Accident Analysis and Prevention 62 (2014) 388396

Contents lists available at ScienceDirect

Accident Analysis and Prevention

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

A review of game-theoretic models of road user behaviour

Rune Elvik
Institute of Transport Economics, Gaustadalleen 21, NO-0349 Oslo, Norway

a r t i c l e i n f o a b s t r a c t

Article history: This paper reviews game-theoretic models that have been developed to explain road user behaviour in
Received 26 October 2012 situations where road users interact with each other. The paper includes the following game-theoretic
Received in revised form 8 May 2013 models:
Accepted 5 June 2013

1. A general model of the interaction between road users and their possible reaction to measures improv-
Keywords:
ing safety (behavioural adaptation).
Game-theory
Road user behaviour
2. Choice of vehicle size as a Prisoners dilemma game.
Interaction 3. Speed choice as a co-ordination game.
Review 4. Speed compliance as a game between drivers and the police.
5. Merging into trafc from an acceleration lane as a mixed-strategy game.
6. Choice of level of attention in following situations as an evolutionary game.
7. Choice of departure time to avoid congestion as variant of a Prisoners dilemma game.
8. Interaction between cyclists crossing the road and car drivers.
9. Dipping headlights at night well ahead of the point when glare becomes noticeable.
10. Choice of evasive action in a situation when cars are on collision course.

The models reviewed are different in many respects, but a common feature of the models is that they
can explain how informal norms of behaviour can develop among road users and be sustained even
if these informal norms violate the formal regulations of the trafc code. Game-theoretic models are
not applicable to every conceivable interaction between road users or to situations in which road users
choose behaviour without interacting with other road users. Nevertheless, it is likely that game-theoretic
models can be applied more widely than they have been until now.
2013 Elsevier Ltd. All rights reserved.

1. Introduction road users interact in complex trafc environments, merely that

The absence of a strong theoretical foundation for road safety
research has often been pointed out. There has been a rapid devel-
opment of statistical methods for studying accident data in recent
years. A large number of statistical models now exist, but it is
not clear that these models reveal causal relationships (Hauer,
2010). Statistical models show that variables are related; they do
not explain why the variables are related. One can measure the
complexity of the trafc environment in various ways and show
statistically that, all else equal, more accidents occur in a com-
plex trafc environment than in a simple one. One may speculate
that in complex trafc environments, road users are faced with a
more demanding task than in simple trafc environments, and are
therefore more likely to make errors that lead to accidents. This con-
jecture is reasonable, but the statistical models do not show how
Tel.: +47 48943353; fax: +47 22 609200.
E-mail address: re@toi.no
their interactions more often are associated with accidents than in
simpler trafc environments.
A number of models of driver behaviour have also been pro-
posed recently (Fuller, 2005; Shinar, 2007; Vaa, 2007). These
models describe driver behaviour at a very general level. It is not
clear that the models provide the needed theoretical foundation
for statistical modelling, nor even that the models can predict how
road users will behave in specic situations where they need to
co-ordinate behaviour in order to avoid accidents. There is, how-
ever, a well-developed body of theory that can be applied to the
study of interaction between road users. Game theory offers a
rich set of models that can fruitfully be applied to the study of
road user behaviour. The purpose of this paper is to review the
use of game-theoretic models of road user behaviour and dis-
cuss how such models can contribute to explaining a number of
commonly observed forms of road user behaviour. More speci-
cally, the following choices of behaviour have all been modelled as
games either between road users or between road users and police
ofcers:

0001-4575/$ see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.aap.2013.06.016
 R. Elvik / Accident Analysis and Prevention 62 (2014) 388396 389

Safe driving speed on 27 road secons according to Dutch drivers maximum safe speed was 100 km/h or more. Only 11 percent of the
45.0
drivers stated a maximum safe speed of between 81 and 99 km/h.
40.0 38.6 Pedersen derived solutions of the game-theoretic model for the
35.0
interaction between doves, for the response of doves to an external
31.6 improvement in safety (i.e. improved safety not brought about by
Percentage of drivers (N = 572)

30.0
the actions of road users, but, for example, by system improvements
25.0 implemented by government); for the interaction between a hawk
20.0
and a dove and their joint response to an external improvement in
safety; and for the interaction between two hawks.
15.0
12.1 11.4
He found that when two doves meet in trafc, both of them ben-
10.0 et from the fact that they behave cautiously; this can in turn induce
6.2
one or both of them to be less careful. Moreover, doves may become
5.0
less careful as a result of an external improvement of safety, i.e. they
0.0
Below 80 80 km/h 81-99 km/h 100 km/h Above 100
may adapt their behaviour by reducing the effort made to maintain
Safe speed in kilometres per hour safety. When a hawk and a dove meet, the dove will tend to become
more cautious and the hawk less cautious than if both were doves.
Fig. 1. Distribution of Dutch drivers according their opinion about safe speed on 27
road sections. As for behavioural adaptation to an external improvement of safety,
the model did not give a unique prediction for the situation where
a dove and a hawk interact. When two hawks meet, they will com-
1. A general model of the interaction between road users and their pete for a rst-mover advantage at the expense of each other and
possible reaction to measures improving safety (behavioural risk can become high. Ultimately, this may induce one of the hawks
adaptation). to behave as a dove to avoid an accident. Finally, it was concluded
2. Choice of vehicle size as a Prisoners dilemma game. that an external safety improvement will tend to increase the share
3. Speed choice as a co-ordination game. of hawks in trafc.
4. Speed compliance as a game between drivers and the police. These model results are at a high level of generality and
5. Merging into trafc from an acceleration lane as a mixed- do not easily lend themselves to presentation in gures and
strategy game. tables. Nevertheless, if concepts such as carefulness and external
6. Choice of level of attention in following situations as an evolu- safety improvement can be measured, the model has empirical
tionary game. implications. Some of the games modelling specic trafc situa-
7. Choice of departure time to avoid congestion as variant of a tions presented below can be viewed as instances of interactions
Prisoners dilemma game. between hawks and doves and as cases of adaptation to external
8. Interaction between cyclists crossing the road and car drivers. safety improvements.
9. Dipping headlights at night well ahead of the point when glare
becomes noticeable.
10. Choice of evasive action in a situation when cars are on collision
course. 3. Choice of vehicle size as a prisoners dilemma game

Tay (2002) modelled the choice of vehicle size as a Prisoners

All these games have been presented in the literature, but an
dilemma. The model is shown in Table 1.
overview and discussion of them has never been presented. The
The owner (driver) of car 1 chooses between the rows of the
game-theoretic models presented in this review have all been
Table. The owner (driver) of car 2 chooses between the columns
tested empirically or are in principle amenable to empirical testing.
of the Table. The entries in the table show relative fatality risk to
drivers of the two cars. Relative fatality risk has been set to 1 for
2. A general model of road user interaction and adaptation accidents involving two large cars. The risk faced by the driver of
to exogenous changes in safety car 1 is shown in the lower left corner of each cell of Table 1. The risk
faced by the driver of car 2 is shown in the upper right corner of each
Pedersen (2001, 2003) developed a general game-theoretic cell. It is seen that the risk to both drivers is lower when two small
model of road user interaction and adaptation to exogenous cars are involved in an accident (relative risk 0.8 for both drivers)
changes in safety, based on the framework developed by Boyer and than when two large cars are involved in an accident (relative risk
Dionne (1987). The model describes interaction between two road 1.0 for both drivers). The relative risks entered in Table 1 are based
users in general terms. Unlike the other games presented in this on an Australian study quoted by Tay (2002).
paper, it does not refer to specic trafc situations. A distinction It is assumed that both drivers prefer a lower fatality risk to a
is made between two types of road users: aggressive road users, higher fatality risk. If the owner of car 1, who presumably is also
referred to as hawks, and cautious road users, referred to as doves. the driver of that car, chooses a small car (upper row), the owner
While the paper does not document the existence of these two of car 2 is better off by choosing a large car since a relative risk of
types of road users, a Dutch study indirectly provides evidence sup- 0.06 is clearly lower than a relative risk of 0.8. By the same token,
porting their existence. Fig. 1 shows an example of results from that if the owner (driver) of car 1 chooses a large car (lower row), the
study. owner of car 2 is better off by also choosing a large car, because
Data from the study (Goldenbeld and van Schagen, 2007), re- the relative risk (1.0) would then be much lower than if a small car
analysed by Elvik (2010) show a distinctly bimodal distribution of was chosen (17.0). In other words, no matter whether the owner
driver preferences with respect to safe speed on Dutch roads that of car 1 chooses a small car or a large car, it is better for the owner
all had a speed limit of 80 km/h. Drivers were shown pictures of of car 2 to choose a large car. The reasoning is entirely analogous
27 roads and asked to state the highest speed they considered to for the owner of car 1, who will also prefer a large car. Thus both
be safe for each of these roads. They were not informed about the car owners end up choosing a large car. Choosing a large car is a
speed limit. 51 percent of the drivers stated that 80 km/h or less was dominant strategy in the game, i.e. it is the best choice irrespective
the maximum safe speed. 38 percent of the drivers started that the of what the other car owner chooses. The solution to the game, and
390 R. Elvik / Accident Analysis and Prevention 62 (2014) 388396
Table 1
Relative driver fatality risk in accidents involving small and large cars.
Small car
Driver of car 1
Large car
Source: Tay (2002).
to the other games presented in tables in this paper, are shown in
bold italics in the tables.
By choosing large cars both drivers end up with a higher risk
than if they had chosen small cars. There is evidence supporting
the assumptions made in this game. Studies by Broughton (1995),
Mizuno and Kajzer (1999) and Fredette et al. (2008) all show that a
uniform reduction of vehicle mass would be associated with fewer
injuries and fatalities, despite the fact that large cars protect the
occupants better in case of an accident than small cars.
4. Speed choice as a co-ordination game
The choice of speed is clearly inuenced by many factors, of
which the speed limit is a very important factor. However, most
drivers do not want to maintain a speed that differs too much
from the speed chosen by other drivers. Lave (1985) argued that
speed limits are an instrument for co-ordinating speed choice and
making it more uniform. The importance of the speed chosen by
other drivers in inuencing ones own speed is conrmed by a
recent stated-preference survey of factors inuencing speed choice
(Ryeng, 2012). One may therefore model speed choice as a co-
ordination game between two main groups of drivers: the fast
movers (those who stated that safe speed on the 27 Dutch roads
was at least 100 km/h; see Section 2) and the slow movers (those
who stated that safe speed on the 27 Dutch roads was 80 km/h or
less). Table 2 shows this game in normal form.
The ordinal preferences of drivers are indicated by numbers; 4
is the most preferred outcome, 1 is the least preferred outcome.
The fast movers choose between columns; their preferences are
shown in the upper right corner of each cell of the Table. The slow
movers choose between rows; their preferences are shown in the
bottom left corner of each cell of the table. Both groups of drivers
are assumed to prefer driving at the same speed as the other group
to driving at a speed that differs from the other group.
For the fast movers, the best outcome is that everybody drives
at a speed of 100 km/h. The slow movers will prefer this outcome
to driving at 80 km/h, given that the fast movers continue to drive
at 100 km/h. However, the most preferred outcome for the slow
movers is that everybody drives at 80 km/h. Both this outcome,
and the outcome in which everybody drives at 100 km/h, are Nash
equilibria, in the sense that no group can get a more preferred out-
come by unilaterally changing its choice. As an example, if the slow
movers decide to drive at 80 km/h when the fast movers are driving
at 100 km/h, their preference rating is reduced from 3 to 2.
Neither of the equilibria is likely to be stable. Connolly and berg
(1993) explore a contagion model of speeding, in which speed
choice is inuenced by other drivers. Depending on how sensitive
drivers are to the choices made by other drivers, the mean speed
of trafc may be determined by the choices made by even a small
minority of drivers. Suppose, for example, that a slow driver main-
tains a speed of 80 km/h. A fast driver may regard this as too slow,
Driver of car 2
Small car Large car
0.8 0.06
0.8 17.0
17.0 1.0
0.06 1.0
but overtaking is a hassle and it may, on balance, be more pleas-
ant to put up with the slow driver and stay in line. If everybody
thinks like that, the speed of trafc will be 80 km/h. A driver who
is less sensitive to the speeds chosen by other drivers may decide
to overtake the slow driver and may thereby induce other drivers
to do the same. Suddenly, many drivers will be overtaking the slow
driver and the speed of trafc will increase.
5. Speeding as a mixed-strategy game
The game presented above had two solutions. One of them,
the fast mover solution, is consistent with widespread speeding,
i.e. with a situation in which most drivers are violating speed
limits. Speeding is very common. Statistics compiled by the Euro-
pean Transport Safety Council (ETSC, 2010) for nearly 30 European
countries show that it is not uncommon that around 50% of trafc
violates speed limits.
Bjrnskau and Elvik (1992) developed a game-theoretic model
of speeding as an interaction between drivers and the police. It
was assumed that drivers prefer to speed when there is no police
enforcement. When there is police enforcement, drivers prefer not
to speed. The police prefer to do enforcement when there is speed-
ing. When there is no speeding, the police will not see a need for
performing enforcement. Table 3 shows this game in normal form
with numerical payoffs that are intended to indicate costs to soci-
ety of the choices made by the police and costs to drivers resulting
from their choice of speed.
This game does not have a pure strategy solution. Suppose that
the police enforce speed limits. Drivers will then prefer not to speed
(50 is better than 300; left cells of Table 3). However, when
drivers do not speed, the police will prefer not to enforce (0 is better
than 10,000; upper row of Table 3). When there is no enforcement,
drivers will prefer to speed (50 is better than 50). Thus, one could
move around the four cells of Table 3 forever; moving from any of
these cells to the one below it, to the one right of it, to the one on
top of it, or to the one left of it will always be preferred by one of
the players, but not the other.
The game does, however, have a mixed-strategy solution. From
the point of view of the police, this is to assign a probability to
enforcing which is such that drivers get the same expected payoff
from speeding as from not speeding:
EU(speed) = p(300) + (1 p)50
EU(not speed) = p(50) + (1 p)(50)
EU(speed) = EU(not speed) = p(300) + (1 p)50
= p(50) + (1 p)(50)
 R. Elvik / Accident Analysis and Prevention 62 (2014) 388396 391

Table 2
Speed choice as a co-ordination game. Ordinal preferences. 4 = most preferred, 1 = least preferred.
The fast movers
80 km/h 100 km/h
3 2
80 km/h
The slow movers 4 2
1 4
100 km/h
1 3

Table 3
Speeding as a mixed-strategy game between the police and drivers. Payoffs to the police (upper right
corner of each cell) represent societal costs of enforcement and accident. Payoffs to drivers (bottom
left corner of each cell) represent the value for drivers of travel time saving and trafc tickets.
The police
Enforce Not enforce
10,000 0
Not speed
Drivers 50 50
10,000 20,000
Speed
300 50

Solving this for p produces a value of 0.2857. It can be veried
that drivers would then be indifferent between speeding and not
speeding:
EU(speed) = 0.2857 (300) + 0.7143 50 = 50
EU(not speed) = 0.2857 (50) + 0.7143 (50) = 50
From the car drivers point of view, they can adopt a mixed strat-
egy which is such that the police obtain the same expected payoff
by enforcing as by not enforcing. Given the payoffs in Table 3, it
can be calculated that the mixed-strategy equilibrium for drivers
is to speed with a probability of 0.50 and not speed with a prob-
ability of 0.50. Thus, the game implies that the best strategy for
drivers to speed half the time, and the best strategy for the police
is to enforce 28.57 percent of the time. These specic values are
obviously determined by the values assumed in Table 3. The result
is, however, stable and it is easy to show that any deviation from
the mixed-strategy equilibrium will elicit a tendency to move back
to it. Suppose, for example, that the police would like to reduce
speeding by stepping up enforcement from 0.2857 to 0.30. It can
then be calculated that the expected payoff to drivers by speeding
is 55, and the expected payoff by not speeding is 50. Drivers
will thus prefer not to speed. This will in turn reduce the payoff
to the police by enforcing and induce a reduction of enforcement.
The model also implies that imposing stricter penalties for speed-
ing will have no effect on the level of speeding, because the police
will adapt to stricter penalties by reducing enforcement. This will in
turn make speeding as attractive as it was before stricter penalties
were imposed.
an acceleration lane and has to decide whether to merge into the
through lane or stop in the acceleration lane to wait for a sufcient
gap in trafc. Player 2 is driving a car on the through lane adjacent to
the acceleration lane and has to decide whether to give way to the
entering car or not give way. Player 2 can give way either by slow-
ing down or by moving to the other through lane. Fig. 2 shows the
players as well as additional cars in the situation that may inuence
choices, in particular car 5, which inuences the choice of player 2
in changing lane.
The drivers are assumed to play mixed strategies, i.e. they will
not always merge into trafc, nor always give way; the choices
between the options facing each driver are made by certain prob-
abilities. The values of these depend on the trafc situation. The
following variables are assumed to inuence choices:
T1: The distance between the merging car (player 1) and the end of
the acceleration lane, divided by the speed of the merging car. This
variable can be interpreted as the time left for player 1 to decide
whether to merge or give it up.
T2: The time to collision between the merging car and the nearest
through car (car 3 in Fig. 2). This time depends on how close behind
the merging car the through car is.

6. Merging into trafc as a mixed strategy game

Kita (1999) developed in game-theoretic model of merging

behaviour when entering a major road from an acceleration lane. Fig. 2. Key variables inuencing choices when merging into trafc from an acceler-
There are two players in the game, as well additional elements that ation lane.
may inuence the outcome of the game. Player 1 is driving a car in Source: Kita (1999), Fig. 2.
392 R. Elvik / Accident Analysis and Prevention 62 (2014) 388396
Table 4
Interaction between a leading driver and a following driver in a following situation.
Attentive
Leading driver
Inattentive
Source: Chatterjee and Davis (2013), Table 1.
T3: The time to collision between the merging car and the second
through car (car 4 in Fig. 2). If car 3 is close and car 4 is far away,
the driver of car 1 may decide to let car 3 pass and merge behind
it.
T4: The time headway between through cars 2 and 3.
T5: The time to collision between through cars 3 and 5, dened
as the space headway between these cars divided by their relative
speed.
Coefcients describing the effects of variables T1 to T5 were
estimated by observing merging and give-way behaviour on an
expressway in Japan. A total of seven coefcients were estimated.
By inserting these in seven equations, the payoffs to the players
of different choices could be estimated. In general payoffs take on
higher values in situations characterised by large safety margins
with respect to the variables T1 to T5. Once payoffs are known for a
given combination of values on T1 to T5, the probabilities of choos-
ing one or the other option can be estimated for each player. The
best strategy is the one in which player 1 gets a higher expected
payoff by choosing an option with probability p*, rather than a dif-
ferent probability p, given that player 2 chooses an option with
probability q*. The game may have multiple equilibrium values for
the probabilities and a numerical example of equilibrium values
is therefore not given here. It should be noted, however, that this
study is a good example of an empirical testing of a game-theoretic
model and that the study should be easy to replicate.
7. Choice of level of attention in following situations
Chatterjee and Davis (2013) developed an evolutionary game-
theoretic model of driver attention in following situations.
Rear-end collisions are very common in dense urban trafc and
driver inattention contributes importantly to these accidents. Chat-
terjee and Davis model driver attention in following situations by
means of evolutionary game theory. This version of game theory
has been developed and applied in biology to study the evolution of
traits in animal populations. The solution to an evolutionary game
is an evolutionary stable strategy. An evolutionary stable strategy
is a strategy, which if adopted by the entire population cannot be
successfully invaded by a different strategy. This means that once
equilibrium has been established, it will be maintained forever.
Chatterjee and Davis assume that an attentive driver will have
a reaction time of 1 s; an inattentive driver will have a reaction
time of 3 s. Sustained attention requires effort; hence driving is
less demanding and more comfortable if less attention is given
to it. Based on assumptions about the utility functions of drivers,
the speed of trafc and the speed of retardation, the interaction
between a leading driver and a following driver results in the pay-
offs shown in Table 4.
The payoffs to the leading driver are shown in the bottom left
corner of each cell; the payoffs to the following driver are shown in
Following driver
Attentive Inattentive
6 10
6 6
6 590
10 590
the upper right corner of each cell. The game has two Nash equilib-
ria in pure strategies: (1) The leading driver is attentive, following
driver is inattentive, and (2) The leading driver is inattentive, the
following driver is attentive. There is also a mixed-strategy solution
to the game, in which being attentive is chosen by a probability of
0.993 and being inattentive by a probability of 0.007. All these solu-
tions imply that the best outcome is that some drivers are always
inattentive, or, equivalently that each driver is inattentive some of
the time. This may seem counterintuitive, since it is associated with
a higher risk than if all drivers were always attentive. Chatterjee and
Davis therefore ask if all drivers being attentive is an evolutionary
stable strategy.
They conclude that it is not. They show that an evolutionary
stable solution is that 94.57 percent of drivers are attentive and
5.43 percent are inattentive. They go on to study whether impos-
ing penalties on inattentive drivers by means of insurance schemes
can inuence the proportion of inattentive drivers in a popula-
tion. They conclude that insurance schemes penalising inattentive
drivers may indeed reduce their share in the population of drivers.
This model is interesting in that it, like some of the other models
presented in this paper, suggests that normatively inappropriate
behaviour, in this case not paying attention to trafc, could actu-
ally be the best choice for a road user, or a certain share of road
users.
8. Choice of departure time to avoid congestion
Trafc congestion is widely regarded as a waste of time and
many drivers try to avoid it. One way of trying to avoid congested
trafc is to leave early, before the majority of drivers have left, or to
leave late, after the rush-hour has started subsiding. Neither choice,
however, guarantees success. Congestion is to some extent unpre-
dictable, so that even drivers who start early or late could be caught
in congested trafc. Levinson (2005) developed a game-theoretic
model of driver choices in congested trafc. The game is modelled
as a two-person game only, but, as argued in the paper, this for-
mulation of the game involves no loss of generality. Each driver
has three options: (1) Leave early, (2) Leave on-time, (3) Leave late.
There are three outcomes associated with these choices: (1) Arrive
early, (2) Arrive on-time, (3) Arrive late. Arriving on-time is the
most preferred outcome. Arriving early or late is assumed to involve
additional costs compared to arriving on-time.
The payoffs associated with the different choices in this game
are shown in Table 5. These payoffs represent disutilities; hence the
lower the numbers, the better the outcome. There are three com-
ponents of the disutilities: arriving early (before the desired time
of arrival), being delayed in trafc (trip takes longer than normal),
arriving late (after the desired time of arrival).
The upper part of Table 5 shows payoffs when arriving early is
assigned the value of 1, being delayed the value of 2 and arriving
late the value of 3. Thus, arriving early is regarded at the smallest
 R. Elvik / Accident Analysis and Prevention 62 (2014) 388396 393

Table 5
Choice of departure time to avoid congestion. Equilibrium depends an assumptions made regarding valuation of outcomes.
Vehicle 2
Early On-time Late
Payoffs: early = 1, delayed = 2, late = 3
1.5 0 1
Early
1.5 1 3
1 2.5 3
Vehicle 1 On-time
0 2.5 0
1 0 5.5
Late
3 3 5.5
Payoffs: early = 3, delayed = 1, late = 4
2 0 4
Early
2 3 3
3 2.5 4
Vehicle 1 On-time
0 2.5 0
3 0 6.5
Late
4 4 6.5
Source: Levinson (2005).

of the three disutilities. As can be seen from Table 5, the game has
two solutions: (1) Driver 1 departs early and driver 2 departs on
time; (2) Driver 1 departs on time and driver 2 departs early.
It is, however, not obvious that the values assigned to arriving
early, being delayed or arriving late are the only possible values that
can be used. In the lower part of Table 5, being early was assigned
the value of 3, getting delayed the value of 1 and arriving late the
value of 4. The game then has only one solution: that both drivers
depart on time. This illustrates an important point about game-
theoretic models: the best outcome depends on the assumptions
made regarding the preferences and valuations of the players. It is
therefore important to justify the values that are used to represent
player preferences; ideally speaking these values should be based
on empirical evidence. However, in most game-theoretic models,
the values used to represent preferences are ordinal only; i.e. they
only represent the rank-ordering of outcomes from best to worst.
If only ordinal preferences are assumed it does not matter whether
these are represented as 4, 3, 2, 1 or as 100, 25, 10, 5. The ranking
will be the same in both cases. It is only when stronger assumptions
need to be made about preferences that the need arises to justify
the particular values used to represent preferences regarding out-
comes. Games that only have mixed-strategy solutions are more
sensitive to the values assigned to the outcomes than games that
have xed-strategy solutions.
9. Interaction between cars and cyclists crossing the road
According to the rules for trafc in Norway, a cyclist who wants
to cross the road at a pedestrian crossing should get off the bicycle
and walk across the crossing or, if he or she stays mounted on the
bike, stop to let cars pass rst. Legally, a cyclist who gets off the bike
will become a pedestrian and have the right-of-way at the crossing.
If the cyclist stays mounted on the bike, he or she does not have the
right-of-way at pedestrian crossings and must wait for a gap in
trafc before riding across the crossing. It has been observed that
in practice, what most often happens when a cyclist approaches a
pedestrian crossing is that car drivers give way to the cyclist who
stays mounted on the bike and rides across the pedestrian crossing.
This will normally take considerably shorter time than if the cyclist
dismounted and walked across the pedestrian crossing.
Bjrnskau (2007) has developed a game-theoretic model to
account for this. The model is shown in normal form in Table 6.
The entries of the Table show the ordinal preferences of the cyclist
and car driver for the various outcomes. The most preferred out-
come has the value of 6; the least preferred outcome has the value
of 1.
There are two solutions to the game. One of them, that the
cyclist stops and the car driver passes the crossing, is consistent
with legal regulations. The other solution prescribed by the law is
the outcome in the bottom left cell of Table 6. This outcome has
a preference ranking of 5 for the cyclist and 4 for the car driver,
suggesting that there exists a better outcome for both parties to
the interaction. In this game, it is reasonable to assume that the
players move in sequence. When approaching the crossing, the car
driver will observe the behaviour of the cyclist and try to predict
what he or she will do. This gives the cyclist the opportunity to
signal his or her intentions, thus giving the cyclist a rst-mover
advantage. The best outcome for the cyclist is simply to continue
riding across, not having to stop or get off the bike. This saves both
time and energy. By maintaining speed, direction and posture, the
cyclist can signal his or her intentions to the driver. This will force
the driver to stop, leading to the outcome in the upper left cell of
Table 6. This outcome, although not the most preferred outcome
from the drivers point of view, is better than the legally prescribed
394 R. Elvik / Accident Analysis and Prevention 62 (2014) 388396

Table 6
Giving way to cyclists at pedestrian crossings as a co-ordination game. Ordinal preferences. 6 = most
preferred, 1 = least preferred.
Driver
Stop Go
5 2
Cycle
6 1
3 6
Cyclist Stop
3 4
4 1
Get off and walk
5 2

outcome, because the cyclist will almost always cross faster when
riding than when walking, and this will save time for both parties. It
is, as shown by Bjrnskau (2007), also the outcome that is observed
in the huge majority of cases (about 75%).
10. Dipping of headlights at night
the series of repeated encounters will only be broken if one
of the cars turns into a side road before they pass each other.
Moreover, both drivers know that they can punish a lack of co-
operative behaviour by using full beam headlights, although this
comes at a cost. Thus, incentives for dipping headlights are fairly
strong.

When cars pass each other in the dark in Norway, it is com-
mon to dip the headlights well before the full beam headlights
start to cause glare. In fact, headlights are normally dipped at such a
great distance, that in driver training in Norway, drivers are encour-
aged to keep their full beam headlights on as long as possible,
because they can see a lot better with full beam headlights than
with dipped headlights. The problem when driving at night, driver
educators argue, is not that there is too much light, but that there
is too little of it. Bjrnskau (1989) has developed a game-theoretic
model intended to explain the widespread use of dipped headlights
when driving at night. He models the choice between full beam and
dipped headlights as a repeated Prisoners dilemma game. The pref-
erence ranking of the various outcomes of this game is shown in
Table 7. The most preferred outcome has the value of 4, the least
preferred outcome the value of 1.
The idea is that when two cars are approaching each other,
drivers will continuously assess just when they think it is right to
dip the headlights. On the one hand, each driver wants to see as
well as possible. On the other hand, each driver wants to avoid glare,
which in effect reduces what a driver can see, because he or she will
tend to look away from the full beam headlights of an oncoming car
in order to reduce the discomfort. The game is repeated continu-
ously. It is repeated every millisecond, and every millisecond you
meet the driver again. You do of course not see the driver nor are
you able to identify him or her. In that sense the game is completely
anonymous. But as the two cars get closer, both of them know that
they can punish the other driver, or offer to co-operate by dip-
ping the headlights. A practice has developed of dipping headlights
well before glare becomes uncomfortable. This can be modelled
as the optimal (i.e. co-operative) solution to a repeated Prisoners
11. Choice of evasive manoeuvre
The nal game to be discussed in this paper involves choice of
evasive manoeuvre when two cars are on collision course. The game
was introduced by John Prentice in 1974 (Prentice, 1974). The trafc
situation modelled in the game is shown in Fig. 3. Miss Petty was
about to cross a junction, but had overlooked the stop sign and
was entering the junction when she saw a car approaching from
her left side. This car was driven by Mister McKay, who had the
right-of-way.
To avoid the accident, there are various options. For Miss Petty,
the options are to accelerate, hoping to clear the junction before
the car driven by Mr. McKay arrives, or to brake, hoping to come to
a stop before entering Mr. McKays driving lane. For Mr. McKay, the
options as modelled by Prentice are: (1) to accelerate and steer left,
trying to pass in front of Miss Pettys car before it enters the junc-
tion; (2) to steer right, trying to pass behind Miss Pettys car after
it has entered the junction; (3) to brake. The payoffs to the drivers
associated with the possible outcomes of the game are shown in
Table 8.
Stop sign
Miss Petty

dilemma game.
The dominant strategy in the Prisoners dilemma game is to Mr McKay
defect, i.e. choose to keep the full beam on no matter what the
other driver does. It is, however, well-known that in repeated Pris-
oners dilemma games, the co-operative solution can emerge, in
particular if there is a high probability of encountering the other
player again (Axelrod, 1984). The players may then remember
that they co-operated the last time and will try to do so again.
Bjrnskau (1989) argues cars approaching each other from oppo- Fig. 3. Sketch of situation in games regarding evasive action for cars on collision
site directions can be thought of as a series of repeated encounters; course.
 R. Elvik / Accident Analysis and Prevention 62 (2014) 388396
Table 7
Dipping of headlights at nights as a repeated Prisoners dilemma game. Ordinal preferences. 4 = most
preferred, 1 = least preferred.
Dipped headlights
First car
Full beam
If the accident is avoided, a payoff of zero has been assigned. If
the accident occurs, a negative payoff is expected. A minor accident
has a small negative payoff; a more serious accident has a larger
negative payoff. It is assumed that both drivers want to minimise
the maximum potential loss. The maximum potential loss is asso-
ciated with the outcome when both drivers accelerate. This will
result in a severe accident, since impact speed will be higher than
if one or both drivers brake. To avoid such an outcome, neither
Miss Petty nor Mr. McKay should accelerate. For Miss Petty, the
only remaining option is to brake. For Mr. McKay, the remaining
options are to steer right or to brake. If steering right, Mr. McKay
will not be braking; hence if this results in an accident, it will be
slightly more severe than if both drivers brake. Consequently, the
best option for Mr. McKay is to brake. The solution to the game
is that both drivers brake. Given the assumptions made, this will
not prevent an accident from occurring, but will reduce its sever-
ity. Although the choice of other options may prevent the accident
from occurring, these options may also result in a more severe acci-
dent. Rather than gambling on these options, both drivers should
therefore simply brake to reduce impact speed.
In this game, it is essential to remember that both drivers have
to act quickly and cannot wait to see what the other driver is doing.
By that time, it may be too late to avoid an accident, although
this is obviously the most desirable outcome from both drivers
point of view. If Mister McKay knew that Miss Petty would accel-
erate, steering right to pass behind her would indeed be the best
choice. But Mister McKay does not know what Miss Petty will do
and does not have the time to wait for her to decide. If she brakes,
they will end up in the Steer right/Brake cell, in which a crash will
occur. Since, in this pair of choices, Mister McKay did not brake,
the crash will occur at a greater speed than when both drivers
brake.
Table 8
Choice of evasive action for cars on collision course. Payoffs represent accident severity. Lowest number
(10) = most severe accident, highest number (0) = accident avoided.
Accelerate and steer left
Mister McKay Steer right
Brake
395
Second car
Dipped headlights Full beam
3 4
3 1
1 2
4 2
12. Discussion
The need for establishing a rmer theoretical foundation for
studying various aspects of road safety has often been pointed out,
see, for example Davis et al. (2008). Does game theory provide
such a theoretical foundation? To some extent it does, but it
does not represent a comprehensive theory that accounts for all
safety-related phenomena that call for a theoretical understanding.
Game-theoretic models are best suited for modelling the interac-
tion between road users, such as the choice of speed, giving way
in junctions or pedestrian crossings and signalling intentions by
means of speed or lateral placement. In particular, game-theoretic
models may explain why informal norms develop and are sustained
by road user behaviour. This behaviour may conict with formal
regulations. Thus, drivers may co-ordinate their choice of speed at
a level exceeding the speed limit; in particular if in doing so repeat-
edly, they have learned that the risk of detection by the police is
low. In junctions between a wide road carrying a relatively large
trafc volume and a narrow road carrying a relatively small trafc
volume, an informal rule of giving way to trafc entering from the
wide road may develop, even if the junction is controlled by the rule
of giving way to trafc entering from the right. A game-theoretic
model suggests that some level of speeding will be tolerated and
police enforcement less than 100 percent effective in deterring
speeding. Evolutionary game theory suggests that it is not rational
for all drivers to pay full attention to trafc at all times. Numer-
ous other examples of informal norms that are sustained by being
equilibrium solutions to games can be imagined.
It seems clear, however, that game theory cannot serve as a
general theory of road user behaviour. It is best suited to situa-
tions where road users interact. It may not be capable of explaining
the behaviour of a single road user not interacting with others.
Miss Petty
Accelerate Brake
10 0
10 0
0 6
0 6
5 5
5 5
396 R. Elvik / Accident Analysis and Prevention 62 (2014) 388396
Furthermore, it is based on an assumption of rationality that may
not apply fully to all road users. One cannot, for example, attribute
the same level of rationality to children as road users as to adults.
Another limitation of game theory is that some games will not have
unique solutions, but may have more than one solution, like the
speed-choice co-ordination game discussed above. In such cases,
explanations of why one among several equivalent solutions tends
be chosen may have to be sought outside the framework of game
theory.
Despite these limitations, game-theoretic models can be fruit-
fully applied in many situations. The examples given in this paper
are not exhaustive. One promising area for the application of game
theory is in modelling the effects of road safety measures, in par-
ticular those whose effects depend on interaction among road
users. Thus Elvik (2013) has proposed a game-theoretic model of
the effects of daytime running lights, based on Schelling-diagrams
(Schelling, 1978). One can easily imagine that a similar model could
be applied to, for example, the use of studded tyres in winter. Game
theory has been applied to compliance with truck weight limits
(Hildebrand et al., 1990; the study was not included in the review
as it was difcult to retrieve). By extending the use of game theory
to as many road safety measures as possible, one may at least be
able to partly develop a theoretical foundation for understanding
the effects of these measures.
13. Conclusions
The main conclusions of the review of game-theoretic models
of road user behaviour presented in this paper can be summarised
as follows:
1. Game-theoretic models have been developed to explain road
user choices in a number of situations when road users interact
or make choices that have external effects on other road users.
The paper reviews ten such models.
2. Game-theoretic models can explain why informal norms of
behaviour develop and are sustained, such as widespread speed-
ing, giving way to cyclists who do not have priority, and dipping
headlights well ahead of the point when glare starts to become
noticeable. A game-theoretic model also shows that it may be
better to brake to reduce impact speed than to try a hazardous
evasive manoeuvre, which, if unsuccessful, will make the acci-
dent more severe.
3. Game-theoretic models cannot explain all aspects of road user
behaviour, not even in situations where road users interact.
These limitations do not, however, imply that game-theoretic
models cannot be applied more widely than they have until
now.
References
Axelrod, R., 1984. The Evolution of Co-operation. New York, Basic Books.
Bjrnskau, T., 1989. Rasjonalitetsmodeller og virkeligheten. Kan laboratorie-
eksperimenter med gjentatte fangens dilemma-spill ha empirisk relevans?
Norsk Statsvitenskapelig Tidsskrift 5, 237249.
Bjrnskau, T., 2007. Spillet om gangfeltet 6 r etter. Samferdsel 4, 2223.
Bjrnskau, T., Elvik, R., 1992. Can road trafc law enforcement permanently reduce
the number of accidents? Accident Analysis and Prevention 24, 507520.
Boyer, M., Dionne, G., 1987. The economics of road safety. Transportation Research
Part B 21, 413431.
Broughton, J., 1995. The likely effects of downsizing on driver casualties in two-car
accidents. TRL Report 171. Transport Research Laboratory, Crowthorne, Berk-
shire.
Chatterjee, I., Davis, G.A., 2013. An evolutionary game theoretic approach to rear-
ending events on a congested freeway. Paper presented at 2013 Annual Meeting
of the Transportation Research Board, Washington, DC.
Connolly, T., berg, L., 1993. Some contagion models of speeding. Accident Analysis
and Prevention 25, 5766.
Davis, G., Ivan, J., Sayed, T., Tarko, A., 2008. Theory, explanation, and prediction in
road safety: identication of promising directions and plan for advancement.
White Paper No. 3. Prepared by Future Directions Subcommittee of the Task
Force for the Development of a Highway Safety Manual. Transportation Research
Board, Washington, DC.
Elvik, R., 2010. A restatement of the case for speed limits. Transport Policy 17,
196204.
Elvik, R., 2013. Can a road safety measure be both effective and ineffective at the same
time? A game-theoretic model of the effects of daytime running lights. Accident
Analysis and Prevention (paper AAP-D-13-00331 submitted for publication).
European Transport Safety Council (ETSC), 2010. Road Safety Target in Sight: Making
Up for Lost Time. 4th Road Safety PIN Report. European Transport Safety Council,
Brussels.
Fredette, M., Mambu, L.S., Chouinard, A., Bellavance, F., 2008. Safety impacts due to
the incompatibility of SUVs, minivans, and pickup trucks in two-vehicle colli-
sions. Accident Analysis and Prevention 40, 19871995.
Fuller, R., 2005. Towards a general theory of driver behaviour. Accident Analysis and
Prevention 37, 461472.
Goldenbeld, C., van Schagen, I., 2007. The credibility of speed limits on 80 km/h rural
roads: the effect of road and personal(ity) characteristics. Accident Analysis and
Prevention 39, 11211130.
Hauer, E., 2010. Cause, effect and regression in road safety: a case study. Accident
Analysis and Prevention 42, 11281135.
Hildebrand, M., Prentice, B., Lipnowski, I., 1990. Enforcement of highway weight
regulations: a game theoretic model. Journal of the Transportation Research
Forum 30, 442452.
Kita, H., 1999. A merginggiveway interaction model of cars in merging section: a
game theoretic analysis. Transportation Research Part A 33, 305312.
Lave, C.A., 1985. Speeding, coordination, and the 55 MPH limit. American Economic
Review 75, 11591164.
Levinson, D., 2005. Micro-foundations of congestion and pricing: a game theory
perspective. Transportation Research Part A 39, 691704.
Mizuno, K., Kajzer, J., 1999. Compatibility problems in frontal, side, single car colli-
sions and car-to-pedestrian accidents in Japan. Accident Analysis and Prevention
31, 381391.
Pedersen, P.A., 2001. A game-theoretical approach to road safety. Department of
Economics Discussion Paper No. 01. University of Kent, Canterbury.
Pedersen, P.A., 2003. Moral hazard in trafc games. Journal of Transport Economics
and Policy 37, 4768.
Prentice, J.W., 1974. The evasive action decision in an intersection accident: a game
theory approach. Journal of Safety Research 6, 146149.
Ryeng, E.O., 2012. The effect of sanctions and police enforcement on drivers choice
of speed. Accident Analysis and Prevention 45, 446454.
Schelling, T.C., 1978. Micromotives and Macrobehavior. W.W. Norton and Company,
New York.
Shinar, D., 2007. Trafc Safety and Human Behavior. Oxford, Elsevier Science.
Tay, R., 2002. The prisoners dilemma and vehicle safety. Some policy implications.
Journal of Transport Economics and Policy 36, 491495.
Vaa, T., 2007. Modelling driver behaviour on basis of emotions and feelings: intel-
ligent transport systems and behavioural adaptations. In: Cacciabue, C. (Ed.),
Modelling Driver Behaviour in Automotive Environments. Springer, Berlin, pp.
208232 (Chapter 12).