@prefix vivo: . @prefix edm: . @prefix ns0: . @prefix dcterms: . @prefix dc: . @prefix skos: . vivo:departmentOrSchool "Applied Science, Faculty of"@en, "Civil Engineering, Department of"@en ; edm:dataProvider "DSpace"@en ; ns0:degreeCampus "UBCV"@en ; dcterms:creator "Lin, Fred C."@en ; dcterms:issued "2009-07-06T22:20:59Z"@en, "2000"@en ; vivo:relatedDegree "Master of Applied Science - MASc"@en ; ns0:degreeGrantor "University of British Columbia"@en ; dcterms:description """This thesis describes the development of an economic-based prediction and evaluation procedure that can be utilized in prioritizing safety improvement projects. The objective of the procedure is to alleviate the problems associated with traditional economic analysis of road safety improvement programs. Traditionally, the overall effectiveness of safety improvement programs is normally based on the benefits anticipated from a reduction in road collisions. Although the procedures for performing the economic analysis of road safety improvements in general are reasonably straightforward and well documented in the literature, these procedures fail to accurately estimate the safety benefits or disbenefits of these improvements on a consistent basis. The problems can be categorized into two parts: system-wide versus project-level analysis, and dealing with the uncertainties in the effectiveness and applicability of the proposed countermeasure. To resolve these problems, this thesis first describes the development of a new safety analysis software known as ISECR (the Information System for Estimating Crash Reductions) which can be used to determine the expected collision reduction due to a specific countermeasure. ISECR is an intelligent database that uses a case-based reasoning approach and consists of past safety research efforts on collision reduction factors (CRFs) associated with different countermeasures. The system can be used to determine the expected CRFs and the associated range and reliability of the proposed countermeasure when applied to a particular problem at hand. The safety benefits of implementing a countermeasure at a location can be represented by the expected reduction in collision frequency, which is normally calculated by the product of CRFs and the expected number of collisions. With the ISECR predictions on CRFs and the expected number of collisions determined by the multivariate and Empirical Bayes methods, this thesis then illustrates the use of the moment approach to evaluate the expected collision reduction and its uncertainty. The results can then be used to assist in evaluating the economic feasibility of a countermeasure prior to its implementation. Specifically, the probability of achieving a preset economic goal (i.e., a specific benefit-cost ratio) by implementing a countermeasure at a specific location can be determined. Finally, the prototype ISECR has been verified and validated using several case studies. The results of the verification and validation have shown that ISECR produced results that are comparable to the results obtained from real cases."""@en ; edm:aggregatedCHO "https://circle.library.ubc.ca/rest/handle/2429/10284?expand=metadata"@en ; dcterms:extent "6698254 bytes"@en ; dc:format "application/pdf"@en ; skos:note "PRE-IMPLEMENTATION EVALUATION OF SAFETY IMPROVEMENT PROGRAMS by FRED C. LIN B.A.Sc. (Civil Engineering), The University of British Columbia, 1998 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF APPLIED SCIENCE in THE FACULTY OF GRADUATE STUDIES DEPARTMENT OF CIVIL ENGINEERING We accept this thesis as conforming to the required standard THE UNIVERSITY OF BRITISH COLUMBIA April, 2000 © Fred C. Lin, 2000 In presenting this thesis in partial fulfillment of the requirements for an advanced degree at the University of British Columbia, I agree that the library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Department of Civi l Engineering The University of British Columbia Vancouver, Canada Date 2 , Abstract 11 ABSTRACT This thesis describes the development of an economic-based prediction and evaluation procedure that can be utilized in prioritizing safety improvement projects. The objective of the procedure is to alleviate the problems associated with traditional economic analysis of road safety improvement programs. Traditionally, the overall effectiveness of safety improvement programs is normally based on the benefits anticipated from a reduction in road collisions. Although the procedures for performing the economic analysis of road safety improvements in general are reasonably straightforward and well documented in the literature, these procedures fail to accurately estimate the safety benefits or disbenefits of these improvements on a consistent basis. The problems can be categorized into two parts: system-wide versus project-level analysis, and dealing with the uncertainties in the effectiveness and applicability of the proposed countermeasure. To resolve these problems, this thesis first describes the development of a new safety analysis software known as ISECR (the Information System for Estimating Crash Reductions) which can be used to determine the expected collision reduction due to a specific countermeasure. ISECR is an intelligent database that uses a case-based reasoning approach and consists of past safety research efforts on collision reduction factors (CRFs) associated with different countermeasures. The system can be used to determine the expected CRFs and the associated range and reliability of the proposed countermeasure when applied to a particular problem at hand. The safety benefits of implementing a countermeasure at a location can be represented by the expected reduction in collision frequency, which is normally calculated by the product of CRFs and the expected number of collisions. With the ISECR predictions on CRFs and the expected Pre-Implementation Evaluation of Safety Improvement Programs Abstract iii number of collisions determined by the multivariate and Empirical Bayes methods, this thesis then illustrates the use of the moment approach to evaluate the expected collision reduction and its uncertainty. The results can then be used to assist in evaluating the economic feasibility of a countermeasure prior to its implementation. Specifically, the probability of achieving a preset economic goal (i.e., a specific benefit-cost ratio) by implementing a countermeasure at a specific location can be determined. Finally, the prototype ISECR has been verified and validated using several case studies. The results of the verification and validation have shown that ISECR produced results that are comparable to the results obtained from real cases. Pre-Implementation Evaluation of Safety Improvement Programs Table of Contents iv T A B L E O F C O N T E N T S A B S T R A C T i i T A B L E OF CONTENTS iv LIST OF FIGURES vi i i LIST OF T A B L E S x A C K N O W L E D G E M E N T S xi 1.0 INTRODUCTION 1 1.1 Background 1 1.2 Problem Definition 2 1.3 Thesis Objectives 3 1.4 Methodology 4 1.5 Thesis Structure 4 2.0 L I T E R A T U R E REVIEW 6 2.1 Introduction 6 2.2 Case-Based Reasoning 6 2.2.1 C B R Applications in Transportation Engineering 8 2.2.2 General C B R Issues and Procedures 9 2.2.3 Representation Issues of CBR 9 2.2.4 Control Issues of CBR 12 2.2.4.1 Case Retrieval 13 2.2.4.2 Case Adaptation 15 2.2.4.3 Case Combination 16 2.2.5 C B R Methodologies 17 Pre-Implementation Evaluation of Safety Improvement Programs Table of Contents v 2.2.5.1 Nearest Neighbour Approach 18 2.2.5.2 Collaborative Approach 18 2.3 Collision Analysis 21 2.3.1 The Generalized Linear Modeling (GLIM) Approach 21 2.3.2 The Empirical Bayes (EB) Approach 23 3.0 C B R APPLICATION IN ISECR 25 3.1 Introduction 25 3.2 Case Representation in ISECR 26 3.3 Case Retrieval in ISECR 33 3.4 Quality Score 34 3.4.1 Changes in Traffic Volume 35 3.4.2 Inclusion of Unrelated Effects 35 3.4.3 Regression to the Mean (RTM) Artifact 36 3.4.4 Calculation of Quality Scores 37 3.5 Relevance Score 39 3.5.1 Calculation of Relevance Scores 39 3.6 Calculating Case Distance in ISECR 41 3.7 Solution Construction in ISECR 43 3.7.1 Nearest Neighbour Approach 43 3.7.2 Collaborative Approach 43 4.0 DESCRIPTION OF ISECR 48 4.1 Main Menu 48 4.2 View A l l Documents 49 4.3 Search Documents 53 Pre-Implementation Evaluation of Safety Improvement Programs Table of Contents vi 4.4 Predict CR Factors (CRFs) 54 4.5 Input New Documents 64 5.0 V A L I D A T I O N OF ISECR 65 5.1 Introduction 65 5.2 ISECR Results in the Context of Published Studies 65 6.0 E X P E C T E D COLLISION REDUCTION 68 6.1 Introduction 68 6.2 Collision Reduction and Its Uncertainty 69 7.0 ECONOMIC A N A L Y S I S 71 7.1 Introduction 71 7.2 Benefit-Cost Ratio (BCR) and Its Uncertainty 71 7.3 Probability Density Function of B C R 72 8.0 APPLICATIONS 74 8.1 Problem Definition 74 8.2 Countermeasure Effectiveness 75 8.3 Expected Collision Reduction 84 8.4 Economic Analysis -. 85 9.0 CONCLUSION 87 B I B L I O G R A P H Y 90 APPENDIX A: C O U N T E R M E A S U R E TYPES 94 APPENDIX B: L O C A T I O N CHARACTERISTICS CONSIDERED FOR E A C H OF THE EIGHT L O C A T I O N TYPES 98 APPENDIX C: ISECR WINDOWS: C O U N T E R M E A S U R E TYPES CONSIDERED FOR DIFFERENT C O U N T E R M E A S U R E CATEGORIES 105 Pre-Implementation Evaluation of Safety Improvement Programs Table of Contents vii APPENDIX D: ISECR WINDOWS: INPUT FORMS USED TO ENTER L O C A T I O N CHARACTERISTICS FOR DIFFERENT LOCATION TYPES 116 APPENDIX E: PROBABILITY DENSITY PLOT OF B C R FOR THE PUBLIC LIGHTING E X A M P L E 122 Pre-Implementation Evaluation of Safety Improvement Programs List of Figures viii LIST O F F I G U R E S Figure 2.1. Representation of a Case in Case Base 12 Figure 2.2. Nearest Neighbour Approach 19 Figure 2.3. Collaborative Approach 20 Figure 3.1. Representation of a Case in ISECR 32 Figure 4.1. ISECR Window: Main Menu 49 Figure 4.2. ISECR Window: View A l l Documents 50 Figure 4.3. ISECR Window: Summary of Documents 51 Figure 4.4. ISECR Window: An Example of One Page Summary 52 Figure 4.5. ISECR Window: Selecting Query Parameters 53 Figure 4.6. ISECR Window: Countermeasure Categories 55 Figure 4.7. ISECR Window: Countermeasure Types for Delineation 56 Figure 4.8. ISECR Window: Location Types 57 Figure 4.9. ISECR Window: No Match Found 58 Figure 4.10. ISECR Window: Selecting to Enter Location Characteristics 59 Figure 4.11. ISECR Window: Input Location Characteristics for Signalized Intersections 59 Figure 4.12. ISECR Window: Selecting the Criteria to Calculate Case Distances 60 Figure 4.13. ISECR Window: Summary of Ranked Documents 61 Figure 4.14. ISECR Window: Predicted CRFs and their Standard Deviations 63 Figure 8.1. Example: Cumulative Distribution Plot of B C R 86 Figure C . l . ISECR Window: Countermeasure Types for Area-Wide Schemes 106 Figure C.2. ISECR Window: Countermeasure Types for Bridge Improvements 106 Figure C.3. ISECR Window: Countermeasure Types for Cyclist/Pedestrian Facilities 107 Figure C.4. ISECR Window: Countermeasure Types for Delineation 107 Pre-Implementation Evaluation of Safety Improvement Programs List of Figures ix Figure C.5. ISECR Window: Countermeasure Types for Geometric Improvements 108 Figure C.6. ISECR Window: Countermeasure Types for Intersection Improvements 109 Figure C.7. ISECR Window: Countermeasure Types for Lane/Shoulder Treatment 110 Figure C.8. ISECR Window: Countermeasure Types for Lighting Improvements 110 Figure C.9. ISECR Window: Countermeasure Types for Object Removal/Relocation I l l Figure C.10. ISECR Window: Countermeasure Types for Parking Improvements I l l Figure C . l 1. ISECR Window: Countermeasure Types for Pavement Treatment 112 Figure C.12. ISECR Window: Countermeasure Types for Railway Improvements 112 Figure C.13. ISECR Window: Countermeasure Types for Regulation Change 113 Figure C.14. ISECR Window: Countermeasure Types for Safety Barriers 113 Figure C.15. ISECR Window: Countermeasure Types for Traffic Controls/Signs 114 Figure C.16. ISECR Window: Countermeasure Types for Traffic Signals 115 Figure D . l . ISECR Window: Input Location Characteristics for General Intersections 117 Figure D.2. ISECR Window: Input Location Characteristics for Signalized Intersections 118 Figure D.3. ISECR Window: Input Location Characteristics for Unsignalized Intersections 119 Figure D.4. ISECR Window: Input Location Characteristics for Road Sections 120 Figure D.5. ISECR Window: Input Location Characteristics for Freeways 121 Figure D.6. ISECR Window: Input Location Characteristics for Bridges, Rails, and Construction Zones 121 Figure E . l . Example: Probability Density Plot of B C R 123 Pre-Implementation Evaluation of Safety Improvement Programs List of Tables x LIST O F T A B L E S Table 3.1. Representation of Location Characteristics for Signalized Intersections 30 Table 3.2. Example: Representation of the Location Characteristics 40 Table 3.3. Example: Determining Non-Weighted and Weighted CRFs 45 Table 4.1. Example: Ranking of Evaluation Studies 62 Table 4.2. Example: Final Output to the Summary Window 62 Table 5.1. Summary of CRFs: ISECR Results vs. Published Results 66 Table 8.1. Example: Retrieved Cases and their Quality Scores 77 Table 8.2. Example: Retrieved Cases and their Relevance Scores 79 Table 8.3. Example: Retrieved Cases (Ranked) and their Case Distances and Results 81 Table A . l . Countermeasure Types 95 Table B . l . Representation of Location Characteristics for General Intersections 99 Table B.2. Representation of Location Characteristics for Signalized Intersections 100 Table B.3. Representation of Location Characteristics for Unsignalized Intersections 101 Table B.4. Representation of Location Characteristics for Road Sections 102 Table B.5. Representation of Location Characteristics for Freeways 103 Table B.6. Representation of Location Characteristics for Bridges, Rails, and Construction Zones 104 Pre-Implementation Evaluation of Safety Improvement Programs Acknowledgements xi A C K N O W L E D G E M E N T S I would firstly like to express my gratitude to my thesis advisor, Dr. Tarek Sayed, for his constant guidance and supervision throughout the duration of this research. He was very patient, helpful, and provided me with constructive suggestions to this research. I would also like to thank the Insurance Corporation of British Columbia for providing the financial support for this research and G.D. Hamilton & Associates Consulting L T D . for providing the document summaries used in the ISECR database. Furthermore, I would also like to thank Joan Ng and Vikki Ngan for locating and copying the evaluation studies, and Shaun Bidulka for entering them into the ISECR database. Most important of all, I would like to specially thank my family for their support and encouragement throughout, not just this research, but my whole life. Their influence and example have inspired me to reach many goals in my life. Thank you, mom and dad, and of course, Frank and Kelly. Last but not least, special thanks to that special person (you know who you are, Gillian!) for the loving support and most important of all, patience during this endeavor. You are simply the best! Pre-Implementation Evaluation of Safety Improvement Programs Chapter 1: Introduction 1.0 INTRODUCTION 1.1 Background In response to limited budgets and growing fiscal constraints, it has become very important to ensure the funding available for road safety improvements is efficiently utilized. Traditionally, funding allocated for road safety improvement programs has been proportionally low compared to other road projects. Nevertheless, due to the increasing public awareness and high social and economic costs of road collisions in recent years, it has become apparent to road agencies responsible for road investment and improvement that the importance of road safely can not be overemphasized. Consequently, it is crucial to improve safety evaluation procedure to ensure an optimal allocation of the available funding. In an attempt to maximize the overall safety benefit to road users, safety professionals have developed and invoked a standard process to evaluate the cost-effectiveness of road safety projects and programs. Typical safety improvement programs, or commonly referred to as Black Spot Programs, usually include the identification, diagnosis, and remedy of collision-prone locations. In evaluating these programs, the overall effectiveness is normally based on the safety benefits anticipated from a reduction in road collision frequency and/or severity following the implementation of a safety improvement. Procedures for performing the economic analysis of road improvement programs in general are reasonably straightforward and are well documented in the literature. However, the problem with these procedures, when applied to estimating the effectiveness of road safety improvements, is that they do not always accurately estimate the safety benefits of these improvements. It is generally felt by many professionals associated with Pre-Implementation Evaluation of Safety Improvement Programs Chapter 1: Introduction 2 road safety, that a comprehensive and systematic economic-based approach for the accurate pre-implementation evaluation of road safety programs does not currently exist 1.2 Problem Definition The problem associated with the traditional economic evaluation procedures arises from the estimation of safety benefits of the proposed improvements. The safety benefits are represented by the expected reduction in the number and/or severity of collisions following the implementation of the improvement. The collision reduction is calculated as the product of the countermeasure effectiveness and the expected number of collisions. The problem can be categorized into two parts: system-wide versus project-level analysis, and dealing with the uncertainties in the effectiveness and applicability of the proposed countermeasure. A very important step in estimating collision reduction is determining the effectiveness of countermeasures, or what is known in the literature as collision reduction factors (CRFs). A collision reduction factor (CRF) can be considered simply as a value representing the percentage of collisions that a safety improvement is expected to eliminate from a location, or a group of locations receiving the same treatment type. Several agencies, such as the Federal Highway Administration (FHWA) and the Institute of Transportation Engineers (ITE), have developed CRFs for different safety improvements. Despite the common usage of these CRFs in practice, there are two major deficiencies in using them to generate project-level safety estimates. Firstly, these CRFs are developed as system-wide factors where the variability of collision patterns and geometric configuration among different Pre-Implementation Evaluation of Safety Improvement Programs Chapter 1: Introduction 3 locations are not considered. Secondly, the use of these CRFs does not account for the treatment of the uncertainty issues related to the effectiveness of the proposed safety improvement. Means to resolve these two problems will be presented in this thesis. 1.3 Thesis Objectives The objective of this thesis is to present a systematic procedure that aims to assist in prioritizing safety improvement projects and ultimately ensure the optimal allocation of funding available for road safety improvements. The followings are the key focal points: 1. Investigate methods to evaluate the effectiveness (CRFs) of various safety improvements on a project-level basis where different collision patterns and site characteristics of a location are considered. 2. Explore techniques to deal with the uncertainty issues involved in determining the effectiveness (CRFs) of various safety improvements. Specifically, the expected values of CRFs and the associated standard deviations will be calculated. 3. Present an economic-based prediction and evaluation procedure to increase the confidence of predictions associated with the current practice of evaluating the expected safety benefits from improvement programs. Specifically, the probability of a proposed safety improvement achieving a specific economic goal (e.g., a pre-determined benefit-cost ratio) prior to its implementation will be calculated. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 1 : Introduction 4 1.4 Methodology To estimate CRFs on a project-level basis, a new safety analysis software called the Information System for Estimating Crash Reductions (ISECR) is developed in this thesis. ISECR is an intelligent database that uses case-based reasoning and consists of past safety research efforts on CRFs associated with the different safety improvement measures. ISECR has the following functions: 1. Data entry: ISECR permits the entry of evaluation studies that report CRFs associated with different safety improvements. 2. Data retrieval and analysis: ISECR employs a case-based reasoning (CBR) approach for data retrieval and analysis. The system is designed to accept queries, analyze and display information from the database that matches the queries. 3. CRFs estimation and reporting: ISECR computes the effectiveness and the associated range and reliability of the proposed countermeasure related to the total number of collisions, and different collision types and severity. 4. Benefit-cost ratio estimation: ISECR provides results that can be used in evaluating the benefit-cost ratio pertaining to the proposed safety improvement. 1.5 Thesis Structure This thesis is divided into nine chapters. Following this introductory chapter, Chapter Two presents the results of a comprehensive literature review on the topics of case-based reasoning and collision analysis. Chapter Three explains how CBR is utilized in the design of the Pre-Implementation Evaluation of Safety Improvement Programs Chapter 1: Introduction 5 intelligent database, ISECR. As well, Chapter Three details the methodology which ISECR uses to estimate CRFs and their uncertainties. Chapter Four describes the various functions provided by ISECR. The validation of SECR is presented in Chapter Five. In Chapters Six and Seven, the uncertainty issues involving the expected reduction in collision frequency and economic analysis are examined respectively. An example illustrating the procedures outlined in this thesis is included in Chapter Eight. Finally, the conclusion and some suggestions for further research are included in Chapter Nine. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 6 2.0 LITERATURE REVIEW 2.1 Introduction This chapter presents the results of a comprehensive literature review on the following two topics: 1. Case-base reasoning (CBR) 2. Collision analysis, with a focus on the generalized linear modeling (GLIM) approach and the Empirical Bayes (EB) technique used to determine the expected number of collisions at a location. The first part of this chapter provides a literature review on CBR, which is utilized in the development of ISECR to predict collision reduction factors (CRFs) and the associated uncertainties. As noted in Chapter One, most of collision analyses utilize the products of CRFs and the expected number of collisions to estimate the potential safety benefits, i.e., the expected collision reduction, of implementing a safely improvement at a location. Thus, the second part of this chapter focuses on describing how the G L I M and EB approaches can be used to estimate the expected number of collisions and the associated uncertainties at a location. 2.2 Case-Based Reasoning Knowledge-based systems (KBS) utilize rule-based and model-based reasoning techniques to build design automation and design decision support systems. Although there was notable success in some areas, difficulties have been encountered with K B S when formalizing Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 7 generalized design experiences such as rules, logic, and domain models. These in turn, have led to the development of case-based reasoning (CBR) systems. C B R systems have proven to be effective in providing desired results in various fields, such as in architectural design, structural design, software specifications design, and etc. C B R is a relatively recent concept, with the original research conducted in the early 1980s as first described conceptually by Schank (1982). It is a methodology used to solve new problems by reusing and adapting (or combining) solutions that worked for similar problems in the past. In CBR, past problem solving experiences (cases) are stored in a database (case base) and upon request, cases similar and/or relevant to the current problem (design case) can be retrieved from the case base. The cases retrieved (retrieved cases) from the case base can then be adapted or combined to better fit the design case. Differentiating from the more traditional rule-based and model-based reasoning techniques, i.e., the former captures knowledge in the form of if-then rules while the latter formulates knowledge in the form of principles and/or models to cover various aspects of a problem domain, CBR is an experience-based method that utilizes prior problem solving experiences as its main knowledge source. C B R is a growing field as it resembles, to some degree, the psychological process at which a person follows when attempting to utilize his/her knowledge and experience in solving a new problem. Using the C B R approach, modeling of domains that are not completely understood and open-ended is possible. Consequently, C B R can be considered as an alternative to paradigms such as rule-based and model-based reasoning. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 8 2.2.1 CBR Applications in Transportation Engineering In the transportation field, there are few applications that employ a C B R approach. For example, Khattak and Kanafani (1996) first developed a planning tool for ITS (Intelligent Transportation Systems) using a C B R approach. As an enhanced system for their first proposed C B R planning tool, Khattak and Renski (1999) then focused their effort on planning H O V (High Occupancy Vehicle) lanes by integrating the C B R approach in a GIS (Geographic Information System) environment. In the field of road safety analysis, only one system was found in the literature that employed a C B R approach. Capus and Tourigny (1998) developed a CBR system (called ROSAC, ROad Safety Analysis with Cases) capable of retrieving similar cases from the case base. Furthermore, ROSAC can reuse, adapt, and save the new problem and adapted solution as a new case in the case base. In ROSAC, each case consists of a problem, as represented by site characteristics and collision statistics, and its solution, as described by collision patterns, collision causes, and the implemented safety improvements. Case retrieval is realized when the site characteristics and collision statistics of the current situation match, to some degree, with the ones stored in the case base. The retrieved cases can then be reused and adapted in ROSAC as the solution to the design case. The following sections provide a brief introduction to some of the issues and procedures involved in a C B R system design, with a focus on the techniques utilized in the development of ISECR. However, for a detailed and comprehensive description on CBR, books and/or reports written by Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 9 authors such as Kolodner (1993), Watson (1997), Leake and Plaza (1997), etc., should be consulted. 2.2.2 General CBR Issues and Procedures Regardless of the domain of application, the followings are the two general issues concerning the application and design of any CBR system: 1. Representation issues: Representation issues concern with the contents (information), representation, indexing, and memory organization of cases. 2. Control issues: Control issues, on the other hand, deal with the general processes of a C B R system, which are comprised of the retrieval, adaptation, and combination of cases. Depending on the context of each individual project, the representation and control issues involved in a C B R system design may vary from one project to another. Nevertheless, the main purpose of any C B R system is to facilitate the solving of a similar problem in a somewhat similar context. Hence, general design issues and procedures involved in all C B R systems can be observed and are further discussed below. 2.2.3 Representation Issues of CBR In any C B R system, the database (case base) contains a representation of a set of previously solved problems (cases). It is extremely important to select an appropriate model for case representation when designing a C B R system. An appropriate model provides the very basis of Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 10 how cases are represented in the CBR case base. Furthermore, the subsequent C B R procedures, including the retrieval of relevant cases and their adaptation or combination, rely heavily on the representation model selected. For that reason, a systematic approach is required to identify and express the uniform representation of specific features that make up a case. Case representation, as mentioned previously, is the first of two issues that needs to be addressed prior to carrying out any CBR system design. In general, issues related to developing an adequate and useful case base can be addressed by answering the following four fundamental questions: 1. What are the contents, or features, that are essential to represent a case? 2. How to express or represent these features (as in numbers, symbols, Boolean variables, texts, models, etc.)? 3. How to structure these features effectively and efficiently to minimize the storage and computation requirements? 4. What case indexing schemes should be used for the retrieval of cases? Essentially, case representation is concerned with how the contents of each case are represented and organized in the case base. There are a number of ways to organize the information in a case, however, only two of these alternatives are relevant to this research and discussed further below, i.e., using a set of attribute-value pairs or part-subpart relationships. These two alternatives are illustrated in Figure 2.1. When representing a case as attribute-value pairs, all features and associated values are represented with these sets of values. Information gathering is made simple in this manner as all Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review aspects of a case are kept in one case. This type of case representation is typically adequate for small and simple cases. When representing a case as a hierarchy of part-subpart relationships, information within a case is broken down into a number of subcases. Therefore, a higher degree of information is required to adequately represent the subcases, and to designate the relationship knowledge within a case. Despite the higher degree of complexity with this type of case representation when compared to the attribute-value pairs, using a part-subpart hierarchy is beneficial as it facilitates the representation of large and complex cases. The representation and organization of a case should be consistent for all cases in the case base. In this regard, all cases are described by the same set of features, represented either by attribute-value pairs and/or part-subpart relationships. An adequate representation scheme is the first step to ensure a successful C B R system design. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 12 Case A attribute 1: value 1 attribute 2: value 2 attribute 3: value 3 Case A attribute 1: value 1 attribute 2: value 2 Subcase A l attribute 3: value 3 attribute 4: value 4 Subcase A2 attribute 5: value 5 attribute 6: value 6 Subcase A3 attribute 7: value 7 attribute 8: value 8 Subcase A4 attribute 9: value 9 attribute 10: value 10 (a) Attribute-value pairs (b) Part-subpart relationships Figure 2.1. Representation of a Case in Case Base 2.2.4 Control Issues of CBR Once the above case representation issues are addressed, the following three general C B R processes, as often referred to as the control issues, can then be carried out sequentially: 1. Case retrieval, 2. Case adaptation, and 3. Case combination. Case retrieval refers to the capability of a CBR system to accept queries and filters information from the case base which matches the queries. It is uncommon that the features in a design case Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 13 match completely with the ones in the retrieved cases. The simplest approach would be to retrieve and reuse the unchanged solution of the most similar case as the solution to the design case. While this simplistic approach reduces the computation and programming requirements needed to design a C B R system, it fails to account for the differences between the design and retrieved cases. Since no two problems are ever the same, it is necessary for any C B R system to have at least one of the adaptation or combination capability (Pu, 1998). Solution to the design case can then be constructed either by adapting or combining old solutions. In this manner, use of a C B R system is more advantageous as the adaptation or combination of previously solved solutions is easier than generating a new solution from scratch. 2.2.4.1 Case Retrieval Case retrieval is a basic operation in any CBR design, but it plays a significant role in the establishment of a C B R system. Retrieval of cases can be done informally or formally. Informal case retrieval refers to the selecting of relevant cases from the case base by the user based on his/her experience and/or judgement. On the other hand, formal case retrieval refers to a C B R design system capable of accepting a new set of definition and/or features from the user and conducting searches on its case base for cases that have the same or similar problem specifications. ISECR utilizes the latter of the two retrieving techniques. Retrieval of cases can be based on a perfect match, where the specifications of the design case are found to match exactly. However, it is more likely to find the features in the design case that do not match completely with the ones in the retrieved cases. Thus, most of the retrieved cases from a C B R system are partial matches. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 14 The degree of partial match depends on the relative similarity and importance of the features that make up a case. The degree of similarity of each feature can be calculated by comparing the specific feature in the design case to the corresponding feature of retrieved cases. Case distance can be computed and used as a measure to determine the degree of similarity between the design and retrieved cases. One of the most frequently used case distances is the Euclidean distance calculated as (Yeh, 1997): 2 (Equation 2.1) Z w, i=\\ • where Dk Case distance of the design case to the k\"1 case n Number of input features Importance factor for feature i ABS Absolute value function fl R Maximum value for feature i Minimum value for feature i fl Value of feature i in the design case fR Ji,k Value for feature i in the k' case Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 15 In Equation 2.1, the importance factor, wi can be represented by a numerical value ranging from 0. 0 to 1.0. A higher importance value (closer to 1.0) indicates that the feature is more important compared to the features having lower importance values. The value of case distance, Dk, as determined by Equation 2.1, can also be any real number varying between 0.0 to 1.0. The more similar the design case to the retrieved case, the smaller the Dk value or shorter the case distance. Case distance can be used to rank cases, i.e., lower scoring cases are used before higher scoring ones. Inherently, cases are ranked based on a weighted sum of features in the design case that match the ones in the retrieved cases. 2.2.4.2 Case Adaptation The main purpose of case adaptation is to modify the solution of the case retrieved from the case base to account for the differences between the design and retrieved cases. There are a number of adaptation methods available, i.e., reinstantiation, parameter adjustment, local search, case-based substitution, commonsense transformation, model-guided repair, special-purpose adaptation and repair, derivational reply, etc. In general, adaptation methods in C B R are classified into the following two categories: 1. Structural adaptation: The retrieved case solution is substituted directly with adaptation rules and/or formulas to generate a new solution for the design case. 2. Derivational reply: The same methods at which the original case solution (from the retrieved case) is derived are reused to derive a new solution for the design case. Specifically, it is assumed that the same rules and/or formulas used to generate the retrieved case solution can be reused to produce the new solution to the design case. With this method, the sequence at Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 16 which the retrieved case is solved must be stored in the case base as an additional attribute to the case. Although adaptation of cases is useful in many applications, many successful C B R systems do not perform adaptation at all or leave this option to the user. Watson (1997) suggested that adaptation should be avoided unless it can be carried out easily using simple adaptation methods such as reinstantiation or parameter adjustment. Complex adaptation of cases is knowledge intensive and can only be applied to domains that are well understood. Since C B R is generally employed to solve problems that are not well understood, a complex adaptation method that is knowledge intensive may not be feasible. 2.2.4.3 Case Combination Case combination aims to derive the solution for the design case by combining several of the original solutions stored in the retrieved cases. Frequently, the solution stored in the most similar case may not be the best solution for the design case. Thus, it becomes necessary to retrieve several similar cases and combine their solutions in order to generate an improved solution to the current problem (Yeh, 1997). Case combination can be carried out in several ways, such as the weighted average approach, constructive approach, or frame approach (Pu, 1993, and Yeh, 1997). However, for the purpose of this research, the weighted average approach is adequate to the application and it is shown below: Tweighted = , 7~ (Equation 2.2) Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 17 where Tu Weighted = Weighted solution for the design case Number of retrieved cases used to generate the weighted solution Case distance of the design case to the k\"1 case, as determined by Equation 2.1 Tk = Solution for the k\"1 case The term (1 — Dk) in Equation 2.2 suggests that solutions obtained from cases with lower case distances, i.e., more similar to the design case, are weighted more significantly compared to the ones with higher case distances. The accuracy of the above weighted solution can be represented by the standard deviation that has the following expression: where Stdev(TWeighted) = Standard deviation of the weighted solution for the design case 2.2.5 CBR Methodologies There are several C B R methodologies used in practice, each employing different combination of the three C B R processes described in Sections 2.2.4.1 to 2.2.4.3. In this research, two of the C B R methodologies are utilized in the design of ISECR and thus, discussed further below. Pre-Implementation Evaluation of Safety Improvement Programs (Equation 2.3) Chapter 2: Literature Review 18 2.2.5.1 Nearest Neighbour Approach The first C B R methodology is called the nearest neighbour approach and is illustrated in Figure 2.2. This is the simplest of all C B R methodologies where the only C B R process used is case retrieval. With this approach, only the most similar case from the case base is retrieved and the result stored in this retrieved case is then utilized as the solution to the design problem. 2.2.5.2 Collaborative Approach Moving to a more sophisticated level, the second C B R methodology is called the collaborative approach as shown in Figure 2.3. Collaborative C B R approach retrieves several cases from the case base and performs case combination to the retrieved cases. With this technique, A>nearest neighbour retrieval is used, where k refers to a predetermined number of cases to be retrieved. The retrieved case solutions are then combined (using Equation 2.2) to yield the solution for the design case. The accuracy of the solution generated for the design case can be estimated by calculating its standard deviation by using Equation 2.3. Essentially, collaborative C B R approach utilizes both case retrieval and case combination to derive a better-fit solution to the design case. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 19 Figure 2.2. Nearest Neighbour A p p r o a c h Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 20 Combination Figure 2.3. Collaborative A p p r o a c h Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 21 2.3 Collision Analysis The need for using the generalized linear regression modeling (GLIM) and the Empirical Bayes (EB) approaches arises as they address and overcome the problems associated with the conventional methods of predicting site-specific collision estimates. The following sections review some of the published research to construct a fundamental understanding on the above two techniques. 2.3.1 The Generalized Linear Modeling (GLIM) Approach Most of the earlier work in collision analysis utilized the conventional linear regression approach to develop prediction models relating collisions to traffic volumes with the assumption of a Gaussian (normal) distributed error structure. However, several researchers (Jovanis and Chang, 1986, Hauer et al., 1988, Saccomanno and Buyco 1988, Miaou and Lum 1993) have all shown that conventional linear regression prediction models lack the distributional property to adequately describe the random, discrete, non-negative, and typically sporadic characteristics of traffic collisions. To overcome these problems, Jovanis and Chang (1986) concluded, with supporting results from their modeling of collisions at highway sections in Indiana, that a Poisson distribution should be used to describe the model error structure. Furthermore, Miaou and Lum (1993) also supported the use of a Poisson distributed error structure with their results. A l l in all, these researchers confirmed the shortcomings associated with the conventional linear regression models and their capabilities in predicting collision estimates. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 22 The G L I M modeling approach, on the other hand, can overcome the shortcomings associated with the conventional methods. Specifically, utilizing the G L I M approach, the flexibility of assuming different error structures is provided and the conversion of non-linear models into linear ones is also feasible. As an example, the followings describe how Hauer et al. (1988) and Kulmala (1995) utilized the G L I M approach to develop collision prediction models for intersections. Assuming that Y is a random variable that describes the number of collisions in a specific time period, y is the observation of Y during a period of time, and the random variable A is regarded as the mean of Y. Thus, for A = X, Y is Poisson distributed with parameter X . If each site has its own regional characteristics with a unique mean collision frequency A , Hauer et al. (1988) have shown that for an imaginary group of sites having similar characteristics, A follows a gamma distribution. The gamma distribution, having parameters K and K//U , has a mean and variance that can be described with the following two equations: E(A) = /u Var{h) = — K (Equation 2.4) (Equation 2.5) Based on Equations 2.4 and 2.5, the point probability function of Y can be given by the negative binomial distribution with an expected value and variance described by Equations 2.6 and 2.7 respectively as shown below (Hauer et al. 1988, and Kulmala, 1995): Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 23 E(Y)=M (Equation 2.6) Var(Y)=Li + (Equation 2.7) K As shown above, unless K —> o o , the variance of the observed number of collisions is greater than its expected value. However, when K —> c o , the variance equals the expected value, which is identical to the Poisson distribution (Kulmala, 1995) and suitable to describe the nature of collisions. 2.3.2 The Empirical Bayes (EB) Approach The EB approach is used to refine the G L I M estimate of the expected number of collisions at a location to yield a more accurate, location-specific safety estimate. Two types of clues of the location are used in the EB approach: its traffic and road characteristics, and its historical collision data (Hauer, 1992, Brude and Larsson, 1988). Utilizing the EB approach, Hauer (1992) calculated the expected number of collisions for any intersection using the following equations: (Equation 2.8) where a 1 (Equation 2.9) 1 + Var(E(A)) E(A) count = observed number of collisions Pre-Implementation Evaluation of Safety Improvement Programs Chapter 2: Literature Review 24 it (A) = predicted number of collisions estimated from the GLIM model = variance of the GLIM estimate Since Var(E(h)) = , Equation 2.8 can be rearranged as: f T 7 ( A \\ \\ EB Safety. Estimate E(A) Krc + E(A); (K + count) (Equation 2.10) Finally, the variance of the EB estimate can be determined using Equation 2.11 as: Var^EB Safety Estimate) — f E(A) ^ K + E(A) • (K + count) (Equation 2.11) Hence, the safety estimate, i.e., the expected number of collisions and the associated uncertainties, can be evaluated by using Equations 2.10 and 2.11 respectively. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 25 3.0 C B R A P P L I C A T I O N I N I S E C R 3.1 Introduction The computerized approach of the Information System for Estimating Crash Reductions (ISECR) is designed to minimize the amount of manual work required in evaluating the effectiveness of different safety improvements on a project-level basis. ISECR consists of historical information extracted from past evaluation studies that reported the performance of different safety improvements. Utilizing a CBR approach, ISECR is capable of retrieving and analyzing appropriate past records in order to assess the range and reliability of the predicted countermeasure effectiveness (CRFs) for a given safety improvement under certain condition. This chapter provides an overview of the implementation of CBR in ISECR by focusing on the followings: 1. Case representation: Determine the contents/features used to represent a case, and their organization and structure in the ISECR case base. 2. Case retrieval: Establish the query parameters used to retrieve relevant cases. 3. Case distance: Determine the criteria used to calculate the case distance for each of the retrieved cases. 4. Solution construction: Establish the adaptation and/or combination strategies employed to modify and/or combine the retrieved solutions to create an improved solution for the design case. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 26 3.2 Case Representation in ISECR Frequently, more than one CRF can be found in an evaluation study, i.e., some studies reported the effectiveness for more than one countermeasure while others evaluated the same countermeasure under different situations, etc. Hence, it is possible to represent an evaluation study with more than one case in the ISECR case base. The information contained within each case is organized using part-subpart relationships, as described in Section 2.2.3. The contents within each case are broken down into the following six subcases: 1. General information, 2. Case quality, 3. Countermeasure type, 4. Location type, 5. Location characteristics, and 6. Case solutions. Within each subcase, attribute-value pairs are used to represent the features and associated values. The representation of each of the above six subcases is organized as follows and shown in Figure 3.1: Subcase 1- General Information: • Case Id: reference point for a given case in the case base • General information on the evaluation study of which the case belongs to: Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 27 - Study Id: an identifier used to reference the evaluation study - Author(s) of the study - Study title Source of the study: journal, volume, pages, publication date, and country Subcase 2- Case Quality: • This subcase stores information regarding the treatment of the following three confounding factors: - Changes in traffic volume - Inclusion of unrelated effects Regression to the mean (RTM) artifact Subcase 3- Countermeasure Type: • A total of 116 countermeasure types are considered in this subcase (see Appendix A for details). They are organized and grouped into the following sixteen countermeasure categories: area-wide schemes, bridge improvements, cyclist/pedestrian facilities, delineation, geometric improvements, intersection improvements, lane/shoulder treatment, lighting improvements, object removal/relocation, parking improvements, pavement treatment, railway improvements, regulation change, safety barriers, traffic controls/signs, and traffic signals. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 28 Subcase 4- Location Type: • This subcase gathers the information on the type of location where the countermeasure effectiveness is investigated. A total of eight location types are considered in this research, i.e., general intersections, signalized intersections, unsignalized intersections, road sections, freeways, bridges, rails, and construction zones. Subcase 5- Location Characteristics: • The location characteristics associated with a given case are extracted from the evaluation study and stored in this subcase. Different location characteristics are considered for each of the eight location types (see Appendix B for details). For example, the following characteristics are considered for signalized intersections (see Table 3.1): 1. Area type: urban, suburban, rural, or other 2. Intersection type: four-legged, t-intersection, or y-intersection 3. Implementation level: isolated location or wide area 4. Total traffic volume: Total entering traffic volume (AADT) recorded for the intersection. The intervals are as follows: 0-4999, 5000-9999, 10000-14999, 15000-19999, 20000-29999, 30000-39999, 40000-49999, 50000-59999, 60000-69999, 70000-79999, and 80000 and more. 5. Average lane width: Average width of all traffici lanes. The distinction is made between less than 12 feet and greater or equal to 12 feet. 6. Provision of left-turn channelization: yes or no Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 29 7. Provision of right-turn channelization: yes or no 8. Left-turn movement: not allowed, permissive, or protected 9. Right-turn movement: not allowed, permissive, or protected 10. Street parking: allowed or not allowed 11. Type of traffic control: fixed-timed, semi-actuated, or fully-actuated 12. Average number of lanes per approach: less than or equal to two lanes or more than two lanes • The above location characteristics are represented in the case base as different sets of attribute-value pairs. For example, Table 3.1 presents a set of attribute-value pairs to represent location characteristics for signalized intersections (see Appendix B for the representation of the location characteristics used for other location types). Representing these features in such a manner does not only minimize the storage requirement, but also facilitates the calculation of case distance, as will be shown later. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 30 Table 3.1. Representation of Location Characteristics for Signalized Intersections Characteristic Feature Value Area type l=urban 2=suburban 3=rural 4=other Intersection type l=four-legged 2=t-intersection 3=y-intersection Implementation level l=isolated location 2=wide area Total entering traffic volume 1=0-4999 (AADT) 2=5000-9999 3=10000-14999 4=15000-19999 5=20000-29999 6=30000-39999 7=40000-49999 8=50000-59999 9=60000-69999 10=70000-79999 11=80000 and more Average lane width l=less than 12 ft 2=greater or equal to 12 ft Provision of left-turn l=yes channelization 2=no Provision of right-turn l=yes channelization 2=no Left-turn movement l=not allowed 2=permissive 3=protected Right-turn movement l=not allowed 2=permissive 3=protected Street parking l=allowed 2=not allowed Type of traffic control l=fixed-timed 2=semi-actuated 3=fully-actuated Average number of lanes per l=less than or equal to 2 lanes approach 2=more than 2 lanes Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 31 Subcase 6- Case Solutions: • The reported collision reduction factors (CRFs) are entered into this subcase as the case solutions. CRFs corresponding to the following collision severity and types are entered into the case base (if available): - Total collisions - Collision severity: fatal, injury, casualty, and property-damage-only. Collision types: angle, bike/pedestrian, fixed-object, head-on, left-turn, off-road, overtaking, parked vehicle, rear-end, right-turn, sideswipe, and other. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 32 Case A •General Information •Case Quality •Countermeasure Type •Location Type •Location Characteristics •Case Solutions Figure 3.1. Representation of a Case in ISECR With the above information entered into the case base, the ISECR user can then query for cases that report the effectiveness of a specific countermeasure at a specific location. Case retrieval strategies used in ISECR are further explained in the next section. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 33 3.3 Case Retrieval in ISECR Case retrieval in any C B R system requires the user to input a set of definitions and/or features, which can be used to uniquely describe the current problem (design case). These specifications become the query parameters for filtering cases from the case base. When using ISECR, the user is first asked to specify both the type and location of the proposed countermeasure by selecting from the same lists used in Subcases 3 and 4 respectively (see Section 3.2). These two features are then used as the query parameters and are matched against the associated attribute-value pairs contained in each case stored in the case base. A case is only retrieved from the case base i f a perfect match of the two query parameters is found. It is likely that more than one case with varying CRFs will match the above query criteria and will be retrieved from the case base. In this instance, it is essential to determine which of the retrieved cases and the associated results are more valid and/or relevant to the design case. As mentioned previously, a credibility factor and an application factor, or referred to in this thesis as the quality and relevance scores respectively, can be evaluated for each case to assess its validity and relevance. The following sections discuss the quality and relevance scores in greater depth, the methods in which they are derived, followed by an explanation on how solution construction is implemented in ISECR. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 34 3.4 Quality Score Quality score is a quantitative measure used to determine the validity of a case and its results. In this thesis, the quality of a case is influenced by the treatment of the three common confounding factors, as considered in Subcase 2 (see Section 3.2) of each case. Confounding factors are factors that may affect the accuracy of the evaluation on the effectiveness of a safety improvement i f they are not accounted for in the retrieved case. The lack of treatment of these factors, in turn, can threaten the validity of cases and their results. Ultimately, by treating these confounding factors, one can decide i f the observed changes in road safety are caused by the implemented countermeasure, the existence of the confounding factors, or a combination of both. Some researchers have attempted to account for the presence of confounding factors in evaluation studies. One approach is to assign arbitrary weights to evaluation studies based on their treatment of these factors. For example, Elvik has assessed the quality of evaluation studies based on several confounding variables (Elvik, 1995, 1996, and 1998). These variables include research design, decade of study, change in traffic volume, regression to the mean, collision migration, etc. In this thesis, the following three confounding factors are considered: 1. Changes in traffic volume, 2. Inclusion of unrelated effects, and 3. Regression to the mean (RTM) artifact. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 35 3.4.1 Changes in Traffic Volume Changes in traffic volume are usually controlled by calculating collision rates (Elvik, 1996). Collision rates can be expressed as collisions per million-entering-vehicles (col/mev) for intersections and collisions per million-vehicle-kilometers (col/mvk) for road sections. Changes in collision rate are often used as a measure of the effectiveness of a countermeasure. By using collision rates, the number of collision is assumed to relate linearly with traffic volumes. However, this assumption is not always valid. In fact, a non-linear relationship between collision frequency and traffic volumes has been shown to be a more suitable assumption (Hauer et al., 1988). This indicates that the use of collision rates does not necessarily, under all circumstances, minimize the effect of traffic volumes on collision frequency. Nevertheless, in this thesis, similar to Elvik's approach (1996), cases that employ changes in collision rate as the measure of countermeasure effect are classified as having accounted for the confounding factor on the changes in traffic volume. 3.4.2 Inclusion of Unrelated Effects Frequently, factors other than the treatment may affect the observed difference in collision frequency. The inclusion of unrelated effects in a case may lead one to believe that the implemented countermeasure is more effective than it really is. The presence of unrelated effects in a case can be controlled by using a comparison group when determining the CRF. Inherently, a comparison group is a group of sites that are somewhat similar to the treatment site. This group often consists of the total number of collisions in the area where the treatment area is located. The change in collision frequency of the comparison group can be compared to the one observed Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 36 for the treatment site. This permits the calculation of the actual treatment effects. Therefore, in this research, i f a case includes a comparison group when estimating the countermeasure effectiveness, it is classified as having treated for the confounding factor of including the unrelated effects. 3.4.3 Regression to the Mean (RIM) Artifact Sites are usually selected for treatment due to their high collision occurrence. However, this high occurrence may be entirely caused by a random up-fluctuation of collision around the location's true mean (collision occurrence) value. If a location is selected for treatment solely because it undergoes an up-fluctuation in collision frequency, it will show a reduction in collision occurrence in the after period regardless of the implementation of the countermeasure. Hence, i f the R T M bias is not accounted for in a case, an overestimation of the effectiveness of a countermeasure can take place. R T M artifact can be controlled with the use of a reference group and/or an appropriate analysis technique. A reference group should be selected to represent the treatment site, i.e., from the same potential treatment population. As for the analysis methods used to address the R T M bias, the Empirical Bayes technique is often used in practice. Thus, i f a case uses one of these two methods to remove the R T M effects, it is considered to have accounted for this confounding factor. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 37 3.4.4 Calculation of Quality Scores In ISECR, quality scores are calculated automatically using the information contained in Subcase 2, i.e., case quality, and hence, not requiring any input from the user. Quality score, employing the Euclidean distance, can be computed using Equation 3.1 as shown below: I 2 (Equation 3-1) where Dk(QUality) = Quality score of the k\"1 retrieved case n = Number of input features Wj = Importance factor for feature i ABS = Absolute value function fit\\ax = Maximum value for feature i fUmm = Minimum value for feature i fl = Value of feature i in the design case flk = Value for feature / in the k\"1 retrieved case In Equation 3.1, values of 3.0 and 0.0 are assigned to fRmm and fRmin as the maximum and minimum numbers of confounding factors that can be accounted for in a case respectively. A Y \" wrABS D fi fi,k fR - fR . k{Quality) Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 38 feature value, ftRk, is assigned to each case based on its treatment of the three confounding factors. For example, i f two of the three confounding factors are accounted for in a case, a feature value of 2.0 is assigned. Similarly, i f a case has accounted for none, one, or all of the confounding factors, a feature value of 0.0, 1.0, or 3.0 is given respectively. In Equation 3.1, a value of 1.0 is assigned to wt as the only feature considered at hand is the treatment of the three confounding factors. Lastly, a value of 3.0 is assigned to ff as the optimum number of confounding factors that can be accounted for in any case. Below is a sample calculation of the quality score (using Equation 3.1) for a case having accounted for two of the three confounding factors: D k(Quality) Y\" w, • ABS fi fi,k fR _ fR J i,max J /,min \\.0-ABS\\ 1=1 1 V 1.0 3 -2 3^0 \\ 2 = 0.33 In the above calculation, a quality score of 0.33 is calculated for the k\"' retrieved case as it has accounted for two of the three confounding factors. Alternatively, i f none, one, or all of the three confounding factors are treated, a quality score of 1.0, 0.66, or 0.0 can be computed respectively. Clearly, the smaller the quality score, the more valid the case and its corresponding solutions. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 39 3.5 Relevance Score Not all cases retrieved from the ISECR case base, based on querying for the same countermeasure and location types, have the same location characteristics as the design case. For example, i f the design case deals with an urban signalized intersection and a total entering traffic volume of 10000 A A D T , not all of the retrieved cases have the same site features as the design case. Hence, it would be reasonable to assign more weight to a case and its results i f it has more similar site features compared to the design case. Relevance score determined for each case is influenced by the degree of similarity in the location characteristics between the design and retrieved cases, as provided by the ISECR user and those already stored in the case base respectively. The location characteristics considered, as mentioned previously, are different for each of the eight location types (see Appendix B for details). 3.5.1 Calculation of Relevance Scores In this thesis, relevance score for each retrieved case is determined by utilizing the Equation 3.1. To calculate the relevance score using Equation 3.1, the values of fRnax and fiRmin assigned to each feature of a case are based on the specific feature encountered. For example, suppose that the encountered feature is traffic volume, values of 11.0 and 1.0 are assigned to fRmM and fRmin to represent the maximum traffic volume of 80000 and more A A D T and the minimum traffic volume of 0-4999 A A D T respectively. As for the values of f/ and fRk, they are based on the Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 40 specific features of the design and retrieved cases respectively (see Appendix B for details). Each feature is weighted equally in Equation 3.1, i.e., the value of w: is dependent on the number of features considered. For instance, a value of 0.5 is assigned i f two features are considered. Similarly, a value of 0.33, 0.25, or 0.20 is assigned to w. i f three, four, or five, features are considered respectively. To demonstrate how Equation 3.1 can be used to determine the relevance score for a retrieved case, consider the example below where the design case deals with an urban signalized intersection with a total entering traffic volume of 10000 A A D T , while the retrieved case concerns with a suburban signalized intersection with a total entering traffic volume of 5000 A A D T . These features are represented in ISECR as follows: Table 3.2. Example: Representation of the Location Characteristics Feature Values Maximum Minimum Design Case Retrieved Case Area type 4 1 1 (urban) 2 (suburban) Total Entering traffic 11 1 3 (10000- 2 (5000-9999) volume (AADT) 14999) Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 41 The relevance score for the retrieved case can be determined by using Equation 3.1 as: D A: (Re levance) Y\" w. • ABS fi fi,k fR - f! J /,max J i, R min J XV W; 0.5 • ABS, r i - 2 V . . . _ J 3 - 2 ^ 2 4 - 1 + 0.5 • ABS 11-1 0.5 + 0.5 Dk(Relevance) ~ 0.246 A lower relevance score signifies that the location characteristics are more similar between the design and retrieved cases, and vice versa. 3.6 Calculating Case Distance in ISECR Case distance is evaluated to determine the relative importance of a case compared to others. When computing case distance, Dk, for each of the retrieved cases, the ISECR user can decide to include either the quality score alone or a combination of the quality and relevance scores. If the user decides to exclude the relevance score and only uses the quality score, Dk is equivalent to the quality score computed for the retrieved case. However, i f the user decides to encompass both the quality and relevance scores in determining Dk , both scores are weighted equally as in the expression below: — _ '{Pk(Quality) + At(Re/erance)) (Equation 3-2) Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 42 where Dk = Case distance of the k\"1 retrieved case For example, with the quality and relevance scores (0.33 and 0.246) computed for the two examples in Sections 3.4.4 and 3.5.1 respectively, case distance for the retrieved case can be determined in the following two ways: Dk ~ ^k(Quality) ~ Dk = \\ \\ D k { Q u a l i t y ) + D m e l e v a n c e ) ) = ± . (0.33 + 0.246) = 0.288 The first result indicates that the ISECR user has decided to exclude the relevance score in determining the case distance for the retrieved case. If this is the case, case distance is exactly the same as the quality score determined, i.e., in this example, 0.33. Conversely, the second result indicates that both the quality and relevance scores are utilized in determining Dk. In this instance, the two scores are weighted equally and determined to be 0.288. Hence, a lower case distance is assigned to a case if it has treated more confounding factors and/or if its location characteristics are more similar to the design case. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 43 3.7 Solution Construction in ISECR In ISECR, solutions to the design problem can be constructed utilizing either the nearest neighbour or collaborative approaches (see Section 2.2.5 for details). 3.7.1 Nearest Neighbour Approach The nearest neighbour approach utilizes the most relevant case (lowest case distance) and its result is used as the solution to the design case. The ISECR user is presented with a list of all matched cases ranked in an ascending order according to the calculated case distances. The user can then select the most appropriate case and adapt the case solutions (CRFs), i f necessary, to account for the differences between the design and retrieved cases. 3.7.2 Collaborative Approach The collaborative C B R approach, on the other hand, utilizes more than one retrieved case, combines their results, and then employs the combined result as the solution to the current problem. With this approach, the A:-nearest neighbour retrieval is used, where k is the number of cases to be retrieved and utilized in solution combination. There is little evidence in the literature indicating what the optimum k value should be to produce the best results. Not only that, the k values found in the literature vary greatly from one application to the next. For example, Gonzalez and Laureano-Ortiz (1992) combined outcomes from the three most relevant cases (if available) while Yeh (1997) utilized the results stored in all of the retrieved cases. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 44 The approach employed in ISECR uses all of the retrieved cases to derive the new solution (the effectiveness of the proposed countermeasure and its uncertainty). Although this approach may include results from cases that have lower case distances, it should be realized that utilizing solutions stored in the most similar cases may not necessarily provide the best solution to the new problem (Yeh, 1997). Furthermore, this report intends to provide both the non-weighted average and weighted average solutions (CRFs) based on the results obtained from all of the retrieved cases. The non-weighted average is simply the arithmetic mean of the retrieved CRFs, while the weighted result is determined by weighting the retrieved CRFs with their case distances. Equations 3.3 and 3.4 are used in ISECR to determine the non-weighted and weighted CRFs respectively: CRF, (Equation 3.3) Non-Weighted N CRF, Weighted (Equation 3.4) where CRF, Non-Weighted = Non-weighted solution (CRF) CRFV Weighted = Weighted solution (CRF) = Number of retrieved cases = Case distance of the k\"1 retrieved case CRFK = CRF of the if* retrieved case Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 45 The accuracy of the above results can be estimated with the standard deviations as: Y^JCRF.-CRF)2 °'CRF-Non-weighted = ]j —~ (Equation 3.5) °CRF'-Weighted Zk^-Dk)(CRFk-CRFf (Equation 3.6) z ; = 1 d - ^ ) - ( / v - i ) where o~CRF_Non_Waighted = Standard deviation of the non-weighted solution 0 CRF-weighted = Standard deviation of the weighted solution The provision of the weighed results attempts to account for either the quality or both the quality and relevance of a case, i.e., solutions obtained from cases with lower case distances are weighted more and vice versa. This is demonstrated in the following example as shown in Table 3.3: Table 3.3. Example: Determining Non-Weighted and Weighted CRFs Ranking of the ri-trk-w-d case ( aso Distance, (1-DA) Case Result. C/?f*(\"..) Weighted Case Ui-siill (\"..) 1 0.25 0.75 23 r.25 2 0.5 0.5 25 12.5 3 0.75 0.25 33 8.25 4 1.0 0.0 40 0 Sum 2.5 1.5 121.0 38.0 Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 46 The above table demonstrates that cases with lower case distances are weighted considerably higher than those with higher case distances. The non-weighted and weighted CRFs can be determined by using Equations 3.3 and 3.4 respectively, as shown below: CRF 121-0 _ L R f Non-Weighted ~ ~ ~ ~ 30.25% J^lfi-D^-CRF, = 38^0 CRFWEIGHTED = , , = — = 25 .33% Hence, the non-weighted and weighted CRFs are 30.25% and 25.33% respectively. In this example, the weighted CRF is less than the non-weighted CRF. This is expected as the weighted CRFs reflect results that have been accounted for some or all of the three confounding factors, as mentioned previously. In some instances, ISECR provides weighted CRFs that are slightly larger than the non-weighted results. This arises mainly due to the insufficient data available and/or the small number of evaluation studies available for the specific safety improvement implemented at the specific location. Continuing with the example shown in Table 3.3, the standard deviations for the non-weighted and weighted CRFs can be estimated with Equations 3.5 and 3.6 respectively as: ZljCRFK-CRF)2 IT8Z75 CRF-Non-Weighted ~ y A^ —1 _ ~ Pre-Implementation Evaluation of Safety Improvement Programs Chapter 3: CBR Application in ISECR 47 <7, CRF-Weighted i (1 - Dk) • (CRFk - CRF)2 (1^83 4.5 = 2.05% Thus, the standard deviations for the non-weighted and weighted CRFs are 7.81% and 2.05% respectively. Clearly, the non-weighted standard deviation is significantly greater when compared to the weighted standard deviation. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 48 4.0 DESCRIPTION OF ISECR 4.1 Main Menu ISECR is designed as an user-friendly intelligent database that can facilitate the daily use of practitioners in the highway safety engineering industry. A prototype of ISECR is currently implemented in a personal computer using Microsoft Access 97 in a Microsoft Windows 95/98 environment. Currently, the database consists of 450 documents. In addition to predicting the effectiveness of safety improvements and their reliability, ISECR is designed to accept queries, filter and display information based on the users' specifications, and permit the entry of new documents. This chapter provides a description of the features available in ISECR. Specifically, the first four of the following five ISECR Main Menu options are discussed (see Figure 4.1): 1. View all documents, 2. Search documents, 3. Predict CR factors, 4. Input new documents, and 5. Exit (ISECR). Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 49 Information System for Estimating Crash Reductions Figure 4.1. ISECR Window: Main Menu 4.2 View All Documents With the View All Documents option, the user can list all evaluation studies stored in the ISECR database. As illustrated in Figure 4.2, ISECR allows the user to list and sort the available studies based on: 1. Author name (including all authors and co-authors), 2. Study source, Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 5 0 3. Study title, or 4. Publication date. View All Documents List / Sort Documents by: Figure 4.2. ISECR Window: View All Documents Once the sorting option is selected, ISECR then presents a summary of the documents including the following information (see Figure 4.3): 1. Authors'names, 2. Study title, 3. Study source, 4. Publication date, 5. Availability of the one page summary for the document, and 6. Availability of the document. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 51 E H Return Print Help '@ Lane / Shoulder treatment. @ Lighting improvements # 0 b j e ct re tn oval / re I o cati o n © Parking improvements # Pavement treatment @ Rai I way i rn p rove rn e nts @ Regulation change # Safety barriers @ Traffic controls / signs % Traffic signals I i Figure 4.6. ISECR Window: Countermeasure Categories By clicking on any one of the sixteen categories, another window consists of different countermeasures (belonging to the selected countermeasure category) is opened. For example, Figure 4.7 is a result of selecting Delineation as the countermeasure category. Appendix C provides the details when other countermeasure categories are selected. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 56 Predict CR Factors (cont.) Select a Countermeasure: Figure 4.7. ISECR Window: Countermeasure Types for Delineation Once the countermeasure type is selected, the user is required to click on the Continue button before the location type window can be opened. This window allows the user to specify where the proposed countermeasure is to be implemented. The location type window, as shown in Figure 4.8, consists of eight different location types. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 57 Predict CR Factors (cont.) Select a Location Type: Intersections # General Intersections # Signalised Intersections ; 1 # Unsignalized Intersections •-. -Road Sections and Freeways & IRoad Sections i # Freeways Others # Bridges •S Rai l Crossings # Construction Zones Figure 4.8. I S E C R Window: Location Types After the location type is specified, ISECR searches its case base for cases dealing with the same countermeasure and location types. A case is only retrieved if a perfect match of the two query parameters is realized. If no matched case is found with the specified query parameters, the user will be notified, as illustrated in Figure 4.9: Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 58 No match found. Click YES to continue search Click MO to quit. Yei No Figure 4.9. ISECR Window: No Match Found On the other hand, i f cases are found and retrieved, the user will then be prompted with a dialog box to enter the location characteristics which would be used later in the calculation of case distances. This is shown in Figure 4.10. Once the Yes button is clicked, the user is presented with an input form, as shown in Figure 4.11 for signalized intersections, to enter location characteristics (see Appendix D for the input forms used for other location types). In ISECR, the implementation level for the design case is assumed to be at an isolated location. This assumption is made i f the user decides to use location characteristics in the analysis. Once the user finishes entering location characteristics and clicks on the Continue button, he/she is warned by ISECR to use quality criterion only i f few location characteristics were entered. As shown in Figure 4.12, the user can then decide i f case distances will be calculated based on the quality scores alone or based on both the quality and relevance scores. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 59 ISECR ef) Do you want (o input site characteristics and use then for analy : i ; ? Yes No Cancel Figure 4.10. ISECR Window: Selecting to Enter Location Characteristics Predict CR Factors (cont.) Select Location Characteristic^): Are a Type Intersection Type Type of Traffic Control Total Traffi c Vo I u rn e (AADT) Ave. Number of Lanes per Approach Average Lane Width Left-turn Channelisation Right-turn Channelisation Left-turn Movement Right-turn Movement Street Parking urban Jra mm Figure 4.11. ISECR Window: Input Location Characteristics for Signalized Intersections Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 60 3 It is advised to use quality criterion only if few location characteristics were entered in the location type window. Click YES to weight results ba;ed on both quality and relevance criteria Click NO to weight results based on qualify criterion Yes No Figure 4.12. ISECR Window: Selecting the Criteria to Calculate Case Distances However, i f No is clicked in the ISECR window (Figure 4.10) to indicate that no location characteristics will be entered, case distances will only be calculated based on considering the quality scores alone. Whether the user decides to use only the quality scores or both the quality and relevance scores in determining the case distances, ISECR then presents a summary of ranked evaluation studies, which are essentially the retrieved cases that match the query parameters. This summary is displayed in Figure 4.13. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 61 Information System for Estimating Crash Reductions - [Search Summary] Search Summa Figure 4.13. ISECR Window: Summary of Ranked Documents Since some evaluation studies may consist of more than one case that matched the query parameters, the ranking of the studies is achieved by assigning each study a case distance only when the calculated case distance for that retrieved case is the lowest among all of its retrieved cases. The studies are then ranked according to the case distance assigned, as shown in the following example: Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 62 Table 4.1. Example: Ranking of Evaluation Studies Study Retrieved Case Case Distance Case Distance (Study Id) (Case Id) assigned to the study? 1 1 0.7 l.s 2 3 0.i25 Fes 1 2 0.25 No 3 6 0.25 Fes 4 7 0.3 Fes 2 4 0.3 No 2 5 0.3 No 4 8 0.3 No 4 9 0.3 No 5 10 0.35 Yes Hence, with the above example, the following five studies are listed in the ISECR summary window (see Figure 4.13): Table 4.2. Example: Final Output to the Summary Window Ranking S t u d y Retrieved Case Case Distance ( S t u d y Id) (Case Id) 1 1 1 0.1 2 2 3 0.125 3 3 6 0.25 4 4 7 0.3 5 5 10 0.35 Finally, to determine the countermeasure effectiveness, the user can click on the Calculate Results button, as shown in Figure 4.13. As mentioned in Chapter 3, ISECR utilizes all of the retrieved cases to derive the solutions (CRFs and the associated standard deviations) required for the current problem. Thus, with the same example as the one shown in Table 4.1, a total of ten cases from five evaluation studies would be used. The predicted CRFs and their standard deviations for total, and various collision severity and types are presented in an ISECR window as Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 63 shown in Figure 4.14. The non-weighted and weighted CRFs are determined by using Equations 3.3 and 3.4 respectively. As for the non-weighted and weighted standard deviations, Equations 3.5 and 3.6 are used respectively. Query Results: Collision Reduct ion Factors and their Standard Deviations Query Results: (66 Documents Found) Countermeasure: Delineation General| Location: Road Sections Collision Non-Weighted Non-Weighted Tjpe/Severity CRF(W) Standard Dev. Total: mBham [Kofi ' Weighted Weighted Standard Dev.. Fatal: Injury: Casualty: PDO: 47.3 34.4 24.4 _ Figure 4.14. ISECR Window: Predicted CRFs and their Standard Deviations Pre-Implementation Evaluation of Safety Improvement Programs Chapter 4: Description of ISECR 64 4.5 Input New Documents The Input New Documents option allows new documents to be added to the ISECR database. However, in order to assure the quality and consistency of data entry, input of new documents in ISECR is limited to the agency responsible for the database maintenance. As far as the types of information entered for each document and how they are represented in the database, Section 3 .2 should be referred. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 5: Validation of ISECR 65 5.0 VALIDATION OF ISECR 5.1 Introduction Validation is an essential step to confirm the success of any computer software. In order to validate the results produced by ISECR, this chapter is devoted to compare the CRFs predicted by ISECR and those extracted from the literature for 15 safety improvements (see Table 5.1 for details). These improvements were selected randomly from the 116 countermeasures considered in ISECR (see Appendix A for details). 5.2 ISECR Results in the Context of Published Studies Pertinent portions of four previously published sources (Tamburri and Smith, 1971, Creasey and Agent, 1985, McFarland et al., 1978, Terry and Watson, 1982) that have compiled and tabulated a summary of CRFs for various countermeasures are summarized in Table 5.1. Table 5.1 also provides both the non-weighted and weighted CRFs produced by ISECR for the chosen 15 countermeasures. Upon inspection of Table 5.1, it is evident that the scope of each study varies greatly from one to another. The CRFs presented in Table 5.1, while representative of the four published studies, should not be considered to be complete as there are additional collision data presented in these reports which are not relevant for the purpose of this thesis and thus, excluded from Table 5.1. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 5: Validation of ISECR 66 Table 5.1. Summary of CRFs: ISECR Results vs. Published Results CRF (lotal Collisions) C'oiink'rmeasure Type Tamburri C'rcascv .McFarlami Terry ISKCK ISKCK & ' ; et al. Non- Weighted Smith Agent 1078 W alson W eiglited l«'71 1985 Bridge improvements Widen bridges 30-65 65 63.5 Geometric improvements Horizontal alignment 20-40 40-88 41 25.6 23.8 Vertical alignment 15-54 30.3 29.1 Hor & vert alignment 50-52 20-21 40.4 32.7 Sight distance 20-31 31 33.5 Intersection improvements Left-turn chan. 15 38.2 31.9 Lighting improvements Install at intersections 75b 75b 50° 43.8,43.8C 36.6, 40.3C Pavement treatment Pavement grooving 75a 10-48 58.3 Resurfacing Skid reduction 12- 42 13- 50 12-44 21 22.4 31.4 19.0 14.7 Railway improvements Flashing beacons 70-94 64.5 60.3 Traffic controls/signals 4-way stops 70 68-70 73 62.8 44.1 Traffic signals Flashing beacons 37 54.0 red-yellow 50 34 26 34.0 New signals 15 15-80 6-29 20 25.6 21.0 'Wet pavement collisions only \"'Night time collisions only cRural area type By examining Table 5.1, it is evident that there is a general agreement between the ISECR results and the published results. In some instances, the ISECR CRFs are considerably higher than those reported by the published literature. For example, while McFarland et al. (1978) reported a CRF of 37% for implementing flashing beacons, ISECR predicted a CRF of 54%. The variation may be due to the difference in the sample size of studies and the approach undertaken by each study in determining the CRFs. For instance, by examining the ISECR outputs, it is noted that a small sample size of studies is used in the determination of the CRF, i.e., only four studies are retrieved Pre-Implementation Evaluation of Safety Improvement Programs Chapter 5: Validation of ISECR 67 from the ISECR database. Not only that, of the four studies retrieved, three (Mayer, 1971, Cribbins and Walton, 1970, Wilson, 1967) of them are outdated and analyzed their collision data by using the simple before and after approach. Although the studies are outdated, the ISECR result is an indication of how effective the countermeasure is, without correcting for the confounding factors. Further examination of the ISECR outputs for other countermeasures, it is also noted that the ISECR predictions are more valid and comparable to the literature results when there is a larger sample size available for the specific countermeasure. Based on the countermeasures examined in Table 5.1, it is apparent that more work is required to enter new evaluation studies into the ISECR database to further increase the reliability of the ISECR results for some countermeasures. Nevertheless, for the majority of the countermeasures examined above, ISECR does provide valid results. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 6: Expected Collision Reduction 68 6.0 EXPECTED COLLISION REDUCTION 6.1 Introduction The safety benefits of implementing a safety improvement can be represented by the expected reduction in collision frequency, which is normally calculated by the product of CRF and the expected number of collisions (denoted by N). This thesis has proposed the use of a C B R approach, one of the recent developments in problem-solving paradigms in artificial intelligence, to assist in developing an intelligent database (ISECR) that is capable of assessing the range and reliability of the predicted countermeasure effectiveness. With ISECR, CRFs are determined on a project-level basis where the variability of geometric configuration among different locations is considered. Moreover, ISECR addresses the uncertainty issues related to the effectiveness of the proposed safety improvement. Specifically, non-weighted and weighted CRFs and their standard deviations can be determined with ISECR as shown previously. The expected number of collisions, N , at a given location can be evaluated by procedures such as the multivariate approach, i.e., the G L I M (generalized linear modeling) approach, and the EB (Empirical Bayes) approach. As discussed in Chapter Two, these techniques can be used to readily provide more accurate site-specific safety estimates compared to the conventional approaches. The outcomes from these techniques are drawn upon in the next section, i.e., the expected value and standard deviation of N . Pre-Implementation Evaluation of Safety Improvement Programs Chapter 6: Expected Collision Reduction 69 The next section outlines the procedures involved in evaluating the collision reduction and its uncertainty expected from implementing a safety improvement. Specifically, the moment approach (Benjamin and Cornell, 1970, and Ang and Tang, 1984) is utilized to combine CRFs and expected number of collisions and their uncertainties. 6.2 Collision Reduction and Its Uncertainty With the availability of the expected values and standard deviations of both CRF and N , the expected reduction in collision frequency, Z, and its variance can now be calculated. Assuming that CRF and N are independent of each other, i.e., no correlation between the two variables, the expected value of Z can be determined by the following equation: E(Z) = NxCRF (Equation 6.1) where E(Z) = Expected reduction in collision frequency in the after period N = Expected number of collisions CRF = Collision reduction factor The accuracy of Z is represented by the variance which can be calculated as: Var(Z) = N2(crCRF)2 + CRF2(CJn)2 + (a, CRF (Equation 6.2) where Var(Z) = Variance of Z Pre-Implementation Evaluation of Safety Improvement Programs Chapter 6: Expected Collision Reduction 70 o~CRF = Standard deviation of CRF o~N = Standard deviation of N The application of the above procedures is illustrated with an example in Chapter Eight. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 7: Economic Analysis 71 7.0 ECONOMIC ANALYSIS 7.1 Introduction There are a number of methods that can be used to evaluate the economic feasibility of implementing a safety improvement. One of the most frequently used evaluation measures is the benefit-cost ratio (BCR). The B C R is a measure of the amount of dollar return expected with every dollar spent on a safety improvement. The expected benefits from implementing a countermeasure can be evaluated by the savings anticipated from the reduction in collisions. With the expected value and variance of Z determined in Chapter Six, the uncertainty issues associated with B C R can now be addressed. Specifically, the probability of a proposed countermeasure achieving a preset B C R can now be determined. 7.2 Benefit-Cost Ratio (BCR) and Its Uncertainty The expected value of B C R for implementing a countermeasure can be determined by the following equation: CoStimpiemen(ation where E(BCR) = Expected value of B C R E(Z) = Expected reduction in collision frequency in the after period Pre-Implementation Evaluation of Safety Improvement Programs Chapter 7: Economic Analysis 72 Col.Cost = Average collision cost t = Payback period (year) = Discount rate (P / A, i, t) = Present worth factor, given the payback period and discount rate k (Col.Cost) x (PI A,i,t) (Equation 7.2) implementation Assuming that the only random variable in Equation 7.1 is Z , i.e., k is a constant variable, the variance of B C R can then be expressed by Equation 7.3: 7.3 Probability Density Function of BCR With the expected value and variance of BCR, the probability density function of B C R can now be established. This thesis utilizes a Gamma distribution to model BCR, as this distribution is suitable for modeling continuous random variable and providing wide variety of shapes. Furthermore, this distribution is also limited to positive values and skewed to the right, which is appropriate to model BCR. The Gamma distribution parameters, a and 6, and the probability density function of B C R can be calculated by the following equations: Var(BCR) = k2 xVar(Z) (Equation 7.3) where Var(BCR) = Variance of B C R Var(Z) = Variance of Z , as defined by Equation 6.2. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 7: Economic Analysis 73 _ E(BCR) p = (Equation 7.4) a = E(BCR) x B (Equation 7.5) f(BCR;a,B) = J—fBCRy-'e™*™ (Equation 7.6) T(a) where a and B = Gamma distribution parameters / (BCR,a, 6) = Probability density function of B C R Once the above probability density function is defined, the cumulative distribution function for B C R can be formulated as follows: F(BCR;a,B) = }f(BCR;a,B)d(BCR) (Equation 7.7) o where F(BCR; a, (3) = Cumulative distribution function of B C R By plotting the cumulative distribution function of BCR, the probability of a countermeasure achieving a specific B C R upon its implementation can be determined. This is illustrated with an example in the next chapter. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 74 8.0 APPLICATIONS 8.1 Problem Definition The procedures outlined in the previous chapters can be used to evaluate the cost-effectiveness of road safety improvements. For example, assume that a traffic safety engineer is interested in determining the probability of achieving a 2:1 return, i.e., a benefit-cost ratio (BCR) of 2.0, by improving public lighting along an arterial street (or known as road sections in the location type feature used in this research) that has the following characteristics: • Area type: suburban • Implementation level: isolated location (this is automatically assumed by ISECR when the user decides to include the relevance criterion in the calculation of case distances) • Total traffic volume: 25000 A A D T With the above problem specifications, the remaining sections in this chapter intend to: 1. Determine the effectiveness of the proposed countermeasure, i.e., CRFs and the standard deviations. 2. Calculate the reduction in collision anticipated once the countermeasure is implemented by utilizing the moment approach, as discussed in Chapter Six. 3. Compute the B C R and its variance for the proposed countermeasure and plot its cumulative distribution function to assess the probability for the proposed countermeasure achieving a B C R of 2.0. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 75 8.2 Countermeasure Effectiveness The effectiveness of the proposed countermeasure is evaluated by utilizing cases currently available in the ISECR case base. Relevant cases are retrieved only i f the same countermeasure and location types match between the design case and cases stored in the ISECR case base. Querying the ISECR case base with general lighting improvements and road sections as the countermeasure and location types respectively, a total of 53 cases from 38 documents are retrieved. Table 8.1 provides a list of the retrieved cases. To determine the case distance for each of the retrieved cases, the quality and relevance scores are considered. Tables 8.1 and 8.2 summarize the quality (treatment of the confounding factors) and relevance (location characteristics) information for each case, as extracted from Subcases 2 and 5 from the ISECR case base respectively. With the information available, both the quality and relevance scores can be determined by using Equation 3.1, as shown below for Case Id 296: D k (Quality) Y \" wrABS fi fi,k fR - fR J i,max J i,n\\] 1.0 -ABS V 1.0 3 - 1 3 - 0 , V 0.667 Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 76 D k(Re levance) y w, • ABS f i f i , k fR - f. J i,max J i R mm J i=l ' Q.5-ABS\\ 4-1 + 0 . 5 - ^ 5 5 1- 2^ 2 - 1 0.5 + 0.5 D^Relevance) ~ 0-745 As mentioned previously, when both the quality and relevance scores are considered when determining a case distance, they are weighted equally. For Case Id 296, case distance can be evaluated using Equation 3.2 as below: Dk=\\- + Dk(Ke l e v m c e ) ) = i (0.667 + 0.745) = 0.706 Using the same approach, case distances are computed for all of the retrieved cases and the results are as listed in Table 8.3. Furthermore, Table 8.3 also presents the results (CRFs) extracted from the retrieved cases, however, in this example, only CRFs for total collisions are considered. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 77 Table 8.1. Example: Retrieved Cases and their Quality Scores Retrieved Study Treatment of the Confounding Factors Case ((use Id) (Study Id) Changes in Inclusion of R I M Total U of Qualit> Traffic 1 II related Artifact Factors Score. Volume Kffccts Treated f^kllJmiliiM 169 54 No No .v. 0 1.000 192 59 Yes No No 1 0.667 238 68 Yes No No 1 0.667 257 73 No No No 0 1.000 267 74 No No No 0 1.000 296 77 No Yes No 1 0.667 297 77 No Yes No 1 0.667 314 423 Yes Yes Yes 3 0.000 319 424 Yes Yes Yes 3 0.000 328 434 No No No 0 1.000 359 448 No No No 0 1.000 406 101 No No No 0 1.000 407 101 No No No 0 1.000 408 101 No No No 0 1.000 463 116 No No No 0 1.000 480 119 No Yes No 1 0.667 498 124 No No No 0 1.000 689 78 Yes No No 1 0.667 690 78 Yes No No 1 0.667 713 196 No No No 0 1.000 766 1 No No No 0 1.000 796 229 No No No 0 1.000 835 376 No No No 0 1.000 854 403 No Yes No 1 0.667 855 403 No Yes No 1 0.667 856 403 No Yes No 1 0.667 861 254 No No No 0 1.000 869 256 Yes Yes Yes 3 0.000 870 257 No Yes No 1 0.667 885 262 No Yes No 1 0.667 893 265 No No No 0 1.000 894 265 No No No 0 1.000 895 265 No No No 0 1.000 896 265 No No No 0 1.000 897 265 No No No 0 1.000 898 265 No No No 0 1.000 899 265 No No No 0 1.000 Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 78 Table 8.1. Example: Retrieved Cases and their Quality Scores (cont.) Retrieved Case Study Treatment of the Confounding Factors (Case Id) (Study Id) Changes in Traffic Volume Inclusion of Unrelated KfTccts RT.M Artifact Total # of Factors 'Created Quality Score. •J01 267 No No No 0 1.000 902 267 No No No 0 1.000 903 267 No No No 0 1.000 923 27 No No No 0 1.000 925 89 No No No 0 1.000 930 282 No No No 0 1.000 942 295 No No No 0 1.000 944 297 No No No 0 1.000 958 311 No No No 0 1.000 1253 292 No No No 0 1.000 2800 1 No No No 0 1.000 2805 285 No No No 0 1.000 2820 409 Yes Yes Yes 3 0.000 2828 413 No No No 0 1.000 2867 493 No No No 0 1.000 2872 496 No No No 0 1.000 Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 79 Table 8.2. Example: Retrieved Cases and their Relevance Scores Retrieved Case Study Location ( liaracteristics (Case Id) (Study Id) Area Type Implementation I ,evel Traffic Volume (AADT) Relevance Score, DkiKtlt-iumil 169 54 Other Wide Area 0 S50 192 59 Other Wide Area 0.850 238 68 Other Wide Area 0.850 257 73 Other Wide Area 0.850 267 74 Other Wide Area 0.850 296 77 Urban Wide Area 0.745 297 77 Urban Wide Area 0.745 314 423 Rural Wide Area 0.745 319 424 Other 0.667 328 434 Urban Wide Area 0.745 359 448 Other Wide Area 0.850 406 101 Rural Wide Area 0.745 407 101 Rural Wide Area 0.745 408 101 Urban Wide Area 0.745 463 116 Urban 0.333 480 119 Urban 0.333 498 124 Urban 0.333 689 78 Urban Wide Area 0.745 690 78 Urban Wide Area 0.745 713 196 Urban 0.333 766 1 Rural Wide Area 0.745 796 229 Rural Wide Area 0.745 835 376 Other Wide Area 0.850 854 403 Urban Wide Area 0.745 855 403 Rural Wide Area 0.745 856 403 Other Wide Area 0.850 861 254 Urban Wide Area 0.745 869 256 Rural Wide Area 0.745 870 257 Rural Wide Area 0.745 885 262 Urban Wide Area 0.745 893 265 Other Wide Area 0.850 894 265 Other Wide Area 0.850 895 265 Other Wide Area 0.850 896 265 Other Wide Area 0.850 897 265 Other Wide Area 0.850 898 265 Urban Wide Area 0.745 899 265 Rural Wide Area 0.745 901 267 Other Wide Area 0.850 Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 80 Table 8.2. Example: Retrieved Cases and their Relevance Scores (cont.) Retrieved Case Study 1 .ocalion Characteristics (Case Id) (Study Id) Area Type Implementation Level Traffic Volume (AADT) Relevance Score. ^klReltvamvi 902 267 Other Wide Area 5000-00')') 0.715 903 267 Other Wide Area 0.850 923 27 Other Isolated Location 20000-29999 0.385 925 89 Urban Isolated Location 0.236 930 282 Other Wide Area 0.850 942 295 Other 0.667 944 297 Other Wide Area 0.850 958 311 Other Wide Area 0.850 1253 292 Other Wide Area 0.850 2800 1 Rural Wide Area 0.745 2805 285 Other Wide Area 0.850 2820 409 Other Wide Area 0.850 2828 413 Other Wide Area 0.850 2867 493 Other Wide Area 0.850 2872 496 Other 0.667 Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 81 Table 8.3. Example: Retrieved Cases (Ranked) and their Case Distances and Results ( use Id Study Id Quality Relevance Case Case Result, Weighted Score, Score. Distance, CRFk <%) Case Result DklQualifyi Dk(Rrh-iance) Dk <%) W) 54 1.000 0.850 0.925 ••.'•75 15.00 1.13 192 59 0.667 0.850 0.758 0.242 25.00 6.04 238 68 0.667 0.850 0.758 0.242 79.00 19.10 257 73 1.000 0.850 0.925 0.075 40.00 3.00 267 74 1.000 0.850 0.925 0.075 60.00 4.50 296 77 0.667 0.745 0.706 0.294 75.13 22.09 297 77 0.667 0.745 0.706 0.294 45.28 13.31 • 314 423 0.000 0.745 0.373 0.627 22.00 13.80 319 424 0.000 0.667 0.333 0.667 5.00 3.33 328 434 1.000 0.745 0.873 0.127 57.00 7.26 359 448 1.000 0.850 0.925 0.075 30.00 2.25 406 101 1.000 0.745 0.873 407 101 1.000 0.745 0.873 0.127 30.00 3.82 408 101 1.000 0.745 0.873 463 116 1.000 0.333 0.667 480 119 0.667 0.333 0.500 0.500 9.00 4.50 498 124 1.000 0.333 0.667 0.333 30.00 10.00 689 78 0.667 0.745 0.706 0.294 58.00 17.05 690 78 0.667 0.745 0.706 713 196 1.000 0.333 0.667 0.333 30.00 10.00 766 1 1.000 0.745 0.873 0.127 58.00 7.38 796 229 1.000 0.745 0.873 0.127 30.00 3.82 835 376 1.000 0.850 0.925 0.075 21.00 1.58 854 403 0.667 0.745 0.706 0.294 63.57 18.69 855 403 0.667 0.745 0.706 0.294 63.33 18.62 856 403 0.667 0.850 0.758 0.242 73.09 17.67 861 254 1.000 0.745 0.873 869 256 0.000 0.745 0.373 0.627 10.00 6.27 870 257 0.667 0.745 0.706 0.294 20.00 5.88 885 262 0.667 0.745 0.706 0.294 50.00 14.70 893 265 1.000 0.850 0.925 894 265 1.000 0.850 0.925 895 265 1.000 0.850 0.925 0.075 30.00 2.25 896 265 1.000 0.850 0.925 0.075 21.00 1.58 897 265 1.000 0.850 0.925 898 265 1.000 0.745 0.873 0.127 38.00 4.84 899 265 1.000 0.745 0.873 0.127 36.90 4.70 901 267 1.000 0.850 0.925 0.075 59.00 4.43 902 267 1.000 0.715 0.858 0.142 14.00 1.99 903 267 1.000 0.850 0.925 0.075 26.00 1.95 Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 82 Table 8.3. Example: Retrieved Cases (Ranked) and their Case Distances and Results (cont.) Case Id Study Id Qualil> Relevance C ase 0-0,.) Case Result. Weighted Score, Score, Distance, CRt\\ (%) Case Result DklQualitv) DkfRelcvancei Dk (%) 923 27 1.000 0.385 0.692 925 89 1.000 0.236 0.618 0.382 -19.00 -7.26 930 282 1.000 0.850 0.925 0.075 24.69 1.85 942 295 1.000 0.667 0.833 944 297 1.000 0.850 0.925 958 311 1.000 0.850 0.925 1253 292 1.000 0.850 0.925 0.075 -18.00 -1.35 2800 1 1.000 0.745 0.873 0.127 17.00 2.16 2805 285 1.000 0.850 0.925 0.075 69.00 5.18 2820 409 0.000 0.850 0.425 2828 413 1.000 0.850 0.925 2867 493 1.000 0.850 0.925 0.075 30.00 2.25 2872 496 1.000 0.667 0.833 Sum 8.263 1327.99 260.38 Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 83 Table 8.3 illustrates again that cases with lower case distances are weighted more significantly than those with higher case distances. The non-weighted and weighted CRFs for total collisions can be determined by using Equations 3.3 and 3.4 respectively, as shown below: > , CRF. 13?7 99 CRF = - = = 34 94% ^-^Non-Weighted J t . ^ H / o C R F TkJ-Dk)-CRFk 260.38 Therefore, for the proposed countermeasure, i.e., general lighting improvement, the non-weighted and weighted CRFs are 34.94% and 31.51% respectively. As expected, the weighted CRF is less than the non-weighted CRF as with the weighted result, the confounding factors have been accounted for. Finally, the accuracy of these estimates can be represented by their standard deviations. The standard deviations for both the non-weighted and weighted CRFs can be determined with Equations 3.5 and 3.6 respectively, as shown below: &CRF-Non-Weighted £ M (CRFk - CRFf _ /212221J ^ % N-l V 38-1 °CRF-Weighted Y^-D^iCRF.-CRFf _ (5287^5 = 4 1 6 % Z L d - ^ ) - ( ^ - l ) V 305.73 Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 84 The standard deviations for the non-weighted and weighted CRFs are 23.94% and 4.16% respectively. Clearly, the non-weighted standard deviation is significantly greater than the weighted standard deviation. 8.3 Expected Collision Reduction From the analysis conducted in Section 8.2, the expected values for the non-weighted and weighted CRFs are 34.94% and 31.51%, while the standard deviations are 23.94% and 4.16% respectively. Assuming that from the collision analysis conducted by the safety engineer, the expected number of collisions, N , for this location is 20.0 collisions per year (col/yr) with a standard deviation of 2.5 col/yr. Equations 6.1 and 6.2 can now be used to calculate the expected value and variance of Z respectively. For example, with the non-weighted results, the expected value of Z and its variance can be computed as follows: E (Z) = Nx CRF = 20.0 x 0.3494 = 6.99col I yr Var(Z) = N2(o-CRF)2+CRF2(o-N)2 + (aCRF x crN)2 = 20 2(0.2394) 2 +0.3494 2(2.5) 2 + (0.2394 x 2.5) 2 = 24.06(co//.yr) 2 As for the weighted results, E(Z) = 6.30col I yr and Var(Z) = \\32{col I yr)2 can be determined in a similar fashion as shown above. The above results represent the expected reductions in total collisions and the uncertainties at the location. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 85 8.4 Economic Analysis With the results determined in Section 8.3, the expected value of B C R and its accuracy can now be evaluated. For the purpose of this example, the variable k in Equation 7.1 is assumed to be 0.25. Hence, the non-weighted B C R and its variance can be determined with Equations 7:1 and 7.3 respectively, as shown below: E(BCR) = kx E(Z) = 0.25 x 6.99 = 1.747 Var(BCR) = k2 x Var(Z) = 0.25 2 x 24.06 = 1.504 Similarly, the weighted B C R and its variance can be calculated and are determined to be 1.576 and 0.083 respectively. Therefore, given the expected value and variance of BCR, the two gamma distribution parameters, a and f3 corresponding to the non-weighted B C R can be determined using Equations 7.4 and 7.5 respectively as follows: Similarly, the a and /? parameters for the weighted B C R are 30.017 and 19.051 respectively. It follows that with the known gamma parameters, a cumulative distribution can be plotted as shown in Figure 8.1 (see Appendix E for the probability density plot). Subsequently, the P = EjBCR) _ 1.747 Var(BCR) ~ 1.504 = 1.161 a = E(BCR) x fi = 1.747 x 1.161 = 2.03 Pre-Implementation Evaluation of Safety Improvement Programs Chapter 8: Applications 86 probability of achieving a specific economic goal can be evaluated prior to the implementation of a countermeasure. For example, i f it is desired to determine the probability of achieving a B C R of 2.0 with the proposed countermeasure (general lighting improvement), Figure 8.1 indicates that the probabilities of achieving this goal (or having a B C R of less or equal to 2.0) are 61.5% and 80.6%o based on the non-weighted and weighted B C R curves respectively. Based on Figure 8.1, it is clear that the expected range of B C R represented by the weighted B C R curve is considerably smaller than the one indicated by the non-weighted curve, i.e., a B C R range of 1.0 to 2.5 compared to a range of 0.0 to 6.0 respectively. Cummulative Distribution of BCR 0 1 2 3 4 5 6 Benefit-Cost Ratio — -Weighted NonWeighted Figure 8.1. Example: Cumulative Distribution Plot of BCR Pre-Implementation Evaluation of Safety Improvement Programs Chapter 9: Conclusion 87 9.0 CONCLUSION This thesis first describes the development of ISECR, which is now a functional intelligent database on the MS Access platform. ISECR maintains a case base that consists of published literature which quantifies crash reduction benefits for various safety improvements. Utilizing a C B R approach, ISECR permits users to query the system for cases similar to the current situation, retrieves and then summarizes the retrieved solutions to estimate the range and reliability of the countermeasure effectiveness on a project level. With the ISECR outputs, this thesis then illustrates the use of the moment approach to determine the expected collision reduction and its uncertainties for specific countermeasures. Lastly, the technique involving the probability assessment of achieving a specific benefit-cost ratio for a specific countermeasure is also presented in the thesis. The evaluation procedure illustrated in this report attempts to increase the confidence of predictions when evaluating the expected benefits from safety improvement programs. Currently, 450 evaluation studies are entered into the ISECR case base. To further increase the usefulness and applicability of ISECR, new evaluation studies should be evaluated and entered into the database on a regular basis. Based on the available cases in the database, ISECR has shown to provide results that are comparable to the results obtained from real cases, given that there is an adequate sample size of retrieved studies that matches the user's query. Nevertheless, further data testing and validation of the system from experts and end-users are still necessary to improve the prototype ISECR. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 9: Conclusion 88 The confounding factors considered in ISECR can be further expanded to include other factors, such as collision migration. A different weighting scheme for the confounding factors can also be introduced to highlight the relative importance of each confounding factor. The same can be applied to the location characteristics considered in this thesis, i.e., other location characteristics, in addition to the ones included in ISECR, can also be introduced and a different weighting scheme can also be employed to emphasize the importance of some characteristics. In addition, the prototype ISECR can be improved by providing its user the capability of predicting CRFs for different combinations of countermeasures, instead of one countermeasure as currently allowed in ISECR. Frequently, more than one countermeasure is considered at a location, i.e., intersection improvements may include a new left-turn lane, new delineation, lane widening. Thus, it would definitely be beneficial to be able to query for the effect of a combination of countermeasures at a location. Additional risk analysis should also be performed for the economic procedure illustrated in this thesis. This is essential as the discount rate, i, and project life, t, used in Equation 7.1 were assumed to be constant over time. Nevertheless, these two variables can fluctuate during the life of the project and are subject to changes in interest rate, risk premium, market condition, government policy, etc. Hence, further risk analysis incorporating the random nature of these two variables should be conducted. Pre-Implementation Evaluation of Safety Improvement Programs Chapter 9: Conclusion 89 The computerized approach which ISECR employs minimizes the amount of manual work required for safety analysts to determine the effectiveness of safety improvements. However, it should be noted that ISECR is not intended to eliminate the use of engineering. In many cases, the safety analyst has to use his/her judgement in evaluating the results produced by ISECR. Pre-Implementation Evaluation of Safety Improvement Programs Bibliography 90 B I B L I O G R A P H Y Aamodt, A . and Plaza, E. (1994). Case-based reasoning: foundational issues, methodological variations, and system approaches, Artificial Intelligence Communications, Vol.7, N o . l , pp.39-59. Ang, A . and Tang, W. (1984). Probability concepts in engineering planning and design, John Wiley and Sons, Toronto, Canada. Benjamin, J. and Cornell, C. (1970). Probability, statistics, and decision for civil engineers, McGraw-Hill , Toronto, Canada. Brude, U . and Larsson, J. (1988). The use of prediction models for eliminating effects due to regression-to-the-mean in road accident data, Accident Analysis and Prevention, Vol.20, No.4, pp.299-310. Capus, L . and Tourigny, N . (1998). Road safety analysis: a case-based reasoning approach, Transportation Research Board 77 t h Annual Meeting, Washington, D.C. Creasey, T. and Agent, K.R. (1985). Development of accident reduction factors, Research Report UKTRP-85-6, Lexington, K Y , Kentucky Transportation Cabinet, Federal Highway Administration. Cribbins, P. and Walton, C. (1970). Traffic signals and overhead flashers at rural intersections: their effectiveness in reducing accidents, Highway Research Record, Vol.325, pp. 1-14. Elvik, R. (1995). Meta-analysis of evaluations of public lighting as accident countermeasure, Transportation Research Record, No.1485, pp.112-123. Elvik, R. (1996). Evaluations of road accident blackspot treatment: a case of the iron law of evaluation studies?, Accident Analysis and Prevention, Vol.28, No.6, pp.685-694. Pre-Implementation Evaluation of Safety Improvement Programs Bibliography 91 Elvik, R. (1998). Are road safety evaluation studies published in peer reviewed journals more valid than similar studies not published in peer reviewed journals?, Accident Analysis and Prevention, Vol.30, N o . l , pp.101-118. Elvik, R. (1998). Evaluating the statistical conclusion validity of weighted mean results in meta-analysis by analyzing funnel graph diagrams, Accident Analysis and Prevention, Vol.30, No.2, pp.255-266. Gonzalez, A.J . and Laureano-Ortiz, R. (1992). A case-based reasoning approach to real estate property appraisal, Expert Systems with Applications, Vol.4, pp.229-246. Hauer, E. (1992). Empirical Bayes approach to the estimation of \"unsafety\": the multivariate regression method, Accident Analysis and Prevention, Vol.24, No.5, pp.457-477. Hauer, E., Ng, J., and Lovell, J. (1988). Estimation of safety at signalized intersections, Transportation Research Record, No . l 185, pp.48-61. Jovanis, P. and Chang, H . (1986). Modeling the relationship of accidents to miles traveled, Transportation Research Record, No. 1068, pp.42-51. Khattak, A . and Kanafani, A . (1996). Case-based reasoning: a planning tool for intelligent transportation systems, Transportation Research Part C-Emerging Technologies, Vol.4, No.5, pp.267-288. Khattak, A . and Renski, H . (1999). Plan HOV: a case-based reasoning planning tool for high-occupancy-vehicle lane analysis in a GIS environment, Transportation Research Board 78 t h Annual Meeting, Washington D.C. Kolodner, J. (1993). Case-based reasoning, Morgan Kaufmann Publishers, Inc. Pre-Implementation Evaluation of Safety Improvement Programs Bibliography 92 Kulmala, R. (1995). Safety at rural three- and four-arm junctions. Development of accident prediction models, Espoo 1995, Technical Research Centre of Finland, V T T 233. Leake, D.B. and Plaza E. (1997). Case-based reasoning research and development: second international conference on case-based reasoning, Providence, RI, USA. Maher, M . L . , Balachandran, M.B. , and Zhang, D . M . (1995). Case-based reasoning in design, Lawrence Erlbaum Associates. Mayer, P. (1971). Relating traffic control and roadway elements to highway safety- the relationship between highway transportation and safety, Traffic Engineering, pp.23-27. McFarland, W.F., Griffin, L . L , Rollins, J.B., Stockton, W.R., Phillips, D.T., and Dudek, C.L. (1978). Assessment of techniques for cost-effective of highway accident countermeasures, Report No. FHWA-RD-79-53, Washington, D . C , Federal Highway Administration. Miaou, S. and Lum, H . (1993). Modeling vehicle accident and highway geometric design relationships, Accident Analysis and Prevention, Vol.25, No.6, pp.689-709. Pu, P. (1993). Introduction: issues in case-based design systems, AI E D A M , Vol.7, No.2, pp.79-85. Pu, P. and Maher, M . (1998). Issues and applications of case-based reasoning in design, Lawrence Erlbaum Associates. Saccomanno, F. and Buyco, C. (1988). Generalized log-linear models of truck accident rates, Paper presented at Transportation Research Board 67 t h Annual Meeting, Washington, D.C. Schank, R.C. (1982). Dynamic memory: a theory of reminding and learning in computers and people, Cambridge University Press, Cambridge, England. Pre-Implementation Evaluation of Safety Improvement Programs Bibliography 93 Tamburri, T.N. and Smith, R.N. (1971). The safety index: a method of evaluating and rating safety benefits, Highway Research Record 332, pp.28-43. Terry, D.A. and Watson, J.E. (1982). Post-implementation evaluation system methodology and output, Traffic and Safety Division, New York State Department of Transportation. Watson, I. (1997). Applying case-based reasoning: techniques for enterprise systems, Morgan Kaufmann Publishers, Inc. Wilson, J. (1967). Simple types of intersection improvements, Highway Research Board, Special Report 93, pp.144-159. Yeh, L C . (1997). Case-based approaches for preliminary design of steel building frames, Microcomputers in Civi l Engineering, Vol.12, pp.327-337. Pre-Implementation Evaluation of Safety Improvement Programs Appendix A: Countermeasure Types 94 A P P E N D I X A : C O U N T E R M E A S U R E T Y P E S Pre-Implementation Evaluation of Safety Improvement Programs Appendix A: Countermeasure Types 95 Table A . l . Countermeasure Types Countermeasure Category Countermeasure Type 1. Area-wide schemes General Blackspot treatment Enforcement New roads Traffic calming: general Traffic calming: speed bumps/humps Traffic planning/Management 2. Bridge improvements General Widen bridges 3. Cyclist/Pedestrian facilities General bicycle safety improvements General pedestrian safety improvements Pedestrian crossings Pedestrian overpasses/underpasses Pedestrian signals 4. Delineation General Pavement markings: general Pavement markings: edge lines Pavement markings: median edge line Pavement markings: right edge line Raised pavement markers Reflected guide posts Strips: general Strips: rumble Strips: transverse 5. Geometric improvements General Alignment: general Alignment: horizontal Alignment: vertical Median: general Median: close median openings Median: concrete/attenuator barriers Median: install new median Median: upgrade existing median Median: widen existing median Sight distance Staggered intersection Superelevation Pre-Implementation Evaluation of Safety Improvement Programs Appendix A: Countermeasure Types 96 Table A . l . Countermeasure Types (cont.) ( ouiitcrmeasiire Category Countermeasure T\\pe 6. Intersection improvements General Channelization: general Channelization: left-turn lane: general Channelization: left-turn lane with left-turn phase Channelization: left-turn lane without left-turn phase Channelization: right-turn lane Turning bay/traffic island Roundabouts: general Roundabouts: installation/upgrade 7. Lane/Shoulder treatment General 2-way left-turn lane Acc./Decel./Passing lane Bicycle lane Bus lane Climbing lane Flattening side-slopes H O V lane Lane/shoulder: general Lane/shoulder: narrowing Lane/shoulder: widening Lane addition 8. Lighting improvements General Install new lighting Upgrade existing lighting 9. Object removal/relocation General Relocation: general Relocation: fixed objects Relocation: utility poles Removal: general Removal: fixed objects Removal: trees Removal: utility poles 10. Parking improvements General Change from angle- to parallel parking Eliminate parking: general Eliminate parking: parallel parking 11. Pavement treatment General Pavement grooving Resurfacing Skid reduction Pre-Implementation Evaluation of Safety Improvement Programs Appendix A: Countermeasure Types 97 Table A . l . Countermeasure Types (cont.) Countermeasure Category Countermeasure Type 12. Railway improvements General Automatic gates Flashing beacons 13. Regulation change General Modify speed limit: general Modify speed limit: decrease Modify speed limit: increase Prohibit turns: general Prohibit turns: left-turns 14. Safety barriers General Crash cushions Guardrails: general Guardrails: double-sided Safety poles/posts 15. Traffic controls/signs General Guidance signs Install new signs/upgrade existing signs Regulatory signs: general Regulatory signs: speed Regulatory signs: stop signs: general Regulatory signs: stop signs: 2-way to 4-way stops Regulatory signs: stop signs: 4-way stops Regulatory signs: minor-leg stops Regulatory signs: yield signs Warning signs: general Warning signs: flashing beacons/signals 16. Traffic signals General Actuated signals Advance warning signs Coordinated signals Flashing beacons/signals: general Flashing beacons/signals: all-way red Flashing beacons/signals: red-yellow Install new signals/upgrade existing signals Phasing: general Phasing: left-turn phase Phasing: pedestrian phase Phasing: timing Removal signals Pre-Implementation Evaluation of Safety Improvement Programs Appendix B: Location Characteristics Considered for Each of the Eight Location Types 98 APPENDIX B: LOCATION CHARACTERISTICS CONSIDERED FOR EACH OF THE EIGHT LOCATION TYPES Pre-Implementation Evaluation of Safety Improvement Programs Appendix B: Location Characteristics Considered for Each of the Eight Location Types 99 Table B.l. Representation of Location Characteristics for General Intersections Characteristic Feature \\ alue Area type l=urban 2=suburban 3=rural 4=other Intersection type l=four-legged 2=T-intersection 3=Y-intersection Implementation level l=isolated location 2=wide area Total entering traffic volume (AADT) 1=0-4999 2=5000-9999 3=10000-14999 4=15000-19999 5=20000-29999 6=30000-39999 7=40000-49999 8=50000-59999 9=60000-69999 10=70000-79999 11=80000 and more Average lane width l=less than 12 ft 2=greater or equal to 12 ft Provision of left-turn channelization l=yes 2=no Provision of right-turn channelization l=yes 2=no Left-turn movement l=not allowed 2=permissive 3=protected Right-turn movement l=not allowed 2=permissive 3=protected Street parking l=allowed 2=not allowed Average number of lanes per approach l=less than or equal to 2 lanes 2=more than 2 lanes Pre-Implementation Evaluation of Safety Improvement Programs Appendix B: Location Characteristics Considered for Each of the Eight Location Types 100 Table B.2. Representation of Location Characteristics for Signalized Intersections Characteristic Feature Value Area type l=urban 2=suburban 3=rural 4=other Intersection type l=four-legged 2=T-intersection 3=Y-intersection Implementation level l=isolated location 2=wide area Total entering traffic volume (AADT) 1=0-4999 2=5000-9999 3=10000-14999 4=15000-19999 5=20000-29999 6=30000-39999 7=40000-49999 8=50000-59999 9=60000-69999 10=70000-79999 11=80000 and more Average lane width l=less than 12 ft 2=greater or equal to 12 ft Provision of left-turn channelization l=yes 2=no Provision of right-turn channelization l=yes 2=no Left-turn movement l=not allowed 2=permissive 3=protected Right-turn movement l=not allowed 2=permissive 3=protected Street parking l=allowed 2=not allowed Type of traffic control l=fixed-timed 2=semi-actuated 3=fully-actuated Average number of lanes per approach l=less than or equal to 2 lanes 2=more than 2 lanes Pre-Implementation Evaluation of Safety Improvement Programs Appendix B: Location Characteristics Considered for Each of the Eight Location Types 101 Table B.3. Representation of Location Characteristics for Unsignalized Intersections ('haractcristic Feature Value Area type l=urban 2=suburban 3=rural 4=other Intersection type l=four-legged 2=T-intersection 3=Y-intersection Implementation level l=isolated location 2=wide area Total entering traffic volume 1=0-4999 (AADT) 2=5000-9999 3=10000-14999 4=15000-19999 5=20000-29999 6=30000-39999 7=40000-49999 8=50000-59999 9=60000-69999 10=70000-79999 11=80000 and more Average lane width l=less than 12 ft 2=greater or equal to 12 ft Provision of left-turn l=yes channelization 2=no Provision of right-turn l=yes channelization 2=no Left-turn movement l=not allowed 2=permissive 3=protected Right-turn movement l=not allowed 2=permissive 3=protected Street parking l=allowed 2=not allowed Type of traffic control 1 uncontrolled 2=2-way stops 3=4-way stops Average number of lanes per l=less than or equal to 2 lanes approach 2=more than 2 lanes Pre-Implementation Evaluation of Safety Improvement Programs Appendix B: Location Characteristics Considered for Each of the Eight Location Types 102 Table B.4. Representation of Location Characteristics for Road Sections Characteristic Feature \\ alue Area type l=urban 2=suburban 3=rural 4=other Implementation level l=isolated location 2=wide area Total traffic volume (AADT) 1=0-4999 2=5000-9999 3=10000-14999 4=15000-19999 5=20000-29999 6=30000-39999 7=40000-49999 8=50000-59999 9=60000-69999 10=70000-79999 11=80000 and more Average lane width l=less than 12 ft 2=greater or equal to 12 ft Provision of left-turn l=yes channelization 2=no Provision of right-turn l=yes channelization 2=no Street parking l=allowed 2=not allowed Average number of lanes 1-less than or equal to 2 lanes 2=more than 2 lanes Passing/Acceleration/ l=yes Deceleration Lanes 2=no Pre-Implementation Evaluation of Safety Improvement Programs Appendix B: Location Characteristics Considered for Each of the Eight Location Types 103 Table B.5. Representation of Location Characteristics for Freeways Characteristic geature \\ aluc Area type l=urban 2=suburban 3=rural 4=other Implementation Level l=isolated location 2=wide area Total traffic volume (AADT) 1=0-4999 2=5000-9999 3=10000-14999 4=15000-19999 5=20000-29999 6=30000-39999 7=40000-49999 8=50000-59999 9=60000-69999 10=70000-79999 11=80000 and more Average lane width l=less than 12 ft 2=greater or equal to 12 ft Average number of lanes l=less than or equal to 2 lanes 2=more than 2 lanes Passing/Acceleration/ Deceleration Lanes l=yes 2=no Pre-Implementation Evaluation of Safety Improvement Programs Appendix B: Location Characteristics Considered for Each of the Eight Location Types 104 Table B.6. Representation of Location Characteristics for Bridges, Rails, and Construction Zones Characteristic Feature Value Area type l=urban 2=suburban 3=rural 4=other Implementation Level l=isolated location 2=wide area Total traffic volume (AADT) 1=0-4999 2=5000-9999 3=10000-14999 4=15000-19999 5=20000-29999 6=30000-39999 7=40000-49999 8=50000-59999 9=60000-69999 10=70000-79999 11=80000 and more Average lane width l=less than 12 ft 2=greater or equal to 12 ft Average number of lanes l=less than or equal to 2 lanes 2=more than 2 lanes Pre-Implementation Evaluation of Safety Improvement Programs Appendix C: ISECR Windows: Countermeasure Types Considered for Different Countermeasure Categories 105 APPENDIX C: ISECR WINDOWS: COUNTERMEASURE TYPES CONSIDERED FOR DIFFERENT COUNTERMEASURE CATEGORIES Pre-Implementation Evaluation of Safety Improvement Programs Appendix C: ISECR Windows: Countermeasure Types Considered for Different Countermeasure Categories 106 Predict CR Factors (cont.) Select a Countermeasure: <& General $ Blackspot treatment $ Enforcement $ New roads $ Traffic calming: general : . ~ • $ Traffic calming: speed bumps/humps # Traffic planning/Management Figure C.l. ISECR Window: Countermeasure Types for Area-Wide Schemes Predict CR Factors (cont.) Select a Countermeasure: Figure C.2. ISECR Window: Countermeasure Types for Bridge Improvements Pre-Implementation Evaluation of Safety Improvement Programs Appendix C: ISECR Windows: Countermeasure Types Considered for Different Countermeasure Categories 107 Predict CR Factors (cont.) Select a Countermeasure: @ General bicycle safety improvements $ G e n e ral p e d e stri an s afety i m p rove m e nts # Pedestrian crossings # Pedestrian overpasses/underpasses $ Pedestrian signals Figure C.3. ISECR Window: Countermeasure Types for Cyclist/Pedestrian Facilities Predict CR Factors (cont.) Select a Countermeasure: @ General $ Pave m e nt m arki n g s: g e n e ral # Pave rn e nt m arki n g s: e d g e I i n e s # Pavement markings: median edge line •S Pave rn e nt m arki n g s: ri g ht e d g e I i n e $ Rai s e d p ave m e nt m arke rs $ Reflected guide posts $ Strips: general # Strips: rumble # Strips: transverse Continue: Figure C.4. ISECR Window: Countermeasure Types for Delineation Pre-Implementation Evaluation of Safety Improvement Programs Appendix C: ISECR Windows: Countermeasure Types Considered for Different Countermeasure Categories 108 Predict CR Factors (cont.) Select a Countermeasure: 9 General # Alignment general # Alignment: horizontal # Alignment: vertical # Median: general # Median: close median openings # M e d i an: co n crete/atte n u ato r b arri e rs # M e d i an: i n stal I n ew median # Median: upgrade existing median W Median: widening existing median # Sight distance # Staggered intersection # Superelevation Figure C.5. ISECR Window: Countermeasure Types for Geometric Improvements Pre-Implementation Evaluation of Safety Improvement Programs Appendix C: ISECR Windows: Countermeasure Types Considered for Different Countermeasure Categories 109 Predict CR Factors (cont.] @ General # Channelization: general $ Channelization: left-turn lane: general # Channelization: left-turn lane with left-turn phase # Channelization: left-turn lane without left-turn phase # Channelization: right-turn lane # Tu rn i n g b ay/traff i c i s I an d $ Roundabouts: general ® Roundabouts: installation/upgrade Exit ! Continue Figure C.6. ISECR Window: Countermeasure Types for Intersection Improvements Pre-Implementation Evaluation of Safety Improvement Programs Appendix C: ISECR Windows: Countermeasure Types Considered for Different Countermeasure Categories 110 Predict CR Factors (cont.) Select a Countermeasure: <3 General $ 2-way left-turn lane # Acc. / Decel. / Passing lane W Bicycle lane % Bus lane # Climbing lane % Flattening sideslopes % HOVIarte # Lane/shoulder: general # Lane/shoulder: narrowing % Lane/shoulder: widening >§t Lane addition Figure C.7. ISECR Window: Countermeasure Types for Lane/Shoulder Treatment Predict CR Factors (cont.) Select a Countermeasure: © G e n e r a l # Install new lighting W Upgrade existing lighting Figure C.8. ISECR Window: Countermeasure Types for Lighting Improvements Pre-Implementation Evaluation of Safety Improvement Programs Appendix C: ISECR Windows: Countermeasure Types Considered for Different Countermeasure Categories 111 Predict CR Factors (cont.) Select a Countermeasure: ® General $ Relocation: general # Re I o cati o n: f ixe d o b j e cts $ Re I o cati o n: uti I ity poles $ Removal : general % Removal : fixed objects # Fie rn oval: trees # Re rn oval: utility poles Figure C.9. I S E C R Window: Countermeasure Types for Object Removal/Relocation Predict CR Factors (cont.) Select a Countermeasure: f & General # Ch an g e fro m an g I e- to p aral I e I p arki n g # Eliminate parking: general # Eliminate parking: parallel parking Figure C.10. I S E C R Window: Countermeasure Types for Parking Improvements Pre-Implementation Evaluation of Safety Improvement Programs Appendix C: ISECR Windows: Countermeasure Types Considered for Different Countermeasure Categories 112 Predict CR Factors (cont.) Select a Countermeasure: <9 General # Pave rn e nt grooving ® Resurfacing # Skid reduction Figure C . l l . I S E C R Window: Countermeasure Types for Pavement Treatment | Predict CR Factors (cont.) Select a Countermeasure: @ General # Automatic gates W Flashing beacons Figure C.12. I S E C R Window: Countermeasure Types for Railway Improvements Pre-Implementation Evaluation of Safety Improvement Programs Appendix C: ISECR Windows: Countermeasure Types Considered for Different Countermeasure Categories 113 Predict GR Factors (cont.) Select a Countermeasure: Figure C.13. ISECR Window: Countermeasure Types for Regulation Change Predict CR Factors (cont.) Select a Countermeasure: Figure C.14. ISECR Window: Countermeasure Types for Safety Barriers Pre-Implementation Evaluation of Safety Improvement Programs Appendix C: ISECR Windows: Countermeasure Types Considered for Different Countermeasure Categories 114 Predict CR Factors (cont.) Select a Countermeasure: f& General # Guidance signs # Install new signs/ jpgre.de existing signs $ Regulatory signs general ® Regulatory signs speed # Regulatory signs stop signs: general $ Regulatory signs stop signs: 2-way to 4^/ay stops $ Regulatory signs sto p s i g n s: 4-way sto p s W Regulatory signs stop signs: minor-leg stops # Regulatory signs yield signs