E!5395), Connect & Drive (agentschap.nl HTASD08002), and
Mobility Intelligence using Load based Lateral Stability (MILLS, agentschap.nl HTAS10000003U/
Eureka E!4513).
Note
1. Michon (1985) used the term adaptive control models to characterise models that include
functional components that capture behavioural variability. A more traditional definition of
adaptive control refers to a class of controllers that are able to modify their own parameters, for
example to facilitate learning (Young 1969, Flach 1990). This article follows Michons definition
and regards an adaptive control model as any system whose components dynamically interact
and which makes assumptions about driver behaviour. This also comprises, for example,
manual control models that only include the drivers simple aim of minimising speed difference
with the car in front.
Task analyses are not evaluated in this article. A task analysis is an inventory of facts about
driving tasks, performance objectives and ability requirements. Task analyses can be used for
driver assessment by comparing the drivers performance with the ability requirements of the
driving task (i.e. the norms). However, driver assessment should go beyond such single
performance measures in order to uncover what lies beneath the data. Therefore, using a trait
model in combination with a task analysis could be a useful strategy for constructing a driver
profile (Quenault et al. 1968, Michon 1985).
References
A

berg, L. and Rimmo , P.-A., 1998. Dimensions of aberrant driver behaviour. Ergonomics, 41 (1),
3956.
Af Wa hlberg, A. and Dorn, L., 2009. Bus driver accident record: the return of accident proneness.
Theoretical Issues in Ergonomics Science, 10 (1), 7791.
Allen, R.W., et al., 2005. Driver assessment with measures of continuous control behavior. In:
Proceedings of the third international driving symposium on human factors in driver assessment,
training and vehicle design, 2730 June, Rockport, ME, 165172.
Bekey, G.A., Burnham, G.O., and Seo, J., 1977. Control theoretic models of human drivers in car
following. Human Factors, 19 (4), 399413.
Blockey, P.N. and Hartley, L.R., 1995. Aberrant driving behaviour: errors and violations.
Ergonomics, 38 (9), 17591771.
Blomquist, G., 1986. A utility maximization model of driver traffic safety behavior. Accident
Analysis and Prevention, 18 (5), 371375.
Boer, E.R., 1999. Car following from the drivers perspective. Transportation Research Part F,
2 (4), 201206.
Boer, E.R., et al., 2000. Experiencing the same road twice: a driver centered comparison between
simulation and reality. In: M.P. Gauriat and A. Kemeny, eds. Driving simulation conference
Europe proceedings, 68 September, Paris, INRETS, 3355.
Theoretical Issues in Ergonomics Science 541
Boer, E.R., et al., 2005. Driver performance assessment with a car following model. In: Proceedings
of the third international driving symposium on human factors in driver assessment, training and
vehicle design, 2730 June, Rockport, ME, 433440.
Brackstone, M. and McDonald, M., 1999. Car-following: a historical review. Transportation
Research Part F, 2 (4), 181196.
Brookhuis, K.A., 2008. From ergonauts to infonauts: 50 years of ergonomics research. Ergonomics,
51 (1), 5558.
Brookhuis, K.A., De Waard, D., and Fairclough, S.H., 2003. Criteria for driver impairment.
Ergonomics, 46 (5), 433445.
Cacciabue, P.C., ed., 2007. Modelling driver behaviour in automotive environments. London:
Springer-Verlag.
Carsten, O., 2007. From driver models to modelling the driver: what do we really need to know
about the driver? In: P.C. Cacciabue, ed. Modelling driver behaviour in automotive
environments. London: Springer-Verlag, 105120.
Chen, C.-F., 2009. Personality, safety attitudes and risky driving behaviors evidence from young
Taiwanese motorcyclists. Accident Analysis and Prevention, 41 (5), 963968.
De Pelsmacker, P. and Janssens, W., 2007. The effect of norms, attitudes and habits on speeding
behavior: scale development and model building and estimation. Accident Analysis and
Prevention, 39 (1), 615.
De Winter, J.C.F. and Dodou, D., 2010. The driver behaviour questionnaire as a predictor of
accidents: a meta-analysis. Journal of Safety Research, 41 (6), 463470.
De Winter, J.C.F., et al., 2006. Driver assessment during self and forced-paced simulation-based
training. In: M.P. Gauriat and A. Kemeny, eds. Driving simulation conference Europe
proceedings, 46 October, Paris, INRETS-Renault, 157166.
De Winter, J.C.F., et al., 2007. Violations and errors during simulation-based driver training.
Ergonomics, 50 (1), 138158.
De Winter, J.C.F., et al., 2008. A two-dimensional weighting function for a driver assistance system.
IEEE Transactions on Systems, Man, and Cybernetics Part B, 38 (1), 189195.
De Winter, J.C.F., et al., 2009. Relationships between driving simulator performance and driving
test results. Ergonomics, 52 (2), 137153.
Elvik, R. and Vaa, T., 2004. Handbook of road safety measures. Oxford: Elsevier Science.
Flach, J.M., 1990. Control with an eye for perception: precursors to an active psychophysics.
Ecological Psychology, 2 (2), 83111.
French, D.J., et al., 1993. Decision-making style, driving style, and self-reported involvement in road
traffic accidents. Ergonomics, 36 (6), 627644.
Fuller, R., 2005. Towards a general theory of driver behaviour. Accident Analysis and Prevention,
37 (3), 461472.
Fum, D., Del Missier, F., and Stocco, A., 2007. The cognitive modeling of human behavior: why a
model is (sometimes) better than 10,000 words. Cognitive Systems Research, 8 (3), 135142.
Gibson, J.J. and Crooks, L.E., 1938. A theoretical field-analysis of automobile-driving. The
American Journal of Psychology, 51 (3), 453471.
Ginzburg, L.R. and Jensen, C.X.J., 2004. Rules of thumb for judging ecological theories. Trends in
Ecology and Evolution, 19 (3), 121126.
Golob, T.F., 2003. Structural equation modeling for travel behavior research. Transportation
Research Part B, 37 (1), 125.
Gottfredson, L.S., 1997. Mainstream science on intelligence: an editorial with 52 signatories, history,
and bibliography. Intelligence, 24 (1), 1323.
Grayson, G.B., et al., 2003. Risk, hazard perception and perceived control. Crowthorne, UK:
Transport Research Laboratory, TRL Report 560.
Groeger, J.A., 2000a. Understanding driving: applying cognitive psychology to a complex everyday
task. Philadelphia: Psychology Press.
Groeger, J.A., 2000b. Fast learners: once a speeder, always a speeder? In: Proceedings of the 10th
seminar on behavioural research in road safety, 35 April, Esher, Surrey, 144151.
542 J.C.F. de Winter and R. Happee
Gru nwald, P., 2000. Model selection based on minimum description length. Mathematical
Psychology, 44 (1), 133152.
Gulian, E., et al., 1989. Dimensions of driver stress. Ergonomics, 32 (6), 585602.
Guo, K. and Guan, H., 1993. Modelling of driver/vehicle directional control system. Vehicle System
Dynamics, 22 (3), 141184.
Haight, F.A., 1964. Accident proneness, the history of an idea. Automobilismo e Autombolismo
Industriale, 12, 534546.
Hale, A.R., Stoop, J., and Hommels, J., 1990. Human error models as predictors of accident
scenarios for designers in road transport systems. Ergonomics, 33 (10/11), 13771387.
Hancock, P.A., 1999. Is car following the real question are equations the answer? Transportation
Research Part F, 2 (4), 197199.
Hartos, J.L., et al., 2000. Can I take the car?: relations among parenting practices and adolescent
problem-driving practices. Journal of Adolescent Research, 15 (3), 352367.
Hatakka, M., et al., 2002. From control of the vehicle to personal self-control; broadening the
perspectives to driver education. Transportation Research Part F, 5 (3), 201215.
Hauber, A.R., 1980. The social psychology of driving behaviour and the traffic environment:
research on aggressive behaviour in traffic. International Review of Applied Psychology, 29 (4),
461474.
Hollnagel, E., Na bo, A., and Lau, I.V., 2003. A systemic model for driver-in-control. In: Proceedings
of the second international driving symposium on human factors in driver assessment, training
and vehicle design, 2124 July, Park City, UT, 8691.
Huguenin, R.D., 1988. The concept of risk and behaviour models in traffic psychology. Ergonomics,
31 (4), 557569.
Huguenin, R.D. and Rumar, K., 2001. Models in traffic psychology. In: P.-E. Barjonet, ed. Traffic
psychology today. Boston: Kluwer Academic, 3159.
Iversen, H. and Rundmo, T., 2002. Personality, risky driving and accident involvement among
Norwegian drivers. Personality and Individual Differences, 33 (8), 12511263.
Iversen, H. and Rundmo, T., 2004. Attitudes towards traffic safety, driving behaviour and accident
involvement among the Norwegian public. Ergonomics, 47 (5), 555572.
Jagacinski, R.J. and Flach, J.M., 2003. Control theory for humans: quantitative approaches to
modeling performance. Mahwah, NJ: Lawrence Erlbaum Associates.
Janssen, W.H., 1994. Seat-belt wearing and driving behavior: an instrumented-vehicle study.
Accident Analysis and Prevention, 26 (2), 249261.
Janssen, W.H. and Tenkink, E., 1988. Considerations on speed selection and risk homeostasis in
driving. Accident Analysis and Prevention, 20 (2), 137142.
Jensen, A.R., 2006. Clocking the mind: mental chronometry and individual differences. Amsterdam:
Elsevier.
Kleinman, D.L., Baron, S., and Levison, W.H., 1970. An optimal control model of human response
part I: theory and validation. Automatica, 6 (3), 357369.
Lajunen, T., Parker, D., and Stradling, S.G., 1998. Dimensions of driver anger, aggressive and
highway code violations and their mediation by safety orientation in UK drivers.
Transportation Research Part F, 1 (2), 107121.
Lajunen, T., Parker, D., and Summala, H., 2004. The Manchester driver behaviour questionnaire:
a cross-cultural study. Accident Analysis and Prevention, 36 (2), 231238.
Lajunen, T. and Summala, H., 1995. Driving experience, personality, and skill and safety-motive
dimensions in drivers self-assessments. Personality and Individual Differences, 19 (3), 307318.
MacAdam, C.C., 2003. Understanding and modeling the human driver. Vehicle System Dynamics,
40 (13), 101134.
Matthews, G. and Desmond, P.A., 1998. Personality and multiple dimensions of
task-induced fatigue: a study of simulated driving. Personality and Individual Differences,
25 (3), 443458.
McRuer, D.T. and Jex, H.R., 1967. A review of quasi-linear pilot models. IEEE Transactions on
Human Factors in Electronics, HFE-8 (3), 231249.
Theoretical Issues in Ergonomics Science 543
McRuer, D.T., et al., 1977. New results in driver steering control models. Human Factors, 19 (4),
381397.
Mesken, J., 2006. Determinants and consequences of drivers emotions. Thesis (PhD). University of
Groningen, the Netherlands.
Michon, J.A., 1985. A critical view of driver behavior models: what do we know, what should we do?
In: L. Evans and R.C. Schwing, eds. Human behavior and traffic safety. New York: Plenum,
485520.
Mulder, M., 1999. Cybernetics of tunnel-in-the-sky displays. Thesis (PhD). Delft University of
Technology, the Netherlands.
Na a ta nen, R. and Summala, H., 1974. A model for the role of motivational factors in drivers
decision-making. Accident Analysis and Prevention, 6 (3/4), 243261.
Neisser, U., et al., 1996. Intelligence: knowns and unknowns. American Psychologist, 51 (2),
77101.
ONeill, B., 1977. A decision-theory model of danger compensation. Accident Analysis and
Prevention, 9 (3), 157165.
ONeill, B. and Williams, A., 2004. Risk homeostasis hypothesis: a rebuttal. Injury Prevention, 4 (2),
9293.
Panou, M., Bekiaris, E., and Papakostopoulos, V., 2007. Modelling driver behaviour in European
Union and International Projects. In: P.C. Cacciabue, ed. Modelling driver behaviour in
automotive environments. London: Springer-Verlag, 325.
Pitt, M.A., Young, I.J., and Zhang, S., 2002. Toward a method of selecting among computational
models of cognition. Psychological Review, 109 (3), 472491.
Plo chl, M. and Edelmann, J., 2007. Driver models in automobile dynamics application. Vehicle
System Dynamics, 45 (7), 699741.
Quenault, S.W., Golby, C.W., and Pryer, P.M., 1968. Age group and accident rate driving behaviour
and attitudes. Crowthorne, UK: Transportation and Road Research Laboratory, Report
LR167.
Ranney, T.A., 1994. Models of driving behavior: a review of their evolution. Accident Analysis and
Prevention, 26 (6), 733750.
Ranney, T.A., 1999. Psychological factors that influence car-following and car-following model
development. Transportation Research Part F, 2 (4), 213219.
Rasmussen, J., 1983. Skills, rules, and knowledge; signals, signs, and symbols, and other distinctions
in human performance models. IEEE Transactions on Systems, Man, and Cybernetics, 13 (3),
257266.
Reason, J.T., et al., 1990. Errors and violations on the roads: a real distinction? Ergonomics,
33 (10/11), 13151332.
Reid, L.D., 1983. A survey of recent driver steering behavior models suited to accident studies.
Accident Analysis and Prevention, 15 (1), 2340.
Roberts, S. and Pashler, H., 2000. How persuasive is a good fit? A comment on theory testing.
Psychological Review, 107 (2), 358367.
Rothengatter, T., 1988. Risk and the absence of pleasure: a motivational approach to modelling
road user behaviour. Ergonomics, 31 (4), 599607.
Rothengatter, T., 2002. Drivers illusions no more risk. Transportation Research Part F, 5 (4),
249258.
Rudin-Brown, C.M. and Parker, H.A., 2004. Behavioural adaptation to adaptive cruise
control (ACC): implications for preventive strategies. Transportation Research Part F, 7 (2),
5976.
Sato, T. and Akamatsu, M., 2008. Modeling and prediction of driver preparations for making a right
turn based on vehicle velocity and traffic conditions while approaching an intersection.
Transportation Research Part F, 11 (4), 242258.
Sheridan, T.B., 2004. Driver distraction from a control theory perspective. Human Factors, 46 (4),
587599.
544 J.C.F. de Winter and R. Happee
Stewart, A.E. and St. Peter, C.C., 2004. Driving and riding avoidance following motor vehicle
crashes in a non-clinical sample: psychometric properties of a new measure. Behaviour
Research and Therapy, 42 (8), 859879.
Summala, H., 1996. Accident risk and driver behaviour. Safety Science, 22 (13), 103117.
Theeuwes, J., 2001. The effects of road design on driving. In: P.-E. Barjonet, ed. Traffic psychology
today. Boston: Kluwer Academic, 241264.
Trimpop, R.M., 1996. Risk homeostasis theory: problems of the past and promises for the future.
Safety Science, 22 (13), 119130.
Ulleberg, P. and Rundmo, T., 2003. Personality, attitudes and risk perception as predictors of risky
driving behaviour among young drivers. Safety Science, 41 (5), 427443.
Vaa, T., 2007. Modelling driver behaviour on basis of emotions and feelings: intelligent transport
systems and behavioural adaptations. In: P.C. Cacciabue, ed. Modelling driver behaviour in
automotive environments. London: Springer-Verlag, 208232.
Van der Molen, H.H. and Bo tticher, A.M.T., 1988. A hierarchical risk model for traffic participants.
Ergonomics, 31 (4), 537555.
Van Winsum, W., 1999. The human element in car following models. Transportation Research Part
F, 2 (4), 207211.
Warner, H.W. and A

berg, L., 2006. Drivers decision to speed: a study inspired by the theory of
planned behavior. Transportation Research Part F, 9 (6), 427433.
Warner, H.W., O

zkan, T., and Lajunen, T., 2010. Drivers propensity to have different types of
intelligent speed adaptation installed in their cars. Transportation Research Part F, 13 (3),
206214.
Wiesenthal, D.L., Hennessy, D., and Gibson, P.M., 2000. The Driving Vengeance Questionnaire
(DVQ): the development of a scale to measure deviant drivers attitudes. Violence and Victims,
15 (2), 115136.
Wilde, G.J.S., 1988. Risk homeostasis theory and traffic accidents: propositions, deductions and
discussion of dissension in recent reactions. Ergonomics, 31 (4), 441468.
Wilde, G.J.S., Robertson, L.S., and Pless, I.B., 2002. For and against: does risk homoeostasis theory
have implications for road safety. British Medical Journal, 324, 11491152.
Young, L.R., 1969. On adaptive manual control. IEEE Transactions on Man-Machine Systems,
10 (4), 292331.
About the author
J.C.F. de Winter received his MSc degree in aerospace engineering, specialising in control and
simulation, from the Delft University of Technology (TU Delft), Delft, The Netherlands, in 2004. He
is currently an assistant professor in the BioMechanical Engineering Department at TU Delft, where
he received his PhD in 2009. His research interests include simulator-based driver training and driver
assessment.
Riender Happee received his MSc and PhD degrees in mechanical engineering from the Delft
University of Technology (TU Delft), Delft, The Netherlands, in 1986 and 1992, respectively. Since
2007, he has been an Assistant Professor with the TU Delft. He is also currently with the Intelligent
Automotive Systems Research Group, TU Delft, and has been coordinating various projects on
human-machine interfacing for extreme driving, cooperative driving, driver observation and
biomedical projects on neuromuscular stabilisation.
Theoretical Issues in Ergonomics Science 545

Vous aimerez peut-être aussi