You are on page 1of 11

Accid. And. & Prev. Vol. 20. No. 3, pp. 187-197. 1988 rJool-4.575!88 s3.w + .

oo
Printed m Great Britam. 0 1988 Pergamon Press plc

AN APPLICATION OF LOGIT MODELS IN


ANALYSING THE BEHAVIOUR OF PEDESTRIANS
AND CAR DRIVERS ON PEDESTRIAN CROSSINGS

VELI ~IMANEN and RISTO KULMALA


TechnicalResearch Centre of Finland, Road and Traffic Laboratory, Espoo, Finland
(Received 7 February 1985; in revised form 7 January 1987)

Abstract-A multinomial logit model is used to examine pedestrian and driver reaction to
“encounters” occurring on pedestrian crossings. The probabilities of a driver braking or weaving,
and of a pedestrian continuing to cross in response to an encounter are identified for a variety
of pedestrian, environmental, and traffic conditions. Results indicate that the most im~rtant
explanatory variables included pedestrian distance from kerb, city size, number of pedestrians
simultaneously crossing, vehicle speed, and vehicle platoon size. It is felt that the model per-
formed well, should be applied in further studies, and could be a useful technique for identifying
the most hazardous situations and locations within an area, for planning relevant safety measures,
and for national research for developing traffic legislation.

INTRODUCTION

Traffic accidents involving pedestrians are a major safety problem throughout the world.
Most of these accidents, at least in Finland, occur in built-up areas, and on marked
pedestrian crossings. Many studies exist on pedestrian crossings. The major interests of
these studies are: traffic accidents, traffic conflicts, delays, gap acceptance, and behav-
iour.
Traffic accidents are analysed and related to various site features in order to improve
knowledge to be used in accident prevention [Fegan, 19781. One particular way to do
this is to make “before and after” accident studies [Landles, 19831.
Traffic conflicts give more information concerning the causes of and the chain of
events leading up to accidents on pedestrian crossings. Studies comparing different cross-
ing arrangements and layouts have been made in addition to “before and after” studies.
Estimates of the effects of various measures and road environment characteristics on
accident and conflict risks are presented in these studies [Girder, 1982; Hyden, 1981;
Kulmala, 19811.
Delays at pedestrian crossings are studied especially in order to derive warrants for
various crossing types [Griffiths, Hunt & Cresswell, 19763. Models describing delays
have also been developed [Griffiths, Hunt & Marlow, 19841.
Studies of gap acceptance can be linked to studies dealing with delays and traffic
safety [Wennell and Cooper, 19811.
Studies of the behaviour of car drivers and pedestrians on pedestrian crossings have
increased our understanding of the phenomenon. These studies, however, are limited
in scope and have not produced models which could be used in varying circumstances
[Camenzind, Hurlimann & Ktigi, 1978; Himanen, 1981; Katz, Zaidel & Elgrishi, 1975;
Keskinen, Koivumtiki & Salmenoksa, 1978; Keskinen, MetsHntghti & VBlitalo, 1979;
Kulmala, 1981; Mannfors, 1978; Riipinen, 19811.
As a conclusion of the studies performed so far on pedestrian crossings we think
that we still lack knowledge of the reasons behind the high accident records for pedestrian
crossings. In our opinion we should try to solve these problems starting from the bottom-
traffic behaviour.
The Finnish road traffic law requires the pedestrians to cross a street by using a
marked crossing whenever it exists in the vicinity. The law does not, however, state
clear priority rules for situations where pedestrians and drivers encounter each other on
these crossings. This results in greatly varying traffic behaviour in these encounters, and
poor anticipation of the behaviour of other road users.

187
188 V. HIMANEN and R. KULMALA

The main purpose of our study was to develop a model that describes the behaviour
of drivers and pedestrians in their encounters on pedestrian crossings. This model should
also enable us to evaluate the risks involved in these situations. A valid model would
help us in identifying the factors contributing to the lack of safety of the pedestrian
crossings, and thus form a basis for future safety work.
Drivers and pedestrians have to choose between different behavioural patterns in
their encounters in order to avoid colliding with each other. Multinomial logit models
deal with choice problems producing probabilities for choosing the various alternatives,
These models have been mainly applied in the choice of travel modes or destinations
[Ben-Akiva & Lerman, 1985; Henscher & Johnson, 1981; Rintamaki, 1980; Stopher et
al., 1984; Talvitie, 19773. We shall present in the following how logit models can be used
in analysing the behavior of drivers and pedestrians on pedestrian crossings.
This is primarily a pilot study but we do, however, also draw some conclusions
about the effects of different factors on behaviour and safety at pedestrian crossings.
These results are still preliminary and should be checked by further studies.

MULTINOMIAL LOGIT MODELS AND OUR APPLICATION

The multinomial logit (MNL) model calculates the probabilities of choosing different
alternative behavioural patterns. In the MNL model, individuals are assumed to choose
the alternative that yields the highest utility. The notion of utility is a convenient the-
oretical construct that provides an index of the relative levels of satisfaction associated
with the consumption of particular commodities per unit of time [Henscher & Johnson,
19811.
We studied the encounters between pedestrians and drivers on pedestrian crossings.
We defined an encounter as a situation where a pedestrian and a driver are approaching
each other in time and space so that they would collide unless one or both of them takes
evasive action to prevent the collision. The encounters between a driver and a pedestrian
can be regarded as a conflict between safety and freedom of movement [Himanen, 19811.
Drivers and pedestrians tend to continue their movement on their chosen course until
the perceived hazard of the encounter makes one or both of them react. The utility
function thus corresponds to the road users’ safety and freedom of movement.
Encounters are dynamic incidents where each participant affects others by his be-
haviour. The behaviour can be studied separately for pedestrians and drivers, but the
so-called nested MNL models [Ben-Akiva & Lerman, 198.51enable us to study the choices
of both road users simultaneously. Encounters result in accidents mainly when the road
users do not take evasive action, or do so too late. As our purpose was primarily safety
oriented we decided to try to find out in which situations the pedestrian and the driver
tend to neglect evasive action. This is only possible with separate models for pedestrians
and drivers, as our data does not include any situations where both road users neglected
evasive action. This is quite natural considering our definition of an encounter.
In the MNL model the utility of the ith alternative for the qth individual U,,, is
calculated on the basis of variables affecting the choice of the decision makers (road
users) :

&=I

The variables are usually socioeconomic (attributes of the decision maker) or traffic
system variables. In addition, alternative specific dummy variables are used.
An example of the use of alternative specific dummy variables is described below.
For alternative i = 1 the third variable x13qis equal to 1 for all individuals q, and x,3q
is equal to 0 for alternatives other than i = 1. Thus B,3 is an alternative-specific constant
(specific to alternative 1). The purpose of these variables is to describe the alternatives’
unmeasured attributes not captured by the other included variables.
Logit models analysing pedestrians and car drivers on pedestrian crossings 189
Other restrictions also exist concerning the variables to be used. If the value of xikq
of an individual q is the same for all alternatives (e.g. the age of the road user), the
variable has to be excluded from the utility of at least one alternative [Henscher &
Johnson, 19811.
We chose the variables to be used by analysing the encounters from different angles.
We divided the possible variables into three different categories:

(i) variables describing the physical environment,


(ii) variables describing the encounter, and
(iii) the characteristics of the road users.

The physical environment means on the one hand the locality, and on the other
hand the road environment of the crossing. Earlier studies in Finland have indicated
significant differences between encounters in various localities [Himanen, 19811. The
road environmental factors affecting the encounters we take as the road width to be
crossed by the pedestrian, the total road width, the existence of a refuge on the crossing,
and the priority regulations prevailing at the junction. These factors can affect the
perceived safety of the crossing, and the road users’ anticipation of different types of
behaviour.
The encounter variables are of different types. Due to Finnish road traffic laws, the
pedestrian’s and the driver’s behaviour probably depend on whether the pedestrian is
using the marked pedestrian crossing. The approach speed of the driver affects the risks
perceived by the pedestrian, and the driver’s possibilities of evasive action. If the driver
is making a turn at the junction, it affects his behaviour, and the driver’s use of direction
indicators helps the pedestrian in anticipating the car’s behaviour. Whether the driver
has the priority at the junction will also affect the driver’s willingness or observational
capabilities to react in the encounter with a pedestrian. The driver’s direction of approach
towards the pedestrian can influence the outcome of the encounter, and the number of
drivers and pedestrians involved in the meeting situation will have some effect on the
amount of risk perceived by the road users.
Pedestrians in a hurry (to catch a bus, to keep an appointment, etc.) probably accept
higher risks than other pedestrians. This variable is the counterpart of the vehicle’s
speed. The pedestrian’s distance from the kerb is a crucial factor in decision making. It
limits the pedestrian’s possible reactions, but also affects the driver’s judgement of the
situation.
The road user variables for the pedestrians include the age and sex of the pedestrian.
Whether the pedestrian is pushing a bicycle, baby carriage, etc., affects not only his
ability to react quickly, but also the driver’s visual observation and his judgements. On
the other hand, vehicles of different type (car, lorry, bus etc.) are observed and possibly
reacted to differently by pedestrians.
It is certain that other factors will also affect these encounters, and that individual
behaviour will be influenced by sudden impulses etc. which can only be treated as random
variation.
In the MNL model the coefficients B are calculated as maximum likelihood estimates
on the basis of the observation data. The observation data consists of observed choices
and the connected variables. The data unit is the choice made by one individual and the
values of the variables prevailing at the time of choosing.
The MNL model gives the probability of choosing alternative i from n possible
alternatives as:

The computer program that we used calculates the coefficients (B) in the utility functions,
and the standard errors and t-values for these coefficients. The program also gives some
190 V. HIMANEN and R. KULMALA
measures of the model’s goodness-of-fit. These include the proportion of correct pre-
dictions (the predicted choice matching the observed one), and the index of determination
or success index analogous with R2 of the regression analysis [Talvitie, 19771. We obtained
the original MNL computer program from the Technical University of Helsinki. We
programmed our own subprograms for the input of data, and for the calculation of choice
probabilities on the basis of the model output.

STUDY MATERIAL

The study material was collected from videotapes recorded in connection with other
studies on seven pedestrian crossings in the towns of Helsinki and Salo. Salo has about
ZO,O~, and Helsinki about 5~,~0 inhabitants. The crossings were at street junctions
in central areas. In Helsinki one of the crossings was in the city center and the others
in the fringe area. The observation hours were 11-12 AM and 4-5 PM. The streets to be
crossed were quite heavily trafficked, at least during the later observation hour.
The characteristics of the crossings are presented in Table 1. The street width to be
crossed by the pedestrian varied between 11 .O and 26.3 m. Three of the seven crossings
had refuges. On these crossings the distance between the refuge and the kerb varied
from 5.9 to 12.3 m.
We collected from the videotapes 799 encounters according to the definition given
earlier. The behaviour of the road users can be seen in Tables 2 and 3. Most of the
pedestrians involved decided to stop before stepping into the path of the approaching
vehicle. About 30% of the pedestrians just continued walking, and the vehicle driver
took the evasive action. In about 4% of the situations the pedestrian either retreated
or ran from the path of the vehicle. The vehicle drivers mostly continued to drive on in
the encounters. 10% of the drivers braked or wove slightly, and 15% braked/wove
clearly. 16% of the drivers stopped because of the pedestrian(s).
The variables collected for the model described the traffic environment, the en-
counter and the road users. Variables describing the traffic environment were locality
(Helsinki or Salo), the width of crossing (from kerb to island or from kerb to kerb),
and the width of the street (from kerb to kerb).
The encounter variables were its severity, possible traffic violation, the speed and
path of the vehicle, whether the pedestrian was crossing the street on a marked crossing,
whether he started his crossing from the kerb or refuge, the distance from his starting
place when one of the road users reacted, the direction from which the vehicle was
approaching, and whether the vehicle had the right-of-way when approaching the junc-
tion.
The road user variables were vehicle type, the number of approaching vehicles, the
age and sex of the pedestrian, the number of pedestrians in a group, whether the
pedestrian seemed to be in a hurry, and whether the pedestrian was pushing a baby
carriage or a bicycle etc.

THE FITTING OF THE MODEL

The first problem was to decide on the choice alternatives for the MNL model. We
first tried four alternatives for both pedestrians and drivers. The alternatives were those

Table 1. The characteristics of the crossings

Street Width of crossing


Crossing Locality width (m) Refuge (m)

1 Sal0 14.0 Yes 6.5 + 6.5


2 Helsinki 20.2 Yes 5.9 + 13.9
3 Helsinki 26.3 Yes 12.3 + 11.4
4 Helsinki 25.5 No 25.5
5 Helsinki 19.9 No 19.9
6 Helsinki 11.0 No 11.0
7 Sal0 14.8 No 14.8
Logit models analysing pedestrians and car drivers on pedestrian crossings 191

Table 2. The behaviour of pedestrians in Table 3. The behaviour of car


encounters drivers in encounters

Behaviour Frequency Behaviour Frequency

Walks on 244 Drives on 472


stops 525 Slight brake/weave 81
Retreats 11 Clear brake/weave 116
Has to run 19 stoos 130
Total 799 Total 799

presented in Tables 2 and 3. The resulting MNL models were very unstable, probably
due to the low number of observations for some alternatives.
Thus we decided to use for both drivers and pedestrians only two choice alternatives:
to continue or to react. Reacting for pedestrian includes stopping, retreating or having
to run. For drivers to react means to brake/weave slightly or clearly, or to stop. The
number of observed choices classified by these alternatives can be seen in Table 4. We
calculated separate models for pedestrian and driver behaviour.
As mentioned before, the value of an explanatory variable cannot be the same for
all alternatives. Because of this we had to set each variable to zero for one of the
alternatives. The choice of which variable should be set to zero for which alternative
has no effect on the final results of the model, but it naturally alters the form of the
utility functions.
The alternative specific dummy variable was used for the alternative of reacting.
This variable was set to zero for the alternative “continues” and to one for “reacts.”
The choice of variables was done in practice so that we kept some of the best
variables in the model and changed systematically the other variables in order to find
out whether they had any effect on the outcome of the model. The models always included
the pedestrian’s distance from starting place, locality, and naturally the dummy variable
for the alternative of reacting. We had five explanatory variables including the three
mentioned above for the models of driver and pedestrian behaviour. Thus we tried
systematically all other variables in turn as the two remaining explanatory variables, and
then selected the best ones for analysis. When selecting the variables we also checked
for the multicollinearity of the explanatory variables, and rejected some variable com-
binations for this reason.

DRIVER BEHAVIOUR

In the following we give an example of the models of driver behaviour. There were
five variables in the model:

Xl the number of pedestrians in the group,


x2 the pedestrian’s distance from starting place,
x3 locality (1 = Helsinki, 2 = Salo),
x4 speed of the vehicle (kmph, 20 m before the crossing),
X5 dummy variable (0 = driver continues, 1 = reacts).

All coefficients differ significantly from zero (see Table 5). The utility functions [formulas
(2) and (3)] and the probabilities for continuing (4) and for reacting (5) can be calculated
on the basis of these coefficients. According to the t-values, the pedestrian’s distance

Table 4. The alternative choices used in the model

Driver

Pedestrian Drives on Reacts

Walks on - 244
Reacts 472 83

A4P 20:3-c
192 V. HIMANENand R. KULMALA

from starting place (x2) is the best explanatory variable, and then are locality (x3) and
the alternative specific dummy variable (xJ.
The utility functions are:

(i) for continuing: U,(X) = - 1.543 xj + 0.40 xJ

(ii) for reacting: U?(x) = 0.874 x, + 0.071 x2 -3.858 .x5.

The probabilities are:

e”)
(iii) for continuing: e”l + eL’z’

e”:
(iv) for reacting: e”l + c”2’

Formulas (2) and (3) for the utility functions show that we separated the different
alternatives by setting variables .x1, x2, and x5 to zero for the alternative “continues,”
and variables x3 and X~ to zero for the alternative “reacts”.
The proportion successfully predicted was 0.72 and the success index was 0.20. An
equally good model was achieved when we replaced the number of pedestrians (variable
x1) with the number of approaching vehicles. All other variables produced models with
lower values for the indices for the goodness of the model.
In the following we study the effects of different variables on driver behaviour. All
such variables were included whose coefficients were statistically significant (see Table
6). The coefficients are from models, where variables x2-x5 remained the same as before
and variable x, was replaced with the variable under study. The coefficients for the
remaining variables in these models differed in the following range:

x2 = 0.066 . . . 0.072,

x3 = - 1.38 . . -1.64,

x4 = 0.037 . . . 0.045.

The proportion successfully predicted was 0.71 in all of these models. The effects
of different variables on the choices of the driver are shown in Fig. 1. The probability
curves were derived by calculating the probability of driver reaction with formula 4 using
different values for the explaining variables.
In the set of plots on the right in Fig. 1 we can see the probability of driver reaction
if the driver is travelling at the speed of 50 kmph. On the left the same plots are shown
for vehicle speeds of 10 kmph. Over 90% of the speeds were within this range (lo-50
kmph).
In Helsinki the reacting probability of a driver travelling at the speed of 50 kmph
is nearly zero if there is only one pedestrian approaching, and the pedestrian has not
stepped on the street from the kerb (the set of plots on the upper right corner). If the

Table 5. The coefficients for different variables in the model


for driver behaviour

Standard error
Variable Coefficient of coefficient f-value

XI 0.874 0.197 4.43


X2 0.071 0.006 12.52
Xl - 1.543 0.204 - 7.55
X4 0.040 0.007 5.63
XI - 3.858 0.478 - 8.08
Logit models analysing pedestrians and car drivers on pedestrian crossings 193
Table 6. The coefficients for different variables in the model for driver
behaviour

Standard error
Variable Coefficient of coefficient f-value

The number of vehicles - 0.545 0.116 -4.69


Vehicle type
(1 = other, 2 = car) 1.357 0.375 3.62
Pedestrian on marked
crossing (1 = no, 2 = yes) 2.476 0.796 3.11
Driving direction
(1 = straight, 2 = other) 0.587 0.249 2.36

10 kmph 50 kmph
s=sdo
H = Helsinki

‘,‘,“,“. 1 pedestrian

E L pedestrians

E 5 vehicles

1,O
/@ #
’ 0’
9’ ,0
“f,, other than car
05 &!J
RN z car

~
r’
OO 5 10

1.0
4
H / z

0.5 ,f ,4 ‘11-
I_--.
outside crossing

/ * - on marked crossing
,‘/’
~
OO 5 10

vehicle driving
_--_
---* straight

- other

PEDESTRIAN’S DlSTANCE FROM STARTING PLACE 1m 1


Fig. 1. The probability of driver reaction and its dependence on different variables.
194 V. HIMANENand R. KULMALA

pedestrian is already on the crossing, the probability of driver reaction increases as the
pedestrian’s distance from the kerb increases. When the pedestrian has already walked
five meters from the kerb (or refuge) the probability of driver reaction is about 0.5. In
a smaller town, Salo, the probability of driver reaction is higher by 0.2-0.4. When the
speed of the vehicle is 10 kmph, the probability of driver reaction is higher by 0.2-0.4
than when the speed is 50 kmph. The results concerning the pedestrian’s distance,
locality, and driving speed are about the same for all sets of plots in Fig. 1.
The top three sets of plots deal with variables that could be classified as power-
ratio variables. They include the numbers of pedestrians and vehicles, and vehicle type.
The probability of driver reaction increases as the number of pedestrians increases. With
four pedestrians the probability of the driver’s reaction is higher by about 0.5 than with
just one pedestrian (the set of plots at the top).
As the number of approaching vehicles increases the probability of the first driver
reacting decreases. In a group of five vehicles the probability of reacting is lower by
about 0.4 than with just one vehicle (the set of plots second from the top). The prob-
abilities of car drivers reacting are higher by about 0.3 than those of heavy vehicle drivers
(the set of plots third from the top).
If the pedestrian or pedestrians are crossing the street on a marked pedestrian
crossing, the probability of the driver reacting is higher by about 0.4 than when the
pedestrian is crossing the street elsewhere. The probability of the driver’s reacting is
especially low in Helsinki for encounters with pedestrians crossing the street outside the
marked crossing (the set of plots second from the bottom). For a driver turning at a
junction the probability of reacting is higher by about 0.05 than for a vehicle driving
straight on, regardless of the speed of the vehicle (the set of plots at the bottom).

PEDESTRIAN BEHAVIOUR

In the following we give an example of the models of pedestrian behaviour. The


following variables were used:
x, the number of pedestrians in the group,
x2 the pedestrian’s distance from starting place,
xi locality (1 = Helsinki, 2 = Salo),
x6 the number of vehicles approaching,
x, dummy variable (0 = pedestrian continues,
1 = pedestrian reacts).
All coefficients in the model differed significantly from zero (Table 7). The best
explanatory variable was again the pedestrian’s distance from starting place (x2), and
the second best was the alternative specific dummy variable (x7).
The utility functions are:

(i) for continuing: U, (x) = 1.005 xX - 0.726 x6, (6)


(ii) for reacting: U2 (x) = - 1.239 xi - 0,069 x7
+ 4.243 x_i. (7)

The choice probabilities for the alternatives were calculated as in formulas (4) and (5)

Table 7. The coefficients for different variables in the model


for pedestrian behaviour

Standard error
Variable Coefficient of coefficient t-value

XI - 1.239 0.212 -5.86


x: - 0.069 0.005 - 12.41
xx l.Oil.5 0.215 4.68
& - 0.726 0.145 -5.02
x7 4.243 0.542 7.83
Logit models analysing pedestrians and car drivers on pedestrian crossings 195
from the utility functions (6) and (7). The proportion successfully predicted was 0.76
and the success index was 0.18. In addition to the variables in Table 7, significant
coefficients were obtained for four other variables as replacements for x1 (Table 8).
Figure 2 shows the probability of pedestrians continuing to walk (and not to react),
and its dependence on the explanatory variables. In Salo the pedestrian’s probability of
continuing is clearly higher than in Helsinki. The same kind of difference in pedestrian
behaviour between Helsinki and other Finnish towns has been noted in previous studies
[Himanen, 19811. The pedestrian’s distance from starting place at the time of decision
making (concerning evasive action) very clearly affects the choice of the pedestrian. If
the pedestrian is still on the kerb, the probability of crossing in an encounter situation
is quite low. The probability increases the further the pedestrian walks from the kerb.
The power-ratio variables also affect the behaviour of the pedestrian. The probability
of a group of four pedestrians continuing to walk is distinctly higher than that of a single
pedestrian, and always higher than 0.7 (the set of plots in the top left of Fig. 2). If more
than one vehicle is involved in the encounter, the probability of continuing is lower than
in the case of only one vehicle (top right). The probability of continuing to walk is lower
by about 0.2 when the approaching vehicle is a lorry or a bus instead of a car (middle
left).
Even though the speed of the vehicle had a significant coefficient it does not have
any major effect on pedestrian behaviour (middle right). This could be due to the
pedestrians’ difficulties in estimating the speeds of approaching vehicles. The driving
direction of the vehicle affects the probabilities of the pedestrian by about 0.2. The
probability of continuing to walk is higher when the vehicle is turning at the junction
than when the vehicle is driving straight through it (bottom left).
Of the pedestrian variables, only the sex of the pedestrian was significant but its
effect is quite small. The probability of male pedestrians continuing to walk is higher
by about 0.1 than that of female pedestrians (bottom right). This is consistent with earlier
studies where male pedestrians have been found to accept higher risks than do females
[Girder, 19821.

CONCLUSIONS

A comparison of Figs. 1 and 2 enables us to assess the safety of encounters between


pedestrians and vehicles. If the probability of the pedestrian continuing to walk is lower
than the probability of a driver reaction, then a safety margin exists in the encounter.
If the relation is the other way round, then a hazard exists. The comparison shows that
a safety margin usually exists when the speed of the approaching vehicle is low. If the
speed of the vehicle 20 m before the pedestrian crossing is 50 kmph, there is no safety
margin in the encounter. In many of these cases the probability of the pedestrian con-
tinuing to walk is higher than the probability of the driver reacting. Encounters are much
safer in Salo than in Helsinki because the safety margins are greater.
The most hazardous situation according to the model (with the values used for the
variables studied) is the one between a group of pedestrians and a vehicle approaching
at the speed of 50 kmph.

Table 8. The coefficients for different variables in the models for pedestrian
behaviour

Standard error
Variable Coefficient of coefficient f-value

The speed of vehicle


(20 m before cross.) 0.039 0.008 4.95
Driving direction
(1 = straight, 2 = other) - 1.102 0.241 -4.57
Vehicle type
(1 = other, 2 = car) - 1.555 0.452 -3.44
Sex of pedestrian
(1 = male. 2 = female) 0.516 0.201 2.57
V. HIMANEN and R. KULMALA

Speed of the vehicle


E 10 kmph

T_=_y 50 kmph

0 5 IOm 50Wm

1.0 1.0

0.5 OS
- female
- pedestrian
“-““other
____.

0 0
- 1Clm 0 5 10 m

PEDESTRIANS DISTANCE FROM STARTING PLACE

Fig. 2. The probability of the pedestrian continuing to walk and its dependence on different variables
(only one vehicle unless stated othewise).

Surprisingly, at least in our opinion, many of the variables that we expected to affect
these encounters did not have significant effects in our models. Such variables were
street width, the existence of a refuge, the priority situation of the vehicle at the junction,
and most of the pedestrian variables. The problem with the pedestrian variables was the
concentration of observations in one category, such as being of adult age, not being in
a hurry, not pushing a bicycle or baby carriage. Correlations between the explanatory
variables could cause the lack of significance of some of these variables, and the sample
size of 799 observations is probably insufficient to assess the significance of the variables.
The alternative specific dummy variables were among the three most significant
variables in all models. This means that our models do not include all relevant explanatory
variables. Probably the most essential factor missing from the models is the reaction of
the other road user. This could be included by using the nested logit models mentioned
earlier, but their use would require a new definition of an encounter, and thus a new
observation data base.
According to this study, the logit model can be applied in the analysis of encounters
between pedestrians and vehicles. Further and more extensive studies should be made
in order to find out the general applicability of the conclusions made in this study.
The model can be used in national studies, i.e. in basic research for traffic legislation.
Logit models analysing pedestrians and car drivers on pedestrian crossings 197
The model can also be used in studies within just one town, in which case the analysis
of the data material could point out the most hazardous situations and locations. Anal-
ogously the data material can be collected at just one pedestrian crossing and then the
most hazardous situations can be analysed with the model for the purpose of planning
relevant safety measures.

Acknowledgements-We thank the Central Organisation for Traffic Safety in Finland for their grants enabling
us to conduct this study, and Graham Grayson for editing the paper.

REFERENCES

Ben-Akiva M. and Lerman S. R. Discrete Choice Analysis: Theory and Applications to Travel Demand. MIT
Press, Cambridge, MA. 1985.
Camenzind J., Hurlimann F. W. and Kigi B. Konfliktsteele Fussglngersstreifen (Teil 2). Zeitschrift fiir Ver-
kehrssicherheit 24, 2, 1978.
Fegan J. C. Major Engineering approaches toward increasing pedestrian safety. Public Roads 42(3), S-90,
1978.
Garder P. Gaendes slkerhet i trafiksignaler (Pedestrian safety at traffic-signals). University of Lund, Lund
Institute of Technology, Department of Traffic Planning and Engineering, Bulletin 43, Lund, 1982.
Griffiths J. D., Hunt J. G. and Cresswell C. Warrants for zebra and pelican crossings using a minimum waiting
cost criterion. Traffic Engng and Control February, 17,59-62, 1976.
Griffiths J. D., Hunt J. G. and Marlow M. Delays at pedestrian crossings. Traffic Engng and Control July/
August, 25, 365-371, 1984.
Henscher D. A. and Johnson L. W. Applied Discrete Choice Modelling. Croom Helm Ltd., London, 1981.
Himanen V. Individual and governmental choices between freedom and safety. International Symposium on
Surface Transportation System Performance. U.S. Department of Transportation, Proceedings, Volume
II, pp 512-520, Washington, 1981.
Hyden C. Oskyddade trafikanters slkerhet pa stomgator (The safety of unprotected road-users on arterial
streets). University of Lund, Lund Institute of Technology, Department of Traffic Planning and Engi-
neering, Lund, 1981.
Katz A., Zaidel C. and Elgrishi A. An experimental study of driver and pedestrian interaction during the
crossing conflict. Human Factors 17, 514-527, 1975.
Keskinen E., Koivumlki T. and Salmenoksa L. Suojatietutkimuksia 1. Turun yliopisto, Psykologian tutki-
muksia, 28, 1978.
Keskinen E., Metslntlhti P. and Valitalo T. Suojatietutkimuksia 2. Turun yliopisto, Psykologian tutkimukisia,
32, 1979.
Kulmala R. Mannerheimintien suojateiden turvallisuus (Traffic safety at pedestrian crossings on Manner-
heimintie). Technical Research Centre of Finland, Research Notes 25, 1981.
Landles J. R. The overall effect on accidents at sites where zebra crossings were installed. Traffic Engng and
Control 24, 9-l 1, 1983.
Mannfors J.-E. Autoilija ja jalankulkija suojatietl llhestymlssl. Helsingin yliopisto, yleisen psykologian
osasto. Pro gradu-tutkielma, 1978.
Riipinen M. Suojatiekayttlytyminen erailll maaseutusuojateilll. Helsinki, 1981. Liikenneturva. Tutkimuso-
saston julkaisuja 42, 1981.
Rintamaki H. Paakaupunkiseudun liikennetutkimus. LITU-76. Otaniemi, 1980. Teknillinen korkeakoulu,
liikennetekniikka, Julkaisu 50.
Stopher P. R., Ohstrom E. G., Kaltenback K. D. and Clause D. L. Logit mode-choice models for nonwork
trips. Transpn. Res. Rec. 987, 75-81, TRB 1984.
Talvitie A. Travel models. Lecture notes at the Helsinki University of Technology, January, 1977.
Wennell J. and Cooper D. F. Vehicle and driver effects on junction gap acceptance. Traffic Engng and Control
22, 628-632, 1981.

You might also like