Content uploaded by Dilum Dissanayake
Author content
All content in this area was uploaded by Dilum Dissanayake on Oct 17, 2018
Content may be subject to copyright.
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 1 of 20
Analysing Clearance Time of Urban Traffic Accidents in
Abu Dhabi using Hazard-based Duration Modelling Method
Abdulla Mohammed Saeed Alkaabi, PhD Student (Corresponding Author)
School of Civil Engineering & Geosciences, Newcastle University
Newcastle, NE1 7RU, UK.
Tel: +44 (0)191 222 6323
Fax: +44 (0)191 222 6502
E-mail: a.m.s.k.al-kaabi@ncl.ac.uk
Dilum Dissanayake, Ph.D.
Lecturer in Transport Modelling
Transport Operations Research Group
School of Civil Engineering & Geosciences, Newcastle University
Newcastle, NE1 7RU, UK.
Tel: +44 (0)191 222 5718
Fax: +44 (0)191 222 6502
E-mail: dilum.dissanayake@ncl.ac.uk
Mr Roger Bird
Lecturer in Highway Engineering
Transport Operations Research Group
School of Civil Engineering & Geosciences, Newcastle University
Newcastle, NE1 7RU, UK.
Tel: +44 (0)191 222 7681
Fax: +44 (0)191 222 6502
E-mail: r.n.bird@ncl.ac.uk
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 2 of 20
ABSTRACT:
Traffic incidents generate a lot of adverse impacts in many areas such as traffic flow, air
pollution, fuel consumption, and secondary crashes, and therefore it will be vital for traffic
incident responders and operators to know how they can improve the efficiency of traffic
incident management. This paper presents the results of investigating the effects of traffic
accident characteristics on the accident clearance time using fully parametric’ hazard-based
duration models with emphasize of Accelerated Failure Time (AFT) metric. Accident
characteristics and clearance times were obtained from Abu Dhabi, the capital of the United
Arab Emirates. The data was obtained from the Federal Traffic Statistics System and the
records of Abu Dhabi Collision Investigation Branch during the period from May, 2009 to
April 2010. For the purpose of this study, clearance time was defined as the length of time
between the arrivals of collision investigator on the accident scene until the collision
investigator leaves the scene. According to the goodness of fit test conducted in the study, the
Weibull model without gamma heterogeneity was utilized in this study. The estimation results
show that various accident characteristics were found to be significantly affecting clearance
time including month of year, location, weather condition, accident type, number of
causalities, and number of vehicles involved. These results highlighted some weakness points
in the current practices of clearing accident in Abu Dhabi. Accordingly, this paper suggested
some mitigation measures.
Keywords: Traffic Accident, Traffic Incident Management, Accident Clearance Time,
Hazard-based Duration Models, Accelerated Failure Time.
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 3 of 20
1. INTRODUCTION
Traffic incidents are regarded as one of the high priority problems to be tackled in many
countries. The reason is that such incidents may have potential to generate adverse effects
such as increasing the possibility of secondary accidents, increasing traffic congestion levels,
air pollution and fuel consumption, and reducing roadway capacity. Studies have shown that
secondary incidents account for about one quarter of all incidents and congestion. The figures
vary from place to place. For example, secondary crashes were found to be 20% of the all
incidents in the US [1] whereas they represent 20%-30% of the total traffic incidents in
Europe [2]. Also, incident related congestion was estimated as 25% of the total congestion in
the United States [3] whereas in the UK, traffic incidents account for 25% of the total traffic
congestion [4].
In response to the severe consequences of traffic incidents, great attention has been directed
to improve the effectiveness of Traffic Incident Management (TIM). TIM can be defined as
applying the available resources to reducing the impacts of traffic incidents and incident
duration [5]. As mentioned in this definition, reducing incident duration is one of the main
targets of TIM. Therefore improving the efficiency of the TIM process requires a clear
understanding of the factors affecting incident duration.
The definition of incident duration varies from one study to another. One of the earliest
studies measured incident duration as the time the incident remained on the travel lane [6].
Another study defined accident duration as the time between a police officer receiving a call
to respond to an accident until it is completely cleared [7]. This definition excludes reporting
time as a component of incident duration. However, most researchers defined incident
duration as the time difference between when the incident occurred and when it was
completely cleared [8-11].Accordingly, the total incident duration can be divided into several
phases or interval times. The Highway Capacity Manual [12], breaks down the total incident
duration into the following four phases:
1. Detection time: the time between the incident occurrence and incident reporting time.
2. Response time: the time between incident reporting time and the time of first
responder to arrive at the scene.
3. Clearance time: the time between the arrival of the first responder on the scene and
the moment when the incident has been cleared from the highway
4. Recovery time: the time taken for traffic flow to return to normal after the incident has
been cleared.
The aim of this paper is to investigate the effects of traffic accident characteristics on the
accident clearance time using fully parametric hazard-based duration models with emphasis
to Accelerated Failure-Time (AFT) metric. This metric implies that the external covariates
rescale the time scale. The study investigates data from Abu Dhabi, the capital of the United
Arab Emirates. The data set contains the characteristics and clearance times of accidents that
happened in the period from May, 2009 to April 2010. To achieve the aim of this study, two
databases were utilized to extract the required data including the Federal Traffic Statistics
System (FTSS) and the records of Abu Dhabi Collision Investigation Branch (ACIB).
This paper begins by reviewing previous research related to accident duration analysis. This
is followed by an illustration of the hazard based duration models concept and modelling
related issues. Attention is then directed to the case study and data description. Next, the
methodology approach is presented. Then, descriptive analyses of accident clearance time
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 4 of 20
and accident data are presented. After that, modelling results are discussed. Finally, the paper
presents the conclusions and recommendations.
2. LITERATURE REVIEW
This section reviews the previous research of modelling traffic incident duration. Then, the
basic concept of Hazard-Based Duration Models (HBDMs) will be presented followed by
modelling concerns.
2.1 Analysis of Accident Duration
In the last few decades, several methods have been used to investigate traffic incident
duration. One of the earlier studies analysed freeway accidents involving trucks in California,
USA, to find out the probable distribution of such accident duration [13]. Another research
developed a model using Linear Regression techniques to estimate incident duration based on
incident characteristics [14]. Furthermore, Time Sequential Models were used to predict
incident duration at the early stages of an incident for the purpose of supporting incident
management [15]. Also, a Nonparametric Regression technique applied to predict traffic
incident duration [16]. Another approach found in the literature is the Decision Tree Method.
The advantage of this method is that no assumption of probable distribution is required for
this model, which can find out patterns in a certain data set. Several studies have employed
the Decision Tree Method over the past few decades [17-20].
A recent research proposed an integrated approach of Discrete Choice Model and the Rule-
based Model to predict incident duration in Maryland [21]. Furthermore, Fuzzy Logic has
been shown to be a suitable approach for modelling transportation and traffic processes. In
terms of modelling incident duration, this method has the advantage of allowing the input of
linguistic or category variables [22]. An initial attempt to use this method to predict incident
duration was carried out in Los Angeles, USA [23]. Furthermore, Artificial Neural Network
Method utilized to study incident duration [8].
In addition to the previous methods, Hazard Based Duration Models (HBDMs) have been
popular in analysing traffic incident duration over the years. These models are based on an
important concept which is the conditional probability of a duration ending at some time,
given that the duration has continued for some specific time. One of the early studies applied
HBDMs to analyse freeway traffic accidents in Seattle, USA [7]. Another study applied
HBDMS to analyse highway incidents considering Washington State as a case study [11].
Furthermore, Cox Proportional Hazard model was developed to analyse the influence on the
Emergency Management Services (EMS) response and clearance times by the independent
variables in Ohio State major freeways[24]. Therefore various studies over the last two
decades have shown that Hazard Based Duration Modelling is an appropriate method for
analysing incident duration. It also has considerable potential for creating a predictive model
of incident duration. The main advantage that HBDMs have over other methods is that they
consider the relationship between the probabilities of duration will end soon and the length of
time that the incident has been lasted, not just the duration of the incident alone. In other
word, HBDMs capture the time dependency more appropriately. This is referred to as
duration dependence and is considered to be the key factor when analysing incident duration
[11].
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 5 of 20
It can be concluded from previous research that there are many methods could be used to
study incident duration. The variables used in the analysis can be classified into categories
including temporal characteristics, environmental characteristics, geographic characteristics,
incident characteristics, operational characteristics and driver characteristics. It is also worth
mentioning that direct comparison between previous studies is complicated by variations
between the definitions used, the aims and objectives, and the estimating, planning,
management and forecasting techniques used in each study. In addition, most of the
investigations that have been conducted to date have been carried out in developed countries,
particularly the USA. This study attempts to transfer the HBDM technique to conduct an in
depth investigation in to the analysis of accident clearance time in Arabian countries, using
Abu Dhabi city as a case study.
2.2 The Concept of Hazard-Based Duration Models (HBDMs)
Hazard-Based Duration Models (HBDMs) are sometimes referred to as Time-to-Event
modelling or survival modelling. As mentioned earlier, these models are based on an
important concept which is the conditional probability of a duration ending at some time,
given that the duration has continued for some specific time. This concept is an important one
because in many instances the probability of a time duration ending depends on the length of
time the duration has lasted [25]. This probability may increase, decrease, or remain constant.
For example, the probability of driver being involved in an accident may vary over time due
to the experience and skills gained over time. So, it appears that conditional probability is an
essential concept when studying duration data [26].
Developing HBDMs starts with the cumulative distribution function ,
(1)
Where donates probability, refers to a random time, and t is some specified time. This
function gives the probability of having an event before some transpired time .
The density function, which is a derivative value of the cumulative distribution with respect
to time, is:
(2)
The survival function presents the probability of the duration being greater than or equal to
some specified time t,
(3)
The hazard function can be written as:
(4)
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 6 of 20
Where represents the conditional probability that an event will occur between time
and t + dt given that the event has not occurred up to time t [26].
In order to investigate the effects of explanatory variables using HBDMs, two alternate
parametric approaches can be used including Proportional Hazard model (PH) and
Accelerated Failure Time models (AFT) [26, 27]. In this study, AFT was used because the
aim of this study is to investigate the effects of traffic accident characteristics on the accident
duration. AFT assumes that the external variables rescale (accelerate) the time scale. In other
words, this approach assumes that covariates act multiplicatively on the timescale [28, 29].
Thus, the model is written as follows:
(5)
Where: donates a specific value of clearance time, is a vector of covariates, is the vector
of the estimated coefficients, and is an error term. Furthermore, parametric assumption in
AFT needs to be made about the error term.
Some problems may appear when applying HBDMs such as unobserved heterogeneity. One
of the assumptions when applying HBDMs is that the survival distribution is homogenous
across all observations. This homogeneity will not exist if there are unobserved factors
affecting the duration and causing heterogeneity. As a result of that, this unobserved
heterogeneity may generate significant problems such as inconsistent estimation of
coefficients and standard errors leading to incorrect inference of hazard function shape, and
wrong estimation of covariate effects. To overcome this problem, a common approach is
introducing a new parameter in the model to capture unobserved heterogeneity and work with
conditional density function [26, 29]. These are referred to as a frailty model and it should
mention that different heterogeneity distributions can be used with the gamma distribution
being the most common.
3. CASE STUDY AND DATA DESCRIPTION
This study is based on the metropolitan network in the city of Abu Dhabi, the capital of the
United Arab Emirates (UAE). Roadways covered in this research have one of three
classifications including Primary Roads (Freeways, Expressways), Secondary Roads
(Arterials, Collectors), and Local Roads [30]. All the intersections are signalized in the
metropolitan network.
Although Abu Dhabi has a good road infrastructure, the number of road traffic accidents is
still growing, and is one of the main concerns for public health. According to the Abu Dhabi
Health Authority [31], road traffic fatalities comprise 68% of all injury-related deaths.
Another study compared road traffic accidents in the UAE to other countries in the region and
some western countries in 2002 [32]. It found that UAE had the highest fatality rate, for
example 21.6 fatalities per 100,000 population, with compared to UK, USA and Qatar which
have 5.7, 15.1, 14.7 fatalities per 100,000 population respectively. These facts show that more
effort and cooperation from the local authorities are required to reduce these accident rates
and to reduce the adverse impacts of road traffic accidents. Improving the efficiency of traffic
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 7 of 20
incident management will reduce some of the adverse effects, and may reduce road accidents
rates.
The operational process of accident management has many stages (FIGURE 1). The process
starts when the Police Operational Centre is notified of a traffic accident. Then a
comprehensive police patrol will be assigned to assess accident severity based on accident
type. If it is a property damage accident, a private company called SAAED will be contacted
to carry out the investigation. However, if the accident is found to be serious, injury or fatal
accident, Abu Dhabi Collision Investigation Branch (ACIB) will move to the scene to start a
comprehensive investigation. Upon the arrival at the scene, collision investigator needs to
accomplish several duties including preserving accident scene, preserving suspects if known,
securing evidences, drawing sketch map of final accident scene, recording witness
statements, and filling accident report. Additionally, other responders will be dispatched to
the accident scene including Ambulance and Rescue Service to move injuries, Traffic Control
Centre to watch out traffic flow and apply traffic diversion when necessary, Crime Scene
Department to deal with suspicious cases and fatal injuries, and Fire Department to clean the
debris [33].
Two databases were utilised to extract the data for this study. The first one is the Federal
Traffic Statistics System (FTSS), which covers all traffic accidents’ reports on the UAE. The
FTSS database has comprehensive accident related information such as temporal
characteristics (time of day, day of week, month of year), geographical characteristics (road
name, location), and accident characteristics (severity level, weather condition, injury details,
and vehicle details). The second database is the records of Abu Dhabi Collision Investigation
Branch (ACIB). These records contain the details of accident duration including reporting
time, response time, clearance time, and the total time. In this study, clearance time is
defined as following:
Clearance time: the length of time between the arrival of collision investigator on the
accident scene until the collision investigator leaves the scene.
4. METODOLOGY APPROACH
The methodology in this study comprises two stages: Data Collection and Data Analysis
(FIGURE 2). In the first stage, data regarding clearance time and accident related information
were collected and manipulated in one database. It is worth mentioning that duration data can
be treated as a continuous or discrete dependant variable [34]. In this study, clearance times
have been measured to the nearest minute and therefore associated with an accident on a
continuous variable.
The second stage, data analysis was conducted. This stage started by selecting which of the
explanatory variables were statistically significant for clearance time. In order to identify the
most relevant variables, three steps as below were conducted [35] :
1. The first step is to fit the models (Exponential, Weibull, Gompertz, Log-normal, Log-
logistic) without any external covariates, to check the value of the log likelihood before
convergence. This is referred to as a null or base model when all external variables are
considered to be zero.
2. The second step aims to identify which variable, on its own, significantly decreasing
this statistic. Thus, the value of minus the logarithm of the maximised likelihood (
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 8 of 20
statistic) was used to compare these models with the null model. In this step,
all variables that are significant at the level of 85% were selected.
3. The third step aims to check that none of the omitted variables are significant when
fitting the models with the significant variables resulted from the first step. Since any of
the excluded explanatory variables from the initial model could be significant when put
in back in the model, the models of the significant variables were fitted with one of the
omitted variables at a time. The same criterion of the first step was applied in this step
to add any of the omitted variables.
Then, the best distribution was fitted based on the value of Akaike Information Criterion
(AIC) test. Akaike “proposed penalizing each model’s log likelihood to reflect the number of
parameters being estimated and then comparing them” [36]. Also, in each model the
covariates were checked for any inconsistency (frailty model), for example accounting for
unobserved heterogeneity and the best fit distribution for model estimation was decided. The
criterion of selecting best distribution is based on the lowest value of AIC.
After that, the covariates effects on clearance time were modelled using AFT models. All
results from the model were interpreted using the sign of the estimated coefficient and the
percentage change in duration. Finally, the estimated results were interpreted in terms of the
covariate effect on clearance time and its’ reasonability to the current practices of clearing
accident in Abu Dhabi. All stages and steps were carried out using STATA Software.
5. DESCRIPTIVE ANALYSIS
5.1 Accident Data
In this paper the dataset collected over a one year period starting from May 2009 are used
accounting to a total of 583 accidents. As described above, the main database used to extract
accident data was FTSS database. This database has comprehensive information related to
accident temporal characteristics, geographical characteristics, environmental characteristics,
and accident characteristics. To achieve the aim of this study, much information was used as
candidate variables for the AFT model. For example, geographical information represents the
location of the accident in terms of the street, intersection, region, place, and road layout.
These variables have been included as may influence the travel time of responders and the
investigation time. Another example is accident type. This variable is included because each
kind of accidents may require different clearance equipments and hence, clearance time may
be affected. TABLE 1 shows the candidate variables used in this research.
Also, some data divided into more variables to understand the relationship between the
interval time and this data. For instance, time of the day was divided into 5 variables
including Morning, Afternoon, Evening, AM peak, and PM peak. Also, day of the week was
divided into 7 variables including Sunday, Monday, Tuesday, Wednesday, Thursday, Friday,
and Saturday. Furthermore, month of the year was divided based on the 12 months of the
year. This has been done because duration often tend to be vary by the traffic flow condition
which is obviously vary per all of these temporal variables.
5.2 Clearance Time Data
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 9 of 20
As previously stated, 583 accidents utilized in this study. The mean duration of clearance
time for these accidents was 26.14 minutes; with a minimum of 1 minute and a maximum of
130 minutes. The standard deviation is 18.18 minutes. This finding is quiet small comparing
to the previous research. For example, the mean duration of clearance time found to be 136
minutes in Washington State [11], and 78 minutes in Ohio State [24]. A possible explanation
for this variation is that in most of the developed countries, collision investigator is required
by the country accident investigation manual to gather extensive details of fatal and serious
injuries accidents using some survey instruments such as Total Station [37]. This requirement
does not apply in Abu Dhabi where the main focus of collision investigation is to fill Traffic
Accident Report and gather some measurements manually.
In addition to that, the density distributions of the total duration and each interval time are
skewed to the right due to the differences between the mean value and the median value of
each time (FIGURE 3).
6. MODELLING RESULTS
6.1 Model Selection
TABLE 2 summarises the results of Akaike Information Criterion AIC test. This test can be
written as the following:
AIC = -2In L + 2 (k + c) (6)
Where In L refers to the models log likelihood at convergence, k denotes the number of
covariates in the model and c is the number of distribution parameters. The criteria is to select
the distribution that has the lowest value of AIC [36, 38].
The results show that Weibull distribution without frailty model is the best fit distribution for
the clearance time. Thus, the following section will explain in details the process of data
analysis for Weibull distribution.
6.2 The Weibull AFT Model
Developing Weibull AFT model was based on the 26 statistical significant explanatory
variables, which were resulted from variables selection steps. TABLE 3 summarises the
results of the fitted Weibull AFT model on accident clearance time. In the AFT model, the
sign of the coefficient specifies how the variable affects the clearance time duration. For
example, a positive coefficient means that the variable increases the clearance time duration
[39].
The percentage change in clearance time by each of the explanatory variables can also be
calculated. This could be done by taking the exponent of the estimated coefficient of the
significant variable [11, 40]. For example, the exponential of a positive coefficient like
number of injuries per accident variable is 1.046. This means that the clearance time is about
4.6% longer when the number of the injuries increases by one. Conversely, the exponential of
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 10 of 20
a negative coefficient, such as Hit pedestrian is 0.873. In this case the interpretation is that the
clearance time for hit pedestrian accidents is about 12.7% shorter than other types of
accidents. Generally, when the exponent coefficient is greater than 1.0, it means that the
explanatory variable adds more time to the accident clearance time and vice versa. All the
percentage changes in clearance are presented in TABLE 3.
6.3 Results Discussion
This section presents the discussion of clearance time model results based on each
characteristics category of accident data. Then, a comparison between the findings and
previous research will be presented.
Temporal Characteristics
Accidents occurred out of peak hours were associated with longer clearance time. This result
was unexpected because normally the traffic condition is heavy during peak hours comparing
to out of peak hours. So, it was expected that less traffic may resulted in lower clearance
time. This result raised an important question which is what are the current practices of
accident clearance during the peak periods that make clearance time appears to be shorter
than out of peak period? The answer for this question was found after examining the current
procedure of accident management and discussing this result with collision investigators. The
reason for this result found to be that there is a common perception among collision
investigators to clear accident scene as soon as possible during the peak period to avoid more
traffic congestion. Unfortunately, this is a weakness point in accident management procedure
that may affect collision investigators work to gather enough evidences from the scenes that
can help them later to explore the main causes of the accident and how it occurred. This
finding agree with the previous finding of the attitude among accident management personnel
in Washington State which was classifying accidents occurred out of rush hours an less
important accidents [7].
Additionally, Four months of the year were found to be significantly affect clearance time.
Two of these months (August, October) were associated with shorter clearance time, whereas
the other two months (January, March) found to be associated with longer clearance time.
This is probably due to the occurrence of fatal injuries in all months of the year with except
of August and October. In general, when fatal injury occurred, Crime Scene personnel will be
involved to accomplish the forensic work. This will make collision investigator’s work much
longer and hence longer clearance time.
Geographical Characteristics
Location variables namely 23rd Street, 10th Street, Hamdan Bin Mohammed Street, and 31st
Street were associated with shorter clearance time. On the other hand, accidents occurred in
Meena Street and 35th Street was associated with longer clearance time. These findings
should be carefully interpreted due to the lower number of accidents occurred in some of
these streets (23ed Street, 31st Street, 35th Street). However, after carefully examining these
accidents data, it could interpret these findings as probably attributed to the high number of
injuries and occurrence of fatal injuries among the streets that associated with longer
clearance time comparing to the other streets. This is logically accepted because in practice
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 11 of 20
clearing accidents that involved high number of injuries or fatal injury will require more time.
These findings are similar to that resulted from previous work [7, 11, 40, 41].
Other location variables that have been found significantly affecting clearance time are
related to the nature of accident location. Accidents that occurred on school area had longer
clearance time and those that occurred on Commercial area, Car park, and Government
authority area had shorter clearance time. It is difficult to interpret these findings, but it could
be related to the traffic condition on these places and the perception of speeding the clearance
time in congested area. In Abu Dhabi, school working hours are between 8am to 2pm
comparing to the local authorities and commercial area which usually work up until 8pm and
10pm respectively. Accordingly school area is less congested after 2pm, whereas the
congestion last longer in the other areas. As a result of that, the occurrence of 50% of
accidents on school area after 2pm can be a possible explanation to longer clearance time in
this area comparing to others.
Environmental Characteristics
Weather conditions found to be significantly affecting the clearance time. Accidents occurred
in windy, clear, and rain conditions were associated with longer clearance time. It was
expected that windy and rain weather conditions may cause some delay in clearing accident
scene, but not the clear condition. However, longer clearance time associated with clear
condition can be attributed to the occurrence of over 95% of the accidents during clear
conditions, where all fatality accidents and accidents involved over 10 casualties are
occurred. Thus, the severity of the accidents happened on clear weather condition could be
the reason for longer clearance time.
Accident Characteristics
As expected, longer accident clearance time observed when there was an increase in the
number of casualties and the number of vehicles involved. Furthermore, “Hit pedestrian”
accidents had shorter clearance time and those with “Hit object” had longer clearance time.
This finding was unexpected because “Hit pedestrian” accidents resulted in 21 fatal injuries
whereas those related to “Hit object” had 1 fatal injury. So, longer clearance time was
expected to be associated with “Hit pedestrian” accidents. However, similarly to the
unexpected result of out of peak variable, the reason found to be related to a weakness point
in accident management procedures. The reason, which resulted from a discussion with
collision investigator in Abu Dhabi, was that collision investigation procedure is varying
among collision investigators in a way that some investigators will spend much time to gather
all the required information and witness statement from the accident scene whereas other
investigators collected the basic critical information in the accident report and left other
information and witness statement to be taken after clearing the scene. Thus, the time spent to
write down the non basic information and witnesses’ statements by the later investigators will
not be considered as a part of the accident clearance time. This fact highlights the absence of
clear guidelines of collision investigation and the need to a standard procedure that should be
followed by all collision investigators.
Comparison to Previous Research
Before presenting the differences between this research and previous research, it is worth
mentioning that it is not an appropriate approach to compare the results of this project with
those research modelled incident duration as one piece of time. This is mainly because an
empirical research [11] has shown that statistically significant variables were not stable, in
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 12 of 20
terms of type and effects, between incident duration interval times (reporting time, response
time, clearance time). As a result of that, the comparison in this section is limited to the
previous research that aimed to model each interval time of the total incident duration
separately [11, 24].
This research shows some differences in terms of the best fit distribution and the resulted
significant explanatory variables. Similarly to [11] and in contrast to [24], environmental
characteristics found to be significantly affect clearance time. However, some of these
variables, for instance clear and windy weather conditions, which are found to be significant
in this study were not significant in the previous research [11, 24]. Also, day of week variable
has been statistically significant in previous research. However this variable it is not fount to
be significant in this study.
On the other hand, the resulted best fit distribution for the clearance time is varying between
this research and previous research. For example, the Log-logistic distribution without
heterogeneity provided the best fit distribution in [11] study, however in this research;
Weibull distribution without heterogeneity found to be the best. Also, it is not possible to
compare the fitted distribution of this study with [24] study because the later study applied
Cox Proportional Hazard Model that does not require distribution assumption. Finally, having
stated these differences in the results, it can be clear that different datasets and case study
areas may yield different results.
7. CONCLUSIONS AND RECOMMENDATIONS
This paper presented the application of AFT metric hazard-based duration models to
investigate the factors affecting accident clearance time on Abu Dhabi metropolitan network.
The model estimation results show that different kinds of factors significantly affecting
clearance time including month of year, location, weather condition, accident type, number of
injuries, and number of vehicles involved. Also, these findings indicated the drawbacks of the
current accident management procedure in Abu Dhabi, particularly the perception of speeding
clearance time during congested area and times, and the diversity of collision investigation
procedure among the personnel. Mitigating them could be done by several measures such as
publishing a standard collision investigation guideline and a strategic clearance time targets
based on the accident severity. These measurements will ensure the fulfilment of high level of
investigation quality per accident which will lead to identifying the main causes of the
accident.
In terms of future work, collecting further data would help to get more insight of clearance
time effects. Examples of these data include the location of the vehicle after the accident,
damage rate, vehicle type, the involvement of hazard material, injuries age, and traffic flow
data. Also, further work is needed to investigate other interval times of the total accident
duration including reporting time and response time. This investigation will assist ACIB to
understand what effect each part of the total incident duration and, as a result, apply the best
measures to improve the efficiency of TIM. Finally, further study is needed to transfer this
study results into a prediction tool that can assist traffic operators in taking decisions.
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 13 of 20
REFERENCES
List of Figures and Tables
Figures:
FIGURE 1 Operational process of accident management
FIGURE 2 Methodology framework
FIGURE 3 Density distribution for accident clearance time
Tables:
TABLE 1 Accident Data
TABLE 2 Comparisons of AIC Values for AFT Models
TABLE 3 Weibull AFT Model of the Accident Clearance Time
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 14 of 20
Police Operational
Centre
Comprehensive
Police Patrol
ACIB
Crime Scene
Department
Traffic Control
Centre
Ambulance and
Rescue Service
SAAED
Serious
Accident
NoYes
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 15 of 20
FIGURE 1 Operational process of accident management
Fire Department
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 16 of 20
D
AT
A
A
N
A
LY
SI
S
DAT
A
COL
LEC
TIO
N &
PRE
PAR
ATI
ON
FTSS Records
- Temporal characteristics
- Geographic characteristics
- Environmental characteristics
- Accident characteristics
ACIB Records
- Arrival time
- Departure time
Analysing covariate effects on clearance times using Fully Parametric Hazard
models by estimating Accelerated Failure Time (AFT) metric
-Estimated Coefficient Sign
-Magnitude in Clearance Time
Data Processing
Data cleaning, coding
and preparation
Investigating the best distribution to represent clearance time data using:
-Akaike Information Criterion
-Unobserved heterogeneity test
Step 3
Step 1
Step 2
Interpreting covariate effects on clearance time and their links to the current
practices of clearing accident in Abu Dhabi
Investigating the significant level of the explanatory Variables
using - 2 log statistic
Investigating the significant level of the explanatory Variables
using - 2 log statistic
Investigating for further significant variables using Stepwise Method
Investigating for further significant variables using Stepwise Method
Developing a base model using clearance time data
Developing a base model using clearance time data
Expone
ntial
Weibu
ll
Gomp
ertz
Log-
normal
Models
estimated
Log-
logistic
Clearance
Time
Database
583 Accident records
185 variables
FIGURE 2 Methodology framework
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 17 of 20
0 .01 .02 .03 .04
Density
0 50 100 150
Clearance time
FIGURE 3 Density distribution for accident clearance time
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 18 of 20
TABLE 1 Accident Data
Database Characteristics
The Federal
Traffic
Statistics
System
(FTSS)
Temporal characteristics
- Time of day (Morning, Afternoon, Evening, AM peak, PM peak )
- Day of week (Sunday, Monday, Tuesday, Wednesday, Thursday,
Friday, Saturday)
- Month of year (January, February, March, April, May, June, July,
August, September, October, November, December)
Geographical characteristics
- Street
- Intersection
- Region
- Road layout
- Place nature
Environmental characteristics
- Weather condition (Clear, Foggy, Rain, Windy)
- Road surface condition (Dry, Wet, Sandy)
- Light condition (Daylight, Darkness)
Accident characteristics
- Severity ( Slight, Serious, Fatal)
- Number of casualties
- Number of vehicles involved
- Accident type (Side Impact, Hit Pedestrian, Hit Object, Overturn,
Rear-end, Head-on, Other types of Accidents)
ACIB
Records
Accident clearance time
- Investigator arrival time
- Investigator departure time
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 19 of 20
TABLE 2 Comparisons of AIC Values for AFT Models
Model -2In LK c AIC
Exponential 1377.94 17 1 1413.94
Weibull 1074.99 26 2 1130.99
Gompertz Without Frailty 1188.08 29 2 1250.08
With gamma Frailty 1125.97 30 2 1187.97
Log-normal 1116.09 33 2 1186.09
Log-logistic 1084.54 36 2 1160.54
Abdulla Alkaabi, Dilum Dissanayake, Roger Bird Page 20 of 20
TABLE 3 Weibull AFT Model of the Accident Clearance Time
Bold figures are significant at 90% level of significance
Variable Estimated
Coefficient
t-statistics Percentage
Change (%)
Temporal Characteristics
Out of peak period 0.15 2.68 16.5
Monday - 0.09 -1.51 -8.9
August - 0.33 -3.97 -28.8
October - 0.17 -2.20 -16.4
December - 0.12 -1.55 -11.6
January 0.26 2.82 30.2
March 0.25 3.18 28.6
Geographical Characteristics
23ed Street - 1.58 -2.99 -79.4
10th Street - 0.39 -2.18 -32.7
Hamdan Bin Mohammed Street - 0.24 -2.03 -21.9
31st Street - 0.73 -1.94 -52.2
Meena Street 0.25 2.29 29.2
35th Street 1.40 2.61 306.8
Bainunah Street 0.28 1.44 32.7
School area 0.20 1.82 22.9
Commercial area - 0.17 -3.23 -15.7
Car park - 0.30 -2.49 -26.1
Government authority area - 0.11 -1.66 -10.7
Environmental Characteristics
Weather condition: Windy 1.16 2.10 221.2
Weather condition: Clear 0.96 2.00 162.3
Weather condition: Rain 1.83 3.25 524.8
Accident Characteristics
Number of injuries per accident 0.04 2.48 4.7
Number of vehicles involved
per accident
0.10 3.43 11.2
Accident type: Hit pedestrian - 0.13 -1.98 -12.7
Accident type: Hit object 0.22 2.17 25.7
Accident type: Angle Collision - 0.09 -1.57 -8.9
Model Structure Parameters
p (distribution shape parameter) 1.90 31.25
λ (the scale parameter) 2.15 4.27
Goodness-of-fit test
Akaike information criterion 1130.99
Initial Log Likelihood -630.05
Log-Likelihood at convergence -537.49
Number of observations 583
