You are on page 1of 22

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

net/publication/340104531

Development of crash modification factors for intersections in Toowoomba


city

Article  in  International Journal of Urban Sciences · March 2020


DOI: 10.1080/12265934.2020.1743739

CITATIONS READS

0 23

3 authors, including:

Mohammad Nour Al-Marafi Kathirgamalingam Somasundaraswaran


Tafila Technical University University of Southern Queensland 
4 PUBLICATIONS   2 CITATIONS    11 PUBLICATIONS   9 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Mohammad Nour Al-Marafi on 23 April 2020.

The user has requested enhancement of the downloaded file.


International Journal of Urban Sciences

ISSN: 1226-5934 (Print) 2161-6779 (Online) Journal homepage: https://www.tandfonline.com/loi/rjus20

Development of crash modification factors for


intersections in Toowoomba city

Mohammad Nour Al-Marafi, Kathirgamalingam Somasundaraswaran &


Frank Bullen

To cite this article: Mohammad Nour Al-Marafi, Kathirgamalingam Somasundaraswaran & Frank
Bullen (2020): Development of crash modification factors for intersections in Toowoomba city,
International Journal of Urban Sciences, DOI: 10.1080/12265934.2020.1743739

To link to this article: https://doi.org/10.1080/12265934.2020.1743739

Published online: 23 Mar 2020.

Submit your article to this journal

Article views: 13

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=rjus20
INTERNATIONAL JOURNAL OF URBAN SCIENCES
https://doi.org/10.1080/12265934.2020.1743739

Development of crash modification factors for intersections in


Toowoomba city
Mohammad Nour Al-Marafi, Kathirgamalingam Somasundaraswaran and Frank Bullen
School of Civil Engineering and Surveying, University of Southern Queensland, Toowoomba, Australia

ABSTRACT ARTICLE HISTORY


This research outlines the analytical approaches that may be Received 10 December 2018
employed to recommend multiple treatments for application to Accepted 4 March 2020
improve safety level at hazardous intersections in regional areas.
KEYWORDS
Data from 106 intersections in the regional city of Toowoomba, Crash prediction models;
Queensland, Australia were used in modelling and subsequent road safety measures;
analysis. Specifically, crash prediction models were developed to Empirical-Bayes method;
estimate crash modification factors that could be used to propose crash modification factors;
remedial measures. A total of sixteen geometric and operational hazardous intersections
conditions were used as explanatory variables including the
number of legs at an intersection, the availability of entering and
exiting via through lanes, the presence of left-turning lanes, right-
turning lanes, slip lanes, median islands, traffic control, and speed
limits on the feed roads. The Empirical Bayes approach was
employed to identify five hazardous intersections in Toowoomba
for further investigation. The four most suitable techniques for
estimating combined crash modification factors were reviewed
and then used to propose effective road safety measures for the
five selected hazardous intersections. Finally, this research
suggests four future areas worthy of further investigation.

Highlights
. Estimated the number of road crashes at intersections using
crash prediction models.
. Applied the Empirical Bayes method to identify the most
hazardous intersections.
. Combined crash modifications factors were calculated using
various techniques.
. Identified the effective countermeasures to reduce the road
crashes at hazardous intersections.

1. Introduction
Road crashes cause substantial socio-economic losses and account for many billions of
dollars’ worth of overall damage (WHO, 2015). In the process of attempts to address
these issues, countries continue to research and apply more radical approaches to
ensure road safety. In Australia, the National Road Safety Strategy (NRSS), introduced
in 2011, targeted a 30% reduction in road fatalities by 2020 (BITRE, 2015).

CONTACT Mohammad Nour Al-Marafi moh.marafi89@gmail.com Faculty of Health, Engineering & Sciences,
School of Civil Engineering and Surveying, University of Southern Queensland, Toowoomba, Queensland 4350, Australia
© 2020 The Institute of Urban Sciences
2 M. N. AL-MARAFI ET AL.

Road engineers and transport planners have used Crash Prediction Models (CPMs) as a
useful tool to investigate and improve level of road safety. A number of studies have been
conducted over the years on the development of CPMs for predicting possible crashes on a
road network. Given that intersections are amongst the most hazardous sites on city road
networks (due to both geometric configuration and traffic concentration), they are among
the important elements that must be considered if the road network is to facilitate the safe
flow of traffic in all directions.
Statistics indicate that 43.5% of all road crashes (fatalities and hospitalized injuries) in
the State of Queensland during the period 2008–2015 occurred at intersections. In Too-
woomba city, it was reported that 50.4% of all road crashes took place at intersections
during the same period (Queensland Government, 2016).
In recent years, intensive studies have been conducted to investigate the impact of
various geometric design parameters and traffic volume at intersections on safety, using
CPMs and Crash Modification Factors (CMFs). CMFs, in particular, can be used to ident-
ify the safety impact after introducing one or more road safety treatments. However, only a
small number of studies have been conducted in order to estimate the combined effect of
several road safety treatments using the combined CMF techniques (Gross & Hamidi,
2011; Park, Abdel-Aty, & Lee, 2014).
The aim of this study is to investigate the expected safety impact after applying com-
bined treatments and comparing them with individual treatments. To achieve this aim,
the study used Negative Binomial (NB) distribution models to investigate and predict
crash frequency at intersections. These models were used to identify the geometric and
traffic factors that would contribute to crashes at those intersections. The Empirical
Bayes (EB) method was then used to identify local hazardous (black spot) intersections.
These locally developed models were then used to estimate CMFs at the hazardous inter-
sections to determine how each treatment could affect road safety. Combined CMFs for
multiple treatments were also estimated using the four existing techniques including
Highway Safety Manual (HSM), Turner, Alabama Method, and the most effective CMF
(lower value) technique.
The remainder of the study is organized as follows. The second section presents the pre-
vious studies related to the development of CPMs and CMFs. The third section describes
the methodology used for estimating the CPMs. The fourth section describes the data used
in the analysis. The fifth section describes the road CMFs derived from the data and CPMs.
The last section draws conclusions from the analysis performed in this study.

2. Literature review
2.1. Crash prediction models (CPMs)
Road crashes are caused by a combination of many factors, including the road environ-
ment, vehicles, and road user behaviour. In their study, Bauer and Harwood (2000)
used lognormal, Poisson and negative binomial regression analyses to develop CPMs to
investigate the relationship between road crashes and highway geometry, traffic control,
and traffic volume variables for at-grade intersections. The form of the statistical distri-
bution selected for modelling any type of intersection should be selected based on a
review of the crash frequency distribution for that type of intersection. Accordingly, the
INTERNATIONAL JOURNAL OF URBAN SCIENCES 3

authors concluded that the Poisson, negative binomial, and lognormal distributions were
more suitable for modelling road crashes than the normal distribution. Greibe (2003)
developed CPMs for urban intersections in Denmark using a Poisson distribution and
found that the most important variables in the models were those describing the
vehicle traffic flow.
Abdel-Aty and Radwan (2000) argued that Poisson models have some limitations such
as the mean must equal the variance of the crash number (dependent variable). In most
crash data, the variance of the crash number exceeds the mean and, in such cases, the
data would be over-dispersed. In order to solve the limitation of overdispersion in a
Poisson regression model, the studies (Chin & Quddus, 2003; Kim, Washington, & Oh,
2006; Lord & Mannering, 2010) have recommended using alternative models to overcome
the limitations of Poisson regression. One alternative is the use of negative binomial
regression which does not require equal mean and variance assumption. Further, Chin
and Quddus (2003) developed random effect negative binomial (NB) models to identify
the contributory factors that affect intersection safety. Crash data from a total of 52 inter-
sections in Singapore were used in the analysis for the data collected between the years
1992 and 1999. In this study, a total of 32 explanatory variables were considered for
including geometric characteristics, regulatory control measures, and traffic volume.
The results revealed that 11 explanatory variables significantly affected the road safety
at the intersections where four of them were considered to be highly significant. These
were the total traffic volume, uncontrolled left turn lane, number of phases per cycle,
and presence of a surveillance camera.
Elvik, Vadeby, Hels, and van Schagen (2019) conducted a meta-analysis to review the
relationship between the road safety and speed of traffic based on studies published after
2000. The findings showed that, using either the Power Model or an Exponential Model, a
strong relationship between the road safety and speed of traffic was achieved. In addition,
it was concluded that the speed of traffic is one of the most important parameters that sig-
nificantly affect road safety. In other work, Yu, Quddus, Wang, and Yang (2018) investi-
gated the impacts of the following data aggregation schemes on the relationships between
road safety and speed of traffic: (1) a segment-based dataset in which crashes are grouped
by roadway segment, and (2) a scenario-based dataset where crashes are aggregated by
traffic operating scenarios. Result revealed that the relationship between crash frequency
and speed of traffic as a U-shape curve (i.e. crash frequencies are relatively high at both
high and low speed conditions).

2.2. Empirical Bayes (EB)


Identification of hazardous road locations, sometimes known as black spots, is normally
considered as the first step in a road crash reduction process. Elvik (2008) refers to
black spots as locations that have a higher predicted number of road crashes than what
would be considered as being normal at other similar locations. According to Hauer,
Kononov, Allery, and Griffith (2002) the identification of hazardous locations signifies a
list of spots being prioritized for further research and engineering investigation which
can distinguish road crash patterns, effective variables, and potential countermeasures.
In those processes, cost-effective remedial projects are often selected to obtain the
optimal outcomes from limited resources. Therefore, researchers have introduced the
4 M. N. AL-MARAFI ET AL.

EB approach as an effective means of solving the Regression-To-Mean (RTM) problem


that has been associated with crash data (Da Costa, Qu, & Parajuli, 2015; Gross,
Persaud, & Lyon, 2010; Tegge, Jo, & Ouyang, 2010). Through calculating the difference
between predicted and actual crashes at the location, this approach identifies high crash
locations (black spots) based on their Potential for Safety Improvement (PSI). A positive
value of PSI shows that the potential exists for safety improvements. According to Elvik
(2007) and Hauer, Harwood, Council, and Griffith (2002) by calculating the weighted
combination of the number of predicted and recorded crashes, the EB approach can
provide an expected crash frequency for a specific road intersection.
Recently, researchers have introduced the use of the Full Bayes (FB) approach for iden-
tifying hazardous road locations and evaluating the impact of safety treatments (Fawcett,
Thorpe, Matthews, & Kremer, 2017; Heydari, Miranda-Moreno, & Amador, 2013; Huang,
Chin, & Haque, 2009; Lan, Persaud, Lyon, & Bhim, 2009). This approach has shown
several advantages over the EB approach, including the ability to account for all uncertain-
ties in the data used, requiring less data, providing more flexibility in selecting crash fre-
quency distributions, providing more detailed causal inferences, and the ability to consider
the effect of one site’s proximity to other sites (i.e. spatial correlation) in the model formu-
lation. However, it is worth mentioning that the complexity of the FB approach makes the
EB approach more attractive to use (Khan, Abdel-Rahim, & Williams, 2015; Persaud, Lan,
Lyon, & Bhim, 2010).

2.3. CMFs for individual treatments


CMFs or Crash Reduction Factors (CRFs) can provide simple and quick arithmetic
methods for estimating crash reductions. The relationship between the CMFs with
CRFs is: CMF = 1 − CRF/100 and CMF = Nw/Nw-o, where Nw is the expected number of
crashes with the safety improvement and Nw-o is expected number of crashes without
the improvement (AASHTO, 2010).
CMFs are used with a road safety prediction model to estimate the expected crash fre-
quencies for a specific site and/or to estimate the effect of a change in conditions relating to
road safety. In this regard, Bonneson and Lord (2005) indicated that CMFs typically range
in value from 0.5 to 2.0, with a value of 1.0 indicating no effect on safety by any change in
geometric design and/or traffic control feature. When the CMF is less than 1.0 the treat-
ment reduces the predicted number of crashes, whereas when it is greater than 1.0, the
treatment increases the predicted number of crashes.
The USA’s Highway Safety Manual (HSM) Volume 3 Part D (AASHTO, 2010) and
other studies (Bahar, 2010; Bonneson & Pratt, 2009; Gross et al., 2010; Li, Lord, &
Zhang, 2010; Oh & Park, 2014; Persaud et al., 2010) used the EB before-and-after
method and/or cross-sectional method for estimating safety effectiveness as well as for cal-
culating the CMFs of specific roadway treatments.
The CMFs in the HSM were estimated using observational before-and-after studies that
account for the regression-to-the-mean bias. Generally, there are five approaches can be
employed to implement observational before-and-after studies: (1) Comparison Group
(CG) approach; (2) naïve approach; (3) Full Bayes (FB) approach; (4) Empirical Bayes
(EB) approach; and (5) yoked comparison approach (Elvik et al., 2017; Lan et al., 2009;
Park, Abdel-Aty, Wang, & Lee, 2015; Persaud et al., 2010; Wang, Abdel-Aty, & Wang,
INTERNATIONAL JOURNAL OF URBAN SCIENCES 5

2017). There are some limitations associated with observational before-and-after studies.
For example, the treatment date should be known to determine the evaluation periods and
several years may need to elapse after implementing any treatment to collect accurate data.
It is also difficult to distinguish safety effects when implementing more than one treatment
at a site. In such cases, the cross-sectional method can be employed to estimate CMFs
because of its simplified approach for obtaining data compared to observational before-
and-after studies. AASHTO (2010) also indicated that the cross-sectional method might
be appropriate when observational before-and-after studies are not practical due to data
restrictions (e.g. crash data in the before period are not available). It is worth mentioning
that the cross-sectional method does not take into account the effects of factors that are not
included in the analysis, i.e. external causal factors (Hauer, 2013). Another limitation is
that a sufficient sample size is especially required when large explanatory variables are
included in the developed model.
The Crash Modification Function (CMFunction) uses the coefficients of prediction
models (Gross et al., 2010; Lee, Abdel-Aty, Park, & Wang, 2015; Lord & Bonneson,
2007; Park et al., 2014, 2015; Sacchi, Sayed, & El-Basyouny, 2014; Wood, Donnell, &
Porter, 2015) to estimate the safety benefits. As a part of the cross-sectional method,
CMFunction has been employed recently to derive CMFs at a specific site. Wood
et al. (2015) compared the CMFs obtained from an observational before-and-after
study using the EB approach and the cross-sectional method using the regression
model. The study revealed that the cross-sectional method appears to be consistent
with the result of the EB approach; therefore, using the cross-sectional method will
yield a reasonable result where data for after treatments are not available. Likewise,
Park et al. (2015) and Sacchi et al. (2014) proposed using CMFunctions based on a
cross-sectional approach to identify the relationship between safety effects and
roadway characteristics.

2.4. CMFs for multiple treatments


There are a number of techniques proposed to estimate the value of combined CMFs for
multiple treatments. The first of these techniques was adopted by the USA’s HSM
(AASHTO, 2010) and this technique assumes that the road safety effect of each treatment
is independent when CMFs for individual treatments are multiplied to estimate combined
CMFs (Park et al., 2014; Wu & Lord, 2016). For this technique, Equation (1) was used to
estimate combined CMF at the i th intersection:
CMFcombined, i = CMFi1 × CMFi2 × . . . × CMFij × . . . × CMFin (1)
Where, CMFin: crash modification factor associated with treatment j ( j = 1, 2, … , n) at i th
intersection.
A second technique was proposed by Turner (2011), where a specific weighted factor of
2/3 (two-thirds) is applied when estimating combined CMFs for two or more treatments.
The combined CMF is estimated using Turner’s technique as in Equation (2).
 
2
CMFcombined, i = 1 − (1 − (CMFi1 × CMFi2 × . . . × CMFij × . . . × CMFin )) (2)
3
6 M. N. AL-MARAFI ET AL.

The third technique was proposed by the U.S. State of Alabama (NCHRP, 2008), which
assumed that the safety effects of the less effective safety treatment are systematically
reduced. This means that the full effect of the most effective safety treatment among all
treatments is used and has an added benefit of additional treatments (i.e. less effective
treatments) as detailed in Equation (3).
1 − CMFi2 1 − CMFij 1 − CMFin
CMFcombined, i = CMFi1 − − ... − − ... − (3)
2 j n
A fourth technique applies only the most effective safety treatment, which is the
lowest CMF. The disadvantage of this technique is in underestimating the combined
effect of safety treatments if the additional safety treatments provide additional
benefits (Gross & Hamidi, 2011). Gross and Hamidi (2011) used the result from two
earlier studies by Pitale, Shankwitz, Preston, and Barry (2009) and Hanley, Gibby,
and Ferrara (2000) to examine the techniques that were used to estimate combined
CMFs. The study used two individual treatments (widening shoulders and installing
shoulder rumble strips) to achieve their study objectives. The results showed that the
combined CMFs that were estimated using the technique adopted by the HSM and
the technique introduced by the state of Alabama were close to the actual CMFs. Park
et al. (2014) examined the existing techniques, and the results showed that the technique
adopted by HSM and the fourth technique (most effective safety treatment technique)
were close to the actual values of CMFs.

2.5. Summary
The conclusion from the previous studies can be summarized as follows:

. Different CPMs have been developed to investigate the impacts of various geometric
and traffic variables on crash frequencies. However, the statistical techniques such as
Poisson and NB regression models have been widely used as suitable techniques for
developing road crash models.
. Various approaches to identify the black spot locations have been developed. The inte-
gration of expected crash frequency into the method of analysis has been highlighted by
researchers for precise investigations. The EB approach can provide an expected crash
frequency for a specific location by calculating the weighted combination of the
recorded and predicted crash frequencies.
. A cross-sectional method has been widely used in recent years to estimate CMFs. In this
method, the CMF value is estimated for a specific site based on its characteristics before
implementation of the treatment by using the coefficients of the prediction models.
According to previous studies, the results from the cross-sectional method seem to
be consistent with the observational before-after study results.
. Most previous studies estimate CMF as a single value by ignoring the variation of
CMF values among different sites characteristics. In most cases, it is not realistic
to assume a uniform safety impact for all treated sites where intersections are
having different geometric and traffic operational characteristics. Recently, a few
studies estimated CMF values through developing a CMFunctions to overcome
INTERNATIONAL JOURNAL OF URBAN SCIENCES 7

this limitation. A CMFunction allows the value of CMF to change based on site
characteristics.

3. Methodology
3.1. Poisson and NB distribution models
The Generalized Linear Model (GLM), which is the Poisson or NB with log link analysis,
was used in this study. The Poisson regression is a distribution that states the probability of
a certain number of rare events occurring during a given period (Caliendo, Guida, &
Parisi, 2007). As stated previously, the Poisson distribution assumes that the mean and
variance are required to be exactly equal.
When the mean and the variance of the model data are not equal, the Poisson distri-
bution becomes unsuitable for analysing the data. This problem can be resolved by the
use of NB regression instead of Poisson regression. The NB regression describes the occur-
rence of random and rare events. This model can be used in the case of the means being
smaller than the variance.
The general form of the predicted model by using Poisson or NB regression for the i th
intersection can be written in the form of Equation (4).

n
b0 + b j Xij
Npre.i = Qamajor,i
1
· Qaminor,i
2
·e j=1
(4)

Where: Qmajor,i and Qminor,i – annual average daily traffic (AADT) on major and minor
approach at i th intersection, respectively; Xij – explanatory variable j at i th intersection;
α1, α2, β0, and βj – are model parameters.

3.2. Goodness-of-fit
The following three performance tests have been used to determine the goodness-of-fit of
the models.

(a) Akaike’s Information Criterion (AIC) test was used to measure the goodness-of-fit of
each model, relative to each of the other models. This test was defined by Akaike
(1974) as shown in Equation (5).
AIC = −2 log L + 2P (5)

Where: log L – maximum log-likelihood of the Model; P – number of independent vari-


ables in the model excluding the constant.

(a) Bayesian Information Criterion (BIC) test was used to measure the goodness-of-fit of
each model, relative to each of the other models. This test was defined by Schwarz
(1978) as shown in Equation (6).
BIC = −2 log L + ln (n)S (6)
8 M. N. AL-MARAFI ET AL.

Where: n – number of data points (sample size); S – number of independent variables in


the model including the constant.
In general, the smaller the AIC and BIC values, the more preferred the model would be
(Abdul Manan, Jonsson, & Várhelyi, 2013; Cafiso, Di Graziano, Di Silvestro, La Cava, &
Persaud, 2010; Young & Park, 2013).

(a) Pearson’s chi-square test was used to establish whether an observed distribution
differs from a theoretical distribution, i.e. predicted crashes. This test was defined
by Pearson (1934) as shown in Equation (7).

n
(yi − ′̂yi )2
X2 = (7)
i=1
′̂yi

Where: ýi – predicted crashes number at i th intersection; yi – observed crashes number at


i th intersection.

3.3. Model validation


Four performance measures were used to assess the ability of models to predict road
crashes over subsequent additional years including: mean squared prediction error
(MSPE); mean absolute deviation (MAD); mean squared error (MSE); and Freeman-
Tukey R-Squared coefficient (R2FT). MSPE is used to determine the variance of the differ-
ence between observed crashes and predicted crashes results. In addition, it is typically
employed to evaluate error associated with a validation dataset. MSE is typically employed
to evaluate error associated with an estimation dataset. Ideally, MSPE and MSE results can
be used to reveal whether the models are over-fitted (MSPE > MSE) or under-fitted
(MSPE < MSE) (Washington, Persaud, Lyon, & Oh, 2005; Young & Park, 2013). When
the values of MSPE and MSE are similar, this indicates that the validation dataset fit
the developed model similar to the estimation dataset. The MAD value provides a
measure of the average magnitude of the prediction variability. In general, a smaller
value (closer to zero) of MSPE, MAD, or MSE refers to lower prediction error. Likewise,
the higher values of R2FT indicate a better prediction performance and vice-versa
(Washington et al., 2005).

4. Data preparation
The crash data obtained from the Department of Transport and Main Roads, Queensland
included all roads and intersections in Queensland and it was necessary to separate the
intersection crashes for Toowoomba city to select sites for the study. As stated by
Corben and Wai (1990), the use of either high or low crash frequency locations for the
data collection process could lead to concerns about the sample being biased towards
high or low crash frequency approaches. A random selection approach was employed
to avoid any bias.
A sample of 106 intersections, resulting in 1108 fatal and injury crashes, were randomly
selected for the study. Severe crashes were adopted to increase the significance to the
analysis, so the damage only type crashes were removed from the dataset. The dataset
INTERNATIONAL JOURNAL OF URBAN SCIENCES 9

includes 62 signalized intersections with 813 crashes and 44 un-signalized intersections


with 295 crashes. The intersections were separated based on the geographic location in
Toowoomba, to represent the Northern, Southern, Eastern and Western intersections in
Toowoomba.
The study area was divided into four quadrants using James Street and Ruthven Street
to provide a uniform distribution for data selection based on the geographic location as
shown in Figure 1. The intersections were identified using their locations in the North-
East (NE), North-West (NW), South East (SE) and South-West (SW) quadrants together
with a reference number (e.g. NE5: James Street with Hume Street). Intersection crashes
were defined as the number of crashes that occurred at the intersection area and within
20 m measured towards upstream from the stop line as shown in Figure 2.
For the scope of this analysis, six years (2008–2013) of crash data was extracted for
modelling purposes. The subsequent two years (2014–2015) of crash data were then
used for model validation. The crash data consisted of information about the crash:
day, time, location, severity level, traffic control type, and speed limit. Road geometric
data were collected from site visits, historical design records, and Google Earth. Traffic
volume data for the selected intersections were obtained from the road authorities of
this jurisdiction namely Toowoomba Regional Council and Department of Transport
and Main Roads, Queensland.
In order to propose effective crash reduction measures, it is important to understand
the main factors that contribute to the occurrence of crashes. Seventeen variables were
identified in this study as the most common factors associated with the road crashes. A
statistical summary of the independent variables considered in the safety models develop-
ment and the manner in which they are defined in the dataset is shown in Table 1.

Figure 1. Toowoomba road network.


10 M. N. AL-MARAFI ET AL.

Figure 2. A typical intersection representing key contributory factors.

From a practical point of view, the speed limits need to be categorized into count or
categorical variables rather than continuous variables. However, many studies used
speed limit as a continuous variable in order to show the effect of every slight change
in speed limits on road safety (Abdel-Aty & Radwan, 2000; Gargoum & El-Basyouny,

Table 1. Statistical summary of dataset.


Variable N Minimum Maximum Mean Std. deviation SPSS labelling Variable type
No. of legs 106 3 4 3.67 0.473 X1 Count
No. of through lanes-entering
Major-approach 106 2 5 3.46 0.886 X2 Count
Minor-approach 106 0 4 2.40 1.478 X3 Count
No. of through lanes-exiting
Major-approach 106 2 5 3.20 0.960 X4 Count
Minor-approach 106 1 4 2.03 0.980 X5 Count
Traffic control type 106 0 1 0.58 0.495 X6 Categorical
No. of left turn lanes
Major-approach 106 0 2 0.12 0.407 X7 Count
Minor-approach 106 0 2 0.16 0.417 X8 Count
No. of right turn lane
Major-approach 106 0 2 0.75 0.906 X9 Count
Minor-approach 106 0 2 0.47 0.783 X10 Count
AADT
Major-approach 106 4,500 21,784 12,546 4,630 Qmajor Continuous
Minor-approach 106 1,600 14,837 5,769 3,199 Qminor Continuous
No. of slip lanes
Major-approach 106 0 2 0.29 0.617 X11 Count
Minor-approach 106 0 2 0.19 0.537 X12 Count
Presence of median island
Major-approach 106 0 1 0.46 0.501 X13 Categorical
Minor-approach 106 0 1 0.28 0.453 X14 Categorical
Speed limit (km/h)Major 106 40 60 59.06 3.787 X15 Continuous
Note: AADT, Annual Average Daily Traffic.
INTERNATIONAL JOURNAL OF URBAN SCIENCES 11

2016; Kim & Lee, 2013; Park & Abdel-Aty, 2016). Also, it should be noted that the inter-
section data were analyzed as one group rather than separating the data into two groups
(i.e. signalized or un-signalized intersection) because one of the strategies would be chan-
ging the traffic control at the intersections, and it was considered best to use the data as one
group (Chen, Cao, & Logan, 2012; Gomes, Geedipally, & Lord, 2012).

5. Results
5.1. Model development
The Pearson correlation matrix is a good tool for determining which explanatory variables
need to be considered for modelling because it can be used to measure the strength of
linear dependence between the individual independent variables (Turner, Singh, &
Nates, 2012). The Pearson correlation between independent variables in prediction
models would be accepted when values are between −0.49 and +0.49 and the variable par-
ameter is considered to be statistically significant at 0.1 significance level (using 90% confi-
dence). Based on these criteria four models were finally identified and developed using
SPSS software version 23 (IBM Corp, 2015) as shown in Table 2.
Firstly, the NB distribution was used, and this assumption was tested using the value of
Chi-square (x2)/degree of freedom (df) and the value of variance/degree of freedom (df).

Table 2. Negative binomial parameter estimates for selected models.


Model I Model II Model III Model IV
Variable β Р-valueb β Р-valueb β Р-valueb β Р-valueb
Intercept −9.251 0.000 −1.536 0.000 −4.094 0.013 −1.300 0.006
No. of legs (X1) 0.622 0.000 – – – – – –
No. of through lanes-entering
Major-approach (X2) – – – – – – 0.398 0.000
Minor-approach (X3) – – – – 0.116 0.028 – –
No. of through lanes-exiting
Major-approach (X4) – – 0.448 0.000 0.146 0.006 – –
Minor-approach (X5) – – 0.166 0.002 – – – –
Traffic controlc (X6) – – – – – – −0.136 0.588
No. of left turn lane
Major-approach (X7) 0.056 0.091 0.298 0.041 – – 0.472 0.031
Minor-approach (X8) – – – – −0.075 0.000 – –
No. of right turn lane
Major-approach (X9) −0.034 0.005 – – – – – –
Minor-approach (X10) – – – – −0.067 0.473 0.231 0.124
Ln(AADT)
Major-approach (Qmajor) 0.283 0.144 – – – – – –
Minor-approach (Qmajor) 0.281 0.098 – – 0.430 0.023 – –
No. of slip lanes
Major-approach (X11) – – −0.068 0.707 – – – –
Minor-approach (X12) 0.316 0.000 – – 0.247 0.000 0.021 0.000
Median islandd
Major-approach (X13) – – −0.560 0.004 −0.154 0.270 −0.597 0.013
Minor-approach (X14) −0.329 0.016 – – – – 0.392 0.149
Speed Limit (km/h)Major (X15) 0.038 0.000 – – – – –
Dispersion (k) 0.210a 0.102a 0.330a 0.271a
a
Computed based on the Pearson Chi-square.
b
Significance at 0.1 level.
c
Traffic control = 1 if signalized; traffic control = 0 if un-signalized.
d
Median island = 1 if present; median island = 0 if not present.
12 M. N. AL-MARAFI ET AL.

These values indicate whether the NB distribution assumption is accepted or not. The NB
distribution was used, and this assumption was tested using the value of Chi-square (x2)/
degree of freedom (df) and the value of variance/degree of freedom (df). These values indi-
cate whether the NB distribution assumption is accepted or not. If the result of these tests
lies between 0.8 and 1.2, the negative binomial model is appropriate, and the model fits the
data well (Abdul Manan et al., 2013; Bauer & Harwood, 2000; Maina, 2009). In the case of
the assumption not being accepted then the Poisson distribution would be used (Abdul
Manan et al., 2013; Ackaah & Salifu, 2011).
Using the goodness-of-fit tests, Akaike’s Information Criterion (AIC) and the Bayesian
Information Criterion (BIC), the models were ranked starting with the best model as following:
Model I, Model II, Model III, and Model IV with AIC and BIC values as shown in Table 3.
Table 4 shows the four models selected as suitable models based on statistical signifi-
cance, goodness-of-fit, and Pearson correlation value.
After developing all CPMs, the prediction ability of each model was tested using the
four performance measures discussed earlier: Mean Squared Prediction Error (MSPE),
Mean Absolute Deviation (MAD), Mean Squared Error (MSE), and Freeman-Tukey R-
Squared coefficient (R 2FT). Table 5 shows the performance for all crash prediction
models based on the estimation dataset (2008–2013) and the validation dataset (2014–
2015). It can be seen that the values of MSPE using the validation dataset and MSE
using the estimation dataset are close to each other. In addition, the values of MAD
using both datasets are similar. The R 2FT test results were slightly different for the esti-
mation datasets compared to the validation datasets. The overall results indicate that

Table 3. Goodness of fit tests for selected models.


Model Parameter Value df Value/df
I Deviance 81.126 96 0.845
Pearson Chi-Square 79.470 0.825
Akaike’s Info. Criterion (AIC) 254.166
Bayesian Info. Criterion (BIC) 280.801
II Deviance 103.509 100 1.035
Pearson Chi-Square 94.263 0.943
Akaike’s Info. Criterion (AIC) 287.110
Bayesian Info. Criterion (BIC) 303.090
III Deviance 91.564 99 0.925
Pearson Chi-Square 80.063 0.809
Akaike’s Info. Criterion (AIC) 294.754
Bayesian Info. Criterion (BIC) 313.398
IV Deviance 92.836 98 0.947
Pearson Chi-Square 79.329 0.809
Akaike’s Info. Criterion (AIC) 295.419
Bayesian Info. Criterion (BIC) 316.727

Table 4. Summary of the selected models to estimate road crashes.


Model no. Model form
(−9.251 + 0.622X1 +0.056X7 −0.034X9 + 0.316X12 − 0.329X14 + 0.038X15 )
I Npre, i = major × Qminor × exp
Q0.283 0.281
(−1.536 + 0.448X2 +0.116X5 + 0.298X7 − 0.068X11 − 0.560X13 )
II Npre, i = exp
(−4.094 + 0.116X3 +0.146X4 −0.075X8 −0.067X10 +0.247X12 −0.154X13 )
III Npre, i = Q0.430
minor × exp
IV Npre, i = exp(−1.300 + 0.398X2 +0.136X6 +0.472X7 +0.231X10 +0.021X12 −0.597X13 +0.392X14 )
Npre,i = predicted crash frequency at i th intersection.
Xj = explanatory variable j, described in Table 1.
INTERNATIONAL JOURNAL OF URBAN SCIENCES 13

Table 5. Performance measures for all crash prediction models.


Model I Model II Model III Model IV
Performance 2008– 2014– 2008– 2014– 2008– 2014– 2008– 2014–
measures 2013a 2015b 2013a 2015b 2013a 2015b 2013a 2015b
MSPE – 0.527 – 1.109 – 0.624 – 1.262
MSE 0.691 – 1.289 – 0.790 – 1.425 –
MAD 0.569 0.516 0.763 0.781 0.585 0.523 0.768 0.785
R 2FT % 49.0 45.7 45.2 41.4 41.9 35.0 18.1 22.7
a
Calculated based on estimation dataset 2008–2013.
b
Calculated based on validation dataset 2014–2015.

the four selected models have demonstrated the ability to estimate the road crashes
reasonably over additional years.
Overall, based on the outcome from the goodness-of-fit measures described previously,
all models can be accepted for further analysis (e.g. estimated CMFs). Model I as the best-
fitted model was subsequently used to calculate the expected road crash frequency.

5.2. Identifying hazardous intersections


The Highway Safety Manual (HSM) (AASHTO, 2010) indicates that the advantage of
using the CPMs is that the user will obtain a value for a long-term predicted crashes
number rather than short-term observed crashes number. Specifically, the expected
number of crashes using EB adjustments was used in this study to increase the accuracy
of safety estimation by accounting for the Regression-To-the-Mean (RTM) bias usually
associated with the road crash data.
Thus, in this section, the EB approach was employed together with Model I (Table 2) to
identify the hazardous intersections. In the first step, Model I was used to estimate the pre-
dicted number of crashes for each intersection. In the second step, the weighting adjust-
ment (ω) was calculated using the over-dispersion parameter (k) and the predicted
number of crashes. In the third step, the number of crashes was estimated by combining
the predicted number of crashes from Model I with the actual number of crashes.
Equations (11) and (12) were used to calculate the weighting adjustment (ω) and the
expected crash numbers (AASHTO, 2010):
1
vi = T (11)
1+k× t=1 Npre, i

Nexp , i = vi × Npre, i + (1 − vi ) × Nobs, i (12)


where Nexp,i – the expected crash frequency at i th intersection; ωi – the weighting adjust-
ment to model prediction; Npre,i – the predicted crash frequency at i th intersection in a
time period t; Nobs,i – the observed crash frequency; k – over-dispersion parameter
from the prediction model.
Finally, the Potential for Safety Improvements (PSI) was calculated for use in ranking
the intersections. The values of PSI were calculated as the difference between the expected
and predicted number of crashes. Based on the PSI values the intersections were ranked,
starting from the most hazardous intersection. Table 6 presents the top five hazardous
intersections in the study area from a total of 106 intersections.
14 M. N. AL-MARAFI ET AL.

Table 6. Ranking intersections for safety improvement.


Intersection Observed Predicted EB weighted adjustment Expected
ID mean (crash/year) (w) (crash/year) PSI Rank
NW9 6.67 2.92 0.21 5.86 2.948 1
NE5 4.83 2.38 0.25 4.22 1.840 2
SE12 3.67 2.10 0.27 3.24 1.136 3
NW15 3.33 1.65 0.32 2.79 1.135 4
NE6 3.00 1.16 0.41 2.25 1.093 5
Note: PSI, Potential for safety improvements.

The locations of these hazardous intersections were also shown in Figure 1. The most
dangerous intersection within the area analyzed was NW9 located between Bridge Street
and Tor Street.

5.3. Estimation of CMFs


The CMF is a value representing the change in road safety after modifying the geometric
design or operation of the facility. In this research, the CMFunction used to estimate the
road safety effect for changing any independent variable at the intersections is given by
Equation (13) (Abdel-Aty et al., 2014; Lord & Bonneson, 2007). After applying this
method to the variables associated with the types of treatment, Standard Error (SE) for
each treatment was calculated using Equation (14) (Bahar, 2010):
CMFi = ebi [Xi −Xib ] (13)

(ebi [Xi −Xib ]+SEbi − ebj [Xi −Xib ]−SEbi )


SEi = (14)
2
Where Xi – observed value for the variable i; Xib – base condition value for the variable i; βi
– model parameters for the variable i; SEβi – standard error of the model parameter βi.
The outcomes of this study are only applicable to those intersections with similar geo-
metric and traffic conditions, i.e. within the range of the datasets used as per the details
given in Table 1.
It is important to consider a base value for using developed CPMs to estimate crashes to
reflect conditions after a treatment. For example, pooled data from 106 intersections
shows that the number of slip lanes to minor approach ranged from 0 to 2 with an
average of 0.19. Therefore, no slip lane condition was assumed as the base value for deter-
mining the CMF from introducing slip lane to the minor approach. Similarly, the base
values for estimating CMFs from various treatments are shown in Table 7. However,
the base conditions for each individual intersection may take different values to accommo-
date specific site conditions, therefore, they need to be adjusted to accommodate the actual
site condition.
When the value of SE equals 0.1 or less, this indicates that an estimated CMF is more
accurate. The geometric features of hazardous intersections and recent operational con-
ditions were incorporated to determine the possible treatments for each intersection.
Initially, CMFs were estimated for single treatments using the previously developed
models (Models I–IV). Table 7 shows the suggested treatments for the top five hazardous
intersections, where CMFs were estimated for a single suggested treatment. The most
INTERNATIONAL JOURNAL OF URBAN SCIENCES 15

Table 7. Estimated CMFs from single treatment and identified intersections.


Suitable for top 5
Proposed treatments Labelling Base value CMF SE intersections
Reduce posted speed on major approaches from V60-50 60 kph 0.68 0.015 NW9, NE5, SE12,
60 to 50 km/h NW15, NE6
Add a median island on minor approaches AMminors 0 (No median) 0.72a 0.099 NW9, NE5, SE12,
NW15, NE6
Add a median island on major approaches AMmajors 0 (No median) 0.58 a 0.132 NE5, SE12, NW15, NE6
Add one left turn lane on one minor approach A1LT1minor 0 (No left lane) 0.96 0.073 NW9
Add one left turn lane on minor approaches A1LTminors 0 (No left lane) 0.92a 0.138 NE5, NW15
Reduce one entering through lane on minor R1TLminors 2-lane/approach 0.88a 0.047 NE5
approaches (i.e. from two to one)
Add one slip lane to one major approach A1SL1major 0 (No slip lane) 0.97 0.050 NE5
Add one slip lane to all major approaches A1SLmajors 0 (No slip lane) 0.94a 0.097 SE12, NW15
Introduce signalization Signal 0 (unsignalized) 0.87 0.221 SE12
a
CMF value was estimated for both road approaches i.e. in two directions.

effective single treatment for intersections NE5, SE12, NW15, and NE6 was to add a raised
median island on the major road for both directions, and the estimated CMF was 0.58. For
the intersection NW9 the optimum was to reduce the posted speed limit on both major
approaches from 60 to 50 km/h, with an estimated CMF of 0.68.
The next step undertaken was to analyze the CMFs for combined treatments using
various techniques. In this section, the CMFs for combined treatments were estimated
using the four existing techniques as discussed earlier, i.e. the HSM technique (technique
1), the Turner technique (technique 2), the systematic reduction of subsequent CMFs
technique (technique 3), and applying only the most effective CMF technique (technique
4). As results shown in Table 8, CMFs for treatments were ranked starting with the most

Table 8. Estimated CMFs for multiple treatments at intersections.


Combined CMFs
Intersections Technique Technique Technique Technique
ID Suggested treatment 1a 2b 3c 4d Average
NW9 V60-50 + AMminors 0.49 0.66 0.54 0.68 0.59
V60-50 + AMminors + A1LT1minor 0.47 0.65 0.53 0.68 0.58
NE5 AMmajors + V60-50 0.39 0.59 0.42 0.58 0.50
AMmajors + V60-50 + AMminors 0.28 0.52 0.33 0.58 0.43
AMmajors + V60-50 + AMminors + RTLminors 0.25 0.5 0.30 0.58 0.41
AMmajors + V60- 0.23 0.49 0.28 0.58 0.39
50 + AMminors + RTLminors + A1LT1minor
AMmajors + V60-50 + AMminors + RTLminors + 0.22 0.48 0.27 0.58 0.39
A1LT1minor + A1SL1major
SE12 AMmajors + V60-50 0.39 0.59 0.42 0.58 0.50
AMmajors + V60-50 + AMminors 0.28 0.52 0.33 0.58 0.43
AMmajors + V60-50 + AMminors + Signal 0.24 0.49 0.29 0.58 0.40
AMmajors + V60-50 + AMminors + Signal + 0.23 0.49 0.28 0.58 0.39
A1SL1major
NW15 AMmajors + V60-50 0.39 0.59 0.42 0.58 0.50
AMmajors + V60-50 + AMminors 0.28 0.52 0.33 0.58 0.43
AMmajors + V60-50 + AMminors + A1LT1minor 0.26 0.51 0.31 0.58 0.41
AMmajors + V60-50 + AMminors + A1LT1minor + 0.24 0.49 0.29 0.58 0.40
A1SL1majors
NE6 AMmajor + V60-50 0.39 0.59 0.42 0.58 0.50
AMmajor + V60-50 + AMminors 0.28 0.52 0.33 0.58 0.38
a
Highway Safety Manual (HSM) technique.
b
Turner technique.
c
Systematic reduction of subsequent CMFs technique.
d
Apply only the most effective CMF technique.
16 M. N. AL-MARAFI ET AL.

effective single treatment and they were later combined to estimate the collective treat-
ments. It can also be seen that the combined CMFs results from the four existing tech-
niques differ from each other. Therefore, to estimate combined CMFs in this study the
average of these four techniques (adjustment approaches) was adopted.
The results also indicated that the effect of treatments on road safety does not depend
on the number of treatments that have been applied but rather depend on the quality and
the suitability of these treatments relative to the intersection’s operating environment. The
final treatment plans for the five hazardous intersections together with the expected crash
reduction are summarized below.

. Three treatments were proposed for intersection NW9. These were reducing the posted
speed limit on major approaches from 60 to 50 km/h, adding a median island on minor
approaches, and adding one left-turn lane on one minor approach. The estimated road
crash reduction after applying the suggested treatments was 42%.
. Six treatments were proposed for intersection NE5. These were adding a median island
on major approaches, reducing the posted speed limit on the major approaches from 60
to 50 km/h, adding a median island on minor approaches, reducing one entering
through lane on minor approaches, adding one left turn lane on one minor approach,
and adding one slip lane to one major approach. The estimated road crash reduction
after applying the suggested treatments was 61%.
. Five treatments were suggested for intersection SE12. These were adding a median
island on major approaches, reducing posted speed limit on the major approaches
from 60 to 50 km/h, adding a median island on minor approaches, introducing signa-
lization, and adding one slip lane to one major approach. The estimated road crash
reduction after applying the suggested treatments was 61%.
. Five treatments were suggested for intersection NW15. These were adding a median
island on the major approaches, reducing the posted speed limit on the major
approaches from 60 to 50 km/h, adding a median island on the minor approaches,
adding one left turn lane on one minor approach, and adding one slip lane to one
major approach. The estimated road crash reduction after applying the suggested treat-
ments was 60%.
. Three treatments were suggested for intersection NE6. These were adding a median
island on the major approaches, reducing the posted speed limit on the major
approaches from 60 to 50 km/h, and adding a median island on minor approaches.
The estimated road crash reduction after applying the suggested treatments was 62%.

6. Conclusions
Practitioners are interested in estimating the expected outcomes from a set of multiple
safety treatments at intersections, as they will provide the information required to make
a comparison between the costs of crash reduction and the costs of treatment. Importantly
the information also allows prioritization of safety improvement projects, which should
provide benefits to the country. This paper outlines how the models can be developed
and used to identify dangerous intersections, then leading to introducing ways of estimat-
ing combined crash modification factors for various treatment plans.
INTERNATIONAL JOURNAL OF URBAN SCIENCES 17

Initially, the NB approach was employed to develop the crash prediction models
based on the data collected from 106 intersections in Toowoomba, Australia. The
most hazardous intersections were identified with the help of the best prediction
model and the EB approach. These models were then used to estimate the combined
CMFs for multiple treatments. The combined CMFs were estimated using four existing
techniques including HSM, Turner, Systematic reduction of subsequent CMFs, and the
most effective CMF technique. It was found that there were variations in the estimation
of combined CMFs using the applied techniques. The average values were adopted as the
simplest approach to estimate the effect of the proposed combined treatments. The
results demonstrated that multiple treatments have higher safety effects than single
treatments.
The developed crash prediction models were also used to estimate CMFs to identify the
effect of the various proposed treatments at the intersections. Through the seventeen vari-
ables used to develop the models and subsequent analysis, it was found that there were
seven variables that would best support the increase of road safety at intersections.
These were presence of signals, separate left-turning lane on minor approach, separate
right-turning lane on major and minor approaches, median island on major and minor
approaches, and slip lanes on major approach. The following ten variables were also ident-
ified that have negative effects on road safety such as: increased legs at intersection,
through lanes on major and minor approaches, through lanes at exits on major and
minor roads, left-turning lane on major approach, traffic volume on major and minor
approaches, increased posted speed limits on major approaches and the presence of slip
lanes on minor approaches.
This study has identified four important aspects for future research. First, calibration of
the prediction models to verify the transferability to different regions with similar road
characteristics. Second, the importance of separating the safety effects of combined treat-
ments on various severity levels and crash types. Third, the inclusion of additional inter-
section geometric features and operational conditions for improving the modelling and
the quality of the analysis. Finally, there is a need to validate the estimation of benefits
from the use of combined CMFs for road safety treatments.

Acknowledgments
The first author gratefully acknowledges the financial support from a Tafila Technical University
scholarship for carrying out his PhD study at the University of Southern Queensland. The
authors acknowledge the data support by the Department of Transport and Main Roads and Too-
woomba Regional Council.

Disclosure statement
No potential conflict of interest was reported by the author(s).

References
AASHTO. (2010). Highway safety manual (1st ed.). Washington, DC: American Association of
State Highway and Transportation Officials.
18 M. N. AL-MARAFI ET AL.

Abdel-Aty, M., Lee, C., Park, J., Wang, J., Abuzwidah, M., & Al-Arifi, S. (2014). Validation and
application of highway safety manual (Part D) in Florida. Orlando, FL. Retrieved from Final
Report.
Abdel-Aty, M., & Radwan, E. (2000). Modeling traffic accident occurrence and involvement.
Accident Analysis & Prevention, 32(5), 633–642.
Abdul Manan, M. M., Jonsson, T., & Várhelyi, A. (2013). Development of a safety performance
function for motorcycle accident fatalities on Malaysian primary roads. Safety Science, 60, 13–20.
Ackaah, W., & Salifu, M. (2011). Crash prediction model for two-lane rural highways in the Ashanti
region of Ghana. IATSS Research, 35(1), 34–40.
Akaike, H. (1974). A new look at the statistical model identification. IEEE Transactions on
Automatic Control, 19(6), 716–723.
Bahar, G. (2010). Methodology for the development and inclusion of crash modification factors in the
first edition of the highway safety manual (Report No. Transportation Research Circular E-
C142). Washington, DC: T. R. Board.
Bauer, K. M., & Harwood, D. W. (2000). Statistical models of At-grade intersection accidents –
Addendum (Report No. FHWA-RD-99-094). Georgetown Pike.
BITRE. (2015). Road trauma Australia. Retrieved from http://bitre.gov.au/publications/ongoing/
road_deaths_Australia_annual_summaries.aspx
Bonneson, J. A., & Lord, D. (2005). Role and application of accident modification factors in the
highway design process (Report No. FHW A/TX-05/0-4703-2). Texas.
Bonneson, J. A., & Pratt, M. P. (2009). Roadway safety design workbook (Report No. FHWA/TX-09/
0-4703-P2). Texas.
Cafiso, S., Di Graziano, A., Di Silvestro, G., La Cava, G., & Persaud, B. (2010). Development of com-
prehensive accident models for two-lane rural highways using exposure, geometry, consistency
and context variables. Accident Analysis & Prevention, 42(4), 1072–1079.
Caliendo, C., Guida, M., & Parisi, A. (2007). A crash-prediction model for multilane roads. Accident
Analysis & Prevention, 39(4), 657–670.
Chen, H., Cao, L., & Logan, D. B. (2012). Analysis of risk factors affecting the severity of intersection
crashes by logistic regression. Traffic Injury Prevention, 13(3), 300–307.
Chin, H. C., & Quddus, M. A. (2003). Applying the random effect negative binomial model to
examine traffic accident occurrence at signalized intersections. Accident Analysis &
Prevention, 35(2), 253–259.
Corben, B. F., & Wai, F. C. (1990). Pro-active traffic engineering safety study, final report: Part 2-
right-turn-against crashes at traffic signals (Report No. 11).
Da Costa, S., Qu, X., & Parajuli, P. M. (2015). A crash severity-based black spot identification
model. Journal of Transportation Safety & Security, 7(3), 268–277.
Elvik, R. (2007). State-of-the-art approaches to road accident black spot management and safety
analysis of road networks (Report No. 883).
Elvik, R. (2008). Comparative analysis of techniques for identifying locations of hazardous roads.
Transportation Research Record, 2083, 72–75.
Elvik, R., Ulstein, H., Wifstad, K., Syrstad, R. S., Seeberg, A. R., Gulbrandsen, M. U., & Welde, M.
(2017). An Empirical Bayes before-after evaluation of road safety effects of a new motorway in
Norway. Accident Analysis & Prevention, 108, 285–296.
Elvik, R., Vadeby, A., Hels, T., & van Schagen, I. (2019). Updated estimates of the relationship
between speed and road safety at the aggregate and individual levels. Accident Analysis &
Prevention, 123, 114–122.
Fawcett, L., Thorpe, N., Matthews, J., & Kremer, K. (2017). A novel Bayesian hierarchical model for
road safety hotspot prediction. Accident Analysis & Prevention, 99, 262–271.
Gargoum, S. A., & El-Basyouny, K. (2016). Exploring the association between speed and safety: A
path analysis approach. Accident Analysis & Prevention, 93, 32–40.
Gomes, S. V., Geedipally, S. R., & Lord, D. (2012). Estimating the safety performance of urban inter-
sections in Lisbon, Portugal. Safety Science, 50(9), 1732–1739.
Greibe, P. (2003). Accident prediction models for urban roads. Accident Analysis & Prevention, 35
(2), 273–285.
INTERNATIONAL JOURNAL OF URBAN SCIENCES 19

Gross, F., & Hamidi, A. (2011). Investigation of existing and alternative methods for combining mul-
tiple CMFs (Report No. T, 06-013). FHWA.
Gross, F., Persaud, B., & Lyon, C. (2010). A guide to developing quality crash modification factors
(Report No. FHWA-SA-10-032). Washington, DC.
Hanley, K., Gibby, A., & Ferrara, T. (2000). Analysis of accident-reduction factors on California
state highways. Transportation Research Record, 1717, 37–45.
Hauer, E. (2013). Even perfect regressions may not tell the effect of interventions. Washington, DC.
Transportation Research Board 92nd Annual Meeting.
Hauer, E., Harwood, D., Council, F., & Griffith, M. (2002). Estimating safety by the empirical Bayes
method: A tutorial. Transportation Research Record, 1784, 126–131.
Hauer, E., Kononov, J., Allery, B., & Griffith, M. (2002). Screening the road network for sites with
promise. Transportation Research Record, 1784, 27–32.
Heydari, S., Miranda-Moreno, L., & Amador, L. (2013). Does prior specification matter in hotspot
identification and before-after road safety studies? Transportation Research Record: Journal of the
Transportation Research Board, 2392, 31–39.
Huang, H., Chin, H., & Haque, M. (2009). Empirical evaluation of alternative approaches in iden-
tifying crash hot spots: Naive ranking, empirical Bayes, and full Bayes methods. Transportation
Research Record, 2103, 32–41.
IBM Corp. (2015). IBM SPSS statistics for windows (version 23). Armonk, NY: IBM Corp.
Khan, M., Abdel-Rahim, A., & Williams, C. J. (2015). Potential crash reduction benefits of shoulder
rumble strips in two-lane rural highways. Accident Analysis & Prevention, 75, 35–42.
Kim, D.-G., & Lee, Y. (2013). Modelling crash frequencies at signalized intersections with a trun-
cated count data model. International Journal of Urban Sciences, 17(1), 85–94.
Kim, D.-G., Washington, S., & Oh, J. (2006). Modeling crash types: New insights into the effects of
covariates on crashes at rural intersections. Journal of Transportation Engineering, 132(4), 282–
292.
Lan, B., Persaud, B., Lyon, C., & Bhim, R. (2009). Validation of a full Bayes methodology for obser-
vational before–after road safety studies and application to evaluation of rural signal conversions.
Accident Analysis & Prevention, 41(3), 574–580.
Lee, C., Abdel-Aty, M., Park, J., & Wang, J.-H. (2015). Development of crash modification factors
for changing lane width on roadway segments using generalized nonlinear models. Accident
Analysis & Prevention, 76, 83–91.
Li, X., Lord, D., & Zhang, Y. (2010). Development of accident modification factors for rural frontage
road segments in Texas using generalized additive models. Journal of Transportation
Engineering, 137(1), 74–83.
Lord, D., & Bonneson, J. (2007). Development of accident modification factors for rural frontage
road segments in Texas. Transportation Research Record, 2023, 20–27.
Lord, D., & Mannering, F. (2010). The statistical analysis of crash-frequency data: A review and
assessment of methodological alternatives. Transportation Research Part A: Policy and
Practice, 44(5), 291–305.
Maina, E. (2009). Developing crash modification factors for operational parameters on urban
freeway. Newark, NJ: New Jersey Institute of Technology.
NCHRP. (2008). Crash reduction factors for traffic engineering and ITS improvements (Report No.
Final Report Project 17-25). Washington, DC.
Oh, J., & Park, D. (2014). Analysis on crash reduction factors for road segment safety. International
Journal of Urban Sciences, 18(3), 396–403.
Park, J., & Abdel-Aty, M. (2016). Evaluation of safety effectiveness of multiple cross sectional fea-
tures on urban arterials. Accident Analysis & Prevention, 92, 245–255.
Park, J., Abdel-Aty, M., & Lee, C. (2014). Exploration and comparison of crash modification factors
for multiple treatments on rural multilane roadways. Accident Analysis & Prevention, 70, 167–177.
Park, J., Abdel-Aty, M., Wang, J.-H., & Lee, C. (2015). Assessment of safety effects for widening
urban roadways in developing crash modification functions using nonlinearizing link functions.
Accident Analysis & Prevention, 79, 80–87.
Pearson, K. (1934). On a new method of determining goodness of fit. Biometrika, 26(4), 425–442.
20 M. N. AL-MARAFI ET AL.

Persaud, B., Lan, B., Lyon, C., & Bhim, R. (2010). Comparison of empirical Bayes and full Bayes
approaches for before–after road safety evaluations. Accident Analysis & Prevention, 42(1),
38–43.
Pitale, J. T., Shankwitz, C., Preston, H., & Barry, M. (2009). Benefit: Cost analysis of in-vehicle tech-
nologies and infrastructure modifications as a means to prevent crashes along curves and shoulders
(Report No. MN/RC 2009-39). Minneapolis.
Queensland Government. (2016). Crash data from Queensland roads. Retrieved from https://data.
qld.gov.au/dataset/crash-data-from-queensland-roads
Sacchi, E., Sayed, T., & El-Basyouny, K. (2014). Collision modification functions: Incorporating
changes over time. Accident Analysis & Prevention, 70, 46–54.
Schwarz, G. (1978). Estimating the dimension of a model. The Annals of Statistics, 6(2), 461–464.
Tegge, R. A., Jo, J.-H., & Ouyang, Y. (2010). Development and application of safety performance
functions for Illinois (Report No. FHWA-ICT-10-066).
Turner, B. (2011). Estimating the safety benefits when using multiple road engineering treatments.
A. R. R. Board. Retrieved from Road Safety Risk Reporter, 11.
Turner, S., Singh, R., & Nates, G. (2012). The next generation of rural road crash prediction models:
Final report.
Wang, J.-H., Abdel-Aty, M., & Wang, L. (2017). Examination of the reliability of the crash modifi-
cation factors using empirical Bayes method with resampling technique. Accident Analysis &
Prevention, 104, 96–105.
Washington, S. P., Persaud, B. N., Lyon, C., & Oh, J. (2005). Validation of accident models for inter-
sections (Report No. FHWA-RD-03-037). Washington, DC.
WHO. (2015). Global status report on road safety 2015. Retrieved from http://www.who.int/
violence_injury_prevention/road_safety_status/2015/en/
Wood, J. S., Donnell, E. T., & Porter, R. J. (2015). Comparison of safety effect estimates obtained
from empirical Bayes before–after study, propensity scores-potential outcomes framework,
and regression model with cross-sectional data. Accident Analysis & Prevention, 75, 144–154.
Wu, L., & Lord, D. (2016). Investigating the influence of dependence between variables on crash
modification factors developed using regression models. Transportation Research Board 95th
Annual Meeting, TRB, Washington, DC.
Young, J., & Park, P. Y. (2013). Benefits of small municipalities using jurisdiction-specific safety
performance functions rather than the highway safety manual’s calibrated or uncalibrated
safety performance functions. Canadian Journal of Civil Engineering, 40(6), 517–527. doi:10.
1139/cjce-2012-0501
Yu, R., Quddus, M., Wang, X., & Yang, K. (2018). Impact of data aggregation approaches on the
relationships between operating speed and traffic safety. Accident Analysis & Prevention, 120,
304–310.

View publication stats

You might also like