See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/304628362

Rural casualty crashes on the Kings Highway: A

new approach for road safety studies

Article in Accident Analysis & Prevention October 2016

DOI: 10.1016/j.aap.2016.06.005

CITATIONS READS

0 116

3 authors:

Sahar Alian Robert G V Baker

University of New England (Australia) University of New England (Australia)
2 PUBLICATIONS 1 CITATION 57 PUBLICATIONS 659 CITATIONS

SEE PROFILE SEE PROFILE

Stephen Wood
University of New England (Australia)
13 PUBLICATIONS 87 CITATIONS

SEE PROFILE

All content following this page was uploaded by Robert G V Baker on 06 July 2016.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
 Accident Analysis and Prevention 95 (2016) 819

Contents lists available at ScienceDirect

Accident Analysis and Prevention

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

Rural casualty crashes on the Kings Highway: A new approach for

road safety studies
Sahar Alian , R.G.V. Baker, Stephen Wood
Geography and Planning, BCSS, University of New England, Armidale 2351, Australia

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

Article history: This paper will consider the contribution that changes in road geometry and driver visual information
Received 9 January 2016 make to the incidence and distribution of road casualties in different driving environments. This relation-
Received in revised form 7 June 2016 ship will be explored specically for the Kings Highway, a major arterial road connecting Queanbeyan
Accepted 7 June 2016
with coastal southern New South Wales, Australia. It introduces and suggests a new empirical approach
of plotting crashes with road segmentation, calculating sinuosity indices and grades as key features of
Keywords:
road geometry, and critical visual points as a behavioural component of road curvature, within a GIS
Casualty crashes
context. It is an approach that might be used when detailed road geometry data is not available. The
Sinuosity index
Critical visual points
visualisation and segmentation approach in this research might be used for summarising crash rates and
Grade road geometry factors, and for comparing day/night and eastbound/westbound driving conditions. The
Road environment results suggest some early interpretations for detailed road safety studies that might be considered at
Road safety local or national levels. The rate of crashes increases according to changes in road geometry factors during
the day and for eastbound travel. This is not the case for night driving where the incidence of crashes is
similar on both straight and curved roads segments due to the headlight effect and limited background
visual eld. Crash clusters at day-time may be due to the stronger effect of road geometry (e.g. combi-
nation of curvature and vertical grade) on driver behaviour travelling eastbound. The outcomes suggest
that it might be essential to consider the effect of environmental factors in any road safety and crash
analysis studies.
2016 Elsevier Ltd. All rights reserved.

1. Introduction
Rural casualty crashes are a major safety problem in New South
Wales (NSW), Australia (Austroads, 2010). Within the substantial
literature that has sought to explore and theorise the factors con-
tributing to road safety there is widespread agreement that human
behaviour, road characteristics and the environment are the main
factors that affect road safety (Shankar et al., 1995; Elvik et al., 2004;
Wang et al., 2013; Yu and Abdel-Aty, 2014). Driver errors, in isola-
tion or in combination with other factors, are involved in about 80%
of crashes (Forkenbrock and Foster, 1997). Crash risks are higher
on rural roads when actual behaviour deviates from appropriate
behaviour (Weller, 2010). In addition to the signicant contribution
of driver behaviour, road geometry (e.g. curvature and the qual-
ity of infrastructure), trafc conditions (e.g. trafc ow and speed)
and lighting conditions (e.g. day/night driving) can also affect road
safety (Shankar et al., 1995; Sagberg, 1999; Golob, 2003; Aarts and
Corresponding author.
E-mail addresses: salian3@myune.edu.au, alian14995@alumni.itc.nl (S. Alian),
rbaker1@une.edu.au (R.G.V. Baker), swood26@une.edu.au (S. Wood).
Van Schagen, 2006; Jones et al., 2012). The mixed effect on road
safety of these factors, especially trafc ow, speed and geometry
(Wang et al., 2013) provides a key rationale for further research in
this area.
Road design, especially the amount of curvature along a road,
might affect driver behaviour and road safety in different ways.
Extra effort may be required to remain within lanes, visibility dis-
tances along road axes may be reduced, and anticipation of the
course of the road and upcoming trafc may be limited, leading to
driver errors (Castro, 2008). There is a need to develop methods
that might allow for better understanding of interactions between
the geometry of curves and behavioural responses of drivers. In this
context, sources such as tangent points offer an enticing way for-
ward in so far as they exist at the interface between drivers and
the driving environment (Chattington et al., 2007; Kandil et al.,
2009). They are at once a geometrical feature of a curve and of
visual importance to drivers when steering into a curve (represent-
ing changes in the direction of a drivers optical eld). They hold the
promise that they might be used to help predict driver behaviour
along the curve (Land and Lee, 1994; Authi and Mestre, 2012).
Situations that involve lacking or reduced visual information also
warrant further exploration. For example, when driving at night

http://dx.doi.org/10.1016/j.aap.2016.06.005
0001-4575/ 2016 Elsevier Ltd. All rights reserved.
 S. Alian et al. / Accident Analysis and Prevention 95 (2016) 819
under headlights, drivers might behave in signicantly different
ways when compared to driving under normal conditions because
drivers might assume objects to be more distant than they actually
are, or may have reduced reaction times (Groeger, 2000).
Against this backdrop, the current paper will consider associ-
ations between key road geometry variables (both physical and
behavioural) and crash rates, under different environmental con-
ditions (day/night, and eastbound/westbound driving), along Kings
Highway, NSW. The main objectives of this research are three-
fold: rst, to measure key geometrical and behavioural features
of road curvature and analyse associations between them; second,
to graphically represent the results and illustrate how crash counts
and rates are differently distributed between equal sections of the
road; and, third, to analyse crash rates and consider changes in
crash rates relative to changes in road geometry variables under
different environmental conditions.
2. Method and background
2.1. Research framework
This research aims to evaluate interactions between changes
in road geometry and behavioural responses to those changes
(changes in driver information), and their relationship to the geo-
graphical distribution and rate of casualty crashes along Kings
Highway, under different driving conditions (day/night and east-
bound/westbound). There is a large literature on crash data analysis
and methods for identifying hazardous road sections; the particu-
lar approach that is employed in particular cases will be dependent
upon data availability and project aims and objectives (Mannering
and Bhat, 2014). In recent years various statistical approaches
have been used and improved and have added considerable value
to the elds ability to model the factors that affect the number
of crashes in a particular geographical space over a certain time
period (Lord and Mannering, 2010). There is considerable agree-
ment that the distributional properties of least square techniques
and mainly linear regressions do not tend to describe random, dis-
crete, nonnegative crash data in an adequate manner. Instead, it
is often recommended that Poisson or Negative Binomial distribu-
tions or other approaches be employed that might more adequately
account for the integer nature of the data (Miaou and Lum, 1993;
Shankar et al., 1995; Savolainen et al., 2011). However, there are
also limitations to Poisson and Negative Binomial distributions in
so far as the theoretical underpinnings assume random distribu-
tion, whereas environmental factors, such as curvature, are not
randomly distributed across the total length of a road. There is a
body of work that indicate nonparametric or nonlinear approaches
might be an alternative to existing methods based on the sample
size, nature of the data and research objective (Lord and Mannering,
2010; Washington et al., 2010). It might also be a possible solution
to consider the effect of trafc, segment length, and aggregation
on crash data, and meet the requirements for linear regression
approaches (Rakha et al., 2010).
In an attempt to address such issues, the current paper adopts
a bottom up approach, proceeding from the data, as such, with no
underlying assumptions of normality or randomness; the goal is to
investigate variations in crash data according to actual geograph-
ical and time constraints. This includes an interest in functional
segments of change, on particular parts of the road, or between
day/night. Rather than linear regression, with its assumptions of
normality, polynomial regression is used, without assumption of
normality, along with nonparametric testing in the initial phases
of exploratory analysis. Rather than random partitioning, the data
is interrogated in the mode of exploratory analysis according to
direction of travel, gender and day/night driving conditions and
the quadratic model is applied within the context of change. The
9
approach to road crashes is geographical and location-specic,
looking at environmental change (such as curvature and perception
information) within the framework of the relationship between
road geometry and human behavioural information processing.
Finally, the use of discrete choice models encounters similar issues
to those described above in so far as one must assume a normal
distribution of human behaviour characteristics over the environ-
ment. The bottom up approach adopted here is not limited by
this condition and is more amenable to public policy development
including initiatives such as straightening roads and speed limit
change.
2.2. Study area
The study area is a main rural road, around 132 km in length,
connecting Queanbeyan, on the border of the Australian Capital
Territory (ACT), to Batemans Bay, a coastal town in NSW. It com-
prises a mixed geometry i.e. both straight and curved sections. The
road provides residents of the ACT with their primary means of
accessing a range of holiday, recreational and leisure destinations
along the scenic south coast of NSW. Bungendore and Braidwood
are the main towns along the road. In recent years, there has been a
slight increase in casualty crash rates along Kings Highway, result-
ing in a road safety review and ongoing upgrades (RMS, 2013). Fig. 1
illustrates the Kings Highway route and casualty crash locations
from 2007 to 2011.
2.3. Data collection
To identify associations between road geometry variables and
changes in crash rates, three data sources were used: crash data,
trafc data and road data. The crash data related to the period
20072011, and was sourced from the NSW Department of Roads
and Maritime Services. The data considered variables relating
to: driver gender, age and speed; travel direction; and time of
day/night, amongst others. Sample sizes were sufcient to enable
statistical analysis, and enabled variations from one year to the
next to be taken into account (Mamcic and Sivilevicius, 2013;
Hosseinpour et al., 2014). While the quality of this data was
generally very good, it did have the limitation of not identify-
ing crash types. Trafc data was 5 year daily trafc count data,
also for the period 20072011, and also sourced from the NSW
Department of Roads and Maritime Services. This data was avail-
able for 736 days, and encompassed both travelling directions
(eastbound/westbound), all days of the week, and all seasons. In
terms of road geometry data, road centreline data was supplied by
the NSW Lands Department. Comprehensive geometric data con-
cerning Kings Highway was not made available to the authors.
While this might be regarded as a limitation of the study which
constrained the range of methods that could be considered, the par-
ticular method that was employed (see Sections 2.42.5) ensures
that road centreline data sufces to enable identication and anal-
ysis of key road geometry variables.
2.4. Data processing
2.4.1. Segmentation
With respect to analysis of road characteristics, segmentation
is almost invariably used to extract certain road geometry vari-
ables, and to identify black spots. Within the literature, the length
of road segments ranges from 500 m to 7 km; the sections may be
of xed length, or sections with homogenous geometric charac-
teristics, depending upon research aims and objectives (Miaou and
Lum, 1993; Ackaah and Salifu, 2011; Mamcic and Sivilevicius, 2013;
Hosseinpour et al., 2014). For the purposes of this study, the road
was divided into n circles where the diameter of each circle was
10 S. Alian et al. / Accident Analysis and Prevention 95 (2016) 819
Fig. 1. Crash distributions along Kings Highway, 20072011.
1 km. The latter corresponded to the straight-line distance for each
road segment, entailing that the actual length of each road segment
kn 1 km. Using this approach, the total road length (about 132 km
from the last roundabout in Queanbeyan to the rst roundabout in
Batemans Bay) was divided into 124 segments, each with differ-
ent lengths but similar straight-line distances. Fig. 2 illustrates the
adopted approach within the Clyde Mountain area.
2.4.2. Road geometry characteristics
Once a road has been segmented, the next task for any analysis of
road data is to determine key geometric characteristics, especially
those relating to curvature. Within the literature, a wide range of
curvature measurements have been used in the analysis of road
crashes, including bend density (the number of bends per kilo-
metre), detour ratio (the ratio of actual road distance to straight
distance), straightness index (the proportion of road segments that
are straight), cumulative angle (the cumulative angle turned per
kilometre), mean angle (the mean angle turned per bend) (Haynes
et al., 2007). It is widely agreed that use of a single measure will
not sufce to capture all of the signicant geometric characteristics
for a given road (Jones et al., 2012; Wang et al., 2013). Accord-
ingly, three different aspects of road geometry are considered in
this study: sinuosity index (actual road distance to straight-line
distance) (Williams, 1986; Jones et al., 2012); grade (ratio of ver-
tical change to horizontal change) (Austroads, 2015); and critical
visual points (focal points, or points in a bent segment where the
visual information of the driver changes due to changes in direc-
tion) (Land and Lee, 1994). Elements of these features are presented
diagrammatically in Fig. 3, and discussed further, below, in turn.
The sinuosity index is measured using Equation (1) whilst the
grade (gradient) of the straight line is measured using Equation (2).
For the sinuosity index (SI),
   
/ AB
SI = AB (1)
where AB is actual road (path) length and AB is shortest road (path)
length. For the gradient (G),
G = H/X (2)
where H is change of height and X is change of distance.
The sinuosity index provides a physical measurement of road
curvature. Where SI = 1, the road segment is straight; where SI > 1,
the road segment is curved; as the index increases, so too does
the curvature of the road segment. The sinuosity index is a rela-
tively straightforward indicator of curvature that might be usefully
employed when detailed geometric data is limited. It is essential
to supplement the sinuosity index with consideration of grades
because crash rates have been shown to be different between uphill
and downhill sections of road (Jurewicza et al., 2014).
Critical visual points are measured by tangent lines and are
visually important for drivers when steering into a turn as a good
predictor of the curvature of the road ahead (Land and Lee, 1994).
They reect tangent points and involve a threshold for angular
change in the direction of the eld of view at a particular point.
Critical visual points are a combination of inection points and sta-
 S. Alian et al. / Accident Analysis and Prevention 95 (2016) 819 11

Fig. 2. Road segmentation in Clyde Mountain area.

Fig. 3. Sinuosity index, critical visual points and grade of the straight line.

tionary points: in mathematics, a stationary point is a point in a
curve that has a horizontal tangent and the derivative of the func-
tion is zero; a point of inection is a point on a curve at which the
curve changes from being concave (concave downward) to convex
(concave upward), or vice versa.
The critical visual point measure provides two key functions
for the methodology. First, it supplements the sinuosity index to
enable differentiation between different types of curves. Consider,
for example, two road segments where the respective sinuosity
indices are identical, but one has three critical visual points, the
other seven. The road segment with seven critical visual points will
have sharper and more frequent curves, than the segment with
three critical visual points. And second, the critical visual points
measure embeds a key aspect of driver behaviour in the crash anal-
ysis since critical visual points relate to the visual eld of drivers.
It provides a proxy to the change of information processing in the
moving optic eld, and may be used as a dynamic spatial reference
to analyse driver behaviour on curves.
12 S. Alian et al. / Accident Analysis and Prevention 95 (2016) 819
2.5. Data analysis
2.5.1. Sinuosity index and critical visual points interaction
As discussed above, a key goal of this study was to examine
interactions between more or less purely geometrical features of
road curvature (sinuosity index), and more behavioural aspects of
road curvature (critical visual points reecting changes in a drivers
visual information on curves). Accordingly, the interaction between
changes in the sinuosity index and critical visual points was a crit-
ical rst step in the analysis. Linear and multiple regressions were
the main statistical tools used to examine the kind and strength
of relationship between interval and ratio dependent and inde-
pendent variables and have the capability of predicting dependent
variables (A et al., 2011). Relationships between sinuosity and
critical visual points along the 124 road segments were plotted.
Quadratic regressions were used to evaluate the rate of change,
bearing in mind that literature suggests linear regressions might
not be suitable for analysing and predicting the distributional prop-
erty of crash data, and nonlinear regressions often better t the data
(Lord and Mannering, 2010).
2.5.2. Crash data analysis
In the literature, the most common methods used for crash
analysis of hazardous road segments are crash frequency (counts
of crashes at specic road segment during a specic time period)
and crash rate (rate of crashes considering the effect of trafc at
a specic road segment during a specic time) methods (Yu et al.,
2014). To analyse the data, crash frequencies and crash rates were
determined, and associations between crash rates and driver gen-
der, age and speed were explored, including differences (if any)
between day/night travel, and eastbound/westbound driving direc-
tions (night was considered as extending from 6pm to 6am i.e. when
headlights are generally used). The nonparametric chi-square test
was used to consider whether two classication variables were sta-
tistically signicant. If required, the Mann-Whitney U Test might be
used to evaluate the equality of two population means as an alter-
native to the two independent samples to test when the assumption
of normal population distribution is not satised (Washington et al.,
2010).
2.5.3. Data visualisation
To investigate how crashes were distributed along the road,
along with variations in crash rates according to different driv-
ing conditions (e.g. day/night, eastbound/westbound), the crash
data was visualised in a diagram. The centreline was divided into
124 segments (see Section 2.4), and the sinuosity index and crit-
ical points for each segment were used to represent changes in
geometry along the road. The distribution of crashes along the
road was plotted in such a way as to allow comparisons between
day/night and eastbound/westbound travel. Urban areas were
excluded because they have a built-up environment, along with
different road geometry features, speed limits and environmental
conditions; the distinction between rural and urban roads is recog-
nised by NSW speed zoning guidelines (RTA, 2011). The diagram
was used for crash cluster identication, and for visual inter-
pretation of differences, where present, between day/night and
eastbound/westbound travel. It was also possible to compare the
rate of crashes instead of crash counts. The rate of crashes was com-
pared between straight and curved sections of the road using the
Mann-Whitney U Test. In this study the total road segments were
divided into two groups, where SI 1.05 was considered straight
and SI > 1.05 was considered curved (Leopold and Wolman, 2013).
2.5.4. Regression analysis
The nonlinear nature of the data was explored with a quadratic
regression to analyse how crash rates varied according to sinuosity
Fig. 4. Sinuosity index and critical visual points regression.
index, grade, and critical visual points (change of driver infor-
mation) between day and night. The total road segments were
aggregated to form 24 equal 5 km sections. This was intended to
reect road safety signage practices (whereby advisory signs refer
to 5 km stretches of road), and to increase the homogeneity of seg-
ments.
The 5 km sections of road with no crashes were then cate-
gorised as safe road sections and excluded from further analysis.
For the remaining sections, crash rates were standardised per
100,000 moving vehicles per section, and were compared by
change of mean road characteristic variables (both geometrical
and behavioural) in aggregate (between day/night) and disaggre-
gate (between day/eastbound, night/eastbound, day/westbound
and night/westbound) levels. R-squared and p-values were com-
pared for both quadratic and linear regressions. Regression results
are relevant for p-values less than 0.1 (greater than 90%) and signif-
icant for p-values less than 0.05 (greater than 95%). For signicant
p-values, tted regression lines/curves were drawn and formulated
for further analysis. Finally, similarities and differences between
day/night and eastbound/westbound were identied and analysed.
If signicant results were recognised a geometric index might be
developed for the study area.
3. Results
3.1. Sinuosity index and critical visual points interaction
Fig. 4 illustrates the relationship between the sinuosity index
and critical visual points. The X-axis illustrates the sinuosity index
per segment and the Y-axis summarises critical visual points per
segment along 124 segments. When linear and quadratic regres-
sions are compared, the results suggest that a quadratic regression
ts the data better than a linear regression (R2 = 0.650 and 0.618,
respectively). There is improvement in the R-squared value, and
the results are signicant (p < 0.001). The graph also suggests that
there is a ceiling to the impact of sinuosity on the number of crit-
ical visual points. It conrms Land and Lees (1994) approach and
its relevance to the study of segmental change in road geometry.
3.2. Crash data analysis
Table 1 summarises crash data results, with signicant out-
comes highlighted in bold. In general, both crash frequency and
rates are different between day/night and eastbound/westbound
travelling directions. The frequency of crashes during the day is
 S. Alian et al. / Accident Analysis and Prevention 95 (2016) 819 13

Table 1
Data exploration: Kings Highway, 20072011.

Crash Freq. Average Daily Trafc Volume Crash Rate Driver Age (yrs.) Driver Gender T. Speed (Km/h)

N % 736 days /100000 MV Mean SD M F M/F Mean SD

Total (24 h.) 180 100.00 4025 2.45 40 18 106 73 1.45 78.51 23.65
Day (12 h.) 144 80.00 3418 2.31 42 19 80 64 1.25 76.84 22.32
Night (12 h.) 36 20.00 607 3.25 33 16 26 9 2.88 85.83 28.03

East Total (24 h.) 115 63.90 2015 3.13 40 18 75 39 1.92 75.91 24.90
Day (12 h.) 93 51.67 1688 3.02 42 18 60 33 1.82 74.44 23.65
Night (12 h.) 22 12.23 327 3.68 31 12 15 6 2.5 83.93 30.60

West Total (24 h.) 65 36.10 2010 1.77 40 20 31 34 0.91 82.80 20.93
Day (12 h.) 51 28.33 1730 1.61 41 19 20 31 0.64 81.22 19.16
Night (12 h.) 14 7.77 280 2.74 36 20 11 3 3.66 87.86 26.00

Signicant results are highlighted in bold (p < 0.05).

about three times more than night-time; but the rate of hav- Table 2
Sections with no crash along the Kings Highway.
ing a crash at night is more than day-time. Crash numbers vary
by age and gender between day/night, and by gender between Section East/Day West/Day East/Night West/Night
eastbound/westbound travelling directions. There is a signicant Crashes Crashes Crashes Crashes
difference in the male/female ratio and younger/older drivers when 1 0
comparing travel during the day and travel at night. Young, male 2 0
drivers are far more likely to have a crash during the night than 3 0 0 0 0
4 0
females. The ratio of the number of crashes by male to female
5 0
drivers is different travelling day/night and eastbound/westbound. 6 0 0 0 0
The ratios are 1.92 travelling eastbound and 0.91 travelling west- 7
bound, suggesting that females are more likely to be involved in 8 0 0
9 0 0
westbound crashes, and males are more likely to be involved in
10 0
eastbound crashes. More specically, the ratio at night is about 11 0
ve times more than day-time travelling westbound, indicating 12
that male drivers are more likely to be involved in night west- 13 0 0
bound crashes and female drivers are more likely to be involved 14 0
15 0 0
in day westbound crashes; this is not the case for eastbound travel.
16 0
Nonparametric testing (see Section 2.5.2) shows that the results 17 0
are signicant (p < 0.05). Finally, the mean travel speed is different 18
between day/night and eastbound/westbound travel with higher 19
20 0
travel speeds at night and lower speeds for eastbound travel.
21
22 0 0
3.3. Data visualisation 23 0 0
24 0
Fig. 5 depicts the distribution of crash counts along Kings High-
way, divided into 124 segments. The X-axis summarises road
centreline data measured using the methodology described above tions, each section incorporating 5 segments. The sections of the
(critical visual points changes for AB = 1, and 1 SI 1.5), and the road with no crashes for eastbound/westbound and day/night are
Y-axis illustrates sinuosity changes (1 SI 1.5). The distribution illustrated by 0. The results conrm that in 2 out of 24 sections
is compared from Queanbeyan to Batemans Bay (above the X-axis) (sections 3 and 6), one before, and the other one after Bungendore,
and Batemans Bay to Queanbeyan (below the X-axis), and between no crashes occurred for both eastbound/westbound and day/night
day and night. driving. The sections of the road with no crashes are most likely to
Excluding the obvious differences in crash counts between day be at night and involving westbound travel.
and night (due to greater volumes of trafc), the most notewor- In general quadratic regression analysis shows signicant R-
thy features of the diagram are clusters during the day travelling squared and p-values between crash rates, and means of sinuosity,
eastbound between segments 90110 in the Clyde Mountain area critical visual points, and gradient. There is an improvement
(the road section with maximum sinuosity and critical points). At in quadratic R-squared values rather than in linear regression.
night-time, there are no clusters; the number of crashes at night Appendix A summarises possible quadratic correlation coefcients
does not seem to vary according to straightness or sinuosity. When between crash rates and road geometry variables in Kings Highway.
crash rates are standardised per trafc per length and compared in The signicant R-squared values are shown in bold.
aggregate, Mann Whitney nonparametric test results suggest that Fig. 6 illustrates associations between sections of the road with
day-time eastbound crashes are signicantly different between more than 0 crashes, and mean road geometry variables at the
straight sections and curves (p < 0.05); this is not the case for night aggregate level, between day/night driving. The X-axis illustrates
crashes. mean sinuosity index in the rst row, mean critical visual points in
the second row, and mean gradient in the third row per 5 road seg-
3.4. Regression analysis ments. The Y-axis summarises crash rates standardised per 100,000
moving vehicles per 5 road segments. The quadratic regression
Table 2 illustrates sections of the Kings Highway with or with- analysis for day-time crashes shows signicant R-squared values
out crashes. The results are compared at an aggregate level, where between changes of crash rates and mean sinuosity index, critical
the total number of road segments is combined to form 24 sec- visual point, and gradient. No signicant results are found between
14 S. Alian et al. / Accident Analysis and Prevention 95 (2016) 819
Fig. 5. Crash frequency distribution through Kings Highway: a comparison between day and night.
night-time crashes and the mean of road geometry factors. The day-
time regression results suggest the sinuosity index R-squared value
is not as signicant as critical visual points and gradient, and there
is a limit to the impact of the sinuosity index on rates of crashes.
Fig. 7 summarises regression analysis between crash rates for
sections with more than 0 crashes and means for the road geom-
etry factors, during the day and night, travelling eastbound and
westbound at the disaggregated level. The X-axis illustrates mean
sinuosity index in the rst and second row, mean critical visual
point in the third and fourth rows and mean gradient in the fth and
sixth rows per 5 road segments. The Y-axis summarises crash rates
standardised per 100,000 moving vehicles per 5 road segments. R-
squared and p-values are compared for both quadratic and linear
regressions to select the better t regression. The results indicate
that quadratic regression is a better t than linear analysis because
the R-squared increases and might better indicate behavioural con-
straints on the perception of highly sinuous sections. In summary,
the results suggest that there is a signicant quadratic correlation
(p < 0.05) between eastbound day-time crashes, mean critical visual
points, and mean gradient. No signicant or relevant correlation is
found between westbound day and night crashes and eastbound
night crashes with mean sinuosity, mean critical visual points and
mean gradient. Signicant R-squared values are found between
eastbound day crashes and mean critical points and gradient. The
regression curves and equations show an increase in the rate of
eastbound day-time crashes and an increase in mean critical visual
points and gradient.
Fig. 8 summarises relationships between crash rates and road
geometry factors along Kings Highway. In this gure, a geometrical
index is used to examine a range of scenarios (day/night and east-
bound/westbound) to take into account any impacts of gradient on
the distribution of crash rates. The geometrical index is a multi-
plicative index, resulting from mean sinuosity sum gradient. The
X-axis illustrates road segments in aggregate, and the Y-axis sum-
marises crash rates standardised per trafc in 5 segments (in the
rst and second column), and the geometry index in the third col-
umn. The highest index corresponds to eastbound travel during
the day, and to section 19 which encompasses the beginning seg-
ments of the Clyde Mountain area (segments 1922). The quadratic
regression analysis results suggest an improvement in R-squared
value from 0.28 to 0.83 by multiplying sinuosity index by gradient,
suggesting the importance of gradient to driver behaviour.
4. Discussion
In this research a simple, new and empirical approach is used to
explore the interplay between behavioural and geometrical mea-
sures of road curvature and its role in crash rate distributions, in
different environmental conditions, along Kings Highway NSW. The
method may be used to consider the change of crash rates along
routes with similar geometrical characteristics and can be fruitfully
employed when detailed road geometry data is unavailable. Sinuos-
ity, gradient, and critical visual points are the three main variables
used to examine associations between geometrical and behavioural
measures of road characteristics and the distribution of crashes
(crash rate), for day/night and eastbound/westbound travel. The
visualisation and segmentation method in this research is useful
for summarising crash data and road geometry factors, and sug-
gests some early interpretations that might be explored through
further research and analysis. The idea of identifying crash clusters
might be used for recognising hazardous road segments or sections
of the road that possess underlying safety problems, in order to
create safer road systems. Summarising and analysing the change
of crash rates between day/night and eastbound/westbound travel
might be used for road safety review studies and different poli-
cies (e.g. speed limits) might be implemented to take account of
differences between day/night driving.
In this research both geometrical and behavioural aspects of
road characteristics are introduced to consider the effect of driver
error on curves. Tangent points are a useful dynamic spatial refer-
ence when analysing gaze strategies during driving, even though it
is not exactly xed (Land and Lee, 1994; Mars and Navarro, 2012).
 S. Alian et al. / Accident Analysis and Prevention 95 (2016) 819 15

Fig. 6. Regression between crash rate and road geometry factors: day and night comparison (aggregated).

They provide perceptual information relevant for steering whilst
driving through curves. Drivers do not need to extract geometric
road characteristics, such as curvature, but may safely rely on con-
tinuously available visual information (Coutton-Jean et al., 2009).
The results of this research follow previous ndings and suggest
that critical visual points might be used as a proxy for identifying
the change of road curvature, and might be considered in future
road safety studies.
A key nding is that rates of crashes do not vary according to
changes in road geometry factors at night, but vary signicantly
during the day. The high crash rates at night may be due to speed
(Fildes et al., 1989), lack of visual eld, fatigue or alcohol (Plainis
et al., 2006a), voluntary risk taking of the driver (Konstantopoulos
et al., 2010), and driver age or experience (Underwood et al., 2002).
The results for day-time, eastbound travel suggest that road geom-
etry has a stronger inuence on crash rates on curves (Jurewicza
et al., 2014), in combination with the road environment and driver
behaviour and because of the difference between drivers visual
cues in different travelling directions (Land and Lee, 1994; Authi
and Mestre, 2012). The absence of a signicant difference between
crash rates on straights and curves at night, may be linked to a
constrained visual eld due to headlights (Aarts and Van Schagen,
2006) and/or limited visual cues in the visual eld (Fildes et al.,
1989; Baker, 1999; Plainis et al., 2006b; Konstantopoulos et al.,
2010). The rate of day crashes increases with increasing sinuosity
and grade (geometrical measures of road characteristics), and criti-
cal visual points (a behavioural measure of road curvature), but it is
not the case for night driving. It is obvious that sinuosity and grade
are constant between day and night, so the difference is likely due
to changes in the visual information of drivers, something which
16 S. Alian et al. / Accident Analysis and Prevention 95 (2016) 819

Fig. 7. Regression between crash rate and road geometry factors: day and night comparison (disaggregated).

is signicantly different between day and night. At night, driv- gest that existing Poisson and Negative Binomial models might be
ing under headlamps with background information absent, drivers more relevant at night when limits of the perceptual eld makes
may behave on curved sections as they would on straights. It sug- the effect of topography less important than day-time.
 S. Alian et al. / Accident Analysis and Prevention 95 (2016) 819
Fig. 8. Summary of crash rate and geometry index distribution in Kings Highway.
Identifying the sections of the road with no crashes is a
good way of recognising safe and unsafe section of the road. In
addition, because crash rates are compared for day/night and east-
bound/westbound travel it is a good way of comparing road safety
in different environmental conditions. The results suggest that the
number of sections with no crashes are different between day/night
and eastbound/westbound travel. The number of sections with no
crashes is higher at night.
The interaction between the sinuosity index and critical visual
points suggest that there is a limit on the impact of geometrical
measures of road curvature (sinuosity) on changes in behavioural
measures of road curvature (critical visual points). As a result,
quadratic regression can better describe the relationship because
quadratic regressions are useful when curvilinear effects are
present, and can be used as an approximation function for pos-
sible complex nonlinear functions. The regression analysis results
suggest that quadratic regression better ts the data than linear
ones due to the improvement in both R-squared and p-values. The
results conrm previous ndings in the literature that non-linear
functions better characterise the relationship between crashes and
explanatory variables than linear regression because linear regres-
sion models lack the distributional property to describe random
and discrete crashes (Miaou and Lum, 1993; Lord and Mannering,
2010). R-squared and p-values comparison suggests that regression
analysis results are signicant between mean road geometry fac-
tors and crash rates for day-time driving, and there is a limit to the
effect of sinuosity on crash statistics during the day. In addition,
crash distribution patterns and regression line slopes are a simple
way to distinguish the variation of crash rates in different sections
of the road between day/night and eastbound/westbound travel.
No signicant results at night-time suggest that quadratic regres-
sion might not be suitable for night driving because of the random
distribution of accidents. It also suggests that night-time crashes
are more likely due to behavioural errors (e.g. fatigue) and do not
change according to changes in road geometry characteristics.
As crash rates or clusters increase in a specic road segment,
road safety might decrease and actions such as road upgrades,
changing speed limits and/or additional police enforcement might
be required. However, the ndings reported here indicate that crash
rates vary signicantly, not only according to different road geome-
tries, but also in different environmental conditions. In addition,
17
due to a close relationship between speed limits and crash rates
(Kloeden et al., 2001; Aarts and Van Schagen, 2006), and the sig-
nicant effect of speed limit restrictions on crash rates (De Pauw
et al., 2014), further consideration might be given to greater util-
isation of variable speed limits. The cumulative effect of curves is
a way of determining advisory speed limits that can be explored
by this method with further research. In addition; mean travel
speeds are different between day/night and eastbound/westbound
travel. Higher travel speed at night might be due to the fewer visual
cues, voluntary risk taking of the driver, and underestimating the
speed at night. Lower travel speed for eastbound travel could be
due to steeper grades, advisory signs, speed limits, maximum sin-
uosity and negative gradient. The results might suggest reviewing
the speed limits and considering speed limit updates specically
between day and night to increase road safety and decrease crash
rates.
Crash numbers for males and females differ signicantly
between day/night and eastbound/westbound travel. The results
are statistically signicant. This may be due to psychological and
physiological differences between male and female drivers as
discussed in previous studies. Young male drivers are more respon-
sible for night and loss-of-control crashes due to speeding, risk
taking, and the way they use roads (Clarke et al., 2006). Different
crash rates for males and females travelling eastbound and west-
bound may be due to differences between male and female driving
skills (Laapotti and Keskinen, 1998). The results raise a question
about whether there are any fundamental differences between
male and female drivers in collecting and processing visual infor-
mation. It might be due to the fact that female information comes
more from the forward eld (i.e. road), whereas male information
receives from greater contributions from the lateral eld (e.g. envi-
ronment). It might also be the cause of a higher number of crashes
by males at night where lateral information is restricted from the
constraints of headlights. Additional research is required in this
area.
The adopted approach and polynomial regression analysis in
this research are capable of predicting and analysing the rate of
changes during day-time, but they cannot quantify the distribution
of crashes at night because no changes were found due to constant
visual cues. It might be considered that all of the results discussed
above are only relevant to Kings Highway, but it may be possible
18 S. Alian et al. / Accident Analysis and Prevention 95 (2016) 819

to begin to generalise for routes with similar geometries if further
similar studies are undertaken. In addition, the adopted approach
might be used for some initial interpretations at local or national
scales, but further analysis and more detailed data is necessary for
predictions. Some of the other limitations of this research are not
considering the effect of some road characteristics (e.g. lane and
shoulder width, and overtaking zones), additional factors such as
animals and alcohol, type of vehicles, and type of crashes (e.g. run-
off, rear end, or turning) due to data limitations. In addition, further
research and some simulation studies might be required to inves-
tigate the differences between male and female drivers and speed
in day/night and eastbound/westbound directions.
More in-depth analysis of crash data, using and applying the
method to new study areas with similar and different geometrical
characteristics are some of the research that is in progress by the
authors, and is required to validate and generalise the results.
5. Conclusions
This research develops an empirical approach to road safety
studies using exploratory data analysis. It considers interactions
between geometrical and behavioural measures of road curvature,
and their impact on the geographical distribution of casualty crash
rates in different driving environments along Kings Highway.
The regression analysis results suggest that crash rates vary
signicantly according to changes in road geometry factors at day-
time, especially for eastbound travel, due to changes in the visual
cues of drivers. However no signicant differences were found
between straight and curved sections of the road at night, due to
the effect of a reduced visual eld and the constraint of vehicular
headlights.
The method considers the differences in crash rates between
day/night and eastbound/westbound driving and might be used
for preliminary analysis of road safety studies by government of-
cials. The results suggest there are complex interactions between
geometrical and behavioural constraints on driver information.
However, at night-time and on straight road segments, modelling
based on the assumption of a random distribution of crashes seems
more appropriate.
The outcomes suggest to national authorities and governmental
departments that different modelling techniques might be required
to distinguish between day/night driving and different travelling
directions. It also suggests that different safety policies such as dif-
ferent speed limits or advisory signs might be required in this type
of road.
This approach might be used for identifying safe and unsafe
sections of the road and differences between day/night and
eastbound/westbound driving. It is also capable of identifying haz-
ardous road sections with crash clusters, and might be used by local
and national authorities in the development of public policies. Most
government policies are location-specic, and this approach might
be used for safety reviews and road upgrades (e.g. road straighten-
ing and speed limit change).
One of the limitations of this research is that it is not able to
quantify the rate of change during night driving when the crashes
are randomly distributed. In addition, the data does not include
detailed road geometry characteristics such as lane and shoulder
width. Finally, it does not specify the role played by use of drugs and
alcohol, type of motor vehicle, and type of crash. Further research
might be required to analyse the effect of driver speed and fatigue,
and differences between day/night driving.
All the outcomes that are drawn for Kings Highway have the
potential to be generalised for roads with similar characteristics.
The adopted approach will be implemented in other study areas
in NSW with similar and different geometries in future studies to
facilitate generalisation and validation of the results.
In summary, apart from the signicant effect of road character-
istics on road safety, it is necessary to evaluate and consider the role
of environmental conditions and driver behaviour, and the interac-
tion between these three in any road safety and crash analysis study
because crash rates are signicantly different in geographical loca-
tions with similar road characteristics but different environmental
conditions.
Acknowledgment
We need to acknowledge the NSW Roads and Maritime Services
for providing crash and trafc volume data, and the Department of
Lands for providing road centreline data. It was not possible to do
this research without their support.
A brief version of this paper was presented at the 2015 Aus-
tralasian Road Safety Conference.
Appendix A.
See Table A1.

Table A1
Quadratic correlation coefcients between crash rates and road geometry variables in aggregated level along the Kings Highway.

Day Crashes Night Crashes East Crashes West Crashes Mean SI Deviation SI Mean CVP Deviation CVP Mean Grade Deviation Grade

Total Crashes 0.937 0.537 0.890 0.236 0.295 0.194 0.482 0.538 0.664 0.172
0.000 0.001 0.000 0.102 0.036 0.130 0.002 0.001 0.000 0.166
Day Crashes 0.432 0.861 0.266 0.291 0.199 0.468 0.512 0.637 0.152
0.008 0.001 0.073 0.044 0.121 0.003 0.001 0.000 0.210
Night Crashes 0.362 0.146 0.185 0.042 0.158 0.206 0.164 0.084
0.028 0.307 0.176 0.693 0.232 0.141 0.219 0.474
East Crashes 0.103 0.321 0.218 0.580 0.716 0.798 0.197
0.421 0.031 0.109 0.000 0.000 0.000 0.139
West Crashes 0.018 0.022 0.161 0.027 0.133 0.013
0.857 0.829 0.224 0.795 0.297 0.896
Mean SI 0.755 0.864 0.590 0.501 0.537
0.000 0.000 0.000 0.001 0.000
Deviation SI 0.585 0.592 0.285 0.613
0.000 0.000 0.030 0.000
Mean CVP 0.576 0.663 0.542
0.000 0.000 0.000
Deviation CVP 0.565 0.454
0.000 0.002
Mean Grade 0.558
0.000

Note. The rst number in each cell is R-squared value and the second number is p-value. Signicant results are highlighted in bold (p < 0.05).
 S. Alian et al. / Accident Analysis and Prevention 95 (2016) 819
References
Aarts, L., Van Schagen, I., 2006. Driving speed and the risk of road crashes: a
review. Accid. Anal. Prev. 38 (2), 215224.
Ackaah, W., Salifu, M., 2011. Crash prediction model for two-lane rural highways in
the ashanti region of ghana. IATSS Res. 35 (1), 3440.
A, A., May, S., Clark, V.A., 2011. Practical Multivariate Analysis. CRC Press.
Austroads, 2010. Rural run-off road crashes. Road Safety Engineering Assessment.
Austroads, 2015. Road Geometry Study for Improved Rural Safety. Austroads Ltd.
Authi, C.N., Mestre, D.R., 2012. Path curvature discrimination: dependence on
gaze direction and optical ow speed. PLoS One 7 (2), e31479.
Baker, R.G.V., 1999. On the quantum mechanics of optic ow and its application to
driving in uncertain environments. Transp. Res. Part F 2, 2753.
Castro, C., 2008. Human Factors of Visual and Cognitive Performance in Driving.
CRC Press.
Chattington, M., Wilson, M., Ashford, D., Marple-Horvat, D., 2007. Eye-steering
coordination in natural driving. Exp. Brain Res. 180 (1), 114.
Clarke, D.D., Ward, P., Bartle, C., Truman, W., 2006. Young driver accidents in the
UK: the inuence of age, experience, and time of day. Accid. Anal. Prev. 38 (5),
871878.
Coutton-Jean, C., Mestre, D.R., Goulon, C., Bootsma, R.J., 2009. The role of edge lines
in curve driving. Transp. Res. Part F: Trafc Psychol. Behav. 12 (6), 483493.
De Pauw, E., Daniels, S., Thierie, M., Brijs, T., 2014. Safety effects of reducing the
speed limit from 90 km/h to 70 km/h. Accid. Anal. Prev. 62 (0), 426431.
Elvik, R., Christensen, P., Amundsen, A., 2004. Speed and road accidents: an
evaluation of the power model.
Fildes, B.N., Leening, A.C., Corringan, J.M., 1989. Speed Perception 2: Drivers
Judgments Os Safety and Travel Speed on Rural Curved Roads at Night. Federal
Ofce of Road Safety.
Forkenbrock, D.J., Foster, N.S.J., 1997. Accident cost saving and highway attributes.
Transportation 24 (1), 79100.
Golob, T.F., 2003. Structural equation modeling for travel behavior research.
Transp. Res. 37B, 115.
Groeger, J.A., 2000. Understanding Driving: Applying Cognitive Psychology to a
Complex Everyday Task. Psychology Press.
Haynes, R., Jones, A., Kennedy, V., Harvey, I., Jewell, T., 2007. District variations in
road curvature in england and wales and their association with road-trafc
crashes. Environ. Plann. A 39 (5), 1222.
Hosseinpour, M., Yahaya, A.S., Sadullah, A.F., 2014. Exploring the effects of
roadway characteristics on the frequency and severity of head-on crashes:
case studies from Malaysian federal roads. Accid. Anal. Preve. 62 (0), 209222.
Jones, A.P., Haynes, R., Harvey, I.M., Jewell, T., 2012. Road trafc crashes and the
protective effect of road curvature over small areas. Health Place 18, 315320.
Jurewicza, C., Chaub, T., Mihailidisb, P., Buic, B., 2014. From research to
practice-development of rural mass curve treatment program. Proceedings of
the Proceedings of the 2014 Australasian Road Safety Research, Policing &
Education Conference.
Kandil, F.I., Rotter, A., Lappe, M., 2009. Driving is smoother and more stable when
using the tangent point. J. Vis. 9 (1), 1111.
Kloeden, C.N., Ponte, G., Mclean, A.J., 2001. Travelling speed and the risk of crash
involvement on rural roads.
Konstantopoulos, P., Chapman, P., Crundall, D., 2010. Drivers visual attention as a
function of driving experience and visibility. Using a driving simulator to
explore drivers eye movements in day, night and rain driving. Accid. Anal.
Prev. 42 (3), 827834.
View publication stats
19
Laapotti, S., Keskinen, E., 1998. Differences in fatal loss-of-control accidents
between young male and female drivers. Accid. Anal. Prev. 30 (4), 435442.
Land, M.F., Lee, D.N., 1994. Where we look when we steer. Nature 369 (6483),
742744.
Leopold, L., Wolman, M., 2013. River channel patterns. Fluv. Geom.: Geom. Crit.
Conc. 3.
Lord, D., Mannering, F., 2010. The statistical analysis of crash-frequency data: a
review and assessment of methodological alternatives. Transp. Res. Part A:
Policy Pract. 44 (5), 291305.
Mamcic, S., Sivilevicius, H., 2013. The analysis of trafc accidents on lithuanian
regional gravel roads. Transport 28 (1), 108115.
Mannering, F.L., Bhat, C.R., 2014. Analytic methods in accident research:
methodological frontier and future directions. Anal. Methods Accid. Res. 1 (0),
122.
Mars, F., Navarro, J., 2012. Where we look when we drive with or without active
steering wheel control. PLoS One 7 (8), e43858.
Miaou, S.P., Lum, H., 1993. Modeling vehicle accidents and highway geometric
design relationships. Accid. Anal. Prev. 25 (6), 689709.
Plainis, S., Murray, I., Pallikaris, I., 2006a. Road trafc casualties: understanding the
night-time death toll. Inj. Prev. 12 (2), 125138.
Plainis, S., Murray, I.J., Pallikaris, I.G., 2006b. Road trafc casualties: understanding
the night-time death toll. Inj. Prev. 12 (2), 125128.
RMS, 2013. Kings highway rout safety review. Transport for NSW.
RTA, 2011. Nsw speed zoning guidlines. In: TRT (Ed.), Authority. NSW Centre for
Road Safety.
Rakha, H., Arafeh, M., Abdel-Salam, A., Guo, F., Flintsch, A., 2010. Linear regression
crash prediction models: issues and proposed solutions.
Sagberg, F., 1999. Road accidents caused by drivers falling asleep. Accid. Anal. Prev.
31, 639649.
Savolainen, P.T., Mannering, F.L., Lord, D., Quddus, M.A., 2011. The statistical
analysis of highway crash-injury severities: a review and assessment of
methodological alternatives. Accid. Anal. Prev. 43 (5), 16661676.
Shankar, V., Mannering, F., Bareld, W., 1995. Effect of roadway geometrics and
environmental factors on rural freeway accident frequencies. Accid. Anal. Prev.
27 (3), 371389.
Underwood, G., Chapman, P., Bowden, K., Crundall, D., 2002. Visual search while
driving: skill and awareness during inspection of the scene. Transp. Res. Part F:
Trafc Psychol. Behav. 5 (2), 8797.
Wang, C., Quddus, M.A., Ison, S.G., 2013. The effect of trafc and road characteristics
on road safety: a review and future research direction. Saf. Sci. 57 (0), 264275.
Washington, S.P., Karlaftis, M.G., Mannering, F.L., 2010. Statistical and Econometric
Methods for Transportation Data Analysis. CRC press.
Weller, G., 2010. The Psycology of Driving on Rural Roads. PhD. Dissertation.
Technische Universitt Dresden.
Williams, G.P., 1986. River meanders and channel size. J. Hydrol. 88 (1), 147164.
Yu, R., Abdel-Aty, M., 2014. Using hierarchical bayesian binary probit models to
analyze crash injury severity on high speed facilities with real-time trafc
data. Accid. Anal. Prev. 62 (0), 161167.
Yu, H., Liu, P., Chen, J., Wang, H., 2014. Comparative analysis of the spatial analysis
methods for hotspot identication. Accid. Anal. Prev. 66 (0), 8088.