Accident Analysis and Prevention 87 (2016) 117–126

Contents lists available at ScienceDirect

Accident Analysis and Prevention

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

Hotspots and social background of urban traffic crashes: A case study

in Cluj-Napoca (Romania)
József Benedek a,b , Silviu Marian Ciobanu b,∗ , Titus Cristian Man b,c
a
University of Miskolc, World and Regional Economics Department, 3515 Miskolc-Egyetemvaros, Hungary
b
Babes-Bolyai University of Cluj-Napoca, Faculty of Geography, Clinicilor Street, 5-7, 400006 Cluj-Napoca, Romania
c
Eastern Wyoming College, Torrington, WY 82240, USA

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

Article history: Mobility practices have changed dramatically in Romanian towns over the last 25 years, following the
Received 12 May 2015 collapse of socialist mobility restrictions. Urban areas like Cluj-Napoca are facing both increasing immi-
Received in revised form 15 October 2015 gration and car mobility, and therefore increasing levels of road traffic crashes.
Accepted 20 November 2015
The analysis of traffic crashes is one of the most important elements for improving the road safety
Available online 9 December 2015
policy. This paper is divided in two parts. In the first one, the authors focus on identifying the traffic crash
hotspots along the street network, while in the second part they discuss the social background of road
Keywords:
traffic crash occurrence.
Traffic crashes
Hotspots
The first step in analyzing traffic crashes is to determine crash hotspots. A four-year record (2010–2013)
SANET provided by the Traffic Department of the General Inspectorate of Romanian Police (GIRPTD) was used.
Kernel Density Estimation As a method of hotspot determination, the Kernel Density Estimation tool was employed, in the frame of
Social vulnerability the spatial analysis along network (SANET). The outcome was the hotspot map of traffic crashes in Cluj-
Napoca. The results have revealed 4 categories of street segments: not-dangerous, low-dangerous, medium-
dangerous and high-dangerous. Based on this classification, at least 4 dangerous areas were identified,
located at the city entrances-exits (in the West, North-West and East) and the city center (the most
dangerous zone).
The second part of the paper focuses on social groups involved in car crashes. The following are con-
sidered: age, gender and blood alcohol concentration of the person (driver or pedestrian) found guilty
for every individual crash.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction injured, 1120 seriously injured and 242 deaths. According to

The issue of traffic crashes is of major interest at national and
international level since these lead to loss of human lives, causing
significant material damages and potential environmental degra-
dation. Traffic crashes analysis one of the most important elements
to improve traffic safety policies.
It is very important to know where and when do crashes occur
and which are the vulnerable social groups in order to help the
enforcement authorities take effective measures to reduce their
number.
According to the General Inspectorate of the Romanian Police,
Traffic Department (GIRPTD), no less than 3792 road crashes were
recorded in Cluj County between January 1st, 2010 and December
31st, 2013. These totaled 4129 casualties of which 2767 slightly
∗ Corresponding author.
E-mail address: silviumarianciobanu@gmail.com (S.M. Ciobanu).
European Commission reports (2015), most of the traffic crashes
in Romania are recorded in urban areas. Because of the low level of
infrastructure improvement and strong increase of traffic, Roma-
nia is the “leading” EU country in traffic crashes statistics (European
Commission, 2015).
In this paper, the authors intend to answer the questions
“where do traffic crashes occur?” and “which are the social groups
vulnerable to the occurrence of traffic crashes?” in the Cluj-
Napoca municipality, the economic core area of Cluj county, one
of the leading areas of Romania in traffic crashes statistics. Thus,
the paper is structured into two distinct parts. The first part
deals with identification of traffic crash hotspots using GIS (Geo-
graphic Information System) techniques, while the second part is
dedicated to delineation of social groups vulnerable to crash occur-
rence.
GIS played an important role in traffic crashes analysis and
prevention in recent years, as it constantly expanded, providing
researchers with more and more diverse opportunities for analysis

http://dx.doi.org/10.1016/j.aap.2015.11.026
0001-4575/© 2015 Elsevier Ltd. All rights reserved.
118 J. Benedek et al. / Accident Analysis and Prevention 87 (2016) 117–126
every year. Therefore, some authors used GIS for simple linear
analyses and development of maps combining various parame-
ters such as the number of crashes, number of deaths and injuries
or material damage (Gundogdu, 2010; Truong and Somenahalli,
2011; Pulugurtha et al., 2007; Qin et al., 2013). Others used it for
spatial analyses and identification of risk areas (Ivan et al., 2015;
Erdogan, 2009; Aguero-Valverde and Jovanis, 2006; Vandenbulcke
et al., 2014), while others employed it for spatio-temporal analy-
ses (Ivan and Haidu, 2012; Plug et al., 2011; Prasannakumara et al.,
2011).
Traffic crashes often form clusters known as hotspots (Xie and
Yan, 2008). Hotspots identification leads to the discovery and
removal of common causes of crashes, thus improving traffic safety
(Steenberghen et al., 2010). Several methods for determining the
traffic crashes hotspots have been analyzed and implemented in
literature over the last decades (Li et al., 2007; Erdogan et al., 2008;
Montella, 2010; Okabe et al., 2009; Okabe and Sugihara, 2012; Xie
and Yan, 2013; Yu et al., 2014). One of the most frequently used
methods to determine traffic crash hotspots is the Kernel Density
Estimation (Anderson, 2009; Thakali et al., 2015; Bíl et al., 2013;
Flahaut et al., 2003). This is known in literature in two forms: Planar
Kernel Density Estimation (PKDE) and Network Kernel Density Esti-
mation (NKDE). The PKDE method is more rarely used in the study
of traffic crashes because it calculates the density within a homoge-
neous 2D space using the Euclidean distance between the crashes,
thereby altering the result. It is a well-known fact that crashes occur
along a network, also known in the literature as “network space”
(Yamada and Thill, 2004). Hence, the NKDE method was developed,
which calculates the density of events strictly along the network
(Xie and Yan, 2008). This method was included in a toolbox (Okabe
et al., 2009), called SANET (spatial analysis along network).
In terms of the theoretical framework of vulnerability, we men-
tion only one of the most influent theory developed by Watts and
Bohle (1993), called the realistic theory of vulnerability. Thereby,
vulnerability is determined by the interconditionings between
institutions, power relations and historical framework. In this con-
text, vulnerability is determined by distribution of commodities in
time and space, by manner of occurrence and by property rights,
which result in material deprivation of some classes to the benefit
of others (Craddock, 2000).
Unlike physical vulnerability, social vulnerability is often
ignored in the vulnerability studies (Flanagan et al., 2011; Siagian
et al., 2014). Social vulnerability refers both to social groups (social
vulnerability) and to territorial entities such as regions (regional
vulnerability). Some authors (Watts and Bohle, 1993; Bohle, 2001;
Oulahen et al., 2015) have even defined a vulnerability space con-
sisting of the structure determined locally and historically by the
interrelations between different social risk factors. However, in
this paper the authors have adopted the perspective represented
by Flanagan, who considered that social vulnerability referred to
socioeconomic factors affecting the resilience of communities to
hazards (Flanagan et al., 2011). In line with this perspective, the
authors consider crash hotspots as being areas which require spe-
cial emergency actions. They operate simultaneously as areas of
high social vulnerability to car crashes. The authors have used
hotspots in order to identify critical spaces (spaces of vulnerability).
2. The study area and the database
The study area of this research is the street network of Cluj-
Napoca municipality. This is the capital of the Cluj County and
it ranks second in the top cities in Romania, after Bucharest, in
terms of number of residents, with 324,576 inhabitants accord-
ing to the General Population Census in 2011 (National Institute
of Statistics, 2012). An important aspect of this research is the
fact that the metropolitan area of the city underwent a strong
population growth over the last decade, as well as an increase in
the level of motorization and urban mobility (Benedek et al., 2013).
This represents the most important economic area of Romania,
after the capital city Bucharest, which generates both a positive
migration balance as important intra-urban “travel to work” type
of mobility as well (Soaita, 2013). Contrasting with the dynamic
economic and demographic development of the examined urban
area, there was little improvement of urban transportation infra-
structure; practically the road structure is the same (Benedek et al.,
2013).
The findings and conclusions of this research are representa-
tive not only for Cluj-Napoca, but for all major cities of Romania
because the main factor of increasing urban mobility is related to
the 1989 system change. Moreover, the results are comparable to
all other cities of the world where similar situations can be found:
old and overcrowded transportation infrastructure, strong increase
of urban mobility over a short period of time, increasing population,
and reduced local administrative capacities.
2.1. Crash data
Statistically speaking, for such a study to have credible results,
traffic crash records for a period of minimum 3 years are needed
(Steenberghen et al., 2010). The traffic crash records in the city of
Cluj-Napoca over a 4-year interval (2010–2013) were used in this
study. These are the only official data made available by the Gen-
eral Inspectorate of Romanian Police, Traffic Department (GIRPTD).
It should be noted that not all traffic crashes are recorded in the
database, but only those of substantial gravity, which require the
response of a traffic police squad.
Thus, the police go to the crash site only if there is at least one
slightly injured person. Otherwise, the insurance companies man-
age the crash. Special cases are represented by collisions between 3
or more vehicles and the auto-pedestrian crashes, when the police
go to the scene even if there are no casualties. The data collected
by the insurance companies (minor traffic crashes – no casualties)
could improve the impact of such a research. However, this data is
unfortunately not available to the public.
The database contains detailed information about each particu-
lar crash, such as GPS coordinates, date and time of occurrence, the
features of the street at the crash site (intersection, road turn, etc.),
signaling means present in the vicinity of crash location, lighting
conditions, main cause, manner of occurrence, number of vehicles
involved, number of injured (slightly or severely), number of dead,
age, gender and blood alcohol concentration of persons involved,
etc. Unfortunately, the database contains no data about ethnicity,
socio-economic status, area type or vehicles type.
It is worth mentioning that data collected by police at the crash
site is updated for 90 days, depending on how the investigation
progresses and depending on how the condition of victims evolves.
For example, a slightly injured victim can become severely injured,
a severely injured can die, a person declared guilty can become
innocent and the other way around, etc.
The 5 municipalities and the town of Huedin in Cluj County
have recorded a number of 2312 crashes over the 4 analyzed years
(according to RDGIRP). The county capital, the city of Cluj-Napoca
is clearly in the leading position, with a number of 1466 crashes.
This study considered both auto-auto crashes and auto-
pedestrian crashes. During the survey period, 1466 crashes were
reported, involving a total of 2510 participants (drivers, passengers,
pedestrians), among which 1331 victims (934 slightly injured, 364
seriously injured and 33 dead). Of the 2510 participants, the police
could determine the guilt for a number of 1159 persons.
The selection of variables for the model was restricted by the
availability of data, which is highly reduced for the local level anal-
ysis in Romania. The authors of this study have considered only
 J. Benedek et al. / Accident Analysis and Prevention 87 (2016) 117–126
those factors affecting the occurrence of crashes, where reliable
and official data was available.
2.2. Traffic data
The traffic was used in this paper as a non-behavioral factor of
traffic crashes occurrence (causality). The database was composed
of the total counts (24 h) and time-interval traffic volume for 56
streets of Cluj-Napoca, including the most important traffic arteries
of the city, along which no less than 782 crashes have occurred
(2010–2013). The total counts (24 h) and time-interval traffic in
the most important 22 intersections were also used (General Urban
Plan Cluj-Napoca, 2013). It is worth mentioning that the passenger
car equivalent (PCE) per hour daily unit was used to quantify the
traffic volume (Al-Kaisy et al., 2002; Yeung et al., 2015).
3. Methodology
3.1. The simple linear regression and the Pearson correlation
coefficient
The two concepts are used in this research to determine the cor-
relation between traffic crashes and road traffic. The simple linear
regression shows the dependence relationship between the cause
and the effect of an event. The first step in calculating this regression
is the distribution of the 2 variables in a XOY coordinate system,
the simple linear regression thus taking the following form (the
regression line):
Y = aX + B
where X is the cause and Y is the effect. In our case X is the amount
of traffic and Y is the number of traffic crashes.
Pearson correlation coefficient shows the intensity of the linear
dependence between two variables (in our case, crashes and traffic)
and it is calculated by the formula (Lee and Wong, 2001):
n
x y /n − X̄
i=1 i i
· Ȳ
r=
Sx Sy
where r is Pearson correlation coefficient, X and Y represent the
average of x and y Sx and Sy represent the standard deviation of x
and y, calculated by the formulas (Lee and Wong, 2001):

n authors decided to use NKDE because this model fitted perfectly the
n=1
x2
Sx = − X̄ 2
n on a linear network (road network).

n The formula that calculates NKDE is very similar to that of PKDE
n=1
y2
Sy = − Ȳ 2
n
The linear simple regression was used as result of some tests
conducted by means of the CurveExpert Pro 2.2 application (Hyams,
2005; Poirier et al., 2008; Hinojosa et al., 2008). More than 100
regression models were applied (on our data set), each one achiev-
ing a certain score. These scores were classified depending on the
determination coefficient R2 , correlation coefficient and standard
error. After their evaluation, the highest scores, which also had the
best statistical significance, were included in the paper and were
represented by the simple linear regression.
3.2. Kernel Density Estimation
Kernel Density Estimation is one of the most common discov-
ery (definition) techniques of the hotspots of point-events. Both
Planar (2-D) Kernel Density Estimation (PKDE) and Network Ker-
nel Density Estimation (NKDE) are useful in identifying clusters, but
the second method has several advantages over the first since it is
less dependent on the input parameters. The largest inconvenience
119
when using Planar Kernel Density Estimation in determining and
analyzing the traffic crash hotspots is that PKDE uses Euclidean
distance between events. This is not quite correct (applicable)
knowing that the traffic crashes occur along a road network. There-
fore, NKDE was used in the present study for the delineation of
traffic crash hotspots, but to better understand the operating prin-
ciple of this method the authors decided to make a comparison
between PKDE and NKDE.
3.2.1. Planar (2-D) Kernel Density Estimation (PKDE)
PKDE analyses the required events within a 2-D homogeneous
space and it is expressed as follows (Xie and Yan, 2008):

n
1
d 
is
(s) = k (5)
r 2 r
i=1
where (s) is the density in the location s, r is the radius (bandwidth)
of KDE (only points with radius r are used to estimate (s), k is the
weight of point i, at a distance dis of the location s. Most times,
k is modeled as a function, also called the kernel function of the
relationship between dis and r. Therefore, instead of being a uniform
function, in which every point i from the radius r has equal weight,
KDE uses a function model that takes into account the decrease
of the i point weight depending on its distance from the location
s (distance decay effect, Xie and Yan, 2008). In other words, the
longer the distance between a point and the location s, the lower
the weight of that point. The final density in location s is calculated
by adding all points in the search radius with a lower or higher rate
(depending on their distance from s).
(1) The literature mentions a series of formulas for the calculation
of the kernel function, such as (Levine, 2004; Schabenberger and
Gotway, 2005; Gibin et al., 2007): the Gaussian function, the Quartic
function or the minimum variation function.
Most of the literature indicates that PKDE has two key param-
eters: the kernel function k and the bandwidth r, and choosing
the search radius is more important than choosing the calculation
function of k (Bailey and Gatrell, 1995; O’Sullivan and Unwin, 2002).
(2)
3.2.2. Network Kernel Density Estimation (NKDE)
NKDE is an extension of PKDE that calculates the density of
point type events on a linear unit (in a network) rather than
on a 2-D homogeneous area unit, as was the case of PKDE. The
(3) research on traffic crashes, which were point-type events occurring
(4) and is given also by Xie and Yan (2008):

n
1
d 
is
(s) = k (6)
r r
i=1
where (s) is the density in the location s and r is the search radius
(bandwidth) of KDE (only points with radius r are used to estimate
(s), k (the kernel function) is the weight of point i, located at a dis-
tance dis of location s. As with PKDE for the calculation of the kernel
function, the literature includes a wide range of formulas such as
Gaussian, Quartic, Conic, negative exponential, and epanichnekov
(Levine, 2004; Gibin et al., 2007).
3.3. Spatial analysis along network (SANET)
SANET includes a toolkit that integrates and works together with
the ArcMap 10.1 software and it is used specifically for network
spatial analysis (Okabe and Sugihara, 2009). As one may predict,
we used SANET, more precisely the “Kernel density estimation” tool
to implement the NKDE method in the calculation of traffic crash
120 J. Benedek et al. / Accident Analysis and Prevention 87 (2016) 117–126

regions, administrative units of different ranks). Efforts are made

to explain the different territorial models resulting from mapping
and possibly to calculate the regression functions and correlation
coefficients between various indices, in order to obtain high validity
explanation models.
At this stage of the research, the authors cannot calculate an
accurate social vulnerability index to road traffic crashes due to the
fact that there is no data available about the social characteristics of
the entire population of drivers and pedestrians. This paper intends
only to provide a description of the social characteristics of drivers
and pedestrians involved in car crashes.
Fig. 1. The three forms of the kernel function (Young and Park, 2014). The second part of this study describes the social groups (pede-
strians and drivers) vulnerable to the occurrence of traffic crashes.
hotspots. To calculate NKDE, SANET uses the basic formula given Therefore, the “pyramids” of vulnerable social groups were built.
by Xie and Yan (2008), but unlike these authors who use formulas The notion of “pyramid” was taken up from demography (Hobbs,
like Gaussian or Quartic for the calculation of the kernel function 2004; Dannefer and Bhatta, 2015; Horiuchi, 2015), more exactly
(k), SANET uses a different formula provided by Okabe et al. (2009): from the stratification of population by age and gender. Thus, only
⎧ ⎫ the persons responsible for causing a traffic crash (drivers or pede-
⎪
⎪ k(x), for − h ≤ x ≤ 2d − h ⎪
⎪ strians) were considered when building these pyramids. They were
⎪
⎨ n−2 ⎪
⎬
k(x) − k(2d − x), for 2d − h ≤ x ≤ d divided by gender and age into 5-year age groups (except for the
Ky (x) = n (7) 0–10 year-old group), and each group was related to the city pop-
⎪
⎪ ⎪
⎪
⎪
⎩ 2 k(x), ⎪
⎭ ulation divided by the same gender and age groups (guilty persons
for f ≤ x ≤ h
n per 10,000 inhabitants).
where k(x) is the basic kernel function, y is the center of the ker-
nel function, x is a point located on the network, h is the bandwidth
(meters), n is the node degree and d is the shortest path from y to
x (meters). 4. Results and discussion
The density in the point of interest is calculated according to
equation (Okabe et al., 2009): 4.1. The traffic–crashes correlation
 2d−h  d  2 − n 
Before establishing the traffic crash hotspots in the city sub-
D(O) = k(−y)dy + k(−y) − k(−2d + y) dy
n jected to study, the authors intended to examine the statistical
−h 2d−h
 relationship between the crashes and one of the most important
h 2 factors triggering the crashes, the traffic, for the purpose of seeing
+ (n − 1) k(−y)dy (8)
n to what extent this could explain the hotspots.
d
The hourly distribution of crashes and traffic volume (Fig. 3)
where D(O) is the density at the origin. shows that the maximum number of crashes is recorded both in
Fig. 2 represents the graphical expression of the equation (12), intersections and the analyzed streets between 17:00 and 17:59 h,
which indicates that the kernel function is divided equally from as the traffic volume is very high in this time interval (reaching the
node v toward the 3 lines that start from it (l2 , l3 , l4 ), and Fig. 1 maximum point in the case of intersections). Another observation is
represents the graphical expression of the 3 ranks of the kernel that during the morning hours, even if the traffic is high, the num-
function (11). ber of crashes does not reach the same level. An explanation for
this could be the “fatigue” factor, which begins to set in during the
3.4. Social background of crashes evening hours. The graph (Fig. 3) also reveals a significant differ-
ence between the number of crashes in intersections and those in
There are no undeniable indicators to express vulnerability the streets, although the traffic volume does not differ very much.
or ambiguous strategies to eliminate it, which creates significant This can be explained by another important factor which influences
methodological difficulties (Flanagan et al., 2011). Thus, there is the occurrence of crashes, namely the speed. In intersections, the
yet no universally accepted methodology to calculate social vulner- speed decreases significantly due to corresponding signaling (traf-
ability to hazards, as some of its dimensions are extremely difficult fic lights, roundabouts, etc.) and therefore the number of crashes is
to quantify, while there is little knowledge on their reliability lower.
(Tate, 2013). To this one might add the large number of indicators To see the dependence relation between the two variables
expressing these dimensions. Most often, complex indices are built (crashes and traffic), we used the simple linear regression as shown
by aggregating multiple indicators (possibly weighting their impor- in the charts of Fig. 4a and b. In the design of the two charts we have
tance). These are mapped at different spatial scales (countries, used the total traffic volume (PCE) on the X-axis and total num-
ber of crashes on Y-axis, both separated by time intervals (every
hour). The results show a direct linear relation between the two;
therefore, with the growth of traffic, the number of crashes also
increases.
In the case of the 56 analyzed streets, the determination coef-
ficient (R2 ) is 0.73, greater than that calculated in intersections,
whose value reaches 0.58. This indicates that 73% of the crashes
occurred on these streets are explained by traffic, while in intersec-
tions the traffic only explains for 58% of the crashes. This difference
is due mainly to the fact that although the traffic volume keeps
Fig. 2. Simplified example of the kernel function (Young and Park, 2014). itself high in intersections, the number of crashes is lower, because
 J. Benedek et al. / Accident Analysis and Prevention 87 (2016) 117–126
Fig. 3. The total counts of traffic (PCE) and crashes in the Cluj-Napoca municipality by time intervals from (a) the monitored intersections and (b) the monitored streets.
most of the intersections are signalized, controlled, and the risk of
causing a crash is lower (low speed).
To see the intensity of linear dependence between the two vari-
ables (traffic and crashes), the Pearson correlation coefficient was
used with the value of 0.86 for streets and 0.76 for intersections.
Since the two values are positive and range between 0.75 and 1, one
can state that the number of crashes is directly proportional to the
traffic volume and the intensity of the linear dependence between
the two (crashes and traffic) is strong.
Considering the fact that the 56 streets and 22 intersections ana-
lyzed include the most important traffic arteries, together with the
result of this section, one can say that traffic explains the crash
hotspots in Cluj-Napoca municipality to a great extent.
4.2. Determining the traffic crash hotspots
To determine the traffic crash hotspots from Cluj-Napoca
municipality, the authors used the NKDE method implemented by
means of the “Kernel density estimation” tool in the SANET toolbox.
NKDE calculates the density of point type events along a network.
In this case, the analyzed events are the traffic crashes in Cluj-
Napoca, while the linear unit is represented by the road network
of the same city. The location of crashes is given by the GPS coordi-
nates recorded with the equipment mounted on the traffic police
car. 1466 crashes were considered along approximately 662 km of
streets over 4 years (2010–2013).
The result is the map of traffic crash hotspots along the streets
network of Cluj-Napoca (Fig. 5). This includes a total number of
38,095 segments, which are associated with minimum 0, maximum
9 crashes, depending on the proximity, with a standard deviation
of 1.1.
Fig. 4. The regression line of crashes and traffic in (a) monitored intersections and (b) monitored streets.
121
Dependent on the kernel density value, the street segments
were classified into 4 categories: not dangerous, low dangerous,
medium dangerous and high dangerous. When the kernel density
is 0 (0 crashes/segment), the respective segment is declared “not
dangerous”. The “low dangerous” segments have values between
0 and 2, the “medium dangerous” between 2 and 3 and the “high
dangerous” segments include values between 4 and 9.
According to this classification, the authors identified several
dangerous areas in the Cluj-Napoca municipality, but the most
important 4 hotspots were highlighted, as shown in Fig. 5.
Area 1 includes the city center and it is by far the most danger-
ous area in terms of traffic crashes density (NKDE) because most
street segments here are included in the “high dangerous” category.
More than 440 crashes occurred in this area and this hotspot com-
prises a significant part of the main transport axis, the West-East
axis (central section) of Cluj-Napoca, including high-traffic streets
(30,000–39,000 PCE/24 h). This area also includes many major high
traffic-volume intersections (55,000–73,000 PCE/24 h).
Area 2 is located in the North-West, at the entrance/exit to/from
the city. It includes streets with traffic volumes ranging between
28,200 and 36,500 PCE/24 h. The high density of crashes can be
explained also by the fact that heavy vehicles are also allowed in
this sector. Moreover, this hotspot includes also a set of very impor-
tant intersections consisting of the streets in the neighborhood of
the railway station, with a traffic volume between 38,800–45,000
PCE/24 h. During the 4 years of the survey, no less than 209 crashes
occurred here.
Area 3 is located at the entrance/exit to/from the East of the city
and includes, among others, the Eastern part of the East-West trans-
port axis. This area covers a small number of streets, but their traffic
volume is very high, between 50,000 and 60,000 PCE/24 h. Besides
the fact that these hotspots are located at the city entrance/exit,
122 J. Benedek et al. / Accident Analysis and Prevention 87 (2016) 117–126

Fig. 5. The traffic crash hotspots in Cluj-Napoca determined based on the “Network Kernel density estimation method – NKDE” applied by the SANET Tool.

another explanation for the high traffic volume would be the fact Table 1
Total of guilty persons by gender.
that the Avram Iancu International Airport is positioned in this area.
Area 4 is located in the West part of the city, and the hotspots Crashes Male Female
are found mostly in junctions. Even if the road traffic is lower in 1466 887 272
this area (12,800–30,000 PCE/24 h), the hotspots are mainly due 76.53% 23.47%
to the high density of intersection. Moreover, this area is a resi-
dential one and the pedestrian traffic is high, and the number of
auto-pedestrian crashes is slightly higher than in other areas. According to the introductory hypothesis, most frequently
By applying the same method, but separately for the auto-auto, (75.23%), the drivers are to blame for causing the crashes, followed
respectively for the auto-pedestrian crashes, the results revealed by pedestrians (22.26%), cyclists (1.64%) and in very few cases
very small differences between the locations of the hotspots for (0.87%), the passengers (see Table 2). Only the first 2 categories
the 2 types of crashes. Thus, auto-pedestrian crashes also occur were taken into consideration in this analysis.
in the places where auto-auto crashes happen regularly, indirectly In order to delimit as precisely as possible each group prone to
confirming the starting hypothesis, according to which the traffic road crashes and to avoid biased results, the city population was
intensity is the main factor in generating crashes. also inserted in the model. Therefore, the number of guilty persons
Compared to NKDE, PKDE computes hotspots using the in each gender and age group was reported to the total number of
Euclidean distance between the crashes. By applying PKDE to the the population of the population in the same groups (guilty persons
available data set, the authors could notice (Fig. 6) that the danger- per 10,000 inhabitants).
ous areas were approximately the same as in the case of NKDE, but During the 2010–2013 period, as one may notice in Table 3, the
it was more difficult to delimit precisely the segments of danger- dynamics of the population in Cluj-Napoca municipality was rel-
ous streets. Therefore, the authors consider that the PKDE method atively constant, and during the 4 years it increased only by 2000
is inadequate, as it lowers the accuracy of results. inhabitants (0.6%). Therefore, the average number of people per
four years was used when building this model. During this period,
the average population was 319,616 inhabitants, of which 151,028
4.3. Social groups vulnerable to crash occurrence men and 168,588 women. Their distribution by age is presented in
Fig. 7.
In order to better identify and describe the social groups vulner-
able to the occurrence of traffic crashes, the vulnerability pyramids Table 2
were built separately for drivers and for pedestrians. One should Total of guilty persons by type of participants.
mention that this study included both auto-auto crashes and auto-
Crashes Pedestrian Driver Passenger Cyclist
pedestrian crashes. Over the 4 analyzed years (2010–2013), there
occurred 1466 crashes involving 2510 people (drivers, passengers, 1466 258 872 10 19
22.26% 75.23% 0.87% 1.64%
pedestrians). Of these, 1159 were found guilty (see Table 1).
 J. Benedek et al. / Accident Analysis and Prevention 87 (2016) 117–126 123

Fig. 6. The traffic crash hotspots in Cluj-Napoca determined based on the “Network Kernel density estimation method – PKDE”.

Table 3 Looking at the general distribution of the total number of guilty

Number of inhabitants by gender.
persons (Fig. 8a), one can notice that the people most prone to the
Year Total Male Female occurrence of crashes, be they pedestrians or drivers and regardless
2010 318,442 150,854 167,588 of their gender, belong to the 21–25 year-old group, i.e. over 200
2011 319,582 151,088 168,494 guilty persons per 10,000 inhabitants (both genders). This is fol-
2012 320,042 151,112 168,930 lowed by the 16–20 year-old group (122 guilty persons per 10,000
2013 320,400 151,059 169,341 inhabitants) and 26–30 year-old group (124 guilty persons per
10,000 inhabitants).
As for the alcohol consumption, this plays an insignificant role
for women groups, while the younger men (mainly the 21–25
year-old group, 14 guilty persons per 10,000 inhabitants) cause
the highest number of alcohol-related crashes (Fig. 8b). The alco-
hol factor has influenced 7.93% of the persons who caused traffic
crashes in Cluj-Napoca between 2010 and 2013. The graph (Fig. 8b)
clearly shows that the persons who cause alcohol-related crashes
are most frequently males. The distribution between the two gen-
ders is extremely unbalanced, respectively 94.5:5.5. The age groups
most prone to causing alcohol-related crashes are the 21–25 year-
old male group (14 guilty persons per 10,000 inhabitants) and the
26–30 year-old group (12.3 guilty persons per 10,000 inhabitants,
1.38% of the total number of guilty persons, 17.4% of the total num-
ber of guilty persons under the influence of alcohol).
As regards the age and gender distribution of the guilty drivers
(Fig. 8c), the ratio between the male and female groups is unbal-
anced, i.e. 86.5:13.5. For the male category, the most exposed
age group is by far the 21–25 year-old group, which accounts for
no less than 17.08% from the total number of guilty drivers and
12.85% from the total number of guilty persons. When report-
ing this group to the number of inhabitants, it represents 172
Fig. 7. Population by gender and age (2010–2013). guilty people per 10,000 inhabitants. For the male category also,
124 J. Benedek et al. / Accident Analysis and Prevention 87 (2016) 117–126
Fig. 8. The frequencies pyramid by age groups and gender for: (a) total of guilty persons, (b) persons convicted under the influence of alcohol, (c) guilty drivers, (d) guilty
pedestrians.
Source: General Inspectorate of Romanian Police, Traffic Department.
it is worth mentioning the 26–30 year-old and 16–20 year-old
groups, which together sum up over 208 guilty persons per 10,000
inhabitants. This distribution shows the high vulnerability of the
young group (21–35 year-old) to the occurrence of crashes, a
phenomenon which can be explained by assuming higher risks
(alcohol consumption) and by little experience, specific to this cat-
egory.
For the female category, the pyramid has a similar distribution
to the male one only for the 16–20, 21–25 and 26–30 age groups.
Thus, the category most prone to causing crashes is the 21–25
year-old group, which includes 3.1% of the total number of guilty
persons (32 guilty persons per 10,000 inhabitants). A significant
difference between the two genders is the fact that for females,
the age groups between 36 and 65 account for similar percentages
ranging between 0.3 and 0.6% of the total number of guilty persons
(3–6 guilty persons per 10,000 inhabitants) while toward the top
of the pyramid (66–80+ groups) the percentage is mostly 0.
The situation is different for the guilty pedestrians (Fig. 8d).
In this case, the male-female distribution is much more bal-
anced, but with a slight advantage in favor of females, respectively
48.07:51.93. While the critical group of drivers was the 21–25
year-old group for both the male and the female category, there
are more pedestrian groups prone to road crashes, with large age
differences between the two genders. The general distribution of
guilty pedestrians (Fig. 8d) reveals that young people under 25 and
senior people over 66 are the groups most prone to car crashes.
The separate analysis of the two genders concludes that the criti-
cal male group is the 80+ group (over 27 guilty persons per 10,000
inhabitants), followed by the 76–80 year-old group (over 23 guilty
persons per 10,000 inhabitants).
It is very interesting that in the female category there are 3 crit-
ical groups with approximately the same value. This time there are
young groups: the 11–15 year-group group, the 16–20 year-group
group and the 21–25 year-group group. Each of these includes
about 19 guilty persons per 10,000 inhabitants. The groups between
66 and 80 years are also worth mentioning, but these do not have
an impact as large as in the case of the male category (on average,
13 guilty persons per 10,000).
5. Conclusions and future research directions
The key findings of this paper are the following. Firstly, by using
innovative analysis tools like SANET toolbox and the NKDE method
the authors were able to define very precisely the road traffic
crashes vulnerability areas, or hotspots in an urban environment
 J. Benedek et al. / Accident Analysis and Prevention 87 (2016) 117–126
characterized by high economic dynamics, intense spatial mobil-
ity and old road infrastructure. The delimited hotspots are located
in geographical areas with the highest values of traffic intensity:
the inner city and the three main entrances/exists of the city. This
major finding can be helpful for the future implementation of better
tailored traffic safety measures and urban transport policies.
Secondly, this research offered for the first time in the Roma-
nian spatial context a description of the social characteristics of
both drivers and pedestrians involved in road traffic crashes. The
authors highlight the most vulnerable drivers groups exposed to
the occurrence of road traffic crashes and the important role of
alcohol consumption in generating road traffic crashes. One find-
ing is that men aged between 21 and 25 are the most vulnerable
drivers exposed to the occurrence of road traffic crashes.
Driving under the influence of alcohol is among the most fre-
quent causes of crashes in this age group, in line with the findings in
the international literature (Maycock, 2002; Scagnolari et al., 2015).
However, in contrast with the findings of studies from developed
countries (Kahane, 2013; Cicchino, 2015), the authors have found
a low level of crash involvement for the older age groups. This
situation can be explained by potentially different demographic
characteristics of this age group: poor health conditions (average
life expectancy in Romania is the lowest in the EU) and there-
fore limited driver license renewal, lower income and hence lower
motorization rate and reduced spatial mobility.
In the case of pedestrians, the research reflects a different pic-
ture: much higher crash involvement of women in general as
compared to men, the younger women (11–25 year-old) being
highly vulnerable. However, high involvement rate on the male
side is also noticed for the older groups, and especially for the 80+
year-old group. It is clear that field methods are needed in order to
explain these variations in gender and age involvement in crashes.
Acknowledgements
This paper is a result of a doctoral research made possible by the
financial support of the Sectoral Operational Programme for Human
Resources Development 2007–2013, co-financed by the European
Social Fund, under the project POSDRU/159/1.5/S/132400 – “Young
successful researchers – professional development in an interna-
tional and interdisciplinary environment”.
References
Aguero-Valverde, J., Jovanis, P., 2006. Spatial analysis of fatal and injury crashes in
Pennsylvania. Accid. Anal. Prevent. 38, 618–625.
Al-Kaisy, A., Hall, F., Reisman, E., 2002. Developing passenger car equivalents for
heavy vehicles on freeways during queue discharge flow. Transp. Res. A: Policy
Pract. 36 (8), 725–742.
Anderson, T.K., 2009. Kernel density estimation and K-means clustering to profile
road accident hotspots. Accid. Anal. Prevent. 41 (3), 359–364.
Bailey, T.C., Gatrell, A.C., 1995. Interactive Spatial Data Analysis. Longman Scientific
& Technical, Harlow Essex, UK.
Benedek, J., Cristea, M., Lang, T., 2013. Fallstudienbericht Rumänien, in: André
Müller, JörgPlöger (szerk.):WiedererstarkteStädte: Strategien,
Rahmenbedingungen und Ansätze der Regenerierung in
europäischenGroß-und Mittelstädten. 16-30. BundesministeriumfürVerkehr,
Bau und Stadtentwicklung, BundesamtfürBauwesen und Raumordnung, Berlin
(Werkstatt: Praxis, Heft 82).
Bíl, M., Andrášik, R., Janoška, Z., 2013. Identification of hazardous road locations of
traffic accidents by means of kernel density estimation and cluster significance
evaluation. Accid. Anal. Prevent. 55, 265–273.
Bohle, H.-G, 2001. Neue Ansätze der geographischen Risikoforschung. Die Erde 2,
119–139.
Cicchino, J.B., 2015. Why have fatality rates among older drivers declined? The
relative contributions of changes in survivability and crash involvement. Accid.
Anal. Prevent. 83, 67–73.
Craddock, S., 2000. Disease, social identity, and risk: rethinking the geography of
AIDS. Trans. Inst. Brit. Geogr. 25, 153–168.
Dannefer, D., Bhatta, T., 2015. Age Stratification, International Encyclopedia of the
Social & Behavioral Sciences, Second Edition., pp. 333–337.
125
Erdogan, S., Yilmaz, I., Baybura, T., Gullu, M., 2008. Geographical information
systems aided traffic accident analysis system case study: city of
Afyonkarahisar. Accid. Anal. Prevent. 40 (1), 174–181.
Erdogan, S., 2009. Explorative spatial analysis of traffic accident statistics and road
mortality among the provinces of Turkey. J. Saf. Res. 40, 341–351.
European Commission, 2015. Road Safety in European Union, Trends, statistics and
Main Challenges, European Commission, Mobility and Transport DG, BE-1049
Brussels.
Flahaut, B., Mouchart, M., San Martin, E., Thomas, I., 2003. The local spatial
autocorrelation and the kernel method for identifying black zones: a
comparative approach. Accid. Anal. Prevent. 35 (6), 991–1004.
Flanagan, B., Gregory, E., Hallisey, E., Heitgerd, J., Lewis, B., 2011. A social
vulnerability index for disaster management. J. Homel. Secur. Emerg. Manage.
8 (1), Article 3.
General Urban Plan of Cluj Napoca “Plan Urbanistic General Cluj Napoca” 2013
Traffic study, “Studiu de trafic”, Annex 1 – Significant network traffic flows on
“Fluxuri de circulatie pe reteaua semnificativa”; Annex 2 – Critical analysis of
the main intersections monitored “Analiza critica a principalelor intersectii
monitorizate”.
Gibin, M., Longley, P., Atkinson, P., 2007. Kernel Density Estimation and
Percent Volume Contours in General Practice Catchment Area Analysis in
Urban Areas. Paper Presented at the GIScience Research UK Conference
(GISRUK).
Gundogdu, I.B., 2010. Applying linear analysis methods to GIS-supported
procedures for preventing traffic accidents: case study of Konya. Saf. Sci. 48,
763–769.
Hinojosa, B., Carreira, J., Rodriguez-Maroto, J., Garcia-Ruiz, R., 2008. Effects of
pyrite sludge pollution on soil enzyme activities: ecological dose–response
model. Sci. Total Environ. 396 (2–3), 89–99.
Hobbs, F., 2004. Age and sex composition. In: Siegel, J.S., Swanson, D.A. (Eds.), The
Methods and Materials of Demography. Elsevier Academic Press, San
Diego/London, pp. 125–173.
Horiuchi, S., 2015. Age Structure, International Encyclopedia of the Social &
Behavioral Sciences, 2nd ed., pp. 338–343.
Hyams, D., 2005. CurveExpert 1.3: A Comprehensive Curve Fitting System for
Windows. http://curveexpert.webhop.biz/.
Ivan, K., Haidu, I., 2012. The spatio-temporal distribution of road accidents in
Cluj-Napoca. Geogr. Tech. 2, 32–38.
Ivan, K., Haidu, I., Benedek, J., Ciobanu, S.M., 2015. Identification of traffic accident
risk-prone areas under low-light conditions. Nat. Hazards Earth Syst. Sci. 15,
2059–2068.
Kahane, C.J., 2013. Injury Vulnerability and Effectiveness of Occupant Protection
Technologies for Older Occupants and Women, Report No. DOT-HS-811-766.
National Highway Traffic Safety Administration, Washington, DC.
Lee, J., Wong, D.W.S., 2001. Statistical Analysis with ArcView GIS. John Wiley, New
York.
Levine, N., 2004. A Spatial Statistics Program for the Analysis of Crime Incident
Locations. The National Institute of Justice, Washington.
Li, L., Zhu, L., Sui, D., 2007. A GIS-based Bayesian approach for analyzing
spatial–temporal patterns of intra-city motor vehicle crashes. J. Transp. Geogr.
15, 274–285.
Maycock, G., 2002. Estimating the Effects of Age and Experience on Accident
Liability Using Stats19 Data. Behavioural Research in Road Safety XII.
Department for Transport, London.
Montella, A., 2010. A comparative analysis of hotspot identification methods.
Accid. Anal. Prevent. 42, 571–581.
National Institute of Statistics, Romania – INS 2012 Online database, www.insse.ro.
Okabe, A., Sugihara, K., 2012. Spatial Analysis along Networks: Statistical and
Computational Methods. John Wiley, Chichester, UK, pp. 177–193.
Okabe, A., Satoh, T., Sugihara, K., 2009. A kernel density estimation method for
networks, its computational method and a GIS-based tool. Int. J. Geogr. Inf. Sci.
23, 7–32.
O’Sullivan, D., Unwin, D.J., 2002. Geographic Information Analysis. John Wiley,
Hoboken, New Jersey, pp. 68–80.
Oulahen, G., Mortsch, L., Tang, K., Harford, D., 2015. Unequal vulnerability to flood
hazards: “ground truthing” a social vulnerability index of five municipalities in
Metro Vancouver, Canada. Ann. Assoc. Am. Geogr. 105 (3),
473–495.
Plug, C., Xia, J.C., Caulfield, C., 2011. Spatial and temporal visualisation techniques
for crash analysis. Accid. Anal. Prevent. 43, 1937–1946.
Poirier, I., Jean, N., Guary, J.C., Bertrand, M., 2008. Responses of the marine
bacterium Pseudomonas fluorescens to an excess of heavy metals: Physiological
and biochemical aspects. Sci. Total Environ. 406 (1–2), 76–87.
Prasannakumara, V., Vijith, H., Charutha, R., Geetha, N., 2011. Spatio-temporal
clustering of road accidents: GIS based analysis and assessment. Proc. Soc.
Behav. Sci. 21, 317–325.
Pulugurtha, S.S., Krishnakumar, V.K., Nambisan, S.S., 2007. New methods to
identify and rank high pedestrian crash zones: an illustration. Accid. Anal.
Prevent. 39 (4), 800–811.
Qin, X., Parker, S., Liu, Y., Graettinger, A.J., Forde, S., 2013. Intelligent geocoding
system to locate traffic crashes. Accid. Anal. Prevent. 50, 1034–1041.
Scagnolari, S., Walker, J., Maggi, R., 2015. Young driver’s night-time mobility and
attitude toward alcohol consumption: a hybrid choice model. Accid. Anal.
Prevent. 83, 74–89.
Schabenberger, O., Gotway, C.A., 2005. Statistical Methods for Spatial Data
Analysis. Chapman & Hall, Boca Raton, Florida, USA.
126 J. Benedek et al. / Accident Analysis and Prevention 87 (2016) 117–126
Siagian, T.H., Purhadi, P., Suhartono, S., Ritonga, H., 2014. Social vulnerability to
natural hazards in Indonesia: driving factors and policy implications. Nat.
Hazards 70, 1603–1617.
Soaita, M., 2013. Romanian suburban housing: home improvement through
owner-building. Urban Stud. 50, 2084–2101.
Steenberghen, T., Aerts, K., Thomas, I., 2010. Spatial clustering of events on a
network. J. Transp. Geogr. 18, 411–418.
Tate, E., 2013. Uncertainty analysis for a social vulnerability index. Ann. Assoc. Am.
Geogr. 103 (3), 526–543.
Truong, L.T., Somenahalli, S.V.C., 2011. Using GIS to Identify Pedestrian-Vehicle
crash hotspots and unsafe bus stops. J. Public Transp. 14 (1), 99–114.
Thakali, L., Kwon, T., Fu, L., 2015. Identification of crash hotspots using kernel
density estimation and Kriging methods: a comparison. J. Mod. Transp. 23 (3),
93–106.
Vandenbulcke, G., Thomas, I., IntPanis, L., 2014. Predicting cycling accident risk in
Brussels: a spatial case–control approach. Accid. Anal. Prevent. 62,
341–357.
Watts, M.J., Bohle, H.G., 1993. The space of vulnerability: the causal structure of
hunger and famine. In: Progress in Human Geography, vol. 17,
pp. 43–67.
Xie, Z., Yan, J., 2008. Kernel Density Estimation of traffic accidents in a network
space. Comput. Environ. Urban Syst. 32, 396–406.
Xie, Z., Yan, J., 2013. Detecting traffic accident clusters with network kernel density
estimation and local spatial statistics: an integrated approach. J. Transp. Geogr.
31, 64–71.
Yamada, I., Thill, J.-C., 2004. Comparison of planar and network K-functions in
traffic accident analysis. J. Transp. Geogr. 12, 149–158.
Yeung, J.S., Wong, Y.D., Secadiningrat, J.R., 2015. Lane-harmonised passenger car
equivalents for heterogeneous expressway traffic. Transp. Res. A: Policy Pract.
78, 361–370.
Young, J., Park, P., 2014. Hotzone identification with GIS-based post-network
screening analysis. J. Transp. Geogr. 34, 106–120.
Yu, H., Liu, P., Chen, J., Wang, H., 2014. Comparative analysis of the spatial analysis
methods for hotspot identification. Accid. Anal. Prevent. 66, 80–88.