Journal of Safety Research 40 (2009) 341–351

Contents lists available at ScienceDirect

Journal of Safety Research

j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / j s r

Explorative spatial analysis of traffic accident statistics and road mortality

among the provinces of Turkey
Saffet Erdogan ⁎
Afyon Kocatepe University, Faculty of Engineering, Department of Surveying, 03200 Afyonkarahisar, Turkey

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

Available online 19 September 2009 Introduction: The aim of the study is to describe the inter-province differences in traffic accidents and
mortality on roads of Turkey. Method: Two different risk indicators were used to evaluate the road safety
Keywords: performance of the provinces in Turkey. These indicators are the ratios between the number of persons killed
Accident rates in road traffic accidents (1) and the number of accidents (2) (nominators) and their exposure to traffic risk
GIS
(denominator). Population and the number of registered motor vehicles in the provinces were used as
Spatial analysis
Geographically weighted regression
denominators individually. Spatial analyses were performed to the mean annual rate of deaths and to the
number of fatal accidents that were calculated for the period of 2001–2006. Empirical Bayes smoothing was
used to remove background noise from the raw death and accident rates because of the sparsely populated
provinces and small number of accident and death rates of provinces. Global and local spatial autocorrelation
analyses were performed to show whether the provinces with high rates of deaths–accidents show
clustering or are located closer by chance. The spatial distribution of provinces with high rates of deaths and
accidents was nonrandom and detected as clustered with significance of P < 0.05 with spatial autocorrelation
analyses. Results: Regions with high concentration of fatal accidents and deaths were located in the provinces
that contain the roads connecting the Istanbul, Ankara, and Antalya provinces. Accident and death rates were also
modeled with some independent variables such as number of motor vehicles, length of roads, and so forth using
geographically weighted regression analysis with forward step-wise elimination. The level of statistical
significance was taken as P < 0.05. Large differences were found between the rates of deaths and accidents
according to denominators in the provinces. The geographically weighted regression analyses did significantly
better predictions for both accident rates and death rates than did ordinary least regressions, as indicated by
adjusted R2 values. Geographically weighted regression provided values of 0.89–0.99 adjusted R2 for death and
accident rates, compared with 0.88–0.95, respectively, by ordinary least regressions. Impact on industry:
Geographically weighted regression has the potential to reveal local patterns in the spatial distribution of rates,
which would be ignored by the ordinary least regression approach. The application of spatial analysis and
modeling of accident statistics and death rates at provincial level in Turkey will help to identification of provinces
with outstandingly high accident and death rates. This could help more efficient road safety management in
Turkey.
© 2009 National Safety Council and Elsevier Ltd. All rights reserved.

1. Introduction explosion in immigration and population, and a corresponding

Road traffic accidents are increasingly being recognized as a
growing public health problem in developing countries. It is estimated
that in the European Union 1.3 million traffic accidents occur every
year, resulting in over 40,000 deaths (CEC, 2003). Turkey, a rapidly
developing country, is a junction point between Asia and Europe in
terms of its social and economic structure (Işıldar, 2006). The
population is 70,586,256, and the number of drivers license holders
is 18,877,354 according to the 2007 population census. There is an
⁎ Kocatepe University, Engineering Faculty, Ahmet Necdet Sezer Campus, Gazligol
Road, 03200, Afyonkarahisar, Turkey. Tel.: +90 272 228 13 11; fax: +90 272 228 14 22.
E-mail addresses: saffet_erdogan@hotmail.com, serdogan@aku.edu.tr.
increase in vehicle numbers. Meanwhile gross domestic product and
income per capita have grown rapidly in recent years. The increase in
motor-vehicle ownership is very high in Turkey, 795,552 new vehicles
being registered in 2007 alone. Although the population increased by
14.70%, motor-vehicle ownership increased by more than 75%
between 1995 and 2007. During that period, the number of traffic
accidents (167.98%) and injuries (30.45%) also increased. The number
of deaths, however, decreased by 42.39%. The amount of material loss
owing to traffic accidents was approximately $1.1 billion for 2007,
excluding health expenses and loss of labor (TUIK, 2007).
Previous statistics have shown that casualty and fatality rates in
Turkey are much higher than in developed countries with comparable
vehicle ownership levels. Compared with other countries, Turkey's
traffic death rates are deplorable. The rate of fatal accidents in Turkey

0022-4375/$ – see front matter © 2009 National Safety Council and Elsevier Ltd. All rights reserved.
doi:10.1016/j.jsr.2009.07.006
342 S. Erdogan / Journal of Safety Research 40 (2009) 341–351
is 20 per 100 billion vehicle-kilometers, whereas the rates in the UK,
United States, and Germany are 0.9, 1.1, and 1.6 respectively. In the
last decade, the rate of accident per 100 million vehicle-kilometers
more than doubled (http://www.asirt.org/). Turkey, having fewer
vehicles than other countries, is recognized as being unlucky in terms
of fatal accident numbers.
Rapid expansion of road construction and increased numbers of
vehicles means that road traffic accidents are becoming an increasingly
serious public health problem in Turkey. Achieving reductions in the
number of traffic accidents and deaths is a national priority. Despite
the significant impact of traffic accidents and mortality on public health,
the magnitude of the effects could be greatly reduced if preventative
measures were taken in the safety-deficient provinces. The identifica-
tion of safety-deficient regions on the highway network is aimed at
comprehensive safety programs. Geographical information systems
(GIS) are a very important and comprehensive management tool for
traffic safety. GIS-aided spatial analysis provides information on
hazardous regions, hot spots, warm spots, and so forth (Liang, Mo'soem,
& Hua, 2005). The spatial component of traffic accidents has always
interested researchers and GIS-users. Empirical examples of spatial
analysis of traffic accidents are mainly concerned with black spot
analysis in local areas (Erdogan, Yilmaz, Baybura, & Gullu, 2008; Levine,
Kim, & Lawrence, 1995; Loo, 2006; Thomas, 1996), with a few
exceptions (Eksler, 2007; Hayakawa, Fischbeck, & Fischhoff, 2000; La
Torre, Beeck, Quaranta, Mannocci, & Ricciardi, 2007; Lassarre & Thomas,
2005) regional data are often neglected. Regional variations in traffic
accidents and mortality in Turkey have also received no attention so far.
This paper examines the regional disparities hidden behind
national statistics on road accidents and fatalities in Turkey using
GIS and spatial analyses. It is a common practice to compare cities or
countries in terms of road safety performance and to rank them in
terms of risk indicators such as the accident or death rates, which are
often expressed as the number of accidents/deaths per 100,000
inhabitants (Lassarre & Thomas, 2005). In Turkey, regional differences
in traffic accidents and traffic accident mortality and their underlying
determinants have not been studied, although Turkey has large
regional differences in socio-economic development. The study
therefore describes preliminary investigations of road accidents and
mortality in Turkey at the province level, adjusting for population and
number of registered motor vehicles owing to socio-economical
differences. The number of accidents and mortality ratio are modeled
with variables through geographically weighted regression. Different
types of software were used for visualization and spatial analyses of
the accident data in the study (i.e., Arc GIS 9.2, developed by ESRI,
GeoDa 0.9.5-I developed by Luc Anselin through the Center for
Spatially Integrated Social Science at the University of Illinois
(Anselin, 2004), CrimeStat 3.1 developed by Ned Levine, with support
from the National Institute of Justice (Levine, 2006) and GWR 3.0
developed by Fotheringham, Brunsdon and Charlton (2002)).
2. Data, methods and analyses
Accident reports are prepared by police departments within the
borders of municipalities and by the gendarmerie headquarters in the
provinces without police responsibility. Spatial analyses were per-
formed on the aggregated data reported by both police and
gendarmerie in the provinces of Turkey from 2001 to 2006.
Population by census year, annual intercensal rate of increase and
mid-year population forecasts, number of traffic accidents, and
records of road deaths (including deaths occurring during the
previous 30 days) in the provinces were taken from the data held
by the Turkish Statistical Institute (TUIK, 2006). Socio-economic index
values of the provinces, calculated by principal component analysis,
including 58 factors regarding demographic, education, employment,
health, industrial, economic, agricultural, and construction status
were taken from the State Planning Agency (DPT) for 2003.
Analyses in this study are based on the province levels of spatial
aggregation. Aggregated area-based data are very important sources
of information for many social science disciplines. Geographical
locations of these data are also an important factor in many aspects
of social and economic policies at the national level because
demographic, social, and economic characteristics of society repre-
sented by aggregated province units show differences throughout the
country. The province unit is the common level for social, economic,
demographic, and administrative data collection by agencies in
Turkey. Usage of province units in analyses therefore allows
comparison of accident and mortality rates with other demographic
and economic features. Accident and death rates were examined with
spatial analyses at the province level in this study. Province units,
however, have important limitations; provinces are administrative
units, and cover large areas with different heterogeneous populations,
and they may not match the ecological scale. It has been suggested
that aggregating the accident and mortality rates for the entire 5-year
period provides the advantage of stability in the province-level
accident and death rates, and also summarizes the phenomenon. A
temporal equilibrium state was therefore examined by means of the
time series of mortality and number of accident cases, and 2001–2006
emerged as a relatively stable period.
Population density and the number of registered motor vehicles
were used as standardization factors in this study. The mortality rate
(MR) standardized with population is the number of deaths based on
traffic accidents in a province during one year divided by the total
number of inhabitants residing in that province in the middle of that
year:
MR Pðstandardized by populationÞ = 100; 000⁎Number of deaths
= Total number of inhabitants:
The mortality rate standardized with numbers of registered
motor-vehicles is the number of deaths based on traffic accidents in
a province during one year divided by the total number of registered
motor-vehicles in that province in the middle of that year:
MR Mðstandardized by number of registered motor  vehiclesÞ
= 100; 000⁎Number of deaths = Number of registered motor  vehicles:
The accident rate (AR) standardized by population is the number
of accidents resulting in death or injury in a province during one year
divided by the total number of inhabitants residing in that province in
the middle of that year:
AR Pðstandardized by populationÞ = 100; 000
⁎Number of accidents resulting in death or injury
= Total number of inhabitants:
The accident rate standardized by number of registered motor-
vehicles is the number of accidents resulting in death or injury in a
province during one year divided by the total number of registered
motor-vehicles in that province in the middle of that year:
AR Mðstandardized by number of registered motor  vehiclesÞ
= 100; 000⁎Number of accidents resulting in death or injury
= Total number of registered motor  vehicles:
These are not true rates, as the populations in the provinces are not
equal to the populations that are exposed to the risk of road death. Some
inhabitants reside in the province but work or travel in another
province, and some motor-vehicles are registered in the province but
are used for work in or travel to another province. Hence, this index does
not consider mobility. It uses the simplest and most available variable as
the denominator: total resident population and total number of
 S. Erdogan / Journal of Safety Research 40 (2009) 341–351
registered motor-vehicles, which are currently the most appropriate
variables for comparative purposes (Lassarre & Thomas, 2005).
A commonly-used concept in rate analysis is that of a standardized
mortality rate, or the ratio of the observed mortality rate to a national
standard. The excess risk is the ratio of the observed rate to the
average rate computed for both the number of accidents and
mortality. This average is not the average of the provincial rates, but
calculated as the ratio of the total sum of all cases over the total sum of
all populations at risk. An excess risk map of numbers of road
accidents is shown in Fig. 1. An excess risk rate greater than 1.0
indicates that more fatal accidents occurred than would have been
expected; a ratio of less than 1.0 indicates fewer deaths than expected.
Each map is a choropleth map where the natural break method for
classification of the data has been applied to reflect the distribution as
accurately as possible. The natural break technique creates ranges
according to an algorithm that uses the mean of each range to
distribute the data more evenly across the ranges.
Since the accident data were aggregated into the areal units of
provinces, an important aspect is the derivation of a spatial weight
matrix (W) for spatial analyses. W is the fundamental tool used to model
Fig. 1. Excess risk maps of number of road accidents: ratios standardized with population (above) and ratios standardized with number of registered motor-vehicles (below).
343
the spatial proximity and interdependence between areal units.
Determination of the proper W matrix is a difficult and controversial
topic in spatial analyses. There are wide ranges of W determination
applications according to the aim of the studies covered in the literature
(Anselin, Cohen, Cook, Gorr, & Tita, 2000; Shi, Zhang, & Liu, 2006). In this
study, three different methods were used to obtain W matrices. The first
and second matrices were calculated by the criterion of contiguity
according to the centroids of the nearest 5–10 neighbors. The third
matrix was formed according to the criterion of general social distance.
When aggregated data are in use, if the population or the number
of cases is relatively small or sparse, rate estimates may not be precise.
In order to overcome the problem of rate instability, various
smoothing methods are usually employed (Anselin, Lozano, &
Koschinsky, 2006; Krivoruchko, Gotway, & Zhigimont, 2003). The
idea of smoothing is to borrow information from other small areas for
the estimation of the relative risk. In this study, Empirical Bayes (EB)
smoothing was used and raw rates were replaced by their globally-
smoothed values calculated by an EB tool in ArcGIS 9.2 (National
Cancer Institute of USA). Smoothed rates can be mapped to identify
more clearly some structural forms different from those induced by
344 S. Erdogan / Journal of Safety Research 40 (2009) 341–351

density (Lassarre & Thomas, 2005). Choropleth maps provide an simply the result of chance, can be calculated through a number of
effective way of visualizing how rates vary across a region. Smoothed formulae that have been derived by either the normal approximation
maps of MR_P, AR_P, MR_M and AR_M for the provinces from 2001 to or by randomization experiments. Mathematical formulae of methods
2006 are shown in Figs. 2 and 3. Smoothed rates provide accurate used are shown below, where X ̅ is the global mean value, calculated as
visual representation of the standardized rates compared with the a simple average value based on all data.
raw mortality and accident rates.
In order to identify and measure the strength of spatial patterns, N N
 
showing how the accident statistics were correlated in the country, N ∑ ∑ Wij ðXi − XÞðXj − XÞ
i=1 j=1
Moran's I and Geary c values were calculated with three W matrices. I= !  ð1Þ
Moran's I and Geary's c statistics measure the deviation from spatial
N N
2
∑ ∑ Wij ∑ ðXi − XÞ
randomness. Moran's I is produced by standardizing the spatial auto i≠j j = 1 i=1

covariance by the variance of the data using a measure of the

connectivity of the data. Geary's c uses the sum of squared differences N N
2
between pairs of data values as its measure of covariation. The range ðN−1Þ ∑ ∑ Wij ðXi −Xj Þ
i=1 j=1
of possible values of Moran's I is −1 to 1. Positive values indicate c= ð2Þ
spatial clustering of similar values and negative values indicate a
N N N
2
2 ∑ ∑ Wij ∑ ðXi − XÞ
clustering of dissimilar values. The range of possible values of c is 0 to i=1 j=1 i=1

2. A value of c close to 0 means the distribution of values clustered;

conversely, a value of c close to 2 means the distribution of values is Mathematical formulae of Z-score methods testing the significance
dispersed. For both Moran's I and Geary's c the statistical significance, of the results are shown below, where co and Io are observed values,
which indicates how confident one can be that any pattern is not cE and IE are expected values, and SD is the square root of variances.

Fig. 2. Smoothed AR_M (above) and MR_M (below) maps of provinces belong to the 2001–2006 period.
 S. Erdogan / Journal of Safety Research 40 (2009) 341–351 345

Fig. 3. Smoothed AR_P (above) and MR_P (below) values of provinces belong to the 2001–2006 period.

Moran's I and Geary's c values and Z-scores of these methods are spots). Mathematical formulae of G statistic, Z-score, and expected G
shown in Table 1. statistic can be written as below.
co −cE
Zc = ZI = Io −IE ð3Þ ∑ ∑ wij ðXi Xj Þ
SDcE SDIE i j GðdÞo −GðdÞE W
GðdÞ = ZGðdÞ = GE ðdÞ = ð4Þ
∑ ∑ðXi Xj Þ SDGðdÞ nðn−1Þ
Moran's I and Geary's c methods indicate clustering of high or low i j

values, but these methods cannot distinguish between situations. The

General G statistic is therefore used to give an understanding of the
clustering of high or low values. The General G statistic shows either
hot spots or cold spots in the region. A larger value of G statistic than
expected means that high values are found together (hot spots), and a
small value of G statistic means low values are found together (cold
Geary's c, Getis–Ord General G, and Moran's I are global measures,
and provide a single value of the spatial autocorrelation, checking the
clustering of the spatial pattern, but they do not show where the
clusters are. To investigate the spatial variation as well as the spatial
associations, it is possible to calculate local versions of Moran's I,

Table 1
Global spatial autocorrelation values of accident rates (NN: Nearest neighbor).

Rates Moran's I Expected Z-score Moran's I P value Moran's I P value Geary's C P value Getis–Ord G Z-score
(W with G. distance) index (5 NN W) (10 NN W )

AR_P 0.323 − 0.0125 20.78 0.695 0.001 0.642 0.001 0.6666 0.001 0.276 6.9
MR_P 0.162 − 0.0125 11.21 0.391 0.001 0.329 0.001 0.918 0.001 0.275 5.57
AR_M 0.075 − 0.0125 5.55 0.281 0.001 0.197 0.002 0.95 0.05 0.265 3.44
MR_M 0.09 − 0.0125 6.44 0.263 0.002 0.008 0.002 0.885 0.001 0.27 3.00
346 S. Erdogan / Journal of Safety Research 40 (2009) 341–351

Geary's c, and the General G statistic for each province unit in the
data. They may be interpreted as indicators of local pockets of non-
stationarity or hot spots, or they may be used to assess the influence of
individual provinces and to identify outliers (Anselin, 1995).
Therefore, local Moran's I (LISA, Anselin, 1995) and Gİ⁎ statistics of
Getis and Ord (1992) were used to explore where the accident ratios
are clustered. Local analyses based on the local Moran statistic are
visualized in the form of significance and cluster maps. Mathematical
formulae of local Moran's I and Z-score can be written as below,
where x ̅ is mean value and S2 is variance.
 fourth, a cluster of provinces with low–low rates. The legend shows
ðXi − XÞ I −EðIi Þ
Ii = ∑ Wij ðXj −XÞ ZðIi Þ = pi ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
S2 j VarðIi Þ
Second, Gİ⁎ statistics were used to detect local pockets of
dependence that may not show up with global spatial autocorrelation
methods, as suggested by Getis and Ord (1992). A high value of Gİ⁎
statistic indicates a spatial cluster of high values, where a low value of
Gİ⁎ statistic indicates a spatial cluster of low values around the area.
Mathematical formulae of Gİ⁎ statistic and Z-score of statistic can be
written as follows, where Wij is the weight for the target neighbor pair
(Mitchell, 2005).
∑ Wij ðdÞXj
j G⁎ −EðG⁎İ Þ
G⁎i ðdÞ = ZðG⁎İ Þ = qİ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi ð6Þ
∑ Xj VarðG⁎İ Þ
j
Four situations are identified through LISA. First, a cluster of
provinces with high–high rates, second, a cluster of provinces with
high–low rates, third, a cluster of provinces with low–high rates and,
the LISA results, and selected objects show the significant provinces
ð5Þ
with a level of significance of 95% determined as a cluster with Gİ⁎
statistic (Fig. 4).
Since many researchers have reported that determinants of
accidents were differentiated according to the urban characteristic
of the region, after the determination of clusters, the relation between
the accident rates and level of socio-economic development of
provinces was explored by means of bivariate spatial autocorrelation
(La Torre et al., 2007; Paulozzi, Ryan, Espitia-Hardeman, & Xi, 2007;

Fig. 4. Choropleth map of local autocorrelation values for AR (below) and MR (above) with LISA and Gİ⁎ statistic.
 S. Erdogan / Journal of Safety Research 40 (2009) 341–351
Traynor, 2008). This shows the correlation between one variable at a
location and a different variable at the neighboring locations. Greater
similarity indicated under spatial randomness implies spatially-
similar clusters in the two variables, while greater dissimilarity
implies a negative relationship between two variables (Anselin,
Syabri, & Smirnov, 2002). Maps of Multivariate LISA statistics for the
accidental rates are shown in Fig. 5.
There are many aspects explaining the observed differences in
accident statistics. Population density and the number of registered
motor-vehicles are two of the many variables. There are, however,
some regional disparities, such as length of highway-province roads
and number of registered motor vehicles leading to various factors. An
attempt was made to model the relationship between accident ratios
and other parameters that are believed to have some impact on
accident rates in Turkey. The reason for choosing the particular
parameters was that they are believed to relate some of the social and
economic characteristics of Turkey with available data in province
level. Given the data availability at the province level, the main
variables affecting the level of provincial accident and death rates
were selected through the technique of step-wise regression. The
number of accidents and the number of deaths as dependent variables
are therefore linked to a set of independent variables with one of the
Fig. 5. Choropleth map of multivariate LISA for AR_P (below) and AR_M (above).
347
main outputs of regression being the estimation of parameters that
link each independent variable to the dependent variables. These
independent determinant variables are length of highways, length of
province roads, and number of registered cars, buses, minibuses,
trucks, small trucks, and motorcycles in the provinces.
Regression model parameters derived by the ordinary least
squares (OLS) method, however, are assumed to apply globally over
the entire country from which measured data have been taken, based
on the assumption of spatial stationarity in the relationship between
the variables under study. A major problem with this technique when
applied to spatial data is that the processes under examination are
assumed to be constant over space. In most cases, however, this
assumption is invalid. If there is spatial non-stationarity, then the
global estimate of spatial relationships will misrepresent the real
phenomena. Accident and death rates were therefore modeled by the
geographically weighted regression (GWR) technique with men-
tioned variables. Alternatively, GWR, a recent refinement of the
ordinary linear regression model, is a local spatial statistical technique
used to analyze spatial non-stationarity, defined as the situation when
the measurement of relationships among variables differs from
location to location. Some relationships between the level of accident
statistics and different variables might also inherently vary over space,
348 S. Erdogan / Journal of Safety Research 40 (2009) 341–351

Table 2
Summary of the local parameter estimates for accidents.

Label Minimum Lower quartile Median Upper quartile Maximum Mean

Intercept − 163.682 8.567495 37.33018 49.58921 75.60523 26.59026279

Length of province roads − 0.03721 0.102428 0.147075 0.317986 0.752646 0.22685059
Number of cars − 0.00444 0.002647 0.005086 0.005224 0.008394 0.00346348
Length of highways − 0.46666 0.550781 1.658829 1.932737 3.271758 1.390518
Number of minibus − 0.0153 − 0.00396 0.016236 0.03184 0.124337 0.02791937
Number of bus − 0.03359 0.010165 0.03444 0.064457 0.139197 0.04137442
Number of minitruck − 0.04496 0.000157 0.007689 0.02452 0.044963 0.00969405
Number of truck − 0.01436 − 0.00016 0.0054 0.011037 0.029644 0.00544748
Number of motorbike − 0.00441 0.000342 0.002086 0.004361 0.007722 0.00231546
Population − 0.00065 −0.00026 0.000016 0.000055 0.000166 − 0.00008393

which could be tested/analyzed by use of the GWR technique. For
example, there are spatial differences in the effect of the levels of
provincial economic conditions (difference in registered number of
motor vehicles) on the provincial accident ratios.
GWR uses the coordinates of province centroids as target points
for a form of spatially-weighted least square regression. With GWR,
observations close to the center of the window are weighted more
heavily than observations further away.
With GWR, the relationship between the variables may be
expressed as:
ZðXi Þ = β0 ðxi Þ + β1 ðxi Þyi + β2 ðxi Þy2 + … + βn ðxi Þyn + ε
where xi indicates that the parameters are to be estimated at a
location, and ε is the random error term. The derived parameter
estimates may be mapped to show the nature of the variation in the
country. This can help to reveal spatial variations in the relationship
between the variables that would pass unnoticed in a global analysis
(Foody, 2003). Parameter estimation is highly dependent on the
weighting function and kernel used. When GWR is applied, key
decisions therefore concern the choice of weighting function with the
type of kernel function and the bandwidth of kernel function. A small
bandwidth results in very rapid distance decay, while a large value
results in a smoother weighting scheme. This parameter can be
optimized by the use of some form of adaptive methods such as cross-
validation minimization or Akaike's information criterion minimiza-
tion (Foody, 2003). A fuller description of geographically weighted
regression is provided by Fotheringham et al. (2002).
In this study, a Gaussian kernel function was used as an adaptive
function with variable bandwidth size with a bi-square function. The
variable bandwidth approach was chosen and optimized by cross-
validation minimization to account for the spatial variation in the size
of the provinces, and hence the density of province centroids. The
main output from GWR software is a set of local parameter estimates
for each relationship. Because of the volume of output of these local
parameter estimates and their local standard errors, as a convenient
indication of the extent of the variability in the local parameter
estimates, a six-number summary of the local parameter estimates is
shown in Tables 2 and 3. This gives a reasonable, although very
informal, means of comparison. The range of values of the local
estimates can be compared between the lower and upper quartiles
with the range of values at ±1 standard deviations of the respective
global estimate (Fotheringham et al., 2002).
To compare the relative performance of the GWR and OLS models
in replicating the observed data, an approximate likelihood ratio
test, based on the F-test, can be used (Fotheringham et al., 2002).
Table 4 shows a panel of results of an ANOVA in which the global
model is compared with the GWR model for accident rate modeling.
The ANOVA tests the null hypothesis that the GWR model represents
no improvement over a global model. According to the results, F-tests
ð7Þ suggest that the GWR model is a significant improvement on the
global model for the AR and MR in Turkey.
The significance of the spatial variability in the local parameter is
estimated by conduction of a Monte Carlo test. The results of the
Monte Carlo test on the local estimates indicate that there is robust
spatial non-stationarity in the local parameter estimates for the
variables length of highway, and number of minibuses, trucks, and
motorbikes over the whole country. The other four variables did not
have significant spatial variation. Hence, spatial variation of the effect
of variables on provincial accident rates was mainly represented by
the factors of the basic level of the other four variables among the
provinces.
One important characteristic of the GWR technique is that maps
generated by this technique play a key role in exploring spatial non-
stationarity and interpreting the results. The spatial distributions of
residuals for AR_P and MR_P are shown in Fig. 6 with an equal step
classification.
3. Results
In this study, exploratory spatial analyses and geographically
weighted regression analyses were conducted for determination and
understanding of the distribution of problematic provinces owing to
accidents and accident-caused mortality.
The risk of becoming an accident victim is given as the number of
such fatalities per population at risk. In its simplest form, risk for any
province is determined by the raw-accident death rate divided by
province population-number of registered motor vehicles. Small

Table 3
Summary of the local parameter estimates for deaths.

Label Minimum Lower Quartile Median Upper quartile Maximum Mean

Intercept −0.567782 0.125833 1.650974 3.810592 5.710883 1.942397

Length of province roads 0.016662 0.017664 0.019835 0.021547 0.023794 0.019965
Number of cars −0.000092 − 0.000076 − 0.000054 − 0.000027 0.000019 − 0.000047
Length of highways −0.016861 − 0.006401 0.034488 0.074572 0.088397 0.035654
Number of minibus 0.000268 0.001307 0.003575 0.005307 0.005958 0.003288
Number of bus 0.006308 0.009114 0.010474 0.010786 0.011673 0.009815
Number of minitruck −0.000838 − 0.000735 − 0.000693 − 0.000521 − 0.000415 − 0.000636
Number of truck 0.000458 0.000663 0.001119 0.001241 0.001370 0.000998
Number of motorbike 0.000114 0.000173 0.000299 0.000386 0.000458 0.000288
Population −0.000007 − 0.000005 − 0.000003 − 0.000001 0.000005 − 0.000002
 S. Erdogan / Journal of Safety Research 40 (2009) 341–351 349

Table 4 excess risk map for AR_P AND AR_M is shown in Fig. 1. The Sakarya,
ANOVA test results for accident rates. Bolu, Düzce, and Ankara provinces on the route of Ankara–İstanbul
Source Residual sum Degrees of Mean of F-test and the Muğla, Burdur, and Antalya provinces, which are tourist
of squares freedom squares value regions, were determined as problematic, with the rates standardized
OLS residuals 4,158,518 10 by population. The Çankırı, Kırıkkale, and Tunceli provinces were
GWR improvement 3,273,575 15.92 205,573.7 determined as problematic, with the rates standardized by the
GWR residuals 884,943.3 55.08 16,067.7 12.7942 number of registered motor vehicles.
While this excess risk map provides solid reference for relative risk
levels within the provinces, the calculations still ignore the large
populations may, however, have few deaths, and yet, owing to the variations in province populations. To address this issue, smoothing
small denominator in the rate calculation, may show very high raw- techniques were introduced. One smoothing method, EBS, utilizes a
accident rates. Conversely, large metropolitan areas may appear to weight inversely proportional to a province's population size,
have relatively low accident rates even though hundreds of deaths essentially giving greater weight to more informed rates. EBS-
take place in a geographically-small area over the same time period. generated rates are a weighted average of the observations' raw
Exclusive reliance on raw rates can therefore be misleading. To rate and the overall sample mean. Adjusting death and accident rates
measure excess risk rate, a straightforward approach for comparing by the population and number of registered motor vehicles enabled us
relative rates across a sample was used. As mentioned above, to identify the high variability of this cause of death and accidents, and
expected risk for the entire country was estimated by computing their relation with some risk factors. The differing population and
the rate for the entire provinces as total events divided by total economic structures of 81 provinces have a major influence on road
population at risk. Then each province's population was multiplied by accident and mortality rates.
the province's expected risk, resulting in the expected number of Global spatial autocorrelation methods were used to indicate the
accident-deaths if the average rates were applied to the observation. absence of clustering of high accident-mortality rates, as described
Finally, excess risk was determined as the actual number of deaths above. As shown in Table 1, global measures of spatial autocorrelation
divided by the expected number of deaths for each province. The indicated clustering with the 0.05 level of significance in the whole

Fig. 6. Spatial distributions of residuals for accidents (above) and deaths (below).
350 S. Erdogan / Journal of Safety Research 40 (2009) 341–351
country, with positive and bigger values than the expected values of
Moran's I values at all W matrices, and with Geary's c values less than
one at all W matrices. Getis–Ord General G values verified the clustering
of high values (hot spots). Positive clustering means that provinces with
high rates of accident and mortality tend to have neighbors that also
have high rates or vice versa. Although accident rates were significantly
autocorrelated, the autocorrelation is, however, quite unstable, with the
aggregation level.
Local approaches were also used for analyzing spatial association
to identify the location of the similar spatial patterns in the country.
İzmir, Aydın, Muğla, and Antalya Denizli, which are tourist provinces
of Turkey, and the corridor provinces, which include Bolu, Sakarya,
Kocaeli, Bilecik, Afyonkarahisar among the İstanbul–Antalya–Ankara
provinces, were determined as clusters with population-standardized
accident rates. The Kırıkkale, Erzincan, Tunceli, and Bingöl provinces,
however, were determined as clusters with motor-vehicle standard-
ized accident rates. According to the local spatial analyses results, the
presence of hot spots with high rates of mortality and accidents
showed that accidents and deaths caused by traffic are significant
public health problems in Turkey.
Multivariate Moran analyses indicated significant correlation
between accidental rates and the socio-economic status index values
declared by the State Planning Agency in 2003, with positive values as
shown in Fig. 5. As many researchers have observed, accidental rates
are higher in developed provinces than undeveloped provinces. The
same corridor in the cluster analysis was determined as clusters with
high–high clustering with population-standardized accident rates.
The Afyonkarahisar and Kırıkkale provinces were determined as
clusters with motor-vehicle standardized accident rates. Eastern
Turkey was determined as a cluster with low–low rates.
Conventional OLS and GWR methods were adapted to analyze
accident rates in relation to some available variables. When the
dependent value was the number of deaths, the best model had AICc
(Akaike's Information Criterion) = 668.170 and adjusted R2 = 0.891,
and when the dependent value was the number of accidents, the best
model had AICc = 1134.238 and adjusted R2 = 0.951 with OLS. The
GWR model had AICc = 679.088 and adjusted R2 = 0.894 with the
dependent value of the number of deaths and had AICc = 1065.252
and adjusted R2 = 0.986 with the dependent value of the number of
accidents. The improvement of model performance was evident in
OLS and GWR, both from the values of AICc and R2 and from the F-test.
The F-test result (Table 4, P < 0.001) suggests that the GWR model
significantly improved model fitting over the OLS model.
In principle, the intercept term or constant parameter measures
the fundamental level of accident rates, excluding the effects of all
factors on regional accident statistics across Turkey. It is henceforth
referred to as ‘the basic level of regional accident statistics’ (Huang &
Leung, 2002). There is a clear spatial variation with higher constant
parameters in the western region and lower ones in the eastern region
with regards to the number of accidents.
When we examine the spatial distribution of the length of
highways parameter in Turkey it can be observed that the west side
of the country had greater parameter estimates while the eastern side
had medium parameter estimates. The length of highways parameter
had a great effect on both the number of accidents and the number of
deaths. Length of province roads also had a positive effect on the
number of accidents and the number of deaths. The parameters of
number of buses, mini-trucks, and trucks had no significant influence
on both the number of accidents and the number of deaths. The
population variable had a negative effect on both accident rates and
mortality.
As a result, spatial analyses and statistics significantly contributed
to determination of the problematic provinces in Turkey. According to
the results, the provinces of Bolu, Düzce, Zonguldak, Sakarya, and
Kocaeli that contain the highway that connects Ankara, the capital
province of Turkey, to İstanbul, the biggest province, were problem-
atic provinces with AR_P and MR_P in the northwest of Turkey. These
provinces are developed provinces with a high density of population
and a high volume of motor vehicles. GWR-predicted values of
accident rates and residuals also confirm this situation. The provinces
of Bolu and Düzce, with number of accidents, and the provinces of
Sakarya and Düzce, with number of deaths, had high residuals. Many
of these provinces are undeveloped provinces with a low density of
population and a low number of motor vehicles. These provinces are
either tourist provinces located in tourist regions or junction
provinces that connect the other provinces with Antalya, which is a
tourist province. The cause of high rates of accidents and mortality in
these provinces is therefore the mobility of motor vehicles because of
tourism and link-roads. The high values of GWR residuals agree with
this. For example, Afyonkarahisar, which is an undeveloped province
but a junction point connecting the developed big cities and tourist
areas in Turkey, had large residual values with both death rates. The
accidents there, especially in summer, have a mobility origin (Erdogan
et al., 2008).
4. Conclusions
Identification of safety-deficient provinces with GIS-aided spatial
analysis will certainly help with the targeting of proper preventative
measures. Application of GIS and spatial statistics gave information
about the distribution of accidents and mortality at the province level
in this study. GIS and spatial statistics facilitate analysis and decisions
at the province level with regard to problematic areas and
precautionary measures. Traffic accident analysis is a very complex
topic owing to the existence of many variables affecting accidents
spatially. Many relationships are spatially non-stationary. When
accident statistics were examined, these values were problematic
for classical statistical tests like OLS regression, which violated the
assumption of independently-distributed errors with positive spatial
autocorrelation. Since spatial non-stationarity was common in the
data, GWR is a technique that allows locational information to be
usefully included in this type of analysis. It has been quantitatively
demonstrated that the relationships between the level of accident
rates and various factors exhibited considerable spatial variability. In
the analysis of relationships between the level of accident rates and
various factors in Turkey, it has been found that the GWR model is
significantly better than the OLS model. When we examined the GWR
residuals, high residual values of the provinces indicated those that
had high potential mobility because of tourism.
Although the decrease in fatalities by 42.39% in recent years is no
coincidence, fatality ratios are still high when compared with
developed countries. This could be because of low rates of seat-belt
use in Turkey (Şimşekoğlu & Lajunen, 2008). Nevertheless, Turkey has
adopted necessary and important measures in order to reduce serious
accidents. The measures taken relate to road traffic engineering,
education, enforcement and rescue, first aid, and emergency health
services (Işıldar, 2006). This study has demonstrated that it is
necessary to implement further preventative measures in the areas
determined in this research.
References
Anselin, L. (1995). Local indicators of spatial association—LISA. Geographical Analysis,
27, 93−115.
Anselin, L. (2004). GeoDa 0.95i release notes. Urbana-Champaign, IL: Spatial Analysis
Laboratory (SAL), Department of Agricultural and Consumer Economics, University
of Illinois.
Anselin, L., Cohen, J., Cook, D., Gorr, W., & Tita, G. (2000). Spatial analyses of crime.
Measurement and Analysis of Crime and Justice, 4.
Anselin, L., Lozano, L., & Koschinsky, J. (2006). Rate transformations and smoothing.
Urbana-Champaign, IL: Spatial Analysis Laboratory Department of Geography
University of Illinois.
Anselin, L., Syabri, I., & Smirnov, O. (2002). Visualizing multivariate spatial correlation
with dynamically linked windows. In L. Anselin, & S. Rey (Eds.), New tools for spatial
 S. Erdogan / Journal of Safety Research 40 (2009) 341–351
data analysis: Proceedings of a workshop Santa Barbara: Center for Spatially
Integrated Social Science.
Association For Safe International Road Travel. http://www.asirt.org/
Commission of the European Communities [CEC]. (2003). European road safety action
programme: Halving the number of road accident victims in the European Union by
2010: A shared responsibility Communication from the Commission Com 311 final:
Author.
Eksler, V. (2007). Road mortality in Europe — How sensitive is it to demographic
structure and population dynamics? IATSS Research, 31(1), 80−88.
Erdogan, S., Yilmaz, İ., Baybura, T., & Gullu, M. (2008). Geographical information
systems aided traffic accident analysis system case study: City of Afyonkarahisar.
Accident Analysis and Prevention, 40(1), 174−181.
Foody, G. M. (2003). Geographical weighting as a further refinement to regression
modelling: An example focused on the NDVI–rainfall relationship. Remote Sensing
of Environment, 88, 283−293.
Fotheringham, A. S., Brunsdon, C., & Charlton, M. (2002). Geographically weighted
regression: The analysis of spatially varying relationships. Chichester: Wiley.
Getis, A., & Ord, J. K. (1992). The analysis of spatial association by use of distance
statistics. Geographical Analysis, 24, 189−206.
Hayakawa, H., Fischbeck, P. S., & Fischhoff, B. (2000). Traffic accident statistics and risk
perceptions in Japan and the United States. Accident Analysis and Prevention, 32,
827−835.
Huang, Y., & Leung, Y. (2002). Analysing regional industrialization in Jiangsu province
using geographically weighted regression. Journal of Geographical Systems, 4,
233−249.
Işıldar, S. (2006). Road accidents in Turkey 1995–2004. IATSS Research, 30(2), 115−118.
Krivoruchko, K., Gotway, C., & Zhigimont, A. (2003). Statistical tools for regional data
analysis, using GIS. Proceeding of ACMGIS'03 New Orleans-Louisiana. http://www.
esri.com/software/arcgis/arcgisxtensions/geostatistical/research_papers.html
La Torre, G., Beeck, E. V., Quaranta, G., Mannocci, A., & Ricciardi, W. (2007).
Determinants of within-country variation in traffic accident mortality in Italy: A
geographical analysis. International Journal of Health Geographics, 6, 49.
351
Lassarre, S., & Thomas, I. (2005). Exploring road mortality ratios in Europe: National versus
regional realities. Journal of the Royal Statistical Society. Series A, 168, 127−144.
Levine, N. (2006). Crime mapping and the CrimeStat program. Geographical Analysis,
38, 41−56.
Levine, N., Kim, K. E., & Lawrence, H. N. (1995). Spatial analysis of Honolulu motor vehicle
crashes: 1. spatial patterns. Accident Analysis and Prevention, 27(5), 663−674.
Liang, L. Y., Mo'soem, D. M., & Hua, L. T. (2005). Traffic accident application using
geographic information system. Journal of the Eastern Asia Society for Transportation
Studies, 6, 3574−3589.
Loo, B. P. Y. (2006). Validating crash locations for quantitative spatial analysis: A GIS-
based approach. Accident Analysis and Prevention, 38(1), 879−886.
Mitchell, A. (2005). The ESRI guide to GIS analysis volume 2: Spatial measurements.
California: ESRI press.
Paulozzi, L. J., Ryan, G. W., Espitia-Hardeman, V. E., & Xi, Y. (2007). Economic development's
effect on road transport-related mortality among different types of road users: A cross-
sectional international study. Accident Analysis and Prevention, 39(3), 606−617.
Shi, H., Zhang, L., & Liu, J. (2006). A new spatial-attribute weighting function for
geographically weighted regression. Canadian Journal of Forest Research, 36, 996−1005.
Şimşekoğlu, Ö., & Lajunen, T. (2008). Why Turks do not use seat belts? An interview
study. Accident Analysis and Prevention, 40(2), 470−478.
Thomas, I. (1996). Spatial data aggregation: Exploratory analysis of road accidents.
Accident Analysis and Prevention, 28(2), 251−264.
Traynor, T. (2008). Regional economic conditions and crash fatality rates — A cross-
county analysis. Journal of Safety Research, 39, 33−39.
TUIK. (2006). www.tuik.gov.tr. Ankara, Turkey: Turkish Statistical Institute.
Saffet Erdogan is an Assistant Professor at Afyon Kocatepe University in Turkey. He
has a doctorate in Geomatic Engineering from Istanbul Technical University. His
primary research expertise is spatial analysis and spatial statistical techniques and
applications in different disciplines. He has been leading the Public Surveying
discipline in the Surveying Engineering Division in Afyon Kocatepe University.