traffic safety along rural mountainous highways in malaysia · variables (56 in total) representing...

Traffic Safety along Rural Mountainous

Highways in Malaysia

Rusdi Bin Rusli

M.Edu in Technical and Vocational Education

Submitted in fulfilment of the requirements for the degree of

Doctor of Philosophy

Centre for Accident Research and Road Safety – Queensland (CARRS-Q)

School of Psychology and Counselling

Faculty of Health

Queensland University of Technology

2017

Keywords i

Keywords

Mountainous highways, Single-vehicle crashes, Multi-vehicle crashes, Random

Parameters Negative Binomial (RPNB), Crash dataset with excess zeros, Crash

severity, decision tree, injury prevention, injury severity, Negative Binomial –

Generalized Exponential (NB-GE), Negative Binomial – Lindley (NB-L), Random

Parameters Logit model, Scobit model, road trauma, road safety, rural highways,

Sabah, Malaysia

Abstract ii

Abstract

Road traffic crashes along mountainous highways are more injurious compared to the

highways on plain topography. Crash statistics show that the ratio of fatalities to road

injuries along mountainous highways in Malaysia is about four times higher than

non-mountainous roads. This might be because of constrained topography

conditions, complex road geometry and adverse weather conditions along

mountainous highways, which often represent a demanding driving situation. In

addition, mountainous highways, particularly those in developing countries such as

Malaysia, often have narrow road shoulders and risky roadside environments due to

steep cliffs and high embankments. Despite the uniqueness of mountainous highways

and the higher fatality rates, there is only limited research on this topic. As a result,

the ways in which different roadway geometric characteristics, cross-sectional

elements, roadside features, spatial characteristics, and traffic factors influence the

crash occurrence and injury severity of traffic crashes along rural mountainous

highways are not well understood.

As such, the objective of this study is to develop an in-depth understanding of

road traffic crashes along rural mountainous highways so that targeted

countermeasures could be developed. To achieve this, this research has identified

four sub-objectives: 1) to examine the characteristics of road traffic crashes on rural

mountainous roads and to compare these with the characteristics of crashes on roads

located in non-mountainous areas; 2) to investigate the effects of roadway

geometries, traffic characteristics, real-weather conditions, cross-sectional elements,

roadside features, and spatial characteristics on Single-vehicle (SV) crashes along

rural mountainous highways; 3) to examine critical factors contributing to Multi-

vehicle (MV) crashes on rural mountainous highways; and 4) to investigate the

injury severity of road traffic crashes along rural mountainous highways.

This research has selected Sabah, Malaysia as the study area as about 60% of

Sabah state has mountainous topography and accommodates many of the steepest

roads in Malaysia. Five different datasets have been merged to construct the final

unique dataset for this research. They are: 1) road traffic crash data obtained from the

Malaysian Institute of Road Safety Research (MIROS); 2) topographical information

Abstract iii

obtained from the Department of Survey and Mapping Malaysia; 3) weather

conditions obtained from the Department of Irrigation and Drainage Sabah (DID)

and National Centers for Environmental Information (NOAA); 4) traffic volume

information obtained from Highway Planning Unit, Ministry of Works Malaysia and

Public Works Department Sabah (PWD); and 5) roadway geometric, cross-sectional

element, roadside features and spatial characteristic data from an extensive field

survey.

Study 1 involved a systematic statistical comparison of general crash

characteristics including the crash type, crash severity, roadway geometric features,

environmental factors, and road user/vehicle characteristics between mountainous

and non-mountainous highways. During the five years period from 2008 – 2012, a

total of 25,439 crashes occurred along federal highways in Sabah, of which 4,875

crashes occurred in mountainous areas. Categorical data analysis techniques were

used to examine the differences between mountainous and non-mountainous crashes.

Results show that the odds ratio of ‘out-of-control’ collisions and the crash

involvement due to speeding are respectively about 4.2 times and 2.8 times higher on

mountainous than non-mountainous roads. Other factors and crash characteristics

that increase the odds of crashes along mountainous roads compared with non-

mountainous roads include horizontal curved sections compared with straight

sections, single-vehicle crashes compared with multi-vehicle crashes and weekend

crashes compared with weekday crashes.

The state-of-the-art count modelling technique was applied to investigate the

statistical relationship between single-vehicle (SV) crashes and a wide range of

variables (56 in total) representing road geometries including horizontal and vertical

alignment, traffic characteristics, real-time weather conditions, cross-sectional

elements, roadside features, and spatial characteristics to achieve sub-objective 2. To

account for structured heterogeneities resulting from multiple observations within a

site and other unobserved heterogeneities, the study applied a random parameters

negative binomial (RPNB) model to develop a safety performance function (SPF) for

SV crashes. Results suggest that rainfall at the time of crash is positively associated

with SV crashes, but real-time visibility is negatively associated. The presence of a

road shoulder, particularly a bitumen shoulder or wider shoulders, along

mountainous highways is associated with a lower likelihood of SV crashes. While

Abstract iv

speeding along downgrade slopes increases the likelihood of SV crashes, the

presence of delineation decreases SV crash frequencies.

Study 3 of this research investigated multi-vehicle (MV) crashes along rural

mountainous highways, which represent about 35% of total crashes along selected

rural mountainous highways. The dataset for MV crashes suffers from

heterogeneities resulting from excess zero counts. To address this, two specialized

modelling techniques for excess zeros including Negative Binomial – Lindley (NB-

L) and Negative Binomial – Generalized Exponential (NB-GE) were employed, and

their performances were compared with a Random Parameters Negative Binomial

(RPNB). Results showed that the RPNB model outperformed NB-L and NB-GE

models in terms of prediction ability and model fit. It was found that heavy rainfall at

the time of crash, presence of horizontal curves along a steep gradient and the

presence of minor junctions along mountainous highways increase the likelihood of

MV crashes, while the presence of an overtaking lane and the presence of road

delineation both decrease the likelihood of MV crashes.

Study 4 examined factors contributing to injury severity of traffic crashes

along rural mountainous highways. This study applied a two-step modelling

approach with a combination of decision tree analysis and discrete outcome model.

While the decision tree identified the possible high order interactions among

explanatory variables and provided an input to the discrete outcome model as a priori

knowledge, the logistic regression developed inferences for contributing factors

influencing injury severity of crashes. This novel methodology has been tested for

three discrete outcome models including standard logit model as a base model,

Scobit model accounting for the imbalance among injury categories, and random

parameters logit model accounting for unobserved heterogeneities. Results showed

that the combination of decision tree and random parameters logit regression model

perform better in terms of model fitness and identifying significant variables. The

likelihood of severe crashes decreases in rear-end collisions, while it increases in

head-on collisions. Crashes involving female drivers are less likely to be severe,

however, crashes involving heavy vehicles are more likely to be severe. Crashes

during rainy conditions decrease the likelihood of severe crashes. The proportion of

segment length with simple curves, combination of horizontal and vertical alignment,

proportion of segment length with unsealed shoulder and proportion of segment

Abstract v

length with cliffs along both sides of highway segments are associated with high

injury severities. A high order interaction term suggests that the severity of crashes

decreases when light and medium vehicles get involved in a single-vehicle crash

along a highway segment with higher proportions of its length covered by curves.

Another interaction term indicates that single-vehicle crashes involving heavy

vehicles are associated with severe injuries along highway segments with a

combination of vertical longitudinal grades over less than 8% and horizontal curves

over less than 50% of its length.

The availability of reliable and accurate data is a common barrier in

conducting road safety research in the context of a developing country like Malaysia.

By overcoming this limitation by an extensive field survey and scrutinizing various

secondary data sources, this research has provided several new insights into road

safety issues along rural mountainous highways in developing countries. This has

both theoretical and practical contributions. First, the developed SPF with random

parameters model to deal with excess zero counts represents a significant

contribution for crash modelling. The econometric modelling technique combining

the decision tree analysis and discrete outcome model also represents a unique

contribution in modelling injury severity of traffic crashes. Second, the findings of

this research will help developing targeted countermeasures to reduce the likelihood

of crashes and/or lessen the injury severities resulting from traffic crashes, and thus

help in designing a safer environment for rural highways in mountainous areas.

Table of Contents vi

Table of Contents

Keywords ...................................................................................................................... i

Abstract ....................................................................................................................... ii

Table of Contents ...................................................................................................... vi

List of Figures ............................................................................................................. x

List of Tables ............................................................................................................. xii

List of Abbreviations ............................................................................................... xiv

Statement of Original Authorship ......................................................................... xvi

Acknowledgements ................................................................................................. xvii

Associated Publications and Presentations ......................................................... xviii

Chapter 1: Introduction ...................................................................................... 1

1.1 Background .................................................................................................................... 1

1.2 Rationale for Research ................................................................................................... 3

1.3 Research Objectives ....................................................................................................... 5

1.4 Scope of the Research .................................................................................................... 6

1.5 Research Questions ........................................................................................................ 6

1.6 Conceptual Framework .................................................................................................. 7

1.7 Research Design ........................................................................................................... 10

1.8 Structure of the Thesis ................................................................................................. 13

Chapter 2: Literature Review ........................................................................... 16

2.1 Introduction .................................................................................................................. 16

2.2 Road Safety along Mountainous Highways ................................................................. 16

2.3 Safety Performance Functions (SPF) ........................................................................... 18

2.4 Factors Influencing Crash Occurrence along Mountainous Highways ........................ 19

2.5 Factors Influencing Injury Severity along Mountainous Highways ............................ 27

Table of Contents vii

2.6 Crash Modelling by Crash Types ................................................................................. 28

2.7 Statistical Modelling Techniques ................................................................................. 29

2.8 Identified Research Gaps .............................................................................................. 34

2.9 Chapter Summary ......................................................................................................... 39

Chapter 3: Methodology and Data ................................................................... 40

3.1 Study Setting and Population ........................................................................................ 40

3.2 Data Collection ............................................................................................................. 42

3.3 Crash Characteristics Analysis ..................................................................................... 60

3.4 Single – Vehicle Crash Model ...................................................................................... 60

3.5 Multi – Vehicle Crash Model ....................................................................................... 64

3.6 Crash Severity Model ................................................................................................... 70

3.7 Health Risk Assessment and Ethics Statement ............................................................. 72

3.8 Chapter Summary ......................................................................................................... 73

Chapter 4: Characteristics of Mountainous Roads Crashes .......................... 74

4.1 Introduction .................................................................................................................. 74

4.2 Objectives ..................................................................................................................... 74

4.3 Data Description ........................................................................................................... 74

4.4 Methodology ................................................................................................................. 75

4.5 Results .......................................................................................................................... 75

4.6 Discussion ..................................................................................................................... 80

Chapter 5: Single Vehicle Crashes ................................................................... 85

5.1 Introduction .................................................................................................................. 85

5.2 Objectives ..................................................................................................................... 85

5.3 Data Description ........................................................................................................... 86

5.4 Methodology ................................................................................................................. 88

5.5 Model Results ............................................................................................................... 89

5.6 Discussion ..................................................................................................................... 93

Chapter 6: Multi Vehicle Crashes .................................................................... 99

Table of Contents viii

6.1 Introduction .................................................................................................................. 99

6.2 Objectives .................................................................................................................... 99

6.3 Data Description ........................................................................................................ 100

6.4 Methodology .............................................................................................................. 103

6.5 Model Results ............................................................................................................ 103

6.6 Discussion .................................................................................................................. 109

Chapter 7: Crash Severity Study .................................................................... 113

7.1 Introduction ................................................................................................................ 113

7.2 Objectives .................................................................................................................. 113

7.3 Data Description ........................................................................................................ 114

7.4 Data Availability ........................................................................................................ 116

7.5 Methodology .............................................................................................................. 117

7.6 Results ........................................................................................................................ 118

7.7 Discussion .................................................................................................................. 123

Chapter 8: Discussion and Conclusions ......................................................... 128

8.1 Introduction ................................................................................................................ 128

8.2 Review of Findings .................................................................................................... 128

8.3 Contribution to Scientific Knowledge and Implications ............................................ 138

8.4 Strengths and Limitations .......................................................................................... 149

8.5 Conclusions ................................................................................................................ 152

8.6 Recommendations for Future Research ..................................................................... 155

Bibliography ........................................................................................................... 157

Appendices .............................................................................................................. 171

APPENDIX A ...................................................................................................................... 171

APPENDIX B ...................................................................................................................... 181

APPENDIX C ...................................................................................................................... 191

APPENDIX D ...................................................................................................................... 193

APPENDIX E ....................................................................................................................... 195

Table of Contents ix

APPENDIX F........................................................................................................................ 197

APPENDIX G ....................................................................................................................... 198

APPENDIX H ....................................................................................................................... 201

APPENDIX I ........................................................................................................................ 203

APPENDIX J ........................................................................................................................ 207

APPENDIX K ....................................................................................................................... 208

APPENDIX L ....................................................................................................................... 209

APPENDIX M ...................................................................................................................... 211

List of Figures x

List of Figures

Figure 1-1: Predicted Road Fatalities in Malaysia up to 2020 (MOT, 2014) .............. 1

Figure 1-2: Five pillars proposed by WHO dealing with road safety (WHO,

2011) .............................................................................................................. 7

Figure 1-3: Overall Framework of Road Safety Plan of Malaysia 2014 – 2020

(MOT, 2014) .................................................................................................. 9

Figure 1-4: Safe System Approach (NRSS, 2011) ....................................................... 9

Figure 1-5: Conceptual Framework ........................................................................... 12

Figure 1-6: An overview of the thesis ........................................................................ 15

Figure 3-1: Location of Sabah in Malaysia. Source: adapted from google maps ...... 41

Figure 3-2: Topography Maps of Sabah and Selected Roads for Study. Source:

adapted from Google maps .......................................................................... 42

Figure 3-3: Location of Study Sites ........................................................................... 43

Figure 3-4: Handheld GPS Garmin Etrex 10 ............................................................. 46

Figure 3-5: A typical horizontal curve along a mountainous road segment. ............. 46

Figure 3-6: Details Plotting in AutoCAD 2015 ......................................................... 47

Figure 3-7: Measuring cross-sectional elements using measuring wheels ................ 47

Figure 3-8: One of the selected highway segments along the Kimanis –

Keningau Highway ...................................................................................... 48

Figure 3-9: Different types of horizontal curves in a typical road in

mountainous areas ........................................................................................ 49

Figure 3-10: Ogive for 85th percentile of a sample of speed ...................................... 50

Figure 3-11: Segment coverage by different rainfall and weather stations ................ 60

Figure 3-12: Resemblance of Lindley distribution to the distribution of MV

crash count with excess zeros. ..................................................................... 65

List of Figures xi

Figure 3-13: Resemblance of Generalized Exponential distribution to the

distribution of MV crash counts with excess zeros. .................................... 67

Figure 4-1: Percentage of crashes by vehicle type for mountainous and non-

mountainous roads ....................................................................................... 80

Figure 5-1: Frequency of speeding-related SV crashes under different visibility

conditions ..................................................................................................... 95

Figure 5-2: The relationship between SV crash frequencies and road shoulders ...... 97

Figure 6-1: Adjusted cumulative residual plots for exposure variable. ................... 104

Figure 7-1 : Decision Tree ....................................................................................... 120

List of Tables xii

List of Tables

Table 3-1: List of District in Sabah ............................................................................ 41

Table 3-2: Total Registered Vehicles and New Registrations in Sabah for 2010 ...... 42

Table 3-3: List of Highway Segments based on Segmentation Criteria .................... 44

Table 3-4: List of available information/ variables in M-ROADS ............................ 51

Table 3-5: Crash Characteristics for all Rural Mountainous Highways in Sabah

vs. along Selected 102 Highway Segments ................................................. 57

Table 3-6: List of Nearest Rainfall Stations ............................................................... 59

Table 3-7: List of Nearest Weather Stations .............................................................. 59

Table 4-1: General crash characteristics .................................................................... 76

Table 4-2: Characteristics of crashes by time of the day, day of the week, and

seasonal variations ....................................................................................... 77

Table 4-3: Driver and vehicle factors ......................................................................... 79

Table 5-1: Summary statistics of explanatory variables included in the model ......... 86

Table 5-2: RPNB model estimates of SV crashes along rural mountainous

highways ...................................................................................................... 89

Table 5-3: Elasticity and pseudo-elasticity estimates of significant variables in

SPF ............................................................................................................... 91

Table 5-4: Cross-tabulation analysis of shoulder type and width for SV crashes ...... 97

Table 6-1: Summary statistics of variables included in the model .......................... 100

Table 6-2: Modelling results for MV crashes along rural mountainous

highways .................................................................................................... 105

Table 6-3: Elasticity and pseudo-elasticity for crash contributing factors of the

RPNB model .............................................................................................. 108

Table 7-1: Summary statistics of variables included in the model .......................... 114

List of Tables xiii

Table 7-2 : Estimation results for standard logit, Scobit, and random

parameters logit models ............................................................................. 121

Table 8-1: A comparison of factors associated with SV and MV crashes ............... 131

Table 8-2: A comparison of factors influencing road safety between developing

and developed countries. ............................................................................ 134

Table 8-3: Recommendations for engineering treatments to improve road

safety along rural mountainous highways in Malaysia .............................. 146

List of Abbreviations xiv

List of Abbreviations

4WD Four Wheel Drive

ADT Average Daily Traffic

AADT Annual Average Daily Traffic

AASTHO American Association of State Highway and Transportation

AIC Akaike Information Criterion

BIC Bayesian Information Criterion

CMF Crash Modification Factors

CURE Cumulative Residual

DIC Deviance Information Criterion

DID Department of Irrigation and Drainage Sabah

DOSM Department of Statistics Malaysia

DTM Digital Terrain Model

DTNB Decision table/Naïve Bayes

ESC Electronic Stability Control

FHWA Federal Highway Administration, U.S.

GOF Goodness-of-fit

GPS Geographical Position System

HEF Hourly Expansion Factors

HPU Highway Planning Unit, Ministry of Works Malaysia

IDS Institute for Development Studies Sabah

JKJR Road Safety Department of Malaysia

MAD Mean Absolute Deviation

MCMC Markov Chain Monte Carlo

MIROS Malaysian Institute of Road Safety Research

MLE Maximum Likelihood Estimation

MOT Ministry of Works Malaysia

M-ROADS MIROS Road Accident Analysis and Database System

List of Abbreviations xv

MSE Mean Squared Error

MSLE Maximum Simulated Likelihood Estimation

MSPE Mean Squared Predictive Error

MV Multi-Vehicle Crashes

NB Negative Binomial

NB-CR Negative Binomial – Crack

NB-GE Negative Binomial – Generalized Exponential

NB-L Negative Binomial – Lindley

NOAA National Centers for Environmental Information

PDO Property Damage Only

PIAM The General Insurance Association of Malaysia

PIG Poisson Inverse Gaussian

PWD Public Works Department

REAM Road Engineering Association of Malaysia

RP Random Parameters

RPNB Random Parameters Negative Binomial

RTMS Remote Traffic Microwave Sensor

RTVM Road Traffic Volume Malaysia

SUV Sport Utilities Vehicle

SV Single-Vehicle Crashes

SVM Support Vector Machine

SPF Safety Performance Function

VKT Vehicle Kilometre Travel

WHO World Health Organization

WinBUGS Windows Software for Bayesian Inference Using Gibbs Sampling

ZINB Zero-Inflated Negative Binomial

ZIP Zero-Inflated Poisson

Statement of Original Authorship xvi

Statement of Original Authorship

The work contained in this thesis has not been previously submitted to meet

requirements for an award at this or any other higher education institution. To the

best of my knowledge and belief, the thesis contains no material previously

published or written by another person except where due reference is made.

Signature: QUT Verified Signature

Date: October 2017

Acknowledgements xvii

Acknowledgements

I would like to express my sincerest gratitude to my supervisor, Dr Md.

Mazharul (Simul) Haque for his continuous support, constructive advice, patient

guidance, and exceptional encouragement throughout my PhD study and research.

His door was always open for me discuss any problem throughout my PhD journey.

This thesis would not have been possible without his invaluable help in all the

research and writing of this thesis. I also learnt precious skills during my research

from him which have been important in helping me in my professional pursuits.

My deepest thanks also goes to my associate supervisors, Dr Mark King and

Prof Wong Shaw Voon. Dr Mark King always used his experience and expertise in

solving problems related to my research candidature. Although Prof Wong was not

physically present at QUT, he contributed significantly to my research, in particular

in relation to crash data in Malaysia. Thanks also to Kat Bowman and Natasha

Kitano, the professional editors of my thesis.

It has been a pleasure to know and work with all of the staff and students in

CARRS-Q – particularly Oscar, Navid, Yusuf, Peter, Wahi, Atiyeh, Kristi and Saif

for their wonderful company and encouragement during my studies.

I would also like to thank a number of relevant agencies in Malaysia for

providing support and help to facilitate the data collection through site surveys. They

include the Malaysian Institute of Road Safety Research (MIROS), the Road Safety

Department Sabah (JKJR), and the Public Works Department (JKR) Sabah. Many

thanks is also dedicated to Mr Herdianshah Abdul Karim, the Director of Road

Safety Department of Malaysia, Sabah Branch for his generous support during my

field surveys.

Finally, I would like also to thank and show gratitude to my family, especially

my wife Rohaida Hayani, because she has been a very patient and a good listener

throughout this period. Also, my children, Hazim, Aida, Aina and Hariz. I would

like to thank them for their patience and understanding as I have been preoccupied

with my thesis. Lastly, I cannot forget my Mother and Father in Malaysia and their

continued prayers for my success.

Associated Publications and Presentations xviii

Associated Publications and Presentations

Publications

1. Rusli, R., Haque, M.M., King, M., & Wong, S. V. (2017). Single-Vehicle

Crashes along Rural Mountainous Highways in Malaysia: An Application of

Random Parameters Negative Binomial Model. Accident Analysis &

Prevention, 102, 153 – 164.

2. Rusli, R., Haque, M.M., Afghari, A. P., King, M., & Wong, S. V. (2017).

Multi-Vehicle Crashes along Rural Mountainous Highways: A Comparison

of Count Models with the Ability to Handle Excess Zeros. Accident Analysis

& Prevention, Article under review.

3. Rusli, R., Haque, M.M., Saifuzzaman, M., King, M., & Wong, S. V. (2017).

Injury severity of traffic crashes along rural mountainous Highways in

Malaysia: An application of combined decision tree and logistic regression

model. Traffic Injury Prevention, Working paper.

Presentations

1. Rusli, R., Haque, M.M., King, M., & Wong, S. V. (2015). A comparison of

road traffic crashes along mountainous and non-mountainous roads in

Sabah, Malaysia. In Australasian Road Safety Conference 2015. 14 – 16

October 2015. Gold Coast Covention and Exhibition Centre, Gold Coast,

Australia.

2. Rusli, R., Haque, M.M., King, M., & Wong, S. V. (2015). An Exploratory

Analysis of Mountainous Road Crashes in Sabah, Malaysia. In Malaysian

PostGraduate Colloquium 2015. 19 – 20 December 2015. Melbourne,

Australia.

3. Rusli, R., Haque, M.M., King, M., & Wong, S. V. (2016). A comparison of

road traffic crashes along mountainous and non-mountainous roads in

Sabah, Malaysia. In Sabah Road Day 2016: Sustainable Road for

Development. 5 – 6 April 2016. Kota Kinabalu, Sabah.

4. Rusli, R., Haque, M.M., King, M., & Wong, S. V. (2016). Traffic Safety

along Mountainous Highways in Malaysia. Annual School of Psychology and

Associated Publications and Presentations xix

Counselling Postgraduate Research Symposium. 2 November 2016. South

Bank, Brisbane, Australia.

Chapter 1: Introduction 1

Chapter 1: Introduction

1.1 BACKGROUND

Road safety represents an important concern because of the increase in road

crashes and fatalities around the world in general, developing countries in particular.

The World Health Organization (WHO) predicts that road traffic injuries will climb

from ninth place in 2004 to fifth place in 2030 as one of the leading causes of death

worldwide (WHO, 2009). In the year 2000, the U.S. lost 1.5 people (fatalities) per

100 million-vehicle-miles in road crashes (Abdel-Aty, 2003). Meanwhile, 28,000

people died due to road crashes in the European Union in the year 2012 (Garrido,

Bastos, de Almeida, & Elvas, 2014). In China, the number of fatalities in road

crashes reached 67,759 in the year 2009 (Chen, Wei, & Zhou, 2011). Statistics from

the Department of Statistics Malaysia (2012) show that the number of fatalities due

to road crashes is increasing every year. For example, the total number of crash-

related fatalities in Malaysia in 2007 was 6,282, and rose to 6,877 in 2011. Sarani,

Rahim, Marjan, and Voon (2012) predict that the total annual road fatalities in

Malaysia could reach nearly 11,000 by 2020. Figure 1-1 shows the increasing trend

of road fatalities in Malaysia, with future predicted if additional road safety

initiatives not introduced.

Figure 1-1: Predicted Road Fatalities in Malaysia up to 2020 (MOT, 2014)

YEAR

FA

TA

LIT

IES

FATALITIES

ARIMA


The road network is an important aspect of a country’s economic

development and modernization. Every country provides an annual budget for the

construction of roadways and related facilities such as intersections, ramps, and

roadside safety barriers in line with population and vehicle growth. The economic

development of a country often requires roadway development across a wide range

of topographies, including mountainous areas. In many countries, mountainous areas

are extensive and account for a high degree of economic activity simply because

many people live there, while in other cases, roads through mountainous areas are

vital links to facilitate freight movement, or are required to support development for

an important resource, such as minerals.

Mountainous road engineering work requires considerable technological

capacities and resources to build safer roads, however, these are not always available,

which can make it difficult to maintain desirable roadway geometric features that

ensure the safety of the road users (Yun, Shui, & Zuo, 2013). Moreover,

topographical features and road environment are closely linked with driving

performance, as driving along high and steep mountain grades is often a challenging

task. In China, many blackspots or high risk locations are located along mountainous

highways (Lin, Jinhai, & Yan, 2013) and more than 75% of fatalities and serious

injury crashes between 2007 and 2013 occurred along mountainous highways (Chen,

Li, et al., 2016). Mountainous sections of the I-70 freeway in Colorado, U.S. also

recorded higher crash rates compared with non-mountainous sections (Yu, Xiong, &

Abdel-Aty, 2015). Thus, it is crucial to investigate factors contributing to traffic

crashes in such environments in order to mitigate the risk associated with those

crashes.

The severity of crashes in mountainous regions is often significantly higher

than in flat areas, mainly because of the hazard associated with vehicles falling down

cliffs or striking high embankments when a crash occurs. ‘Right of way’ along

mountainous roads is limited and sometimes only meets the minimum engineering

standard. For example, buffer zones, especially for run-off roadway crashes, may not

be enough and therefore, additional support from other road safety facilities, such as

guardrails and fences is often required.

A study in India (Rautela & Pant, 2007) reported that traffic safety along

mountainous roads is largely dependent on three basic roadway parameters:


sinuosity, gradient, and road width. This study also reported that the Fatality Index

(FI), the ratio of fatalities to road injuries, is significantly higher in mountainous

regions than in flat areas. Similar patterns are also observed in Malaysia; road crash

statistics for the year 2011 suggest that Ranau, a district in Sabah, Malaysia with

almost 90% mountainous areas, has an FI of about 0.4 compared with Kota

Kinabalu, another district in Sabah with about 20% mountainous areas which has an

FI of only about 0.11 (DOSM, 2012). Furthermore, Lin, et al. (2013) reported that

road traffic crashes along mountainous roads are more severe in nature and are

generally associated with greater difficulties for rescue operations to retrieve the

injured victims. Therefore, an appropriate understanding of injury patterns in

mountainous road crashes is required to develop appropriate countermeasures to

improve safety.

In mountainous regions, driving speed along longitudinal grades is an

important issue as drivers often face difficulties in controlling their vehicle speed.

Continuous braking along downgrades may make brake pads hot, cause them to lose

their grip, and eventually may lead to a loss of brake function. A study (Yuan, Fu,

Guo, Feng, & Shi, 2008) on traffic crashes on longitudinal grades reported that

crashes are usually concentrated in the latter half of the down-grade, and that risk

multiplies if there are additional horizontal curves. Thus, it is important to

understand how speed characteristics influence traffic crashes in mountainous

regions.

1.2 RATIONALE FOR RESEARCH

Many factors are associated with road traffic crashes, including traffic

conditions, roadway geometric features, environmental factors and driver/vehicle

characteristics. Many researchers have argued that the effect of these factors on crash

occurrence and injury severity vary across location types. For instance, using a

disaggregate approach Qin, Ivan, and Ravishanker (2004) demonstrated that the

relationships between crashes and traffic volumes are different across different

locations such as road segments and intersections. Milton, Shankar, and Mannering

(2008) argued that the injury-severity outcomes are likely to be different across

geographical locations. For motorcycle crashes, Haque, Chin, and Debnath (2012)

demonstrated that the crash characteristics vary across location types such as

intersections, expressways and other road sections away from intersections. For a


similar reason, crash characteristics in urban and rural areas were found to be

compared to identify the impact of geographic differences on crash outcome, and

interventions were suggested for rural and urban areas separately (Li, Doong, Chang,

Lu, & Jeng, 2008). Although mountainous regions have unique characteristics in

terms of road design and weather condition, only few studies have investigated road

safety issues along mountainous highways.

Economic and population growth around the world has increased

development in rural and mountainous areas, however, the complex topography

specific to mountainous regions often make it challenging to construct roads

following the engineering standards to ensure safety. This problem is worse in the

context of a resource-constrained developing country. For example, many

mountainous highways in developing countries are constructed with insufficient

shoulder width or without a clear zone because of the costs involved in providing

these features. This means that there is insufficient distance between the roadway

lanes and dangerous roadside features such as cliffs and embankments. Despite the

fact that constrained topography and complex road geometries are among major

issues in designing and constructing roads in mountainous regions, research

examining the effects of road traffic parameters on mountainous road safety is

relatively scant.

Speed is one of the most important parameters of road safety and has a direct

relationship with crash occurrence. Posted speed limits are often used as a proxy

measure of traffic speed along road segments, but the posted speed limit may not be

a good indicator, particularly along mountainous roads, because traffic speed on

these roads may be influenced by roadway geometric characteristics and driver

perceptions of comfort and safety (Castro, Sánchez, & Sánchez, 2012). In fact,

Ahmed, Huang, Abdel-Aty, and Guevara (2011) reported that crash frequencies

along mountainous roads are not significantly associated with the posted speed limit.

In contrast, Yu and Abdel-Aty (2013b) reported that the average speed recorded by

downstream detectors has significant associations with vehicle crashes along

mountainous roads. Other driver behaviour factors also play a pivotal role in traffic

safety; however, their effects are not much known for rural mountainous highways.

Weather in mountainous areas plays a vital role in road safety. A recent study

on the relationship between real-time weather and crash occurrences along a


mountainous freeway in the United States demonstrated that, depending on weather

conditions, the same traffic parameters along a mountainous road section might

influence driver behaviour, and thus safety, differently (Ahmed, Abdel-Aty, & Yu,

2012). In this study, real-time visibility refers to visibility during the 30 minutes

before and after the time of the crashes. Visibility cannot directly be captured in

weather stations; however, it was calculated based on the measurement of light

extinction, which includes the scattering and absorption of light by particles and

gases that have been automatically captured by three weather stations along the I-70

for the same time period. In a subsequent study that examined hazardous factors

involved in single- and multi-vehicle crashes along the same freeway (I-70) in the

United States, Yu, Abdel-Aty, and Ahmed (2013) reconfirmed that the crash

occurrence along mountainous roads is highly influenced by weather conditions, and

suggested the adoption of different active management strategies across different

seasons. In particular, visibility and precipitation were reported to increase crash

risks along mountainous section of the I-70 (Yu et al., 2015). Ma, Chen, and Chen

(2015b) also reported that poor visibility along mountainous roads increases the

crash risk. It is pertinent to point out that all of these studies were conducted in the

US, and their findings may not be generally applicable to Malaysia, which is located

in the equatorial region and has a tropical rainforest climate.

1.3 RESEARCH OBJECTIVES

The main objective of this research is to develop an in-depth understanding of

road traffic crashes along rural mountainous highways and propose possible

countermeasures to improve traffic safety along those highways. This main objective

is divided into four sub-objectives as follows:

1) To examine the characteristics of road traffic crashes on rural

mountainous roads and to compare these with the characteristics of

crashes on non-mountainous roads.

2) To investigate the effects of roadway geometries, traffic characteristics,

real-weather conditions, cross-sectional elements, roadside features, and

spatial characteristics on Single-vehicle (SV) crashes along rural

mountainous highways.

3) To examine critical factors contributing to Multi-vehicle (MV) crashes

along rural mountainous highways.


4) To investigate the injury severity of road traffic crashes along the rural


1.4 SCOPE OF THE RESEARCH

The state of Sabah in Malaysia was selected as the study area for this research

because more than 60% of its area is mountainous. It also has the steepest roads in

Malaysia, which include roads with more than 15% slope. Sabah is the easternmost

state in Malaysia and is located on the island of Borneo, i.e. physically separated

from Peninsular Malaysia, which is where the majority of Malaysia’s population live

and where Kuala Lumpur, the capital city of Malaysia is located. It has a population

of 3.2 million population scattered over 73,634 km2 area (DOSM, 2010). The total

number of vehicles registered in 2010 was 862,181, with the highest proportion

being passenger cars (56.5%), followed by motorcycles (23.7%) (MOT, 2010).

1.5 RESEARCH QUESTIONS

The following research questions were identified after conducting an in-depth

literature review. The main research questions to be addressed in the area of traffic

safety along rural mountainous highways are:

1. What are the characteristics of road traffic crashes along rural

mountainous roads?

2. What are the factors that lead to SV crashes along rural mountainous

highways?

3. What are the factors that contribute to the occurrence of MV crashes

along rural mountainous highways?

4. What are the factors related to the crash severity of traffic crashes along

rural mountainous highways?

Following the above four main research questions, the sub-questions of this research

are:

1. What are the most common types of crashes occurring along rural

mountainous roads?

2. What are the proportions of vehicle types involved in crashes along rural

mountainous roads?


3. What is the effect of weather and traffic conditions on SV and MV crash

occurrence? Does weather at the time of incident play a role?

4. Does speeding contribute to road crashes along upgrade and downgrade

sections on rural mountainous highways?

5. What are the variables of road geometry, cross-sectional elements,

roadside features, and spatial characteristics that influence SV and MV

crashes along rural mountainous roads?

6. Is there any difference between the likelihood of crashes in SV and MV

crashes along rural mountainous highways?

7. What driver and vehicle factors are associated with greater or lesser crash

severity?

8. What is the relationship between weather conditions and crash severity?

9. What is the relationship between road geometry, cross-sectional elements,

roadside features, and crash severity?

10. Are the road safety factors along rural mountainous highways different in

developing and developed countries?

1.6 CONCEPTUAL FRAMEWORK

Five pillars have been highlighted by the World Health Organization (WHO)

in dealing with road safety problems around the world (WHO, 2011). They are:

1. Road Safety Management

2. Safer Roads and Mobility

3. Safer Vehicles

4. Safer Road Users

5. Post – Crash Response

Figure 1-2: Five pillars proposed by WHO dealing with road safety (WHO, 2011)


Based on these pillars, the Ministry of Transportation (MOT) in Malaysia launched

the Road Safety Plan of Malaysia 2014 – 2020, which is a national level action

supporting a United Nations road safety program (MOT, 2014).

Figure 1-3 presents the basic framework and approach of the Road Safety

Plan of Malaysia 2014 – 2020. In addition to both plans, the Safe System approach

was also referred to in designing the framework for this research. Figure 1-4 shows

the diagram of the Safe System approach that has been implemented in Australia and

New Zealand. In this approach, road safety is viewed as a system that involves with

four main components: safe roads and roadsides, safe speed, safe vehicles and safe

road use. This system aims to create a forgiving road system based on four principles

including (NTZA, 2012):

1. People make mistakes

2. People are vulnerable

3. We need to share responsibility

4. We need to strengthen all parts of the system

In this research, both plans and approach were referred to for preparing this

research towards the improvement of road safety management along rural

mountainous areas. Referring to road safety plan introduced by WHO and MOT,

three pillars have been used in developing this research’s conceptual framework: 1)

Road Safety Management; 2) Safer Roads and Mobility; and 3) Safer Road Users.

The principles of the Safe System approach have also been included to strengthen the

conceptual framework constructed from these three pillars.

The first pillar highlighted in both plans is road safety management. To

manage road safety, the factors contributing to crashes need to be identified before

any action can be considered. In the Safe System approach recommends improving

understanding of crashes and associated risks by conducting research to identify the

contributing factors that are significant at different locations. In this study, three

modelling techniques have been suggested to identify the influence of the various

factors on SV, MV crashes, and crash severity. After completing the modelling

process, potential countermeasures will be proposed based on the identified

contributing factors. These countermeasures are related to innovation, legislation and

enforcement that can be implemented.


Figure 1-3: Overall Framework of Road Safety Plan of Malaysia 2014 – 2020 (MOT,

2014)

Figure 1-4: Safe System Approach (NRSS, 2011)

In the previous literature, mountainous areas are known to have complex

conditions with various combinations of road alignments. In addition, roadsides in

these areas are more dangerous compared to non-mountainous areas due to

dangerous cliffs and embankments. This affects the performance of the driver, who

reacts with different driving behaviours when travelling in these areas. Consistency

of design is also difficult to maintain in these areas because of space constraints.

Although crash rates on mountainous roads are not higher than the non-mountainous


roads, the percentage of fatal and serious injury crashes is much higher on

mountainous roads. Based on pillar 2, the stakeholders should consider the safety and

mobility in the context of the crash characteristics (i.e. severity on mountainous

roads) of the roads, but not only on higher traffic volumes and crash records. In the

Safe System approach, roads and the roadside are one of the important component.

The identification of contributing factors in this study will help authorities to prepare

countermeasures that can make the road safer for all road users.

The literature shows that speeding is an important contributing factor to

crashes along mountainous roads. Various road geometric factors such as

downgrades and upgrades influence driving speed along and therefore can increase

the likelihood of crashes. For example, on downgrade sections, conversion of

potential energy to kinetic energy under the acceleration of gravity increases vehicle

speed through external means rather than driver acceleration. In addition, drivers will

often increase their speed at the bottom of a downgrade in order to climb the

following slope without losing too much speed. Increased speed at these points can

increase the risk of crash occurrence. Pillar 3 is referred to in providing a safe

environment for road users. In the Safe System approach, safe speed is the most vital

component related to crash risk, and thus it is important to understand how it relates

to crash risk along mountainous highways.

Models are one of the techniques used to simplify descriptions or to represent

an understanding about complex phenomenon (Hughes, Anund, & Falkmer, 2014).

They help to create a mental picture, facilitate questioning, establish rules, check,

evaluate, analyse, identify and assess countermeasures and communication (Hughes,

Newstead, Anund, Shu, & Falkmer, 2014). Modelling has been used for study 2, 3

and 4 because it has proven useful for identifying contributing factors to crash

occurrence and crash severity. The in-depth literature review on crash modelling

presented in Chapter 2 shows that some challenges remain, such as developing count

models with excess zero counts and accounting for unobserved heterogeneities,

which are also addressed in this research.

1.7 RESEARCH DESIGN

An extensive field survey has been conducted to collect data for road

geometry, cross-sectional elements, roadside features, and spatial characteristics for a


sample of sites. This study also explores four secondary datasets including road

traffic crash data, topographical information, weather conditions, and traffic volume

data. The secondary data was collected directly from agencies such as the Malaysian

Institute of Road Safety Research (MIROS), Department of Survey and Mapping

Malaysia (JUPEM), the Department of Irrigation and Drainage Sabah (DID), the

National Centers for Environmental Information (NOAA), the Highway Planning

Unit (HPU), the Ministry of Works Malaysia and the Public Works Department

Sabah (PWD). MIROS source the crash data annually from the Royal Malaysian

Police, who use the POL27 form in reporting road traffic crashes (see Appendix A).

The most challenging task when conducting scientific research in the context

of developing countries is the availability of reliable data. In addition to utilizing

existing data sources, an extensive field survey was carried out to collect information

about roadway and traffic conditions in the study areas. Roadway geometric, cross-

sectional elements, roadside features and spatial characteristics data were mainly

collected by field surveys, as these data were not readily available from any agency

in Malaysia. Two survey teams were formulated to capture GPS coordinates along

the selected highway segments, using a handheld GPS device (Garmin Etrex 10).

Additional information such as cross-sectional elements of the road sections,

including the length of shoulder width were measured using measurement wheels.

Appropriate statistical models and analytical techniques were developed to achieve

the objectives of this research. Figure 1 shows the conceptual framework of this

research.

The selected mountainous highways for this research are all in Sabah state,

Malaysia. For Study 1, all Federal Highways in Sabah have been selected. In this

study, a comparison between mountainous and non-mountainous highways was

conducted to aid in understanding the different crash characteristics at both locations.

For Study 2, 3 and 4, four highways were selected based on their topography and

availability for field survey. They are: 1) Kimanis – Keningau highway, 2)

Penampang – Tambunan highway, 3) Tamparuli – Ranau highway, and 4) Ranau –

Sandakan highway. These highways are mainly two-lane, two-way highways (along

approximately 99% of their total length), and the posted speed limit along these

highways is mainly 90 km/h. These highways are the main mountainous highways in

the region with some stretches having average natural ground cross slopes as much


as 25% (REAM, 2002). Segmentations of these highways was done based on three

main criteria including gradient changes, the presence of a major intersection, and

changes in number of lanes. A sample of 102 highway segments were then randomly

selected for detailed data collection and subsequent crash modelling.

Figure 1-5: Conceptual Framework

Traffic Modelling

Crash Count

Injury Severity

Countermeasures

Road Safety Management

Safer Road Users

Speed Behaviour

(Operating speed)

Crash Occurrence

Safer Road and Mobility

Road Geometry

(Horizontal alignment)

(Longitudinal grades)

(Cross-sectional elements)

Environment

(Real-time weather)

(Roadway and roadside features)

(Spatial characteristics)

Single – Vehicle Crashes Multi – Vehicle Crashes

Traffic characteristics

(Traffic volume)

(Speed limit)


1.8 STRUCTURE OF THE THESIS

This thesis has eight main chapters as follows:

Chapter 1: This section presents the introduction of the research area and topic. It

highlights the background of the research and the rationale for

selecting this topic for study. This section also contains research aims

and objectives, scope of the research, research questions, conceptual

framework that applies to this research and research design.

Chapter 2: This chapter presents a review of the literature relevant to this

research. It describes the previous road safety research conducted in

mountainous areas and the application of Safety Performance

Functions. A literature review on factors contributing to the crash

occurrence and injury severity along mountainous highways is then

discussed. The importance of separating crash modelling for SV and

MV crashes is then presented. Then, the development of modelling

techniques related to crash occurrence and crash severity is described.

Finally, the research gaps are highlighted.

Chapter 3: This chapter focuses on data collection and the development of the

methodology for all studies in this research. It starts with an

explanation of the study setting and population. Then, the data

collection process is described in detail. This is followed by the

explanation of a descriptive analysis for crash characteristics study.

The model development and parameters estimates for SV crashes

have been described. This is followed by model development for all

three selected MV crash models, model estimations, goodness-of-fit,

and elasticities. Then, the decision tree and development of crash

severity is described. Finally, health risk assessment and the ethics

statement are discussed.

Chapter 4: This chapter presents the results of Study 1 on crash characteristics

along mountainous highways. The chapter starts with an introduction

to the study and the study objectives, which are followed by a

description of the data. Then, the results and discussion are presented.


Chapter 5: This chapter discusses the results of Study 2 on single-vehicle crashes.

The first section introduces the study and study objective. Then, the

data is described. Finally, the results of the model and the discussion

section are presented.

Chapter 6: The results of Study 3 on multi-vehicle crashes are presented in this

chapter. This chapter has the same arrangement as the previous

chapter, starting with an introduction to the study, study objectives,

data description. Then, the results of the modelling and a discussion

are presented.

Chapter 7: This chapter presents the results of Study 4 which is the crash severity

study. This chapter starts with an introduction to the study, followed

by an explanation of the availability of data. Then, the results of the

model and the discussion section are presented.

Chapter 8: The last chapter in this thesis presents a synthesis of research findings

followed by a discussion of the implications of this research and its

contribution to scientific knowledge. The strengths and limitations of

the research are also discussed in this chapter. Then, the conclusion of

this research is presented and finally, recommendations for future

research are included.


A Conclusions and Recommendations

Figure 1-6: An overview of the thesis

Study 2 Safety Performance Function

(Single-Vehicle Crashes)

Random Parameters Negative Binomial

Study 3 Safety Performance Function

(Multi-Vehicle Crashes)

Random Parameters Negative Binomial

Negative Binomial – Generalized Exponential

Negative Binomial – Lindley

Study 4 Injury Severity Modelling

Standard Logit

Non - Severe crashes Severe crashes

Random Parameters Logit

Skewed Logit (Scobit)

Environmental factorsGeneral Crash Characteristics Driver and vehicle factors

Study 1 Mountainous Road Crash Characteristics

(Descriptive Analysis)

Field survey

National Centers for Environmental

Information (NOAA)

Highway Planning Unit,

Ministry of Works Malaysia and Public Works

Department (PWD), Sabah

Department of Irrigation and Drainage Sabah

(DID)

Topographical Maps

Malaysian Institute of Road Safety Research

(MIROS)

Chapter 2: Literature Review 16

Chapter 2: Literature Review

2.1 INTRODUCTION

According to the WHO Global Road Safety Report, about 1.2 million people

die and up to 50 million people sustain road traffic injuries on the world’s roads

every year (WHO, 2015). Low and middle-income countries account for nearly 90%

of these road fatalities and injuries. Malaysia is a middle-income country and, with a

traffic death rate of 24 per 100,000 inhabitants, is the second worst country in

Southeast Asia in terms of road safety. To tackle this problem, the Malaysian

Government has recently launched a 15 year Road Safety Plan (MOT, 2014).

However, existing research to date is not sufficient to inform government agencies

and road authorities on the best ways to initiate targeted countermeasures. This

scarcity of scientific research is more acute for rural regions of Malaysia than for

urban areas.

This chapter presents a critical review of studies relevant to this research.

Although there are few studies which identify factors contributing to crash

occurrence and injury severity along mountainous highways, this chapter describes

comprehensive studies on other road types that explore the effects of road geometry,

roadside features, road cross-sectional elements, weather conditions, speeding

behaviour on crash occurrence and crash severity.

2.2 ROAD SAFETY ALONG MOUNTAINOUS HIGHWAYS

Constrained topography and complex road geometries are among major issues

in designing and constructing roads in mountainous regions to meet appropriate

roadway standards. This problem is worse in the context of a resource-constrained

developing country. Substandard cross-section elements and dangerous roadside

environments coupled with adverse weather conditions in mountainous areas

significantly increase the level of risk on mountainous highways compared with

highways in flatter areas and generally represent a risky road traffic situation.

In recent years, there have been some studies on road safety along

mountainous highways for crash occurrence (Ahmed et al., 2011; Ahmed et al.,


2012; Chen, Wei, et al., 2011; Chen, 2014; Ma et al., 2015b; Yu & Abdel-Aty,

2014a; Yu et al., 2013; Yu et al., 2015; Zhang & Zhu, 2011) and injury severity

(Chen & Chen, 2013; Li, Wang, & He, 2014; Ma, Chen, & Chen, 2015a; Yu &

Abdel-Aty, 2014a). For example, Wang, He, Sun, and Hu (2009) identified the

effects of geometric features on rear-end crash incidence on mountainous two-lane

highway in Gansu Province, China between 2003 and 2005. Li, Sun, and He (2010)

developed a crash prediction model for mountainous freeways including Yunaan

Luofu freeway, Shaanxi Xihan freeway, and Chongqing Yusui highway. Another

study in China’s mountainous regions was conducted by Wang, Chen, Hu, and Pei

(2010) which focussed on Chengyu freeway in Chongqing, Jingzhu freeway in

Guandong and Shenda freeway. Safety performance along mountainous freeways in

the U.S. has also received attention from road safety researchers. For instance,

Ahmed et al. (2011) obtained vehicle crash data from the I-70 in the State of

Colorado for 6 years (2000-2005), together with roadway geometry, traffic

characteristics, and adverse weather in order to develop safety performance functions

for a mountainous freeway. Using mountainous freeway geometry, real-time

weather, real-time traffic data, and crash data for 3 years (2007-2009) on the same

freeways, Ahmed et al. (2012) investigated the effect of the interaction between these

variables on the occurrence of crashes on a mountainous freeway. Yu et al. (2013)

used Bayesian random effect models to investigate mountainous freeway hazardous

factors along the I-70 using one year crash data. Then, Ma et al. (2015b) divided this

freeway into 104 homogeneous roadway segments with a length of about 1.08 miles

each to model crash rates using Refined-Scale Panel Data for crashes between

January, 2010 and December, 2010. Another study by Yu et al. (2015), also

conducted along the I-70 freeway, used three years of crash data (from Jan, 2008 to

Dec, 2010) to investigate the effect of weather conditions on crash risk. Most of the

studies mentioned above were conducted in developed countries, and findings from

these studies may not be directly applicable to developing countries because of

differences in roadway design, roadside environment, enforcement practice, and

driver behaviour.

In terms of injury severity, Chen and Chen (2013) conducted research to

identify the differences in injury severity of crashes on mountainous and non-

mountainous highways using data from three major interstate highways in Colorado

including the I-70, I-25 and I-76. Using crash data from a mountainous freeway


section along the I-70, Yu and Abdel-Aty (2014a) analysed crash severity

incorporating real-time traffic and weather data for a mountainous freeway. Using

four years of crash data from two major interstate highways that both cross Colorado:

the I-70 and I-25, Ma et al. (2015a) identified different characteristics of injury

severity on mountainous and non-mountainous freeways. In China, research was

conducted by Li et al. (2014) to identify geometric factors contributing to crashes

and injuries on mountainous freeways.

Although there is some research on mountainous road safety, much more

remains to be known about how crashes along mountainous highways are different

from those along highways in flatter areas.

2.3 SAFETY PERFORMANCE FUNCTIONS (SPF)

Safety performance functions (SPFs) have been used in road safety analysis

to predict the average number of crashes per year at a location as a function of

exposure and in some cases, roadway or intersection characteristics (e.g., number

of lanes, traffic control and median type) (AASHTO, 2010). The Highway Safety

Manual (HSM) elaborates on the procedures and application of SPFs. Srinivasan

and Bauer (2013) explore in detail the development of specific SPFs, and many

other researchers in road safety have also used these SPFs to identify factors

contributing to crash occurrence (e.g., Ahmed et al., 2011; Dell’Acqua & Russo,

2011; Ibrahim & Sayed, 2011; Liang, Wan, Zheng, Yang, & Guo, 2014; Oh,

Washington, & Lee, 2009). In addition, a lot of research has been conducted with

the primary focus on developing modelling techniques to derive the SPFs (e.g.,

Lord & Geedipally, 2011; Lord, Washington, & Ivan, 2005; Ma, Kockelman, &

Damien, 2008; Yu & Abdel-Aty, 2013a; Zou, Lord, Zhang, & Peng, 2013). The

main application of SPFs in the project development process has been divided into

three categories: 1) network screening, 2) countermeasure comparison, and 3)

project evaluation (FHWA, n.d.).

2.3.1 Network Screening

SPFs are used by different agencies as a tool to identify the safety performance

of certain locations compared with the average safety performance functions of other

sites that have the same characteristics and exposure. This process is useful in the

safety management process for identifying potential safety improvements.


2.3.2 Countermeasure Comparison

Another application of SPFs is to predict the baseline crash frequency for

certain site conditions and to compare potential countermeasures. In this situation,

the estimation of long-term crash frequency is made using crash history for baseline

conditions (without treatment) and crash modification factors (CMFs) are applied to

estimate the crashes with treatment. This is useful to identify the effectiveness of

each treatment for a location where there are multiple alternatives to address safety

concerns.

2.3.3 Project Evaluation

Another function of SPFs is to evaluate the effectiveness of roadway

improvements. This evaluation focuses on safety in order to provide input for future

planning information for makers, and guidance for programming. Future crash

numbers can be predicted using a SPF developed on the basis of historical crash data

and expected characteristics of future roadways.

2.4 FACTORS INFLUENCING CRASH OCCURRENCE ALONG

MOUNTAINOUS HIGHWAYS

The conventional road engineering approach to road safety has been to

establish relationships between crash frequency and the traffic characteristics,

environmental conditions, and geometry of roadways (Ahmed et al., 2011; Bester &

Makunje, 1998; Fu, Guo, Yuan, Feng, & Ma, 2011; Geurts & Wets, 2003; Goodwin,

2002). In addition, some road engineering studies have examined other contributing

factors to road crashes, such as driver or human behaviour (Uchida, Kawakoshi,

Tagawa, & Mochida, 2010; Wang, Guo, et al., 2011; Yu et al., 2013), vehicle

characteristics (Lin et al., 2013) and improvements in road safety policy (Yu et al.,

2013). All studies have shown the significance of these variables to traffic safety.

This section will focus on factors that contribute to road traffic crashes along

mountainous and non-mountainous highways, including weather conditions, traffic

characteristics, speed, road geometry, cross-sectional elements, roadway and

roadside features and spatial characteristics.

2.4.1 Road Geometry

Roadway geometry is one of the most important factors in traffic safety, as

evidenced by numerous research articles suggesting proper consideration of


geometric factors to avoid negative outcomes such as road traffic crashes (Ahmed et

al., 2011; Bester & Makunje, 1998; Fu et al., 2011; Geurts & Wets, 2003; Goodwin,

2002; Yu et al., 2013). Three aspects of road geometry are explored in this thesis: 1)

horizontal alignment; 2) vertical alignment; and 3) the combination of horizontal and

vertical alignment.

2.4.1.1 Horizontal Alignment

Horizontal curve is one aspect of road geometry that is often highlighted by

researchers when conducting research in road safety. Most studies find that this

alignment significantly contributes to crash occurrence (Kim, Lee, Washington, &

Choi, 2007; Milton & Mannering, 1998; Oh, Washington, & Lee, 2010).

Studies into the effects of horizontal curves on crash occurrence in

mountainous areas show that the existence of horizontal curves along mountainous

roads contributes to crashes (Bauer & Harwood, 2013; Yu et al., 2015; Zhang, Liu, &

Mi, 2010). Eck (1983) states that a driver who is unable to reduce vehicle speed to

negotiate a curve will lose control of their vehicle. Heavy vehicles have been shown

to be more at risk of crashing on curved sections (Chen, Chen, & Wu, 2011). Chen

(2014) also found horizontal curves to be a significant factor in crashes in

mountainous regions in China. In addition, the number of crashes increases when

horizontal curves are located on continuous downgrade segments (Zhang, Liu, et al.,

2010). Curve characteristics also play an important role in crash occurrence. Zhang,

Tang, and He (2010) show that sharp horizontal curves have a positive relationship

with crash occurrence. In contrast, research from Ahmed et al. (2011) found that a

higher degree of curvature appears to be associated with a lower crash rate and other

studies report the same finding using a degree of curvature (Ahmed et al., 2012; Guo

& Sun, 2013). However, Yu and Abdel-Aty (2013b) found that only multi-vehicle

crash rates varied according to the degree of curvature. Road traffic crashes along

mountainous highways are reported to increase with the increase in curve deflection

angles (Ma et al., 2015b). Curve radius has been negatively related to crash rate

(Bauer & Harwood, 2013; Li et al., 2014; Wang et al., 2010). Horizontal curve

length has been found to be negatively related to crash frequency, but only on

tangent and horizontal curves on straight grades (Bauer & Harwood, 2013).

Another factor affecting road safety along mountainous roads is design

consistency. While driving, the driver usually expects consistency in road alignment


(i.e. how the driver actually wants roads to be aligned) and curve design. Studies

have found that using similar horizontal curve radiuses is safer than curves with

different radiuses (Li et al., 2014; Wang et al., 2010). Chen, Wei, et al. (2011)

concluded that if the ratio of curvature radius, curvature change rate and ratio tangent

length are underestimated, crash rates will be higher. They found more crashes occur

when the ratio of an individual curve radius to the average curve radius is between

0.4 and 0.6, and the curvature change rate is between 100 and 190 gon/km, and the

ratio of the individual tangents to the average tangent length is under 0.9. Other

research has found that curvature length ratios were significantly associated with

more multi-vehicle crashes (Yu & Abdel-Aty, 2013b).

In summary, many results indicating the contribution of horizontal curvature

to crash occurrence along mountainous highways have been reported. First, the

existence of horizontal curves in a mountainous area contributes to crash occurrence.

Design characteristics of those curves, such as sharpness, length, and radius can

further increase the danger, although the effect of the degree of curvature is

debatable. Consistency in horizontal curve design along a particular road segment

also has a significant effect on crash occurrence, as drivers can become comfortable

driving along similar and therefore predictable curve radiuses.

2.4.1.2 Vertical Alignment

The variables such as gradient used to represent vertical alignment have

positive relationships with crash occurrence. For instance, Fu et al. (2011) found that

crash rates increase exponentially with the average gradient. This finding is

supported by Chen, Chen, et al. (2011) and Li et al. (2014) who found that graded

curves were associated with more crashes. However, Yu and Abdel-Aty (2013b)

found only single vehicle crashes are positively associated with longitudinal grades.

Guo and Sun (2013) found that longitudinal grades from 6 to 8% and -6 to -8%

percent have significant effects on road crashes, i.e. both upgrades and downgrades.

On the other hand, in some studies, road segments with steep downgrades were

found to exhibit increased crash occurrence (Ahmed et al., 2011; Ahmed et al., 2012;

Li et al., 2014; Wang et al., 2010; Yu et al., 2013; Yu et al., 2015). Trucks are the

vehicles most affected by steep gradient segments (Gui, Wang, & Fang, 2011).

Another factor affecting the crash rate along the vertical alignment is the length of

the section, as crashes have a positive relationship with the slope length (Chen, 2014;


Eck, 1983; Li et al., 2010). In addition, a study conducted by Yuan et al. (2008)

shows that crashes are concentrated in the lower half of downgrade segments. This

finding is supported by research conducted by Wu, Yang, and Mi (2011). However,

it contradicts findings from Zhang, Liu, et al. (2010), who found three obvious crash

peaks near the bottom, middle and top of a continuous downgrade. Upgrade

segments are safer compared to downgrade segments with the same slope range. A

comparison between crest and sag curves show that crest curves are generally

associated with greater crash rate values (Li et al., 2014; Wang et al., 2010; Yu et al.,

2013).

In summary, all available research shows that vertical grades are one of the

contributing factors to road crashes along mountainous roads. However, there is

some debate about the nature of this effect on single and multi – vehicle crashes and

on crash locations along downgrade areas. An increase in gradient will increase crash

occurrence on that section and the risk is increased when this gradient is more than

6% or below -6%. Another vertical alignment factor that contributes to higher crash

occurrence is the length of the segment. A long steep gradient is dangerous to traffic,

and especially heavy vehicles. Meanwhile, due to sight distance problems, crest

locations witness more crashes compared to sag locations.

2.4.1.3 Combination of Horizontal and Vertical Alignment

The relationship between horizontal and vertical alignments and road safety

has not been widely studied. There is little research that focuses on both the

alignments together, and their combined effect on crash occurrence. Wang et al.

(2010) found that crash rates increase in grade and a reduction in the radius of a

horizontal curve. A study conducted by Chen, Chen, et al. (2011) also identified that

a combination of steep grades and sharp curves leads to a higher crash frequency. In

contrast, Zhang, Liu, et al. (2010) found that sections with a large gradient were

helpful to balance the eccentric power of sharp horizontal curves. Meanwhile, Bauer

and Harwood (2013) used different models at different locations along mountainous

roads, such as locations with a combination of horizontal curves with tangents on

straight grades and horizontal curves with tangents at type 1 crest vertical curves

(crest with connected positive and negative grades or vice versa). They found that

crash frequency is higher at locations with a combination of short horizontal curves

with sharp crest vertical curves, as well as at places with a combination of short


horizontal curves with a sharp sag vertical curve. Research conducted by Yu et al.

(2015) reveals that there is a positive correlation between crash rates and steep

downgrade slopes with horizontal curves. Ma et al. (2015a) have reported that steep

gradients and sharp curves along mountainous roads induce different driving

behaviour compared to roads in flatter areas.

To summarise, a combination of these alignments has been found to

contribute to crash occurrence. However, findings also show that the gradient helps

in balancing horizontal force created on horizontal curves, although the effect is

slightly different at different combinations of alignments along mountainous roads.

More research is needed to understand the exact relationship between these two

alignments and their combined effect on crash occurrence.

2.4.2 Cross-sectional Elements

The number of lanes has some impact on crash occurrence, as an increase in

the number of lanes appears to be associated with lower crash rates (Ahmed et al.,

2011; Yu & Abdel-Aty, 2013b). Other significant road cross-section elements related

to road crashes are the presence of a median (Wang et al., 2010) and median width

(Ahmed et al., 2012; Guo & Sun, 2013; Yu & Abdel-Aty, 2013b). All of the research

available to date shows that both median presence and wider medians have a

negative relationship with road crash occurrence. Visibility or sight distance is also

highlighted in some research, as poor visibility has been shown to increase crash

occurrence (Ahmed et al., 2012).

In summary, there are important elements in the cross-section of a road which

need to be considered in road safety studies along mountainous highways. The first

element is the width of the roadway or lane. The literature shows that this element

contributes to reducing crash occurrence along mountainous highways. Moreover, an

increase in the number of lanes will reduce crash occurrence. Presence of median and

the width of the median also have an impact on road safety in mountainous areas.

However, arguments about the significance of these variables have been highlighted

by researchers because of contradictory findings. Greater visibility and shoulder

width also reduce crashes along mountainous roads. Although the results of these

studies are informative, it seems there is little research that specifically examines the

importance of cross-section elements to road safety along mountainous highways. In


particular, the effects of other elements stated by FHWA (2014), such as drainage

channels and side slope (embankment) remain unexplored in this area.

2.4.3 Roadway and Roadside Features

The effect of roadway and roadside features on road safety has been found to

be significant in recent research (Holdridge, Shankar, & Ulfarsson, 2005; Lee &

Mannering, 2002), however, studies on the effects of the roadside features along

mountainous roads are rare. Roadside features are mainly relevant to run-off crashes.

For vehicles involved in run-off crashes, the probability of collisions with roadside

objects, and especially fixed objects, is quite high. Miaou (2013) emphasized that a

model predicting run-off crashes should pay attention to the relationship among

variables related to roadside design, vehicle speed, traffic volume, and vehicle types.

Lee and Mannering (2002) reported that run-off crashes could be reduced by

avoiding cut side slopes, decreasing the distance from the outside shoulder edge to

the guardrail, decreasing the number of isolated trees along roadway sections, and

increasing the distance from the outside shoulder edge to light poles.

On mountainous roads buffer zones are often limited, and therefore roadside

features play an important role in road safety. Wang, Chen, Ci, and Hu (2011)

highlighted the importance of a safe roadside environment and suggest these issues

be considered in freeway design and maintenance in China. Providing safety

facilities along the tops of roadside cliffs and other dangerous sections will help

prevent vehicles becoming involved in serious run-off crashes (Zhi-yun, Shui, &

Zuo, 2013). Zhang, Liu, et al. (2010), in their research about the safety characteristics

of continuous segments, found that the level of roadside hazard tends to lower crash

occurrence. The importance of roadside safety facilities along this type of road is

revealed in research conducted by Zhou, Zhao, Jiang, Zhu, and Zhou (2005), which

showed that after the installation of guardrails, traffic crashes, injuries and fatalities

decreased significantly.

Findings from this review reveal that appropriate safety facilities along

mountainous roads are important to reduce serious crashes involving roadside

features. Some solutions, such as the introduction of guardrails, have been proposed

and trialled in China, and have been shown to be very helpful in reducing crash

occurrence.


2.4.4 Weather Conditions

There are two different types of weather data sets; 1) general weather data,

for example monthly rainfall, and daily precipitation; and 2) real-time weather data,

for example rainfall both upstream and downstream of the crash locations. A variety

of weather measurements have been used in different studies based on their research

objectives and the availability of weather data in their study locations. For example,

Yu et al. (2015) used temperature, visibility (an effect of weather conditions) and

precipitation in their study. Yu and Abdel-Aty (2014a) used snow season indicators

and temperature as their explanatory variables. The effects of weather variables are

not necessarily the same for crash occurrence and injury severity. For instance,

Theofilatos and Yannis (2014) concluded that precipitation affects crash occurrence

but not injury severity.

In mountainous areas, weather plays a vital role in road safety. A recent study

on the relationship between real-time weather and crash occurrence along a

mountainous freeway in the United States demonstrated that, depending on weather

conditions, the same traffic parameters along a mountainous road section might

influence driver behaviour differently (and thus safety) (Ahmed et al., 2012). In this

study, real-time visibility refers to visibility during the 30 minutes before and after

the time of the crash. Visibility could not directly be captured in weather stations but

was calculated based on the measurement of light extinction, which includes the

scattering and absorption of light by particles and gases that were automatically

captured by three weather stations along the I-70 for the same time period. In a

subsequent study that examined hazardous factors involved in single- and multi-

vehicle crashes along the same freeway (I-70) in the United States, Yu et al. (2013)

reconfirmed that crash occurrence along mountainous roads is highly influenced by

weather conditions and suggested adoption of different active management strategies

across different seasons. In particular, visibility and precipitation were reported to

increase crash risks along mountainous sections of the I-70 (Yu et al., 2015). Ma et

al. (2015b) also reported that poor visibility along mountainous roads increases the

crash risk. Once again it is important to point out that all of these studies have been

conducted in the US, and their findings may not be generally applicable to Malaysia,

which is located in the equatorial region and has a tropical rainforest climate.


2.4.5 Speed

Design speed along a stretch of a road must be selected to match the

geometric design elements of that stretch of the highway (REAM, 2002). Drivers

always use a posted speed limit as an indicator for selecting an appropriate speed at

different segments of the road. A study from Li et al. (2010) and Guo and Sun (2013)

reveal that increasing a posted speed limit will increase the crash rate. However,

Ahmed et al. (2011) found that speed limits were not significantly associated with

crash frequency in their model. On the other hand, Eck (1983) confirms that a driver

is likely to lose control of the vehicle and either run off the road or strike another

vehicle when the speed is not properly managed along curvatures. Moreover, Zhang

and Zhu (2011) used the local standard to relate speed and alignment variables. They

concluded that the selection of an inappropriate speed by the driver will lead to crash

occurrence. This is also true of over-speeding behaviour among drivers along these

roads. Over-speeding has been identified as a significant crash contribution factor

along mountainous roads; Zhou et al. (2005) found 31.7% of traffic crashes in this

area come from this behaviour and more recent research confirms this relationship

(Chen, 2014). A safety performance audit conducted in the Chinese province of

Jiangxi indicates that speeding is one of the main causes of run-off crashes along

mountainous roads (Lin et al., 2013). High speed tests the limits of the braking

capacity of vehicles as well as the driver’s ability, and contributes to run-off,

turnover and rear-end crashes (Wang et al., 2010). Some studies have highlighted

speed variance as a crash contributing factor. For example, Hou, Han, Sun, and He

(2010) correlated the crash rate with dispersion of speed. Yu et al. (2013) used real-

time traffic data detected by 30 Remote Traffic Microwave Sensor (RTMS) radars to

show that the average speed of vehicles on the road segment that vehicles were

travelling along during the 5 – 10 minutes prior to crashing is significant.

Speed variation also affects the likelihood of severe crashes along

mountainous roads (Yu & Abdel-Aty, 2014a). Yu and Abdel-Aty (2013b) reported

that the average speed recorded by downstream detectors is significantly associated

with vehicle crashes along mountainous roads. Instead of using the posted speed

limit or average speed, Ma et al. (2015b) used the speed gap, that is, the difference

between the posted speed limit and the mean traffic speed, and found that an increase

in speed gap leads to a higher crash rate. The 85th percentile speed is often the main

interest among studies examining the speed selection behaviour of drivers simply


because of its widespread use as an indicator of the appropriate speed for the road

segments, and its greater sensitivity than mean speed when engineering and other

interventions are implemented. Multiple linear regression models are often used to

examine the relationship between the driving speed and roadway characteristics

(Abbas, Adnan, & Endut, 2011; Gibreel, Easa, & El-Dimeery, 2001; Semeida, 2013).

To summarise, the literature shows that speed has a significant effect on road

crash occurrence. Posted speed limits are a guide for the driver in selecting a safe

speed, and therefore an increase in that speed limit will increase the crash rate.

However, some research found this factor is not related to the crash. Changes in

speed limits along a road segment also have a positive impact in reducing crash

occurrence, particularly on roads with higher traffic density. The selection of a

suitable speed by drivers in a certain area is also important to help reduce crashes.

Driving speed variations are also identified as factors contributing to crashes along

mountainous roads. Posted speed limits are often used as a proxy measure of traffic

speed along road segments, but the posted speed limit may not be a good indicator,

particularly along mountainous roads, because traffic speed on these roads may be

dominated by roadway geometric characteristics and driver perceptions of comfort

and safety (Castro et al., 2012).

2.5 FACTORS INFLUENCING INJURY SEVERITY ALONG

MOUNTAINOUS HIGHWAYS

Constrained topography and complex road geometry along rural mountainous

highways often represent a demanding driving situation. Consequently, traffic

crashes along mountainous highways are likely to have different characteristics to

crashes on highways in flatter areas. Similarly, the variables that impact injury

severity in mountainous highway crashes are substantially different than in non-

mountainous highway crashes (Chen & Chen, 2013). Previous research has identified

several key factors that increase injury severity along mountainous freeways,

including, but not limited to, large speed variance, snow seasons, dry seasons,

daytime driving, steep grades, low temperature, bad visibility and use of passenger

cars (Yu & Abdel-Aty, 2014a, 2014b). In contrast, the presence of a large outside

shoulder reduces the severity of crash (Yu & Abdel-Aty, 2014b). Meanwhile, Ma et

al. (2008) found that a wider road shoulder will reduce severe crashes along

mountainous roads. The presence of guardrails and better road surface are reported to


reduce both fatal and serious injury crashes along a mountainous freeway in

Southwest China (Zhou et al., 2005). Yamamoto and Shankar (2004) found the type

of fixed object to be one of the factors significantly affecting injury severity for both

drivers and passengers. Unprotected ends of guardrails and bridge rails with large

wooden poles increase the probability of fatal injury (Holdridge et al., 2005). Li, Ma,

Niu, and Wang (2008) created a roadside safety protection technology system

(RSPTS) to provide protection to vehicles and minimizes the risk of severe injury to

vehicle occupants, and it was found suitable for use in mountainous areas in China.

Jehle, Connolly, Godzala, and Cole (2010) reported that motor vehicle speed is

a key determinant of the severity of injury in an individual crash in the highway

system. This is supported by research by Dell’Acqua and Russo (2011) which

concluded that speed has a positive correlation with the number of injury crashes per

year per kilometre. In addition, Ma et al. (2008) found that an increase in the posted

speed limit significantly increases all types of injury severity. Speed variation also

affects the likelihood of severe crashes along mountainous roads (Yu & Abdel-Aty,

2014a). In a recent study, Ma et al. (2015a) found eleven factors that increase the

likelihood of severe injury crashes along mountainous highways, including highway

interchanges, driving under the influence of alcohol or drugs, careless/reckless

driving, driving while fatigued, lack of driver insurance/no proof of insurance, multi-

vehicle crashes, speeding, overturn, collision with parked motor vehicle, collision

with embankment and collision with delineator post.

All of the above-mentioned studies were conducted in developed countries

where the road and traffic characteristics barely match those of developing countries.

For example, there are differences in roadway designs, roadside environment, traffic

mix, enforcement schemes, and, most importantly, driver behaviour. Unfortunately,

there are no existing studies that examine the injury severity in rural mountainous

highways in developing countries. Hence, this study could give some important

insights into the underlying factors related to the severity of mountainous highways

crashes in developing countries.

2.6 CRASH MODELLING BY CRASH TYPES

Most crash modelling studies combine all types of crashes in a single model.

However, combining both single (SV) and multi-vehicle (MV) crashes could give


different outcomes. For example, Ivan, Pasupathy, and Ossenbruggen (1999)

conducted research to identify casualty factors for SV and MV crashes on two-lane

roads. They found that increasing traffic intensity (lower level of service - LOS),

shoulder width and sight distance decrease SV crashes, while MV crashes increase

with the number of signals, the daily single-unit truck percentage and the shoulder

width. MV crashes decreased on principal arterials compared to other roadway

classes and LOS was not found to be significant in MV crash models. Geedipally and

Lord (2010) investigated the effect of modelling SV and MV crashes separately on

confidence intervals using Poisson-gamma models. They found that the predicted

confidence interval was larger in separate models compared with modelling SV and

MV crashes in a single model. Research along mountainous freeways the I-70, in

Colorado also found that SV crashes have different crash mechanisms compared with

MV crashes (Yu et al., 2013). Yu and Abdel-Aty (2013b) conducted research using

multi-level Bayesian analysis for SV and MV freeway crashes on the same freeway.

They found that speed limits and longitudinal grades were only related to SV crashes

and degree of curvature, curve length ratios and AADT were only significant in the

MV crash model. Variable lane numbers, median width, and segment length were

significant in both models.

This review shows that modelling SV and MV crashes separately increases the

explanatory power of the model. The results in each model help practitioners to

identify potential countermeasures in tackling different types of crashes. Is should be

noted that there are also studies only focusing on single types of collisions, such as

rear-end (Hosseinpour, Yahaya, Ahadi, Asoode, & Momeni, 2016) and head-on

collisions (Zhang & Ivan, 2005).

2.7 STATISTICAL MODELLING TECHNIQUES

In general, the methodology for the analysis of crashes in road safety can be

divided into two main categories: 1) descriptive analysis; and 2) inference analysis.

Some researchers have used descriptive analysis in their study to describe the

relationship between crash occurrence and injury severity with factors of

contribution (Chen, Wei, et al., 2011; Lin et al., 2013; Wang, Chen, et al., 2011; Wu

et al., 2011; Zhang, Liu, et al., 2010; Zhou, Chen, & Xiang, 2014). They used

histograms, graphs and other statistical figures to present their data. Other research

has applied advanced analysis methods using a statistical approach, and methods


such as the chi-square test and t-test has been used to analyse the proportion between

all groups and to compare measurement data respectively (Zhou et al., 2005). In

addition, inference analysis has been introduced in road safety research with in-depth

analyses of data for identifying factors that contribute to crash occurrence and injury

severity.

One of the most popular methods among road safety researchers is modelling.

Crash count or crash frequency modelling attempts to establish a relationship

between observed crash frequencies and existing geometric, roadway and traffic

conditions of a roadway (Geurts & Wets, 2003). When conducting modelling, there

are two general focuses in crash frequency and injury severity models: modelling

methods and parameters for dependent and independent variables. Modelling

methods can be roughly divided into two categories as follow: 1) general linear

regression model; and 2) generalized regression models such as Poisson distribution

and negative binomial regression (Li et al., 2010).

2.7.1 Crash Modelling

Crash modelling is a process of using advanced statistical techniques to

understand crash-contributing factors and to predict crash frequencies and/or

severities at different intersections and road segments. Crash frequency (also referred

to as crash count) is one of the most common indicators of risk used in the literature.

It has been shown that non-negative crash count observations follow a Poisson

distribution, and as a result, Poisson regression models were used as the first crash

modelling technique in the 1960s (e.g., Kemp, 1967). In the late 1980s, however, it

was argued that crash data are usually over/under-dispersed (i.e. the variance is

more/less than the mean) and the Poisson models are not able to capture this

over/under dispersion. To address this problem, the Negative Binomial (NB)

regression model was introduced in which the variance may not be equal to the

mean. Another analytical challenge confronting crash modelling procedures which

that was highlighted in the 1990s was the presence of excess zeros in crash

observations because crashes are rare events (e.g.,Miaou, 1994; Shankar, Milton, &

Mannering, 1997). This problem particularly arises in circumstances where crash

data are collected in short periods of time (e.g. monthly). The argument revolved

around the fact that there is a vital difference between a safe site (with zero crashes)

and a site (with a non-zero crash risk) that has not experienced any crashes in a

particular period of time. To overcome this analytical challenge, several


methodological approaches have been proposed by researchers including Zero-

Inflated and Markov-Switching models. Zero-Inflated Poisson (ZIP) and Zero-

Inflated Negative Binomial (ZINB) modelling techniques divide the road segment

into two states; zero state (safe state) and non-zero state (normal state) (Dong, Shi,

Huang, Chen, & Ma, 2016; Lee & Mannering, 2002; Lord, Washington, & Ivan,

2007; Malyshkina & Mannering, 2010b; Shankar et al., 1997). However, this model

was criticized because the distribution of the safe state (with zero observations) has a

long-term mean, equal to zero, which is not theoretically plausible (Lord et al., 2007;

Lord et al., 2005). In addition, it is difficult to identify safe road segments from the

ones that have experienced zero crashes because crashes are random events and can

occur in different segments at different times. Consequently, researchers developed a

new model, the Markov switching model, in which they replaced the zero state with

another normal state and assumed that all roadway segments are in the same state at

the same time (Malyshkina, Mannering, & Tarko, 2009). However, some of the road

characteristics (e.g. traffic volume) might vary across time, and this was not

accounted for in such a model. Scholars therefore continued the application of

Markov switching models for zero excess crash data by allowing road segments to

switch between zero and normal state over time (Malyshkina & Mannering, 2010b).

The advantage of the Markov-switching model was that it allows a direct statistical

estimation of the specific road segment’s state, whereas this is not possible in

traditional zero-inflated models. However, model specification (i.e. formulation) of

this model is complex, and so its application to large datasets is computationally

heavy (Behnood & Mannering, 2016). In addition, the combination of Markov-

switching with other models may cause biased estimates, which may in turn lead to

misinterpretation of factors, because the components of the model are restricted to a

single family (Ma, Wang, Yan, & Weng, 2016).

Discussions about crash data with an excess number of zeros have continued

with the introduction of new distributions that are capable of handling observations

with small counts and combining them with the parent distributions for crash data

(i.e. Poisson and NB distributions). Negative Binomial – Lindley (NB-L)

(Geedipally, Lord, & Dhavala, 2012; Lord & Geedipally, 2011), Negative Binomial-

Crack (NB-CR) (Saengthong & Bodhisuwan, 2013), Sichel – also known as Poisson

generalized Gaussian (Zou et al., 2013), Poisson-weighted Exponential (Zamani &

Ismail, 2010b; Zamani, Ismail, & Faroughi, 2014), Poisson Inverse Gaussian (PIG)


(Zha, Lord, & Zou, 2016), Negative Binomial – Generalized Exponential (NB-GE)

(Vangala, Lord, & Geedipally, 2015) and Negative Binomial with Dirichlet process

(Shirazi, Lord, Dhavala, & Geedipally, 2016) are recently developed models for

crash data with excess zeros. However, the Negative Binomial with Dirichlet process

is more flexible in capturing dispersion data where there is a heavy tail and a smaller

percentage of zero observations (Shirazi et al., 2016).

Recently, some studies have compared the performance of NB, ZINB, NB-L

and NB-GE models in dealing with excess zeros data. For example, Lord and

Geedipally (2011) used two crash datasets with excess zeros containing 89% and

90% zero observations respectively using Poisson, NB and NB-L. They found that

the NB-L model had a superior statistical fit to the other models. Furthermore,

Geedipally et al. (2012) used crash data for Indiana and Michigan where 36% and

70% of the observations were zero, respectively. They also found that the NB-L

model outperformed the NB and ZINB models. Vangala et al. (2015) conducted

research using the same data as Lord and Geedipally (2011) and found that the NB-

GE model was comparable with the NB-L, and significantly outperformed the

traditional NB model. However, they found that the NB-GE resulted in smaller

standard errors for parameter estimates of crash contributing factors, implying that

this model may be capturing more variance in parameters than the NB-L model.

Another important challenge in crash modelling, which was noticed in the late

2000s, is the unobserved heterogeneity in crash data. This property of the data stems

from the fact that crashes are not homogeneous across all highway segments and

during all time periods. There are two sources of heterogeneities; structured and

unstructured. Structured heterogeneity may arise from spatial correlation among

crashes across the network. Temporal correlations of crash data can also create

heterogeneity in a panel data setting where one segment is observed in multiple time

periods. The second source of heterogeneity is unstructured heterogeneity, which

comes from model misspecification, uncertainty in covariates, and omitted

independent variables. These two sources of heterogeneities may lead to estimation

bias in regression coefficients and ultimately lead to incorrect inferences about model

parameters. One promising solution to overcome this issue is the Random Parameters

(RP) modelling approach in which regression parameters are allowed to vary across

observations (different locations and/or different time periods) (Mannering, Shankar,


& Bhat, 2016). Anastasopoulos and Mannering (2009) found that the Random

Parameters Negative Binomial (RP-NB) model provides superior overall fit relative

to the standard NB model. Ma et al. (2015b) showed that the Random Parameters

Tobit model could properly deal with serial correlations in panel data and left-

censoring effects. Despite recent advances in crash modelling methodologies, there

have been few studies to compare the prediction performance of the Random

Parameters models for panel data with excess zeros. Furthermore, although the NB-L

model is sometimes identified as a form of RP model where the only random

parameter used is the constant term (Geedipally et al., 2012), it is still required to

compare a full RP model with other candidate models to identify an appropriate

model in handling crash data with heterogeneity and an excess number of zeros.

2.7.2 Injury Severity Modelling

Injury severity models attempt to establish the relationship between injury

severities and various roadway geometric, traffic and driver/vehicle related factors.

Injury severity models are different from crash count models because discrete

outcomes will be generated, such as fatality, serious injury, light injury and property

damage only. There are many methodological approaches that have been used by

researchers to analyse crash-injury severity (Savolainen, Mannering, Lord, &

Quddus, 2011). For example, a multinomial logit model was widely applied in road

safety research because it can consider three or more outcomes (Carson &

Mannering, 2001; Lee & Mannering, 2002; Shankar, Albin, Milton, & Nebergall,

2000). An ordered probit model has also been introduced, with the argument that

injury severity is discrete and in a natural order (Abdel-Aty, 2003; Garrido et al.,

2014; Kockelman & Kweon, 2002; O'Donnell & Connor, 1996). It is claimed that a

multinomial logit and probit model as well as a nested logit would fail to account for

the ordinal nature of the injury classes. However, Jung, Jang, Yoon, and Kang (2014)

examined the multinomial and ordered probit models in their study on injury severity

in single vehicle crashes and found that the multinomial logit model is preferable to

the ordered model of injury severities. Meanwhile, nested logit has been introduced

because the ordered model has a restriction on the effect of the explanatory factors,

causing those factors to either increase the probability of greater severity or increase

the probability of lesser severity (Holdridge et al., 2005).

Furthermore, a mixed logit model solves the problems of the multinomial

model and considers the random effects of variables (Chen & Chen, 2013). In


addition, Milton et al. (2008) argued that every geographical location has a different

effect on injury severity outcomes, which supports the mixed logit model.

Meanwhile, Ye and Lord (2014) identified that, for small samples research, an

ordered probit model was the best fit of the three models: multinomial logit, ordered

probit and mixed logit. In 2014, Yu and Abdel-Aty (2014a) compared three different

models; 1) the fixed parameter logit model; 2) the support vector machine (SVM)

model with radial-basis kernel function to detect non-linearity; and 3) the random

parameter logit model with an unrestricted variance-covariance matrix to account for

individual heterogeneity, and also to investigate the potential correlations between

the explanatory variables. They found that the SVM model and random parameter

model provide more superior fits than the fixed parameter logit model.

In summary, the choice of analysis method or modelling can give different

results for the same variables. For example, Theofilatos and Yannis (2014) in their

review found that using different count regression models to examine the effect of

traffic flow gives different results. This shows that selecting a proper model is

important to answer the research question accurately.

2.8 IDENTIFIED RESEARCH GAPS

The above literature review has identified research gaps in two main areas: 1)

road safety issues along rural mountainous highways, and 2) crash modelling

methodology.

2.8.1 Road Safety along Rural Mountainous Highways

The literature shows that rural mountainous roads have higher fatality rates

compared to flat areas (Ahmed et al., 2011; Rautela & Pant, 2007; Wang et al., 2009;

Zhang, Tang, et al., 2010). Based on in-depth literature review, the following

problem areas and gaps are identified.

Much road safety research, both for crash occurrence and injury severity,

focuses on locations with a higher crash record. These locations mostly belong to the

non-mountainous areas because of high traffic volumes in these areas. Appendix B

shows the previous road safety studies conducted in mountainous and non-

mountainous areas. It is evident that road safety research identifying the effect of

roadway geometric factors on crash occurrence and roadside features on injury

severity has mainly been conducted on highways in non-mountainous areas.


However, mountainous area records show a higher fatality index, probably due to the

complexity of their topography and the dangerous roadside environment, but very

limited research has been conducted for mountainous highways.

To increase understanding about crash characteristics along rural

mountainous roads, the following research objectives and research questions have

been developed in this research.

RO1: To examine the characteristics of road traffic crashes on rural mountainous

roads and to compare these with the characteristics of crashes on non-

mountainous roads.

RQ1: What are the characteristics of road traffic crashes along rural

mountainous roads?

The few studies on road crashes along mountainous roads have only

considered a few variables such as the radius of horizontal curves, curvature length,

and gradient. Appendix B shows a list of variables used in previous studies along

both mountainous and non-mountainous roads. In order to understand traffic safety

along mountainous highways, a comprehensive set of variables need to be considered

and tested in crash prediction models.

There are some conflicting findings in the relationship between traffic safety

and geometrical factors. For instance, some research (e.g., Ahmed et al., 2011;

Ahmed et al., 2012; Guo & Sun, 2013) indicated that a higher degree of curvature is

associated with lower crash rates. However, Yu and Abdel-Aty (2013b) found that

only multi-vehicle crash rates relate to the degree of curvature. For vertical

alignment, a study conducted by Yuan et al. (2008) shows that traffic crashes are

concentrated in the latter half of downgrade segments. This finding is also supported

by research conducted by Wu et al. (2011). However, it contradicts the research

conducted by Zhang, Liu, et al. (2010), which reported three obvious crash peaks

near the bottom, middle and top of continuous downgrades. Moreover, a study

conducted by Chen, Chen, et al. (2011) showed that a combination of steep grades

and sharp curves resulted in a higher crash frequency. In contrast, Zhang, Liu, et al.

(2010) found that downgrade or upgrade sections is helpful to balance horizontal

force when vehicles are entering the horizontal curve. Therefore, it is very important

to establish a comprehensive model so that the relationship between geometrical

factors and road safety along rural mountainous roads can be investigated. In


addition, the relationship between each explanatory variable should be identified as

well.

Another important research gap is that most of the crash occurrence studies

analysed single (SV) and multi-vehicle (MV) crashes in a single crash data.

However, these types of crashes are quite different to each other. For example, head-

on and rear-end crashes are the most common MV crash types, while run-off-road

and hit-object crashes are common in SV crashes. Thus, there is reason to separate

MV from SV crashes and analyse them separately, as in some studies in the literature

(e.g.,Geedipally & Lord, 2010; Islam, Jones, & Dye, 2014; Yu & Abdel-Aty, 2013b;

Yu et al., 2013).

Speed is one of the most important parameters of road safety and has a direct

relationship with crash occurrence. Posted speed limits are often used as a proxy

measure of traffic speed along road segments, but the posted speed limit may not be

a good indicator, particularly along mountainous roads because traffic speed on these

roads may be dominated by roadway geometric characteristics and driver perceptions

of comfort and safety (Castro et al., 2012). In fact, Ahmed et al. (2011) reported that

crash frequencies along mountainous roads are not significantly associated with

posted speed limits. In contrast, Yu and Abdel-Aty (2013b) reported that the average

speed recorded by downstream detectors is significantly associated with vehicle

crashes along mountainous roads. Instead of using posted speed limit or average

speed, Ma et al. (2015b) used the speed gap—the difference between the posted

speed limit and the mean traffic speed—and found that an increase in speed gap

leads to a higher crash rate. In addition, the potential energy, which transfers to

kinetic energy while a vehicle is travelling from the top to the bottom of a gradient,

also impacts on speed selection. Drivers sometimes lose control of their speed and

steering, which can lead to a crash occurrence; however, the effect of driving speed

on crashes on mountainous highways is not very well known.

Another important research gap is the need for the development of Safety

Performance Function (SPF) tools for rural mountainous highways. It is strongly

suggested that SPFs for different areas and conditions should be developed explicitly

using the dataset from the corresponding jurisdiction for greater reliability (Brimley,

Saito, & Schultz, 2012; Young & Park, 2013). Until now, no SPFs have been

developed along rural mountainous roads in developing countries, so developing


SPFs in these areas is one way to further explore the significant roadway geometry

and traffic related variables related to road crashes along mountainous highways.

Most of the research along mountainous highways was conducted in western

countries such as the United State. Findings from such developed countries may not

be directly applicable, as developing countries are likely to have differences in

roadway designs, roadside environment, presence of roadside furniture, traffic mix,

enforcement practices, and, most importantly, driver behaviour. In addition, traffic

safety research in developing countries often suffers from limited data due to low

levels of police reporting of crashes and poor data quality.

To address the above research gaps, two research objectives and two research

questions have been developed for this research.

RO2: To investigate the effects of roadway geometries, traffic characteristics, real-

weather conditions, cross-sectional elements, roadside features, and spatial

characteristics on Single-vehicle (SV) crashes along rural mountainous

highways.

RO3: To examine critical factors contributing to Multi-vehicle (MV) crashes along

rural mountainous highways.

RQ2: What are the factors that lead to SV crashes along rural mountainous

highways?

RQ3: What are the factors that contribute to the occurrence of MV crashes

along rural mountainous highways?

The injury severity of crashes along mountainous highways have not received

much attention in the past. Some previous studies (e.g., Wang et al., 2009; Yu &

Abdel-Aty, 2014a, 2014b) have only focused on particular crash types/injury

severity, such as rear-end crash incidence on two-lane mountainous highways.

However, comprehensive injury modelling is needed in order to develop an in-depth

understanding of the injury severity of crashes along rural mountainous highways

and to identify significant road geometry factors, roadside features and weather

conditions. This will help in designing different control strategies to reduce different

types of crashes, and to reduce injury severity in the case of a collision along



RO4: To investigate the injury severity of road traffic crashes along the rural


RQ4: What are the factors that are related to the crash severity of traffic

crashes along rural mountainous highways?

2.8.2 Crash Modelling Methodology

First, unobserved heterogeneities represent a major challenge in developing

SPFs. Heterogeneities could be structured or unstructured depending on the sources

they arise from. In the context of this research, structured heterogeneities may result

from data clustering or because of temporal correlations, as the same road segment

was observed for multiple time periods. The NB model cannot take into account

location specific effects and potential serial correlation associated with the use of

time-series cross-sectional panel data for crashes in this study. This may lead to

incorrect inferences of model parameters as the estimated standard errors of

regression coefficients may be underestimated. On the other hand, unstructured

heterogeneities may arise from model misspecifications, uncertainty in exposure and

covariates, and omitted variables.

Second, the crash data also has problems with the presence of excess zeros in

crash observations (e.g.,Miaou, 1994; Shankar et al., 1997). Despite the wide

applications of random parameters models to address heterogeneities arising from

various factors like road geometrics, traffic characteristics, socioeconomic factors

and driver behaviour, their capability of addressing the excess zeros problem is still

not known.

Finally, injury severity models attempt to establish the relationship between

injury severity and various contributing factors including driver, traffic and vehicle

characteristics, weather conditions, road geometry, roadside features, and crash

types. Injury severity models are different from crash-count-models due to their

discrete outcomes. Some frequently reported outcomes are: fatality, serious injury,

slight injury, and property damage only (PDO). Many studies have combined these

outcomes into two categories (severe and non-severe) due to small counts for certain

severity levels (e.g., Yu & Abdel-Aty, 2014a, 2014b). The standard binary logit

model is mostly used to model these binary outcomes. However, a low share of

severe injuries may create an imbalance in the response variables. A proper selection

of a model that can handle this problem can avoid misinterpretation of the modelling


results. Another issue related to injury severity modelling is the specification errors

where a model represents an incorrect relationship between dependent and

independent variables, non-additive effects, and nonlinearities. To avoid these errors,

a two-step modelling approach has been suggested, in which the interactions among

a set of variables from the decision tree analysis will be used in the logistic

regression model along with other prospective variables to improve the model’s

predictive power.

2.9 CHAPTER SUMMARY

This literature review briefly discusses the existing research related to the

thesis topic. It has looked at previous research conducted along mountainous

highways, the explanation of safety performance functions, factors influencing crash

occurrence and injury severity, including road geometry, cross-sectional elements,

roadway and roadside features, weather conditions and speeding behaviour. This

chapter also explores the important aspects of separating models for SV and MV

crashes. The recent development in crash count and injury severity modelling was

also discussed. Finally, the identified research gaps were highlighted.

Chapter 3: Methodology and Data 40

Chapter 3: Methodology and Data

This chapter provides a brief description of the plan and methodologies in

this research to achieve the aims and objectives stated in Section 1.3 of Chapter 1.

The current research consists of four distinct studies. Section 3.1 discusses the study

setting and population; Section 3.2 discusses the data collection process; Section 3.3

outlines the analysis technique used in Study 1; Section 3.4 describes the modelling

technique for single-vehicle crashes; Section 3.5 elaborates on the procedures in

modelling multi-vehicle crashes; Section 3.6 explains the crash severity modelling.

Then, Section 3.7 explains the health risk assessment and ethical considerations

relevant to the research and Section 3.8 summarises all of the sections included in

this chapter.

3.1 STUDY SETTING AND POPULATION

Sabah is one of the 13 member states of Malaysia. The total area of Sabah is

73,632 square kilometres with a population of approximately 3.2 million people in

2010 (DOSM, 2010). Relevant details about the district of Sabah and its

characteristics are shown in Table 3-1. Statistics from the Ministry of Transport

Malaysia in 2010 showed that the total number of motor vehicles registered in Sabah

was 863,181(MOT, 2010). Details about the total number of vehicles registered and

new registration by type are shown in Table 3-2. The total length of roads in Sabah is

22,646.47 km, including federal roads, state roads and local government roads (HPU,

2013). Figure 3-1 shows the location of Sabah in the Asian region, while Figure 3-2

shows the topography and location of the roads that will be studied.

In Malaysia, roads have been mainly divided into two groups: urban and rural

roads (REAM, 2002). Urban roads are subdivided into four categories, namely

expressways, arterial roads, collector roads, and local roads; and roads located in

rural areas are subdivided into five categories, namely expressways, highways,

primary roads, secondary roads, and minor roads. Based on the road authority’s

definitions, federal roads are those constructed and maintained by the Ministry of

Works in Malaysia through funding from the federal government. Federal roads are


mainly rural highways which contribute to nearly 50% of injury crashes in Sabah

(MIROS, 2014).

Figure 3-1: Location of Sabah in Malaysia. Source: adapted from google maps

Table 3-1: List of District in Sabah

No District Area (km2) Population (‘000)

(Census 2010)

Crash Frequency

(2008-2012)

1 Kota Kinabalu 351 436.1 23,444

2 Tuaran 1,166 97.8 2,087

3 Lahad Datu 7,444 213.1 4,103

4 Tawau 6,125 402.4 7,368

5 Kota Belud 1,386 89.2 1,942

6 Kudat 1,287 85.4 851

7 Papar 1,243 111.4 2,811

8 Keningau 10,969 261.4 2,998

9 Tenom 2,409 54.4 529

10 Beaufort 2,188 95.9 1,690

11 Sandakan 2,266 453.5 8,882

12 Semporna 1,145 140.4 921

13 Penampang 465 159.6 5,697

14 Kunak 1,134 72.0 1,063

15 Kota Marudu 3,336 114.2 2,204

16 Ranau 3,608 88.8 1,775

17 Beluran 7,719 105.4 1,661

18 Sipitang 2,732 35.5 332

19 Kinabatangan 16,659 197.6 1,800

Total 73,632 3,214.1 72,158

Source: DOSM (2012);MIROS (2014)


Table 3-2: Total Registered Vehicles and New Registrations in Sabah for 2010

Types of vehicle Total Registered New Registration

Amount Percentage (%) Amount Percentage (%)

Passenger Car 487,510 56.5 35,310 51.0

Motorcycle 204,662 23.7 26,898 38.9

Bus 6,783 0.8 186 0.3

Taxi 5,096 0.6 116 0.2

Hire and Drive Car 1,233 0.1 96 0.1

Goods Vehicle 104,495 12.1 4,078 5.9

Others 53,402 6.2 2,617 3.8

Total 863,181 100.0 69,301 100.0

Source: MOT (2010)

Figure 3-2: Topography Maps of Sabah and Selected Roads for Study. Source: adapted from Google maps

3.2 DATA COLLECTION

To achieve research objectives and answer the research questions, this research

used two sources of dataset: 1) primary datasets through field surveys, and 2)

secondary datasets obtained from relevant agencies and authorities in Malaysia

including crash data obtained from the Malaysian Institute of Road Safety Research

(MIROS), topographical information obtained from the Department of Mapping and

Survey Malaysia (JUPEM), weather information obtained from the Department of

Irrigation and Drainage Sabah (DID) and National Centers for Environmental

Information (NOAA), and traffic volume information obtained from the Highway


Dongongon

Kimanis Beluran

Telupid

Ranau

Tamparuli

Keningau

Tambunan Roundabout

Site 1

Site 3

Site 2

Planning Unit under the Ministry of Works Malaysia and Public Works Department

Sabah (PWD).

3.2.1 Selection of Mountainous Highways

The dataset for this study was collected for four highways passing through

rural mountainous areas in Sabah, Malaysia, including 1) Kimanis – Keningau

highway, 2) Penampang – Tambunan highway, 3) Tamparuli – Ranau highway, and

4) Ranau – Sandakan highway. These highways were mainly two-lane two-way

roads (approximately 99% of total highway length), and the posted speed limit along

these roads was mainly 90 km/h. Topographical information for these highways was

obtained from the Digital Terrain Model (DTM) provided by the Department of

Survey and Mapping Malaysia. The geographical information system (GIS) software,

ArcGIS was used to overlap road maps with topographical information. Following

the geometric design guidelines of Malaysia, mountainous highway sections were

selected as those located in areas where the average natural ground slope is more

than 25% (REAM, 2002). As a result, mountainous highways in Sabah were divided

into three study sites; 1) Kimanis – Keningau, 2) Penampang – Tambunan, and 3)

Tamparuli – Beluran with a total length of 312.3 km. Figure 3-3 shows the locations

of study sites along the four selected rural mountainous highways in Sabah.

Figure 3-3: Location of Study Sites


3.2.2 Road Segmentation

Segmentation of the road is important to target important variables and answer

the research questions. For this research, road segmentation has been made for the

whole road based on coordinates from the first site visit (please refer Appendix C-

G). Table 3-3 shows a list of highway segments for each road study area. The

identified sections were then segmented mainly based on guidelines provided by

AASHTO (AASHTO, 2011). Three main criteria for segmentation include:

1. Presence of major intersections.

2. Changes in number of lanes.

3. Changes in longitudinal grades of more than 2%.

These resulted in about 375 mountainous highway segments along the four

highways mentioned above. Mountainous highways in Sabah are full of different

types (e.g. simple, compound, reverse, broken-back) of horizontal curves with large

variations in curve radius. Horizontal alignment was not used as a criterion for

segmentation because it may lead to segments with very short lengths.

Table 3-3: List of Highway Segments based on Segmentation Criteria

Item From Kimanis Penampang Tamparuli

Total To Keningau Tambunan Beluran

No. of Lane 2 3 2 3 2 3

Gradient

0 to < 2 , 0 to >-2 17 0 15 0 97 0 129

2 to < 4, -2 to > -4 6 1 11 0 47 0 65

4 to < 6, -4 to > -6 7 0 21 0 45 1 74

6 to < 8, -4 to > -8 7 4 16 0 26 8 61

8 to < 10, -8 to >-10

5 3 4 1 7 7 27

>=10, <=-10 0 15 1 0 1 2 19

Total Segments 42 23 68 1 223 18 375

Overall Total Segments

65 69 241 375

Total Length (km) 48.3 62 202 312.3

3.2.3 Sampling Technique

Using a randomly assigned draw number, 102 out of these 375 segments were

selected for detailed data collection through field surveys. The characteristics of the

selected samples in terms of gradient and number of lanes were similar to those of


the 375 segments reported in Table 3-3. The crash mapping uses the location variable

in M-ROADS as an indicator in selecting segments.

3.2.4 Primary Datasets

3.2.4.1 Field survey information

Roadway geometric, cross-sectional elements, roadside features and spatial

characteristics data were mainly collected by field surveys, as these data were not

readily available. To capture the data for horizontal and vertical alignments, the

survey team took GPS coordinates (x, y, and z) every five meters along the road

segment using a handheld GPS device (Garmin Etrex 10) (see Figure 3-4). The

accuracy of this device is +/- 3m (Garmin, 2011). GPS coordinates at two lateral

points were recorded every five meters along the segment (See Figure 3-5). These

coordinates were used to construct the whole road segment in AutoCAD 2015.

Figure 3-6 shows details of plotting in AutoCAD after transfer coordinates at both

sides along the segment to the centre line. To crosscheck accuracy, the constructed

segments were overlapped with the map available in Google Earth (Google, n.d).

Cross-section elements like road width and shoulder width were measured by

measurement wheels, and longitudinal distances of these elements were extracted

from GPS coordinates. In addition, both directions of each segment were driven and

filmed with a video camera. These videos were extremely helpful, not only for

collecting data on roadside features like pavement marking, signs, and guardrails, but

also for crosschecking data accuracy for almost all variables. Spatial characteristic

data, such as the number of houses and commercial developments was manually

collected and crosschecked with the aerial view on Google Earth. Figure 3-8 shows

one of the selected highway segments along the Kimanis – Keningau Highway.


Figure 3-4: Handheld GPS Garmin Etrex 10

Figure 3-5: A typical horizontal curve along a mountainous road segment.


Figure 3-6: Details Plotting in AutoCAD 2015

Figure 3-7: Measuring cross-sectional elements using measuring wheels


Figure 3-8: One of the selected highway segments along the Kimanis – Keningau Highway

Variables related to horizontal alignment included a proportion of segment

length with horizontal curve, a proportion of segment that was simple curve,

compound curve, reverse curve, or broken back curve, maximum and minimum

degrees of curvature, curve radius, length of circular curve, and length of tangent.

Variables related to horizontal alignments are shown in Figure 3-5, and the details of

horizontal curve types in a typical mountainous highway segment are illustrated in

Figure 3-9. In Figure 3-5, the length of the tangent refers to the distance along the

tangent from the Point of Intersection (point where the back and forward tangents

intersect) to the point where the circular curve begins. A deflection angle or an

intersecting angle is the amount of angle change from the first tangent line to the

second tangent line. The proportion of segment length with horizontal curve

represents the length of highway segment with horizontal curve over the total length

of the segment. Gradient was calculated as the change in vertical distance divided by

the change in horizontal distance times 100%. Vertical distance is the different of z

coordinates between two points. The Pythagorean theorem was used to compute the

horizontal distance between two points from the corresponding vertical and

measurement distances. Variables for longitudinal grades include a proportion of

segments with longitudinal grades greater than zero, intensity of changes in vertical

alignment (number of vertical curves per km) and indicators for different levels of


longitudinal grades ranging from 2% to 8%. A similar range of longitudinal grades

was also used in Ahmed et al. (2011).

Figure 3-9: Different types of horizontal curves in a typical road in mountainous areas

Explanatory variables related to cross-section elements include a proportion

of segment lengths with a wide shoulder, a proportion of segments with a concrete

shoulder, bitumen shoulder or unpaved shoulder, a proportion of segments with

unbroken centre lines, rumble strips, marginal strips (the area between edge line and

edge drop of pavement) more than 0.5m wide, edge drop-offs of more than 100mm

and the presence of an overtaking lane. In the context of two-lane two-way

mountainous highways in Sabah, an additional lane is only provided occasionally for

overtaking, therefore the number of lanes is not included as an additional variable.

Roadside features include a number of minor intersections, trees, roadside culverts,

or lighting poles per km, proportion of segment with embankments, cliffs, or

guardrails, presence of bridge and presence of road delineation, spatial characteristics

which include the proportion of segments with forest, farm/agricultural land, or

roadside housing and commercial premises. In addition, the number of houses or

commercial premises within 100m of the road edge was also captured in order to

examine the effect of adjacent land used for mountainous road safety.


3.2.4.2 Speed

A posted speed limit may not be a good indicator of driving speed, so a two-

hour spot speed study was conducted for each highway segment. Two speed related

variables were created from this speed survey data: upgrade and downgrade speeding

indicators. For instance, a downgrade speeding indicator refers to a condition when

the 85th percentile speed along a downgrade segment is greater than the posted speed

limit. SV and MV crashes studies were used for this variable.

Methods of Spot Speed Study

1. Speed detection stations were set up for each selected highway segment. 2. Two-hour observations for each speed station were conducted. This technique

was adapted from research by Shankar and Mannering (1998) in their study on lane speeds and speed deviations.

3. A proper radar gun was used for data collection for every point. 4. The radar guns were operated by research assistants in a vehicle parked on

the side of the road or in the bush outside the road shoulder, so as not to impact the natural behaviour of the drivers.

5. Only free-flow vehicles, with leading headways of at least 5s, were considered when collecting individual speed data (Fitzpatrick, Miaou, Brewer, Carlson, & Wooldridge, 2003; Hashim, 2011).

6. Data was collected under normal conditions (dry roadway, no adverse weather, adequate sunlight) during daytime (Morris & Donnell, 2014).

7. Speed study forms were filled in (Appendix K). 8. Ogive (cumulative of percentage vs. speed) for each section were then

presented to get the 85th percentile speed (refer to Figure 3-10).

Figure 3-10: Ogive for 85th percentile of a sample of speed

85%

79km/h


3.2.5 Secondary Datasets

3.2.5.1 Crash Data

Five years of road crash data from 2008 to 2012 were obtained from the

Malaysian Institute of Road Safety Research – Road Accident Analysis and Database

System (M-ROADS) (MIROS, 2014). The main source of this data is from the

Malaysian Royal Police. The Malaysian Royal Police use the POL27 (Appendix A)

to report road traffic crashes. This form contains more than 63 pieces of information,

including a detail report and the time of crash, road information, environmental

information, crash location, vehicle information, driver information, comments,

sketch of the crash and sketch of crash location. Despite the variety of information

recorded for each crash, information on all variables above is not available in M-

ROADS, particularly for many property damage only crashes. In total, only 30 types

of information/ variables from M-ROADS were used in this study. Table 3-4 shows

the available information/ variables in M-ROADS related to crashes at the study

sites.

Table 3-4: List of available information/ variables in M-ROADS

No. Variables Description Availability for

crash sites A. Details report / Time of

crash

1. State The state in which the crash occurs. For this research, only crashes in Sabah were used.

Yes

2. District There are 19 different districts in Sabah state. Seven districts were involved in this research including Tuaran, Papar, Keningau, Sandakan, Penampang, Ranau and Beluran.

Yes

3. Police station This study uses crash reports from 9 of the 32 police stations in Sabah: Balai Polis Tuaran, Balai Polis Tamparuli, Balai Polis Papar, IPPD Keningau, Balai Tambunan, Balai Polis Sandakan, Balai Polis Penampang, Balai Polis Ranau, and Balai Polis Togod.

Yes

4. Report number This is a unique number that identifies each crash

Yes

5. Year Only crashes from 2008 to 2012 were considered in this research.

Yes

6. Month January to December Yes7. Date Date of crash8. Time The time of crash was recorded in 24-h

time.Yes

9. Day Sunday to Saturday Yes



crash sites 10. Number of vehicles involved Total number of vehicles involved in the

crashYes

11. Number of vehicles damaged

Total number of vehicles damaged in the crash

Yes

12. Number of drivers killed Total number of fatalities among the drivers in the crash

Yes

13. Number of drivers injured Total number of injuries among the drivers in the crash

Yes

14. Number of passengers killed Total number of fatalities among the passengers in the crash

Yes

15. Number of passengers injured

Total number of injuries among the passengers in the crash

Yes

16. Number of pedestrians killed Total number of fatalities among the pedestrians in the crash

Yes

17. Number of pedestrians injured

Total number of injuries among the pedestrians in the crash

Yes

18. Type of accident/ crash The four types of crash are fatal, serious injury, slight injury and property damage only.

Yes

B. Road Information 19. Road surface type There are five types of road surface

including crasher run (gravel), interlocking block (brick), bitumen/ Tar pavement, concrete pavement and earth surfacing.

No

20. Traffic system There are four categories of traffic system including one way, two way, three lane and dual carriageway.

No

21. Road geometry There are seven categories of road geometry including straight, bend, roundabout, cross section, T/Y junction, staggered junction and interchange.

Yes

22. Quality of road surface There are four qualities of road surface including smooth, corrugated, potholed and rutted.

No

23. Road condition This variable relates to the longitudinal grade and is divided into two categories, flat and slope road.

No

24. Line marking There are six categories of line markings: double, single, one way, divider (median), U – Turn and no marking.

No

25. Hit run This variable relates to information about the act of causing the crash and not stopping afterwards.

No

26. Control type There are nine categories of control types: police, other agencies, traffic light, pedestrian crossing, pedestrian crossing with traffic light, train crossing, yellow line, yellow box and no control.

No

27. Road width The width of road measured in metres. No 28. Shoulder width for both

sides The width of the shoulder on both sides, left and right.

No

29. Type of road shoulder Road shoulder can be paved or unpaved. No 30. Road defect There are twelve categories of physical

defects that may be present: shoulder drop/ raise, main hole drop / raise, loose gravel, dusty road, pothole, polished surface, defective traffic light, narrow railway crossing, narrow bridge, no guard rails, no/insufficient street lights and no

No



crash sites defect.

31. Speed limit There are six categories of speed limit in this variable: 50, 70, 80, 90, 110 km/h and Other.

No

32. Road surface condition There are six categories of road surface condition: dry, flood, wet, oily, sandy and under reconstruction.

No

33. Collision type There are thirteen collision types: head-on, rear-end, right angle side, angular, side swipe, forced, hitting animal, hitting object off road, hitting object on road, hitting pedestrian, overturned, out-of-control and others.

Yes

C. Environmental Information 34. Weather conditions There are three categories of weather

condition available: clear, foggy and rain. No

35. Light conditions There are four categories of lighting conditions: day, dawn/dusk, dark with street lighting and dark without street lighting.

No

D. Crash Location 36. Road type Road types are divided into 5 categories:

expressway, federal road, state road, municipal and others.

Yes

37. Route no. Each of the gazetted roads in Malaysia has its own unique route number.

Yes

38. Type of location There are four categories of location types: city, urban, built-up area and rural area.

No

39. Type of area There are seven categories of area types: residential, office, commercial, construction/industrial, bridge/foot bridge, school and others. The information about the nearest kilometre post and the distance from nearest location also included in this variable.

No

E. Vehicle Information 40. Vehicle Make and Model The make and model of each vehicle

involved in the crash.No

41. Year of Manufacture The year of manufacture for each vehicle. No42. Registration number The unique registration number of each

vehicle.No

43. Type of vehicle There are twenty types of vehicle: express bus, stage bus, factory bus, mini bus, tour/excursion bus, school bus, four-wheel drive, special duty vehicle, bullock cart, lorry trailer, rigid lorry, small lorry, passenger car/wagon, motorcycle > 250 cc, motorcycle <251 cc, taxi, trishaw, van, hired car and bicycle.

Yes

44. Type of ownership There are six types of ownership: private, goods, service, government, police and army.

No

45. Part of damage There are seven categories of this variable: no damage, front, rear, left side, right side, top and various damage.

No

46. Vehicle movement This variable represents the movement of vehicle during crash. This variable has twelve categories: parked, suddenly stopped, diverging, converging, slippery,

No



crash sites right turn, left turn, overtaking, U-turn, forward, reverse and others.

47. Vehicle defect There are eleven factors related to vehicle defects that contribute to crash. These are brake malfunction, broken windscreen, vehicle without light, light not working, steering, old tyres, recycled tyres, bald tyres, wipers malfunction and not related/relevant.

No

48. Vehicle modified This variable clarifies either vehicle was modified or not.

No

49. Length of brake marks The length of brake marks on the road in metres. Brake marks only appears in some conditions.

No

50. Tyre burst This variable identifies whether any tyres were burst.

No

51. Foreign vehicle This variable identifies any foreign vehicles from Singapore, Thailand or Brunei and Diplomatic vehicles.

No

F. Driver Information 52. Sex The gender of the driver. Yes 53. Aged The age of the driver in years. Yes 54. Race There are eleven categories for this

variable: Malay, Chinese, Indian, Kadazan, Murut, Melanau, Bajau, Bidayuh, Iban, Foreigner and others.

Yes

55. Licence process Three types of process for obtaining a licence: private, driving school and not applicable.

No

56. Licence status There are seven licence statuses: no licence, learner licence, provisional licence, full licence, international licence, police licence and army licence.

No

57. Driver injury There are four types of injuries: fatal, serious, slight and no injuries.

Yes

58. Seat belt This variable represents whether seat belts were fastened or not. There are six categories: seat belt fastened, seat belt unfastened, wearing helmet, serban (turban), wearing a helmet improperly properly tight and not wearing helmet or serban.

No

59. Part of body injured Parts of the body injured in the crash: head, neck, chest, arm, back, buttock, leg, various and no injury.

Yes

60. Type of driver fault This variable identifies any driver faults, divided into fifteen categories: in/out vehicle, negligent signalling, overloading (goods), overloading (passenger), wrong parking, drugs, careless driving, dangerous driving, dangerous turning/wrong turning, dangerous overtaking, driving too close, speeding, traffic light violation, other offences and not at fault.

Yes

61. Driver qualification The highest qualification of the driver, divided into four categories: no schooling, primary school, secondary school and higher education.

No

62. Drunk driver This variable identifies whether driver was No



crash sites drunk or not. The three categories for this variable are not tested, tested positive and tested negative.

63. Driver employment status The employment status of the driver may be working, student, or not applicable.

Yes

The crash data form is divided into six main sections, as in Table 3-4: details

report/ time of crash, road information, environmental information, crash location,

vehicle information, and driver information. Although all the data from the first

section (details report/time of crash) was available, data in other sections is often

missing. For example, in the road information section, only two of the fifteen

variables, road geometry and collision type, were available for all crashes. None of

the weather information data was complete for the crashes in the study sites. Of the

four variables in the crash location section, only two were complete, including road

type and route number. Type of vehicle is only one of the 12 variables in the vehicle

information section with complete data. Lastly, only seven of the variables in the

driver information section were complete, and these were sex, age, race, driver

injury, part of body injury, type of driver fault and driver status.

During 2008-2012, a total of 25,439 crashes occurred along federal roads in

Sabah. Among them, about 19% (4,875) crashes were identified as occurring along

roads in mountainous areas and the other 81% occurred along non-mountainous

roads. Study 1 was the only study that used these statistics.

Due to budget limitations and reduced labour resource needs, this research

(Study 2, 3 and 4) only focussed on the 102 selected segments that have been

described in Section 3.2.1. Table 3-5 shows the different crash characteristics on all

mountainous highways in Sabah and on the selected 102 segments. For other studies,

between 2008 and 2012, a total of 715 single-vehicle crashes (SV) and 257 multi-

vehicles (MV) crashes (including injury and property damage only crashes) occurred

along the selected segments. To account for monthly variations in traffic volume and

to ensure the accuracy of real-time weather information, SV crashes on each segment

were counted at monthly intervals. This led to a panel dataset of 6,120 observations

for 102 mountainous highway segments. For MV crashes, yearly intervals were

made, which represent 510 observations for the same highway segments. In the crash


severity study (Study 4), the combined data of SV and MV crashes were used, with

the total observations numbering 972 crashes.

3.2.5.2 Topographical Information

The topographical information was obtained from the Digital Terrain Model

(DTM) provided by the Department of Survey and Mapping Malaysia. The DTM

includes all the topographical information in digital form, including contour lines, for

all districts in Sabah (Series No.DTM_T738).

3.2.5.3 Exposure Variables and Traffic Characteristics

The two exposure variables used for the SV crashes study were average

traffic volume (ADT) and length of road segment. Traffic volume data was collected

from the Road Traffic Volume Malaysia (RTVM) database (HPU, 2013) maintained

by the Public Works Department in Sabah and the Highway Planning Unit, Ministry

of Works Malaysia. This database includes historical traffic volume data for some

predefined census locations along the roads and may not have exact volume data for

the selected segments. To circumvent this problem, two-hour vehicle counts were

conducted for each segment, and the ADT for each segment was estimated by using

hourly expansion factors and seasonal variation factors following the procedure

mentioned in Garber and Hoel (2009). Traffic volume data from RTVM was divided

into two seasons; January to June (census in March) and July to December (census in

September) for each year. In these two seasons, RTVM provided the details of traffic

survey for a 16-hour period at the selected survey stations. This 16-hour census data

was then transformed into 24-hour data using factors provided by the Public Works

Department in Sabah (HPU, 2013). From this 24-hour traffic volume, the hourly

expansion factors (HEF) for each census station were calculated. Using these

seasonal HEFs, the two-hour vehicle counts of each segment were converted to

ADT. For MV crashes study, ADT and segment length were multiplied to estimate

the daily vehicle kilometre travel (VKT) and to reflect the crash exposure for each

segment.


Table 3-5: Crash Characteristics for all Rural Mountainous Highways in Sabah vs. along Selected 102 Highway Segments

Site From To Length (km) Number of

Crashes

Crash Severity

PDO SLI SOI FA

1 Kimanis Keningau 48.3 (12.8) 506 (120) 459 (113) 17 (2) 13 (1) 17 (4)

2 Dongongon Tambunan 62.0 (14.9) 690 (241) 641 (224) 36 (10) 23 (6) 8 (1)

3 Tamparuli Beluran 202.0 (62.2) 2,185 (611) 2,042 (569) 38 (14) 63 (18) 42 (10)

Total 312.3 (89.9) 3,381 (972) 3,142 (906) 91 (26) 99 (25) 67 (15)

* Value in parenthesis represents crash characteristics for 102 selected segments * PDO = Property Damage Only, SLI = Slight Injury, SOI = Serious Injury, FA = Fatality


3.2.5.4 Weather Conditions

Hourly rainfall and visibility information for the same five year period were

collected from the Department of Irrigation and Drainage, Sabah, Malaysia and

National Centers for Environmental Information (NOAA). While the rainfall data

were available from twelve rainfall stations next to the study locations, the visibility

data were available from only two weather stations. Detailed information about

rainfall stations and weather stations are presented in Table 3-6 and Table 3-7,

respectively. Using proximity measures in AutoCAD - Geolocation, hourly weather

information for crashes on each road segment was obtained following the

aggregation procedure developed by Yu et al. (2015). Figure 3-11 shows the location

of rainfall and weather stations and boundaries for each of them. In this aggregation

procedure, the reported crash times were first matched with weather information

recorded at the nearest weather stations. Real-time weather information (e.g. rainfall,

visibility at the time of crash) was then converted into a segment level variable

following three criteria: exact value if a segment had only one crash, an average

value if a segment had more than one crash and annual average if a segment had no

crash. Using the same procedure, the time stamp of every crash in this study was

matched with the meteorological dataset to extract the corresponding weather

information. Real-time weather data was converted to the segment level by using

exact values if the segment had only one crash. An average value (a single value for

each segment) was used if the segment had more than one crash. The monthly

average weather information was used if the segment had no crash in that month. For

study 2 – SV crashes study, real-time weather information included average visibility

at the time of crash, and average hourly rainfall at time of crash. For study 3 – MV

crashes study, an indicator rainfall variable; heavy rainfall (more than 5.08 mm) at

the time of MV crashes and heavy rainfall during one hour before MV crashes were

used as the criteria to construct the real-time rainfall indicator variable. Meanwhile,

visibility conditions during MV crashes were collected from two weather stations

available near the study area. For Study 4 – the crash severity study, a rain indicator

was used to represent weather conditions at the time of the crash.


Table 3-6: List of Nearest Rainfall Stations No Station Station Name Location Information Data Period 1 5361002 Keningu Met Stn Latitude = 5.3455 N 01/2008 – 12/2012 Longitude = 116.1595E

2 5460001 Ulu Pampang Latitude = 5.4700 N 01/2008 – 12/2012 Longitude = 116.0583E

3 5558001 Bongawan Latitude = 5.5187 N 01/2008 – 12/2012 Longitude = 115.8733 E 4 5663001 Tambunan Agr.stn Latitude = 5.6298 N 01/2008 – 12/2012 Longitude = 116.3246 E

5 5862002 Ulu Moyog Latitude = 5.8714 N 01/2008 – 12/2012 Longitude = 116.2504 E

6 5961005 Babagon Latitude = 5.9044 N 01/2008 – 12/2012 Longitude = 116.1790 E

7 6062001 Kiulu Latitude = 6.0623 N 01/2008 – 12/2012 Longitude = 116.2722 E

8 6064001 Dalas Latitude = 6.0339 N 01/2008 – 12/2012 Longitude = 116.4553 E

9 5965003 Kinasaraban Kundasang Latitude = 5.9994 N 01/2008 – 12/2012 Longitude = 116.3469 E

10 5966001 Ranau Agr Stn Latitude = 5.9528 N 01/2008 – 12/2012 Longitude = 116.6691 E

11 5768001 Tampias Latitude = 5.7156 N 01/2008 – 12/2012 Longitude = 116.8641 E

12 5663001 Telupid Latitude = 5.6289 N 01/2008 – 12/2012 Longitude = 117.1247 E

Source: DID (2016)

Table 3-7: List of Nearest Weather Stations No Station Station Name Location Information Data Period 1 964710 Kota Kinabalu Latitude = 5.937 N 01/2008 – 12/2012 Airport Longitude = 116.051E

2 964910 Sandakan Airport Latitude = 5.901 N 01/2008 – 12/2012 Longitude = 118.059 E


Figure 3-11: Segment coverage by different rainfall and weather stations

3.3 CRASH CHARACTERISTICS ANALYSIS

This study applied disaggregate-analysis techniques to examine the

differences in crash characteristics between mountainous and non-mountainous

roads. Two types of outcome variables were used in the analysis: 1) crash frequency;

and 2) crash percentage. Several explanatory variables were tested by the proposed

technique, including collision type, crash severity, roadway geometric features, time

and day of crashes, vehicle characteristics, driver attributes and driving manoeuvres

prior to the crash. A series of chi-square tests in the form of contingency tables were

conducted to compare the statistical differences between mountainous and non-

mountainous road crashes across the range of explanatory variables. In addition, odds

ratios, which provide a relative likelihood of occurrence of events for a given

category in comparison with other categories, were calculated to measure effect size

and the strength of the relationship between pairs of categorical variables (McHugh,

2009).

3.4 SINGLE – VEHICLE CRASH MODEL

This section describes the modelling technique that was used in the single-

vehicle crashes study.

1

2

3 4

5 6

7 8 9 10

11

12

1

2

Weather stations

Rainfall stations

Road


3.4.1 Model Development

Count data modelling techniques were applied to establish the relationship

between observed SV crashes and explanatory variables like road geometries, traffic

characteristics, real-time weather conditions, cross-sectional elements, roadside

features, and spatial characteristics of rural mountainous roads. As crash counts at

transportation entities are often over-dispersed, the Negative Binomial (NB)

regression model is generally preferable to the Poisson regression model.

Let Yit represent SV crash counts on the ith highway segment in the tth time

period. Following the Poisson process, the probability of segment i and period t

having Yit crashes is expressed as follows:

( ) = ( )! , i = 1, 2, ……., N and t = 1, 2, ….., T (1)

where λit is the Poisson mean for highway segment i in time period t. In Negative

Binomial (NB) regression model, the Poisson mean is specified as follows:

)exp( ititit ελ +′= βX (2)

where ),......,,1( ,1, ′= kititit XXX is a vector of covariates representing segment-specific

attributes of mountainous highways, ),.....,( 0 ′= kβββ is a vector of unknown

regression parameters, and itε is the model error that is independent of all covariates.

The stochastic component, itε allows for over-dispersion in the crash data. The NB

model assumes that )exp( itε is Gamma distributed with mean 1 and variance θ. The

parameter θ is often referred to as an over-dispersion parameter, which leads to

following probabilistic distribution for observed crashes on mountainous highways

segments:

( ) = Γ (1 ⁄ ) +Γ(1 ⁄ ) ! 1 ⁄1 +⁄ ⁄ (1 ⁄ ) + (3)

where Γ (⋅)is a gamma function.


The simplistic mean structure in equation 2 cannot take into account possible

non-linear relationships between exposure (e.g. traffic flow) and crashes. It also does

not ensure that in the absence of exposure there should not be any crash. To address

these fundamental issues, logarithmic transformations of major and minor road

traffic flows have been used in the development of safety performance functions for

intersection crashes (Mitra & Washington, 2007; Tulu, Washington, Haque, & King,

2015; Washington & Haque, 2013). Following the same principle, the mean of SV

crashes along highway segments is structured as follows:

)exp(21

0 ititiitit LF εαλ αα +′= βX (4)

where Fit represents traffic flows in Average Daily Traffic (ADT) along ith highway

segment in tth time period, Li is the length of segment i, and α0, α1 and α2 are

regression parameters to be estimated.

Unobserved heterogeneities represent a major challenge in developing SPFs.

Heterogeneities could be structured or unstructured depending on the sources they

arise from. In the context of this study, structured heterogeneities may result from

data clustering or because of temporal correlations, as the same road segment was

observed for multiple time periods. The above NB model cannot take into account

location specific effects and potential serial correlation associated with the use of

time-series cross-sectional panel data for SV crashes in this study. This may lead to

incorrect inferences of model parameters as the estimated standard errors of

regression coefficients may be underestimated. On the other hand, unstructured

heterogeneities may arise from model misspecification, uncertainty in exposure and

covariates, and omitted variables. Although an extensive effort has been made to

collect relevant data that may influence SV crashes along rural mountainous roads in

Malaysia, there may remain some unobserved variables. For example, driver

behaviour factors like aggressiveness and risk-taking have a strong relationship with

traffic crashes, but they are unobserved in this study. In the absence of driver

behaviour factors, it may be unrealistic to assume that the effects of available

explanatory variables are fixed across all observations. This misspecification may

lead to biased and inconsistent parameter estimates, and erroneous inferences

(Mannering et al., 2016).


To account for these unobserved heterogeneities, this paper has applied a

random parameters Negative Binomial model (RPNB) for SV crashes along rural

mountainous roads. The random constant term of RPNB model acts as a location-

specific parameter and allows for structured heterogeneities or within-subject

correlations. The randomness specification of regression parameters allows

parameters to vary across road segments to account for unstructured heterogeneities.

The regression coefficients in RPNB model can be expressed as follows:

itit ωββ += (5)

where is a randomly distributed term (e.g. a normally distributed term with mean

zero and variance ). With this equation, the negative binomial parameter

becomes | = ( + ), and the corresponding log likelihood

function can be expressed as follows:

= ( )∀ ( | ) (6)

where g(⋅) is the probability density function of . Theoretically, a wide range of

probability distributions could be specified for . In this study, regression

parameters are specified to be normally distributed as this is often found to be

suitable in SPFs. As the log-likelihood function in equation 6 is computationally

cumbersome with random parameters, simulation-based maximum likelihood

techniques are typically employed with Halton draws (Bhat, 2003; Milton et al.,

2008; Train, 1999). A parameter was defined as random if the estimated standard

deviation was significantly different from zero; otherwise it was estimated as a fixed

parameter. To obtain the parsimonious model with the best subset of regression

parameters, preliminary multicollinearity and backward stepwise techniques were

employed in this study.

3.4.2 Parameter Estimates and Effects

To estimate the effects of estimated regression parameters on SV crashes, two

types of elasticities were computed: elasticities for continuous variables (equation 7)

and pseudo-elasticities for indicator variables (equation 8).


= Χ = (7)

= ( ) − 1( ) (8)

where E represents the elasticity, PE represents pseudo-elasticity, is the value of

the independent variable for observation , is the estimated parameter for the

independent variable, and is the expected SV crash frequency for observation i.

Elasticity for a continuous variable indicates the percentage in expected SV crash

frequencies for a percent change in the continuous variable while holding all other

variables at their mean. The pseudo-elasticity for an indicator variable indicates the

percentage change in SV crash counts for the condition change (0 to 1) in the

indicator variable while holding all other variables at their mean (Washington,

Karlaftis, & Mannering, 2010).

3.5 MULTI – VEHICLE CRASH MODEL

To analyse MV crash data with excess zeros, three state-of-the-art modelling

approaches are employed and compared in this study. These models include Random

Parameters Negative Binomial (RPNB), Negative Binomial – Lindley (NB-L) and

Negative Binomial – Generalized Exponential (NB-GE) models. The following

sections present the details of the above-mentioned models respectively.


3.5.1.1 Random Parameters Negative Binomial (RPNB)

The development of the RPNB model for MV crashes also used the same

procedures as described for SV crashes in section 3.4.1.

3.5.1.2 Negative Binomial – Lindley (NB-L)

Negative Binomial – Lindley (NB-L) distribution has been introduced in the

literature as a promising distribution to handle crash data with excess zeros

(Geedipally et al., 2012) and thus is selected in this study as an option to analyse MV

crashes. The appealing characteristic of the NB-L stems from the property of its core

distribution, the Lindley distribution, whose mean is close to zero and has a long tail

for observations that are far away from zero. Figure 3-12 shows the probability


density of Lindley against NB distributions overlaid with the histogram of MV crash

data used in this study. It can be seen that the density of the Lindley distribution

around zero fits the histogram of MV crashes more appropriately and thus a

combination of NB and Lindley distributions can shift the sole NB distribution to the

left and can capture excess zeros more effectively.

Figure 3-12: Resemblance of Lindley distribution to the distribution of MV crash count with excess zeros.

The probability density function for the Lindley distribution can be presented as

(Zamani & Ismail, 2010a):

0,0;)1(1

2);( >>−+

+= YYeYYf ψψ

ψψψ (9)

where ψ is the parameter of the Lindley distribution. The NB-L is a combination of

NB and Lindley distributions and thus the probability density function of the NB-L

can be expressed as (Geedipally et al., 2012):

== εψεεμφψφμ dLindleyyNByYP );(),:(),,,( (10)

where f(u;a,b) means that f is the distribution of the variable μ, with parameters a

and b.


Accordingly, the likelihood function of the NB-L can be determined by

computing the product of the density function (Equation 10) over the entire

observations, i.e. N highway segments and T time periods. This product, however,

involves a double integration which does not have a closed form. On the other hand,

the unique property of the NB-L distribution is that it can be expressed as a

hierarchical representation of Bernoulli and Gamma distributions (Geedipally et al.,

2012). From Equation (10), it can been that the Lindley distribution is a mixture of

exponential distribution with parameter ψ and gamma distribution with parameter p

where = and thus the NB-L model for MV crashes can be constructed as

follows (Geedipally et al., 2012):

== ),:(),,,( εμφψφμ yNByYP

);(~ ψεε Lindley (11)

An alternative expression of the above-mentioned model specification is that

it is assumed that MV crashes (Yit) follow a Negative Binomial distribution with the

mean of λit, dispersion parameter φ and random effect (ɛ ):

),,;()|,,( iititititit yNByYP εφλεφλ ==

),1(~ ψε zGamma +

)1

1(~

ψ+Bernoulliz

(12)

3.5.1.3 Negative Binomial – Generalized Exponential (NB-GE)

Another appealing distribution introduced in the literature for observations

with excess zeros is the Negative Binomial – Generalized Exponential (NB-GE)

distribution (Aryuyuen & Bodhisuwan, 2013; Vangala, Lord, & Geedipally, 2014).

Similar to the NB-L approach, the motive to use this distribution for crashes with

excess zeros comes from the unique property of the GE distribution which has dense

probabilities for values around zero and long tail for values farther from zero. Figure

3-12 shows the probability density of Generalized Exponential against NB

distributions overlaid with the histogram of MV crash counts. The same conclusion


obtained from Figure 3-13 can be drawn here about the combination of NB and GE

distributions to capture crash data with excess zeros.

Figure 3-13: Resemblance of Generalized Exponential distribution to the distribution of MV crash counts with excess zeros.

The probability density function (pdf) of the GE distribution can be presented

as follows (Vangala et al., 2014):

),,;()|,,( iititititit yNByYP εφλεφλ ==

),1(~ ψε zGamma +

)1

1(~

ψ+Bernoulliz

(13)

where α is the shape parameter and λ is the scale parameter of the GE distribution.

Similar to the NB-L, the NB-GE can also be presented as a combination of NB, and

GE distributions and thus its probability density function can be stated as (Vangala et

al., 2014):

== dzzGEzxNByYP ),;(),;(),,,,( λαμφλαφμ (14)

where the notations are as previously stated. The likelihood function of the NB-GE is

the product of its pdf over the entire observations (highway segments and time


periods). Although the likelihood function for the NB-GE can be expressed as a

closed form, this model has also been estimated in a Bayesian approach. Thus, the

hierarchical representation of the NB-GE is presented in the following without going

any further into the details of the likelihood function. The interested reader is

referred to Aryuyuen and Bodhisuwan (2013) for more detailed discussion about the

NB-GE model specification.

It is assumed that MV crashes at site i and time period t (Yit) follow a

Negative Binomial distribution with the mean of λit, dispersion parameter Φ and a

random effect ɛ which follows a Generalized Exponential distribution with

parameters a and b. Accordingly, the complete multi-level hierarchical model can

now be given as:

),,;()|,,( iitityNBitityitYP εφλεφλ ==

ε ~ Generalized Exponential (a,b)

(15)

3.5.2 Model Estimation

The integral involved in the likelihood function of the RPNB model (Equation

6) cannot be solved directly and thus RPNB model specification impedes the use of

conventional Maximum Likelihood Estimation (MLE); rather, it is required to apply

a simulation approach to estimate model parameters (Hensher, Rose, & Greene,

2005). In RPNB model of this study, the Maximum Simulated Likelihood Estimation

(MSLE) with Halton draws is applied to estimate model parameters.

On the other hand, since the NB-L and NB-GE models have hierarchical

structure, the Bayesian inference (BI) is employed to estimate the corresponding

regression parameters. According to the BI, the posterior probability is estimated as a

product of likelihood and prior over the marginal probability. Thus, model estimation

can be achieved by maximizing the posterior probability as follows:

[ ] [ ] [ ]

=ityditym

itYPitYP

)(

||

μπμμ

(16)

where | is referred to as the posterior probability, | is the likelihood of

MV crashes and ( ) is the marginal distribution of MV crashes. is the


prior term which includes any prior information about regression parameters (αi,βi)

In this study, non-informative priors have been used for all parameters. Markov

Chain Monte Carlo (MCMC) simulation with two chains has been employed to

determine the posterior for MV crashes along rural mountainous highways.

3.5.3 Goodness-of-fit (GOF) Measures

As the candidate models of this study were estimated in different platforms (i.e.

MSLE and Bayesian approaches), their prediction abilities could not be directly

compared using local measures such as Akaike Information Criterion (AIC),

Bayesian Information Criterion (BIC) or Deviance Information Criterion (DIC). As

such, the candidate models were compared based on global Goodness-of-Fit (GOF)

criteria. In this study, three commonly used prediction-based model selection criteria

were used, including 1) Mean Absolute Deviation (MAD) 2) Mean Squared

Predictive Error (MSPE), and 3) Mean Squared Error (MSE). Suppose, iY and iY

are the predicted and the observed MV crash counts for site i, respectively. The

MAD, MSPE and MSE can be calculated as follows (Oh, Lyon, Washington,

Persaud, & Bared, 2003):

|1

ˆ|1

ii yn

iy

nMAD −

== (17)

2)1

ˆ(1

ii yn

iy

nMPSE −

== (18)

2)1

ˆ(1

ii yn

iy

pnMSE −

=−= (19)

where n and p donate the total number of observations (highway segments and time

periods) and the number of degrees of freedom, respectively. The model with smaller

values of MAD, MSPE and/or MSE is selected as the superior model in terms of

prediction ability (goodness-of-fit).


3.5.4 Elasticity effect

One way to gain deeper understanding of the effects of crash contributing

factors is to determine their elasticities. The elasticities determine the percentage

change in expected MV crash counts for a one percent change in any of the crash

contributing factors utilized in the model, while holding all other factors constant.

The explanation of the computation of this elasticity can be found in Section 3.4.2.

3.6 CRASH SEVERITY MODEL

This study aimed to identify factors influencing crash severity along rural

mountainous highways. Three models will be tested to obtain a best model that

produces the highest goodness-of-fit statistics with meaningful parameters. The

models are: 1) standard logit model that will act as a base model, 2) Scobit model to

consider the imbalance in the response variable, and 3) random parameters logit

model to consider the unobserved heterogeneity in the dataset. Moreover, a decision

tree will be formed to identify significant interactions among a set of independent

variables in relation to the crash severity.

3.6.1 Decision Tree

A decision tree is one of the nonparametric methods that have been used to

identify the relationship between dependent and independent variables. The

advantages of this method that it is easy to interpret the complicated association in

crash severity modelling, and there is no need to identify the relationship between

dependent and independent variables. In addition, this method has the capability to

capture the interaction between independent variables through the structure of the

tree (Rashidi, Ranjitkar, & Hadas, 2014). However, it is a nonparametric method and

may suffer from type I errors (i.e. incorrect rejection of a true null hyphothesis).

Hence, a two-step modelling approach is used in this study with a combination of a

decision tree and logistic regression. The results from the decision-tree are used as a

prior knowledge for the logistic regression model. In short, the possible higher order

interactions are determined in the decision tree and inferences of parameters are

identified in the logistic regression.



3.6.2.1 Standard Logit Model

In the crash severity study, there are many models that have been introduced

such as binary logit/probit, ordered logit/probit, multinomial logit/probit, mixed logit

and latent class models. In this study, binary logit was used in the analysis of crash

severity of crashes along rural mountainous highways in Malaysia, because of the

imbalance proportion of injury severity categories in crash data.

Let Yi represent injury severity (severe and non-severe) in the ith crashes. In

the logistic regression model, the marginal expectation of ( ) = satisfies

logit( ) = ,where = ( ,… , ) donate a × 1vector of explanatory

variables, and β is a vector of estimable regression parameters. The probability of

crashes i having severe crashes is expressed as follows:

Pr( = 1) = exp( )1 + exp( ) (20)

3.6.2.2 Scobit Model

In crash data with an imbalanced dependent variable, the Scobit model is an

alternative to a standard logit model. This model was introduced by Nagler (1994)

and used in an injury severity study by Tay (2016). In the standard logit model, ε has

a logistic distribution. If ε has a Burr-10 distribution, the skewed logistic is as

follows (Nagler, 1994):

F( ;) = 1 − 11 + (21)

where = and α serves as a measure of skewness. Note that if α =1 or ln (α) =

0 then the Burr-10 distribution is equivalent to the logistic distribution and the Scobit

model is reduced to the standard logistic model. The probability of skewed logistic is

defined as follows:

Pr( = 1) = 1 − 11 + exp( ) (22)


3.6.2.3 Random Parameters Logit Model

Another issue relating to injury severity modelling is unobserved

heterogeneity. The standard logit model has limitations in handling this issue.

Heterogeneities could arise from model misspecification, uncertainty in exposure and

covariates, and omitted variables. The random parameters logit or mixed logit model

was introduced to allow parameter estimates to randomly vary across the

observations. The probability of random parameters logit model are defined as

follows (Milton et al., 2008):

= exp( )1 + exp( ) ( | ) (23)

where ( | ) is the density function of β with ϕ referring to a vector of parameters

of the density function (mean and variance), and all other terms are as previously

defined in the standard logit model. For model estimation, β can now account for

crashes-specific variations of the effect of X on severe crashes, with the density

function ( | ) used to determine β. This model was estimated using 200 Halton

draws with logit distribution.

3.7 HEALTH RISK ASSESSMENT AND ETHICS STATEMENT

This research was conducted with respect for health and safety in order to

protect the researchers involved at every stage of the research. Road geometry

characteristics and traffic operational data were collected with due care. This

research also involved discreet observations of speeding behaviours of drivers on

public roads in Malaysia. Initial meetings with the CARRS-Q Health and Safety

officer were conducted in order to develop safe procedures during data collections.

The application for ethical approval was considered by the QUT Human Research

Ethics committee and was approved on 8th April 2015 (QUT Ethics Approval

Number 1500000272).

The major portion of this research used secondary data such as road traffic

crash data for Sabah, Malaysia. An Ethics Exemption application for the crash

characteristics study (Study 1) was submitted and approved in October 2014. An

Ethics Exemption application for other datasets (weather information, topographical

information, and traffic volume) was approved on 8th April 2015.


3.8 CHAPTER SUMMARY

This chapter described the process of collecting data and the methodology of

studies in this research. The study setting and population for this study is described at

the beginning. Then, the data collection process is described in detail, including the

road segmentation process and the sampling technique. The analysis method for

every study is then described. There are three modelling processes involved in this

research; 1) Random Parameters Negative Binomial – SV crashes study; 2) Random

Parameters Negative Binomial, Negative Binomial – Lindley (NB-L), Negative

Binomial – Generalized Exponential (NB-GE) – MV crashes study; 3) Standard

logit, Scobit and Random Parameters Logit Models – crash severity study. Finally,

the health risk assessment and ethics approval related to this research was explained.

Chapter 4: Characteristics of Mountainous Roads Crashes 74

Chapter 4: Characteristics of

Mountainous Roads Crashes

4.1 INTRODUCTION

This chapter presents the results of Study 1, which a descriptive study that

addresses Research Question 1 (What are the characteristics of road traffic crashes

along rural mountainous roads?). This chapter is structured into five major sections.

Section 4.1 gives a brief introduction to the study and the chapter, Section 4.2

presents the objective of this study. Section 4.3 discusses the data description.

Section 4.4 presents the study results and Section 4.5 discusses the findings of this

study.

4.2 OBJECTIVES

The objective of this study was to examine the characteristics of road traffic

crashes on rural mountainous roads and to compare these with the characteristics of

crashes on non-mountainous roads.

4.3 DATA DESCRIPTION

This study compares the crash characteristics along federal roads in

mountainous and non-mountainous areas using two types of data; 1) traffic crash

data from 2008 to 2012 from MIROS, and 2) topographical information obtained

from Department of Survey and Mapping Malaysia. The selection of mountainous

federal roads in this study was different from the selection of highway segments

described in Section 3.2.1. Using the same criteria and software to identify

mountainous highways in Section 3.2.1, mountainous roads were found to constitute

about 208 km (14.5%) out of 1,428 of km Federal roads in Sabah (IDS, 2007). After

identifying mountainous and non-mountainous roads, crash data were allocated to

these two types of roads by using the ‘route number’ variable in the M-ROADS

dataset. During 2008-2012, a total of 25,439 crashes occurred along federal roads in


Sabah. Of these, about 19% (4,875) were identified as occurring along roads in

mountainous areas, and the other 81% occurred along non-mountainous roads.

4.4 METHODOLOGY

This study applied descriptive analysis to compare crash characteristics along

mountainous highways and non-mountainous highways in Sabah. A series of chi-

square tests were conducted for different explanatory variables from the crash

dataset. The odds-ratio was calculated to identify the relative likelihood occurrence

of an event in certain categories relative to other categories with a 95% confidence

interval.

4.5 RESULTS

Results are discussed based on the differences in general crash characteristics,

environmental factors and driver/vehicle characteristics of crashes along

mountainous and non-mountainous roads.

4.5.1 General Crash Characteristics

Table 4-1 presents a univariate analysis comparing crash characteristics

between mountainous and non-mountainous roads. In the M-ROADS database, there

are 39 variables describing the general crash characteristics of every crash, such as

the month of crash, day of the week, road geometry, intersection type and area type.

Among these, four variables were found to be statistically significant in

distinguishing crashes between mountainous and non-mountainous roads. These

included horizontal alignment, collision type, crash type and injury severity. As

shown in Table 4-1, the collision type variable has eight categories, which are

including rear-end, out-of-control, head-on, angle, side swipe, vehicle-pedestrian,

overturn and other crashes. While rear-end crashes were the most frequent (nearly

38%) collision type along non-mountainous roads, ‘out-of-control’ crashes were the

most common (about 48%) collision type along mountainous roads. Compared to

rear-end crashes, the odds of ‘out-of-control’ crashes along mountainous roads were

about 4.2 times (95%CI: 3.89 – 4.60) higher than on non-mountainous roads. The

odds of head-on, side swipe and overturn crashes were also significantly higher along

mountainous roads than non-mountainous roads, with the corresponding odds

respectively about 3.6 times (95%CI 3.09 – 4.23), 3.1 times (95%CI 2.64 – 3.64) and


3.4 times (95%CI 2.84 – 4.10) higher. Differences in the likelihood of vehicle-

pedestrian collisions and angle collisions were not statistically significant across

mountainous and non-mountainous roads. In addition, the number of collisions

involving animals was very small compared to other type of collisions along

mountainous roads, and thus they were grouped together with the ‘other collision

type’ category.

Fatal and serious injury, slight injury, and property damage only crashes

represented respectively about 5.8%, 2.4%, and 91.8% of crashes along mountainous

roads. Similar shares of injury crashes were also observed among non-mountainous

road crashes. Therefore, only the odds of a slight injury crash were statistically

significantly different between mountainous and non-mountainous roads. In general,

crashes along mountainous roads were slightly more severe, as the fatality index

(ratio of fatalities to road injuries) for mountainous roads was 0.21 whereas the

fatality index for non-mountainous roads was only 0.18.

Table 4-1: General crash characteristics

Variable Mountainous

, n (%)

Non-Mountainous, n

(%)OR (95% CI) 2, p-value

Collision Type Rear-end* 900 (18.5) 7708 (37.5) 1.00 Out-of-control 2317 (47.5) 4697 (22.8) 4.23 (3.89 – 4.60) 1204.893, p < 0.01 Head-on 271 (5.6) 642 (3.1) 3.62 (3.09 – 4.23) 282.901, p < 0.01 Angle and right angle side

546 (11.2) 4438 (21.6) 1.05 (0.94 – 1.18) 0.829, p = 0.36

Side swipe 248 (5.1) 686 (3.3) 3.10 (2.64 – 3.64) 206.287, p < 0.01 Vehicle-pedestrian 50 (1.0) 448 (2.2) 0.96 (0.71 – 1.29) 0.087, p = 0.77 Overturn 187 (3.8) 469 (2.3) 3.42 (2.84 – 4.10) 191.765, p < 0.01 Others 356 (7.3) 1476 (7.2) 2.07 (1.81 – 2.36) 115.017, p < 0.01 Crash Severity Property damage only* 4474 (91.8) 18765 (91.3) 1.00 Slight injury 118 (2.4) 657 (3.2) 0.75 (0.62 – 0.92) 7.861, p < 0.01 Fatal & serious injury 283 (5.8) 1142 (5.6) 1.04 (0.91 – 1.19) 0.318, p = 0.573 Crash Type Multi-Vehicle * 1695 (34.8) 11948 (58.1) 1.00 Single-Vehicle 3139 (64.4) 8485 (41.5) 2.61 (2.44 – 2.78) 862.435, p < 0.01 Unknown 41 (0.8) 131 (0.6) 1.18 (0.83 – 1.68) 0.863, p = 0.35 Horizontal Alignment Straight* 2071 (45.2) 13748 (84.4) 1.00 Bend 2507 (54.8) 2546 (15.6) 6.54 (6.08 – 7.02) 2983.346, p < 0.01 * Reference category

Among crash types, single-vehicle crashes were the most frequent crash type

in mountainous areas, representing about 64.4% of all crashes along mountainous

roads. Compared to multi-vehicle crashes, the odds of single-vehicle crashes along

mountainous roads were about 2.6 times (95%CI 2.44 – 2.78) higher than for non-


mountainous roads. Moreover, the single-vehicle crashes represented about 97% of

out-of-control crashes.

The horizontal alignment of roads appeared to have more influence along

mountainous roads than non-mountainous roads. Nearly 55% of crashes along

mountainous roads occurred along roads with a horizontal curve, whereas only 16%

of crashes in flat areas occurred along a road bend. The corresponding odds ratio

suggested that the presence of horizontal curves, compared to straight road segments,

increased the likelihood of crashes as much as 6.5 times (95%CI 6.08 – 7.02) along

mountainous roads compared to non-mountainous roads.

4.5.2 Environmental factors

Table 4-2: Characteristics of crashes by time of the day, day of the week, and seasonal variations

Variable Mountainous, n

(%) Non-Mountainous, n


Time of day Day time* 3037 (62.3) 13573 (66.0) 1.00 Night time 1838 (37.7) 6991 (34.0) 1.18 (1.10 –

1.25) 28.888, p < 0.01

Day of week Weekdays* 3411 (70.0) 14779 (71.9) 1.00 Weekend 1464 (30.0) 5785 (28.1) 1.10 (1.02 –

1.17) 6.975, p < 0.01

Season of year Dry Season* 4013 (82.3) 16873 (82.1) 1.00 Wet Season 862 (17.7) 3691 (17.9) 0.98 (0.91 –

1.07) 0.191, p = 0.66

School seasons School Days* 3711 (76.1) 16305 (79.3) 1.00 School Holidays

1164 (23.9) 4259 (20.7) 1.20 (1.12 – 1.29)

23.549, p < 0.01

* Reference category

Table 4-2 presents the results of disaggregate analyses which compared

environmental factors such as time of the day, day of the week and seasonal

variations between mountainous and non-mountainous road traffic crashes. As

reported in Table 4-2, night time crashes represented about 38% of mountainous

crashes. Compared to day time crashes, the odds of night time crashes along

mountainous roads were about 18% (95%CI 1.10 – 1.25) higher than on non-

mountainous roads. Compared to weekdays, weekend crashes were slightly

overrepresented along mountainous roads, with the corresponding odds about 10%

(95%CI 1.02 – 1.17) higher. In terms of seasons of the year, crash occurrences on

mountainous and non-mountainous roads were not significantly different across dry

and wet seasons. In general, the dry season for central parts of Sabah, where most of


the mountainous roads are located, occurs between February and April, while the wet

season is between May and January (MET, 2017). However, in terms of school

seasons, the odds of crashes along mountainous road were about 20% (95%CI 1.12 –

1.29) higher during school holidays.

4.5.3 Driver and vehicle factors

Table 4-3 presents the distribution of crashes along mountainous and non-

mountainous roads across various driver/vehicle factors. The age distributions of

drivers involved in crashes along mountainous and non-mountainous roads were

marginally different, with young (less than 25 years old) and older drivers (more than

64 years old) being slightly overrepresented in crashes along non-mountainous roads.

Female drivers were also less represented in crashes along mountainous roads, with

the corresponding odds about 43% lower (OR0.57, 95%CI 0.52 – 0.63).

About 60% of drivers involved in crashes along mountainous roads were

engaged in risky driving activities (e.g. speeding, dangerous overtaking, etc.) prior to

the crash, while the corresponding percentage for non-mountainous roads was about

51%. Speeding was the most frequent risky driving behaviour among drivers

involved in crashes along mountainous roads with the corresponding percentage

about 32% of all mountainous road crashes. In the crash database, speeding is

defined as driving over the posted speed limit. Compared to not-at-fault crashes, the

odds of crash involvement due to speeding were 2.78 times (95%CI 2.62 – 2.96)

higher along mountainous roads than non-mountainous roads. Risky driving

behaviour like dangerous overtaking was also more evident among drivers involved

in crashes along mountainous roads than non-mountainous roads, with the

corresponding odds about 14% higher, but this estimate was only significant at 10%

significance level. Other risky driving activities like ‘driving too close’ and

‘dangerous turning’ were more frequent among crash involved drivers along non-

mountainous roads, with the corresponding odds respectively 25% (OR 0.80, 95%CI

0.74 – 0.87) and 70% (OR 0.59, 95%CI 0.50 – 0.69) higher than mountainous roads.

Note that ‘driving too close’ and ‘dangerous turning’ are identified and recorded by

the traffic police based on their evaluation of driving manoeuvers prior to a crash.

The distribution of crashes along mountainous and non-mountainous roads

across vehicle types is presented in Figure 4-1. For both mountainous and flat areas,

passenger cars represented most of the crashes. Four-wheel drive (4WD) vehicles


were overrepresented in crashes along mountainous roads, representing about 29% of

all crashes in mountainous areas. Compared to passenger cars, the odds of crash

involvement for 4WDs were about 67% (95%CI 1.57 – 1.78) higher along

mountainous roads than non-mountainous roads. Similarly, heavy vehicles were also

overrepresented in crashes along mountainous roads, with the corresponding odds

about 43% (95%CI 1.32 – 1.55) higher compared to passenger cars and non-

mountainous roads. Small lorries and vans were also overrepresented in crashes

along mountainous roads, with the corresponding odds about 72% (95%CI 1.57 –

1.89) and 13% (95%CI 1.00 – 1.28) higher. In contrast, motorcycles only represented

about 1.7% of crashes along mountainous roads but 5.8% of crashes along non-

mountainous roads, resulting in the odds of motorcycle crashes for non-mountainous

roads about 3 times higher compared to passenger cars.

Table 4-3: Driver and vehicle factors

Variable Mountainous, n

(%)

Non-Mountainous, n


Driver Age <15 103 (1.4) 878 (2.8) 0.46 (0.37 – 0.56) 57.493, p < 0.0115-24 771 (10.3) 3914 (12.6) 0.77 (0.71 – 0.84) 37.789, p < 0.0125-44* 4491 (60.1) 17534 (56.6) 1.0 45-64 1989 (26.6) 8091 (26.1) 0.96 (0.91 – 1.02) 1.860, p = 0.17>64 113 (1.5) 586 (1.9) 0.75 (0.61 – 0.92) 7.484, p < 0.01Driver Gender Male* 4258 (87.3) 16365 (63.2) 1.00 Female 597 (12.2) 4027 (25.2) 0.57 (0.52 – 0.63) 145.532, p < 0.01Unknown 20 (0.4) 172 (11.6) 0.45 (0.28 – 0.71) 12.193, p < 0.01Driver Errors Not at fault* 3175 (40.1) 15807 (48.6) 1.0 Speeding 2497 (31.5) 4466 (13.7) 2.78 (2.62 – 2.96) 1091.885, p <

0.01Driving too close 892 (11.3) 5521 (17.0) 0.80 (0.74 – 0.87) 28.284, p < 0.01Dangerous turning 188 (2.4) 1591 (4.9) 0.59 (0.50 – 0.69) 45.447, p < 0.01Dangerous overtaking

227 (2.9) 993 (3.1) 1.14 (0.98 – 1.32) 2.890, p = 0.09

Other offences 945 (11.9) 4130 (12.7) 1.14 (1.05 – 1.23) 10.124, p = < 0.01

Types of vehicle Passenger car* 3534 (47.3) 16624 (53.6) 1.0 Four Wheel Drive 2170 (29.1) 6110 (19.7) 1.67 (1.57 – 1.78) 275.539, p < 0.01Heavy Vehicle 1032(13.8) 3391(10.9) 1.43 (1.32 – 1.55) 80.703, p < 0.01Van 367 (4.9) 1526 (4.9) 1.13 (1.00 – 1.28) 4.093, p = 0.04Small lorry 174(2.3) 699(2.3) 1.17 (0.99 – 1.39) 3.318, p = 0.069Motorbike 127 (1.7) 1789 (5.8) 0.33 (0.28 – 0.40) 150.349, p < 0.01Other vehicles 27 (0.4) 278 (0.9) 0.46 (0.31 – 0.68) 15.745, p < 0.01Unknown 34 (0.5) 586 (1.9) 0.27 (0.19 – 0.39) 61.382, p < 0.01*Reference category


Figure 4-1: Percentage of crashes by vehicle type for mountainous and non-mountainous roads

4.6 DISCUSSION

This study examined the characteristics of crashes along mountainous roads

and compared them with the characteristics of crashes along non-mountainous roads.

The results have brought several new insights into the characteristics of crashes

along mountainous roads which would be useful for designing countermeasures as

well as targeting more focused in-depth research.

It was found that ‘out-of-control’ was the most frequent collision type and

single-vehicle crashes were the most frequent crash type among crashes along

mountainous roads compared to non-mountainous road crashes. Further, about 97%

of ‘out-of-control’ crashes involved a single vehicle only. Mountainous roads often

represent a demanding driving situation due to their constrained topography and

complex road geometry. Chen and Chen (2013) also argued that mountainous roads

with steep gradients and horizontal curves represent a unique situation and impose

significant challenges to driving tasks. The second most frequent collision type along

mountainous roads was rear-end crashes. Rear-end crashes may have a variety of

contributing factors, including roadway and traffic characteristics and driver factors,

and therefore investigating their contributing factors on mountainous roads would be

0

10

20

30

40

50

60P

erce

nta

ge o

f cr

ash

es

Vehicle Type

Mountainous Roads

Non- Mountainous Roads


a worthwhile future research pursuit. Other types of crashes, such as vehicle-

pedestrian collisions, were less common along mountainous roads mainly because

these roads are located in rural areas where there are fewer pedestrians. Similarly,

angle or right-angle crashes were infrequent in mountainous areas because there are

fewer intersections along rural mountainous roads.

The presence of a horizontal curve was more associated with crashes along

mountainous roads than non-mountainous roads. It appears that horizontal curves

represent a relatively risky situation in mountainous areas mainly because of the

constrained topographical conditions. In addition, a larger proportion in the road

network in mountainous areas results in a higher exposure to along horizontal curves

on mountainous roads compared to non-mountainous roads. Consistency of

horizontal curves throughout the road bend is important to ensure comfortable and

safe driving. Wang et al. (2010) reported that a road bend with a constant radius

horizontal curve is safer than a road bend that consists of road curves with varying

radii. Li et al. (2014) also reported that driving along mountainous road curves with

different radii represents a dangerous situation mainly because drivers often fail to

calibrate driving speed with the sudden change of road alignments. Due to

constrained geometry and limitation of space, financial and technical resources, it is

often challenging to construct roads following engineering standards. As a result,

many horizontal curves along mountainous roads have substandard designs

compared to those in flat areas, which may contribute to the higher crash occurrence

along mountainous roads.

Night time crashes were more frequent in mountainous areas. Mountainous

roads in Sabah generally do not have street lighting. Complex road geometries of

mountainous roads in dark conditions may make the driving task more complex, and

drivers may face difficulties in negotiating roads with tight curves and steep slopes.

It is not clear from this research how the night time environment makes the driving

task complex in mountainous areas. Future research should investigate the influence

of various road geometrical elements on night time driving behaviour and safety.

Given that night time represents an unsafe environment along mountainous roads,

road authorities should target provision of street lights along these roads, at least in

black spot areas.


The odds of road crashes along mountainous roads were found to be higher

during weekends and school holiday seasons. This may simply reflect a higher

exposure of traffic along mountainous roads during times when people travel longer

distances for holidays or to visit family, friends and relatives. Further research on the

relationship between weekend/school holiday exposure and road crashes would be

useful in this regard. Despite previous research indicating that precipitation can

influence for crash occurrences along mountainous roads, the difference in crash

frequencies between wet and dry seasons was not found to be significant in this

research. Comparing only the data for different seasons may not yield sufficient

information to examine the influence of precipitation on road safety, and hence

future research should include meteorological data and compare crash frequencies in

wet and dry conditions.

Young and older drivers had lower odds of involvement in crashes in

mountainous areas compared with middle aged drivers. Similarly, female drivers

were involved less in road traffic crashes along mountainous roads compared with

male drivers. In the cultural context of Malaysia, middle-aged male drivers often

choose to drive when they travel with their family, and such longer trips are more

likely and this may be the case when they drive in mountainous areas which impose

complicated driving tasks and require long distance travel. Therefore, the high crash

involvement of middle-aged males along mountainous roads may be due to their

higher exposure. Other than that, it is well known that males are generally more

involved in crashes than females at all ages (e.g., McGwin Jr and Brown (1999)). In

addition, male drivers are generally noted to partake in higher risk driving.

With respect to driver actions, speeding and dangerous overtaking were

significant factors among drivers involved in crashes along mountainous roads. Lin

et al. (2013) also found that speeding was the main illegal driving action in

mountainous areas. As reported by Lee, Nam, and Abdel-Aty (2015), low traffic

volume in rural mountainous areas may encourage drivers to increase their speed. In

addition, many tourist spots are located in mountainous areas in Sabah, which

attracts weekend and holiday traffic. Illegal speeding and dangerous overtaking may

be initiated by drivers in holiday mode, and this merits further investigation. Due to

dangerous driving actions or manoeuvres of drivers along mountainous roads which


generally have complex road geometries, many researchers have suggested strict

speed enforcement along mountainous highways (e.g. Chen and Chen (2013)).

According to the Ministry of Transportation Malaysia, in 2012, more than

half of vehicles registered in Sabah were passenger cars (MOT, 2012). However, the

odds of crash involvement for 4WDs were much higher along mountainous roads in

Sabah. 4WDs, including sport utilities vehicles (SUVs), were also reported to be over

involved in crashes in earlier research elsewhere (McGinnis, Davis, & Hathaway,

2001). Research from Keall, Newstead, and Watson (2006) highlighted that 4WDs

are more liable to rollover crashes because of their higher centre of gravity relative to

the width of the wheel track. Recently, technology such as Electronic Stability

Control (ESC) has been introduced to solve this problem (Chatzikomis & Spentzas,

2014), however much of the vehicle fleet in Sabah Malaysia does not yet have ESC.

More research is also needed to investigate the performances of ESC along

mountainous roads with tight curves and steep slopes. In addition, the exposure of

4WDs is higher along mountainous roads because people may simply prefer using a

4WD for traveling in mountainous areas. Other than 4WDs, the odds of crash

involvement were also higher for small lorries and vans along mountainous roads

compared with non-mountainous roads.

Heavy vehicles, such as rigid lorries, lorry trailers and buses, represented a

substantial 13.8% of crashes along mountainous roads. Their odds of crash

involvement were higher on mountainous than non-mountainous roads. Due to their

size and manoeuvrability, heavy vehicle drives face an even greater challenge on

mountainous roads with steep slopes and tight curves. Some past research has

attempted to examine the effects of heavy vehicles on mountainous road safety. For

example, Chen, Chen, et al. (2011) found that vertical alignments of roads and

pavement surface condition influence the crash risk of trucks along mountainous

roads. Lee et al. (2015) reported that the proportion of trucks in the traffic volume is

negatively associated with crash rates. Li et al. (2010) reported that the crash risk of

trucks is likely to increase with the increase in speed limits along the roads in

mountainous areas. Heavy vehicles often face difficulties in maintaining driving

speed both along upgrade and downgrade sections of a road with a steep slope. The

slow speed of heavy vehicles often interrupts the flow of other traffic, particularly

along a road where no overtaking or relief lane is provided. In addition, continuous


braking while travelling along a downgrade section may impose additional hazards

for heavy vehicles, as continuous braking may cause brake-fade in which the braking

capability of the heavy vehicle significantly reduces due to over-heating.

In Sabah, motorcycles are the second most common travel option after

passenger cars (MOT, 2012). The odds of motorcycle crashes were found to be lower

along mountainous roads. Motorcycles are cheaper, easy to ride and require little

parking space, which makes them a good choice for middle and low income earners.

However, motorcycles are less likely to be used for long distance travel and climbing

up and down mountainous roads, and thus the exposure of motorcycles along

mountainous roads is less.

Chapter 5: Single Vehicle Crashes 85

Chapter 5: Single Vehicle Crashes

5.1 INTRODUCTION

Chapter 5 presents the results of Study 2, which aims to identify factors

contributing to single-vehicle (SV) crashes along rural mountainous highways in

Sabah, Malaysia. Study 2 addresses Research Question 2 of the current research

(What are the factors that lead to SV crashes along rural mountainous highways?).

To answer this research question, this study used five datasets including crash data,

topographical information, weather information, traffic characteristics, and field

survey data such as road geometry, cross-sectional elements, roadside features, and

spatial characteristics.

This chapter is organised into five sections. The first section (Section 5.1) is

the introduction section for this study. Then, Section 5.2 presents the objectives of

the study. Section 5.3 and 5.4 present the description of the data and modelling

results, respectively. The last section, Section 5.5, discusses the findings of this

study.

5.2 OBJECTIVES

The overall objective of this study is to investigate the effects of roadway

geometries, traffic characteristics, real-weather conditions, cross-sectional elements,

roadside features, and spatial characteristics on SV crashes along rural mountainous

highways. This is achieved by developing a safety performance function (SPF) for

SV crashes along rural mountainous highways. The contribution of this study is

threefold. First, it addresses an important road safety issue in a developing country –

SV crashes along rural mountainous highways. While there has been considerable

research on this topic in western countries, little is known in the context of a

developing country. According to the findings of Study 1, SV crashes represent

nearly 65% of road crashes along rural mountainous roads in Malaysia. The findings

from western countries may not be directly applicable, as there are differences in

roadway designs, roadside environment, presence of roadside furniture, traffic mix,

enforcement practices, and driver behaviour in developing countries compared to


developed countries. Second, road engineers often face challenges in designing and

ensuring adequate cross-sectional elements (e.g. shoulder, overtaking lane, etc.) due

to constrained road reserves along mountainous roads. A proper understanding of

their effects on safety is much needed in this regard. Third, the development of an

SPF for SV crashes along rural mountainous highways is a unique contribution of

this study. In addition to common roadway and traffic factors, the SPF developed in

this study includes a wide range of roadway factors including cross-sectional

elements, roadside features and spatial characteristics (e.g. adjacent land use factors)

in the context of a developing country. The SPF developed can provide insights into

SV crash occurrences along rural mountainous highways in developing countries in

general and in Malaysia in particular.


The details explanation of data collection for this study was made in the

Section 3.2. Table 5-1 represents descriptive statistics of explanatory variables

included in the SV model.

Table 5-1: Summary statistics of explanatory variables included in the model

Variables Mean Std. dev.

Min, Max Count* Percentage*

Exposure Variables

ADT (veh/day) 1682.70 856.10 297.60 5728.92

- -

Segment length (m) 881.58 384.89 30.00, 2039.00 - - Real– time weather information Average visibility at the time of crash (km)

12.15 1.395 1.50, 28.00 - -

Average hourly rainfall at time of crash (mm)

0.49 1.60 0, 90.00 - -

Traffic characteristics Upgrade speeding indicator (1 if 85th percentile vehicle operating speed greater than the posted speed limit)

- - - 539 8.81

Downgrade speeding indicator (1 if 85th percentile vehicle operating speed greater than the posted speed limit)

- - - 479 7.83

Horizontal alignment Proportion of segment with horizontal curve

0.52 0.17 0, 1 - -

Proportion of segment with simple curve

0.40 0.20 0, 1 - -

Proportion of segment with reverse curve

0.06 0.13 0, 0.66 - -

Proportion of segment with compound 0.04 0.09 0, 0.45 - -




curve Proportion of segment with broken back curve

0.03 0.12 0, 0.61 - -

Maximum degree of curvature (°) 17.54 17.44 0, 104.00 - -Minimum degree of curvature (°) 64.23 42.15 0, 330.00 - -Maximum radius of curvature (km) 0.84 0.96 0, 5.99 - -Minimum radius of curvature (km) 0.23 0.50 0, 3.66 - -Maximum length of circular curve (km)

0.15 0.08 0, 0.40 - -

Minimum length of circular curve (km)

0.09 0.07 0, 0.38 - -

Maximum length of tangent (km) 0.12 0.13 0, 1.10 - -Minimum length of tangent (km) 0.04 0.03 0, 0.17 - - Longitudinal grades Proportion of segment with longitudinal grades greater than zero

0.65 0.34 0, 1 - -

Number of vertical curves per km 4.32 3.96 0.54, 33.33 - -Maximum longitudinal grade <2% indicator (1 if maximum longitudinal grade <2%, 0 otherwise)

- - - 360 5.90

Maximum longitudinal grade 2 - 4% indicator (1 if maximum longitudinal grade 2-4%, 0 otherwise)

- - - 900 140.70


- - - 960 15.70


- - - 1200 19.60

Maximum longitudinal grade >8% indicator (1 if maximum longitudinal grade >8%, 0 otherwise)

- - - 2700 44.12

Cross-section elements Proportion of segment with concrete shoulder

0.09 0.20 0, 1 - -

Proportion of segment with bitumen shoulder

0.03 0.10 0, 0.52 - -

Proportion of segment with gravel and earth shoulder

0.14 0.23 0, 1 - -

Proportion of segment with turf shoulder

0.73 0.31 0, 1 - -

Proportion of segment with one side shoulder width >1.5m

0.21 0.33 0, 1 - -

Proportion of segment with both sides shoulder width >1.5m

0.65 0.39 0, 1 - -

Proportion of segment with both sides shoulder width <1.5m

0.14 0.28 0, 1 - -

Proportion of segment with unbroken centre line

0.72 0.32 0, 1 - -

Proportion of segment with rumble strip

0.01 0.04 0, 0.24 - -

Proportion of segment with marginal strip > 0.5m

0.01 0.05 0, 0.5 - -

Proportion of segment with edge drop-offs >100mm

0.01 0.02 0, 0.12 - -

Presence of overtaking lane (1 if there is an overtaking lane along the

- - - 900 14.70




segment, 0 otherwise) Roadway and roadside features Number of minor intersections 0.80 1.25 0, 7.00 - - Number of appropriate emergency stop areas

0.06 0.24 0, 1 - -

Number of trees per km 2.98 8.51 0, 58.67 - - Number of culverts per km 0.84 1.37 0, 7.81 - - Number of electric poles per km 23.86 23.82 0, 122.81 - - Number of roadway lighting poles per km

1.52 4.32 0, 28.83 - -

Proportion of segment with guardrails along one side

0.17 0.17 0, 0.72 - -

Proportion of segment with guardrails along both sides

0.02 0.05 0, 0.34 - -

Proportion of segment with embankments along one side

0.62 0.34 0, 1 - -

Proportion of segment with embankments along both sides

0.16 0.24 0, 1 - -

Proportion of segment with cliffs along one side

0.49 0.32 0, 1 - -

Proportion of segment with cliffs along both sides

0.05 0.09 0, 0.39 - -

Presence of bridge (1 if there is a bridge along the segment, 0 otherwise)

- - - 600 9.80

Presence of road delineation (1 if there are road delineations such as guide posts and chevron signs along the segment, 0 otherwise)

- - - 2520 41.20

Spatial characteristics Number of houses/shops/commercial buildings within 100m buffer zone from each road edge along the road segment per km

12.87 18.84 0, 113.87 - -

Proportion of segment with forest within 10m of the road edge

0.75 0.30 0, 1 - -

Proportion of segment with farm/ agricultural activity within 10m of the road edge

0.10 0.21 0, 1 - -

Proportion of segment with houses/shops/commercial buildings within 10m of the road edge

0.14 0.20 0, 1 - -

*count and percentage are reported for indicator variables

5.4 METHODOLOGY

In this study, random parameter negative binomial (RPNB) has been selected

for modelling single-vehicle crashes. This model was selected after considering the

effect of unobserved heterogeneity in longitudinal crash data. A detailed explanation

of the development of this model can be found in Section 3.4.


5.5 MODEL RESULTS

The RPNB model estimates of SV crashes along rural mountainous highways

are presented in Table 5-2. Reported random parameters were estimated using 200

Halton draws. A likelihood ratio test comparing the log likelihood values between

the fitted model and the null model indicates the overall significance (LR statistic =

384.92, p-value < 0.001) of the fitted model in explaining SV crashes. Akaike

Information Criteria (AIC) were used in the backward stepwise technique to derive

the parsimonious model by removing insignificant variables one by one. The over-

dispersion parameter was significant (95% CI: 0.49, 4.40) at a 5% significance level,

suggesting the greater appropriateness of the Negative Binomial regression model

than the Poisson regression model. The statistical significance of the variance of

location-specific effect (95% CI: 0.23, 0.39) suggests the existence of strong

structural temporal correlation (structured heterogeneity) in the dataset.

Table 5-2: RPNB model estimates of SV crashes along rural mountainous highways

Variables Estimate Std. Err.

z Prob. |z| >Z

[95% conf. interval]

Constant term 14.259 0.955 -14.93 0.000 [-16.131,-

12.388] Standard deviation of distribution 0.313 0.041 7.69 0.000 [0.233, 0.393] Exposure Variables Log ADT 0.451 0.091 4.94 0.000 [0.272, 0.631]Log of segment length 1.071 0.109 9.84 0.000 [0.857, 1.284] Real-time weather information Average visibility at the time of crash (km) 0.083 0.013 6.40 0.000 [0.058, 0.109]Average hourly rainfall at time of crash (mm)

0.116 0.014 8.30 0.000 [0.089, 0.143]

Traffic Characteristics Downgrade speeding indicator (1 if 85th percentile vehicle speed along downgrade greater than the posted speed limit, 0 otherwise)*

0.604 0.120 5.03 0.000 [0.369, 0.839]

Standard deviation of distribution 0.536 0.117 4.57 0.000 [0.307, 0.766] Horizontal alignment Maximum radius of curvature -0.183 0.060 -3.07 0.002 [-0.301,-0.066] Longitudinal grades Proportion of segment with longitudinal grades greater than zero*

0.086 0.155 0.55 0.582 [-0.219, 0.390]

Standard deviation of distribution 0.789 0.065 12.10 0.000 [0.661, 0.916]Maximum longitudinal grade > 8% indicator (1 if maximum longitudinal grade >8%, 0 otherwise)

0.216 0.087 2.49 0.013 [0.046, 0.387]

Cross-sectional elements Proportion of segment with bitumen -2.363 0.695 -3.40 0.001 [-3.725,-1.001]


Variables Estimate Std. Err.

z Prob. |z| >Z


shoulder Proportion of segment with one side shoulder width >1.5m

-0.540 0.191 -2.82 0.005 [-0.915,-0.165]

Roadway and roadside features Proportion of segment with embankments along one side

0.525 0.175 3.00 0.003 [0.182,0.868]


-0.477 0.094 -5.07 0.000 [-0.661,-0.292]

Spatial characteristics Number of houses/shops/commercial buildings per km*

0.005 0.002 2.12 0.034 [0.000, 0.010]

Standard deviation of distribution 0.014 0.002 7.33 0.000 [0.010, 0.018] Dispersion parameter negative binomial distribution

2.445 0.997 2.45 0.014 [0.490,4.399]

Log likelihood at zero -2194.91Log likelihood convergence -2002.45AIC 4042.90Mean Squared Error (MSE) 0.14Chi-sq. / DF 384.92/4P-value 0.000*random parameter

The parsimonious model identified 13 explanatory variables influencing SV

crashes along rural mountainous highways – all have plausible signs and magnitudes.

Ten of them were estimated as fixed parameters, and the other three turned out to be

random parameters. Variables estimated as fixed parameters include (1) ADT, (2)

segment length, (3) average visibility at the time of crash, (4) average hourly rainfall

at the time of crash, (5) maximum radius of curvature, (6) maximum longitudinal

grade, (7) proportion of segment with bitumen shoulder, (8) proportion of segment

with wide shoulder, (9) proportion of segment with roadside embankment, and (10)

presence of road delineation. The standard deviations of three parameters were found

to significantly differ from zero and thus they were estimated as random parameters.

They include (1) downgrade speeding indicator, (2) proportion of segment with

longitudinal grades greater than zero, and (3) number of houses, shops or commercial

buildings per km. The presence of random parameters indicates the existence of

unobserved heterogeneities around these parameters, and thus further indicates the

appropriateness of the RPNB model in the context of this study. Elasticity estimates

of significant variables in the SPF are presented in Table 5-3.


Both exposure variables, annual daily traffic (95% CI:0.27, 0.63) and segment

length (95% CI: 0.86, 1.28), are significant and positively associated with SV

crashes along rural mountainous highways. Elasticity estimates suggest that a 1%

increase in log of ADT is associated with about a 3.29% increase in SV crashes. SV

crashes are also found to increase by about 7.11% for a 1% increase in log of

segment length.

Table 5-3: Elasticity and pseudo-elasticity estimates of significant variables in SPF

Variables Elasticity / Pseudo-

elasticity (%)Log ADT 3.294 Log of segment length 7.106 Average visibility at the time of crash 1.008 Average hourly rainfall at time of crash 0.057 Downgrade speeding indicator (1 if 85th percentile vehicle speed along downgrade greater than the posted speed limit)

45.34

Maximum radius of curvature -0.154 Proportion of segment with longitudinal grades greater than zero 0.056 Maximum longitudinal grade > 8% indicator (1 if maximum longitudinal grade >8%, 0 otherwise)

19.43

Proportion of segment with bitumen shoulder -0.071 Proportion of segment with one side shoulder width >1.5m -0.113 Proportion of segment with embankments along one side 0.326 Presence of road delineation (1 if there are road delineations such as guide posts and chevron signs along the segment, 0 otherwise)

-61.12

Number houses/shops/commercial buildings within 100m buffer zone at each end of road segment per km

0.064

Average visibility at time of crash is found to be positively associated with

SV crashes (95% CI:0.06, 0.11). The elasticity estimate suggests that a 1% increase

in visibility is associated with about 1.01% increase in SV crashes.

The average hourly rainfall at time of crash is positively associated (95% CI:

0.09, 0.14) with SV crashes along rural mountainous highways. SV crashes are found

to increase by 0.06% for a 1% increase in average rainfall.

The parameter estimate for the variable downgrade speeding indicator is

found to be a normally distributed random parameter with mean 0.60 and standard

deviation 0.54, suggesting that the coefficient for this variable is positive for 87% of

road segments and negative for the other 13% of samples. The corresponding

elasticity estimate indicates that SV crashes increase by about 45% if the 85th

percentile driving speed along a mountainous highways segment is higher than the

posted speed limit.

The maximum radius of curvature is negatively associated (95% CI: -0.30, -

0.07) with SV crashes along rural mountainous roads. Results suggest that a one


percent increase in radius of curvature is associated with about a 0.15% decrease of

SV crashes on mountainous highways.

The parameter estimate for proportion of segment length with longitudinal

grades greater than zero was found to be normally distributed, with mean 0.09 and

standard deviation 0.79, implying that the corresponding relationship is positive for

54% of highway segments and negative for the remaining 46%. On average SV

crashes are found to increase by about 0.06% for every percent increase in segment

length with longitudinal grades greater than zero.

Among the indicator variables for magnitudes of longitudinal grade along

mountainous roads, a highway segment with the maximum longitudinal grade higher

than 8% is positively associated (95% CI: 0.05, 0.39) with SV crashes. A pseudo-

elasticity estimate of this variable indicates that a steep highway segment with

longitudinal grade more than 8% increases SV crashes as much as 19% compared to

segments with milder gradients.

Among the cross-sectional elements, the proportion of segment length with

bitumen shoulder is a significant (95% CI:-3.72, -1.00) predictor in the SPF and

negatively associated with SV crashes. Elasticity suggests that SV crashes decrease

by 0.07% for one percent increase in proportion of segment length with bitumen

shoulder.

Proportion of segment length with wide shoulder (>1.5m) along one side is

negatively associated (95% CI: -0.92, -0.17) with SV crashes. SV crashes are found

to decrease by 0.11% per one percent increase in the proportion of the segment with

wide shoulder along a side.

Proportion of segment length with embankment along a side is significant

(95% CI: 0.18, 0.88) in explaining SV crashes along rural mountainous roads. A 1%

increase in proportion of segment with one side embankment is associated with about

0.33% increase in SV crashes.

The presence of road delineation like chevron signs and guide posts is a

significant predictor (95% CI: -0.66, -0.29) and positively associated with safety.

The presence of road delineation along rural mountainous highways has been found

to reduce SV crashes as much as 61%.


Among the spatial factors, the number of houses, shops or commercial

buildings per km within 100m of the road edge of mountainous highway segments is

significant in explaining SV crashes. The parameter estimate has been found to be

normally distributed, with mean 0.004 and standard deviation 0.01, implying that SV

crashes increase with an increase in the density of houses or commercial buildings

for 66% of samples, but decrease for the other 34% of road segments.

5.6 DISCUSSION

SV crashes along rural mountainous highways are associated with a wide

range of factors, including horizontal and vertical alignment, real-time weather

conditions, traffic characteristics, cross-sectional elements, roadside features, and

spatial characteristics. Effects of these variables are comprehensively discussed and

contrasted with the findings from developed countries in the following subsections.

Average daily traffic (ADT) and the length of segment are positively

associated with SV crashes. The corresponding elasticity estimates suggest that they

are associated with about 3.29% and 7.11% of SV crashes, respectively, implying

that the risk of SV crashes increases with exposure. In general, traffic crashes are

positively associated with exposure (e.g., Ceder & Livneh, 1982; Chang, 2005), but

the relationship between SV crashes and exposure is not very straightforward. For a

set of two-lane rural highway segments (non-mountainous), Ivan et al. (2000) found

that SV crashes are higher along segments with low volume/capacity ratios, and

along segments with a better level of service (e.g., LOS A). Using hourly traffic

volume data from a set of non-mountainous road segments in the United States, Qin

et al. (2006) demonstrated that the relationship between exposure and SV crashes

may change from positive to negative depending on the time of day. However,

exposure measured as average annual daily traffic (AADT) was not significantly

associated with SV crashes along mountainous freeways in the United States (Yu &

Abdel-Aty, 2013b). In contrast, the relationship between SV crashes and exposure

(i.e., ADT and segment length) along rural mountainous highways in Malaysia, as is

the case of in this study, is positive and non-linear. As the relationship between

exposure and SV crashes is complex, and different levels of traffic volume along

mountainous roads may lead to different responses and adaptations by local drivers, a

proper understanding of other factors that contribute to SV crashes is required.


Mountainous areas are well known for their adverse weather conditions. This

study has identified that the average rainfall at the time of a crash increases the

likelihood of SV crashes, with a 1% increase in the average rainfall increasing SV

crashes by 0.06%. Similar findings are also reported elsewhere (Ma et al., 2015b; Yu

et al., 2013; Yu et al., 2015). Wet pavements offer less skid resistance (Colonna,

Berloco, Intini, Perruccio, & Ranieri, 2016), and the effects of wet pavements may

be more prominent for SV crashes along mountainous highways, as vehicles are

harder to control while negotiating curves or driving along steep slopes.

Visibility is another important weather factor in mountainous areas, as a

number of studies in the United States have reported that visibility is negatively

associated with crashes on mountainous roads (Ahmed et al., 2012; Ma et al., 2015b;

Yu et al., 2013; Yu et al., 2015) and elsewhere (Ahmed, Abdel-Aty, Lee, & Yu,

2014). It is argued that poor visibility increases the total crash rate along

mountainous roads mainly because of drivers facing difficulties in car-following or

lane-changing manoeuvers, which may lead to rear-end or side-swipe collisions. In

contrast, this study has found that visibility is positively associated with SV crashes

along rural mountainous highways in Malaysia. There are two explanations for this

condition. First, this may suggest that better visibility encourages higher speed along

rural roads, which may in turn increase the likelihood of SV crashes. In general,

driving speed increases with visibility (Colonna, Intini, Berloco, & Ranieri, 2016).

To illustrate the effect of visibility on SV crash risk, speeding-related SV crashes are

plotted across visibility at the time of the crash in Figure 5-1. It is evident that about

93% of speeding-related SV crashes occurred when visibility was higher than 10km.

Second, this finding may be related to the exposure variable ADT. It is possible that

traffic volume decreases as visibility goes down and vice versa. This could explain

why the finding of this research is different from other research where real-time

traffic together with real-time weather was considered (e.g., Ahmed et al., 2014;

Chen, Chen, & Ma, 2016; Yu & Abdel-Aty, 2013b). This possibility could be better

examined if real-time traffic data were available for this study.

The likelihood of SV crashes is found to be about 45% higher for those

highway sections where the 85th percentile driving speed is higher than the posted

speed limit. As such, downgrade speeding represents a significant safety concern for

rural mountainous highways. Drivers generally increase their speed along rural roads


because of low traffic volume (Lee et al., 2015), and the higher speed may more

common along downgrade sections. Therefore, appropriate countermeasures should

be targeted to control driving speed along downgrade sections of rural mountainous

highways. This variable is also estimated as a random parameter in the model,

suggesting that this parameter is negative for some road segments. This may suggest

the presence of unobserved heterogeneities around this parameter, perhaps indicating

the fact that some road segments are well designed to accommodate speeds higher

than the posted speed limit.

Figure 5-1: Frequency of speeding-related SV crashes under different visibility conditions

This study has identified that the likelihood of SV crashes is lower for

horizontal road segments with a large radius of curvature, with 1% increase in

maximum radius of curvature associated with about 0.15% decrease in SV crashes.

This is expected as a curve with a large radius represents a less complex situation

than curves with small radii. Other studies have also reported that a road section with

uniform horizontal curve radius is safer than a road with varying radii (Li et al.,

2014; Wang et al., 2010).

The proportion of segment length with longitudinal grades greater than zero

and the presence of a longitudinal grade higher than 8% are associated with increased

SV crashes. These findings are consistent with other research conducted in

3

30

359

87

9

12

0 50 100 150 200 250 300 350 400

< 5

5 - 10

10 - 15

15 - 20

20 - 25

>25

Crash count

Vis

ibil

ity

at t

he

tim

e of

cra

sh (

km

)


mountainous areas elsewhere. For example, Ahmed et al. (2011) reported that a

downgrade segment with slope 6 to 8% is more hazardous than smaller gradients. the

most hazardous compared to other gradients such as 4 to 6% and 2 to 4%. Similarly,

Yu et al. (2015) reported that the presence of a steep downgrade slope increases

crash risk. There are at least two possible reasons for the increased likelihood of SV

crashes along steep gradients. First, speed is likely to be higher for vehicles travelling

along downgrades. Second, continuous braking along downgrades of mountainous

roads is likely to increase the temperature of brake pads, which may eventually

degrade the efficiency of vehicle brakes and may increase crash risk. Results also

suggests that the variable ‘proportion of segment length with longitudinal grades

greater than zero’ is a normally distributed random parameter, and thus has an

opposite safety effect for about 46% of highway segments. This might again indicate

the presence of unobserved heterogeneities around this variable. For example, drivers

may be more cautious while driving down longitudinal grades; however, the SPF of

this study does not capture any driving behaviour factors. Investigating driving

behaviour along longitudinal grades of mountainous highways could be a worthwhile

future research topic.

Among the cross-sectional elements, the presence of a bitumen shoulder and

the presence of wide shoulders along rural mountainous highways have a positive

effect on road safety and are respectively associated with about 0.07% and 0.11%

less SV crashes. To examine the overall effect of road shoulders, the expected crash

frequencies of SV crashes are plotted against the proportion of segment length with

bitumen shoulder or wide shoulders (≥1.5m) and presented in Figure 5-2. It appears

that SV crashes decrease with the increase in proportion of segment length with a

bitumen shoulder, and the SV crash frequencies could be about 85% lower if

bitumen shoulders were present along 100% of mountainous highways, given all

other variables being equal. Sealed shoulders allow drivers to recover and redirect

their errant vehicle back onto the travelling lanes. In Australia, sealed shoulders are

reported to reduce casualty crashes (Jurewicz et al., 2015). The safety effect of wide

shoulders is also evident as Figure 5-2 clearly shows that the likelihood of SV

crashes is about 29% lower if a wide shoulder is present along the entire highway

segment. Similar safety effects of wide shoulders along non-mountainous highways

are reported elsewhere (e.g., Ivan et al., 1999). To further understand the combined

effect of shoulder type and width on SV crashes along rural mountainous highways


in Malaysia, a cross-table frequency analysis (Table 5-4) was conducted, accounting

for shoulder type (turf, paved, gravel & earth) and shoulder width (narrow vs. wide).

It appears that the odds of SV crashes along mountainous highway segments with

narrow shoulders (less than 1.5m) is about 41% lower (OR 0.59, 95%CI 0.04 – 0.87)

for paved shoulders compared to turf or unpaved shoulders. These results suggest the

safety benefits of paved shoulders along constrained road geometries of mountainous

highways, which often do not allow sufficient space for wider shoulders.

Figure 5-2: The relationship between SV crash frequencies and road shoulders

Table 5-4: Cross-tabulation analysis of shoulder type and width for SV crashes

Road Shoulder Type

Shoulder widthOR (95% CI) Chi-sq., p-value

Narrow (<1.5m) Wide (≥1.5m)

Turf* 119 (48.0) 548 (42.7) 1.00

Paved 39 (15.7) 304 (23.7) 0.59 (0.40-0.87) 2.66, p < 0.01

Gravel & earth 90 (36.3) 431 (33.3) 0.96 (0.71-1.30) 0.26, p = 0.799

* Reference category

The presence of embankments along the mountain side is a typical feature of

rural mountainous roads, but they represent a safety concern, as the proportion of

segment length with embankments is positively associated with SV crashes, with an

elasticity estimate of 0.33%. First, an embankment along a highway segment may

decrease sight distance, particularly if they are located along curved road sections.

Second, surface run-off is higher over the pavement next to an embankment, which

0.00

0.02

0.04

0.05

0.07

0.09

0 25 50 75 100

Exp

ecte

d c

rash

fre

qu

ency

Proportion of segment length

Bitumen shoulder


may pose additional risks because of hydroplaning or less skid resistance. This

finding merits further investigation.

The presence of road delineations such as chevron signs and guide posts

decreases the likelihood of SV crashes by about 61% along rural mountainous

highways. Proper curve delineations through chevron signs, curve warning signs, and

repeater arrows are well-established treatment options for improving safety along

non-mountainous roads (Charlton, 2007; Montella, 2009). The safe benefit of these

treatments might be more acute in mountainous areas where road geometries are very

tight and the visibility remains a major issue because of mountainous weather.

The number of houses or commercial buildings within 100m of the road edge

is heterogeneously associated with SV crashes, with positive association for 66% of

segments and negative association for the other 34%. This is an interesting finding

which may have captured the randomness of driver behaviour which was unobserved

in this study. Further investigation of driving behaviours and traffic movements

through residential or commercial developments in mountainous areas is, therefore,

required.

Chapter 6: Multi Vehicle Crashes 99

Chapter 6: Multi Vehicle Crashes

6.1 INTRODUCTION

Chapter 6 presents the results of Study 3, which aims to identify factors

contributing to Multi-vehicle (MV) crashes along rural mountainous highways in

Sabah, Malaysia. Study 3 addresses Research Question 3 of the current research

(What are the factors that contribute to occurrence of MV crashes along rural

mountainous highways?). Five datasets were used in this study: crash data,

topographical information, weather information, traffic characteristics, and field

survey data, such as road geometry, cross-sectional elements, roadside features, and

spatial characteristics.

This chapter is structured into five main sections. The first section (Section

6.1) gives a brief introduction to the study, followed by the objectives of this study in

Section 6.2. Section 6.3 and 6.4 present a description of the data and the modelling

results, respectively. The last section, Section 6.5, discusses the findings of this

study.

6.2 OBJECTIVES

The main objective of this study is to examine critical factors contributing to

MV crashes along rural mountainous highways. This is achieved by developing an

appropriate SPF to model MV crashes as a function of roadway geometrics, traffic

characteristics, real-time weather conditions, cross-sectional elements, roadside

features, and spatial characteristics along a set of rural mountainous highways in

Malaysia. There are several important contributions of this research. First, this paper

investigates MV crashes along rural mountainous highways, which have received

less attention in the literature compared to other types of highways. Second, this

study investigates an important road safety problem in the context of developing

countries, as MV crashes represent a substantial 35% of total crashes in mountainous

areas in Malaysia. The findings from similar research in western countries may not

be directly applicable to a developing country like Malaysia because of differences in

roadway designs, traffic characteristics, roadside environments and driver behaviour


factors. Third, the traffic safety research in developing countries often suffers from

limited data. An extensive effort has been made to collect relevant data to examine

MV crashes along selected rural mountainous highway segments in Malaysia. The

findings from this research can provide insights into the MV crashes along rural

mountainous highways in developing countries in general and in Malaysia in

particular. Fourth, this research investigates the suitability of a random parameters

count model to deal with excess zeros by comparing its performance with the state-

of-the-art models for excess zeros including NB-L and NB-GE. Despite the wide

applications of random parameters models to address heterogeneities arising from

various sources, their capability of addressing the excess zeros problem is still not

known.


Five different datasets have been merged to construct the final unique datasets

shown in Table 6-1. These include crash data, topographical information, weather

conditions, traffic volume, and data from field surveys, such as road geometry, cross-

sectional elements, roadside features, and spatial characteristics. Section 3.2 details

the procedures used for collecting these datasets.

Table 6-1: Summary statistics of variables included in the model



Exposure Variables

Average daily traffic (veh/day) 1682.70 830.76 315.73,527

5.13- -

Road segment length (m) 881.58 385.24 30.00,

2039.00- -

ADT x segment length 13.94 0.876 10.12, 15.93

- -

Real – time weather condition Average visibility during the crashes (km)

12.21 1.69 0.50, 24.00 - -

Heavy rainfall indicator at time of crash (if rainfall in 1-hour is greater than 5.08mm, 0 otherwise)

- - - 10 1.96

Heavy rainfall indicator during the hour before the crash (1 if 1-hour amount of rainfall during the hour before the crash is greater than 5.08mm, 0 otherwise)

- - - 7 1.37


- - - 45 8.82




Downgrade speeding indicator (1 if 85th percentile vehicle operating speed greater than the posted speed limit)

- - - 40 7.84

Horizontal alignment Proportion of segment with horizontal curve

0.52 0.17 0, 1 - -


0.40 0.20 0, 1 - -

Proportion of segment with reverse curve

0.06 0.13 0, 0.66 - -

Proportion of segment with compound curve

0.04 0.09 0, 0.45 - -

Proportion of segment with broken back curve

0.03 0.12 0, 0.61 - -

Maximum degree of curvature (°) 17.54 17.44 0, 104.00 - -Minimum degree of curvature (°) 64.23 42.15 0, 330.00 - -Maximum radius of curvature (km) 0.84 0.97 0, 5.99 - -Minimum radius of curvature (km) 0.23 0.50 0, 3.66 - -Maximum length of circular curve (km)

0.15 0.08 0, 0.40 - -

Minimum length of circular curve (km)

0.09 0.07 0, 0.38 - -

Longitudinal grades Proportion of segment with longitudinal grades greater than zero

0.65 0.34 0, 1 - -

Maximum longitudinal grade <2% indicator (1 if maximum longitudinal grade <2%, 0 otherwise)

- - - 30 5.88


- - - 75 14.71


- - - 80 15.69


- - - 100 19.60

Maximum longitudinal grade >8% indicator (1 if maximum longitudinal grade >8%, 0 otherwise)

- - - 225 44.12

Combination of horizontal and vertical alignment indicator;

Category 1: (1 if 50% or less of a segment have horizontal curve and absolute gradient ≤ 4%, 0 otherwise)

- - - 60 11.76

Category 2: (1 if more than 50% of a segment has horizontal curve and absolute gradient ≤ 4%, 0 otherwise)

- - - 45 8.82




Category 3: (1 if 50% or less of a segment has horizontal curve and absolute gradient >4%, 0 otherwise)

- - - 120 23.53

Category 4: (1 if more than 50% of a segment has horizontal curve and absolute gradient >4%, 0 otherwise)

- - - 285 55.88

Cross-sectional elements Proportion of segment with broken centre line

0.72 0.32 0, 1 - -

Proportion of segment with concrete shoulder

0.09 0.20 0, 1 - -


0.03 0.10 0, 0.52 - -

Proportion of segment with crusher run and earth shoulder

0.14 0.23 0, 1 - -

Proportion of segment with turf shoulder

0.73 0.31 0, 1 - -

Proportion of segment with one side shoulder width >1.5m

0.21 0.33 0, 1 - -

Proportion of segment with both sides shoulder width >1.5m)

0.65 0.39 0, 1 - -

Proportion of segment with both sides shoulder width <1.5m

0.14 0.28 0, 1 - -

Presence of overtaking lane (1 if there is an overtaking lane along the segment, 0 otherwise)

- - - 900 14.70

Roadway and roadside features Number of minor intersections 0.80 1.25 0, 7.00 - - Number of roadway lighting poles per km

1.52 4.32 0, 28.83 - -

Proportion of segment with guardrails along one side

0.17 0.17 0, 0.72 - -

Proportion of segment with guardrails along both sides

0.02 0.05 0, 0.34 - -

Proportion of segment with embankments along one side

0.62 0.34 0, 1 - -


0.16 0.24 0, 1 - -

Proportion of segment with cliffs along one side

0.49 0.32 0, 1 - -


0.05 0.09 0, 0.39 - -

Presence of bridge (1 if there is a bridge along the segment, 0 otherwise)

- - - 50 9.80


- - - 210 41.18

Spatial characteristics Proportion of segment with forest within 100m from the road edge

0.75 0.30 0, 1 - -

Proportion of segment with farm/ agricultural within 100m from the

0.10 0.21 0, 1 - -




road edge Proportion of segment with house/shop/commercial building within 100m from the road edge

0.14 0.20 0, 1 - -


6.4 METHODOLOGY

In this study, three different models were developed to compare the

performance of each models for handling the excess zero problem. Out of 510

observations, this dataset contains 356 (70%) zero observations. Two different

models were selected to address this problem, including the Negative Binomial –

Lindley (NB-L) and Negative Binomial – Generalized Exponential (NB-GE).

Previous research has shown that these models performed better than the Negative

Binomial when the data has a high percentage of zeros. Another issue related to the

crash dataset is unobserved heterogeneity. The Random Parameters Negative

Binomial (RPNB) model was selected in this study for observing heterogeneity. As

mentioned in Section 2.7.1, heterogeneities can be divided into two categories;

structured and unstructured. Structured heterogeneities occur in panel datasets

(longitudinal data) where the same segment was observed multiple times (in this

study is 5 times). The unstructured heterogeneities occur when there is

misspecification in the model, uncertainty in exposure and covariates, and omitted

variables. These models used different analysis platforms, so the ordinary estimation

cannot be used to compare these models. As an alternative, three different global

goodness-of-fit estimations were applied here including Mean Absolute Deviation

(MAD), Mean Squared Predictive Error (MSPE) and Mean Squared Error (MSE).

6.5 MODEL RESULTS

The RPNB model was estimated using a MSLE approach with 200 Halton

draws in Nlogit while the NB-L and NB-GE models were estimated using the

Bayesian approach and MCMC simulation in WinBUGS. The MCMC simulation for

the NB-L and NB-GE models resulted in two Markov chains converging after 30,000

iterations. The convergence was ensured by visual monitoring (obtaining stabilized

and well-mixed chains) as well as assessing the Gelman-Rubin statistics (RGelman-

Rubin→1). The simulation was continued for another 50,000 iterations for making

reliable inferences about regression parameters. Table 6-2 presents the regression


results of the three candidate models. The results are discussed in two parts. The first

part presents a comparison between the three modelling options in terms of their fit

and prediction ability. The second part provides more details about model parameters

and presents a deeper insight into factors contributing to MV crashes along rural


Figure 6-1: Adjusted cumulative residual plots for exposure variable.


Table 6-2: Modelling results for MV crashes along rural mountainous highways

Variables

RPNB NB-L NB-GEMean (Std. Dev)


Mean (Std. Dev)

[95% cred. interval]

Mean

(Std. Dev) [95% cred.

interval]

Constant -8.245 [-11.222,-

5.267] -7.666 [-10.490,-

3.927] -9.187 [-12.190,-

6.045] Exposure variable Log (ADT x segment length) 0.514 [0.304,0.725] 0.473 [0.257,0.674] 0.586 [0.010,0.349] Real-time weather conditions Heavy rainfall indicator at time of crash (if rainfall in 1-hour is greater than 5.08mm, 0 otherwise)

0.900 [-0.042,1.842] 0.912 [0.029,0.065] 0.882 [0.188,1.52]

Longitudinal grades Combination of horizontal and vertical alignment indicator; Category 4: (1 if more than 50% of a segment have horizontal curves and absolute gradient >4%, 0 otherwise)*

0.017 [-0.292,0.326] - - - -

Standard deviation of distribution 0.419 [0.187,0.652] - - - - Cross-sectional elements Proportion of segment with bitumen shoulder - - -2.997 [-6.632,-0.646] -3.139 [-6.276,-0.367] Presence of overtaking lane (1 if there is an overtaking lane along the segment, 0 otherwise)

-1.055 [-1.868,-0.241] -1.072 [-1.902,-0.092] -0.955 [-1.851,-0.166]

Roadway and roadside features Number of minor intersections 0.210 [0.124,0.297] 0.189 [0.012,0.373] 0.165 [0.019,0.318] Presence of road delineation (1 if there are road delineations such as guide posts and chevron signs along the segment, 0 otherwise)*

-0.213 [-0.526,0.099] - - - -

Standard deviation of distribution 0.609 [0.336,0.883] - - - -


Variables

RPNB NB-L NB-GEMean (Std. Dev)


Mean (Std. Dev)

[95% cred. interval]

Mean

(Std. Dev) [95% cred.

interval]

Number of parameters 6 5 5MAD 0.568 0.575 0.581MSPE 0.793 0.801 0.809MSE 0.796 0.814 0.822


To visually demonstrate the prediction ability of the three models, Cumulative

Residuals (CURE) were plotted against increasing order of the exposure factor.

CURE plots are helpful tools for demonstrating a model fit with respect to its

covariates (e.g. exposure) and identifying any potential and systematic bias (e.g.

over/under prediction) (Hauer, 2015). A superior fit occurs when the plots are

oscillating more closely to zero (horizontal axes). Furthermore, excess oscillations

above/under the zero axes can be a sign of under/over prediction and thus a less

biased model has an equal amount of positive and negative residuals. Figure 6-1

presents the adjusted CURE plots for the three modelling options and shows that all

of the three models resulted in cumulative residuals oscillating around zero, and

maintained a balance between positive and negative sides. Furthermore, the CURE

plots of all of the three models stayed within the 95% boundaries ( σ2± ) of

cumulative residuals which indicates their good fit with respect to exposure.

As reported in Table 6-2, the RPNB model yields lower values of MAD

(0.568), MSPE (0.793) and MSE (0.796) compared with the other two models,

showing that this model outperformed the others in terms of prediction ability. NB-L

is the second ranked model in terms of fit with MAD, MSPE and MSE values equal

to 0.575, 0.801 and 0.814, respectively.

Four of the variables are common between the three models: exposure, real-

time weather conditions, the presence of an overtaking lane, and the number of

junctions. While the sign of estimated parameters for such variables is the same

across models, the magnitude of the parameters is slightly different from one model

to another. While the NB-L and NB-GE models estimated an additional variable,

proportion of segment with bitumen shoulder, the RPNB model identified two

additional significant explanatory variables, which were combination of horizontal

and vertical alignment and presence of road delineation. The notable finding of this

study is that the parameters of these two variables were estimated as random, which

elegantly shows the ability of the RPNB model to capture the unobserved

heterogeneity across observations.

The RPNB was therefore selected as the superior model based on its goodness-

of-fit and capability of capturing heterogeneities. To investigate the effects of

contributing factors on MV crashes along rural mountainous highways, their

elasticities were calculated using the RPNB estimates and presented in Table 6-3.


Table 6-3: Elasticity and pseudo-elasticity for crash contributing factors of the RPNB model

Variables Elasticity/ Pseudo-

elasticity

Log (ADT x Segment length) 7.17

Heavy rainfall indicator (if rainfall in 1-hour is greater than 5.08mm, 0 otherwise)

59.34*



1.69

Presence of overtaking lane (1 if there is an overtaking lane along the segment, 0 otherwise)

-187.20*

Number of minor intersections 0.17


-23.74*

*Pseudo-elasticity

Among three variables related to real-time weather conditions, only heavy

rainfall indicator at time of crash was found to be statistically significant (95% CI:

0.04, 1.84) and positively associated with MV crashes. The corresponding pseudo-

elasticity indicates that heavy rainfall increases the likelihood of MV crash

frequencies as much as 59% along rural mountainous highways.

The combination of horizontal and vertical alignment (with the grade higher

than 4%) is found to be randomly distributed with the mean of 0.017 and standard

deviation of 0.419. Given these distributional parameters, 52% of highway segments

are positively associated with MV crashes and the remaining 48% are negatively

associated. This distinction in the heterogeneous effects of this variable is a direct

consequence of using Random Parameters modelling technique. The corresponding

pseudo-elasticity indicates that MV crashes increase by an average of 2% if there is a

combination of horizontal and vertical alignments on mountainous highway

segments.

The presence of an overtaking lane along highway segments is found to be a

significant predictor (95% CI: -1.87, -0.07) which is negatively associated with MV

crash frequencies in mountainous areas. The pseudo-elasticity estimate suggests that

the presence of an overtaking lane along rural mountainous highways can reduce MV

crash frequencies by 46.5%.

Number of junctions is another significant (95% CI: 0.12, 0.30) predictor of

MV crashes along rural mountainous highways. The estimate of this parameter is


found to be positively associated with MV crash frequencies, and the corresponding

elasticity suggests that a one percent increase in the number of junctions results in

about a 0.17% increase in MV crashes.

The presence of road delineation is found to be normally distributed with mean

-0.213 and standard deviation 0.609, implying that the corresponding relationship is

positive for 64% of highway segments and negative for the rest. The corresponding

elasticity estimate indicates that the presence of road delineation along rural

mountainous highways reduces MV crash frequencies by about 24%.

6.6 DISCUSSION

This study has compared three state-of-the-art count modelling techniques,

Random Parameters Negative Binomial (RPNB), Negative Binomial – Generalized

Exponential (NB-GE) and Negative Binomial – Lindley (NB-L) models, on a dataset

with excess zeros. In addition to the methodological contribution, this study has also

examined MV crashes using a comprehensive set of exogenous variables, including

real-time weather conditions, traffic characteristics, roadway geometric

characteristics, cross-sectional elements, roadside features, and spatial

characteristics. The following subsections discuss the findings of this study, with an

emphasis on contrasting the effects of significant variables with findings from

developed countries.

6.6.1 Excess Zeros

In handling a dataset with excess zeros (containing 70% of zero observations),

this study has found that the RPNB significantly outclasses the NB-GE and NB-L

models. The estimates of three global goodness-of-fit criteria; MAD, MSE and

MPSE show that RPNB has a smaller value compared to other models. In addition,

the CURE plots also show that the cumulative residual for this model is closer to

zero and maintains balance between positive and negative sides. In a previous study

by Vangala et al. (2015) comparing the performance of NB-L, NB-GE and NB

models for a dataset containing about 90% zero observations, reported that NB-L and

NB-GE perform significantly better than the NB model, and that these two models

have similar performance. The suitability of the NB-L model in handling excess

zeros has also been reported elsewhere (Lord & Geedipally, 2011; Zamani & Ismail,

2010a). The present study shows that the RPNB model offers an even a better


statistical fit than the NB-L model. A worthwhile future research direction would be

testing the performance of the RPNB model with other datasets with excess zeros.

6.6.2 Exposure Variable Effect

Average daily traffic (ADT) and segment length were multiplied to estimate

the daily VMT and to capture the crash exposure for every highway segment. The

positive association of log of exposure and MV crashes is intuitive and consistent

with other studies in the literature about the effects of exposure on crashes along

mountainous freeways (Yu & Abdel-Aty, 2013b). For non-mountainous road

segments in Michigan and Connecticut states in the United States, Qin et al. (2006)

have also reported the non-linear relationship between traffic volume and three

different types of MV crashes, including same direction, opposite direction and

intersecting direction. In contrast, Anastasopoulos and Mannering (2009) reported

that Annual Average Daily Traffic (AADT) is a random parameter for crashes along

rural interstate highways in Indiana, with an increasing effect on crash frequencies

for the majority of road segments and a decreasing effect for a small proportion of

road segments. It indicates that the different level of exposure may lead to different

safety outcomes along rural highways. As such, a proper understanding of crash

contributing factors is important particularly for rural highways.

6.6.3 Real-Time Weather Conditions

Weather conditions play a vital role in road safety in mountainous areas (Yu

et al., 2015). This study has identified that heavy rainfall (hourly rainfall greater than

0.508mm) at the time of crash increases the likelihood of MV crashes along rural

mountainous highways. Heavy rainfall was also found to increase the crash risk

along mountainous sections of the I-70 freeway in Colorado in US (e.g., Yu et al.,

2013; Yu et al., 2015). The rainfall data from the twelve rainfall stations around the

mountainous highway segments of this study showed that that more than 60% of the

months of the study period recorded more than 15 rainy days per month. Driving in

rainy conditions may be normal for local drivers, but heavy rainfall may increase the

risk of crashes, as it reduces sight distance and limits vehicle manoeuvrability along

horizontal and vertical curve. A further analysis of MV crash types along the selected

segments of this study shows that head-on crashes are 4.45 times (95%CI 1.03-19.16,

p<0.05) higher during heavy rainfall compared to rear-end crashes, indicating the


increasing likelihood of lane encroachment along two-lane two-way mountainous

highways.

6.6.4 Horizontal alignment and Longitudinal Grades

The combined effect of horizontal alignment and longitudinal grades on MV

crashes is interesting. It is found that mountainous segments with an absolute

gradient more than 4% and that having more than 50% of their length with

horizontal curve have a mixed effect on MV crashes, with about 52% of sample

segments being associated with a higher likelihood of MV crashes, and the remaining

48% associated with lower likelihood of MV crashes. Yu et al. (2015) reported that

curve segment indicator and steep downgrade indicator (more than 4% in absolute

gradient) as separate variables are associated with a higher probability of crashes

along the I-70 freeway in Colorado. Interestingly, both variables were identified as

random parameters. The negative effect of both horizontal curves and steep

longitudinal grades can be explained by the fact that the presence of a horizontal

curve acts as a traffic calming measure, controlling driving speed along downgrades

and thus reducing the likelihood of crashes. An in-depth field study on driving speed

along mountainous road segments would be an interesting future research project.

6.6.5 Cross-Sectional Elements

The presence of an overtaking lane along rural mountainous highways is

negatively associated with MV crashes. Overtaking lanes provide the required space

for vehicles in a queue (mainly towards upgrade directions along mountainous

highways) to safely overtake slow heavy vehicles without trying to find a gap in the

traffic in the opposite lane of a two-lane two-way highway. Overtaking lanes are

reported to decrease fatal and injury crashes on two-lane two-way non-mountainous

highways (Frost & Morrall, 1995; Schumaker, Ahmed, & Ksaibati, 2016).

Overtaking lanes along mountainous highways might be even more effective, as it is

generally difficult to ensure sufficient sight distance required for safe overtaking

using the opposite lane along two-lane two-way mountainous highways.

6.6.6 Roadside Features

The number of minor junctions increases the likelihood of MV crashes along

rural mountainous highways. Minor junctions along rural highways in a developing

country like Malaysia provide access to local villages. These minor junctions may


increase conflicts of traffic with a large speed differential and hence increase the

likelihood of crashes. Berhanu (2004) reported that the number of minor junctions

per unit length of road significantly increased the MV crash frequencies along

undivided arterial roads in Addis Ababa, demonstrating the relationship between

access and safety in the context of a developing country.

Road delineations are very important to guide drivers through mountainous

highways where the road alignment changes very quickly. This study has shown that

the presence of road delineation (e.g. chevron signs and guideposts) decreases the

likelihood of MV crashes along rural mountainous roads in Malaysia. In Wyoming,

US, the delineators are reported to reduce all types of injury crashes, including fatal,

injury and property damage only crashes along rural roads (Ksaibati, Evans, &

Shinstine, 2015). A study in Spain (Gross, Eccles, & Nabors, 2011) also reported that

delineations like advanced curve warnings, chevrons, and post mounted delineators

are effective in reducing crashes along horizontal curves on low-volume roads. The

delineations are more important for mountainous roads to guide drivers through

changing road alignments, particularly at night and during adverse weather

conditions (e.g. rain, fog). In the context of a developing country, road safety

treatments using delineations may be particularly attractive, due to their low cost.

In conclusion, this study successfully developed an appropriate safety

performance function for MV crashes along rural mountainous highways. The

parsimonious RPNB model identified six significant variables affecting MV crashes

along rural mountainous roads in Malaysia. While exposure (ADT x segment length),

heavy rainfall at time of crash, presence of horizontal curves along a steep gradient

and number of junctions were positively associated with MV crashes, the presence of

an overtaking lane and the presence of road delineations were negatively associated.

Among these, the presence of horizontal curves along a steep gradient and the

presence of road delineations were identified as random parameters, suggesting

unobserved heterogeneities around them. Findings of this study provide considerable

insights into the factors affecting MV crashes along rural mountainous highways in

Malaysia, which will be helpful for developing effective countermeasures to improve

road safety in mountainous areas.

Chapter 7: Crash Severity Study 113

Chapter 7: Crash Severity Study

7.1 INTRODUCTION

This chapter presents the results of Study 4, which is the last study of the

current research. Previous studies examined crash characteristics and identified

factors influencing single- and multi-vehicle crashes along rural mountainous

highways in Sabah, Malaysia. The current study is a further investigation that

addresses the Research Question 4 (What are the factors related to the crash severity

of traffic crashes along rural mountainous highways?). This study used a

combination of information from crash reports, secondary data from related agencies,

and field survey data to answer this research question.

This chapter is organised into six sections. Section 7.1 gives a brief

introduction to the current study and chapter. Section 7.2 and 7.3 present the

objectives of this study and data description. Section 7.4 discusses the limited

availability of data for conducting this study and the solution to this problem. Section

7.5 details the results for the decision tree analysis and crash severity modelling.

Section 7.6 discusses the findings of this study.

7.2 OBJECTIVES

The objective of this study is to investigate the injury severity of road traffic

crashes along the rural mountainous highways. The factors investigated include

variables related to collision types, crash temporal characteristics, driver and vehicle

characteristics, weather conditions, road geometry and cross-sectional elements, and

roadside features. To consider the imbalance in the response variable and unobserved

heterogeneity in crash data, three logistic models are employed including standard

logit, Scobit and random parameters logit models. To avoid specification errors, a

two-step modelling approach will be used where the interactions among a set of

variables from the decision tree will be used in the logistic regression model, along

with other prospective variables to improve the model’s predictive power.



The three different data sources used in this study are crash reports, weather

information, and field survey data. Crash reports are obtained from the Malaysian

Institute of Road Safety Research – Road Accident Analysis and Database System

(M-ROADS) for a five-year period (2008-2012). This dataset contains crash

information on crash severity, collision type, time of the crash, and driver(s) and

vehicle(s) involved in the crash. Detailed information about the weather and field

survey data are provided in Section 3.2, above. Table 7-1 presents the descriptive

statistics of the explanatory variables included in this study.

A total of 972 crashes occurred within the study period on the selected

highways segments. Injury severity of traffic crashes in Malaysia is classified into

four categories; fatal, serious injury, slight injury, and property damage only (PDO).

The severity of a crash refers to the injury severity of the most severely injured

person in the crash. The dataset contains a total of 15 fatal crashes, 25 serious

injuries, 26 crashes with slight injuries and 906 PDO crashes. As the number of fatal,

serious, and slight injuries is very small compared to PDO crashes, they have been

combined for purpose of this study into a single category, severe injury, while PDO

crashes are considered non-severe. Therefore, this study categorized injury severity

into two main categories where severe injuries represent only 7% of total crashes and

non-severe injuries represent 93% of total crashes. The high portion of PDO crashes

are probably explained by the fact that in Malaysia it is compulsory to report a crash

to the police for purposes such as insurance claims (PIAM, 2013). This creates an

unbalanced distribution between the two categories of injury severity.

Table 7-1: Summary statistics of variables included in the model


Min, Max

Count* Percentage*

Horizontal alignment

Proportion of segment with horizontal alignment

0.50 0.17 0, 1 - -

Proportion of segment with simple curve 0.40 0.18 0, 1 - - Proportion of segment with reverse curve 0.05 0.12 0, 0.66 - - Proportion of segment with compound curve

0.03 0.09 0, 0.45 - -

Proportion of segment with broken back curve

0.02 0.10 0, 0.61 - -

Longitudinal grades

Proportion of segment with longitudinal grades

0.55 0.32 0, 1 - -



Min, Max

Count* Percentage*

Steep gradient indicator (1 if maximum longitudinal grade >8%, 0 otherwise)

- - - 413 42.49


Category 1: (1 if 50% or less of a segment has horizontal curve and absolute gradient ≤ 8%, 0 otherwise)

- - - 283 29.12


- - - 276 28.40


- - - 148 15.23


- - - 265 27.26

Cross-sectional elements

Proportion of segment with both sides shoulder width >1.5m

0.19 0.32 0, 1 - -

Proportion of segment with unsealed shoulder

0.88 0.22 0, 1 - -

Roadway and roadside features Proportion of segment with guardrails along both sides

0.02 0.04 0, 0.23 - -


0.14 0.20 0, 1 - -


0.06 0.10 0, 0.39 - -

Weather conditions

Rain indicator (1 if rain at time of crash, 0 otherwise)

- - - 265 27.26

Temporal characteristics

Nighttime indicator (1 if crash in nigthtime, 0 otherwise)

- - - 384 39.51

Weekend indicator (1 if crash in weekend, 0 otherwise)

- - - 296 30.45

Crash type

Single-vehicle crashes (1 if single-vehicle crashes, 0 otherwise)

- - - 715 73.56

Collision type

Rear-end collision indicator (1 if rear-end collision, 0 otherwise)

- - - 128 13.17

Out-of-control collision indicator (1 if out-of-control collisions,0 otherwise)

- - - 535 55.04

Head-on collision indicator (1 if head-on collision, 0 otherwise)

- - - 43 4.42

Angle and right angle side (1 if angle and right angle side collision, 0 otherwise)

- - - 87 8.95

Side swipe (1 if side swipe collision, 0 otherwise)

- - - 42 4.32

Driver and vehicle factors Young driver involvement indicator (1 if young driver involve, 0 otherwise)

- - - 137 14.09



Min, Max

Count* Percentage*

Female driver involvement indicator (1 if female driver involve, 0 otherwise)

- - - 134 13.79

Heavy vehicle involvement indicator (1 if heavy vehicle involve, 0 otherwise)

- - - 227 23.35

Speeding indicator (1 if speeding, 0 otherwise)

- - - 562 57.82


7.4 DATA AVAILABILITY

Injury severity models attempt to establish the relationship between injury

severity and various contributing factors, including driver, traffic and vehicle

characteristics, weather conditions, road geometry, roadside features, and crash

types. Many studies have combined these outcomes into two categories (severe and

non-severe) due to the small counts for certain severity levels (e.g., Yu & Abdel-Aty,

2014a, 2014b). A standard binary logit model is often used to model these binary

outcomes. However, a low share of severe injuries might create imbalance in the

response variable, in which case, a Scobit model might outperform a standard logit

model (Tay, 2016).

The Scobit model was introduced by Nagler (1994) as an alternative to the

standard binary logistic model in order to allow for imbalance among the categories

of the response variable. The argument was made that the distribution of error term

in the standard binary logistic model follows logistic (logit) or normal (probit)

distribution where it is unimodal and symmetric. In these distributions, it is assumed

that the sensitivity of observational changes in explanatory variables is highest for

the crashes where indifferent preferences over severe and non-severe (i.e., the choice

of probability is 0.5). However, in reality, the proportion of severe injury crashes is

often substantially less than half the total. Estimation for marginal effects could then

be miss-specified because these estimates are derived not only from estimated

parameters, but also the form of choice probability (Wu, Zhang, & Fujiwara, 2013;

Wu, Zhang, Fujiwara, & Chikaraishi, 2012). An alternative to this is to provide a

more flexible distribution that allows for skewness or non-symmetry (Tay, 2016).

Furthermore, crash data contains heterogeneity between observations, that is,

the same factors might not lead to the same outcome for a different crash. To

circumvent the unobserved heterogeneity in the injury severity model, a random

parameters or mixed logit model is often used. For example, Chen and Chen (2013)


discovered that 56.5% of the mountainous crashes that occurred on snow-covered

roads result in a decrease in possible injury/non-incapacitating injury crashes, while

the rest result in an increase in possible injury/non-incapacitating injury crashes.

They concluded that it is difficult to observe the complex interaction and random

nature of the parameters without application of a random parameters model. Based

on goodness-of-fit estimation, this model performs better than a fixed parameter

model (e.g., Chen & Chen, 2013; Christoforou, Cohen, & Karlaftis, 2010; Milton et

al., 2008). Although random parameter models can capture heterogeneity in the

model, their performance for imbalanced response variables is not fully identified in

previous research.

Another issue in injury severity modelling is the specification errors

(Washington, 2000). It occurs when a model represents an incorrect relationship

between dependent and independent variables, and includes incorrect variables

including interactions between variables, non-additive effects, and nonlinearities. To

avoid this error, the analyst should have prior knowledge based on the theoretical

background and empirical findings from previous research to develop a model that

has acceptable function form and plausible interaction. However, this process

requires great needs a huge effort to rank interactions, and to identify interactions

between variables. Another way to avoid such error is to create a decision tree to

systematically identify the relationship between dependent and independent

variables. Each tree branch shows the interactions among a set of independent

variables available in that particular branch, and their relationship with the dependent

variable. This method was recently introduced to road safety studies to determine

factors influencing injury severity (e.g., Kashani & Mohaymany, 2011; Li, Ranjitkar,

Zhao, Yi, & Rashidi, 2016; Prato, Bekhor, Galtzur, Mahalel, & Prashker, 2010).

7.5 METHODOLOGY

This study used a two-step approach in analysis; decision tree and logistic

regression. The decision tree was used to identify the complexities of variables and

injury severity, while logistics regression was used to identify inference of the

parameters. Chi-Squared Automatic Interaction Detection (CHAID) was selected for

the decision tree analysis. This technique was developed by Kass (1980) as a data

mining algorithm. In this analysis, the performance of CHAID was also compared

with other decision tree communities such as Classification and Regression Trees


(CRT) and Quick, Unbiased, Efficient, Statistical Trees (QUEST). However,

presenting the results of this analysis is not in the scope of this research.

In terms of logistic regression, three different models were developed to

handling imbalance response variable and unobserved heterogeneity. These models

including standard logit model as a base model, Scobit or skewness model for

imbalance response variable and random parameters logit model for observed

heterogeneity. Details explanation about these models can be refer in Section 3.6.2.

7.6 RESULTS

Results from the decision tree analysis revealed some interesting interactions among

the variables. A total of six interactions were found to be significant among all the

prospective variables listed in Table 7-1. Each of these interactions are represented as

a branch of the tree diagram shown in Figure 7-1. Interaction 1 (the left-most branch

of the tree diagram) represents a crash when light and medium vehicles get involved

in a single-vehicle crash along a highway segment with higher proportions of its

length covered by curves. Interaction 2 is similar to the former except that it is a

multi-vehicle crash. Interaction 3 also involves light and medium vehicles, but along

a highway segment with a lower proportion of simple curves. Crashes involving

heavy vehicles on steep highways with the presence of horizontal curves is

represented by interaction variable 4. Interaction 5 and 6 both represent a crash when

heavy vehicles get involved in a single-vehicle and multi-vehicle crashes, along

segments with less than 8% of vertical longitudinal grades and less than 50% of their

length being horizontal curves, respectively.

Table 7-2 presents the estimation results of the standard logit, Scobit and

random parameters logit models. Based on the goodness-of-fit measures (e.g. AIC

and Log likelihood), the random parameter logit model performs better than the other

models. In addition, this model also captures five random parameters. The following

discussion in Section 7.7 will be based on the results of the random parameters logit

model.

The proportion of segment length with simple curves, presence of horizontal

curves along a steep gradient, unpaved shoulder and proportion of segment length

with cliffs along both sides are found to be positively associated with severe crashes.

Crashes occurring during rainy conditions are found to be negatively associated with


severe crashes. Among the parameters of collision types, the head-on collision is

found to be positively associated with severe crashes while the opposite relation is

observed for rear-end collision. Female drivers are found to be negatively associated

with severe crashes. On the other hand, crashes involving heavy vehicles are

positively associated with severe crashes. Among the six interaction variables found

from the decision tree analysis, interaction variable 1 is negatively associated with

severe crashes, while interaction variable 5 is found to be positively associated. A

detailed discussion of these parameter estimates are presented in the next section.


Figure 7-1 : Decision Tree


P value corrected = 0.001, Chi-square =14.518 df = 1

Combination of horizontal and vertical alignment (>8%)

P value corrected = 0.017, Chi-square = 5.688, df = 1

Heavy vehicle involvement

P value corrected = 0.000, Chi-square = 29.002, df = 1

<=0.4932 >0.4932 Yes No

No Yes

[1] [2]

[4][3]

[5] [6]

Single-vehicle crashes


Single-vehicle crashes


Yes Yes NoNo


Table 7-2 : Estimation results for standard logit, Scobit, and random parameters logit models

Variables Standard logit Scobit Random Parameters logit

Coeff. S.E Marginal

effect Coeff. S.E

Marginal effect

Coeff. S.E Marginal

effect Constant -3.042 0.313 - -2.288 1.443 - -3.591 0.908 - Horizontal alignment Proportion of segment with simple curve* - - - - - - 1.238 0.750 0.013 Standard deviation of distribution - - - - - - 1.889 0.394 - Longitudinal grades Combination of horizontal and vertical alignment indicator; Category 4: (1 if more than 50% of a segment has horizontal curve and absolute gradient >8%, 0 otherwise)

1.266 0.328 0.070 1.393 0.479 0.070 1.060 0.330 0.011

Cross-sectional elements Proportion of segment with unsealed shoulder* - - - - - 0.278 0.822 0.003 Standard deviation of distribution - - - - - 1.218 0.205 - Roadway and roadside features Proportion of segment with cliffs along both sides 4.147 1.445 0.230 4.809 2.293 0.241 4.381 1.379 0.045 Weather conditions Rain indicator (1 if rain at time of crash, 0 otherwise)* - - - - - - -0.216 0.304 -0.002 Standard deviation of distribution - - - - - - 0.994 0.359 - Collision type Rear-end collision indicator (1 if rear-end collision, 0 otherwise) -1.199 0.503 -0.066 -1.309 0.602 -0.066 -1.131 0.468 -0.012 Head-on collision indicator (1 if head-on collision, 0 otherwise) 1.382 0.397 0.077 1.646 0.852 0.082 1.453 0.387 0.015 Driver and vehicle factors Female driver involvement (1 if female driver involve, 0 otherwise)*

- - - - - - -5.534 2.401 -0.056


Variables Standard logit Scobit Random Parameters logit

Coeff. S.E Marginal

effect Coeff. S.E

Marginal effect

Coeff. S.E Marginal

effect Standard deviation of distribution - - - - - - 6.046 2.000 - Heavy vehicle involvement indicator (1 if heavy vehicle involve, 0 otherwise)

0.749 0.327 0.041 0.806 0.390 0.040 0.703 0.326 0.007

Interaction Interaction 1* - - - - - - -1.754 0.809 -0.018 Standard deviation of distribution - - - - - - 2.884 0.665 - Interaction 5 1.158 0.447 0.064 1.250 0.553 0.063 1.094 0.431 0.011 Ln (α) - - - -0.770 1.425 - - - - Number of observations 972 972 972Number of significant parameters 7 7 12AIC 408.823 410.667 402.400Log likelihood -196.411 -196.333 -188.074* random parameter


7.7 DISCUSSION

7.7.1 Effects of horizontal alignment along mountainous highways

The proportion of segment length with simple curves is positively associated

with severe crashes; a 1% increase in the proportion of segment length with simple

curve leads to a 0.013 unit increase in the probability of severe crashes, when the

other factors remain constant. This finding may seem odd, as negotiating a simple

curve is much easier than negotiating other complex curves such as broken back,

compound and reverse curves. Higher driving speed on the simple curve compared to

the other curves could be responsible for this outcome.

7.7.2 Effects of longitudinal grades along mountainous highways

The presence of horizontal curves along a steep gradient is also found to have a

positive relation with crash severity. The probability of the severity of crashes

increases by 0.011 unit if it occurs on a highway segment that contains horizontal

curve for more than half of the segment and has longitudinal grades greater than 8%.

Highway segments with this combination are very challenging due to the fact that

drivers have to manage their speed and control their steering at the same time.

Schneider IV, Savolainen, and Zimmerman (2009) found that fatal crashes increased

by 560% when crashes occurred along combination of horizontal and vertical

curvature on curves with a medium radius. Reduced sight distance along these types

of segment is suspected as a major cause behind the crashes.

7.7.3 Effects of cross-sectional elements along mountainous highways

The parameter estimate for the variable proportion of segment length with

unpaved shoulders is normally distributed with mean 0.278 and standard deviation

1.218. This implies that the probability of severe crashes increases with the increase

in the proportion of segment length with unpaved shoulders in 59% of crashes,

whereas the effect is the opposite for the other 41% of crashes. Most of the unsealed

shoulder along mountainous highways is turf and soil. These types of shoulder are

very soft and dangerous for run-off-road vehicles. Research conducted in Australia

found that a sealed shoulder of at least 0.6 – 1.0m was observed to reduce run-off-

road casualty crashes rate by 33 – 64 % when compared with similar roads with

unsealed shoulder only (Jurewicz et al., 2014). However, for some crashes, perhaps


those during dry season, the proportion of segment length with unsealed shoulders

increases the likelihood of severe crashes.

7.7.4 Effects of roadway and roadside features along mountainous highways

Mountainous highways often have dangerous roadside features such as cliffs

and embankments. The proportion of segment length with cliffs along both sides is

found to increase the probability of severe crashes. This is because the segments with

cliffs are dangerous for run-off-road crashes. Vehicles involved in run-off-road

crashes along this section might go down to the bottom of cliffs and thus increase the

probability of severe crashes.

7.7.5 Effects of rainy conditions along mountainous highways

The parameter estimate for rainy conditions was found to be normally

distributed with mean -0.216 and standard deviation 0.994. This implies that the

relationship is negative for 59% of crashes and positive for the remaining 41%. Rain

decreases the probability of a severe crash by 0.002 unit. This finding is consistent

with the finding of Donnell and Mason Jr (2004), who reported that a wet or icy

pavement surface significantly decreases crash severity. On the wet pavement

surface, drivers tend to reduce their speeds and be more careful, thus decreasing

crash severity level when crashes occur during this condition (Quddus, Wang, &

Ison, 2009; Yamamoto & Shankar, 2004). Similarly, Yu and Abdel-Aty (2014a)

found that during snow season, when there is often low visibility and high

precipitation, drivers may lower their speed and become more cautious. Ibrahim and

Hall (1994) found that drivers reduce their speed by an average of 2km/h during light

rain and 5 to 10 km/h during heavy rain. Moreover, Rahman and Lownes (2012)

found that drivers also increase the time gap from the preceding vehicle when the

weather changes from clear to rainy conditions. Hence, wet condition make drivers

more cautious, which helps reduce crash severity. However, the opposite is also

found, where the rainy condition increases the probability of severe crashes in 41%

of crashes. This might be due to the decrease of pavement skid resistance during wet

conditions. Hence, a complex relationship is observed between pavement surface

condition and injury severities.


7.7.6 Relationship between collision type and injury severity

The model suggests that the head-on collision increases the likelihood of

severe crash. In fact, the marginal effect shows that the head-on collision increases

the probability of a severe crash by 0.015 unit. Previous research also supports this

finding. For example, O'Donnell and Connor (1996) found that a head-on collision is

more dangerous than any other type of collision in Australia. Similarly, Abu-Zidan

and Eid (2015) found the injury severity for front impacts was double that of rear

impacts.

In contrast, the rear-end impacts are found to decrease the likelihood of

severe crashes. Marginal effects of this variable suggest that for a rear-end collision,

the probability of a severe crash decreases by 0.012 units. The low severity of rear-

end collisions is also reported in previous research (e.g., Kockelman & Kweon,

2002). Driving on mountainous highways is a challenging task due to steep gradients

and sharp curvatures. On a specific study about rear-end collisions, Chen, Zhang,

Yang, Milton, and Alcántara (2016) observed that drivers become more aware of

potential risks under adverse driving conditions, which reduce the severity of rear-

end collisions.

7.7.7 Effects of driver gender

Interestingly, female drivers are found to be less involved in severe crashes

than male drivers. The parameter estimate for female drivers is randomly distributed

with mean -5.53 and standard deviation 6.05, suggesting that the coefficient for this

variable is negative for 82% of samples and positive for the other 18%. Research has

found that, compared to males, female drivers are more concerned about safety, have

a safer attitude towards in drink driving, and have a better attitude towards rule

violations and speeding (Nordfjærn & Rundmo, 2013). These attitudes play a vital

role in reducing crash severity.

7.7.8 Effects of heavy vehicles along mountainous highways

Another important predictor of crash severity is heavy vehicle involvement,

which is found to be positively associated with severe crashes. The marginal effects

show that heavy vehicle involvement increases the probability of a severe crash by

0.007 unit. The size and weight of the heavy vehicle is likely to increase the impact

of the crash and make it more severe than other type of vehicles. Chen and Chen

(2013) reported that trucks (10001 lbs or over) are more likely to be involved in


severe crashes along mountainous highways compared with non-mountainous

highways. Similarly, Qin, Wang, and Cutler (2013) found that crashes involving

heavy trucks are generally more severe because of their size, weight and speed

differential.

7.7.9 Effects of high interaction along mountainous highways

The tree diagram also presents some insightful findings. It reports that the

heavy vehicle involvement in a crash is a dominant factor which increases the

percentage of severe crash by 3 times compared to crashes where no heavy vehicle is

involved. All six interactions from the tree diagram (branches of the tree) are given

as input to the statistical model and two of them (interaction 1 and 5) are found to be

significant. Interaction variable 1 suggests that the crash severity significantly

decreases when a non-heavy vehicle is involved in a single vehicle crash on a road

segment that contains less than 50% of simple curves. The corresponding marginal

effect shows that for a crash of this combination, the probability of a severe crash

decreases by 0.018 units. This interaction is very complex and further research is

needed to clarify how this interaction in contributes to severe crashes. Heavy vehicle

involvement is found to have a significant positive impact on crash severity, and

interaction variable 5 further emphasizes that heavy vehicle involvement in multi-

vehicle crashes on highway segments without horizontal curves along steep gradients

significantly increases the probability of a severe crashes. While heavy vehicle

involvement alone increases the probability of a severe crash by 0.007, this

combination increases it by 0.011 unit. In related research, Brodie, Lyndal, and Elias

(2009) found that most of the heavy vehicle fatal crashes in Victoria, Australia

occurred on straight road segments. Hence interaction 5 rightly highlighted the fact

that heavy vehicle involvement is more likely to result in severe crashes on less

complex road segments in rural mountainous highways. The underlying reason could

be related to higher speed and/or lack of attention resulting from less complex road

geometry.

7.7.10 Challenges and implications

There were two big challenges in conducting this study: 1) the availability of

information in the crash database and 2) the imbalance response variable and

unobserved heterogeneity in injury severity data. To circumvent this first issue, a

field survey was conducted along selected highways to obtain more information


about road geometry within the crash prone locations. To address the second issue,

three models were developed and the random parameter logit model was found to

perform better than Scobit and standard logit models. An effort was made to collect

as much information as possible about the crash sites. However, some vital crash

related information could not be collected, such as the direction of the vehicles

involved in the crash and speed of the vehicle(s). These types of data could give a

better justification of the model outcomes. Finally, the model outcomes are justified

based on the limited information available and related previous research. Findings

from this study will help practitioners to provide appropriate countermeasures for

reducing the number of severe crashes along rural mountainous highways.

Chapter 8: Discussion and Conclusions 128

Chapter 8: Discussion and Conclusions

8.1 INTRODUCTION

This chapter presents an overall discussion of the different studies of this

thesis. First, the findings from the studies are reviewed with reference to the research

questions. Second, the contributions of this research to scientific knowledge and its

implications are explored. The strengths and limitations of this research are then

discussed, followed by the conclusion. Finally, some recommendations for future

research are presented.

8.2 REVIEW OF FINDINGS

8.2.1 Research Question One

“What are the characteristics of road traffic crashes along rural mountainous

roads?”

8.2.1.1 What are the most common types of crashes occurring along rural mountainous roads?

The finding of this research reveals that ‘out-of-control’ collisions are the most

frequent collision type, and that single-vehicle crashes are the most frequent crash

type among crashes along rural mountainous roads.

8.2.1.2 What are the proportions of vehicle types involved in crashes along rural mountainous roads?

The findings from Study 1 show that the higher proportion of vehicle type

involved in crashes along rural mountainous roads is the passenger car, which is

followed by four wheel drives, heavy vehicles, small lorries, motorcycles, and other

types of vehicles. An analysis from Study 1 shows that the odds of crash involvement

for heavy vehicles, which in this research, include rigid lorries, lorry trailers and

buses, are much higher along mountainous roads in Sabah. The conditions of

mountainous roads with steep slopes and tight curves are particularly a challenging

for heavy vehicles due to their size and manoeuvrability.


8.2.2 Research Question Two

“What are the factors that lead to SV crashes along rural mountainous highways?”

8.2.2.1 What is the effect of weather and traffic conditions on SV crash occurrence? Does weather at the time of incident play a role?

The SPF for SV crashes suggests that both average hourly rainfall at the time

of crash and average visibility at the time of crash are positively associated with SV

crashes along rural mountainous highways. In contrast, findings from previous

studies in the U.S. found that visibility is negatively associated with crash occurrence

(Ahmed et al., 2012; Ma et al., 2015b; Yu & Abdel-Aty, 2013b; Yu et al., 2015).

There are two explanations for these contradictory results; 1) better visibility

encourages higher speed along rural highways; and 2) it is possible that traffic

volume decreases as visibility goes down and vice versa.

8.2.2.2 Does speeding contribute to crashes along upgrade and downgrade sections on rural mountainous highways?

In this research, the speeding indicator refers to a condition when the 85th

percentile speed along downgrade or upgrade segments is greater than the posted

speed limit, and is used to represent speeding behaviour along rural mountainous

highways. The downgrade speeding indicator is found to be positively associated

with SV crashes. This finding is related to the condition of rural mountainous

highways with low traffic volume and less enforcement that can encourage higher

driving speed along these highways.

8.2.2.3 What are the variables of road geometry, cross-sectional elements, roadside features, and spatial characteristics that influence SV crashes along rural mountainous roads?

There are eight variables from these groups which are associated with SV

crashes: maximum radius of curvature, proportion of segment length with

longitudinal grades greater than zero, maximum longitudinal grade >8%, proportion

of segment length with bitumen shoulder, proportion of segment length with one side

shoulder width > 1.5m, proportion of segment length with embankments along one

side, presence of road delineation and the number of houses/shops/commercial

buildings. Among these variables, two variables, the proportion of segment length

with longitudinal grades greater than zero and the number of

houses/shops/commercial buildings, turned out to be random parameters which are.


The presence of random parameters indicates the existence of unobserved

heterogeneities around these parameters in explaining SV crashes. Variables that

were found to increase the likelihood of SV crashes are the proportion of segment

length with grades greater than zero, maximum longitudinal grade >8%, proportion

of segment length with embankments along one side and number of houses/shops/

commercial buildings; The other variables were found to decrease the likelihood of

SV crashes.

8.2.3 Research Question Three

“What are the factors that contribute to the occurrence of MV crashes along

rural mountainous highways?”

8.2.3.1What is the effect of weather and traffic conditions on MV crash occurrence? Does weather at the time of incident play a role?

The SPF for MV crashes suggests that heavy rainfall (if rainfall in 1-hour is

greater than 5.08mm) at the time of crash increases the likelihood of MV crashes

along rural mountainous highways. This finding is in line with previous studies along

freeways in Colorado, U.S. (Yu et al., 2015), where it was found that heavy

precipitation increases the probability of crash occurrence.

8.2.3.2 Does speeding contribute to road crashes along upgrade and downgrade sections on rural mountainous highways?

Both speeding indicators along upgrade and downgrade sections are not found

to be significant in the SPF of MV crashes along rural mountainous highways.

8.2.3.3 What are the variables of road geometry, cross-sectional elements, roadside features, and spatial characteristics that lead to MV crashes along rural mountainous roads?

Among 43 variables that represent road geometry, cross-sectional elements,

roadside-features, and spatial characteristics examined in the MV crash model, only

four variables are found to be statistically significant. While the presence of

horizontal curves along steep gradients and the number of minor intersections along

highway segments are positively associated with MV crashes, the presence of

overtaking lanes and road delineations are found to decrease the likelihood of MV

crashes.


8.2.3.4 Is there any difference between the likelihood of crashes in SV and MV crashes along rural mountainous roads?

Table 8-1 shows the list of variables that have been statistically significant in

the SPFs of SV and MV crashes along rural mountainous highways. The

parsimonious models identified 13 and 6 explanatory variables influencing

respectively SV and MV crashes along rural mountainous highways with plausible

signs and magnitudes. Exposure factors have similar positive effects in both SV and

MV crashes, but the function or effect of exposure is different. While the logarithm

of both ADT and segment length are significant predictors as separate variables in

the SPF for SV crashes, the logarithm of ADT x segment length is identified as a

proper exposure metric for MV crashes. Rainy conditions in mountainous areas

increase the crash likelihood of both types of crashes. It is also quite clear that other

factors influencing the SV and MV crashes along rural mountainous highways are

different. Speeding, radius of horizontal curve, longitudinal grades and roadside

features such as type and width of road shoulder, embankments and number of

houses/shops/commercial buildings are shown to influence SV crashes, while the

presence of horizontal curves along a steep gradient, overtaking lane and minor

intersection influence MV crashes. The presence of road delineation along rural

mountainous highways decrease the likelihood of both SV and MV crashes.

Table 8-1: A comparison of factors associated with SV and MV crashes Variables SV MV

Exposure Variables Log ADT ↑ NA Log of segment length ↑ NA Log (ADT x segment length) NA ↑ Real– time weather information Average visibility at the time of crash (km) ↑ - Average hourly rainfall at time of crash (mm) ↑ NA Heavy rainfall indicator at time of crash (1 if rainfall in 1-hour is greater than 5.08mm, 0 otherwise)

NA ↑

Heavy rainfall indicator during the hour before the crash (1 if 1-hour amount of rainfall during the hour before the crash is greater than 5.08mm, 0 otherwise)

NA -


- -

Downgrade speeding indicator (1 if 85th percentile vehicle operating speed greater than the posted speed limit) ↑ -

Horizontal alignment Proportion of segment with horizontal curve - -Proportion of segment with simple curve - -


Variables SV MV Proportion of segment with reverse curve - - Proportion of segment with compound curve - - Proportion of segment with broken back curve - - Maximum degree of curvature (°) - - Minimum degree of curvature (°) - - Maximum radius of curvature (km) ↓ - Minimum radius of curvature (km) - - Maximum length of circular curve (km) - NA Minimum length of circular curve (km) - NA Maximum length of tangent (km) - NA Minimum length of tangent (km) - NA Longitudinal grades Proportion of segment with longitudinal grades greater than zero ↑ - Number of vertical curves per km - NA Maximum longitudinal grade <2% indicator (1 if maximum longitudinal grade <2%, 0 otherwise)

- -


- -


- -


- -

Maximum longitudinal grade >8% indicator (1 if maximum longitudinal grade >8%, 0 otherwise) ↑ -

Combination of horizontal and vertical alignment indicator; Category 1: (1 if 50% or less of a segment has horizontal curve and absolute gradient ≤ 4%, 0 otherwise)

NA -


NA -


NA -


NA ↑

Cross-sectional elements Proportion of segment with concrete shoulder - - Proportion of segment with bitumen shoulder ↓ - Proportion of segment with gravel and earth shoulder - - Proportion of segment with turf shoulder - - Proportion of segment with one side shoulder width >1.5m ↓ - Proportion of segment with both sides shoulder width >1.5m - - Proportion of segment with both sides shoulder width <1.5m - - Proportion of segment with broken centre line - - Proportion of segment with rumble strip - NA Proportion of segment with marginal strip > 0.5m - NA Proportion of segment with edge drop-offs >100mm - NA Presence of overtaking lane (1 if there is an overtaking lane along the segment, 0 otherwise)

- ↓

Roadway and roadside features Number of minor intersections - ↑ Number of appropriate emergency stop areas - NA Number of trees per km - NA Number of culverts per km - NA Number of electric poles per km - NA Number of roadway lighting poles per km - - Proportion of segment with guardrails along one side - NA Proportion of segment with guardrails along both sides - - Proportion of segment with embankments along one side ↑ NA


Variables SV MV Proportion of segment with embankments along both sides - -Proportion of segment with cliffs along one side - NAProportion of segment with cliffs along both sides - -Presence of bridge (1 if there is a bridge along the segment, 0 otherwise) - -Presence of road delineation (1 if there are road delineations such as guide posts and chevron signs along the segment, 0 otherwise) ↓ ↓

Spatial characteristics Number of houses/shops/commercial buildings within 100m buffer zone from each road edge in the road segment per km ↑ NA

Proportion of segment with forest within 10m of the road edge - -Proportion of segment with farm/ agricultural activity within 10m of the road edge

- -

Proportion of segment with houses/shops/commercial buildings within 10m of the road edge

- -

*NA – not included in the parsimonious model

8.2.4 Research Question Four

“What are the factors related to the crash severity of traffic crashes along

rural mountainous highways?”

8.2.4.1 What driver and vehicle factors are associated with greater or lesser crash severity?

Crashes involving female drivers are found to be less severe than crashes

involving male drivers. Previous research identified that female drivers are more

aware of road safety and have a safer attitude towards drink driving, rules violations,

and speeding (Nordfjærn & Rundmo, 2013). In terms of vehicle type, crashes

involving heavy vehicles are more severe. This is because of the fact that the impact

of the crashes involving heavy vehicles is higher compared to other types of vehicles

because of their size and weight.

8.2.4.2 What is the relationship between weather conditions and crash severity?

Crashes during rainy conditions are found to be negatively associated with

severe crashes, probably due to driving speed during rainy conditions. Previous

studies had identified that most drivers reduce their speed (Ibrahim & Hall, 1994;

Quddus et al., 2009; Yamamoto & Shankar, 2004) and increase the time gap between

preceding vehicles (Rahman & Lownes, 2012) during rainy conditions.

8.2.4.3 What is the relationship between road geometry, cross-sectional elements, roadside features, and crash severity?

Study 4 has identified that among 16 variables in these categories, four of

them are statistically significant in the crash severity model with plausible signs and

magnitudes. These variables include the proportion of segment length with simple


curves, presence of horizontal curves along a steep gradient, proportion of segment

length with unsealed shoulder and proportion of segment length with cliffs along

both sides. These variables are found to be positively associated with severe crashes

along rural mountainous highways.

8.2.4.4 Are the road safety factors along rural mountainous highways different in developing and developed countries?

Findings from this research show that there are similarities and differences of

factors contributing to road safety along rural mountainous highways between

developing and developed countries. As mentioned before, there are differences in

highway characteristics along rural mountainous highways in term of roadway

designs, roadside environment, presence of roadside furniture, traffic mix,

enforcement practices, and driver behaviours. Table 8-2 shows the list of factors that

are found to significantly influence crash and crash severity along mountainous

highways in developing country, compared with developed countries. It is shows that

most of the identified factors have the same affect in developing and developed

countries. However, the effect of visibility on SV crashes is different between these

countries. There are two explanations for these contradictory findings. First, better

visibility along rural mountainous highways in developing countries may encourage

speeding behaviour, while in developed countries, where speed enforcement more

prevalent and effective, better visibility does not encourage speeding as much.

Second, it is possible that traffic volume decreases as visibility increases or vice

versa, which is observed in the context of a developed country with real time traffic

data. However, this visibility and exposure relationship is not very well known for

developing countries as real time traffic data is generally not available. Third, the

effects of minor junctions and the effects of simple curves or cliffs on traffic safety

along rural mountainous highways in developing countries are considerably different

than that of developed countries.

Table 8-2: A comparison of factors influencing road safety between developing and developed countries.

Factors Findings from this research (context of a developing country)

Findings in the context of developed countries

Comparison

Single-vehicle crashes Average daily traffic (ADT)

Positively associated with SV crashes along rural mountainous highways.

Traffic crashes are positively associated with exposure (e.g., Ceder & Livneh, 1982; Chang, 2005);

Consistent




Comparison

however, exposure was measured as average annual daily traffic (AADT) was not significantly associated with SV crashes along mountainous freeways in the United States and not significant (Yu & Abdel-Aty, 2013b)

Visibility Average visibility at the time

of crash is found to be positively associated with SV crashes.

A number of studies in the United States reported that visibility is negatively associated with crashes on mountainous roads (Ahmed et al., 2012; Ma et al., 2015b; Yu & Abdel-Aty, 2013a; Yu et al., 2015)

Inconsistent

Average rainfall at the time of crash

The average hourly rainfall at the time of crash is positively associated with SV crashes along rural mountainous highways.

Average rainfall at the time of crash increases the likelihood of SV crashes (Ma et al., 2015b; Yu & Abdel-Aty, 2013a; Yu et al., 2015).

Consistent

Downgrade speeding The corresponding elasticity

estimate indicates that SV crashes increase by about 45% if the 85th percentile driving speed along a mountainous highways segment is higher than the posted speed limit.

Ma et al. (2015b) found that the speed gap significantly contributes to crash occurrence along mountainous highways.

Partly consistent

Maximum radius of curvature

The maximum radius of curvature is negatively associated with SV crashes along rural mountainous highways.

Crash counts are reported to decrease with the decrease in horizontal curve radius and curve length (Bauer & Harwood, 2013).

Consistent

Proportion of segment with longitudinal grades greater than zero

On average SV crashes are found to increase by about 0.06% for every percent increase in segment length with longitudinal grades greater than zero.

Yu et al. (2015) reported that the presence of a steep downgrade slope increases the crash risk.

Partly consistent

Maximum longitudinal grade>8% indicator

Maximum longitudinal grade higher than 8% is positively associated with SV crashes.

A downgrade segment with slope 6 to 8% is the most hazardous compared to other gradients such as 4 to

Consistent




Comparison

6% and 2 to 4% (Ahmed et al., 2011; Yu et al., 2015).


Proportion of segment length with a bitumen shoulder is negatively associated with SV crashes.

Sealed shoulders are reported to reduce casualty crashes (Jurewicz et al., 2015).

Partly consistent

Proportion of segment with one side shoulder width>1.5mz

Proportion of segment length with wide shoulder (>1.5m) along one side is negatively associated with SV crashes.

The effect of wide shoulders decrease the likelihood of crashes in Connecticut, U.S. (e.g., Ivan et al., 1999)

Consistent

Presence of road delineation

The presence of road delineation like chevron signs and guide posts is a significant predictor, and negatively associated with SV crashes.

Proper curve delineations through chevron signs, curve warning signs, and repeater arrows are well-established treatment options for improving safety along non-mountainous roads (Charlton, 2007; Montella, 2009).

Partly consistent

Multi-vehicle crashes Vehicle kilometres travel

The logarithm of ADT x Segment length is associated with approximately 7.2% increase in the frequency of MV crashes along rural mountainous highways.

Average annual daily traffic (AADT) is found to increase the probability of MV crashes occurrence along 15-mile mountainous highways in Colorado, U.S. (Yu & Abdel-Aty, 2013b).

Consistent

Heavy rainfall indicator at time of crash

Heavy rainfall indicator at time of crash was found to be statistically significant and positively associated with MV crashes.

Heavy rainfall was found to increase the crash risk along mountainous section of the I-70 freeway in Colorado in US (e.g., Yu et al., 2013; Yu et al., 2015).

Consistent

The combination of horizontal and vertical alignment (with the grade higher than 4%)

MV crashes increase by 2% on an average if there is a combination of horizontal and vertical alignments on mountainous highway segments.

Yu et al. (2015) reported that curve segment indicators and steep downgrade indicators (more than 4% in absolute gradient) as separate variables are associated with the higher probability of crashes along the I-70 freeway

Partly consistent




Comparison

in Colorado. Presence of overtaking lane

The presence of overtaking lane along highway segments is found to be a significant predictor and negatively associated with MV crash frequencies.

Overtaking lanes are reported to decrease fatal and injury crashes on two-lane two-way non-mountainous highways (Frost & Morrall, 1995; Schumaker et al., 2016).

Partly consistent

Number of minor junctions

The parameter estimate for number of junctions is found to be positively associated with MV crash frequencies.

No study in developed countries reported this variable in their model.

-

Presence of road delineation

The corresponding elasticity estimate indicates that the presence of road delineation along rural mountainous highways reduce MV crash frequencies by about 24%.

In Wyoming, US, the delineators are reported to reduce all types of injury crashes including fatal, injury and property damage only along rural roads (Ksaibati et al., 2015).

Consistent

Injury Severity Proportion of segment with simple curve

Proportion of segment length with simple curves is positively associated with severe crashes.

There is no study in developed countries using this variable in their research.

-

The combination of horizontal and vertical alignment (with the grade higher than 8%)

Presence of horizontal curves along a steep gradient is also found to have a positive relation with the crash severity.

Schneider IV et al. (2009) found that fatal crashes increased by 560% when crashes occurred along combination of horizontal and vertical curvature on curves with medium radius.

Consistent

Proportion of segment with unsealed shoulder

The probability of severe crashes increases with the increase in the proportion of segment length with unpaved shoulder in 59% of crashes.

Sealed shoulders are reported to reduce casualty crashes (Jurewicz et al., 2015).

Consistent


Proportion of segment length with cliffs along both sides is found to be positively associated with the severity of crashes.

There is no study in developed countries using this variable in their research.

-

Rainy condition Rainy conditions decrease the

probability of a severe crash by 0.002 unit.

On a wet pavement surface, drivers tend to reduce their speeds and be more careful where

Consistent




Comparison

decreasing crash severity level when crashes occur during this condition (Quddus et al., 2009; Yamamoto & Shankar, 2004).

Rear-end collision Rear-end collisions are found

to decrease the likelihood of severe crashes.

Low severity of rear-end collisions are also reported in previous research (e.g., Kockelman and Kweon (2002).

Consistent

Head-on collision Head-on collisions increase

the likelihood of a severe crash.

O'Donnell and Connor (1996) found that head-on collisions are more dangerous than any other type of collisions in Australia.

Consistent

Female driver Female drivers are found to be

less involved in a severe crash than male drivers.

Research has found that, compared to males, female drivers are more concerned about safety, and have a safer attitude towards drink driving, rule violations, and speeding (Nordfjærn & Rundmo, 2013).

Consistent

Heavy vehicle Heavy vehicle involvement is

found to be positively associated with severe crashes.

Chen and Chen (2013) reported that trucks (10001 lbs or over) are more likely to be involved in severe crashes along mountainous highways than on non-mountainous highways.

Consistent

8.3 CONTRIBUTION TO SCIENTIFIC KNOWLEDGE AND

IMPLICATIONS

This research represents a systematic scientific contribution to the study of

road safety along rural mountainous highways, particularly in the context of a

developing country. By collecting an extensive dataset through field surveys and

utilizing existing databases of traffic crashes and weather information, this research

successfully modelled SV and MV crashes, and crash severity as a function of wide


ranging variables including road geometries, traffic characteristics, real-time weather

conditions, cross-sectional elements, roadside features, and spatial characteristics.

8.3.1 Methodological Contribution

The methodological contribution of this research is the application of the

random parameters model to identify unobserved heterogeneity in explaining SV and

MV crashes along rural mountainous highways. There are two types of

heterogeneities, structured and unstructured, based on the sources of they arise. In the

context of this research, structured heterogeneities may result from data clustering

and temporal correlations, as the highway segments were observed multiple times.

Unstructured heterogeneities, however, might arise from misspecification in the

model, uncertainty in exposure or covariates and omitted variables. Modelling results

show that the standard deviations of three parameters in the SV crash model and two

parameters in the MV crash model were found to be significantly different from zero

and thus they were estimated as random parameters. The presence of three and two

random parameters further confirmed the existence of unobserved heterogeneities in

the crash dataset and the appropriateness of the Random Parameters Negative

Binomial (RPNB) model for the development of safety performance functions for SV

and MV crashes.

In addition to the application of the random parameters model, the comparison

of the predictive performance of three state-of-the-art modelling methodologies was

undertaken in the MV crash study to capture crash observations with excess zeros.

This involved the development of Negative Binomial–Lindley (NB-L), Negative

Binomial–Generalized Exponential (NB-GE) and RPNB models. The prediction

performance of these models was compared based on several global goodness-of-fit

measures, including Mean Absolute Deviation (MAD), Mean Squared Predictive

Error (MSPE) and Mean Square Error (MSE). For the first time in literature, it was

found that the RPNB performed better for handling a crash dataset with excess zeros

than other models.

Injury severity of traffic crashes in Malaysia is classified into four categories:

fatal, serious injury, slight injury, and property damage only (PDO). However,

because the numbers of fatal, serious and slight injuries are very small compared to

PDO crashes in this dataset, they were combined into a single severe injury category,

while PDO crashes were classified as non-severe. Despite combining fatal, serious,


and slight injuries into one category, the data remained imbalanced with 7% of total

crashes representing severe crashes and 93% representing non-severe crashes. To

alleviate this problem, this research developed and compared three different models,

including the standard logit model to act as a base model, the skewness (Scobit)

model to account for the imbalance between the response variable, and the Random

Parameters Logit model to account for the unobserved heterogeneity in the dataset.

Based on the goodness-of-fit (AIC and Log Likelihood), the Random Parameters

Logit model performs better than the other models.

Modelling processes can suffer from specification errors (Washington, 2010).

These errors occur when a model represents an incorrect relationship between

dependent and independent variables or includes irrelevant variables and/or

inappropriate forms of main effects, interactions among variables having non-

additive effects, and nonlinearities. In this research, a two-step modelling approach

has been used to address this issue in injury severity modelling. This approach

involves the use of a decision tree to identify the higher order interaction among the

variables, and logistic regression to identify the inference of the parameters. Six

interaction variables were identified in the decision tree process. Out of these, two

interaction variables were found to be statistically significant, with the other ten

prospective variables in random parameters logit regression.

8.3.2 Empirical Contribution

This research provides an in-depth understanding of the causes and

contributing factors related to SV and MV crashes, and the injury severity of crashes

that occurred along rural mountainous highways in Malaysia. This sub-section is

divided into two parts: general discussion of research contribution and a specific

discussion for developing countries.

8.3.2.1 General Contribution

Roadway geometric and cross-sectional elements

There are various roadway geometric and cross-sectional elements influencing

road traffic crashes. Steep gradients and horizontal curves are two typical features of

highways in mountainous areas because of their complex topography. Findings from

this research identified that the proportion of longitudinal grade greater than zero and

steep gradients both influence SV crashes along rural mountainous highways. In

terms of horizontal alignment, the radius of curvature plays a vital role in road safety.


SV crashes were found to decrease when the maximum radius of curvature increases.

The effect of simple curves is also important, as this type of curvature was found to

increase severe crashes. The risk of crash is more serious when there is a

combination of steep gradients and horizontal curves in the road alignment. This

research identified that this combination increases both MV crashes and severe

crashes. Heavy vehicles are particularly affected along steep longitudinal grades. A

higher proportion of heavy vehicles increases queuing, particularly along highway

sections with steep gradients, which may encourage drivers of other vehicle types to

perform risky overtaking manoeuvres. Road authorities should consider putting

additional lanes along steep longitudinal grades to separate slow moving heavy

vehicles from other vehicles in the traffic stream. In addition, speed advice and

warning signs should be posted along steep and/or curved highway segments to

advise drivers of the appropriate speed for the segment. Road authorities should also

consider redesigning existing horizontal curves with small radii, and avoiding such

curves in future construction.

In terms of cross-section elements, the road shoulder was identified as one of

the more important elements for mountainous road safety, as both paved shoulders

(bitumen) and wider shoulders are associated with the reduced likelihood of SV

crashes. A subsequent investigation into their effects identifies that the likelihood of

SV crashes along highway sections with narrow shoulders is significantly less for

paved shoulders than unpaved ones. As constrained geometric conditions often do

not allow an increase in the road reserves to accommodate wide shoulders, paving

the road shoulders or extending the bitumen road to cover narrow shoulders could

improve safety along mountainous highway segments with narrow shoulders. The

subsequent injury severity study also found that unpaved shoulder increases the

probability of severe crashes. Although many previous studies related narrow

shoulders to side swipe and head-on crashes, findings from this research is differ for

MV crashes. Unpaved and narrow shoulders were not found to be statistically

significant in the MV crash model. There are many factors related to this finding,

including that the number of MV crashes along mountainous highways in Sabah is

lower than the number of SV crashes (26% vs. 74%). Out of 257 MV crashes, only

37 (14%) were head-on and side-swipe collisions.


Weather conditions

Real-time weather factors like rainfall and visibility are linked with SV crashes

in mountainous areas. As rainfall is quite common in mountainous areas, road

authorities should prioritize frequent pavement management and maintenance

programs to ensure sufficient pavement quality. Colonna, Berloco, et al. (2016)

found that wet pavement offers less skid resistance. Findings from this research

reveal that the effects of wet pavement may be more prominent for SV crashes along

mountainous highways, as vehicles are harder to control while negotiating curves or

driving along steep slopes.

The MV crash model shows that heavy rain increases the number of MV

crashes along rural mountainous highways. Although driving in rainy conditions is

common for local drivers, which may mean they have improved driving skills in

such an environment, heavy rainfall (larger than 0.508mm) may have negative

impacts on all drivers. This might be due to visibility conditions during heavy

rainfall. There are a few engineering countermeasures that can help drivers during

heavy rainfall. For example, road studs are a low-cost countermeasure that might be

useful for preventing drivers from veering into the opposite lane during heavy rain.

Driver and vehicle factors

Among these factors, driver and vehicle factors play an important role in road

safety. Findings from this study show that crashes involving female drivers are less

likely to be severe compared to crashes involving male drivers. As mentioned before

in Section 8.2.2.2, there is less enforcement from road authorities on rural

mountainous highways than in other areas, however, rules and road regulations

should be followed by road users not only when there is an enforcement. This finding

could indicate that female drivers might be more concerned about road safety and

have better attitudes towards rules and speeding along rural mountainous highways,

where a lower level of enforcement often means that driver attitudes are of greater

importance. Nordfjærn and Rundmo (2013) make a similar argument. Although the

current research does not explore driver behaviour, driver’s road safety awareness,

especially for male drivers, is of concern.

In terms of the type of vehicle, it seems that the involvement of heavy vehicles

in mountainous highway crashes should be of concern, and deserve more attention

from road authorities. This research has found that this type of vehicle is more likely

to crash along mountainous highways than non-mountainous highways. Crashes


involving heavy vehicles also showed an increased likelihood of being severe. The

design and capability of heavy vehicles should also be considered when designing

the highways, particularly with regards to vertical alignment components such as

maximum gradient. In addition, manufacturers of heavy vehicles must also consider

the capability of their vehicles in mountainous conditions.

8.3.2.2 Contributions for Developing Countries

Roadway and roadside features

This research has shown that minor intersections along highway segments

intuitively increase the number of conflicts among vehicles leading to a higher

potential for MV crashes. Minor intersections are common in developing countries to

connect residential roads and major highways. However, most of these intersections

receive less consideration from road authorities in terms of safety. For example, most

of the safety requirements for these intersections, such as sufficient sight distance,

proper road markings and warning signs, were not considered in the design or

construction phases. Road authorities should increase their concern about these

safety requirements in order to reduce MV crashes at minor intersections.

A major portion of highways along in rural mountainous areas in developing

countries, especially tropical countries, passes through forest areas with many trees

along the roadside. Such circumstances create a darker road environment,

particularly at night. This is compounded by a general lack of street lighting along

rural mountainous highways due to budget limitations. Proper road delineation could

be helpful in the reduced visibility condition. Results from this study showed that

providing road delineators such as chevron signs and guide posts reduces the number

of SV and MV crashes along mountainous highways. This low-cost treatment option

should be prioritized to improve mountainous road safety, particularly in resource-

constrained developing countries.

Cross-sectional elements

Mountainous areas have complex topographic conditions that influence the

design of the roadway alignment. This issue is more serious in developing countries

where there are constraints on budgetary and technological resources. Two common

features of rural mountainous highways are steep gradients and horizontal curves.

Heavy vehicles are the most affected vehicles along steep vertical gradients and a

higher proportion of heavy vehicles increase queuing, especially along upgrade


sections. This problem is more serious when other drivers have a limited gap to

overtake due to constraints in sight distance along horizontal curves. Consequently,

drivers may overtake in places where it is not safe, which increases the probability of

MV crashes. Findings from this research show that providing overtaking lanes is

another countermeasure that can reduce the number of MV crashes along rural


This research also identified the importance of paved and wide road shoulders.

Providing paved and wide road shoulders decreases the likelihood of SV crashes and

reduces crash severity. However, many highways in developing countries do not

meet engineering standards, especially for ‘right of way’. This problem is more

serious in mountainous areas where topographic conditions limit the space available

for wide road shoulders. This research found that paving narrow shoulders can

decrease the likelihood of crashes.

Weather Conditions

This research also shows that weather conditions play an important role in road

safety along rural mountainous highways. There are two issues that have been

highlight by previous researchers related to weather conditions: visibility and wet

pavement. This research has identified that good visibility increases the likelihood of

SV crashes. This finding contradicts previous research in developed countries which

found that visibility is negatively associated with crash occurrence. A likely

explanation for this contradictory finding is a difference in driver behaviour, as

drivers along rural mountainous highways are likely to increase their speed in good

visibility conditions, which leads to an increase in SV crashes. Providing an

Automated Enforcement System (AES) may be a good option to enforce speeding

regulation along rural mountainous highways.

Although variables related to pavement conditions are not examined in this

research due to research resources and the availability of that data, the effect of wet

pavement on crash occurrence should be of concern. Results from crash modelling

identified that rainy conditions increase the likelihood of SV and MV crashes. The

performance of pavement surfaces along rural mountainous highways should be

monitored by road authorities regularly to make sure it provides a good skid-

resistance, especially during rainy conditions.

Speeding


Speeding behaviour along rural mountainous highways is another issue that

should be of concern to road authorities. Rural mountainous highways have less

speed enforcement because of various factors, including distance from the nearest

police station and limitations in budget. This creates an opportunity for the drivers to

speed when travelling along these highways. In this research, two types of speed

were observed; upgrade and downgrade speed. From these observations, upgrade and

downgrade speeding indicators (85th percentile of free flow speed higher than the

posted speed limit) were created. Downgrade speeding is found to be associated with

higher a likelihood of SV crashes. There are several effective countermeasures for

reducing speed along downgrade sections, including yellow transverse bars, speed

advisor, and an Automated Enforcement System (AES) to capture speeding

behaviour.

Driver and vehicle factors

In this research, female drivers were found to have a lower the likelihood of

severe crashes. In developing countries, the reduced traffic enforcement along rural

mountainous highways creates an opportunity for drivers to violate traffic rules such

as speed limits. This problem is more serious among male drivers, as previous

research has identified that female drivers are more concerned about road safety, and

less likely to violate traffic rules. Road safety campaigns should particularly target

male drivers to increase their awareness about road safety and safe driving

behaviour.

Involvement of heavy vehicles is found to be increase the likelihood of severe

crashes along rural mountainous highways. Although road design standards usually

consider heavy vehicles in their design, not all the highway segments studied were

constructed based on this standard. This is due to budget and technology limitations

in developing countries, especially along rural mountainous highways with complex

topographical conditions. This is a serious problem because most supplies for rural

areas are transported by heavy vehicle. Considering this, road authorities should

implement countermeasures for reducing the probability of severe crashes involving

heavy vehicles by providing proper rest areas that consider the needs of heavy

vehicles, and arrestor beds on downgrade sections for stopping out of control heavy

vehicles.


8.3.2.3 Implications for Road Safety in Malaysia

The Malaysian government is continuously working with various road

authorities and agencies to improve road safety. Findings from this research might be

useful to road authorities in Malaysia to help achieve their mission to reduce traffic

crashes, particularly in terms of rural mountainous highways.

Although some research on road safety has been conducted in Peninsular

Malaysia, relatively little or no research has examined road safety issues in Sabah

and Sarawak, despite their mountainous topography and different traffic mix. For

example, motorcycles are the most common vehicles in Peninsular Malaysia,

according to registration data for 2010, while for Sabah and Sarawak it is motor car

(passenger car) (MOT, 2010). In addition, topographic conditions in both states, but

especially in Sabah, are slightly different, as more than 60% of the topography in

Sabah is mountainous. These differences might have different effects on to road

safety. This means that successful road safety countermeasures in Peninsular

Malaysia might not be effective in Sabah and Sarawak. The same may be true for

countermeasures on non-mountainous highways versus mountainous highways.

Study 1 of this research identified that crash characteristics in mountainous and non-

mountainous highways are quite different.

The implications of this research can be divided into three categories:

engineering, driver behaviour, and vehicle factors. Table 8.3 presents some possible

engineering treatments derived from the findings of this research to improve road

safety along rural mountainous highways.

Table 8-3: Recommendations for engineering treatments to improve road safety along rural mountainous highways in Malaysia

Factor Safety involvement Suggestion countermeasures

Weather conditions Rainfall SV, MV and injury severity Increase illumination along

highways with street lighting, road delineation, road marking and chevron signs.

Traffic characteristics Downgrade speeding SV Increase speed enforcement and

provide speed reduction countermeasures such as speed advisors signs and transverse yellow bars.

Horizontal alignment Simple curve Injury severity Post speed advisory signs along


Factor Safety involvement Suggestion countermeasures simple curves to advise drivers of the appropriate speed along each section.

Radius of curvature SV Redesign horizontal curves with small radii to increase the radius. Avoid designing horizontal curves with small radii in future construction.

Longitudinal grades Longitudinal grade greater than zero

SV Post speed advisors along these sections. Post traffic warning signs such as ‘Please use low gear’ and ‘reduce your speed’.

Steep gradient SV Post speed advisors along these sections. Post traffic warning signs such as ‘Please use low gear’ and ‘reduce your speed’.

Combination of horizontal and vertical alignment

MV and injury severity Create a distance or gap between two opposite lanes, by replacing the centre line with a wide rumble strip or road median.

Cross-sectional elements Bitumen /sealed shoulder SV and injury severity Increase length of highway

segments with bitumen/ sealed shoulder.

Shoulder width > 1.5m SV Increase proportion of highway segments with wide road shoulders.

Overtaking lane MV Provide overtaking lanes along upgrade sections to decrease MV crashes.

Minor intersections MV Increase sight distance by removing landscape, provide good road markings and post intersection signs along major highways.

Roadway and roadside features

Embankments SV Increase the sight distance along segments with embankments and post speed advisory signs to help drivers in choose appropriate speed along these sections.

Cliffs Injury severity Provide an appropriate type of guardrails to make sure vehicles involved in crashes do not go down cliffs.

Road delineation SV and MV Road authorities should install road delineation along rural mountainous highways, as a good, low cost countermeasure for reducing SV and MV crashes.

Spatial characteristics Number of houses/ shops/ commercial buildings

SV Decrease roadside hazards such as vehicle parking along highways in these areas.


Although this research was not focussed on driver behaviour, some of the

findings suggest enforcement and/or driver education to improve driver behaviour,

leading to better safety outcomes. It is found that speeding is one of the problems

along rural mountainous highways in Malaysia. Based on the crash report data, most

SV crashes are related to speeding behaviour. There are two factors that might

increase speeding behaviour, low traffic volume and less enforcement along these

highways. Road authorities like the Royal Malaysian Police (RMP) should increase

their enforcement along the identified sections where speeding behaviour might

contribute to crashes. An automated Enforcement System (AES) should also be

placed along sections where speeding is common. This research has clarified that

most of the speeding behaviour occurs along downgrade sections. Another issue

related to driver behaviour is different level of injury severity for crashes involving

female and male drivers. Previous research identified that female drivers are more

aware road safety compared to male drivers. This awareness is important in the case

of rural highways because there is less traffic enforcement to force drivers to follow

traffic rules such as speed limits. Male drivers should be targeted in the road safety

campaigns that are conducted regularly by the Road Safety Department of Malaysia

nationwide.

Two types of vehicles are found to have a higher crash risk along mountainous

highways; heavy vehicles and four-wheel drives. Heavy vehicles were found to be

more often involved in crashes along mountainous highways than non-mountainous

highways. The injury severity study also found that heavy vehicle involvement

increases the probability of severe crashes. Traffic volume data shows that heavy

vehicle traffic is the second largest type of vehicular traffic on mountainous

highways in Sabah, representing approximately 25% of total traffic (HPU, 2013).

Heavy vehicles have reduced capability to follow the topographical conditions along

mountainous areas with steep gradients and horizontal curves. Proper rest areas or

arrestor beds are two countermeasures for reducing crashes involving heavy vehicles

along mountainous highways. The regulations in Malaysia require all commercial

vehicles (including heavy vehicles) to undergo routine inspections at Malaysia’s

Vehicle Inspection Specialist (PUSPAKOM) every six months (PUSPAKOM, 2017),

but owners should not rely only on this inspection to ensure their vehicle is safe to

operate along mountainous highways. Four-wheel drives are another type of vehicle

found to have increased risk along rural mountainous highways. There are several


factors that increase the likelihood of 4WD involvement in crashes, including a

higher centre of gravity compared to other types of vehicle. However, most newer

4WDs are equipped with Electronic Stability Control (ESC) which would increase

control along horizontal curves. In addition, 4WDs are different from passenger cars

in the size, structures of passenger seats and head support structures (Broyles,

Narine, Clarke, & Baker, 2003), mass, and profile (Broyles, Clarke, Narine, & Baker,

2001). They also seem to be very stable, and strong, and are generally higher

performance vehicles in terms of engine capacity than passenger cars (Bener,

Razzak, Crundall, & Allen, 2014).

These suggestions and proposed countermeasures can be successful if all the

parties involved work together. Road safety is not the responsibility of government

agencies, but also requires participant from road users as well. This research was

fully supported by the Road Safety Department of Malaysia (JKJR), the Public

Works Department Sabah (PWD Sabah) and the Royal Malaysian Police. The first

findings of this research were presented at the Sabah Road Day that was conducted

by the PWD Sabah in April 2016. The Road Safety Department of Malaysia, Sabah

Branch, is also in the progress of organizing a road safety audit of highways in

Sabah. Findings from this research might be useful for this audit process.

8.4 STRENGTHS AND LIMITATIONS

8.4.1 Strengths of the research

There are a number of strengths associated with this research. First, an

intensive effort has been made to collect relevant data through field surveys. These

surveys involved data collection for road geometry, cross-sectional elements,

roadway and roadside features and spatial characteristics along 102 highway

segments (approximately 89.9km). In addition, two-hour vehicle counts were

conducted at each segment. This value was then converted to ADT using a seasonal

Hourly Expansion Factor (HEF) calculated from the Road Traffic Volume Malaysia

(RTVM) obtained from the Highway Planning Unit, Ministry of Works Malaysia.

Another strength of this research is the application of real-time weather

information in the crash count model. Real-time weather variables have been used

widely in road safety research to identify the effect of real weather conditions on

crash occurrence and injury severity (e.g., Ahmed et al., 2014; Ahmed et al., 2012;


Chen, Chen, et al., 2016; Pande & Abdel-Aty, 2006; Yu & Abdel-Aty, 2014a, 2014b;

Yu et al., 2013; Yu, Abdel-Aty, Ahmed, & Wang, 2014). These variables are

important to capture the time-varying nature of environmental factors which cannot

be captured using monthly average rainfall. Hourly rainfall and visibility conditions

from twelve rainfall stations and two weather stations were used to estimate these

variables for SV and MV crashes and crash severity modelling along rural


The third strength is the application of operating speed instead of the posted

speed limit in crash count modelling. The operating speed represents the real speed

selection along upgrade and downgrade sections in each of the highway segments. A

speeding indicator has been used as a variable for this purpose. For example, the

downgrade speeding indicator refers to a condition when the 85th percentile speed

along a downgrade segment is greater than the posted speed limit. A two-hour spot

speed study was conducted along each of the highway segments for this purpose.

A total of 56 and 51 variables were examined in SV and MV crash models,

respectively, which represents another strength of this study. This research is the first

that has included a range of variables including traffic characteristics, weather

conditions, road geometry, cross-sectional elements, roadside features, and spatial

characteristics in a single model.

A fifth strength is that this research also develops count models with random

parameters that can observe heterogeneity. A comparison between models has also

been made to identify the performance of NB-L, NB-GE and RPNB for crash data

with excess zeros.

Another strength is that this research has examined 32 variables in the crash

severity models to further understand the risk factors related to crash severity in road

traffic crashes along rural mountainous highways. Moreover, it has identified the

major risk factors in relation to collision types, driver and vehicle factors, weather

conditions, road geometry, cross-sectional elements, roadway and roadside features.

A seventh strength is that this research has also used a two-step modelling

approach in the crash severity study. This approach is a combination of decision tree

and logistic regression. The results from the decision-tree are used as a priori

knowledge for the logistic regression model. Two interaction variables capturing the


complex relationship between heavy vehicles and road alignments resulting from the

decision-tree were found to be significant in the random parameters logit model.

Finally, this research has evaluated the performance of three different models

for crash severity data with imbalanced response variables. The result indicates that

the Random Parameters Logit model performs better than standard logit and

skewness (Scobit) models for this type of crash severity data.

8.4.2 Limitations of the research

This study also has some limitations. One of the limitations of this study is that

there is lack of information about some factors such as quality of road surface and

direction of the crashes. Table 3-4 shows the list of the variables and information in

the crash report form (POL27), but only 35 out of 63 variables were available for

crashes at the selected study sites. To overcome this limitation, an extensive field

survey was carried out; however, a richer crash dataset could provide more insights.

Traffic volume data for the selected segments were collected with brief traffic

counts. Although appropriate hourly expansion factors and seasonal variation factors

were used to convert these traffic counts to Annual Daily Traffic, the use of Average

Annual Daily Traffic (AADT) collected through year-round counts with loop

detectors may be more appropriate to capture the relationship between exposure and

SV crashes. Moreover, manual traffic counts may be subject to error, and were used

only because they were the best alternative available for this research, given resource

and time constraints.

Visibility information was obtained from only two weather stations because

there are only two weather stations in Sabah which record hourly visibility

information. Different topographic conditions in mountainous areas give a variety of

fog conditions that influence visibility conditions along each highway segment. The

variety in conditions might not be captured using only these two weather stations.

Although spiral transition curves have been reported to have safety benefits on

mountainous roads (Council, 1998), it was not possible to capture this variable

because of the lack of available of road geometry design data from the road

authority, and limitations on the reconstruction of the geometry of highway segments

using GPS coordinates. While GPS coordinates were collected with a handheld GPS

(Garmin eTrex10) at 5m intervals, the accuracy of this device is +/- 3m (Garmin,


2011). This degree of uncertainty did not allow the identification of spiral transition

curves. Ai and Tsai (2014) also acknowledge mentioned this shortcoming of using

GPS data to capture spiral transition curves.

Despite the fact that the pavement condition of mountainous roads may have a

significant relationship with safety, the dataset does not include variables related to

road surface conditions (e.g. skid resistance, rutting, and international roughness

index). The pavement condition and surface run-off after rainfall are two important

factors for traffic safety along rural mountainous highways.

In summary, the researcher believes this research is unique due to intensive

data collection along rural mountainous areas. Further evidence of unique research

can be found in the application of advanced methodologies, the identification factors

that influence SV and MV crashes and crash severity in the context of rural

mountainous highways in developing countries. However, there are also some

limitations due to the short time frame collecting traffic volume, the availability

information in crash data and other data related to road geometry and pavement

conditions. Overall, the strengths of this research outweigh the limitations.

8.5 CONCLUSIONS

Crash characteristics along rural mountainous highways are different from non-

mountainous highways. For example, out-of-control collisions and the crashes that

occur due to speeding are more frequent along rural mountainous highways than

non-mountainous highways. Other factors that increase the odds of crashes along

mountainous highways compared to non-mountainous highways is the horizontal

curved sections compared with straight sections, single-vehicle crashes compared

with multi-vehicle crashes and weekend crashes compared with weekday crashes.

Despite the fact that mountainous highways have fewer crashes, crashes along these

highways are slightly more severe, as the fatality index (ratio of fatalities to road

injuries) was higher than on non-mountainous highways.

From the wide range of explanatory variables which were examined in the

SV and MV crash models, the parsimonious models identified 13 significant

variables in the SV crash model and 6 significant variables in the MV crash model.

Variables positively associated with SV crashes included traffic flow, segment

length, visibility during the crash, average hourly rainfall at the time of crash,


downgrade speeding indicator, proportion of segment length with longitudinal

grades greater than zero, the presence of a steep grade (>8%), the proportion of

segment length with embankment along one side and number of

houses/shops/commercial buildings. On the other hand, variables negatively

associated with SV crashes included maximum radius of curve, proportion of

segment length with bitumen shoulder, proportion of segment length with wide

shoulder (≥1.5m) and the presence of road delineation. Among these variables,

downgrade speeding indicator, proportion of segment length with longitudinal

grades greater than zero and the number of houses/shops/commercial buildings were

identified as random parameters capturing unobserved heterogeneities in explaining

SV crashes around these factors. For MV crashes, four out of six significant variables

were positively associated with MV, crashes including exposure (ADT x segment

length), heavy rainfall at time of crash, presence of horizontal curves along a steep

gradient and number of minor intersections. Two other variables, the presence of

overtaking lane and the presence of road delineation, were negatively associated

with MV crashes. Presence of horizontal curves along a steep gradient and presence

of road delineation were identified as random parameters in the MV crash model.

The findings of both studies (for SV and MV crashes) shed considerable light on the

factors affecting SV and MV crashes along rural mountainous highways in Malaysia.

These findings would be helpful for road engineers, road safety professionals, and

relevant authorities to design appropriate countermeasures.

Crash severity for rural mountainous highways was modelled to investigate the

risk factors. Several factors were found to increase the likelihood of crash severity

along these highways. The findings show evidenced that the likelihood of severe

crashes decreases in rear-end crashes, yet increase in head-on collisions. Female

drivers decrease the likelihood of severe crashes; however, crashes involving heavy

vehicles increase the probability of severe crashes. In terms of highway segment

characteristics, proportion of segment length with simple curves, presence of

horizontal curves along a steep gradient, proportion of segment length with unsealed

(unpaved) shoulder and proportion of segment length with cliffs along both sides of

highway are associated with an increase in severe crashes. This research also

identified a higher interaction between variables that influence occurrence of severe

crashes. The severity of crashes decreases when non-heavy vehicles are involved in

single-vehicle crashes along highway segments with a higher proportion of simple


curves. It was also found that the probability of severe crashes increases with single-

vehicle crashes involving heavy vehicles along highway segments with less

combination of horizontal curve and steep gradient.

This study also used an operating speed instead of a posted speed limit in SV

and MV crash modelling to explain the real speed at different locations. Two

speeding indicators (upgrade and downgrade) were identified and refer to a condition

when the 85th percentile operating speed is greater than the posted speed limit.

Downgrade speeding in the SV crash model is found to increase the likelihood of

severe crashes. These variables are not significant in MV crash model, however.

These results provide valuable information for the road authorities for managing

speeding behaviour along rural mountainous highways.

The developed safety performance function along rural mountainous highways

is one way to further explore the significant roadway geometry, weather conditions,

and traffic characteristics related to variables that are connected to road crashes along

rural mountainous highways. This is the first attempt that has been made to develop

SPF especially for rural areas in developing countries. As suggested by previous

researchers (Brimley et al., 2012; Young & Park, 2013), SPFs that are developed

based on conditions in a particular area are more reliable. The developed safety

performance function will also help to identify blackspots or high-risk sites in

mountainous areas.

The application of a random parameters model is widely used in the crash

count and crash severity models. The random parameters modelling approach is an

alternative to allow regression parameters to vary across observations (e.g., different

location and/or different time periods). The results from three different models in this

research, including the SV crash model, MV crash model, and crash severity model,

show that random parameters exist in these models. This indicates the existence of

unobserved heterogeneity around these parameters in explaining SV and MV crashes

and crash severity along rural mountainous highways. The performance of the

random parameters model is also superior to other models for accounting for excess

zeros in developing SPFs. In addition, the random parameter specification in the

injury severity model is found to better account for imbalanced response variables in

the discrete outcome dataset. This further indicates the appropriateness of the random

parameters models in the context of road safety research.


8.6 RECOMMENDATIONS FOR FUTURE RESEARCH

A major challenge in conducting road safety research in the context of

developing countries is the availability of reliable and accurate data. The most

important data in this research is related to road segment characteristics, including

road geometry, cross-sectional elements, roadside features, and spatial

characteristics. In this study, an extensive effort was made to collect relevant data

through field surveys. All of this data were collected by research assistants and were

monitored completely by the first researcher. For this purpose, GPS coordinates were

collected with a handheld GPS (Garmin eTrex10) at 5m intervals, but the accuracy of

this device is +/- 3m. This degree of uncertainty did not allow the identification of

spiral transition curves, which are reported to have safety benefits on mountainous

highways. Thus, it is advisable to conduct future field surveys using more accurate

equipment, such as the application of LIDAR etc. for capturing highway segment

characteristics in greater details.

The visibility data was obtained from the National Centers for Environmental

Information (NOAA) and it is only available in two weather stations, Kota Kinabalu

International Airport and Sandakan Airport. Using proximity measures in AutoCAD

- Geolocation, hourly visibility information for crashes on each road segment was

obtained following the aggregation procedure developed by Yu et al. (2015). Most of

the mountainous areas have localized fog due to the different topography, which is

not reflected in the NOAA data. Future research should consider local visibility

conditions that might affect drivers’ sight distance. It would also be better if the crash

rate per km is used in the analysis.

As discussed in Section 8.4.2, the pavement condition (e.g., skid resistance,

rutting, international roughness index, etc.), and surface run-off after rainfall are two

important factors for traffic safety along rural mountainous highways. However, due

to the unavailability of this data, the effects of these variables on crash occurrence

could not be investigated. Future research should include these factors in the crash

modelling along rural mountainous highways.

This research found that a few of the variables, such as downgrade speeding

indicator, proportion of segment length with longitudinal grades greater than zeros,

number of houses/shops/commercial buildings, presence of horizontal curves along a

steep gradient and presence of road delineation, are normally distributed in SV and


MV crashes. This implies that these variables have different effects at different

highway segments. Further research is needed to identify the complex relationship

between these variables for both SV and MV crashes along rural mountainous

highways. In addition, the developed SPFs could be tested to predict SV and MV

crashes with different levels of injury severity.

Study 3 in this research compared the performance of three different models in

handling excess zero problems. The result shows that a Random Parameters Negative

Binomial (RPNB) model performs better than a Negative Binomial – Lindley (NB-L)

and Negative Binomial – Generalized Exponential (NB-GE) model. Besides NB-L

and NB-GE, and there are many other models or distributions that have been used to

handle the excess zeros, such as Sichel (Zou et al., 2013), Poisson-weighted

Exponential (Zamani & Ismail, 2010b; Zamani et al., 2014), Poisson Inverse

Gaussian (Zha et al., 2016) and many more. Thus, comparing the performance of

RPNB with these models and with other datasets would be a worthwhile research

direction.

Bibliography 157

Bibliography

AASHTO. (2010). Highway Safety Manual (1st ed.). Washington DC: American Association of Satate Highway and Transportation Officials.

AASHTO. (2011). A policy on geometric design of highways and streets ED 6th ed. Washington, DC, USA: American Association of State Highway and Transportation Officials (AASHTO).

Abbas, S. K. S., Adnan, M. A., & Endut, I. R. (2011). Exploration of 85th Percentile Operating Speed Model on Horizontal Curve: A Case Study for Two-Lane Rural Highways. Procedia - Social and Behavioral Sciences, 16(0), 352-363. doi:http://dx.doi.org/10.1016/j.sbspro.2011.04.456

Abdel-Aty, M. (2003). Analysis of driver injury severity levels at multiple locations using ordered probit models. Journal of safety research, 34(5), 597-603.

Abdel-Aty, M. A., & Radwan, A. E. (2000). Modeling traffic accident occurrence and involvement. Accident Analysis & Prevention, 32(5), 633-642.

Abu-Zidan, F. M., & Eid, H. O. (2015). Factors affecting injury severity of vehicle occupants following road traffic collisions. Injury, 46(1), 136-141. doi:http://dx.doi.org/10.1016/j.injury.2014.10.066

Ahmed, M., Huang, H., Abdel-Aty, M., & Guevara, B. (2011). Exploring a Bayesian hierarchical approach for developing safety performance functions for a mountainous freeway. Accident Analysis & Prevention, 43(4), 1581-1589.

Ahmed, M. M., Abdel-Aty, M., Lee, J., & Yu, R. (2014). Real-time assessment of fog-related crashes using airport weather data: A feasibility analysis. Accident Analysis & Prevention, 72, 309-317. doi:http://dx.doi.org/10.1016/j.aap.2014.07.004

Ahmed, M. M., Abdel-Aty, M., & Yu, R. (2012). Assessment of interaction of crash occurrence, mountainous freeway geometry, real-time weather, and traffic data. Transportation Research Record: Journal of the Transportation Research Board, 2280(-1), 51-59. doi:10.3141/2280-06

Ai, C., & Tsai, Y. (2014). Automatic horizontal curve identification and measurement method using GPS data. Journal of Transportation Engineering, 141(2), 04014078.

Anastasopoulos, P. C., & Mannering, F. L. (2009). A note on modeling vehicle accident frequencies with random-parameters count models. Accident Analysis & Prevention, 41(1), 153-159. doi:http://dx.doi.org/10.1016/j.aap.2008.10.005

Aryuyuen, S., & Bodhisuwan, W. (2013). The negative binomial-generalized exponential (nb-ge) distribution. Applied Mathematical Sciences, 7(22), 1093-1105.

Bauer, K., & Harwood, D. (2013). Safety effects of horizontal curve and grade combinations on rural two-lane highways. Transportation Research Record: Journal of the Transportation Research Board(2398), 37-49.

Behnood, A., & Mannering, F. L. (2016). An empirical assessment of the effects of economic recessions on pedestrian-injury crashes using mixed and latent-class models. Analytic Methods in Accident Research, 12, 1-17. doi:http://dx.doi.org/10.1016/j.amar.2016.07.002

Bibliography 158

Bener, A., Razzak, J. A., Crundall, D., & Allen, K. A. (2014). The relationship between four-wheel drives and risky driving behaviours. International journal of medicine and public health, 4(3).

Berhanu, G. (2004). Models relating traffic safety with road environment and traffic flows on arterial roads in Addis Ababa. Accident Analysis & Prevention, 36(5), 697-704. doi:http://dx.doi.org/10.1016/j.aap.2003.05.002

Bester, C., & Makunje, J. (1998). The effect of rural road geometry on safety in Southern Africa. Transportation research circular(E-C003), 15-11.

Bhat, C. R. (2003). Simulation estimation of mixed discrete choice models using randomized and scrambled Halton sequences. Transportation Research Part B: Methodological, 37(9), 837-855. doi:http://dx.doi.org/10.1016/S0191-2615(02)00090-5

Brimley, B. K., Saito, M., & Schultz, G. G. (2012). Calibration of Highway Safety Manual Safety Performance Function. Transportation Research Record: Journal of the Transportation Research Board, 2279(1), 82-89. doi:10.3141/2279-10

Brodie, L., Lyndal, B., & Elias, I. J. (2009). Heavy vehicle driver fatalities: Learning's from fatal road crash investigations in Victoria. Accident Analysis & Prevention, 41(3), 557-564. doi:http://dx.doi.org/10.1016/j.aap.2009.02.005

Broyles, R. W., Clarke, S. R., Narine, L., & Baker, D. R. (2001). Factors contributing to the amount of vehicular damage resulting from collisions between four-wheel drive vehicles and passenger cars. Accident Analysis & Prevention, 33(5), 673-678.

Broyles, R. W., Narine, L., Clarke, S. R., & Baker, D. R. (2003). Factors associated with the likelihood of injury resulting from collisions between four-wheel drive vehicles and passenger cars. Accident Analysis & Prevention, 35(5), 677-681.

Carson, J., & Mannering, F. (2001). The effect of ice warning signs on ice-accident frequencies and severities. Accident Analysis & Prevention, 33(1), 99-109.

Castro, M., Sánchez, J. F., & Sánchez, J. A. (2012). Operating speed models for two-lane rural highways. Paper presented at the Proceedings of the Institution of Civil Engineers-Transport.

Ceder, A., & Livneh, M. (1982). Relationships between road accidents and hourly traffic flow—I: analyses and interpretation. Accident Analysis & Prevention, 14(1), 19-34.

Chang, L.-Y. (2005). Analysis of freeway accident frequencies: negative binomial regression versus artificial neural network. Safety science, 43(8), 541-557.

Charlton, S. G. (2007). The role of attention in horizontal curves: A comparison of advance warning, delineation, and road marking treatments. Accident Analysis & Prevention, 39(5), 873-885. doi:http://dx.doi.org/10.1016/j.aap.2006.12.007

Chatzikomis, C. I., & Spentzas, K. N. (2014). Comparison of a vehicle equipped with Electronic Stability Control (ESC) to a vehicle with Four Wheel Steering (4WS). Forschung im Ingenieurwesen, 78(1-2), 13-25. doi:10.1007/s10010-014-0172-z

Chen, C., Zhang, G., Yang, J., Milton, J. C., & Alcántara, A. D. (2016). An explanatory analysis of driver injury severity in rear-end crashes using a decision table/Naïve Bayes (DTNB) hybrid classifier. Accident Analysis & Prevention, 90, 95-107. doi:http://dx.doi.org/10.1016/j.aap.2016.02.002

Bibliography 159

Chen, F., & Chen, S. (2013). Differences in injury severity of accidents on mountainous highways and non-mountainous highways. Procedia-Social and Behavioral Sciences, 96, 1868-1879.

Chen, F., Chen, S., & Ma, X. (2016). Crash frequency modeling using real-time environmental and traffic data and unbalanced panel data models. International journal of environmental research and public health, 13(6), 609.

Chen, S., Chen, F., & Wu, J. (2011). Multi-scale traffic safety and operational performance study of large trucks on mountainous interstate highway. Accident Analysis & Prevention, 43(1), 429-438.

Chen, T., Wei, L., & Zhou, W.-X. (2011). Study on the relationship between the horizontal alignment indices and traffic safety in mountainous freeway. Paper presented at the 11th International Conference of Chinese Transportation Professionals (ICCTP).

Chen, T., Zhou, W.-x., & Wei, L. (2011). Study on the Relationship between the Horizontal Alignment Indices and Traffic Safety in Mountainous Freeway ICCTP 2011 (pp. 2319-2325).

Chen, Y. (2014). Traffic crash modeling and driver behavior analysis on mountainous highways in China. Paper presented at the CICTP 2014@ sSafe, Smart, and Sustainable Multimodal Transportation Systems.

Chen, Y., Li, Y., King, M., Shi, Q., Wang, C., & Li, P. (2016). Identification methods of key contributing factors in crashes with high numbers of fatalities and injuries in China. Traffic injury prevention, 17(8), 878-883.

Choi, J., Kim, S., Heo, T.-Y., & Lee, J. (2011). Safety effects of highway terrain types in vehicle crash model of major rural roads. KSCE Journal of Civil Engineering, 15(2), 405-412.

Christoforou, Z., Cohen, S., & Karlaftis, M. G. (2010). Vehicle occupant injury severity on highways: An empirical investigation. Accident Analysis & Prevention, 42(6), 1606-1620. doi:http://dx.doi.org/10.1016/j.aap.2010.03.019

Colonna, P., Berloco, N., Intini, P., Perruccio, A., & Ranieri, V. (2016). Evaluating skidding risk of a road layout for all types of vehicles. Transportation Research Record: Journal of the Transportation Research Board, 2591, 94-102. doi:10.3141/2591-11

Colonna, P., Intini, P., Berloco, N., & Ranieri, V. (2016). The influence of memory on driving behavior: how route familiarity is related to speed choice. An on-road study. Safety science, 82, 456-468.

Council, F. (1998). Safety benefits of spiral transitions on horizontal curves on two-lane rural roads. Transportation Research Record: Journal of the Transportation Research Board(1635), 10-17.

Dell’Acqua, G., & Russo, F. (2011). Safety performance functions for low-volume roads. The Baltic Journal of Road and Bridge Engineering, 6(4), 225-234.

Department, P. W. (1986). Technical Instruction (Road) 8/86 A guide on geometric design of roads. Kuala Lumpur.

DID. (2016). Rainfall Stations Data Inventory. Retrieved from http://www.did.sabah.gov.my/content.php?q=StesenCurahanHujan

Dong, C., Shi, J., Huang, B., Chen, X., & Ma, Z. (2016). A study of factors affecting intersection crash frequencies using random-parameter multivariate zero-inflated models. International journal of injury control and safety promotion, 1-14.

Bibliography 160

Donnell, E., & Mason Jr, J. (2004). Predicting the severity of median-related crashes in Pennsylvania by using logistic regression. Transportation Research Record: Journal of the Transportation Research Board(1897), 55-63.

DOSM. (2010). Basic Population Characteristics by Administrative Districts. Retrieved from Kuala Lumpur: http://www.statistics.gov.my/portal/download_Population/files/BPD/ad_2010.pdf

DOSM. (2012). State / District Data Bank. Retrieved from Kuala Lumpur: http://www.statistics.gov.my/portal/index.php?option=com_content&view=article&id=1871&lang=en

Eck, R. W. (1983). Technique for identifying problem downgrades. Journal of Transportation Engineering, 109(4), 604-610.

FHWA. (2014). Flexibility in Highway Design. Cross Section Elements. Retrieved from http://www.fhwa.dot.gov/environment/publications/flexibility/ch06.cfm

FHWA. (n.d.). Introduction to Safety Performance functions. Retrieved from http://safety.fhwa.dot.gov/tools/crf/resources/cmfs/docs/safety_performance_funtions.pdf

Fitzpatrick, K., Miaou, S.-P., Brewer, M., Carlson, P., & Wooldridge, M. D. (2003). Exploration of the relationships between operating speed and roadway features. Paper presented at the Proc. of 82nd annual meeting of transportation research board. Washington (DC).

Frost, U., & Morrall, J. (1995). A comparison and evaluation of the geometric design practices with passing lanes, wide-paved shoulders and extra-wide two-lane highways in canada and germany. Paper presented at the International Symposium on Highway Geometric Design Practices.

Fu, R., Guo, Y., Yuan, W., Feng, H., & Ma, Y. (2011). The correlation between gradients of descending roads and accident rates. Safety Science, 49(3), 416-423. doi:10.1016/j.ssci.2010.10.006

Garber, N. J., & Hoel, L. A. (2009). Traffic and highway engineering (Fourth Edition ed.). USA: Cengage Learning.

Garmin. (2011). eTrex Owner's Manual for use with models 10, 20, 30. Retrieved from http://static.garmin.com/pumac/eTrex_10-20-30_OM_EN.pdf

Garrido, R., Bastos, A., de Almeida, A., & Elvas, J. P. (2014). Prediction of road accident severity using the ordered probit model. Transportation Research Procedia, 3, 214-223.

Geedipally, S. R., & Lord, D. (2010). Investigating the effect of modeling single-vehicle and multi-vehicle crashes separately on confidence intervals of Poisson–gamma models. Accident Analysis & Prevention, 42(4), 1273-1282.

Geedipally, S. R., Lord, D., & Dhavala, S. S. (2012). The negative binomial-Lindley generalized linear model: Characteristics and application using crash data. Accident Analysis and Prevention, 45, 258-265. doi:10.1016/j.aap.2011.07.012

Geurts, K., & Wets, G. (2003). Black spot analysis methods: literature review. Gibreel, G. M., Easa, S. M., & El-Dimeery, I. A. (2001). Prediction of operating

speed on three-dimensional highway alignments. Journal of Transportation Engineering - ASCE, 127(1), 21-30. doi:10.1061/(ASCE)0733-947X(2001)127:1(21)

Goodwin, L. C. (2002). Analysis of weather-related crashes on US highways. Weather, 2(32), 4,064.

Google (Cartographer). (n.d). Google Maps of Sabah, Malaysia

Bibliography 161

Gross, F., Eccles, K., & Nabors, D. (2011). Low-volume roads and road safety audits: lessons learned. Transportation Research Record: Journal of the Transportation Research Board(2213), 37-45.

Gui, Y., Wang, J., & Fang, S. (2011). Study on the mountainous freeway vertical alignment safety based on typical truck climbing characteristics in China. Paper presented at the Multimodal Approach to Sustained Transportation System Development: Information, Technology, Implementation.

Guo, Y., & Sun, Q. (2013). Modeling crash frequency of a typical mountainous freeway ICTIS 2013: Improving Multimodal Transportation Systems-Information, Safety, and Integration (pp. 1417-1425).

Haque, M. M., Chin, H. C., & Debnath, A. K. (2012). An investigation on multi-vehicle motorcycle crashes using log-linear models. Safety science, 50(2), 352-362.

Haque, M. M., Chin, H. C., & Huang, H. (2010). Applying Bayesian hierarchical models to examine motorcycle crashes at signalized intersections. Accident Analysis & Prevention, 42(1), 203-212.

Hashim, I. H. (2011). Analysis of speed characteristics for rural two-lane roads: A field study from Minoufiya Governorate, Egypt. Ain Shams Engineering Journal, 2(1), 43-52. doi:http://dx.doi.org/10.1016/j.asej.2011.05.005

Hauer, E. (2015). The art of regression modeling in road safety (Vol. 38): Springer. Hensher, D. A., Rose, J. M., & Greene, W. H. (2005). Applied choice analysis: a

primer: Cambridge University Press. Holdridge, J. M., Shankar, V. N., & Ulfarsson, G. F. (2005). The crash severity

impacts of fixed roadside objects. Journal of Safety Research, 36(2), 139-147.

Hosseinpour, M., Yahaya, A. S., Ahadi, M. R., Asoode, R., & Momeni, H. (2016). Determining contributory factors affecting rear-end crashes using Hurdle Count Models. Paper presented at the Transportation Research Board 95th Annual Meeting.

Hou, D., Han, J., Sun, X., & He, Y. (2010). Study on the relationship between speed difference and crash rate on freeway. Paper presented at the ICCTP 2010: Integrated Transportation Systems: Green, Intelligent, Reliable.

HPU. (2013). Road Traffic Volume Malaysia 2013. Huang, H., Chin, H. C., & Haque, M. (2009). Empirical evaluation of alternative

approaches in identifying crash hot spots. Transportation Research Record: Journal of the Transportation Research Board, 2103(1), 32-41.

Hughes, B. P., Anund, A., & Falkmer, T. (2014). System theory and safety models in Swedish, UK, Dutch and Australian road safety strategies. Accident Analysis & Prevention(0). doi:http://dx.doi.org/10.1016/j.aap.2014.07.017

Hughes, B. P., Newstead, S., Anund, A., Shu, C. C., & Falkmer, T. (2014). A review of models relevant to road safety. Accident Analysis & Prevention(0). doi:http://dx.doi.org/10.1016/j.aap.2014.06.003

Ibrahim, A. T., & Hall, F. L. (1994). Effect of adverse weather conditions on speed-flow-occupancy relationships.

Ibrahim, S. E.-B., & Sayed, T. (2011). Developing safety performance functions incorporating reliability-based risk measures. Accident Analysis & Prevention, 43(6), 2153-2159.

IDS. (2007). Socio-Economic Blueprint 2008-2025. Kota Kinabalu: Institute for Development Studies (Sabah).

Bibliography 162

Islam, S., Jones, S. L., & Dye, D. (2014). Comprehensive analysis of single-and multi-vehicle large truck at-fault crashes on rural and urban roadways in Alabama. Accident Analysis & Prevention, 67, 148-158.

Ivan, J. N., Pasupathy, R. K., & Ossenbruggen, P. J. (1999). Differences in causality factors for single and multi-vehicle crashes on two-lane roads. Accident Analysis & Prevention, 31(6), 695-704.

Ivan, J. N., Wang, C., & Bernardo, N. R. (2000). Explaining two-lane highway crash rates using land use and hourly exposure. Accident Analysis & Prevention, 32(6), 787-795. doi:http://dx.doi.org/10.1016/S0001-4575(99)00132-3

Jehle, D., Connolly, S., Godzala, M., & Cole, A. (2010). Speed kills? Not always: the New York State thruway experience. J Trauma, 69(3), 708-714. doi:10.1097/TA.0b013e3181ec6777

Jung, S., Jang, K., Yoon, Y., & Kang, S. (2014). Contributing factors to vehicle to vehicle crash frequency and severity under rainfall. Journal of Safety Research, 50, 1-10. doi:10.1016/j.jsr.2014.01.001

Jurewicz, C., Aumann, P., Bradshaw, C., Beesley, R., Lim, A., & O'Callaghan, N. (2015). Road Geometry Study for Improved Rural Safety (1925294420). Retrieved from https://www.onlinepublications.austroads.com.au/items/AP-T295-15

Jurewicz, C., Steinmetz, L., Phillips, C., Cairney, P., Veith, G., & McLean, J. (2014). Improving roadside safety: Summary report (1925037487). Retrieved from Sydney:

Kashani, A. T., & Mohaymany, A. S. (2011). Analysis of the traffic injury severity on two-lane, two-way rural roads based on classification tree models. Safety Science, 49(10), 1314-1320. doi:http://dx.doi.org/10.1016/j.ssci.2011.04.019

Kass, G. V. (1980). An exploratory technique for investigating large quantities of categorical data. Applied statistics, 119-127.

Ke, Z., & Jian, R. (2010). Research on defining the geometry index for safe driving at the curving and sloping sections for freeway. Paper presented at the ICCTP 2010: Integrated Transportation Systems: Green, Intelligent, Reliable.

Keall, M., Newstead, S. V., & Watson, L. (2006). Four-wheel drive vehicle crash involvement patterns.

Kemp, C. (1967). On a contagious distribution suggested for accident data. Biometrics, 241-255.

Kim, D.-G., Lee, Y., Washington, S., & Choi, K. (2007). Modeling crash outcome probabilities at rural intersections: Application of hierarchical binomial logistic models. Accident Analysis & Prevention, 39(1), 125-134.

Kim, D.-G., & Washington, S. (2006). The significance of endogeneity problems in crash models: an examination of left-turn lanes in intersection crash models. Accident Analysis & Prevention, 38(6), 1094-1100.

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.

Kockelman, K. M., & Kweon, Y.-J. (2002). Driver injury severity: an application of ordered probit models. Accident Analysis & Prevention, 34(3), 313-321. doi:http://dx.doi.org/10.1016/S0001-4575(01)00028-8

Ksaibati, K., Evans, B., & Shinstine, D. S. (2015). Implementation of Wyoming rural road safety program. Transportation Research Record: Journal of the Transportation Research Board(2472), 109-116.

Lee, J., & Mannering, F. (2002). Impact of roadside features on the frequency and severity of run-off-roadway accidents: an empirical analysis. Accident

Bibliography 163

Analysis & Prevention, 34(2), 149-161. doi:http://dx.doi.org/10.1016/S0001-4575(01)00009-4

Lee, J., Nam, B., & Abdel-Aty, M. (2015). Effects of pavement surface conditions on traffic crash severity. Journal of Transportation Engineering, 141(10), 04015020.

Li, D., Ranjitkar, P., Zhao, Y., Yi, H., & Rashidi, S. (2016). Analyzing pedestrian crash injury severity under different weather conditions. Traffic injury prevention, 1-4.

Li, M., Wang, Y. G., & He, X. (2014). Multivariate geometric factors contributing to crashes and injuries in mountainous freeways: A case study from Jiangxi, China. Paper presented at the Applied Mechanics and Materials.

Li, M. D., Doong, J. L., Chang, K. K., Lu, T. H., & Jeng, M. C. (2008). Differences in urban and rural accident characteristics and medical service utilization for traffic fatalities in less-motorized societies. Journal of safety research, 39(6), 623-630.

Li, W., Sun, X., & He, Y. (2010). Research on traffic accident prediction model for mountainous freeways. Paper presented at the Proceedings of the 10th International Conference of Chinese Transportation Professionals.

Li, Y., Ma, R., Niu, Y., & Wang, L. (2008). A safety protection technology system on highway roadside in China. Analysis, 26(7.3484), 30.4919.

Liang, Q., Wan, Q., Zheng, B., Yang, T., & Guo, Y. (2014). Evaluating the safety performance of freeway traffic safety facilities. Paper presented at the CICTP 2014@ sSafe, Smart, and Sustainable Multimodal Transportation Systems.

Lin, L., Jinhai, L., & Yan, W. (2013). Traffic crash characteristic analysis on mountain roads. Paper presented at the Measuring Technology and Mechatronics Automation (ICMTMA), 2013 Fifth International Conference

Lord, D., & Geedipally, S. R. (2011). The negative binomial–Lindley distribution as a tool for analyzing crash data characterized by a large amount of zeros. Accident Analysis & Prevention, 43(5), 1738-1742.

Lord, D., Washington, S., & Ivan, J. N. (2007). Further notes on the application of zero-inflated models in highway safety. Accident Analysis & Prevention, 39(1), 53-57.

Lord, D., Washington, S. P., & Ivan, J. N. (2005). Poisson, Poisson-gamma and zero-inflated regression models of motor vehicle crashes: balancing statistical fit and theory. Accident Analysis & Prevention, 37(1), 35-46. doi:http://dx.doi.org/10.1016/j.aap.2004.02.004

Lu, J., Haleem, K., Alluri, P., & Gan, A. (2013). Full versus simple Safety Performance Functions. Transportation Research Record: Journal of the Transportation Research Board, 2398(1), 83-92.

Ma, J. M., Kockelman, K. M., & Damien, P. (2008). A multivariate Poisson-lognormal regression model for prediction of crash counts by severity, using Bayesian methods. Accident Analysis and Prevention, 40(3), 964-975. doi:DOI 10.1016/j.aap.2007.11.002

Ma, L., Wang, G., Yan, X., & Weng, J. (2016). A hybrid finite mixture model for exploring heterogeneous ordering patterns of driver injury severity. Accident Analysis & Prevention, 89, 62-73. doi:http://dx.doi.org/10.1016/j.aap.2016.01.004

Ma, X., Chen, F., & Chen, S. (2015a). Empirical analysis of crash injury severity on mountainous and non-mountainous interstate highways. Traffic Injury Prevention, 16(7), 715-723. doi:10.1080/15389588.2015.1010721

Bibliography 164

Ma, X., Chen, F., & Chen, S. (2015b). Modeling crash rates for a mountainous highway by using refined-scale panel data. Transportation Research Record: Journal of the Transportation Research Board(2515), 10-16.

Malyshkina, N. V., & Mannering, F. L. (2010a). Empirical assessment of the impact of highway design exceptions on the frequency and severity of vehicle accidents. Accident Analysis & Prevention, 42(1), 131-139.

Malyshkina, N. V., & Mannering, F. L. (2010b). Zero-state Markov switching count-data models: an empirical assessment. Accident Analysis & Prevention, 42(1), 122-130.

Malyshkina, N. V., Mannering, F. L., & Tarko, A. P. (2009). Markov switching negative binomial models: an application to vehicle accident frequencies. Accident Analysis & Prevention, 41(2), 217-226.

Mannering, F. L., Shankar, V., & Bhat, C. R. (2016). Unobserved heterogeneity and the statistical analysis of highway accident data. Analytic methods in accident research, 11, 1-16.

McGinnis, R., Davis, M., & Hathaway, E. (2001). Longitudinal analysis of fatal run-off-road crashes, 1975 to 1997. Transportation Research Record: Journal of the Transportation Research Board, 1746, 47-58. doi:doi:10.3141/1746-07

McGwin Jr, G., & Brown, D. B. (1999). Characteristics of traffic crashes among young, middle-aged, and older drivers. Accident Analysis & Prevention, 31(3), 181-198.

McHugh, M. L. (2009). The odds ratio: calculation, usage, and interpretation. Biochemia Medica, 19(2), 120-126.

Mehta, G., & Lou, Y. (2013). Calibration and development of Safety Performance Functions for Alabama. Transportation Research Record: Journal of the Transportation Research Board, 2398(1), 75-82.

MET. (2017). Malaysia's Climate. Seasonal rainfall Variation in Sabah and Sarawak. Retrieved from http://www.met.gov.my/web/metmalaysia/education/climate/generalclimateofmalaysia?p_p_id=56_INSTANCE_zMn7KdXJhAGe&p_p_lifecycle=0&p_p_state=normal&p_p_mode=view&p_p_col_id=column-1&p_p_col_pos=1&p_p_col_count=2&_56_INSTANCE_zMn7KdXJhAGe_page=3

Miaou, S.-P. (1994). The relationship between truck accidents and geometric design of road sections: Poisson versus negative binomial regressions. Accident Analysis & Prevention, 26(4), 471-482. doi:http://dx.doi.org/10.1016/0001-4575(94)90038-8

Miaou, S.-P. (2013). Some limitations of the models in the highway safety manual to predict run-off-road crashes. Transportation Research Record: Journal of the Transportation Research Board, 2377(1), 38-48.

Miaou, S.-P., & Lum, H. (1993). Modeling vehicle accidents and highway geometric design relationships. Accident Analysis & Prevention, 25(6), 689-709. doi:http://dx.doi.org/10.1016/0001-4575(93)90034-T

Milton, J., & Mannering, F. (1998). The relationship among highway geometrics, traffic-related elements and motor-vehicle accident frequencies. Transportation, 25(4), 395-413. doi:10.1023/A:1005095725001

Milton, J. C., Shankar, V. N., & Mannering, F. L. (2008). Highway accident severities and the mixed logit model: An exploratory empirical analysis. Accident Analysis & Prevention, 40(1), 260-266. doi:http://dx.doi.org/10.1016/j.aap.2007.06.006

Bibliography 165

MIROS. (2014). Malaysian Insitute of Road Safety Research - Road Accident Analysis and Database System (M-ROADS).

Mitra, S., & Washington, S. (2007). On the nature of over-dispersion in motor vehicle crash prediction models. Accident Analysis & Prevention, 39(3), 459-468. doi:10.1016/j.aap.2006.08.002

Montella, A. (2009). Safety evaluation of curve delineation improvements: empirical Bayes observational before-and-after study. Transportation Research Record: Journal of the Transportation Research Board(2103), 69-79.

Morris, C. M., & Donnell, E. T. (2014). Passenger Car and Truck Operating Speed Models on Multilane Highways with Combinations of Horizontal Curves and Steep Grades. Journal of Transportation Engineering, 4014058. doi:10.1061/(ASCE)TE.1943-5436.0000715

MOT. (2010). Transport Statistics Malaysia 2010. Retrieved from Kuala Lumpur: http://www.mot.gov.my/my/Statistik%20Tahunan%20Pengangkutan/Statistik%20Pengangkutan%20Malaysia%202010.pdf

MOT. (2012). Transport Statistics Malaysia 2012. Retrieved from http://www.mot.gov.my/my/Statistik%20Tahunan%20Pengangkutan/Statistik%20Pengangkutan%20Malaysia%20Bagi%20Tahun%202012.pdf

MOT. (2014). Road Safety. Road Safety Plan for 2014 - 2020. Retrieved from http://www.mot.gov.my/en/lands/public%20safety-roads/road-security-plan-2014-2020

Nagler, J. (1994). Scobit: an alternative estimator to logit and probit. American Journal of Political Science, 230-255.

Nordfjærn, T., & Rundmo, T. (2013). Road traffic safety beliefs and driver behaviors among personality subtypes of drivers in the Norwegian population. Traffic injury prevention, 14(7), 690-696.

NRSS. (2011). National Road Safety Strategy 2011 - 2020. Retrieved from http://roadsafety.gov.au/nrss/safe-system.aspx

NTZA. (2012). Embedding The Safe System approach to Road Safety. In N. Z. T. Agencies (Ed.). New Zealand: New Zealand Transport Agencies.

O'Donnell, C., & Connor, D. (1996). Predicting the severity of motor vehicle accident injuries using models of ordered multiple choice. Accident Analysis & Prevention, 28(6), 739-753.

Oh, J., Lyon, C., Washington, S., Persaud, B., & Bared, J. (2003). Validation of FHWA crash models for rural intersections: Lessons learned. Transportation Research Record: Journal of the Transportation Research Board(1840), 41-49.

Oh, J., Washington, S., & Lee, D. (2009). Expected safety performance of rural signalized intersections in South Korea. Transportation Research Record: Journal of the Transportation Research Board, 2114(1), 72-82.

Oh, J., Washington, S., & Lee, D. (2010). Property damage crash equivalency factors to solve crash frequency-severity dilemma. Transportation Research Record: Journal of the Transportation Research Board, 2148(1), 83-92.

Pande, A., & Abdel-Aty, M. (2006). Assessment of freeway traffic parameters leading to lane-change related collisions. Accident Analysis & Prevention, 38(5), 936-948. doi:http://dx.doi.org/10.1016/j.aap.2006.03.004

PIAM. (2013). The Motor Claims Guide. Retrieved from http://www.bnm.gov.my/documents/Accident_Assist/What%20is%20Motor%20Claim%20Guide%20Combine.pdf

Prato, C. G., Bekhor, S., Galtzur, A., Mahalel, D., & Prashker, J. N. (2010). Exploring the potential of data mining techniques for the analysis of accident

Bibliography 166

patterns. Paper presented at the The 12th World Conference on Transport Research, Lisbon, Portugal.

PUSPAKOM. (2017). Routine Inspection. Retrieved from http://www.puspakom.com.my/en/inspections-a-services/types-of-inspection/routine-inspection.html

Qin, X., Ivan, J. N., & Ravishanker, N. (2004). Selecting exposure measures in crash rate prediction for two-lane highway segments. Accident Analysis & Prevention, 36(2), 183-191.

Qin, X., Ivan, J. N., Ravishanker, N., Liu, J., & Tepas, D. (2006). Bayesian estimation of hourly exposure functions by crash type and time of day. Accident Analysis & Prevention, 38(6), 1071-1080.

Qin, X., Wang, K., & Cutler, C. (2013). Logistic regression models of the safety of large trucks. Transportation Research Record: Journal of the Transportation Research Board(2392), 1-10.

Qingpan, X., Qiaojun, X., & Haiyang, L. (2014). Assessment of roadside safety on mountainous rural highway targeted by traffic engineering facilities. Bridges, 10, 9780784412442.9780784412223.

Quddus, M. A., Wang, C., & Ison, S. G. (2009). Road traffic congestion and crash severity: econometric analysis using ordered response models. Journal of Transportation Engineering, 136(5), 424-435.

Rahman, A., & Lownes, N. E. (2012). Analysis of rainfall impacts on platooned vehicle spacing and speed. Transportation Research Part F: Traffic Psychology and Behaviour, 15(4), 395-403. doi:http://dx.doi.org/10.1016/j.trf.2012.03.004

Rashidi, S., Ranjitkar, P., & Hadas, Y. (2014). Modeling Bus Dwell Time with Decision Tree-Based Methods. Transportation Research Record: Journal of the Transportation Research Board(2418), 74-83.

Rautela, P., & Pant, S. S. (2007). Delineating road accident risk along mountain roads. Disaster Prevention and Management, 16(3), 334-343. doi:10.1108/09653560710758288

REAM. (2002). REAM Guideline 2/2002 A guide on geometric design of roads. Shah Alam: Road Engineering Association of Malaysia.

Saengthong, P., & Bodhisuwan, W. (2013). Negative binomial-crack (NB-CR) distribution. International Journal of Pure and Applied Mathematics, 84(3), 213-230.

Savolainen, P. T., Mannering, F. L., Lord, D., & Quddus, M. A. (2011). The statistical analysis of highway crash-injury severities: A review and assessment of methodological alternatives. Accident Analysis & Prevention, 43(5), 1666-1676. doi:http://dx.doi.org/10.1016/j.aap.2011.03.025

Sawalha, Z., & Sayed, T. (2001). Evaluating safety of urban arterial roadways. Journal of Transportation Engineering, 127(2), 151-158.

Schneider IV, W., Savolainen, P., & Zimmerman, K. (2009). Driver injury severity resulting from single-vehicle crashes along horizontal curves on rural two-lane highways. Transportation Research Record: Journal of the Transportation Research Board(2102), 85-92.

Schumaker, L., Ahmed, M. M., & Ksaibati, K. (2016). Policy considerations for evaluating the safety effectiveness of passing lanes on rural two-lane highways with lower traffic volumes: Wyoming 59 case study. Journal of Transportation Safety & Security, 1-19. doi:10.1080/19439962.2015.1055415

Bibliography 167

Semeida, A. M. (2013). Impact of highway geometry and posted speed on operating speed at multi-lane highways in Egypt. Journal of Advanced Research, 4(6), 515-523.

Shankar, V., & Mannering, F. (1998). Modeling the endogeneity of lane-mean speeds and lane-speed deviations: a structural equations approach. Transportation Research Part A: Policy and Practice, 32(5), 311-322.

Shankar, V., Milton, J., & Mannering, F. (1997). Modeling accident frequencies as zero-altered probability processes: an empirical inquiry. Accident Analysis & Prevention, 29(6), 829-837.

Shankar, V. N., Albin, R. B., Milton, J. C., & Nebergall, M. (2000). In-service, performance-based roadside design policy: Preliminary insights from Washington State's Bridge Rail Study. Transportation Research Record: Journal of the Transportation Research Board, 1720(1), 72-79.

Shirazi, M., Lord, D., Dhavala, S. S., & Geedipally, S. R. (2016). A semiparametric negative binomial generalized linear model for modeling over-dispersed count data with a heavy tail: Characteristics and applications to crash data. Accident Analysis & Prevention, 91, 10-18.

Srinivasan, R., & Bauer, K. (2013). Safety Performance Function Development Guide: Developing Jurisdiction Specific SPFs. Retrieved from

Tay, R. (2016). Comparison of the binary logistic and skewed logistic (Scobit) models of injury severity in motor vehicle collisions. Accident Analysis & Prevention, 88, 52-55.

Theofilatos, A., & Yannis, G. (2014). A review of the effect of traffic and weather characteristics on road safety. Accident Analysis & Prevention, 72, 244-256.

Train, K. (1999). Halton sequences for mixed logit. Technical paper, Department of Economics. University of California, Berkeley.

Tulu, G. S., Washington, S., Haque, M. M., & King, M. J. (2015). Investigation of pedestrian crashes on two-way two-lane rural roads in Ethiopia. Accident Analysis & Prevention, 78(0), 118-126. doi:http://dx.doi.org/10.1016/j.aap.2015.02.011

Uchida, N., Kawakoshi, M., Tagawa, T., & Mochida, T. (2010). An investigation of factors contributing to major crash types in Japan based on naturalistic driving data. IATSS Research, 34(1), 22-30. doi:http://dx.doi.org/10.1016/j.iatssr.2010.07.002

Vadlamani, S., Chen, E., Ahn, S., & Washington, S. (2010). Identifying large truck hot spots using crash counts and PDOEs. Journal of Transportation Engineering, 137(1), 11-21.

Vangala, P., Lord, D., & Geedipally, S. R. (2014). An application of the Negative Binomial-Generalized Exponential Model for analyzing traffic crash data with excess zeros. Retrieved from

Vangala, P., Lord, D., & Geedipally, S. R. (2015). Exploring the application of the negative binomial–generalized exponential model for analyzing traffic crash data with excess zeros. Analytic Methods in Accident Research, 7, 29-36.

Wang, H., He, Y., Sun, X., & Hu, X. (2009). Effects of geometric features on rear-end crash incidence on mountainous two-lane highway. Paper presented at the International Conference on Transportation Engineering 2009.

Wang, W., Guo, W., Mao, Y., Jiang, X., Guo, H., Wets, G., & Zhang, W. (2011). Model-based simulation of driver expectation in mountainous road using various control strategies. International Journal of Computational Intelligence Systems, 4(6), 1187-1194. doi:10.1080/18756891.2011.9727867

Bibliography 168

Wang, Y.-G., Chen, K.-M., Hu, L.-W., & Pei, Y.-L. (2010). Voluntary killer: multivariate highway geometric factors contributing to crashes and collisions in China's mountainous regions. Technics Technologies Education Management-TTEM, 5(3), 531-543.

Wang, Y., Chen, K., Ci, Y., & Hu, L. (2011). Safety performance audit for roadside and median barriers using freeway crash records: case study in Jiangxi, China. Scientia Iranica, 18(6), 1222-1230.

Washington, S. (2000). Iteratively specified tree-based regression: theory and trip generation example. Journal of Transportation Engineering, 126(6), 482-491.

Washington, S., & Haque, M. M. (2013). On the commonly accepted assumptions regarding observed motor vehicle crash counts at transport system locations. Paper presented at the 92nd Annual Meeting of Transportation Research Board (TRB), Washington, D.C.

Washington, S., Haque, M. M., Oh, J., & Lee, D. (2014). Applying quantile regression for modeling equivalent property damage only crashes to identify accident blackspots. Accident Analysis & Prevention, 66, 136-146.

Washington, S. P., Karlaftis, M. G., & Mannering, F. L. (2010). Statistical and econometric methods for transportation data analysis: CRC press.

WHO. (2011). Global plan for the Decade of Action for Road Safety 2011–2020. Geneva: WHO.

WHO. (2015). Global status report on road safety 2015. Retrieved from Geneva: http://www.who.int/violence_injury_prevention/road_safety_status/2015/en/

Wu, J., Yang, X., & Mi, X. (2011). Safety treatment technique for continuous downgrade slope. Paper presented at the ICTIS 2011: Multimodal Approach to Sustained Transportation System Development: Information, Technology, Implementation, Wuhan, China.

Wu, L., Zhang, J., & Fujiwara, A. (2013). Tourism participation and expenditure behaviour: Analysis using a scobit based discrete–continuous choice model. Annals of Tourism Research, 40, 1-17.

Wu, L., Zhang, J., Fujiwara, A., & Chikaraishi, M. (2012). Analysis of tourism generation incorporating the influence of constraints based on a Scobit model. Asian Transport Studies, 2(1), 19-33.

Xu, H., Liu, G., & Chen, X. (2010). Safety evaluation of highway in mountainous area based on Fuzzy Neural Network ICCTP 2010 (pp. 777-783).

Yamamoto, T., & Shankar, V. N. (2004). Bivariate ordered-response probit model of driver’s and passenger’s injury severities in collisions with fixed objects. Accident Analysis & Prevention, 36(5), 869-876. doi:http://dx.doi.org/10.1016/j.aap.2003.09.002

Ye, F., & Lord, D. (2014). Comparing three commonly used crash severity models on sample size requirements: multinomial logit, ordered probit and mixed logit models. Analytic methods in accident research, 1, 72-85.

Ye, X., Pendyala, R. M., Washington, S. P., Konduri, K., & Oh, J. (2009). A simultaneous equations model of crash frequency by collision type for rural intersections. Safety science, 47(3), 443-452.

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.

Yu, R., & Abdel-Aty, M. (2013a). Investigating different approaches to develop informative priors in hierarchical Bayesian safety performance functions. Accident Analysis & Prevention, 56, 51-58.

Bibliography 169

Yu, R., & Abdel-Aty, M. (2013b). Multi-level Bayesian analyses for single-and multi-vehicle freeway crashes. Accident Analysis & Prevention, 58, 97-105.

Yu, R., & Abdel-Aty, M. (2014a). Analyzing crash injury severity for a mountainous freeway incorporating real-time traffic and weather data. Safety Science, 63, 50-56. doi:10.1016/j.ssci.2013.10.012

Yu, R., & Abdel-Aty, M. (2014b). Using hierarchical Bayesian binary probit models to analyze crash injury severity on high speed facilities with real-time traffic data. Accident Analysis & Prevention, 62, 161-167.

Yu, R., Abdel-Aty, M., & Ahmed, M. (2013). Bayesian random effect models incorporating real-time weather and traffic data to investigate mountainous freeway hazardous factors. Accident Analysis & Prevention, 50, 371-376. doi:10.1016/j.aap.2012.05.011

Yu, R., Abdel-Aty, M. A., Ahmed, M. M., & Wang, X. (2014). Utilizing microscopic traffic and weather data to analyze real-time crash patterns in the context of active traffic management. IEEE Transactions on Intelligent Transportation Systems, 15(1), 205-213.

Yu, R., Xiong, Y., & Abdel-Aty, M. (2015). A correlated random parameter approach to investigate the effects of weather conditions on crash risk for a mountainous freeway. Transportation research part C: emerging technologies, 50, 68-77.

Yuan, W., Fu, R., Guo, Y., Feng, H., & Shi, J. (2008). Influences of longitudinal gradient on traffic accident rate considering length of downgrade. Journal of Highway and Transportation Research and Development (English Edition), 3(2), 122-126.

Yun, Z. Z., Shui, F. W., & Zuo, W. (2013). Present situation of Yunnan Province mountain area highway traffic safety facilities investigation and comprehensive analysis. Paper presented at the 2013 Fifth International Conference on Measuring Technology and Mechatronics Automation (ICMTMA).

Zamani, H., & Ismail, N. (2010a). Negative binomial-Lindley distribution and its application. Journal of Mathematics and Statistics, 6(1), 4-9.

Zamani, H., & Ismail, N. (2010b). Poisson-weighted exponential distribution and its application on claim count data. Journal of Quality Measurement and Analysis, 6(2), 57-65.

Zamani, H., Ismail, N., & Faroughi, P. (2014). Poisson-weighted exponential univariate version and regression model with applications. Journal of Mathematics and Statistics, 10(2), 148.

Zha, L., Lord, D., & Zou, Y. (2016). The Poisson inverse Gaussian (PIG) generalized linear regression model for analyzing motor vehicle crash data. Journal of Transportation Safety & Security, 8(1), 18-35.

Zhang, C., & Ivan, J. (2005). Effects of geometric characteristics on head-on crash incidence on two-lane roads in Connecticut. Transportation Research Record: Journal of the Transportation Research Board(1908), 159-164.

Zhang, G., & Zhu, R. (2011). The highway design of mountainous and the research of traffic security. Paper presented at the ICTIS 2011: Multimodal Approach to Sustained Transportation System Development: Information, Technology, Implementation Wuhan, China.

Zhang, T., Liu, D., & Mi, X. (2010). A research on the safety characteristic of continuous downgrade segment in two-lane highway. Paper presented at the International Conference on Logistics for Sustained Economic Development.

Bibliography 170

Zhang, T., Tang, C., & He, Y. (2010). A research on the traffic safety characteristics of two-lane highway in China’s mountain area. Paper presented at the 10th International Conference of Chinese Transportation Professionals.

Zhang, T., Tang, C., & Kang, Y. (2012). Safety characteristics of two-lane highway sections passing through towns/villages in mountainous area based on Negative Binomial prediction model. Journal of Highway and Transportation Research and Development, 6, 020.

Zhi-yun, Z., Shui, F.-w., & Zuo, W. (2013). Present situation of Yunnan Province mountain area highway traffic safety facilities investigation and comprehensive analysis. Paper presented at the 2013 Fifth International Conference on Measuring Technology and Mechatronics Automation (ICMTMA).

Zhou, G. X., Chen, X. W., & Xiang, Q. J. (2014). Reach on value of Crash Reduction Factor (CRF) in safeguard technology of mountainous rural highway. Applied Mechanics and Materials, 505, 1067-1070.

Zhou, J.-H., Zhao, X.-C., Jiang, Z.-Q., Zhu, S.-Y., & Zhou, J.-C. (2005). Investigation on road traffic safety of the mountain areas in southwest China based on Wulong County, Chongqing Municipality. Traffic injury prevention, 6(2), 193-196.

Zou, Y., Lord, D., Zhang, Y., & Peng, Y. (2013). Comparison of Sichel and negative binomial models in estimating empirical Bayes estimates. Transportation Research Record: Journal of the Transportation Research Board(2392), 11-21.

Appendices 171

Appendices

APPENDIX A

Crash Report Form - POL 27 Royal Malaysian Police Original Form in Malay

Appendices 172

Appendices 173

Appendices 174

Item Translation in English

A. Details report/ Time of crash 1. State (Code) 2. District (Code) 3. Police Station (Code) 4. Report No. 5. Year 6. Month (01 – 12) 7. Date (01 – 31) 8. Time (0 – 2359) 9. Day

1. Sunday 2. Monday 3. Tuesday 4. Wednesday 5. Thursday 6. Friday 7. Saturday

10. No of vehicle involved 11. No of vehicle damaged 12. No of driver killed 13. No of driver injuries 14. No of occupant killed 15. No of occupant injuries 16. No of pedestrian killed 17. No of pedestrian injuries 18. Type of accident

1. Fatal 2. Serious Injury 3. Slight Injury 4. Property Damage Only

B. Road Information 19. Road surface type

1. Crasher run (Gravel) 2. Interlocking block (brick) 3. Bitumen / Tar Pavement 4. Concrete pavement 5. Earth

20. Traffic System 1. One way 2. Two way 3. Three lane 4. Dual carriageway

21. Road geometry 1. Straight 2. Bend 3. Roundabout 4. Cross section 5. T/Y junction

Appendices 175

6. Staggered junction 7. Interchange

22. Quality of road surface 1. Smooth 2. Corrugation 3. Potholes 4. Rutted

23. Road Condition 1. Flat 2. Slope (gradient)

24. Line Marking 1. Double 2. Single 3. One way 4. Divider (median) 5. U-Turn 6. No marking

25. Hit run 1. Yes 2. No

26. Control Type 1. Police 2. Other agencies 3. Traffic light 4. Pedestrian crossing 5. Pedestrian crossing with traffic light 6. Train crossing 7. Yellow line 8. Yellow box 9. No control

27. Road width (meter) 28. Shoulder width for both sides (meter) 29. Type of road shoulder

1. Paved 2. Unpaved

30. Road Defect 1. Road shoulder drop / raise 2. Main hole drop / raise 3. Loose gravel 4. Dusty road 5. Pothole 6. Slippery 7. Defective traffic light 8. Narrow railway crossing 9. Narrow bridge 10. No guard rails 11. No/ Insufficient street lights 12. Not relevant

31. Speed limit 1. 50km/h 2. 70 km/h

Appendices 176

3. 80 km/h 4. 90 km/h 5. 110 km/h 6. Others

32. Road surface condition 1. Dry 2. Flood 3. Wet 4. Oily 5. Sandy 6. Reconstruction work

33. Collision Type 1. Head-on 2. Rear-end 3. Right angle side 4. Angular 5. Side swipe 6. Forced 7. Hitting Animal 8. Hitting object off road 9. Hitting object on road 10. Hitting Pedestrian 11. Overturned 12. Out-of-control 13. Others

C. Environmental Information 34. Weather condition

1. Clear 2. Foggy 3. Rain

35. Lighting condition 1. Day 2. Dawn/ Dusk 3. Dark with street light 4. Dark without street light

D. Crash Location 36. Road type

1. Expressway 2. Federal road 3. State road 4. Municipal 5. Others

37. Route No Name of Road / Intersection

38. Type of location 1. City 2. Urban 3. Built-up area 4. Rural

Appendices 177

39. Type of area 1. Residential 2. Office 3. Commercial 4. Construction / Industrial 5. Bridge / Foot bridge 6. School 7. Others Nearest kilometre post Distance from _________

E. Vehicle Information 40. Vehicle brand

Manufacturing Model

41. Year manufacturing 42. Registration number 43. Type of Vehicle

1. Express bus 2. Stage bus 3. Factory bus 4. Mini bus 5. Tour/excursion bus 6. School bus 7. Four-wheel drive 8. Special duty vehicle 9. Bullock cart 10. Lorry trailer 11. Rigid lorry (>2.5tonne) 12. Small lorry (<2.5 tonne) 13. Passenger car / Wagon 14. Motorcycle > 250 cc 15. Motorcycle < 251 cc 16. Taxi 17. Trishaw 18. Van 19. Hired car 20. Bicycle

44. Type of ownership 1. Private 2. Goods 3. Service 4. Government 5. Police 6. Army

45. Part of damage 46. Vehicle movement

1. Parked 2. Suddenly stopped 3. Diverging 4. Converging

Appendices 178

5. Slippery 6. Right turn 7. Left turn 8. Overtaking 9. U-turn 10. Forward 11. Reverse 12. Others

47. Crash factors related to vehicle 1. Break 2. Broken windscreen 3. Vehicle without light 4. Light damage 5. Steering 6. Old tyre 7. Recycle tyre 8. Bold tyre 9. Wiper 10. Over smoke 11. Not applicable

48. Vehicle modified? 1. Yes 2. No

49. Length of break marked (meter) 50. Tire burst

1. Yes 2. No

51. Foreign vehicle 1. Singapore 2. Thailand 3. Diplomat 4. Brunei 5. Not applicable

F. Driver Information 52. Sex

1. Male 2. Female

53. Aged (Year) 54. Race

1. Malay 2. Chinese 3. Indian 4. Kadazan 5. Murut 6. Melanau 7. Bajau 8. Bidayuh 9. Iban 10. Foreigner 11. Others

Appendices 179

55. Licence process 1. Private 2. School driving 3. Not applicable

56. Licence Status 1. No licence 2. “L” licence 3. “P” licence, 4. Full licence 5. International licence 6. Police licence 7. Army licence

57. Driver injury 1. Fatal 2. Serious injury 3. Slight injury 4. No injury

58. Seat Belt 1. Seat belt fasten 2. Seat belt unfasten 3. Wearing Helmet 4. Serban (Turban) 5. Wearing helmet but not properly tight 6. Not wearing helmet/ Serban

59. Part of Body Injury 1. Head 2. Neck 3. Chest 4. Arm 5. Back 6. Buttock 7. Leg 8. Various 9. No injury

60. Type of Driver Fault 1. In/out vehicle 2. Negligent signalling 3. Overloading (Goods) 4. Overloading (Passenger) 5. Wrong parking 6. Drugs 7. Careless driving 8. Dangerous driving 9. Dangerous turning/ wrong turning 10. Dangerous overtaking/ wrong 11. Driving too close 12. Speeding 13. Traffic light violation 14. Other offences 15. Not at fault

61. Driver qualification

Appendices 180

1. Not schooling 2. Primary school 3. Secondary school 4. Higher education

62. Drunk Driver 1. Not suspicious 2. Tested positive 3. Tested negative

63. Driver status 1. Working 2. Student 3. Not applicable

Comments Sketch of the crash (Record of road name, direction of vehicle, last position of vehicle and others) Sketch of crash location (Show the position and distance of crash to the junction, main road, building or others)

Appendices 181

APPENDIX B

Variables used in Safety Performance Function

No Variables Mountainous Non-Mountainous

Dependent Variables

1

Crash Amount /

Crash Frequency

(number of crashes)

- Eck (1983) –D - Yuan et al. (2008) – M

(L&E) - Li, Ma, et al. (2008) – D - (Wang et al., 2010)– D,

M (L&PLN) - (Zhang, Liu, et al.,

2010)- D, M (P) - Xu, Liu, and Chen (2010)

– D(F) - Zhang and Zhu (2011) –

D - Chen, Chen, et al.

(2011) - M (SBA) - Wu et al. (2011) – D - Zhang, Tang, and Kang

(2012) – M (NB) - (Lin et al., 2013)– D - Zhi-yun et al. (2013) – D - (Chen, 2014)- D (C) - Zhou et al. (2005) – D - (Ahmed et al., 2012)– M

(BLR) - (Guo & Sun, 2013)- M

(ZINB) - Yu et al. (2013) – M (FE

& P) - Yu and Abdel-Aty

(2013b) – M (BPLN & BHP)

- Oh et al. (2009) – M (G) - Vadlamani, Chen, Ahn,

and Washington (2010) – M (NB)

- Haque, Chin, and Huang (2010) – M (PG,HPG, HPLN & HP)

- Fu et al. (2011) – D & M (E)

- Dell’Acqua and Russo (2011) – M (E)

- Wang, Chen, et al. (2011) - D

- Srinivasan and Bauer (2013) – D & M (E)

- Yu and Abdel-Aty (2013b) – M (PG & PLN)

- Oh et al. (2010) – M (NB) - Washington, Haque, Oh,

and Lee (2014) – M (NB)

2 Crashes Incident

(Fatal, Serious Injury, Slight Injury, Property Damage Only, Major injury, minor injury)

- - Kim and Washington (2006) – M (NB)

- Oh et al. (2009) – M (G) - Vadlamani et al. (2010)–

M (NB) - Oh et al. (2010) – M (NB) - Choi, Kim, Heo, and Lee

(2011) – M (OL) - Washington et al. (2014)

– M (NB) 3 Crash Type

(Angle, head-on, rear-end, sideswipe)

- Wang et al. (2009) – M (NB)

- Kim, Washington, and Oh (2006) – M (P & NB)

- Kim et al. (2007) – M (SML & MBL)

4 Crash Rate

(100 Million Vehicle per km, 100 million entering vehicle 100

- Zhang, Tang, et al. (2010) – D & M (ML, P & NB)

- (Wang et al., 2010)- D, M (L&PLN)

- Hou et al. (2010) – M (L, E, LG)

- Fu et al. (2011) – D & M (E)

Appendices 182


million vehicle travelled per segment, crashes per year, crashes per year per km)

- Li et al. (2010) – M (ML) - Ahmed et al. (2011) – M

(P, RE & S) - Li et al. (2014) – D - (Yu et al., 2015)– M

(FPT, URPT&CRPT) 5 AADT

(Number of vehicle)

- Zhang, Tang, et al. (2010) – D & M (L, P & NB)

- Ahmed et al. (2011) – M (P, RE & S)

- Miaou and Lum (1993) – M (L & P)

- Milton and Mannering (1998) – M (L, P & NB)

- Abdel-Aty and Radwan (2000) – M (NB)

- Kim and Washington (2006) – M (NB)

- Mitra and Washington (2007) – M (NB)

- Oh et al. (2009) – M (G) - Vadlamani et al. (2010)

– M (NB) - Haque et al. (2010) – M

(PG,HPG, HPLN & HP) - Malyshkina and

Mannering (2010a) – M (MML & NB)

- Ibrahim and Sayed (2011) – M (NB)

- Brimley et al. (2012) – M (NB)

- Lu, Haleem, Alluri, and Gan (2013) – M (NB)

- Ye, Pendyala, Washington, Konduri, and Oh (2009) – M (MVP)

6 ADT10k

(Average daily traffic volume in tens of thousands of vehicles)

- - Oh et al. (2010) – M (NB) - Washington et al. (2014)

– M (NB)

7 Daily vehicle miles travelled

(number)

- Ahmed et al. (2011) – M (P, RE & S)


8 Proportion of truck (%)

- (Yu et al., 2015)– M (FPT, URPT&CRPT)

- Vadlamani et al. (2010) – M (NB)

9 Truck Miles

(miles)

- - Miaou and Lum (1993) – M (L & P)

10 HVADT / HVADT 1k

(Average daily heavy vehicle volume in thousands of vehicles)

- - Oh et al. (2009) – M (G) - Oh et al. (2010) – M (NB) - Washington et al. (2014)

– M (NB)

11 Length of section / segment

- Zhou et al. (2005) – D - Zhang, Tang, et al.

(2010) – D & M (L, P &


- Milton and Mannering

Appendices 183


(meter, km, mile) NB) - Ahmed et al. (2011) – M

(P, RE & S)

(1998) – M (L, P & NB) - Yuan et al. (2008) - M

(L&E) - Abdel-Aty and Radwan

(2000) – M (NB) - Sawalha and Sayed

(2001) – M (P & NB) - Malyshkina and

Mannering (2010) – M (MML & NB)


- Fu et al. (2011) – D & M (E)



- Oh et al. (2010) – M (NB) - Washington et al. (2014)

– M (NB) Season

12 Date - - Oh et al. (2009) – M (G)

13 Time - - Oh et al. (2009) – M (G)

14 Season

(dry or snow)

- Ahmed et al. (2011) – M (P, RE & S)

- Ahmed et al. (2012)– M (BLR)

- Oh et al. (2009) – M (G) - Yu et al. (2014) – M (BL

& BREL)

15 Weather - - Kim et al. (2007) – M (SML & MBL)

- Oh et al. (2009) – M (G)

16 Large Precipitation (greater than 0.02 in)


-

Traffic Characteristic - -

17 Proportion of trucks (%)


- Ahmed et al. (2011) – M (P, RE & S)



- Lu et al. (2013) – M (NB) 18 Numbers of truck

involving in crashes

(number)

- - Miaou and Lum (1993) –M (L & P)

19 Driveway density

(driveway per km)


- Sawalha and Sayed (2001) – M (P & NB)


20 Traffic flow / Volume per lane

(vehicles/day)

- Zhou et al. (2005) – D - Choi et al. (2011) – M (OL)

21 Speed limit - Eck (1983) – D - (Ma et al., 2008)- M


Appendices 184


(km/h, mile/h) (UVP, UVNB, MVPLN) - Li et al. (2010) - M (ML)

(Wang et al., 2010)- D, M (L&PLN)

- Ahmed et al. (2011) – M (P, RE & S)

- Chen, Chen, et al. (2011) – M (SBA)

- (Guo & Sun, 2013)– M (ZINB)

- Yu and Abdel-Aty (2013b) – M (BPLN & BHP)

- Kim et al. (2006) – M (P & NB)

- Kim and Washington (2006) - M (NB)


- Haque et al. (2010) – M (PG,HPG, HPLN & HP)



- Lu et al. (2013) – M (NB) - Mehta and Lou (2013) –

M (NB) - Ye et al. (2009) – M

(MVP) - Oh et al. (2010) – M (NB) - Huang et al. (2009) – M

(PG, PLN & HP) - Washington et al. (2014)

– M (NB) 22 Mean Speed /

Average Speed (km/h, mile/h)

- Chen, Chen, et al. (2011)- M (SBA)

- (Ahmed et al., 2012)– M (BLR)

- Yu et al. (2013) – M (FE & P)

- Hou et al. (2010) – M (L, E, LG)

- Yu et al. (2014) – M (BL & BREL) 6-12 minutes before crash


23 Spot Speed / Operating Speed / Travel Speed (km/h, mile/h)

- Zhou et al. (2005) – D - (Wang et al., 2010)- D,

M (L&PLN) - Zhang and Zhu (2011) –

D - (Lin et al., 2013)– D - (Chen, 2014) – D (C)


- Ke and Jian (2010) – M (L)


- Choi et al. (2011) – M (OL)

24 Speed Control Device - - Oh et al. (2010) – M (NB) - Washington et al. (2014)

– M (NB) 25 Visibility / sight

Distance (km, mile, indicator more than 2 mile)

- Ahmed et al. (2012)– M (BLR)

- (Chen, 2014) – D (C) - (Yu et al., 2015)– M

(FPT, URPT&CRPT)

- Kim et al. (2006) – M (P & NB)


- Oh et al. (2009) – M (G) - Yu et al. (2014) – M (BL

& BREL) Vertical Alignment

26 Gradient (%) - (Ma et al., 2008)- M (UVP, UVNB, MVPLN)

- Wang et al. (2009) – M (NB)

- Li et al. (2010) - M (ML)



- Kim et al. (2007) – M

Appendices 185



- Choi et al. (2011) – M (OL)

- Chen, Chen, et al. (2011) - M (SBA)

- Ahmed et al. (2011) – M (P, RE & S)

- (Ahmed et al., 2012)– M (BLR)

- (Guo & Sun, 2013) – M (ZINB)

- Yu and Abdel-Aty (2013b) – M (BPLN & BHP)

- Yu et al. (2013) – M (FE & P)

- Li et al. (2014) – D - Yu and Abdel-Aty

(2014a) – M (FPL, SVM, RPL)

(SML & MBL) - Yuan et al. (2008) – M

(L&E) - Ke and Jian (2010) – M

(L) - Choi et al. (2011) – M

(OL) - Dell’Acqua and Russo

(2011) – M (E) - Ibrahim and Sayed

(2011) – M (NB) - Brimley et al. (2012) – M

(NB) - Srinivasan and Bauer

(2013) – D & M (E) - Yu and Abdel-Aty

(2013b) – M (PG & PLN) - Qingpan, Qiaojun, and

Haiyang (2014) – M (L) - Yu et al. (2014) – M (BL

& BREL) - Oh et al. (2010) – M (NB) - Washington et al. (2014)

– M (NB) 27 Average gradient (%) - Wu et al. (2011) – D

- Yuan et al. (2008) – M

(L&E) - Fu et al. (2011) – D & M

(E) 28 Peak gradient (%) - - Fu et al. (2011) – D & M

(E) 29 Vertical curve length

(feet, meter)

- - Miaou and Lum (1993) –M (L & P)


30 Radius of curvature (feet , meter)


- Li et al. (2014) – D

31 Change rate of vertical curvature (-)

- Wang et al. (2009) – M (NB)

32 Vertical slope length (feet, meter)

- Eck (1983) - D - Wang et al. (2009)– M

(NB) - Li et al. (2010) - M (ML) - Wu et al. (2011) – D - (Chen, 2014)– D (C)

- Fu et al. (2011) – D & M (E)

33 The algebraic different between the initial and final grade, A (-)

- - Srinivasan and Bauer (2013) – D & M (E)

34 A measure of the sharpness of vertical curvature, K (-)

- Wang et al. (2009) – M (NB)


35 Continuous Downgrade Segment (km, meter, feet, mile)

- (Zhang, Liu, et al., 2010)- D, M (P)

-

Appendices 186


36 Steep downgrade segments (≤-4%)


-

37 Site (metropolitan area, rural segments located in level and rolling areas)

- - Eck (1983) – D - Oh et al. (2010) – M (NB)

38 Altitude (m) - - Fu, Guo [22] – D & M (E)

Horizontal Alignment

39 Horizontal curvature

(horizontal curve weighted by length, consecutive curve)

- Wang et al. (2009) – M (NB)



- (Chen, 2014) – D (C) - (Wang et al., 2010)- D,

M (L&PLN) - (Yu et al., 2015)– M

(FPT, URPT&CRPT) -


- Kim et al. (2006) – M (P & NB)

- Ke and Jian (2010) – M (L)

- Oh et al. (2010) - M (NB)

- Washington et al. (2014) – M (NB)

40 Radius of horizontal curvature

(feet, meter)

- Choi et al. (2011) – M (OL)

- Ahmed et al. (2011) – M (P, RE & S)

- Zhang and Zhu (2011) – D

- (Guo & Sun, 2013)– M (ZINB)

- Li et al. (2014) – D


- Kim et al. (2007) – M (SML & MBL)

- Choi et al. (2011) – M (OL)




– M (NB) 41 Degree of curvature

(degree per feet, degree per 100feet, degree per segment)

- Eck (1983) - (Ma et al., 2008) - M

(UVP, UVNB, MVPLN) - Ahmed et al. (2011) – M

(P, RE & S) - (Ahmed et al., 2012)– M

(BLR) - (Guo & Sun, 2013)– M

(ZINB) - Yu and Abdel-Aty

(2013b) – M (BPLN & BHP)


- Malyshkina and Mannering (2010a) – M (MML & NB)

-

42 Curvature length

(feet , curve length per km, mile)

- (Ma et al., 2008) – M (UVP, UVNB, MVPLN)

- Zhang, Tang, et al. (2010) – D M (L, P & NB)

- Ahmed et al. (2011) – M (P, RE & S)


- Choi et al. (2011) – M (OL)


43 Curvature length ratio -

Appendices 187


(-)

44 Change rate of horizontal curve

(grad/km)

- Wang et al. (2009) – M (NB)

- Chen, Zhou, and Wei (2011) – D


-

45 Deflection angle

(radian, degree)

- Wang et al. (2009) – M (NB)


- Ahmed et al. (2011) – M (P, RE & S)


- (Guo & Sun, 2013)– M (ZINB)

- Li et al. (2014) – D

46 Tangent length

(meter, km, feet, mile)

- Choi et al. (2011) – M (OL)

-


- Choi et al. (2011) – M (OL)

- 47 Combination

horizontal and vertical alignment (-)

- - Ke and Jian (2010) – M (L)

48 Terrain Type

(Flat, rolling, mountainous,

- - Kim et al. (2006) – M (P & NB)

- Kim and Washington (2006) – M (NB)


- Oh et al. (2009) – M (G)


– M (NB) 49 Type of road - - Vadlamani et al. (2010)

– M (NB) Road side Features

50 Guardrail

(km, number, m)

- Zhou et al. (2005) – D - Zhi-yun et al. (2013) – D

- Lee and Mannering (2002) – M (NB & ZINB)

- Liang et al. (2014) – D (C)

- Qingpan et al. (2014) – M (L)

51 Road sign

(number, m-distance)

- Zhou et al. (2005) – D - Liang et al. (2014) – D (C)

52 Shoulder - - Kim et al. (2007) – M (SML & MBL)

53 Shoulder width

(m)



- Miaou and Lum (1993) –M (L & P)


Appendices 188


- Choi et al. (2011) – M (OL)

- Ahmed et al. (2011) – M (P, RE & S)



- Kim et al. (2006) – M (P & NB)


- Malyshkina and Mannering (2010) – M (MML & NB)

- Choi et al. (2011) – M (OL)



- Mehta and Lou (2013) – M (NB)

- Ye et al. (2009) – M (MVP)


– M (NB) 54 Shoulder type

(-)

- - Lu et al. (2013) – M (NB) - Oh et al. (2010) – M (NB) - Washington et al. (2014)

– M (NB) 55 Shoulder drop

(cm)

-

56 Embankment / hill


- - Lee and Mannering (2002) – M (NB & ZINB)


- Qingpan et al. (2014) 57 Ditch

(cm-deep, m-distance)



58 Tree




59 Concrete barrier

(cm-height, m-distance)


- Y. Wang et al. (2011) 60 Utility pole




61 Culvert



-

Appendices 189


62 Boulder



- 63 Electric pole



- Qingpan et al. (2014) 64 Fence



- 65 House/ Shop /

building




66 Miscellaneous fixed object


- - Qingpan et al. (2014) –M (L)

67 Delineation - - Oh et al. (2010) – M (NB) - Washington et al. (2014)

– M (NB) 68 Roadside Hazard

Rating - - Kim et al. (2006) – M (P

& NB) - Mitra and Washington

(2007) – M (NB) -

Cross-Section Elements

69 Lane

(number)

- Zhou et al. (2005) – D - Ahmed et al. (2011) – M

(P, RE & S) - (Guo & Sun, 2013)– M

(ZINB) - Yu and Abdel-Aty

(2013b) – M (BPLN & BHP)



- Oh et al. (2009) – M (G) - Sawalha and Sayed

(2001) – M (P & NB) - Oh et al. (2010) – M (NB) - Haque et al. (2010) – M

(PG,HPG, HPLN & HP) - Brimley et al. (2012) – M

(NB) - Mehta and Lou (2013) –

M (NB) - Yu et al. (2014) – M (BL

& BREL) - Huang, Chin, and Haque

(2009) –M (PG, PLN & HP)

- Washington et al. (2014) – M (NB)

70 Lane width

(m)


- Choi et al. (2011) – M (OL)

- (Lin et al., 2013)– D


- Oh et al. (2010) – M (NB) - Choi et al. (2011) – M

(OL)

Appendices 190



- Lu et al. (2013) – M (NB) - Washington et al. (2014)

– M (NB) 71 One way traffic - - Huang et al. (2009) –M

(PG, PLN & HP) - Haque et al. (2010) – M

(PG,HPG, HPLN & HP) 72 Turn left lane

(number, indicator)

- - Eck (1983) – D - Kim and Washington

(2006) – M (NB) - Kim et al. (2006) – M (P

& NB) - Mitra and Washington

(2007) – M (NB) - Ye et al. (2009) – M

(MVP) - Huang et al. (2009) –M

(PG, PLN & HP) - Oh et al. (2009) – M (G) - Haque et al. (2010) – M

(PG,HPG, HPLN & HP)

Appendices 191

APPENDIX C

List of Road Segments from Keningau to Kimanis

Segments Coordinate Distance (m) Altitude Gradient

1 5.32311886 116.143078 0 253 0

2 5.33646324 116.136608 2100 301 2

3 5.35290847 116.126227 2400 335 1

4 5.35330638 116.126444 69 335 0

5 5.361694 116.129472 1100 382 4

6 5.369806 116.124583 1100 435 5

7 5.380028 116.122944 1300 516 6

8 5.38735 116.121876 850 634 14

9 5.392134 116.1176 800 721 11

10 5.393056 116.114222 600 754 6

11 5.396056 116.109667 850 821 7

12 5.399944 116.10475 750 902 10

13 5.39662185 116.10161796 750 983 11

14 5.39343348 116.09985843 700 1043 9

15 5.396917 116.096944 500 1033 -2

16 5.408222 116.097056 1550 1039 0

17 5.4101442 116.09593302 250 1056 7

18 5.41335652 116.09123245 750 1164 14

19 5.41909752 116.09146044 650 1088 -12

20 5.42146333 116.09083548 280 1046 -15

21 5.42589586 116.08980015 600 1010 -6

22 5.427444 116.087611 200 1005 -3

23 5.43381026 116.08318046 950 1098 10

24 5.440944 116.080194 1000 1108 1

25 5.44265642 116.08031318 190 1125 9

26 5.44540396 116.07952997 280 1108 -3

27 5.45162527 116.07734397 600 1180 12

28 5.454722 116.074083 500 1210 6

29 5.46109863 116.06881186 1000 1268 6

30 5.46200911 116.06655344 300 1292 9

31 5.46504226 116.0642226 450 1315 5

32 5.46562165 116.06209427 300 1330 5

33 5.46880163 116.05768606 700 1370 6

34 5.46838244 116.05114952 800 1253 -10

35 5.464972 116.048556 450 1284 7

36 5.45789992 116.03977963 1500 1265 -1

37 5.45852204 116.03310093 900 1238 8

38 5.45754213 116.02860287 700 1239 0

39 5.46285017 116.0213609 1100 1054 -17

40 5.469306 116.01775 850 1039 -2

41 5.47550062 116.0162513 700 988 -7

42 5.48365733 116.01636127 1100 797 -17

43 5.48767022 116.01661339 450 766 -7

44 5.49670515 116.01394191 1100 613 -14

45 5.50145751 116.01326063 550 572 -7

46 5.50460526 116.00906298 750 472 -13

47 5.50638604 116.00844875 250 448 -10

48 5.51168296 116.00687966 600 388 -10

49 5.51612284 116.00547418 550 392 1

50 5.51963093 116.00598648 400 365 -7

51 5.52059739 116.00569144 120 356 -8

52 5.52565657 116.00421891 650 275 -12

53 5.52597694 116.0065712 270 250 -9

54 5.52904714 116.00354031 750 139 -15

55 5.53210931 116.00058988 450 127 -3

56 5.534972 115.99525 700 175 7

Appendices 192


57 5.543556 115.981639 1900 137 -2

58 5.554833 115.98325 1300 119 -1

59 5.570944 115.982944 1800 106 -1

60 5.574615 115.97804725 800 95 -1

61 5.575722 115.975306 350 81 -4

62 5.581694 115.968889 1400 71 -1

63 5.587639 115.966778 750 75 1

64 5.597861 115.965778 1200 61 -1

65 5.60125 115.958528 1000 58 0

Appendices 193

APPENDIX D

List of Road Segments from Donggongon to Tambunan


1 5.91536684 116.11633927 0 1263 0

2 5.91474721 116.12174861 600 1297 1

3 5.91569966 116.12822346 1400 1296 -1

4 5.90893443 116.13749921 3000 1342 1

5 5.91635931 116.1441806 4100 1406 0

6 5.92138031 116.15785986 5800 1458 0

7 5.91582706 116.1661157 7100 1507 4

8 5.90818473 116.1747846 8600 1532 -1

9 5.90401236 116.18468296 9900 1565 0

10 5.90111893 116.19380314 11000 1546 1

11 5.90335303 116.20525181 12700 1530 4

12 5.89817545 116.21968176 14700 1398 0

13 5.88852184 116.22633699 16100 1381 4

14 5.88841745 116.23450533 17100 1306 0

15 5.88405383 116.23872176 19000 1357 6

16 5.88024447 116.24056779 20600 1307 8

17 5.8749316 116.24108814 22600 1257 7

18 5.87126259 116.2506358 24500 1245 9

19 5.86513017 116.25794113 25700 1167 7

20 5.858083 116.270361 27700 1121 6

21 5.857333 116.277167 28600 1082 7

22 5.856167 116.281417 29150 1056 7

23 5.859694 116.287361 29950 1069 9

24 5.855278 116.292111 30700 1018 3

25 5.854694 116.29725 31450 1028 6

26 5.85675 116.299333 31850 1010 3

27 5.857306 116.303833 32450 968 8

28 5.856417 116.308111 33000 921 8

29 5.855611 116.313139 34400 911 3

30 5.851944 116.317472 35050 854 4

31 5.846 116.319667 35850 834 5

32 5.841833 116.327389 37050 823 6

33 5.836611 116.334472 38150 830 5

34 5.833583 116.335583 38550 870 5

35 5.830722 116.331556 39200 891 8

36 5.828472 116.329944 39850 917 6

37 5.828417 116.330417 39903 881 2

38 5.826083 116.337389 40903 852 3

39 5.823278 116.337667 41203 851 0

40 5.819806 116.337083 41803 782 -2

41 5.816639 116.339611 42303 770 -3

42 5.811972 116.337833 42853 746 0

43 5.805472 116.336111 43753 712 -5

44 5.799028 116.335 44553 685 -6

45 5.794333 116.335361 45103 691 -5

46 5.787972 116.337139 45953 702 -5

47 5.785639 116.340167 46453 672 -5

48 5.781083 116.341167 47203 740 -4

49 5.776611 116.343083 47803 810 -6

50 5.7744371 116.343834 48203 817 -1

51 5.772139 116.347083 48653 844 -12

52 5.769472 116.348694 49353 784 -6

53 5.76775 116.344306 49953 777 -3

54 5.76625 116.343861 50403 807 -6

55 5.766 116.346944 50753 825 -6

56 5.764861 116.349694 51103 813 -9

Appendices 194


57 5.760389 116.361167 52803 817 -5

58 5.760167 116.361028 52831 868 -7

59 5.754944 116.363389 53831 836 -6

60 5.750028 116.367833 54731 767 -6

61 5.742083 116.367417 55731 737 -7

62 5.738222 116.368333 56431 642 -7

63 5.738083 116.372333 56931 549 -7

64 5.740139 116.376889 57481 490 -4

65 5.7365 116.384639 58481 432 -4

66 5.730111 116.390917 59481 380 -4

67 5.727139 116.397806 60481 352 -1

68 5.724306 116.402972 61181 362 -2

69 5.724306 116.402972 61981 338 -3

Appendices 195

APPENDIX E

List of Road Segments from Kundasang to Tamparuli (Tamparuli – Beluran)

Segments Coordinate Distance (m) Altitude Gradient 1 5.980194,116.583278 0 1263 0

2 5.98580094 116.57429427 900 1297 4

3 5.98613706 116.57243818 1120 1296 0

4 5.98538079 116.56794682 1670 1342 8

5 5.98125135 116.56164899 2670 1406 6

6 5.98567156 116.5587683 3420 1458 7

7 5.99216176 116.55604318 4420 1507 5

8 5.99940412 116.54730991 5920 1532 2

9 6.00352009 116.54384717 6570 1565 5

10 6.00570876 116.53256178 7970 1546 -1

11 6.00454707 116.52628943 8770 1530 -2

12 6.01132246 116.51370451 10470 1398 -8

13 6.01332839 116.51236609 10740 1381 -6

14 6.01790972 116.51139915 11590 1306 -9

15 6.02007033 116.50231585 12390 1357 6

16 6.02098525 116.50015935 12960 1307 -9

17 6.02058781 116.4984186 13810 1257 -6

18 6.01932345 116.49955854 14000 1245 -6

19 6.020105 116.49437115 14850 1167 -9

20 6.01892867 116.49233535 15260 1121 -11

21 6.01356446 116.48185328 16660 1082 -3

22 6.01397524 116.47842273 17060 1056 -7

23 6.01727487 116.47145972 17910 1069 2

24 6.02634939 116.46371886 19310 1018 -4

25 6.029611 116.458861 19960 1028 2

26 6.034139 116.456667 20660 1010 -3

27 6.039056 116.455167 21460 968 -5

28 6.045028 116.455167 22360 921 -5

29 6.048417 116.455194 22760 911 -3

30 6.04975 116.453778 23560 854 -7

31 6.054861 116.447694 24560 834 -2

32 6.058222 116.446194 24960 823 -3

33 6.062765 116.444409 25510 830 1

34 6.064867 116.443143 25950 870 9

35 6.068083 116.441778 26350 891 5

36 6.071389 116.441722 26750 917 7

37 6.077639 116.440167 27550 881 -5

38 6.081694 116.438722 28050 852 -6

39 6.088333 116.436861 28850 851 0

40 6.095694 116.429833 30150 782 -5

41 6.095945 116.426581 30550 770 -3

42 6.0977113 116.425113 30810 746 -9

43 6.101056 116.422889 31510 712 -5

44 6.104106 116.42379 31910 685 -7

45 6.105417 116.423306 32070 691 4

46 6.109917 116.419917 32770 702 2

47 6.111694 116.414361 33520 672 -4

48 6.108333 116.405361 34820 740 5

49 6.108306 116.402083 35920 810 6

50 6.109814 116.402703 36110 817 4

51 6.112861 116.403194 36860 844 4

52 6.117944 116.396083 38060 784 -5

53 6.120222 116.387917 39060 777 -1

54 6.119806 116.383333 39610 807 5

55 6.118944 116.380111 40010 825 5

56 6.116528 116.374083 40810 813 -2

Appendices 196

Segments Coordinate Distance (m) Altitude Gradient 57 6.113528 116.37175 41260 817 1

58 6.107972 116.367917 42060 868 6

59 6.110167 116.362556 42860 836 -4

60 1085 116.351528 44260 767 -5

61 6.109304 116.3444329 45110 737 -4

62 6.118917 116.341389 46510 642 -7

63 6.125917 116.348306 47710 549 -8

64 6.12825 116.342083 48560 490 -7

65 6.128172 116.338086 49010 432 -13

66 6.125825 116.334522 49510 380 -10

67 6.127047 116.330716 49960 352 -6

68 6.129083 116.328083 50360 362 3

69 6.132806 116.3255 50860 338 -5

70 6.134861 116.324778 51100 336 -1

71 6.137528 116.323389 51450 316 -6

72 6.139556 116.321194 51800 330 4

73 6.144417 116.316833 52550 344 2

74 6.146944 116.313056 53050 325 -4

75 6.150361 116.308361 53750 334 1

76 6.148992 116.298921 54850 257 -7

77 6.150283 116.296587 55150 230 -9

78 6.145806 116.292472 56050 219 -1

79 6.143889 116.287167 56750 177 -6

80 6.143770 116.285971 56880 170 -5

81 6.142376 116.284040 57150 155 -6

82 6.141694 116.280944 57500 131 -7

83 6.14192219 116.27670653 58000 96 -7

84 6.14178418 116.27365686 58350 75 -6

Appendices 197

APPENDIX F

List of Road Segments from Kundasang to Ranau (Tamparuli – Beluran)


1 5.980194,116.583278 0 1238 0

2 5.98580094 116.57429427 450 1216 -5

3 `5.98613706 116.57243818 1550 1145 -6

4 5.98538079 116.56794682 1900 1124 -6

5 5.98125135 116.56164899 2230 1096 -8

6 5.98567156 116.5587683 2430 1081 -8

7 5.99216176 116.55604318 3430 1045 -4

8 5.99940412 116.54730991 3930 1022 -5

9 6.00352009 116.54384717 3995 1022 0

10 6.00570876 116.53256178 4255 999 -9

11 6.00454707 116.52628943 4805 970 -5

12 6.01132246 116.51370451 5505 929 -6

13 6.01332839 116.51236609 6205 876 -8

14 6.01790972 116.51139915 7205 818 -6

15 6.02007033 116.50231585 8105 805 -1

16 6.02098525 116.50015935 8905 827 3

17 6.02058781 116.4984186 9855 785 -4

18 6.01932345 116.49955854 10805 728 -6

19 6.020105 116.49437115 11805 661 -7

20 6.01892867 116.49233535 12705 606 -6

Appendices 198

APPENDIX G

List of Road Segments from Ranau to Beluran (Tamparuli – Beluran)


1 5.954083 116.662472 0 565 0

2 5.958306 116.672861 1000 525 -4

3 5.957935 116.674596 1200 531 3

4 5.966417 116.677722 2300 530 0

5 5.974086 116.679785 3400 558 3

6 5.979028 116.683667 4150 567 1

7 5.986167 116.690806 5250 631 6

8 5.986917 116.699694 6350 597 -3

9 5.989611 116.710417 7550 553 -4

10 5.9905 116.713611 8000 564 2

11 5.985306 116.722444 9400 551 -1

12 5.984056 116.727806 10150 527 -3

13 5.977889 116.732528 11350 461 -6

14 5.970056 116.736583 12350 426 -4

15 5.964722 116.744972 13650 427 0

16 5.963056 116.752972 14750 407 -2

17 5.955722 116.753222 15600 388 -2

18 5.951278 116.753556 16100 373 -3

19 5.953417 116.759778 16950 397 3

20 5.949944 116.76575 17700 377 -3

21 5.945361 116.770639 18500 361 -2

22 5.938222 116.77625 19500 369 1

23 5.931944 116.783861 20600 387 2

24 5.931917 116.784722 20695 379 -8

25 5.923722 116.790111 21895 402 2

26 5.918972 116.789472 22445 420 3

27 5.909111 116.789167 23545 475 5

28 5.903806 116.793083 24395 533 7

29 5.901167 116.799583 25495 501 -3

30 5.893083 116.802722 26695 493 -1

31 5.890139 116.795611 27595 488 -1

32 5.886833 116.797056 28695 403 -8

33 5.881333 116.798694 29645 320 -9

34 5.876194 116.80325 30445 307 -2

35 5.867833 116.806056 31645 277 -3

36 5.860611 116.809278 32595 284 1

37 5.851639 116.812194 33695 281 0

38 5.845167 116.815333 34495 273 -1

39 5.830833 116.816889 36295 262 -1

40 5.824361 116.818917 37045 258 -1

41 5.823833 116.819222 37113 261 4

42 5.813139 116.820667 38313 263 0

43 5.804167 116.817722 39413 251 -1

44 5.791944 116.817389 40813 247 0

45 5.785889 116.818917 41513 244 0

46 5.776333 116.822417 42713 251 1

47 5.754636 116.829237 45413 241 0

48 5.747556 116.831806 46263 225 -2

49 5.742528 116.834861 46913 231 1

50 5.73625 116.841528 47913 228 0

51 5.731611 116.847333 48763 225 0

52 .725583 116.853333 49863 239 1

53 5.721 116.860528 50863 237 0

54 5.713111 116.861583 52063 246 1

55 5.705167 116.860556 53063 272 3

56 5.698778 116.858389 53863 263 -1

Appendices 199


57 5.690028 116.863278 55063 304 3

58 5.683028 116.868111 56063 344 4 59 5.68 116.876278 57063 394 5

60 5.677444 116.885389 58163 424 3

61 5.679028 116.894 59163 366 -6

62 5.678833 116.901111 60263 391 2

63 5.675333 116.904833 61013 427 5

64 5.670111 116.909528 62013 450 2

65 5.670556 116.918167 63113 485 3

66 5.672861 116.926639 64213 445 -4

67 5.677806 116.933306 65213 427 -2

68 5.679944 116.939167 66313 440 1

69 5.679083 116.947111 67263 429 -1

70 5.680583 116.956083 68263 460 3

71 5.688444 116.961889 69363 446 -1

72 5.688861 116.970694 70563 377 -6

73 5.689 116.976722 71263 348 -4

74 5.685333 116.984944 72263 313 -4

75 5.677611 116.988361 73363 333 2

76 5.678056 116.995278 74213 281 -6

77 5.670667 117.00075 75313 248 -3

78 5.665222 117.00675 76513 190 -5

79 5.666722 117.012028 77163 162 -4

80 5.664444 117.022389 78363 162 0

81 5.660528 117.032667 79563 160 0

82 5.662611 117.038889 80463 160 0

83 5.661278 117.046167 81363 167 1

84 5.661472 117.055111 82363 153 -1

85 5.659917 117.064333 83363 147 -1

86 5.657278 117.070667 84213 138 -1

87 5.647417 117.075167 85413 187 4

88 5.638361 117.076917 86513 158 -3

89 5.630222 117.07525 87513 155 0

90 5.620583 117.078056 88713 137 -2

91 5.620444 117.079417 88863 125 -8

92 5.622583 117.087333 89763 135 1

93 5.623833 117.094222 90563 132 0

94 5.621278 117.102 91663 123 -1

95 5.62175 117.108972 92513 122 0

96 5.629667 117.109111 93463 198 8

97 5.630361 117.108611 93558 193 -5

98 5.631667 117.113444 94408 131 -7

99 5.630778 117.1195 95108 127 -1

100 5.627306 117.129833 96308 123 0

101 5.624806 117.135 96958 97 -4

102 5.6205 117.143778 98058 125 3

103 5.622 117.151139 98908 133 1

104 5.626333 117.158944 100008 118 -1

105 5.631139 117.166111 101008 129 1

106 5.638361 117.174222 102208 112 -1

107 5.642083 117.183972 103408 113 0

108 5.648778 117.187444 104258 142 3

109 5.657861 117.194556 105558 88 -4

110 5.658278 117.202861 106658 83 0

111 5.652194 117.212389 107958 80 0

112 5.655222 117.217806 108658 80 0

113 5.66025 117.228667 110058 130 4

114 5.659528 117.240056 111358 113 -1

115 5.656833 117.251167 112658 118 0

116 5.657778 117.259083 113608 166 5

117 5.661222 117.268028 114908 167 0

Appendices 200


118 5.659639 117.273417 115558 111 -9

119 5.659333 117.274139 115645 109 -2

120 5.657694 117.285417 116945 127 1

121 5.65825 117.292917 117795 117 -1

122 5.659361 117.301806 118795 118 0

123 5.659917 117.309361 119645 120 0

124 5.660685 117.313848 120145 115 -1

125 5.660528 117.318556 120645 110 -1

126 5.662139 117.323778 121245 72 -6

127 5.663222 117.328333 121745 71 0

128 5.668083 117.335611 122745 57 -1

129 5.672694 117.344306 123845 55 0

130 5.676 117.353833 124945 56 0

131 5.677083 117.361639 125795 58 0

132 5.680417 117.371917 126995 57 0

133 5.681806 117.377417 127645 53 -1

134 5.682556 117.380583 127995 49 -1

135 5.689361 117.385556 128995 47 0

136 5.693556 117.393806 129995 44 0

137 5.69838466 117.40443561 131195 50 1

Appendices 201

APPENDIX H

Appendices 202

Appendices 203

APPENDIX I

DATA SHEET FORM (ROAD GEOMETRY AND ROADSIDE FEATURES)

ENUMERATORS NAME :

1)…………….........................................………………2).............................................................................3)............................................................

LOCATION :........................................................................................................... DRAWING NO:............................................................

DATE:.......................................... START TIME:…………….................... END TIME:........................…………

LOCATION SKETCH:

Shared portion of length of horizontal and vertical

curves in same combination, X

YES / NO Lane width, LW m

Centre Line

Road shoulder width, RSW m

Broken, CLB YES / NO Length of section

Unbroken, CLU YES / NO Upgrade, LUS m

Edge Line, EL YES / NO Downgrade, LDS m

Guardrail Length,

GRL

Height,

GRH

Distance from

outside shoulder edge

(OSE) to guardrail, GD

Shoulder drop

G1 m m m Height, HSD m

Appendices 204

G2 m m m Length, LSD m

Relief Lane ,

RL

YES / NO Length,

LRL

m

G3 m m m Ditch Depth, DTT Distance from OSE, DTD

G4 m m m DT1 m m

G5 m m m DT2 m m

Average m m m DT3 m m

Earth bank / hill Length,

EBL

Height,

EBH

Distance from OSE, EBD DT4 m m

EB1 m m m DT5 m m

EB2 m m m Average m m

EB3 m m mTree

Number,

TRN

Distance from OSE, TRD

EB4 m m m TR1 m m

EB5 m m m TR2 m m

Average m m m TR3 m m

Concrete barrier Length,

CBL

Height,

CBH

Distance from OSE, CBD TR4 m m

CB1 m m m TR5 m m

CB2 m m m Average m m

CB3 m m m

CB4 m m m

CB5 m m m

Appendices 205

Average m m m

Bridge rail Length,

BRL

Height,

BRH

Distance from OSE, BRD Utility pole

Number,

UPN

Distance from OSE, UPD

BR1 m m m UP1 m m

BR2 m m m UP2 m m

BR3 m m m UP3 m m

BR4 m m m UP4 m m

BR5 m m m UP5 m m

Average m m m Average m m

Fence Length,

FCL

Height,

FCH

Distance from OSE, FCD Road sign

Number,

RSN

Distance from OSE, RSD

FC1 m m m UP1 m m

FC2 m m m UP2 m m

FC3 m m m UP3 m m

FC4 m m m UP4 m m

FC5 m m m UP5 m m

Average m m m Average m m

House/shop/building Number, HSN Distance from OSE, HSD Culvert Number,

CVN

Distance from OSE, CVD

HS1 m m CV1 m m

HS2 m m CV2 m m

HS3 m m CV3 m m

HS4 m m CV4 m m

Average m m Average m m

Appendices 206

Light poles Number, LPN Distance from OSE, LPD

Boulder Number,

BLN

Distance from OSE, BLD

LP1 m m BL1 m m

LP2 m m BL2 m m

LP3 m m BL3 m m

LP4 m m BL4 m m

LP5 m m BL5 m m

Average m m Average m m

Miscellaneous fixed

object

Numbers, MON Distance from OSE,

MOD

MO1 m m MO4 m m

MO2 m m MO5 m m

MO3 m m Average m m

Appendices 207

APPENDIX J

DATA SHEET FORM (VOLUME STUDY)

ENUMERATORS NAME :1)……………......................…………………………

2)...................................................................................

3)...................................................................................

LOCATION :...........................................................................................................

START TIME:…………….................... END TIME:........................…………

TRAFFIC APPROACH: ………………………………………………………….

Time (Hour/

Minute)

Class (P.C.U’s Conversion factors )*

…….. hour/s ………hour/s

15 30 45 60 15 30 45 60

Motorcycles (1.00)

Passenger Cars (1.00)

Light Vans (2.00)

Medium Lorries (2.50)

Heavy Lorries (3.00)

Buses (3.00)

*road class and conversion factors to passenger car unit (P.C.U’s) based on

(Department, 1986), rural standards.

Appendices 208

APPENDIX K

DATA SHEET FORM (SPOT SPEED STUDY)

ENUMERATORS NAME :1)……………......................………………………… 2)................................................................................... 3)................................................................................... DATE:............................................ WEATHER :....................................... LOCATION :........................................................................................................... START TIME:…………….................... END TIME:........................………… TRAFFIC APPROACH: ………………………………………………………….

NO SPEED (km/h) VEHICLE REG. NO

Appendices 209

APPENDIX L

Modelling Results of Random Parameter Negative – Binomial for Single – Vehicle Crashes (NLogit 5) +-----------------------------------------------------------------+ | Variable = ____________ Variable Groups Max Min Average | | TI Group sizes SEGMENTS 103 60 1 59.4 | +-----------------------------------------------------------------+ -------------------------------------------------------------------- Poisson Regression Start Values for SINGLE Dependent variable SINGLE_C Log likelihood function -2194.90830 Restricted log likelihood -2341.70455 Chi squared [ 13 d.f.] 293.59249 Significance level .00000 McFadden Pseudo R-squared .0626878 Estimation based on N = 6120, K = 14 Inf.Cr.AIC = 4417.8 AIC/N = .722 Model estimated: Jan 23, 2017, 15:28:57 --------+----------------------------------------------------------- Standard Prob. 95% Confidence SINGLE_C| Coefficient Error z |z|>Z* Interval --------+----------------------------------------------------------- LNADT | .49439*** .08444 5.85 .0000 .32889 .65989 LNLENGTH| 1.00724*** .10243 9.83 .0000 .80647 1.20800 MAX_GRE| .26474*** .07935 3.34 .0008 .10922 .42026 PRO_SH_B| -2.87789*** .62328 -4.62 .0000 -4.09950 -1.65628 PR_OS_S1| -.38222** .15575 -2.45 .0141 -.68749 -.07695 PO_OS_EM| .72221*** .16205 4.46 .0000 .40459 1.03983 NO_RDEAL| -.52517*** .09073 -5.79 .0000 -.70300 -.34735 MAX_R1| -.13109** .05315 -2.47 .0136 -.23525 -.02692 RAINFALL| .11559*** .01092 10.58 .0000 .09418 .13700 VISIBILI| .09575*** .01676 5.71 .0000 .06291 .12859 Constant|-14.4196*** .88589 -16.28 .0000 -16.1559 -12.6833 PRO_CDS| .33439*** .12786 2.62 .0089 .08379 .58498 D_O_SH2| .00772*** .00194 3.98 .0001 .00391 .01152 SP_UP_DW| .63818*** .10983 5.81 .0000 .42291 .85345 --------+-----------------------------------------------------------Note: ***, **, * ==> Significance at 1%, 5%, 10% level. -------------------------------------------------------------------- Normal exit: 40 iterations. Status=0, F= 2002.447 -------------------------------------------------------------------- Random Coefficients NegBnReg Model Dependent variable SINGLE_C Log likelihood function -2002.44749 Restricted log likelihood -2194.90830 Chi squared [ 4 d.f.] 384.92163 Significance level .00000 McFadden Pseudo R-squared .0876851 Estimation based on N = 6120, K = 19 Inf.Cr.AIC = 4042.9 AIC/N = .661 Model estimated: Jan 23, 2017, 15:32:51 Unbalanced panel has 103 individuals Negative binomial regression model --------+----------------------------------------------------------- Standard Prob. 95% Confidence

Appendices 210

SINGLE_C| Coefficient Error z |z|>Z* Interval --------+----------------------------------------------------------- |Nonrandom parameters LNADT| .45137*** .09140 4.94 .0000 .27224 .63050 LNLENGTH| 1.07070*** .10882 9.84 .0000 .85741 1.28399 MAX_GR5| .21642** .08679 2.49 .0126 .04632 .38653 PRO_SH_B| -2.36317*** .69499 -3.40 .0007 -3.72532 -1.00102 PR_OS_S1| -.54008*** .19147 -2.82 .0048 -.91536 -.16480 PO_OS_EM| .52529*** .17501 3.00 .0027 .18227 .86831 NO_RDEAL| -.47691*** .09413 -5.07 .0000 -.66139 -.29242 MAX_R1| -.18336*** .05976 -3.07 .0022 -.30050 -.06622 RAINFALL| .11589*** .01396 8.30 .0000 .08852 .14325 VISIBILI| .08329*** .01300 6.40 .0000 .05780 .10878 |Means for random parameters Constant|-14.2593*** .95484 -14.93 .0000 -16.1307 -12.3878 PRO_CDS| .08552 .15526 .55 .5817 -.21879 .38984 D_O_SH2| .00502** .00237 2.12 .0340 .00038 .00966 SP_UP_DW| .60398*** .12003 5.03 .0000 .36872 .83924 |Scale parameters for dists. of random parameters Constant| .31320*** .04075 7.69 .0000 .23333 .39307 PRO_CDS| .78854*** .06518 12.10 .0000 .66079 .91628 D_O_SH2| .01391*** .00190 7.33 .0000 .01019 .01764 SP_UP_DW| .53643*** .11728 4.57 .0000 .30658 .76629 |Dispersion parameter for NegBin distribution ScalParm| 2.44485** .99715 2.45 .0142 .49047 4.39923 --------+----------------------------------------------------------- Note: ***, **, * ==> Significance at 1%, 5%, 10% level. -------------------------------------------------------------------- Partial derivatives of expected val. with respect to the vector of characteristics. They are computed at the means of the Xs. Conditional Mean at Sample Point .0637 Scale Factor for Marginal Effects .0637 --------+----------------------------------------------------------- | Partial Prob. 95% Confidence SINGLE_C| Effect Elasticity z |z|>Z* Interval --------+----------------------------------------------------------- LNADT| .02877*** 3.29678 4.94 .0000 .01735 .04018 LNLENGTH| .06824*** 7.10454 9.92 .0000 .05476 .08171 MAX_GR | .01379** .09548 2.49 .0129 .00292 .02466 PRO_SH_B| -.15061*** -.08039 -3.39 .0007 -.23759 -.06363 PR_OS_S1| -.03442*** -.11336 -2.81 .0050 -.05843 -.01041 PO_OS_EM| .03348*** .32447 2.99 .0028 .01153 .05543 NO_RDEAL| -.03039*** -.19637 -5.08 .0000 -.04213 -.01866 MAX_R1| -.01169*** -.15366 -3.07 .0021 -.01915 -.00422 FINAL_RA| .00739*** .04953 8.26 .0000 .00563 .00914 VISI_FIN| .00531*** 1.01144 6.40 .0000 .00368 .00693 PRO_CDS| .00545 .05588 .55 .5818 -.01395 .02485 D_O_SH2| .00032** .06462 2.11 .0347 .00002 .00062 SP_UP_DW| .03849*** .04727 4.80 .0000 .02279 .05420 --------+----------------------------------------------------------- z, prob values and confidence intervals are given for the partial effect Note: ***, **, * ==> Significance at 1%, 5%, 10% level. --------------------------------------------------------------------

Appendices 211

APPENDIX M

Modelling Results of Random Parameter Negative – Binomial for Multi – Vehicle Crashes (NLogit 5) +-----------------------------------------------------------------+ | Variable = ____________ Variable Groups Max Min Average | | TI Group sizes SEGMENTS 102 5 5 5.0 | +-----------------------------------------------------------------+ +------------------------------------------------------+ | Frequency count for group sizes of TI | | Group size = 5 Pct = 100.00% CumPct = 100.00% | +------------------------------------------------------+ -------------------------------------------------------------------- Poisson Regression Start Values for MULTI_ Dependent variable MULTI_CR Log likelihood function -557.13540 Estimation based on N = 510, K = 7 Inf.Cr.AIC = 1128.3 AIC/N = 2.212 Model estimated: Jan 23, 2017, 20:08:08 ----------+--------------------------------------------------------- Standard Prob. 95% Confidence SINGLE_C | Coefficient Error z |z|>Z* Interval ----------+---------------------------------------------------------Constant | -9.18406*** 1.54301 -5.95 .0000 -12.20831 -6.15981 Log(ADTXL)| .57970*** .10867 5.33 .0000 .36672 .79268 RAIN_IND | .88829*** .27078 3.28 .0010 .35757 1.41902 LANE | -1.00744*** .36785 -2.74 .0062 -1.72841 -.28648 NO_JUNC | .21183*** .04154 5.10 .0000 .13041 .29326 NO_RDEAL | -.09640 .13442 -.72 .4733 -.35987 .16707 CVAH4_1 | .16493 .13982 1.18 .2381 -.10910 .43897 ----------+--------------------------------------------------------- Note: ***, **, * ==> Significance at 1%, 5%, 10% level. -------------------------------------------------------------------- Normal exit: 22 iterations. Status=0, F= 437.8771 -------------------------------------------------------------------- Random Coefficients NegBnReg Model Dependent variable MULTI_CR Log likelihood function -437.87710 Restricted log likelihood -557.13540 Chi squared [ 2 d.f.] 238.51659 Significance level .00000 McFadden Pseudo R-squared .2140562 Estimation based on N = 510, K = 10 Inf.Cr.AIC = 895.8 AIC/N = 1.756 Model estimated: Jan 23, 2017, 20:08:22 Unbalanced panel has 102 individuals Negative binomial regression model ----------+--------------------------------------------------------- Standard Prob. 95% Confidence SINGLE_C | Coefficient Error z |z|>Z* Interval ----------+--------------------------------------------------------- |Nonrandom parameters Constant | -8.24464*** 1.51919 5.43 .0000 -11.22221 -5.26708 Log(ADTXL)| .51437*** .10730 4.80 .0000 .30436 .72498 RAIN_IND | .89987** .48073 1.87 .0612 -.04235 1.84209 LANE | -1.05455** .41523 -2.54 .0111 -1.86837 -.24072 NO_JUNC | .21033*** .04424 4.75 .0000 .12362 .29703 |Means for random parameters NO_RDEAL | -.21328 .15931 -1.34 .1806 -.52552 .09895

Appendices 212

CVAH4_1 | .01668 .15763 .11 .9157 -.29227 .32562 |Scale parameters for dists. of random parameters NO_RDEAL | .60933*** .13968 4.36 .0000 .33556 .88310 CVAH4_1 | .41933*** .11862 3.53 .0004 .18683 .65182 |Dispersion parameter for NegBin distribution ScalParm | 2.06248*** .77285 2.67 .0076 .54771 3.57724 --------+----------------------------------------------------------- Note: ***, **, * ==> Significance at 1%, 5%, 10% level. -------------------------------------------------------------------- -------------------------------------------------------------------- Partial derivatives of expected val. with respect to the vector of characteristics. They are computed at the means of the Xs. Conditional Mean at Sample Point .3281 Scale Factor for Marginal Effects .3281 ----------+--------------------------------------------------------- | Partial Prob. 95% Confidence MULTI_CR | Effect Elasticity z |z|>Z* Interval ----------+--------------------------------------------------------- Log(ADTXL)| .16887*** 7.17708 4.36 .0000 .09294 .24480 RAIN_IND | .29525** .01764 1.90 .0570 -.00874 .59925 LANE | -.34600** -.15508 -2.43 .0151 -.62508 -.06693 NO_JUNC | .06901*** .16909 4.13 .0000 .03624 .10178 NO_RDEAL | -.06998 -.08782 -1.30 .1942 -.17563 .03567 CVAH4_1 | .00547 .00932 .10 .9165 -.09684 .10778 --------+----------------------------------------------------------- z, prob values and confidence intervals are given for the partial effect Note: ***, **, * ==> Significance at 1%, 5%, 10% level. --------------------------------------------------------------------

traffic safety along rural mountainous highways in malaysia · variables (56 in total) representing...

Documents