Examination of crash trends in the southeastern US : analysis of fatal crashes / prepared by Karen Dixon, Hong Zhu, Simon Washington, Kangwon Shin

Final Report
Examination of Crash Trends in the Southeastern US: Analysis of Fatal Crashes
GDOT Project 07-01
June 2009
Prepared for: Georgia Department of Transportation
Prepared by: Karen Dixon, Ph.D., P.E., Oregon State University
Hong Zhu, Oregon State University Simon Washington, Ph.D., Arizona State University
Kangwon Shin, Arizona State University
School of Civil and Construction Engineering Oregon State University Corvallis, OR 97331

This page blank intentionally

Table of Contents
Table of Contents ............................................................................................................................. i List of Figures ................................................................................................................................ iv List of Tables ................................................................................................................................ vii 1.0 Introduction and Background ............................................................................................. 1 2.0 Literature Review Summary ............................................................................................... 3
2.1 Driver and Passenger Related Characteristics ..................................................................... 3 2.1.1 Gender and Age ............................................................................................................ 4 2.1.2 Alcohol and Drug Use .................................................................................................. 8 2.1.3 Safety Restraints ........................................................................................................... 9 2.1.4 Seating Position .......................................................................................................... 12 2.1.5 Speeding...................................................................................................................... 12
2.2 Vehicle Related Characteristics ......................................................................................... 13 2.2.1 Vehicle Weight, Size, and Type ................................................................................. 14 2.2.2 Vehicle Occupancy ..................................................................................................... 15
2.3 Roadway and Roadside Related Characteristics................................................................ 16 2.3.1 Roadway Alignment and Grades ................................................................................ 18 2.3.2 Lane and Shoulder ...................................................................................................... 19 2.3.3 Roadside Characteristics ............................................................................................. 19 2.3.4 Speed Limit ................................................................................................................. 20 2.3.5 Wet versus Dry Pavement Conditions ........................................................................ 22 2.3.6 Traffic Volume............................................................................................................ 22
2.4 Crash Related Characteristics ............................................................................................ 23 2.4.1 Crash Type .................................................................................................................. 23 2.4.2 Number of Involved Vehicles ..................................................................................... 25
2.5 Environment Related Characteristics................................................................................. 25 2.5.1 Weather Conditions .................................................................................................... 26 2.5.2 Lighting Conditions .................................................................................................... 27 2.5.3 Urban versus Rural ..................................................................................................... 28
2.6 Literature Summary ........................................................................................................... 28 3.0 Summary Statistics of Study Crash Data .......................................................................... 31
3.1 Data Description ................................................................................................................. 31 3.2 Data Representation of Larger Population.......................................................................... 33 3.3 Descriptive Statistics........................................................................................................... 40
3.3.1 Crash Data Characteristics ........................................................................................... 40
3.3.1.1 Crash Distribution by Month ............................................................................................. 42 3.3.1.2 Crash Distribution by Day of Week................................................................................... 44
3.3.2 Roadway and Roadside Related Characteristics.......................................................... 46
3.3.2.1 Horizontal Alignment Direction and Curve Radius........................................................... 46 3.3.2.2 Vertical Grade .................................................................................................................... 49 3.3.2.3 Cross Section Configuration .............................................................................................. 51 3.3.2.4 National Highway System ................................................................................................ 53 3.3.2.5 Road Functional Classification .......................................................................................... 54 3.3.2.6 Lane Width......................................................................................................................... 55 3.3.2.7 Shoulder Type .................................................................................................................... 57
i

3.3.2.8 Auxiliary Lane Configuration ............................................................................................ 59 3.3.2.9 Road Surface Type............................................................................................................. 59 3.3.2.10 Average Daily Traffic ...................................................................................................... 61 3.3.2.11 Roadway Junction Proximity ........................................................................................... 62 3.3.2.12 Number of Driveways per Mile ....................................................................................... 64 3.3.2.13 Regulatory Speed Limit ................................................................................................... 66 3.3.2.14 Traffic Control Device ..................................................................................................... 68 3.3.2.15 Roadside Hazard Rating .................................................................................................. 70 3.3.2.16 Guardrail and Bridge Rail Type....................................................................................... 72 3.3.2.17 Terrain.............................................................................................................................. 72 3.3.2.18 Relation to Roadway ....................................................................................................... 74
3.3.3 Environment Related Characteristics........................................................................... 76
3.3.3.1 Ambient Light Condition ................................................................................................... 76 3.3.3.2 Weather Condition ............................................................................................................. 78 3.3.3.3 Road Surface Condition ..................................................................................................... 80
3.3.4 Driver and Passenger Related Characteristics ............................................................. 82
3.3.4.1 Gender................................................................................................................................ 82 3.3.4.2 Driver Age Group .............................................................................................................. 84 3.3.4.3 Ejection Status of Vehicle Occupants ................................................................................ 85 3.3.4.4 Occupant Protection System Use ...................................................................................... 87
3.3.5 Vehicle Related Characteristics ................................................................................... 89
3.3.5.1 Vehicle Configuration........................................................................................................ 89 3.3.5.2 Vehicle Travel Speed ......................................................................................................... 91 3.3.5.3 Vehicle Maneuver .............................................................................................................. 93 3.3.5.4 Extent of Damage Towing Status ................................................................................... 95 3.3.5.5 Vehicle Model Year ........................................................................................................... 95
4.0 Ten-Year Historic Fatal Crash Trends in the Southeastern United States........................ 99 4.1 Fatal Crash Trends ............................................................................................................. 99 4.1.1 Fatal Crash Frequency ................................................................................................ 99 4.1.2 Vehicle Miles Traveled............................................................................................. 100 4.1.3 Fatality Rates ............................................................................................................ 102 4.2 Potential Contributing Factors Affecting Fatal Crash Trends ......................................... 106 4.2.1 Temporal Factors ...................................................................................................... 106
4.2.1.1 Day of Week .................................................................................................................... 106 4.2.1.2 Time of Day .................................................................................................................... 110
4.2.2 Spatial Factors........................................................................................................... 112
4.2.2.1 Roadway Function Class................................................................................................. 112 4.2.2.2 Traffic Control ................................................................................................................ 117
4.2.3 Behavioral Factors .................................................................................................... 122
4.2.3.1 Alcohol Involvement ...................................................................................................... 122 4.2.3.2 Restraint Usage ............................................................................................................... 125
4.3 Discussion of Results ....................................................................................................... 130 5.0 Modeling Methodology and Strategy ............................................................................. 133
5.1 Regression Model for Crash Type Prediction................................................................... 133 5.1.1 Safety Predictive Models ........................................................................................... 133 5.1.2 Crash Type Prediction Model Application ................................................................ 134 5.1.3 Crash Types ............................................................................................................... 135
5.2 Binary Logit Models ......................................................................................................... 136 5.2.1 Single-vehicle vs. Multiple-vehicle Crash ................................................................. 137
ii

5.2.2 Head-on vs. Other Multiple-Vehicle Fatal Crashes ................................................... 139 5.3 Model Selection ................................................................................................................ 140 6.0 Crash Type Model Evaluation ........................................................................................ 141 6.1 Single-vehicle Run-off-road Fatal Crash .......................................................................... 141
6.1.1 Combined-State Models............................................................................................. 141
6.1.1.1 Four-State Model (AL, GA, MS, SC) .............................................................................. 141 6.1.1.2 Three-State Model (AL, GA, SC) .................................................................................... 146
6.1.2 Models by State......................................................................................................... 149
6.1.2.1 Alabama .......................................................................................................................... 150 6.1.2.2 Georgia............................................................................................................................ 152 6.1.2.3 Mississippi ...................................................................................................................... 155 6.1.2.4 South Carolina ................................................................................................................ 157
6.1.3 Summary of Single-Vehicle Fatal Crash Models ...................................................... 160 6.1.4 Analysis of Variables for Single-Vehicle Run-off-Road Crashes ............................. 163
6.1.4.1 Lane Width...................................................................................................................... 164 6.1.4.2 Paved and Graded Shoulder Width ................................................................................. 167 6.1.4.3 Roadside Condition......................................................................................................... 173 6.1.4.4 Horizontal and Vertical Alignment................................................................................. 175 6.1.4.5 Road Junction/Intersection.............................................................................................. 177 6.1.4.6 Land Use Type ................................................................................................................ 178 6.1.4.7 Time of Day .................................................................................................................... 179 6.1.4.8 Lighting Condition.......................................................................................................... 181
6.2 Head-on Fatal Crash ........................................................................................................ 181 6.2.1 Combined-State Models (AL, GA, MS, and SC) ...................................................... 181 6.2.2 Models by State.......................................................................................................... 185
6.2.2.1 Alabama .......................................................................................................................... 185 6.2.2.2 Georgia............................................................................................................................ 187 6.2.2.3 Mississippi ...................................................................................................................... 189
6.2.3 Summary of Head-on Fatal Crash Models................................................................. 191 6.2.4 Variable Analysis....................................................................................................... 193
6.2.4.1 Lane Width...................................................................................................................... 194 6.2.4.2 Curve Direction............................................................................................................... 198 6.2.4.3 Road Segment ................................................................................................................. 200 6.2.4.4 Number of Driveways..................................................................................................... 203 6.2.4.5 Restraint System (GA-only Model) ................................................................................ 204
7.0 Practical Applications of Crash Type Prediction Models................................................ 207 7.1 Application Methodology ................................................................................................ 207 7.2 Application Example ....................................................................................................... 210 7.3 Application Limitation..................................................................................................... 217
8.0 Expert Panel Findings: Rural Road Fatal Crash Countermeasures ............................... 219 8.1 Description of Expert Evaluation..................................................................................... 219 8.2 Cumulative Advice from the Experts .............................................................................. 221 8.3 Summary of Expert Opinion Findings ............................................................................. 225
9.0 Conclusions..................................................................................................................... 227 10.0 Reference List ................................................................................................................. 231 11.0 Appendix......................................................................................................................... 239
iii

List of Figures
Figure 1: Alabama Driver Population Distribution by Age Group.............................................. 33 Figure 2: Georgia Driver Population Distribution by Age Group ............................................... 34 Figure 3: Mississippi Driver Population Distribution by Age Group.......................................... 34 Figure 4: South Carolina Driver Population Distribution by Age Group .................................... 35 Figure 5: Alabama Male Driver Sample Distribution by Age Group.......................................... 36 Figure 6: Alabama Female Driver Sample Distribution by Age Group ...................................... 37 Figure 7: Georgia Male Driver Sample Distribution by Age Group ........................................... 37 Figure 8: Georgia Female Driver Sample Distribution by Age Group........................................ 38 Figure 9: Mississippi Male Driver Sample Distribution by Age Group ...................................... 38 Figure 10: Mississippi Female Driver Sample Distribution by Age Group ................................ 39 Figure 11: South Carolina Male Driver Sample Distribution by Age Group .............................. 39 Figure 12: South Carolina Female Driver Sample Distribution by Age Group .......................... 40 Figure 13: Single vs. Multiple-Vehicle Fatal Crashes ................................................................. 41 Figure 14: Fatal Crash Type Distribution .................................................................................... 42 Figure 15: Single-Vehicle Fatal Crashes by Month..................................................................... 43 Figure 16: Multiple-Vehicle Fatal Crashes by Month ................................................................. 44 Figure 17: Single-Vehicle Fatal Crashes by Day of Week .......................................................... 45 Figure 18: Multiple-Vehicle Fatal Crashes by Day of Week ...................................................... 46 Figure 19: Single-Vehicle Fatal Crashes and Associated Horizontal Alignment........................ 48 Figure 20: Multiple-Vehicle Fatal Crashes and Associated Horizontal Alignment .................... 48 Figure 21: Single-Vehicle Fatal Crashes and Associated Horizontal Curvature ......................... 49 Figure 22: Multiple-Vehicle Fatal Crashes and Associated Horizontal Curvature ..................... 49 Figure 23: Single-Vehicle Crashes by Vertical Grade................................................................. 51 Figure 24: Multiple-Vehicle Crashes by Vertical Grade ............................................................. 51 Figure 25: Single-Vehicle Crashes by Cross Section Configuration........................................... 52 Figure 26: Multiple-Vehicle Crashes by Cross Section Configuration ....................................... 53 Figure 27: Single-Vehicle Fatal Crashes by Road Functional Classification.............................. 55 Figure 28: Multiple-Vehicle Fatal Crashes by Road Functional Classification .......................... 55 Figure 29: Single-Vehicle Fatal Crashes by Lane Width ............................................................ 57 Figure 30: Multiple-Vehicle Fatal Crashes by Lane Width......................................................... 57 Figure 31: Single-Vehicle Fatal Crashes by Shoulder Type........................................................ 58 Figure 32: Multiple-Vehicle Fatal Crashes by Shoulder Type .................................................... 59 Figure 33: Single-Vehicle Fatal Crashes by Road Surface Material ........................................... 60 Figure 34: Multiple-Vehicle Fatal Crashes by Road Surface Material........................................ 61 Figure 35: Single-Vehicle Fatal Crashes by Average Daily Traffic............................................ 61 Figure 36: Multiple-Vehicle Fatal Crashes by Average Daily Traffic ........................................ 62 Figure 37: Single-Vehicle Crashes by Roadway Junction Proximity.......................................... 63 Figure 38: Multiple-Vehicle Crashes by Roadway Junction Proximity ...................................... 64 Figure 39: Single-Vehicle Fatal Crashes for Number of Driveways per Mile ............................ 65 Figure 40: Multiple-Vehicle Fatal Crashes for Number of Driveways per Mile......................... 66 Figure 41: Single-Vehicle Fatal Crashes per Regulatory Speed Limit........................................ 67 Figure 42: Multiple-Vehicle Fatal Crashes per Regulatory Speed Limit .................................... 68 Figure 43: Single-Vehicle Fatal Crashes and Associated Traffic Control Devices..................... 69
iv

Figure 44: Multiple-Vehicle Fatal Crashes and Associated Traffic Control Devices ................. 70 Figure 45: Single-Vehicle Fatal Crashes and Associated Roadside Hazard Ratings .................. 71 Figure 46: Multiple-Vehicle Fatal Crashes and Associated Roadside Hazard Ratings............... 72 Figure 47: Single-Vehicle Fatal Crashes and Associated Terrain ............................................... 73 Figure 48: Multiple-Vehicle Fatal Crashes and Associated Terrain............................................ 74 Figure 49: Single-Vehicle Fatal Crashes and Relation to Roadway............................................ 75 Figure 50: Multiple-Vehicle Fatal Crashes and Relation to Roadway ......................................... 76 Figure 51: Single-Vehicle Fatal Crashes and Ambient Lighting Conditions .............................. 77 Figure 52: Multiple-Vehicle Fatal Crashes and Ambient Light Conditions................................ 78 Figure 53: Single-Vehicle Crashes by Weather Conditions ........................................................ 79 Figure 54: Multiple-Vehicle Crashes by Weather Conditions..................................................... 80 Figure 55: Single-Vehicle Crashes and Associated Road Surface Conditions............................ 81 Figure 56: Multiple-Vehicle Crashes and Associated Road Surface Conditions ........................ 82 Figure 57: Single-Vehicle Fatal Crashes by Driver and Passenger Genders............................... 83 Figure 58: Multiple-Vehicle Fatal Crashes by Driver and Passenger Genders ........................... 84 Figure 59: Single-Vehicle Fatal Crashes by Driver Age Group .................................................. 85 Figure 60: Multiple-Vehicle Fatal Crashes by Driver Age Group .............................................. 85 Figure 61: Single-Vehicle Fatal Crashes and Associated Occupant Ejection Status................... 86 Figure 62: Multiple-Vehicle Fatal Crashes and Associated Occupant Ejection Status ............... 87 Figure 63: Single-Vehicle Fatal Crashes and Occupant Protection System Use ......................... 88 Figure 64: Multiple-Vehicle Fatal Crashes and Occupant Protection System Use ..................... 89 Figure 65: Single-Vehicle Fatal Crashes and Associated Vehicles ............................................. 90 Figure 66: Multiple-Vehicle Fatal Crashes and Associated Vehicles ......................................... 91 Figure 67: Single-Vehicle Fatal Crashes by Vehicle Travel Speed............................................. 92 Figure 68: Multiple-Vehicle Fatal Crashes by Vehicle Travel Speed ......................................... 93 Figure 69: Single-Vehicle Fatal Crashes and Associated Vehicle Maneuver ............................. 94 Figure 70: Multiple-Vehicle Fatal Crashes and Associated Vehicle Maneuver.......................... 94 Figure 71: Single-Vehicle Fatal Crashes and Associated Vehicle Model Year .......................... 96 Figure 72: Multiple-Vehicle Fatal Crashes and Associated Vehicle Model Year....................... 97 Figure 73: Fatal Crash Trend in the Southeastern US (1997 to 2006)........................................ 100 Figure 74: VMT (Billions) Trend in the Southeastern US (1997 to 2006)................................ 101 Figure 75: The Relationship between VMT and Fatal Crashes (Lowess Curve) ....................... 102 Figure 76: Fatality Rate per 100 Million VMT by State and Year ............................................. 104 Figure 77: Change in Fatality Rate per 100 Million VMT from 1997 to 2006 .......................... 105 Figure 78: Fatal Crashes by State and Day of Week .................................................................. 108 Figure 79: Fatal Crashes by State and Day of Week (1997 to 2006) ......................................... 109 Figure 80: Percent of Fatal Crashes by State (1997 to 2006) ..................................................... 114 Figure 81: Change in Fatal Crashes by Roadway Class and State (1997 to 2006)..................... 116 Figure 82: Portion of Fatal Crashes by State and Traffic Control (1997 to 2006) ..................... 119 Figure 83: Change in Fatal Crashes by Traffic Control and State (1997 to 2006) ..................... 121 Figure 84: Fatal Crashes by State and Alcohol-involvement ..................................................... 123 Figure 85: Change in the Proportion of Alcohol-impaired Fatal Crashes (1997 to 2006) ......... 125 Figure 86: Drivers in Fatal Crashes by State and Restraint Usage (1997 to 2006) .................... 127 Figure 87: Change in the Portion of Drivers in Fatal Crashes by Restraint Usage and Year ..... 129 Figure 88: Fatal Crash Type Classification ............................................................................... 136
v

Figure 89: Dark without Street Lights -- Lane width by ADT (Three-State Model, Single-Vehicle) ................................................................................................................... 165
Figure 90: Daylight, Dark with Lighting, Dusk, or Dawn -- Lane width by ADT (Three-State Model, Single-Vehicle).................................................................................. 166
Figure 91: Dark without Street Lights -- Graded and Paved Shoulder Width (Three-State Model, Single-Vehicle).................................................................................. 168
Figure 92: Daylight, Dark with Lights, Dusk, or Dawn -- Graded and Paved Shoulder Width (Three-State Model, Single-Vehicle)....................................................................... 168
Figure 93: Dark without Street Lights -- Paved Shoulder Width and Safety Restraint Used (GA only Model, Single-Vehicle) ............................................................................. 170
Figure 94: Daylight, Dark with Lights, Dusk, or Dawn -- Paved Shoulder Width and Safety Restraint Used (GA only Model, Single-Vehicle)................................................... 170
Figure 95: Dark without Street Lights -- Paved Shoulder Width and Safety Restraint Not Used (GA only Model, Single-Vehicle) ...................................................................... 171
Figure 96: Daylight, Dark with Lights, Dusk, or Dawn -- Paved Shoulder Width and Safety Restraint Not Used (GA only Model, Single-Vehicle)............................................ 172
Figure 97: Dark without Street Lights -- Paved Shoulder Width (Three-State Model, Single-Vehicle) ................................................................................................................... 172
Figure 98: Daylight, Dark with Lights, Dusk, or Dawn -- Paved Shoulder Width (Three-State Model, Single-Vehicle).................................................................................. 173
Figure 99: Roadside Hazard Rating (Three-State Model, Single-Vehicle) ............................... 174 Figure 100: Dark without Street Lights -- Road Alignment (Three-State Model,
Single-Vehicle) ................................................................................................................... 176 Figure 101: Daylight, Dark with Lights, Dusk, or Dawn -- Road Alignment
(Three-State Model, Single-Vehicle).................................................................................. 176 Figure 102: Road Junction (Three-State Model, Single-Vehicle Crash) ................................... 178 Figure 103: Land Use Type (Three-State Model, Single-Vehicle)............................................ 179 Figure 104: Time of Crash (Three-State Model, Single-Vehicle) .............................................. 180 Figure 105: Lane Width Crash Probability by ADT (Four-State Model, Head-on) .................. 195 Figure 106: Lane Width by ADT and Used Safety Restraints (GA only Model, Head-on)...... 197 Figure 107: Lane Width by ADT and No Safety Restraints (GA only Model, Head-on) ......... 197 Figure 108: Curve Direction (Four-State Model, Head-on) ...................................................... 198 Figure 109: Curve Direction and Used Safety Restraints (GA only Model, Head-on) ............. 199 Figure 110: Curve Direction and No Safety Restraints (GA only Model, Head-on) ................ 200 Figure 111: Road Junction (Four-State Model, Head-on) ......................................................... 201 Figure 112: Road Junction and Used Safety Restraints (GA only Model, Head-on) ................ 202 Figure 113: Road Junction and No Safety Restraints (GA only Model, Head-on) ................... 203 Figure 114: Number of Driveways (Four-State Model, Head-on) ............................................ 204 Figure 115: Restraint System (GA only Model, Head-on) ........................................................ 205 Figure 116: Single-Vehicle Fatal Crash Type Model Application Six-Step Procedure ............ 209 Figure 117: Safety Evaluation for Plan B1 and B2.................................................................... 215
vi

List of Tables
Table 1: Summary of Driver and Passenger Related Characteristics Literature ........................... 4 Table 2: Summary of Vehicle Related Characteristics Literature .............................................. 14 Table 3: Summary of Roadway and Roadside Related Characteristic Literature ...................... 17 Table 4: Summary of Crash Related Characteristic Literature .................................................... 23 Table 5: Summary of Environment Related Characteristic Literature ....................................... 26 Table 6: Pearson's Chi-Square Test Results of Driver Distribution............................................ 40 Table 7: Crash Type Distribution for Study Crashes................................................................... 41 Table 8: Crash Distribution by Month ......................................................................................... 43 Table 9: Crash Distribution by Day of Week .............................................................................. 45 Table 10: Crash Type by Horizontal Alignment Direction and Horizontal Curvature................ 47 Table 11: Crash Type by Vertical Grade ..................................................................................... 50 Table 12: Cross Section Configuration by Crash Type ............................................................... 52 Table 13: Crash Occurrence on National Highway System ........................................................ 53 Table 14: Crash Occurrence per Road Functional Classification ................................................ 54 Table 15: Crash Type by Lane Width Distribution ..................................................................... 56 Table 16: Crashes by Shoulder Type ........................................................................................... 58 Table 17: Crashes by Auxiliary Lane Configuration ................................................................... 59 Table 18: Crash Type by Road Surface Material......................................................................... 60 Table 19: Crash Type based on Roadway Junction Proximity .................................................... 63 Table 20: Crash Type by Number of Driveways per Mile .......................................................... 65 Table 21: Crash Type per Regulatory Speed Limit ..................................................................... 67 Table 22: Crash Type and Associated Traffic Control Devices .................................................. 69 Table 23: Crash Type and Associated Roadside Hazard Rating ................................................. 71 Table 24: Fatal Crashes and Associated Guardrail/Bridge Rails................................................. 72 Table 25: Crash Type and Associated Terrain............................................................................. 73 Table 26: Crash Type and Relation to Roadway ......................................................................... 75 Table 27: Crash Type and Ambient Light Conditions................................................................. 77 Table 28: Crash Type by Weather Conditions............................................................................. 79 Table 29: Crash Type and Associated Road Surface Conditions ................................................ 81 Table 30: Crash Type by Driver and Passenger Gender.............................................................. 83 Table 31: Crash Type by Driver Age Group ............................................................................... 84 Table 32: Crash Type and Associated Occupant Ejection Status ................................................ 86 Table 33: Crash Type Compared to Occupant Protection System Use ....................................... 88 Table 34: Crash Type and Associated Vehicles .......................................................................... 90 Table 35: Crash Type by Vehicle Travel Speed .......................................................................... 92 Table 36: Crash Type and Associated Vehicle Maneuver........................................................... 93 Table 37: Crash Type and Associated Vehicle Condition (Towing Status) ................................ 95 Table 38: Crash Type and Associated Vehicle Model Year........................................................ 96 Table 39: Fatal Crash Frequency from 1997 to 2006 in the Southeastern US ........................... 100 Table 40: Vehicle Miles Traveled (Billions) in the Southeastern US (1997 to 2006)................ 101 Table 41: Fatality Rate per 100 Million VMT by State and Year .............................................. 103 Table 42: Percent Change in Fatality Rate per 100 Million VMT (1997 to 2006)..................... 105 Table 43: Fatal Crashes by State and Day of Week.................................................................... 108 Table 44: Change in Fatal Crashes between 1997 and 2006 by Day of Week and State ........... 110
vii

Table 45: Fatal Crashes by State and Time of Day (1997 to 2006)............................................ 111 Table 46: Change in Fatal Crashes between 1997 and 2006 by Time of Day and State ............ 112 Table 47: Roadway Functional Class Code (used in FARS) ...................................................... 113 Table 48: Simplified Roadway Functional Class........................................................................ 113 Table 49: Fatal Crashes by State and Roadway Functional Class (1997 to 2006) ..................... 113 Table 50: Change in Fatal Crashes by Roadway Class and State (1997 to 2006) ...................... 115 Table 51: Traffic Control Description ........................................................................................ 118 Table 52: Fatal Crashes by State and Traffic Control (1997 to 2006)........................................ 119 Table 53: Change in Fatal Crashes by Traffic Control and State (1997 to 2006) ...................... 120 Table 54: Fatal Crashes by State and Alcohol Involvement (1997 to 2006) .............................. 123 Table 55: Change in Fatal Crashes by Alcohol-involvement and State (1997 to 2006)............. 124 Table 56: Restraint Usage Description ....................................................................................... 126 Table 57: Drivers in Fatal Crashes by State and Restraint Usage (1997 to 2006)...................... 127 Table 58: Change in Proportion of Drivers in Fatal Crashes (1997 and 2006) .......................... 128 Table 59: Relative Rank of Safety Performance in 2006 of four Southeastern States ............... 132 Table 60: Variable Description (Combined-State Models, Single-Vehicle) ............................. 142 Table 61: Continuous Variable Descriptive Statistics (Four-State Model, Single-Vehicle) ..... 142 Table 62: Distribution of Categorical Variables (Four-State Model, Single-Vehicle) .............. 143 Table 63: Model Estimation (Four-State Model, Single-Vehicle)............................................. 145 Table 64: Continuous Variable Descriptive Statistics (Three-State Model, Single-Vehicle) ... 146 Table 65: Distribution of Categorical Variables (Three-State Model, Single-Vehicle) ............ 147 Table 66: Model Estimation (Three-State Model, Single-Vehicle)........................................... 148 Table 67: Variable Description (AL only Model, Single-Vehicle) ........................................... 151 Table 68: Continuous Variable Descriptive Statistics (AL only Model, Single-Vehicle)......... 151 Table 69: Distribution of Categorical Variables (AL only, Single-Vehicle)............................. 151 Table 70: Model Estimation (AL only Model, Single-Vehicle) ................................................ 152 Table 71: Variable Description (GA only Model, Single-Vehicle) ........................................... 153 Table 72: Continuous Variable Descriptive Statistics (GA only Model, Single-Vehicle) ........ 153 Table 73: Distribution of Categorical Variables (GA only Model, Single-Vehicle) ................. 154 Table 74: Model Estimation (GA only Model, Single-Vehicle)................................................ 154 Table 75: Variable Description (MS only Model, Single-Vehicle) ........................................... 155 Table 76: Continuous Variable Descriptive Statistics (MS only Model, Single-Vehicle) ........ 155 Table 77: Distribution of Categorical Variables (MS only Model, Single-Vehicle) ................. 156 Table 78: Model Estimation (MS only Model, Single-Vehicle)................................................ 157 Table 79: Variable Description (SC only Model, Single-Vehicle)............................................ 158 Table 80: Continuous Variable Descriptive Statistics (SC only Model, Single-Vehicle) ......... 158 Table 81: Distribution of Categorical Variables (SC only Model, Single-Vehicle).................. 158 Table 82: Model Estimation: Single-Vehicle Fatal Crash Model (SC only Model,
Single-Vehicle) ................................................................................................................... 159 Table 83: Model Comparison (Single-Vehicle)......................................................................... 161 Table 84: Illustration of Effects (Single-Vehicle) ..................................................................... 162 Table 85: Description of Road Nominal Condition for Evaluating Single-Vehicle Models ..... 164 Table 86: Crash Probabilities based on Lane Width and ADT (Three-State Model,
Single-Vehicle) ................................................................................................................... 165 Table 87: Crash Probability for Paved and Graded Shoulder Width (Three-State Model,
Single-Vehicle) ................................................................................................................... 167
viii

Table 88: Crash Probability for Paved Shoulder and Restraint System Usage (GA only Model, Single-Vehicle)....................................................................................... 169
Table 89: Roadside Hazard Rating (Three-State Model, Single-Vehicle) ................................ 174 Table 90: Sensitivity to Curve Direction and Vertical Alignment (Three-State
Model, Single-Vehicle)....................................................................................................... 175 Table 91: Sensitivity to Road Junction/Intersection (Three-State Model, Single-Vehicle) ...... 177 Table 92: Sensitivity to Land Use Types (Three-State Model, Single-Vehicle) ....................... 179 Table 93: Sensitivity to Time of Crash (Three-State Model, Single-Vehicle) .......................... 180 Table 94: Variable Description (Four-State Model, Head-on) .................................................. 182 Table 95: Continuous Variables Descriptive Statistics (Four-State Model, Head-on) .............. 182 Table 96: Distribution of Categorical Variables (Four-State Model, Head-on) ........................ 183 Table 97: Model Estimation (Four-State Model, Head-on)....................................................... 183 Table 98: Variable Description (AL only Model, Head-on)...................................................... 185 Table 99: Continuous Variable Descriptive Statistics (AL only Model, Head-on) ................... 185 Table 100: Distribution of Categorical Variables (AL only Model, Head-on).......................... 186 Table 101: Model Estimation (AL only Model, Head-on) ........................................................ 186 Table 102: Variable Description (GA only Model, Head-on) ................................................... 187 Table 103: Continuous Variable Descriptive Statistics (GA only Model, Head-on) ................ 187 Table 104: Distribution of Categorical Variables (GA only Model, Head-on) ......................... 188 Table 105: Model Estimation (GA only Model, Head-on)........................................................ 188 Table 106: Variable Description (MS only Model, Head-on) ................................................... 189 Table 107: Continuous Variable Descriptive Statistics (MS only Model, Head-on) ................ 190 Table 108: Distribution of Categorical Variables (MS only Model, Head-on) ......................... 190 Table 109: Model Estimation (MS only Model, Head-on)........................................................ 191 Table 110: Model Comparison (Head-on) ................................................................................. 192 Table 111: Illustration of Effects (Head-On)............................................................................. 193 Table 112: Description of Road Nominal Conditions (Head-on) .............................................. 194 Table 113: Crash Probability for Lane Width (Four-State Model, Head-on)............................ 195 Table 114: Crash Probability for Lane Width (GA Only Model, Head-on) .............................. 196 Table 115: Crash Probability for Curve Direction (Four-State Model, Head-on) ..................... 198 Table 116: Crash Probability for Curve Direction (GA only Model, Head-on) ........................ 199 Table 117: Crash Probability for Road Junction (Four-State Model, Head-on)........................ 201 Table 118: Crash Probability for Road Junction (GA only Model, Head-on)........................... 202 Table 119: Number of Driveways (Four-State Model, Head-on).............................................. 203 Table 120: Crash Probability for Restraint System Use (GA only Model, Head-on) ............... 205 Table 121: Countermeasure Categories vs. Model Significant Effects ..................................... 208 Table 122: Sample Problem -- Existing Road Conditions for Georgia Site .............................. 210 Table 123: Existing Condition and Proposed Improvement Plan.............................................. 213 Table 124: Safety Evaluation (Three-State Model, Single-Vehicle) ......................................... 214 Table 125: Safety Evaluation Used Safety Restraint (GA only Model, Single-Vehicle) ....... 214 Table 126: Safety Evaluation No Safety Restraint (GA only Model, Single-Vehicle) .......... 214 Table 127: Countermeasures examined in Southeastern US Fatal Crash Study ........................ 220 Table 128: Expert Opinions on the Effectiveness of 12 Fatal Crash Countermeasures ............. 222 Table 129: Statistics on the Expert Derived Countermeasure Effectiveness.............................. 224
ix

This page blank intentionally x

1.0 Introduction and Background
In October 2005, the Georgia Department of Transportation (GDOT) and researchers with the Georgia Institute of Technology completed a summary report that identified a series of rural two-lane road safety assessments for states in the southeastern United States. Included in this report was a summary of findings for state-based research efforts including completed research efforts in Alabama, Florida, Georgia, Kentucky, Mississippi, North Carolina, and South Carolina. Each of these efforts were performed by research teams within the designated state, but the collected crash data for Alabama, Georgia, Mississippi, North Carolina, and South Carolina was also compiled into a similar database format and provided to researchers for GDOT so that a cross-sectional comparison could be performed. Though the data quality and sample size varied somewhat for each state, the overall data can be analyzed to determine the regional differences in crash conditions and how the individual states may benefit from this information. Data for the state of North Carolina did not include all of the key variables present in the crash data for the other four states and, as a result, ultimately was not included in the subsequent final cross-sectional analysis.
The information in this document is an initial task in evaluating the southeast crash data. Specifically, Chapter 2 of this report identifies published research for the evaluation of cross-sectional crash data and how this information may be applicable to an additional assessment of the southeast crash database. Chapter 3 provides an overview of the available crash information and various statistical representations of the crash data available from the previous study. Since the crash data available from the previous study is now over ten years old, Chapter 4 includes an evaluation of a ten-year historic crash trend for the southeastern states of Alabama, Georgia, Mississippi, and South Carolina. This trend analysis enabled the research team to determine if crash trends have dramatically changed since the initial data collection effort.
1

This report includes a cross-sectional analysis of the crash data to help illuminate candidate treatments that can enhance road safety for rural two-lane highways. Chapter 5 introduces the methodology used for this analysis and focuses this analysis on crash types as these predictive models produced the most meaningful results. Chapter 6 then reviews the specific crash type models, while Chapter 7 presents a practical application example of the crash type models. Chapter 8 further reviews the findings from the expert panel assessments performed as part of the earlier Georgia study so as to provide crash modification factors as a supplemental source of information for Georgia crashes. Finally, Chapter 9 summarizes the conclusions for this study. Chapters 10 and 11 include References and the Appendix, respectively.
2

2.0 Literature Review Summary
This chapter provides a review of previous studies and findings of the impact on crash injury severity as well as association with crash types for various factors that are related to driving behavior, vehicle occupancy, vehicle related attributes, roadway design features, roadside characteristics, and environmental conditions. The literature review includes national and international research from the past two decades.
Sections 2.1 through 2.5 introduce recent research for potential contributing crash factors. Some of the factors, such as driver and vehicle related attributes, have been well researched, while others have not yet been explored in depth. For most well studied variables, the reported findings are relatively consistent. Studies with limited data or analysis, however, often provided conflicting results. In addition to the study efforts across potential contributing factors, a majority of the studies focused on the impact on crash injury severities, with relatively less research investigating the corresponding associations with crash types.
2.1 Driver and Passenger Related Characteristics One of the crash impact factor categories that has been under extensive study in the published literature is vehicle occupant related attributes. Numerous previous studies identified vehicle occupant related characteristics as influential factors on crash injury severity and crash type. Common human-related factors include gender, age group, alcohol and drug use, the presence and use of a safety restraint system, passenger seating position, and driver speed choice. Table 1 provides an overview of the published driver and passenger characteristics further summarized in Sections 2.1.1 through 2.1.5 of this summary.
3

Table 1: Summary of Driver and Passenger Related Characteristics Literature

Attributes
Gender and Age
Alcohol and Drug Use
Safety Restraints
Seating Positions
Speeding

Severity

Crash Type

Increase injury severity for

Older drivers are prone to more

female and older drivers

serious injury in side-impact

Dependant on other associated factors
Younger drivers associated with higher severity

crashes Insignificant gender association
with injury severity of older drivers in fixed-object crashes

Inconsistent findings in studies

Increase injury severity consistently
Interactions with gender and age groups

Less severe injury for older drivers in fixed-object crashes suggesting abusive use more common for younger drivers

Reduce injury severity effectively

Reduce injury severity in single-vehicle crashes, twovehicle crashes, and fixed object crashes

Greater risk for occupants seated on side of impact
Greater risk of head and neck

Most frequently an issue for two-vehicle side-impact crashes

injuries for elder drivers on far

side of vehicle



Increase injury severity consistently with speed increase High speed differentials increase severity risk



Speed related fixed-object crashes frequently associated with severe injuries

2.1.1 Gender and Age Researchers have performed extensive investigations on the relationship between injury severity level and vehicle occupants' gender and age. Most studies demonstrated significant evidence of the gender and age effects on crash injury severity after accounting for other influential analysis variables (Ulfarsson and Mannering, 2004; O'Donnell and Connor, 1996; Kweon and Kockelman, 2003; Farmer et al., 1997; Bedard et al., 2002; Khattak et al., 2002; Savolainen and Mannering, 2007; Jones and Jorgensen, 2003; Abdel-Aty et al., 1998). The well known physical and psychological differences
4

among gender and age provide logical support for the reported significant impacts on crash injury severity; however, a few studies provided unexpected results (Dissanayake and Lu, 2002; Kim et al., 1995).
Most studies reported a higher risk of serious injury associated with female drivers, as well as older and younger drivers, if a crash occurred. Ulfarsson and Mannering (2004) distinguished the injury severity differences for male and female drivers involved in crashes categorized by vehicle types, including the sport-utility vehicle, minivan, pickup, and passenger car. They found that there is a significant gender difference of how other factors influence injury severities. Abdelwahab and Abdel-Aty (2001) found that female drivers are more likely to suffer from severe injuries than male drivers based on a study of two-vehicle crashes that occurred at signalized intersections. Abdel-Aty et al. (1998) studied the relationship between driver age and injury severity and reported an increasing risk of involvement in crashes for older and younger driver groups. Srinivansan (2002) found that older drivers as well as female drivers face greater risk of mild injury while involved in crashes. The study did not find significant differences of gender and age for severe injuries.
Considering that rollover crashes tend to lead to more severe injuries than non-rollover crashes, Kweon and Kockelman (2003) classified crash data for rollover and nonrollover crashes and incorporated traffic exposure in the analysis. Their study presented a statistically significant association between injury severity and driver's age and gender. The researchers emphasized the importance of incorporating traffic exposure across driver groups stratified by age and gender. The major findings included:
Younger drivers (younger than age 20) tend to be associated with higher injury severity risk than middle-aged (20 to 60 years of age) or older drivers (above 60 years old).
Female drivers tend to sustain more severe injuries in crashes than male drivers.
5

Based on their findings, Kweon and Kockelman recommended providing more restricted driving policies for younger persons through various approaches, such as raising the legal driving age and prohibiting freeway driving.
As for crash type classification, Farmer et al. (1997) analyzed the relationships of occupant, vehicle, and crash characteristics to injury severities for side-impact crashes. The researchers found that older vehicle occupants were three times as likely as younger occupants to be seriously injured in similar crashes.
Several other international studies also reported consistent findings. Based on a 10-year crash database from Norway, Jones and Jorgensen (2003) reported an increasing risk of fatality for older and female drivers. In Australia, O'Donnell et al. (1996) determined that female drivers are more likely to have higher serious injury probabilities than male drivers involved in crashes. Ryan et al. (1998) investigated the potential of driver's age related patterns with crash type. They found that the younger driver group is more likely to be involved in single-vehicle crashes, the middle age driver group has a greater proportion of same direction crashes, and the older age driver group is over-represented in the angle crash category.
While most of the studies focused on crash severity, Bedard et al. (2002) investigated the effects of driver, crash, and vehicle characteristics on fatalities. The model that was estimated was based on data from the Fatal Analysis Reporting System (FARS) and predicted the likelihood of involvement in fatal crashes will increase as the driver's age increases. For example, drivers older than age 80 are about five times as likely to have fatal injuries in a crash as drivers aged 40-49 years old. Female drivers are about 1.5 times more likely to sustain fatal injuries compared to male drivers.
Aiming to address safety issues of the increasingly older driver population in the U.S., Khattak et al. (2002) investigated influential factors that are associated with severe crash injuries to older drivers (age above 65 years). The study found that an increase in the age of older drivers tends to increase the likelihood of having more severe injuries. The male
6

older drivers are more likely to be injured compared to female older drivers. This finding appears to conflict with the previous findings based on drivers from all age groups (O'Donnell and Connor, 1996; Ulfarsson and Mannering, 2004; Kweon and Kockelman, 2003, and Wang and Kockelman, 2005). A study from Hong Kong had a similar finding: male drivers faced a greater risk of involvement in serious or fatal crashes (Yau et al., 2006).
In general, researchers suggested that several physical deteriorations associated with the aging process increase the vulnerability of older drivers on the road and contribute to the increased injury level witnessed in the older driver population. Some commonly mentioned factors include declining visual functions, cognitive impairment, and increasing difficulty of focus.
While most of the studies focused on the injury level of motor vehicle occupants involved in a traffic crash, researchers also found consistent evidence of the strong relationship between age and injury severity for motorcyclists. The older the motorcyclists, the more likely they are to sustain more severe injuries in crashes (Savolainen and Mannering, 2007).
Even though most of these human-related studies showed evidence of statistically significant gender and age differences on injury severity for vehicle occupants and motorcyclists, a few additional studies reported inconsistent findings. Some researchers investigated potential influential factors on injury severity for older drivers by focusing on fixed-object passenger car crashes. In this case, Dissanayake and Lu (2002) did not identify the gender of older drivers as a statistically significant factor associated with injury severity. Kim et al. (1995) studied the relationships of driver characteristics and behavior with crash and injury severity and did not find statistically significant effects of driver's age and gender.
Overall, the majority of previous research reached consistent findings on the significant effects of age and gender on injury severity and crash types. These studies showed strong
7

evidence of the importance and necessity of a gender based injury severity study while taking into account the vehicle occupants' age groups. Many authors recommended taking into consideration the interactions of age and gender with other factors in future research on various safety issues.
Even though researchers focused on the effects of driver related attributes, select studies incorporated either passenger characteristics in the study or treated vehicle occupants as a whole to present a full picture of crash injury severity outcomes for all individuals involved in crashes. It is a common understanding that drivers and passengers have different influences on crash occurrence. Thus, it may be meaningful to explore and compare the similarities and differences of effects from driver and passenger related characteristics on injury severity level and crash type.
2.1.2 Alcohol and Drug Use
Previous studies of vehicle crash injury severity predominantly provided consistent evidence that alcohol use is one of the contributing factors leading to crashes and severe injuries (Kim et al., 1995; Bedard et al., 2002; Khattak et al., 2002; Traynor, 2005; Krull et al., 2000; Keall et al., 2004; Dissanayake and Lu, 2002; Duncan et al., 1998; Jones and Jorgensen, 2003; Shibata and Fukuda, 1994; Srinivasan, 2002).
Kim et al. (1995) investigated the relationships of driver characteristics and behavior with crash injury severity. The study showed that the driver's alcohol and drug use dramatically increases the odds of having a more severe crash injury. Bedard et al. (2002) presented evidence indicating that drivers with blood alcohol concentration (BAC) greater than 0.3 g/dl will triple their likelihood of fatal injuries. Aiming to address safety issues regarding the increasingly older driver population in the United States, Khattak et al. (2002) investigated influential factors associated with severe crash injuries for drivers over the age of 65. The study found that alcohol consumption will increase the probability of being seriously injured in a crash. Traynor (2005) studied the impact of driver alcohol use on crash severity and found that alcohol consumption by the driver atfault not only increases the likelihood of injury severity, but also increases the number of
8

injuries and fatalities per crash. Krull et al. (2000) also reported an increasing trend of driver injury severity with alcohol use.
International studies also reported similar results. A fatal injury study from New Zealand reported that the increasing BAC exponentially increases the risk for drivers of experiencing fatal injuries while driving at night (Keall et al., 2004). Based on crash data from Norway, Jones and Jorgensen (2003) reported an increasing trend of fatalities associated with alcohol use. Shibata and Fukuda (1994) also had similar findings based on a crash severity study in Japan.
While focusing on fixed-object passenger car crashes, Dissanayake and Lu (2002) reported an inconsistent result. Finding that older drivers under the influence of drugs or alcohol have less likelihood of severe injuries, they offered the explanation that older drivers may pay extra attention to driving while under the influence of alcohol or drugs than drivers in other age groups.
2.1.3 Safety Restraints
Previous researchers have recognized the important role of safety restraints in reducing crash severity. These studies primarily focused on safety restraint usage for motor vehicle occupants as well as helmet use for motorcyclists. Kim et al. (1995) investigated the relationship between driver characteristics and behavior with crash and injury severity. They found that drivers who did not use safety restraints significantly increased their chance of experiencing more severe crashes and injuries. Wang and Kockelman (2005) also reported dramatic reductions in the probability of being injured (36.5%) or killed (47.3%) for vehicle occupants who wear safety restraints compared to those who do not use safety restraints and are involved in a crash with two vehicles. In singlevehicle crashes, those who wear safety restraints face reductions of 90.2% for injury and 71.9% for fatality compared with those who do not. In a study concentrating on older drivers involved in fixed-object passenger car crashes, Dissanayake and Lu (2002) presented similar findings for the effectiveness of safety restraint usage on injury severity reduction. Abdelwahab and Abdel-Aty (2001) studied the two-vehicle crashes that
9

occurred at signalized intersections and reported that drivers wearing safety restraints reduced the likelihood of any occupant having severe injuries. After taking into account other factors, Krull et al. (2000) and Srinivasan (2002) also found an increasing trend of driver injury severity associated with those who failed to use safety restraints. As for motorcyclists' safety concerns, more severe injuries are associated with motorcyclists who do not wear helmets than those who do (Savolainen and Mannering, 2007).
While many studies confirm the effectiveness of safety restraint usage on crash and injury severity reduction, some point out the role of selective recruitment, also called sampling bias, in the evaluation process (Evans, 1996; Derrig el al., 2002; Nakahara et al., 2006). Drivers who have more safety awareness will be more likely to wear safety restraints; drivers who are more aggressive and less safety conscious will be more likely not to wear safety restraints, even in locations with primary and secondary seat belt laws. This second group, taken as a whole, includes the drivers who are more likely to be involved in more severe crashes and injuries. Therefore, selective recruitment may discount the alleged effectiveness of safety restraint usage for assessment studies.
Evans (1996) challenged the study method by measuring the safety restraint effectiveness with overall crash reduction. He recognized that the better measure of the effectiveness of safety restraint usage is the reduction of injury severity. The study showed that using safety restraints can more efficiently prevent fatalities than for less severe injuries in crashes. Evans also indicated that the drivers who do not wear safety restraints are riskier drivers, and that the safety restraint wearing rate dropped as the injury severity increased. The drivers who should benefit the most from safety restraints are more likely not to wear them. The study advocated reinforcing drivers' education program, especially targeting the aggressive driver population.
Using the FARS database from 1983 to 1996, Derrig et al. (2002) studied the impact on fatalities from the overall increasing safety restraint usage rate. The study showed that the increasing rate of safety restraint usage had little impact on reducing the fatality rate (fatalities per million populations and per 10 billion vehicle miles traveled). Primary
10

enforcement laws effectively increased the overall safety restraint usage rate; however, non-aggressive drivers tend to adhere to the law better than riskier drivers. The riskier drivers, however, are the population who are most likely to be involved in fatal crashes. The study by Derrig et al. (2002) provided analysis results consistent with those found by Evans (1996).
A recent study echoed Derrig and Evans's findings. Research by Eluru and Bhat (2007) supported the selective recruitment (sample selection) hypothesis. They applied a mixed joint binary logit model--an ordered logit model with random coefficients--focusing on the analysis of non-commercial driver safety restraint usage and crash severity. The study showed those drivers who have more awareness of safety will be more likely to wear safety restraints. Drivers who are more aggressive and less safety conscious, on the other hand, will be more likely not to wear safety restraints, even in jurisdictions with a primary safety restraint law. According to the authors, defensive driving habits and driver consciousness will result in less severe injuries, given a crash occurrence. Eluru and Bhat further suggested that safety restraint usage and a self-conscious driving attitude both contribute to fatality reduction and recommended that policy makers advocate giving non-safety restraint users' fines as a punishment as well as providing a mandatory defensive driving course.
The majority of the analyses for safety restraint usage are based on police report crash records. Therefore the accuracy of the police report influences the analysis results. While police reports are the source for major crash databases, researchers should be aware of the level of precision of this crash data and make corresponding adjustments to their findings whenever necessary. Some reporting patterns regarding safety restraint usage in police report crash databases include: Police were least likely to make errors in instances of safety restraint usage regarding
fatally injured occupants (Schiff and Cummings, 2004). However, Roberson (2002) found that police tend to report an unbelted status for dead occupants. This reporting trend may lead to an overestimation of the effectiveness of safety restraint usage in reducing crash severity.
11

Police tend to categorize unbelted survivors as belted when they were not.
2.1.4 Seating Position A study by Farmer et al. (1997) for two-vehicle side-impact crashes showed an increasing risk of sustaining serious injuries for vehicle occupants seated on the side of the vehicle struck during the crash. Meanwhile, Farmer also noted the tendency of serious injury for elder occupants seated on the far side of the crash. Most of the far side occupants were suffering from head and neck injuries of a greater magnitude than those on the side of the vehicle hit. The study provided evidence to advocate vehicle structural design improvements in order to provide better protection to occupants in side-impact crashes. Huelke and Compton (1995) examined the effect of safety restraint usage on injury severity for front and rear seat vehicle occupants and indicated that rear seat occupants are safer than front seat occupants.
2.1.5 Speeding As with alcohol consumption, speeding is another well known contributing factor that increases crash injury severity. Based on physical law, Shinar (1998) indicated that the vehicle that has a greater momentum with a faster speed will introduce a larger amount of exchange energy in a crash. This greater energy could cause more damage to vehicles and significant harm to occupants. In addition, the higher speed will also increase the required stopping sight distance and make it difficult for driver's to make adjustments and corrections during a crash event.
Driving with speeds over 69 mph (111 km/h) will double the odds of having a fatality (Bedard et al, 2002). Joksch (1993) found that injury severity increases at a higher rate than the increasing rate of speed. Higher travel speed also increases the likelihood of causing more severe injuries for older driver groups involved in fixed-object passenger car crashes (Dissanayake and Lu, 2002).
12

Not only does the speed level contribute to increased injury severity, the speed difference of two vehicles also increases the likelihood of experiencing severe injuries (Abdelwahab and Abdel-Aty, 2001). In an earlier study examining collisions between heavy trucks and passenger cars, Duncan et al. (1998) also indicated that a high speed differential increases the risk of more serious injuries.
High speeds are also a safety risk for motorcyclists. Given a crash occurrence, driving at an unsafe speed is one of the factors that increase the likelihood for motorcyclists to suffer from severe injuries (Savolainen and Mannering, 2007). Shibata and Fukuda (1994) reported the significant impact on injury severity from speeding. Their study also found that injury severity is due to the interaction between speed and motorcycle helmet protection use. Helmets will perform more effectively under a speed range less than 31 mph (50 km/h) and will be less effective at speeds greater than 31 mph (50 km/h).
2.2 Vehicle Related Characteristics Vehicle related factors that have been discussed in the literature that focus on injury severity include vehicle type, weight, size, and vehicle occupancy. The number of vehicles involved along with the effects from interactions with other factors such as crash types also influence these vehicle related characteristics (Farmer et al., 1997; Chang and Mannering, 1999; Abdel-Aty and Abdelwahab, 2004a; Abdel-Aty and Abdelwahab, 2004b; Wang and Kockelman, 2005; Ulfarsson and Mannering, 2004; Zhang et al., 2000; Chang and Mannering, 1998 ). Table 2 provides an overview of the literature for vehicle related characteristics as further summarized in Sections 2.2.1 and 2.2.2.
13

Table 2: Summary of Vehicle Related Characteristics Literature

Attributes
Vehicle Weight, Size,
and Type
Vehicle Occupancy

Severity
Less severe injuries associated with occupants of heavy vehicles, pickups, SUVs, and vans when compared to those in passenger cars or on motorcycles
Dependant also on rollover likelihood due to other factors
Acting as interactions with other factors
Greater severity risk for multioccupant vehicles

Crash Type Significant impact on two-
vehicle side impact, rear-end and rollover collisions
Increase injury severity while involving turning movement and rear-end collisions
Increase in rear-end and angle crashes due to vehicle mix with larger vehicles
Light trucks, SUVs, and pickups are more likely to be involved in roll-over crashes than other vehicle types

2.2.1 Vehicle Weight, Size, and Type
According to a two-vehicle side-impact study conducted by Farmer et al. (1997), the likelihood of serious injuries tends to increase for the occupants in lightweight passenger vehicles. This finding is consistent with the fundamental principles of physics. Vehicles with a lighter weight and smaller size tend to be more vulnerable in a crash. Krull et al. (2000) also reported that passenger car drivers have a higher probability of sustaining more severe injuries in crashes compared to pickup truck drivers. Abdelwahab and Abdel-Aty (2001) studied two-vehicle crashes occurring at signalized intersections and also reported similar findings about injuries and vehicle type. By examining fatal crash records, Evans and Frick (1993) suggested that lighter weight vehicles pose less risk to others, while heavier vehicles offer less risk to their own occupants. Sirinivasan's (2002) study reinforced previous findings in terms of how strong impacts on injury severity result from the physical protection of vehicles. The researchers reported that sports utility vehicles (SUVs) and light-duty trucks only reduce risk for moderate injuries while heavy-duty trucks are much safer for all their occupants at all injury severity levels.

14

For crashes in Hong Kong involving motorcycles, the motorcyclists exhibited higher risks of sustaining severe injuries than drivers of passenger cars and taxis (Yau et al., 2006). Motorcyclists are more vulnerable on the road than drivers of motor vehicles. Other driver related characteristics and driving behaviors such as age, speed, gap acceptance, and blind spot positioning also contribute to the higher risk of experiencing more serious injuries.
Abdel-Aty and Abdelwahab (2004b) investigated rear-end collisions and the drivers' visibility across various types of light truck vehicles (e.g., light trucks, vans, and SUVs). The study found that drivers' inattention and sight obstruction due to a leading light truck vehicle contributed to vehicle involvement in rear-end crashes. This finding included the age, gender, traffic control devices, and actions initiated by leading vehicles. Abdel-Aty and Abdelwahab (2004a) also investigated the fatality trend of angle crashes with an increasing percentage of light truck vehicles. The study predicted an increase in fatalities due to angle crashes as the number of light truck vehicles increases.
With respect to crash types, Farmer and Lund (2002) discovered that light trucks are twice as likely to experience rollover crashes compared to passenger vehicles. Kweon and Kockelman's (2003) study also revealed that drivers of SUVs and pickups are more likely to be involved in rollover crashes than are drivers of passenger cars. According to a study of SUV safety issues, Khattak and Rocha (2003) reported that SUVs have a higher likelihood of rollover and these crash types lead to more severe injuries. Alternatively, SUVs provide more protection to the occupants in a crash, which should make the SUV safer. This effect agreed with recent study results indicating that SUVs, vans, and pickups tend to be more crashworthy than passenger cars (Toy and Hammitt, 2003).
2.2.2 Vehicle Occupancy
Chang and Mannering (1999) studied crash and injury severities for a variety of vehicle occupancy for crashes and specifically evaluated whether trucks were involved in the
15

crash. After accounting for vehicle occupancy, Chang and Mannering identified influential factors that significantly increase injury severity for truck-involved crashes. These factors included high speed limits, turning maneuver associated crashes, or rearend collisions. The study also indicated that truck-involved crashes tend to more significantly increase crash and injury severities for multi-occupant vehicles than for single-occupant vehicles. Vehicles with a higher numbers of occupants will tend to display a greater likelihood of having someone either injured or seriously injured should a crash occur. In another earlier study, Chang and Mannering (1998) also reported a significant association between injury severity and multi-occupant vehicles.
2.3 Roadway and Roadside Related Characteristics The role of roadway geometric features and roadside characteristics on crash and injury severity provides valuable information for roadway designers, safety engineers, and policy decision makers. Physical roadway data is rarely included comprehensively in a crash database so adequate evaluation of these features is limited when assessing their influence on injury severity and crash types. Table 3 depicts a summary of the literature regarding roadway and roadside characteristics. This information is further summarized in Sections 2.3.1 through 2.3.6.
16

Table 3: Summary of Roadway and Roadside Related Characteristic Literature

Attributes
Alignment and Vertical Grade
Lane and Shoulder

Severity
Increased injury severity with horizontal curves present and steep grades
Various additional dependencies due to urban versus rural location characteristics and associated geometric design
Paved or wider shoulder and lane width reduce crash severity and crash rate

Guardrail use has a mixed

Roadside

effect on crash severity

Characteristics

Speed Limit

Increasing speed limits increases injury severities

Crash Type Various associations with
crash types: single-vehicle and two-vehicle crashes
Various associations with crash types: side-swipe, singlevehicle and two-vehicle crashes
Wider paved shoulders reduce single vehicle crashes
Centerline rumble strips on two-lane rural roads reduce front- and opposing-direction crashes
Increasing speed limit increases injury severities for single-vehicle, two-vehicle, and multiple-vehicle crashes

Wet vs. Dry Pavement Condition

Conflicting findings regarding the association of severity to wet road conditions

Traffic Volume Significant impact factor

Wet pavement conditions interact with crash types: hitobject crashes, multiplevehicle crashes (specifically sideswipe and opposite direction crashes)
Dry pavement and daylight associated with rear-end conditions
Not significant to crash type

17

2.3.1 Roadway Alignment and Grades
Previous studies reviewed the effects due to horizontal or vertical alignments and grades at various road junction types and traffic control devices, such as road segments, intersections, or signaled and unsignalized traffic controls. Most of these studies reported that horizontal and vertical curvature presence and steeper grades were associated with a higher risk of severe injuries.
Horizontal curves located downstream of long straight roadway segments on level terrain tend to increase the likelihood of injuries in crashes for older drivers (Khattak et al., 2002). Researchers recommended the installation of curve warning signs or rumble strips on long section of highways in order to alert older and younger drivers of the oncoming geometric changes. Wang and Kockelman (2005) found that negotiating at curve sections which have high speed limits tends to increase the chance of experiencing injury and fatality for older and female vehicle occupants. For two-vehicle crashes, they estimated a 56.7% increasing in fatalities for curves to the left (in the direction of travel) and a 39.2% increase in fatalities for curves to the right when compared to straight road segments. Single-vehicle crashes exhibited similar crash characteristics at curve locations. These findings are also consistent with study reports from Dissanayake and Lu (2002) and Abdel-Aty (2003).
The presence of horizontal and vertical curvatures is also associated with more severe injuries for motorcycle-operators in single- and multi-vehicle crashes (Savolainen and Mannering, 2007). Kim et al. (2007) determined that for crash types, for example, the presence of horizontal curvature is among the contributing factors that increase the likelihood of angle crashes.
While the majority of previous studies presented convincing evidence that curvilinear roads increase the chance of crash occurrences, one New Zealand fatal crash study found that there was no significant evidence of the association between curves and fatal crashes on rural state highways (Haynes et al., 2008). The study did show, however, that curvilinear roads reduce the likelihood of fatal crashes on urban roads.
18

Wang and Kockelman (2005) also examined the grade effect on injury severity for onevehicle and two-vehicle crashes. The study distinguished between the effects on injury severity for uphill as well as downhill conditions. Wang and Kockelman did not find the grade effect to be significant for single-vehicle crashes. Meanwhile, vertical grade played a significant role for two-vehicle crashes for both uphill and downhill conditions. They found a downhill grade was associated with a 37.3% increase in fatalities and a 13.3% increase in all forms of injury. The 2002 study by Dissanayake and Lu also provided evidence to support the association of grades and more severe injuries.
2.3.2 Lane and Shoulder Numerous studies have identified significant reductions in overall crash rate as a result of wider travel lanes or shoulders (Foody and Long,1974; Shannon and Stanley,1976; Jorgensen, 1978; Zegeer et al., 1979; Zegeer et al., 1988; Griffin and Mak, 1989). Few of these studies focused on the effects on crash injury severity and crash type. Heimbach et al. (1974) performed a study in North Carolina and found a significant decrease in crash severity by paving 3 to 4 ft. (0.9 to 1.2 m) of unpaved shoulders. Kim et al. (2007) predicted that same direction sideswipe crashes have less chance to occur at intersections with shoulders compared to intersections without shoulders.
According to a study by Ivan et al. (1999), increasing shoulder width, sight distance and traffic intensity are more likely to reduce single-vehicle crash rates than for multi-vehicle crashes. Meanwhile, factors that may increase the multi-vehicle crash rate include the increasing number of traffic signals, daily single-unit truck percentage, and shoulder width.
2.3.3 Roadside Characteristics Based on a three-year vehicle crash database of northbound State Route 3 in Washington State, Lee and Mannering (2002) investigated the relationship between crash severity, roadside features, and driver behavior. The various roadside features examined in the
19

study included guardrail, fixed objects, sign supports, tree groups, and utility poles. The study revealed that the impact on crash injury severity is a complex interaction for combined roadside features. A tree group along the roadside will increase the likelihood of having evident injuries by 43.7%. In spite of the fact that the primary purpose of a guardrail is to reduce crash injury severity by preventing crashes as cars run off the road at curvature locations, the study showed that guardrail implementation increased the likelihood of a disabling injury or fatality by 90%. The somewhat counter-intuitive findings imply that there are potential interactions with other observed and unobserved factors. The study findings were also based on data for only one direction of travel for one study corridor in the State of Washington.
By taking into account the number of vehicles involved in crashes, Wang and Kockelman (2005) reported that manufactured barriers slightly reduce injury severity while dividers and one-way roads have the opposite effects for single-vehicle crashes. With two-vehicle crashes, dividers, medians and manufactured barriers reduced injury severity. Among them, manufactured barriers have the strongest effect and can reduce fatalities by 53.7%. Another promising result is that the installation of centerline rumble strips on two-lane rural roads significantly decreased front and opposing-direction sideswipe crashes by 25% (Persaud et al., 2004)
2.3.4 Speed Limit
Wang and Kockelman (2005) reported that roads with higher speed limits increase the percentage of fatalities for two-vehicle crashes, with a similar trend of lesser significance exhibited for single-vehicle crashes. This result implies the difference in the crash mechanism for single-vehicle vs. two-vehicle crashes. A separate modeling effort is, therefore, worthwhile in order to separate the confounding factors and to show the significant effects for variables that tend to be crash-type specific. Yau et al. (2006) also identified an association of higher speed limits and crash severity for multiple-vehicle crashes. Ossiander and Cummings (2002) concluded that speed limit increases from 55 mph (89 km/h) to 65 mph (105 km/h) contributed to the increase in the fatal crash rate on
20

freeways in Washington State. Duncan et al. (1998) also reported a trend of increasing crash severity with high speed limits.
By treating the regulatory speed limit as a categorical variable with two conditions--legal speed limit greater than 53 mph (85 km/h) or below 53 mph (85 km/h), Lee and Mannering (2002) found that a crash will be more likely to have evident injury for speed limits above 53 mph (85 km/h).
The study by Khattak et al. (2002) found that older drivers have a lower probability of severe injury in crashes preceding 1997. Iowa increased speed limits on many of its highways in 1996 after the United States Congress repealed the national maximum speed limit in 1995. It is possible, therefore, that increasing speed limits may be one of the contributing factors to increased injury levels for older drivers after 1997. The Iowa Safety Management System Task Force on Speed Limits (1998) also recorded a trend of increasing fatalities, injuries, and total crashes after the speed limit change. The researchers indicated, however, the need for an in-depth analysis and assessment for the impact of the policy that permitted the increase in legal speed limit.
While many previous studies seemingly correlated increasing fatalities to higher speed limits, some researchers reported study results with different perspectives. Renski et al. (1999) reported an inconclusive result of the effects of speed limit changes on fatalities due to having a small sample size of fatal crashes. They also argued that limitation and bias may have been present from factors such as study site selection (only sites with good safety records were chosen), crash type (single-vehicle crashes only), road functional classification (Interstate only), and time period (two-years only).
Balkin and Ord (2001), after accounting for seasonal patterns, assessed the influence of speed limit increases on fatal interstate crashes for a variety of United States rural and urban interstates. The influence of higher speed limits had different impacts on fatalities for rural and urban interstates and also varied from state to state. After many speed limits increased nationwide in 1995, higher speed limits were believed to slightly increase
21

fatalities on rural interstates and to exhibit either a slight increase or no change for fatalities on urban interstates. Balkin and Ord (2001) emphasized the various effects of speed limit increases on injury severity for each individual state.
2.3.5 Wet versus Dry Pavement Conditions The research focusing on pavement conditions has primarily addressed wet or dry pavement rather than pavement type. According to research by Kim et al. (2007), wet road conditions increase the likelihood of sideswipe, opposite direction crashes. Golob and Recker (2003) studied the relationships of crash types and traffic volume for prevailing weather and lighting conditions for an urban freeway in southern California. They found that wet road conditions increase the chance of crash occurrences that involve lane changing maneuvers, such as hit-object collisions and multiple-vehicle collisions. Meanwhile, dry road and day light conditions increase the occurrence of rearend collisions.
Duncan et al. (1998) reported an increase in injury severity under wet pavement conditions while Krull et al. (2000) reported the opposite findings. Savolainen and Mannering (2007) found that wet pavement conditions were associated with less severe driver injuries. This finding is consistent with observations by Krull et al. (2000).
2.3.6 Traffic Volume A study by Golob and Recker (2003) revealed that traffic volume has a greater impact on crash severity than does speed. This result highlights the importance of incorporating traffic exposure in crash severity studies. Furthermore, Washington et al. (1999) reported the interactions of the traffic exposure as vehicle miles traveled (VMT) for various road functional classifications. Duncan et al. (1998) also reported that congested conditions reduce injury severity.
While some studies confirmed the significant impact traffic exposure has on injury severity, Kim et al. (2007) found the influence of average annual daily traffic (AADT) on
22

the probability of a specific crash type to be insignificant. They estimated models to predict the probability of crash occurrence categorized by crash types (including angle, rear-end, and same and opposite direction sideswipe crashes) using crash data from rural signalized and unsignalized intersections in Georgia. They also indicated that AADT is highly associated with crash frequency with no direct relationship to with crash types.

2.4 Crash Related Characteristics
Though many agencies historically have studies total crashes, it is clear that crash types and number of involved vehicles are directly associated with crash causation. Table 4 briefly summarizes the published literature for these crash related characteristics. This information is further summarized in Sections 2.4.1 and 2.4.2.

Table 4: Summary of Crash Related Characteristic Literature

Attributes
Crash Type
Number of Involved Vehicles

Severity
Associated directly with specific road features such as horizontal curves or paved shoulders
Rollover, head-on crashes most fatal
Side-impact also more severe Increased number of occupants
increases likelihood of severe injury

2.4.1 Crash Type
A variety of crash types such as rear-end, angle, single-vehicle, etc. are addressed in the published literature. Previous researchers identified the important role of crash type for safety analysis (Kim et al., 2007; Retting et al., 1994). Various studies have associated crash type with other potential contributing factors for crash injury severity and crash rate reduction.

23

Side-impact crashes tend to double the chance of a fatality compared to front impact crashes according to a study report by Bedard et al. (2002). Evans and Frick (1993) also reported a higher driver fatality risk for side-impact crashes. For two-vehicle crashes at signalized intersections, Abdelwahab and Abdel-Aty (2001) found that side-impact crashes increase injury severities compared to other impact points while an angle crash is one of the major crash types that will occur at intersections. Jones and Jorgensen (2003) analyzed a crash database collected in Norway and reported that fatality risk increased with head-on collisions compared to rear-end and side-impact collisions, a finding that is inconsistent with those of Bedar et al. (2002). Srinivasan (2002) found that crash severity increases for front impact, head-on collisions when compared to side-impact crashes. Krull et al. (2000) studied the effect of driver injury severity for rollover single-vehicle crashes and reported an increasing trend of driver injury severity with rollover crashes.
A few studies have targeted the older driver population. Under the condition of fixedobject passenger car crashes involving older drivers, the probability of having more serious injuries tends to increase with front impact crashes, while side-impact crashes appear to increase minor injuries for the older driver population (Dissanayake and Lu, 2002).
In order to determine the influence on the probabilities of various crash types, Kim et al. (2007) estimated five separate multilevel models to predict the probability of crash occurrence categorized by crash types, including angle, head-on, rear-end, and same and opposite direction sideswipe crashes. This study used crash data from rural signalized and unsignalized intersections in Georgia. The study identified geometric features and signal control types as influential factors that impact the occurrence of corresponding types of crashes. Factors that contributed to increasing the odds of angle crash occurrence included horizontal curve alignments compared to straight segments, and unsignalized intersections compared to signalized intersections. Rear-end crashes also were more likely to occur at unsignalized rural intersections compared to signalized rural intersections. Sideswipe same direction crashes occurred less at horizontal curves near intersections and skewed intersections. Sideswipe opposite direction crashes are also less
24

likely to occur at horizontal or vertical curves near intersections. The study produced inconclusive results for head-on crashes due to a small sample size (n=16).
In addition to the studies devoted to the influence on crash types due to crash severity, road geometric features, and traffic control devices, Golob and Recker (2003) studied the relationships between crash types and traffic volume after accounting for weather and lighting conditions for urban freeways in southern California. Their study found a strong association between collision types and typical traffic speed as well as temporal speed variations in the left and interior lanes.
According to a motorcycle fatal crash study (Preusser et al., 1995), crash types where vehicles leave their travel lanes and the crash led to either hitting a roadside fixed object or colliding with opposing traffic, were more likely to occur in rural areas, at high speed roadways, and at curve conditions. Alcohol was often associated with these types of crashes. Meanwhile, urban intersections, where more traffic interaction occurs, were more often associated with angle crashes. Motorcycle crash patterns are generally consistent with motor vehicle crash patterns.
2.4.2 Number of Involved Vehicles As stated previously, single-vehicle crashes have different crash dynamics compared to multiple-vehicle crashes. Studies have shown different impacts on crash severity. For one Hong Kong study, Yau et al. (2006) found that the number of vehicles involved in traffic crashes had a significant impact on injury severities. The study determined that more vehicle involvement increases the likelihood of more serious or fatal injuries.
2.5 Environment Related Characteristics The location and conditions under which a crash occurs can directly influence the likelihood of a crash as well as associated crash severity. Table 5 briefly summarizes the literature for environmental issues that is reviewed in detail in Section 2.5.1 through 2.5.3.
25

Table 5: Summary of Environment Related Characteristic Literature

Attributes
Weather Conditions
Lighting Conditions Urban and
Rural

Severity
Conflicting results regarding adverse weather injury severity depending on weather type
Dependant on other factors such as clearance policies and vehicle weather devices (tire traction, chains, etc.)
Inconsistent results Dependant on other factors
such as reflective pavement marking, signage, etc. Rural crashes increase injury severities

Crash Type Influence crash types: one-
vehicle and two-vehicle crashes
Influence crash types: singlevehicle and multiple-vehicle crashes
N/A

2.5.1 Weather Conditions
An interesting question in safety analysis is whether adverse weather conditions increase crash severity. Several studies reported a trend that adverse weather conditions tend to decrease injury severities, given a crash occurrence (Wang and Kockelman, 2005; Edwards,1998). Khattak et al. (1998) found that adverse weather crashes occur more frequently but tend to be less severe.
Wang and Kockelman (2005) had a general finding that injury probability for vehicle occupants is lower during adverse weather conditions. They estimated a reduction by 16.4% and 32.5% for injury probability and fatality, respectively, for two-vehicle crashes during adverse weather conditions. The effect is even more significant for single-vehicle crashes. Wang and Kockelman interpreted the effect as possibly more cautious driving behavior during poor bad weather conditions. Researchers have applied potential adjustments and compensations made by drivers corresponding to driving condition changes as a commonly used interpretation for some of these counterintuitive results (Khattak et al., 1998). Edwards (1998) also reported that rain significantly decreases

26

crash severity compared to during dry weather conditions. Meanwhile the study showed the effect from fog varies across geographical locations.
Some studies provide different findings based on weather and driver age groups. A study by Zhang et al. (2000) found that crashes during snow substantially increase the likelihood of fatal crashes by 60% for older drivers while the snowy/icy road surface (often after the snow has ceased to fall) has opposite effects. One hypothesis is that older drivers may tend to be more cautious, alert, or avoid driving entirely during icy road conditions. For younger drivers, Mao et al. (1997) did not find a significant impact from snow weather on crash severity. Kim et al. (2007) found clear weather conditions among the contributing factors that increase the odds of angle crash occurrence.
2.5.2 Lighting Conditions Research results regarding road lighting conditions and their role on crash occurrence present inconsistent findings for crash severity and crash types. These conflicting and sometimes counterintuitive results demonstrate the complexity of the role of lighting on safety and further suggest that there may be potential interactions with other unknown confounding factors.
Several studies suggest that reduced lighting conditions may significantly increase injury severity (Abdel-Aty, 2003; Khattak et al., 2002; Duncan et al., 1998). Other studies reported inconsistent findings regarding lighting conditions while taking into account the number of vehicles involved in crashes and crash types.
Wang and Kockelman (2005) determined that insufficient lighting conditions decrease injury severity for single-vehicle crashes while increasing injury severity for two-vehicle crashes. Yau et al. (2006) demonstrated a different observation for multiple-vehicle crashes in Hong Kong. At night, the poor street lighting appeared to decrease the risk of severe and fatal injuries while good light conditions at night presented the greater risk involving severe or fatal traffic crashes. Regional differences, such as driving behavior, roadway design standards, and geographical features might contribute to these
27

inconsistent results. Kim et al. (2007) reported an increasing trend for all types of crashes when lighting is sufficient.
2.5.3 Urban versus Rural Various studies revealed a consistent pattern across roadway segments and intersections that rural locations tend to have an increased risk for more severe injuries than in urban environments (Krull et al., 2000; Jones and Jorgensen, 2003; Dissanayake and Lu, 2002; Abdelwahab and Abdel-Aty, 2001; Lee and Mannering, 2002). By their nature alone, urban and rural roadways differ in many ways including design standards, physical features, and traffic exposure. It is also likely that driving behavior differences also contribute to the safety performance for rural versus urban areas.
Krull et al. (2000) presented evidence that rural roads are more likely to increase the chances of drivers sustaining severe injury. In Norway, Jones and Jorgensen (2003) reported fatality risks increased in rural areas. After accounting for horizontal and vertical curvature conditions, crashes that occurred at rural locations still tended to be associated with more severe crashes compared to their in urban counterparts (Dissanayake and Lu, 2002). According to the study by Abdelwahab and Abdel-Aty (2001), drivers were more likely to sustain severe injuries at rural signalized intersections than at urban signalized intersections when the crash involved two vehicles.
2.6 Literature Summary Table 1 through Table 5 provided an overview of the primary research findings in each of the five categories: driver and passenger related characteristics, vehicle related characteristics, roadway and roadside related characteristics, crash related characteristics, and environment related characteristics.
Many of the previous studies were based on crash databases from one state or select study corridors, while other studies focused on national fatality data. The various studies demonstrate many contradictions about crash type and severity and associated factors justifying the need to further explore both crash conditions and the rural road
28

environment. This study will shed light on this issue based on the crash database from select southeastern states (specifically Alabama, Georgia, Mississippi, and South Carolina). The research will investigate the cross state differences and similarities in terms of the impact on crash conditions and from potential contributing factors. The study's findings will ultimately help to explain the various safety performances on twolane rural highways among the four states and offer guidance of generating countermeasures to mitigate the situation.
29

This page blank intentionally 30

3.0 Summary Statistics of Study Crash Data
3.1 Data Description Previously, researchers from several states in the southeastern part of the United States participated in a rural fatal crash study. Of these states, final data is available for five states including Alabama, Georgia, North Carolina, South Carolina, and Mississippi. The database provided by North Carolina, however, does not include all of the study variables included in databases by the other four states. As a result, this summary focuses on fatal crash data statistics for the remaining states. Alabama and Georgia randomly selected 150 fatal crashes from FARS that occurred during 1997 for two-lane rural highways (Dixon, 2005). Mississippi has a smaller crash population and so researchers from that state developed a random sample of 100 fatal crashes for the year 1997. South Carolina evaluated 157 fatal crashes in their final analysis. These South Carolina fatal crashes occurred during 1998 (Dixon, 2005). Each participating state identified the candidate crashes and performed physical or video site visits. These site visits helped identify physical road features that are not commonly available in police crash reports.
Generally, each database included five types of data: crash data elements, site data elements, environmental data elements, person data elements, and vehicle data element. Though certain overlap can be expected between these data categories, the variables common to each are generally summarized as follows:
31

Crash Data Element Crash data elements contain general crash characteristics such as crash date, time, location, numbers of involved vehicles, drivers, occupancy, severity, drug or alcohol usage, and similar information unique to each crash.
Site Data Element Included in the crash databases is information unique to each specific crash location. Common variables included as site data elements are horizontal and vertical alignment information, cross-slope, functional classification, lane width, shoulder type, lane configuration, road surface type and condition, average daily traffic, intersection or driveway information, speed limit, roadside hazard ratings, and similar information unique to a specific location.
Environmental Data Element
The prevailing weather and lighting may influence crash causation. As a result, the environmental data elements included in the database include weather conditions, ambient light conditions, and road surface conditions.
Person Data Element
Data about the individuals involved in a crash can help illuminate how the crash occurred or why certain occupants survived the crash. This driver and passenger characteristic information includes gender, age, occupant protection equipment use, injury status, seating position, ejection status, and similar information unique to the vehicle occupants.
Vehicle Data Element
The type of vehicle as well as vehicle status can directly influence crash conditions and survivability. As a result, the vehicle data elements included in the crash database incorporate vehicle characteristics include vehicle make, model, year, configuration, vehicle maneuvers, and vehicle towing status as examples.
32

3.2 Data Representation of Larger Population The FARS database contains all fatal traffic crashes in the United States including those that occur in the 50 states, the District of Columbia, and Puerto Rico. For a crash to be included in the FARS database, all resulting fatalities of vehicle occupants and nonmotorists must have occurred within 30 days of the crash. Figures 1 through 4 depict a distribution of the four-state driver population distributions for nine age groups based on the driver gender. The male drivers in the age group of 16 to 20 years old represent a relatively high frequency across all four. For all states except Mississippi, both male and female drivers have the highest fatal crash frequency for the age group of 25 to 34 years old, while in Mississippi both the 25 to 34 years old age group and the 35 to 44 years old groups have similar high level frequencies for both male and female drivers.
Figure 1: Alabama Driver Population Distribution by Age Group
33

Figure 2: Georgia Driver Population Distribution by Age Group Figure 3: Mississippi Driver Population Distribution by Age Group
34

Figure 4: South Carolina Driver Population Distribution by Age Group

To determine whether the fatal crash sample data used for this study represents the

general crash data population in FARS, the research team applied a Pearson's chi-square

test to the observed crash data. In this study, the Pearson's chi-square test is used to

evaluate a null hypothesis that the frequency distribution of driver by age group for the

database sample is consistent with that of the driver population for male and female

drivers involved in fatal crashes for each individual state. Equation (1) demonstrates the

equation for calculating this chi-square statistic. The degree of freedom of the chi-square

statistic is the number of outcome categories (age groups of drivers = 9) minus one (df =

9-1 = 8).

2 = n (Observedi - Expectedi )2

(1)

i =1

Expectedi

Given: Observedi = an observed frequency; Expectedi = an expected (theoretical) frequency under null hypothesis; n = the number of outcome categories, n = 9 age groups.

Figures 5 through 12 demonstrate the observed and expected male and female driver sample frequency distribution by age group for the four individual states. The graphs

35

present the pattern of a consistent frequency distribution of drivers by age group for the study sample and the total FARS population. Table 6 presents chi-square test results for male and female drivers for the four focus states. The majority of the p-values from the tests are greater than the 0.05 statistically significant level (assuming 95% acceptance). This means that the statistical chi-square test does not reject the null hypothesis and the sample data is representative of the larger FARS data for the individual states. Overall, the test demonstrates convincing evidence that the sample data has a good representation of the data from FARS in this study. Noticeably, the p-value from male drivers of Georgia slightly exceeds but is very close to 0.05 (p-value=0.051). This suggests that the frequency of Georgia male drivers do not conform as closely to the FARS data as shown for other states and genders; however, the statistical test still gives compelling evidence that the data from all four states is a good representative sample of the overall FARS data in these states.
Figure 5: Alabama Male Driver Sample Distribution by Age Group
36

Figure 6: Alabama Female Driver Sample Distribution by Age Group Figure 7: Georgia Male Driver Sample Distribution by Age Group 37

Figure 8: Georgia Female Driver Sample Distribution by Age Group Figure 9: Mississippi Male Driver Sample Distribution by Age Group
38

Figure 10: Mississippi Female Driver Sample Distribution by Age Group
Figure 11: South Carolina Male Driver Sample Distribution by Age Group 39

Figure 12: South Carolina Female Driver Sample Distribution by Age Group

Table 6: Pearson's Chi-Square Test Results of Driver Distribution

State
Alabama Georgia Mississippi South Carolina

Male Driver

Chi-square Statistic

p-value

8.76

0.36

15.44

0.051

12.22

0.14

6.82

0.56

Female Driver

Chi-square Statistic

p-value

5.44

0.71

6.71

0.57

11.87

0.16

9.61

0.29

3.3 Descriptive Statistics
3.3.1 Crash Data Characteristics
For the four states, an average of 58% of fatal crashes was single-vehicle crashes and 42% were multiple-vehicle crashes (see Table 7 and Figure 13). As shown in Figure 14, head on crashes and angle crashes were two common multiple-vehicle crash types accounting for 36% of total fatal crashes. The study data clearly demonstrates that single-vehicle crashes, head on crashes, and angle crashes are three major fatal crash

40

types that tend to have a greater risk for fatalities compared to rear end and sideswipe crashes.

Table 7: Crash Type Distribution for Study Crashes

Crash Type

South

Alabama Georgia Mississippi Carolina Average

(%) (%)

(%)

(%) (%)

Single-Vehicle Crash

64

59

50

61

58

Multiple-Vehicle Crash

36

41

50

39

42

Head On

18

17

19

15

17

Angle

12

18

21

24

19

Rear End

5

2

4

0

3

Sideswipe Same Direction

1

2

1

0

1

Sideswipe Opposite Direction 0

2

5

0

2

Others

0

0

0

1

0

Figure 13: Single vs. Multiple-Vehicle Fatal Crashes 41

Figure 14: Fatal Crash Type Distribution
3.3.1.1 Crash Distribution by Month The distribution patterns for single-vehicle fatal crashes and multiple-vehicle fatal crashes differed. For single-vehicle fatal crashes, most states exhibited a relatively higher crash frequency around April and May, while the crash frequency of multiple-vehicle fatal crashes peaked around May and October (see Table 8, Figure 15 and Figure 16). More notably, single-vehicle fatal crash frequency for Georgia and Mississippi peaked in April with 11% and 9% of total fatal crashes respectively. Meanwhile multiple-vehicle fatal crash frequency for Georgia peaked in November with 7% of total fatal crashes. Mississippi experienced a peak of 8% of total fatal crashes in October.
42

Table 8: Crash Distribution by Month

State

Crash Type

Jan (%)

Month Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%)

SingleVehicle

3

7

5 5 735 6 7 5 5

4

AL MultipleVehicle

2

1

5 1 533 3 3 4 3

5

AL Total 5 8 10 6 12 6 8 9 10 9 8 9

SingleVehicle

5

5

7 11 5 3 4 4 3 4 4

5

GA MultipleVehicle

4

3

1 3 551 4 5 2 7

2

GA Total 9 8 8 14 10 8 5 8 8 6 11 7

SingleVehicle

2

2

5 9 533 3 5 5 4

4

MS MultipleVehicle

3

3

4 3 543 5 3 8 3

4

MS Total 5 5 9 12 10 7 6 8 8 13 7 8

SingleVehicle

5

5

4 6 984 4 6 5 2

3

SC MultipleVehicle

1

1

2 4 635 4 1 6 2

3

SC Total 6 6 6 10 15 11 9 8 7 11 4 6

Figure 15: Single-Vehicle Fatal Crashes by Month 43

Figure 16: Multiple-Vehicle Fatal Crashes by Month 3.3.1.2 Crash Distribution by Day of Week Table 9 and Figure 17 present a day-of-week pattern for single-vehicle fatal crash distribution across the four states when about 10% to 15% of the total fatal crashes occurred during the weekend compared to 3% to 9% during weekdays. Figure 18, however, demonstrates that multiple-vehicle fatal crashes occurred regularly on Friday and Saturday. The crash percentage was as high as 14% on these days compared to a high of 9% for weekdays. South Carolina data did not provide adequate information for a day of week evaluation.
44

Table 9: Crash Distribution by Day of Week Day of Week (percent per state)
State Crash Type

Total Percent

Monday Tuesday Wednesday Thursday Friday Saturday Sunday

(%) (%) (%) (%) (%) (%) (%) (%)

Single-Vehicle 7

8

7

9

7

14 13 65

AL Multiple-Vehicle 3

3

6

3

9

7

5 36

AL Total

10 11 13 12 16 21 18 --

Single-Vehicle 5

7

4

6

8

13 15 58

GA Multiple-Vehicle 4

5

6

8

5

9

4

41

GA Total

9

12 10 14 13 22 19

--

Single-Vehicle 6

6

6

4

4

14 10 50

MS Multiple-Vehicle 4

6

4

6

9

11 10 50

MS Total

10 12 10 10 13 25 20 --

Figure 17: Single-Vehicle Fatal Crashes by Day of Week 45

Figure 18: Multiple-Vehicle Fatal Crashes by Day of Week
3.3.2 Roadway and Roadside Related Characteristics 3.3.2.1 Horizontal Alignment Direction and Curve Radius The horizontal alignment variable includes three conditions: straight alignment, horizontal curve to the left, and horizontal curve to the right. A second variable referred to as horizontal curvature further separates curves into two groups: sharp and mild curve. Sharp curves are those that require drivers to make speed adjustments, while mild curves do not require drivers to reduce their speeds.
For the four states with comprehensive data, more than 50% of the rural two-lane road fatal crashes occurred on road segments with straight horizontal alignments (see Table 10). At horizontal curve locations, curves to the left tended to have a stronger association with the single-vehicle fatal crashes than did similar crashes at curves to the right (see Figure 19). Meanwhile, as shown in Figure 20, curves to the right were more frequently associated with multiple-vehicle fatal crashes.
In Alabama and South Carolina, mild horizontal curves (curves with large radii) were often associated with single-vehicle fatal crashes (see Figure 21). In Georgia and Mississippi, sharp horizontal curve locations were strongly associated with single-vehicle fatal crashes. Many of the multiple-vehicle fatal crashes occurred at straight alignment
46

locations, but Mississippi and South Carolina had stronger associations between mild curve locations and multiple-vehicle fatal crash occurrences than those that occurred in from Alabama and Georgia.

Table 10: Crash Type by Horizontal Alignment Direction and Horizontal Curvature

State

Crash Type

Curve Curve Unknown

Unknown

Straight to the to the Curve Sharp Mild Curve

Alignment Left Right Direction Curve Curve Radius

(%) (%) (%) (%) (%) (%) (%)

Single-Vehicle 31

23 10

-

10 23

-

AL Multiple-Vehicle 28

5

3

-

3

5

-

AL Total

59

28 13

-

13 28

-

Single-Vehicle 31

17 12

-

17 11

-

GA Multiple-Vehicle 20

6 15

-

10 11

-

GA Total

51

23 27

-

27 22

-

Single-Vehicle 22

14 14

-

18 10

-

MS Multiple-Vehicle 29

9 12

-

6 15

-

MS Total

51

13 26

-

24 25

-

Single-Vehicle 27

17 16

1

3 31

1

SC Multiple-Vehicle 22

8

9

1

2 15

-

SC Total

49

25 25

2

5 46

1

47

Figure 19: Single-Vehicle Fatal Crashes and Associated Horizontal Alignment Figure 20: Multiple-Vehicle Fatal Crashes and Associated Horizontal Alignment
48

Figure 21: Single-Vehicle Fatal Crashes and Associated Horizontal Curvature
Figure 22: Multiple-Vehicle Fatal Crashes and Associated Horizontal Curvature 3.3.2.2 Vertical Grade A steep vertical grade could have a potential influence on the likelihood of crashes, so the database included a vertical grade variable that included the direction of vertical slope (uphill, downhill, and flat) and percentage of grade (level: 1%; mild slope: 2-6%; steep slope: >6%).
49

Table 11 demonstrates the distribution of single-vehicle and multiple-vehicle fatal crashes located at uphill, downhill, and flat locations (see also Figure 23 and Figure 24). In Alabama, Georgia, and Mississippi, single-vehicle fatal crashes were more often associated with downhill slopes than uphill slopes. Figure 24 demonstrates that in Alabama and Georgia, multiple-vehicle fatal crashes were also more often associated with downhill slopes that their uphill counterparts. As depicted in Table 11, most of the study crashes occurred at mild or level vertical grade locations.

Table 11: Crash Type by Vertical Grade

State

Crash Type

Direction of Slope

Vertical Grade

Uphill Downhill Flat NA Level Mild Steep NA

(%) (%) (%) (%) (%) (%) (%) (%)

Single-Vehicle 7

33

25 - 13 26

1

23

AL Multiple-Vehicle 6

11

19 - 11 11

0

13

AL Total

13

44

44 - 24 37

1 36

Single-Vehicle 19

20

21 - 12 25

1

21

GA Multiple-Vehicle 13

19

9

- 13 17

1

9

GA Total

32

39

30 - 25 42

2

30

Single-Vehicle 9

17

24 - 18 20

1

11

MS Multiple-Vehicle 12

12

26 - 18 20

1

11

MS Total

21

29

30 - 36 40

2

22

Single-Vehicle 15

14

31 - 15 15

1

30

SC Multiple-Vehicle 11

10

18 1 10 12

1

17

SC Total

26

24

49 1 25 27

2

47

50

Figure 23: Single-Vehicle Crashes by Vertical Grade
Figure 24: Multiple-Vehicle Crashes by Vertical Grade 3.3.2.3 Cross Section Configuration The cross section configuration variable describes the cross-slope of the road segment at the crash location. Single-vehicle fatal crashes occurred more often at locations with typical rooftop configurations than at superelevated locations (see Table 12 and Figure 25). Highway design principals typically require the use of rooftop cross sections at straight (tangent) alignment locations, while curved locations are generally superelevated. Multiple-vehicle crashes in Alabama, Mississippi, and South Carolina occurred more
51

often at locations with rooftop cross sections. In Georgia, multiple-vehicle fatal crashes occurred similarly at rooftop and superelevated locations (see Figure 26).

State AL GA MS SC

Table 12: Cross Section Configuration by Crash Type

Crash Type
Single-Vehicle Multiple-Vehicle
AL Total Single-Vehicle Multiple-Vehicle
GA Total Single-Vehicle Multiple-Vehicle
MS Total Single-Vehicle Multiple-Vehicle
SC Total

Rooftop (%) 35 27 62 32 21 53 26 34 60 34 24 58

Cross Section Superelevated
(%) 29 8 37 26 19 45 23 16 39 26 16 42

Others (%) 1 1 1 1 1 2 1 0 1 1 0 1

Figure 25: Single-Vehicle Crashes by Cross Section Configuration 52

Figure 26: Multiple-Vehicle Crashes by Cross Section Configuration
3.3.2.4 National Highway System As shown in Table 13, all four states had 79% or more of the fatal crashes that occurred on two-lane rural highways that were not a part of the national highway system. A higher percentage of the multiple-vehicle fatal crashes occurred on the national highway system than for the single-vehicle fatal crashes.

Table 13: Crash Occurrence on National Highway System

State AL GA MS SC

Crash Type
Single-Vehicle Multiple-Vehicle
AL Total Single-Vehicle Multiple-Vehicle
GA Total Single-Vehicle Multiple-Vehicle
MS Total Single-Vehicle Multiple-Vehicle
SC Total

National Highway System

Yes

No

(%)

(%)

2

62

9

27

11

89

9

50

12

29

21

79

0

50

3

47

3

97

4

57

7

32

11

89

53

3.3.2.5 Road Functional Classification
The functional classification for the two-lane rural highways included major arterial, minor arterial, major collector, minor collector, local roads, and unknown. As shown in Table 14, Figure 27, and Figure 28, the road functional classifications were distributed differently between single-vehicle and multiple-vehicle fatal crashes. In addition, the study observed a unique functional classification distribution for Mississippi by the complete absence of arterials for the study sites. The notable fatal crash distribution difference for the functional classification between Mississippi and the other three southeastern states may be caused by differing standards applied to classify rural twolane highways in each state.

Table 14: Crash Occurrence per Road Functional Classification

State

Crash Type

Road Functional Classification

Major Minor Major Minor

Arterial Arterial Collector Collector Local Unknown

(%) (%) (%)

(%) (%) (%)

Single-Vehicle 5

11

28

11

9

1

AL Multiple-Vehicle 10

8

13

1

4

0

AL Total

15

19

41

12

13

1

Single-Vehicle 4

11

21

9

14

0

GA Multiple-Vehicle 9

13

15

3

2

0

GA Total

13

24

36

12

16

0

Single-Vehicle 0

0

0

27

23

0

MS Multiple-Vehicle 0

0

1

41

8

0

MS Total

0

0

1

68

31

0

Single-Vehicle 4

11

30

4

1

10

SC Multiple-Vehicle 9

15

13

1

1

1

SC Total

13

26

43

5

2

11

54

Figure 27: Single-Vehicle Fatal Crashes by Road Functional Classification
Figure 28: Multiple-Vehicle Fatal Crashes by Road Functional Classification 3.3.2.6 Lane Width Table 15, Figure 29, and Figure 30 clearly demonstrate a pattern that the association of lane width is different for single-vehicle fatal crashes than for the study multiple-vehicle fatal crashes. As the lane widths narrow from 12 ft (3.6 m) to 10 ft (3.0 m), the associated percentages of fatal crashes that were single-vehicle crashes increased from 12% to 21% on average, while multiple-vehicle crashes decreased from 21% to 7% on average.
55

Many researchers have suggested that wider lanes are more likely to enable speeding behavior. This increase in speed it thought to lead to many potential crash causation factors. Even though the data can be interpreted that 12 ft (3.6 m) wide lanes appear to be safer compared to the 10 ft (3.0 m) lanes for single-vehicle fatal crashes, lane widening may not always assure crash reduction due to the possible increased speed environment.

Table 15: Crash Type by Lane Width Distribution

Lane Width
(ft)
7 8 8.5 9 9.5 10 10.5 11 11.5 12 12.5 13 14 17 NA

AL

GA

MS

SC

Single- Multiple- Single- Multiple- Single- Multiple- Single- Multiple

Vehicle Vehicle Vehicle Vehicle Vehicle Vehicle Vehicle Vehicle

(%)

(%) (%) (%) (%) (%)

(%) (%)

0

0

0

0

2

0

0

0

1

1

1

0

0

0

2

1

0

0

1

1

0

0

0

0

8

1

5

2

4

2

6

0

0

0

6

1

0

0

0

0

21

9

15

3

18

7

29

7

0

0

3

3

0

0

0

0

19

5

12

9

12

15

16

14

0

0

1

1

0

0

0

0

15

21

11

20

14

25

7

17

0

0

1

0

0

0

0

0

0

0

1

0

0

0

1

0

0

0

0

0

0

1

0

0

0

0

0

0

0

0

1

0

0

0

1

0

0

0

0

1

56

Figure 29: Single-Vehicle Fatal Crashes by Lane Width
Figure 30: Multiple-Vehicle Fatal Crashes by Lane Width 3.3.2.7 Shoulder Type Overall, approximately 50% of the single-vehicle crash study locations were characterized by the presence of graded shoulders, while multiple-vehicle crash study locations had graded shoulders at around 30% of the sites (see Table 16, Figure 31, and Figure 32). Paved shoulders were minimal at the crash sites, but a few locations did have a combination of narrowly paved shoulders combined with graded shoulders.
57

State AL GA MS SC

Table 16: Crashes by Shoulder Type

Crash Type
Single-Vehicle Multiple-Vehicle
AL Total Single-Vehicle Multiple-Vehicle
GA Total Single-Vehicle Multiple-Vehicle
MS Total Single-Vehicle Multiple-Vehicle
SC Total

Paved (%) 5 7 12 3 1 4 0 6 6 1 3
4

Shoulder Type

Combination

Graded Paved and Graded

(%)

(%)

45

5

21

5

66

10

44

9

21

16

65

25

44

3

34

7

78

10

57

1

34

2

91

3

Others (%) 9 3 12 4 3 7 3 3 6 1 1 1

Figure 31: Single-Vehicle Fatal Crashes by Shoulder Type 58

Figure 32: Multiple-Vehicle Fatal Crashes by Shoulder Type

3.3.2.8 Auxiliary Lane Configuration
As indicated in Table 17, over 90% of fatal crashes occurred at locations without either turning lanes, passing lanes, or emergency lanes. The research team used a two-lane rural road as a key requirement for study crash selection, so the minimal presence of auxiliary lanes at the study locations is to be expected as a result of the experimental design.

State
AL GA MS SC

Table 17: Crashes by Auxiliary Lane Configuration

Number of

Turning Lanes

(%)

0

1

2

4

95 5

0

1

96 4

0

0

93 3

4

0

100 0

0

0

Number of

Passing Lanes

(%)

0

1

2

97

1

3

97

3

0

100 0

0

100 0

0

Number of

Emergency Lanes

(%)

0

1

99

1

100

0

100

0

100

0

3.3.2.9 Road Surface Type The road surface type variable depicts the driving surface material at each crash site, including concrete, blacktop (asphalt), slag gravel or stone, and dirt. As shown in Table 18, Figure 33, and Figure 34, more than 95% of total fatal crashes in Alabama and
59

Georgia occurred on road segments with blacktop surfaces. Fatal crashes in Mississippi and South Carolina had a slightly smaller percentage of blacktop surfaces at 68% and 65% respectively.

Table 18: Crash Type by Road Surface Material

State

Crash Type

Road Surface Type

Slag, Gravel

Concrete Blacktop or Stone

Dirt

(%)

(%)

(%)

(%)

Single-Vehicle

0

61

1

1

AL Multiple-Vehicle

1

35

1

0

AL Total

1

96

2

1

Single-Vehicle

1

57

1

1

GA Multiple-Vehicle

1

39

0

1

GA Total

2

96

1

2

Single-Vehicle

1

28

21

0

MS Multiple-Vehicle

0

40

10

0

MS Total

1

68

31

0

Single-Vehicle

23

38

0

0

SC Multiple-Vehicle

12

27

0

0

SC Total

35

65

0

0

Figure 33: Single-Vehicle Fatal Crashes by Road Surface Material 60

Figure 34: Multiple-Vehicle Fatal Crashes by Road Surface Material
3.3.2.10 Average Daily Traffic The average daily traffic (ADT) was provided to the researchers for the study sites. Though an ADT value is often an approximated traffic volume (particularly for lowvolume roads), the ADT value can still provide a context for the approximate traffic exposure per day. As shown in Figure 35, a large portion of the single-vehicle fatal crashes (about 40% of total fatal crashes) occurred at locations with average daily traffic (ADT) below 2000 vehicles per day (vpd). An additional 15% of fatal crashes occurred at sites with ADT values ranging from 2000 vpd to 6000 vpd. Figure 36 demonstrates that the majority of the multiple-vehicle fatal crashes (approximately 30% of the total fatal crashes) occurred at locations with ADT values below 6000 vpd.
Figure 35: Single-Vehicle Fatal Crashes by Average Daily Traffic 61

Figure 36: Multiple-Vehicle Fatal Crashes by Average Daily Traffic 3.3.2.11 Roadway Junction Proximity For the fatal crashes in this study, single-vehicle fatal crashes predominately occurred at non-junction locations (see Table 19 and Figure 37) with percentages ranging from 31% to 59% of total fatal crashes. Similar to the single-vehicle fatal crashes, most of multiplevehicle fatal crashes also occurred at non-junction locations at about one fourth of total fatal crashes on average (see Figure 38). More than 10% of the total fatal crashes, however, were multiple-vehicle crashes located in the proximity of intersections, such as four-way, T- and Y-intersections.
62

Table 19: Crash Type based on Roadway Junction Proximity

Roadway Junction Type

State

Crash Type

Non- Four-way

T-

Y-

Junction Intersection Intersection Intersection Others

(%)

(%)

(%)

(%)

(%)

Single-Vehicle

59

1

3

1

0

AL Multiple-Vehicle 23

7

3

1

2

AL Total

82

8

6

2

2

Single-Vehicle

52

1

7

0

0

GA Multiple-Vehicle 23

11

5

1

0

GA Total

75

12

12

1

0

Single-Vehicle

47

0

2

0

1

MS Multiple-Vehicle 32

11

2

2

3

MS Total

79

11

4

2

4

Single-Vehicle

31

7

10

3

11

SC Multiple-Vehicle 19

8

5

0

8

SC Total

50

15

15

3

19

Figure 37: Single-Vehicle Crashes by Roadway Junction Proximity 63

Figure 38: Multiple-Vehicle Crashes by Roadway Junction Proximity 3.3.2.12 Number of Driveways per Mile Due to the rural nature of the dataset, fatal crash sites were generally characterized by infrequent driveways. Though an increase in driveway density can be correlated with increased interactions between vehicles, a common single-vehicle driveway consideration is the disruption of traversable longitudinal roadside grading resulting in the potential launching of errant vehicles. As shown in Table 20 as well as Figure 39 and Figure 40, there is no substantial difference between crashes and driveway density for the singlevehicle or multiple-vehicle fatal crashes.
64

Table 20: Crash Type by Number of Driveways per Mile

State

Crash Type

0

(%)

Single-Vehicle

17

AL Multiple-Vehicle 13

AL Total

30

Single-Vehicle

18

GA Multiple-Vehicle 14

GA Total

32

Single-Vehicle

17

MS Multiple-Vehicle 16

MS Total

33

Single-Vehicle

24

SC

Multiple-Vehicle

8

SC Total

32

Number of Driveways per Mile

1

2

3

4

5+

(%) (%) (%) (%) (%)

15

11

10

8

3

6

7

5

1

5

21

18

15

9

8

17

7

6

4

7

11

9

2

3

3

28

16

8

7

10

16

5

6

2

4

9

10

9

0

6

25

15

15

2

10

15

11

4

3

2

11

11

3

4

2

26

22

7

7

4

Figure 39: Single-Vehicle Fatal Crashes for Number of Driveways per Mile 65

Figure 40: Multiple-Vehicle Fatal Crashes for Number of Driveways per Mile
3.3.2.13 Regulatory Speed Limit The regulatory speed limit assigned to a road at the time of a crash may not directly reflect the operating speeds during the crash, but can help identify the nature of the crash site (high speed road versus low speed road). As shown in Table 21, the majority of the study crashes for Georgia, Mississippi, and South Carolina occurred on roads with a 55 mph speed limit. Alabama single-vehicle crashes tended to occur at roads with 45 mph speed limits and multiple-vehicle crashes occurred on roads with 55 mph speed limits. Figure 41 and Figure 42 graphically demonstrate this observed trend.
66

Table 21: Crash Type per Regulatory Speed Limit

State Crash Type
Single-Vehicle AL Multiple-Vehicle
AL Total Single-Vehicle GA Multiple-Vehicle
GA Total Single-Vehicle MS Multiple-Vehicle
MS Total Single-Vehicle SC Multiple-Vehicle
SC Total

<=30 (%)
1 0 1 2 1 3 3 3 6 0 0 0

Posted Speed Limit (mph) 35 40 45 50 55 Unknown (%) (%) (%) (%) (%) (%)

5

3

23

1 13

0

2

3

18

0 31

0

7

6

41

1 44

0

5

0

4

0 28

4

3

1

6

0 45

0

8

1

10

0 73

4

3

0

1

1 22

3

0

2

5

0 56

1

3

2

6

1 78

4

3

1

9

0 29

0

3

3

10

1 40

0

6

4

19

1 69

0

Figure 41: Single-Vehicle Fatal Crashes per Regulatory Speed Limit 67

Figure 42: Multiple-Vehicle Fatal Crashes per Regulatory Speed Limit 3.3.2.14 Traffic Control Device The traffic control device variable was only available in the database for the Alabama, Georgia, and Mississippi fatal crashes. As shown in Table 22, Figure 43 and Figure 44, from 44% up to 88% of total fatal crashes occurred on roadways without traffic control devices in the immediate vicinity of the crash sites. Georgia sites did have warning signs present (as reported in the database) more than in the other two states. The researchers acquired this information from site inspections and visits.
68

Table 22: Crash Type and Associated Traffic Control Devices

State

Crash Type

Traffic Control Device Type

Flashing

Traffic Traffic

Railway

No Control Control Stop Warning Crossing

Control Signal Sign Sign Sign Device Unknown

(%) (%) (%) (%) (%)

(%)

(%)

Single-Vehicle 55

0

0

4

1

0

0

AL Multiple-Vehicle 33

0

1

7

0

0

0

AL Total

88

0

1

11

1

0

0

Single-Vehicle 31

0

0

2

18

1

4

GA Multiple-Vehicle 13

0

0

9

22

0

0

GA Total

44

0

0

11 40

1

4

Single-Vehicle 42

0

0

2

0

0

0

MS

Multiple-Vehicle 43

1

0

12

0

0

0

MS Total

85

1

0

14

0

0

0

Figure 43: Single-Vehicle Fatal Crashes and Associated Traffic Control Devices 69

Figure 44: Multiple-Vehicle Fatal Crashes and Associated Traffic Control Devices
3.3.2.15 Roadside Hazard Rating The original study included a roadside hazard rating (RHR) consistent with values ranging from one (for easily traversable roadside conditions) up to seven (roadsides that are extremely hazardous). Each research team assigned a RHR based on sample photographs of each respective rating value. Across the states as the RHR increased from one to five, the fatal crash occurrence steadily increased single-vehicle and multiplevehicle crashes (see Table 23, Figure 45, and Figure 46). RHR values for six and seven varied for the individual states.
70

Table 23: Crash Type and Associated Roadside Hazard Rating

Roadside Hazard Rating

State

Crash Type

1

2

3

4

5

6

7

(%) (%) (%) (%) (%) (%) (%)

Single-Vehicle

0

1

8

20 23 11 1

AL Multiple-Vehicle 1

3

7

14 11

1

0

AL Total

1

4

15 34 34 12

1

Single-Vehicle

1

7

16 17 16

3

1

GA Multiple-Vehicle 0

5

16 12

7

1

0

GA Total

1

12 36 29 23

4

1

Single-Vehicle

0

1

4

7

14 24 0

MS Multiple-Vehicle 0

2

3

16 21

8

0

MS Total

0

3

7

23 35 32 0

Single-Vehicle

0

0

0

12 43

6

0

SC Multiple-Vehicle 0

0

1

10 28

1

0

SC Total

0

0

1

22 71

7

0

Figure 45: Single-Vehicle Fatal Crashes and Associated Roadside Hazard Ratings 71

Figure 46: Multiple-Vehicle Fatal Crashes and Associated Roadside Hazard Ratings
3.3.2.16 Guardrail and Bridge Rail Type A guardrail or bridge rail type variable was only available for the three states of Alabama, Georgia, and Mississippi. The individual state researchers acquired this variable from site inspection. As presented in Table 24, more than 90% of the fatal crashes for each state occurred at locations without guardrails and bridge rails.

Table 24: Fatal Crashes and Associated Guardrail/Bridge Rails

State
AL GA MS

None (%)
91 98 96

Guardrail/Bridge Rail Type

Steel

Concrete Concrete

Breakaway Barrier Bridge Rail

(%)

(%)

(%)

3

1

1

1

0

1

3

0

1

Others (%)
3 0 0

3.3.2.17 Terrain
The road terrain variable describes the general regional terrain at crash locations. Terrain is defined loosely as flat terrain, rolling terrain, and mountainous terrain. As shown in

72

Table 25 and further reflected in Figure 47 and Figure 48, most of the crashes occurred in flat or rolling terrain regions.

State AL GA MS SC

Table 25: Crash Type and Associated Terrain

Crash Type

Flat

(%)

Single-Vehicle

39

Multiple-Vehicle

22

AL Total

61

Single-Vehicle

35

Multiple-Vehicle

16

GA Total

41

Single-Vehicle

8

Multiple-Vehicle

6

MS Total

14

Single-Vehicle

28

Multiple-Vehicle

19

SC Total

47

Terrain Rolling
(%) 24 14 38 22 24 46 42 44 86 32 20 52

Mountainous (%) 1 0 1 2 1 3 0 0 0 0 1 1

Figure 47: Single-Vehicle Fatal Crashes and Associated Terrain 73

Figure 48: Multiple-Vehicle Fatal Crashes and Associated Terrain 3.3.2.18 Relation to Roadway The relation to roadway variable identifies the vehicle location where the first harmful event occurred as it related to the traffic-way. This variable includes the roadway, shoulder, roadside, or other unknown off-road locations. According to this "relation to roadway" variable, a majority of single-vehicle fatal crashes occurred either at off-road or roadside locations, while multiple-vehicle fatal crashes occurred primarily on the roadway (see Table 26, Figure 49, and Figure 50).
74

State AL GA MS SC

Table 26: Crash Type and Relation to Roadway

Crash Type
Single-Vehicle Multiple-Vehicle
AL Total Single-Vehicle Multiple-Vehicle
GA Total Single-Vehicle Multiple-Vehicle
MS Total Single-Vehicle Multiple-Vehicle
SC Total

Relation to Roadway

Off-road or

Roadway Shoulder Roadside Unknown Location

(%)

(%)

(%)

(%)

5

0

0

59

36

0

0

0

41

0

0

59

7

8

2

43

41

0

0

0

48

8

2

43

0

1

49

0

47

0

3

0

47

1

52

0

4

4

11

42

38

0

2

0

42

4

13

42

Figure 49: Single-Vehicle Fatal Crashes and Relation to Roadway 75

Figure 50: Multiple-Vehicle Fatal Crashes and Relation to Roadway 3.3.3 Environment Related Characteristics 3.3.3.1 Ambient Light Condition The ambient light condition variable describes the lighting conditions at the time of the crashes. This variable includes daylight, dark unlit, and others (dawn, dusk, dark lighted, and unknown). As shown in Table 27, approximately one-third of the total fatal crashes were single-vehicle crashes that occurred when conditions were dark with no lighting (see also Figure 51), while only about one-tenth (ranging from 8% to 13%) of the total multiple-vehicle fatal crashes observed under dark unlit conditions (see Figure 52).
76

State AL GA MS SC

Table 27: Crash Type and Ambient Light Conditions

Crash Type
Single-Vehicle Multiple-Vehicle
AL Total Single-Vehicle Multiple-Vehicle
GA Total Single-Vehicle Multiple-Vehicle
MS Total Single-Vehicle Multiple-Vehicle
SC Total

Ambient Light Condition

Daylight

Dark Unlit

Others

(%)

(%)

(%)

24

37

3

27

8

1

51

45

4

27

31

1

27

12

1

54

43

2

13

33

4

34

13

3

47

46

7

22

36

3

24

11

4

46

47

7

Figure 51: Single-Vehicle Fatal Crashes and Ambient Lighting Conditions 77

Figure 52: Multiple-Vehicle Fatal Crashes and Ambient Light Conditions 3.3.3.2 Weather Condition The weather condition variable described the prevailing atmospheric conditions at the time of the crashes. This variable includes weather conditions ranging from clear, cloudy, and rain, to other less common conditions (fog, smoke, smog, sleet, hail, snow, etc). Single-vehicle and multiple-vehicle fatal crashes were similarly distributed for the various weather conditions (see Table 28, Figure 53 and Figure 54). More than twothirds (ranging from 64% to 83%) of the fatal crashes occurred during the clear weather condition.
78

Table 28: Crash Type by Weather Conditions

State AL GA MS SC

Crash Type
Single-Vehicle Multiple-Vehicle
AL Total Single-Vehicle Multiple-Vehicle
GA Total Single-Vehicle Multiple-Vehicle
MS Total Single-Vehicle Multiple-Vehicle
SC Total

Clear (%) 40 24 64 47 36 83 33 34 67 46 31 77

Weather Condition

Cloudy Rain

(%)

(%)

15

5

6

5

21

10

1

11

1

3

2

14

11

4

7

7

18

11

7

6

4

3

11

9

* Others: fog, smog, smoke, snow, etc.

Others* (%) 3 1 4 1 0 1 2 2 4 1 1 2

Figure 53: Single-Vehicle Crashes by Weather Conditions 79

Figure 54: Multiple-Vehicle Crashes by Weather Conditions
3.3.3.3 Road Surface Condition The road surface condition variable describes the crash time surface of the road condition due to current or previous weather (for example, the crash may have occurred during a cloudy weather condition but the road surface could still have been wet due to a previous rain). This variable includes three major conditions - dry surfaces, wet surfaces, and all other conditions (snow, ice, etc). As presented in Table 29, Figure 55, and Figure 56, 78% up to 89% of the fatal crashes occurred during dry road surface conditions.
80

Table 29: Crash Type and Associated Road Surface Conditions

State

Crash Type

Single-Vehicle AL Multiple-Vehicle
AL Total Single-Vehicle GA Multiple-Vehicle
GA Total Single-Vehicle MS Multiple-Vehicle
MS Total Single-Vehicle SC Multiple-Vehicle
SC Total

Road Surface Condition

Dry

Wet

Others

(%)

(%)

(%)

49

13

2

29

7

0

78

20

2

45

14

0

35

5

0

80

19

0

42

7

1

42

8

0

84

15

1

54

7

0

35

4

0

89

11

0

Figure 55: Single-Vehicle Crashes and Associated Road Surface Conditions 81

Figure 56: Multiple-Vehicle Crashes and Associated Road Surface Conditions 3.3.4 Driver and Passenger Related Characteristics 3.3.4.1 Gender Table 30, Figure 57 and Figure 58 present the gender distribution when drivers and passengers are shown collectively for single-vehicle and multiple-vehicle fatal crashes. More than twice as many male drivers were involved in the fatal crashes than female drivers. Male passengers still show more involvements in single-vehicle fatal crashes compared to their counterparts, while the pattern is no longer prominent or diminished in multiple-vehicle fatal crashes.
82

State AL GA MS SC

Table 30: Crash Type by Driver and Passenger Gender

Crash Type
Single-Vehicle Multiple-Vehicle
AL Total Single-Vehicle Multiple-Vehicle
GA Total Single-Vehicle Multiple-Vehicle
MS Total Single-Vehicle Multiple-Vehicle
SC Total

Person Involved

Driver

Passenger

Non-Motorist

Not Male Female Male Female Reported Male Female (%) (%) (%) (%) (%) (%) (%)

20 7

9

6

1

0

0

22 9

8 12

5

0

0

42 16 17 18

6

0

0

19 7

9

6

0

2

1

27 11 10 9

0

0

0

46 18 19 15

0

2

1

12 6 10 5

0

0

0

26 11 16 15

0

0

0

38 17 26 20

0

0

0

22 5 10 5

0

1

0

26 10 9 10

0

0

0

48 15 19 15

0

1

0

Figure 57: Single-Vehicle Fatal Crashes by Driver and Passenger Genders 83

Figure 58: Multiple-Vehicle Fatal Crashes by Driver and Passenger Genders

3.3.4.2 Driver Age Group

As shown in Table 31, Figure 59, and Figure 60, drivers ages 25 to 34 and ages 35 to 44 were involved more often in the single-vehicle and multiple-vehicle fatal crashes selected for this study. In addition, Alabama drivers in the 35 to 44 year old age group were involved in single-vehicle fatal crashes more often than those in a similar age group in the other three study states.

Table 31: Crash Type by Driver Age Group

Driver Age Group (years old) State Crash Type <16 16-20 21-24 25-34 35-44 45-54 55-64 65-69 >69
(%) (%) (%) (%) (%) (%) (%) (%) (%)

Single-Vehicle 0

5

4 12 15 5

1 1

2

AL Multiple-Vehicle 0 8

3 13 13 9 4 2 3

AL Total

0 13 7 25 28 14 5 3 5

Single-Vehicle 1

6

7 11 6 4

3 1

1

GA Multiple-Vehicle 1 7

5 14 11 7 5 3 5

GA Total

2 13 12 25 17 11 8 4 6

Single-Vehicle 1

6

2

8

7

3

3 2

1

MS Multiple-Vehicle 1 10 3 18 11 10 9 3 2

MS Total

2 16 5 26 18 13 12 5 3

Single-Vehicle 1

5

8 12 11 4

1 0

0

SC Multiple-Vehicle 0 7

3 12 13 10 8 2 3

SC Total

1 12 11 24 24 14 9 2 3

84

Figure 59: Single-Vehicle Fatal Crashes by Driver Age Group
Figure 60: Multiple-Vehicle Fatal Crashes by Driver Age Group 3.3.4.3 Ejection Status of Vehicle Occupants For a small percentage of crashes, a vehicle occupant was partially or completely ejected from the vehicle (see see Table 32, Figure 61, and Figure 62). With the exception of the South Carolina crashes, the percentage of ejected occupants involved in single-vehicle fatal crashes was almost twice as that of multiple-vehicle fatal crashes.
85

Table 32: Crash Type and Associated Occupant Ejection Status

State AL GA MS SC

Crash Type
Single-Vehicle Multiple-Vehicle
AL Total Single-Vehicle Multiple-Vehicle
GA Total Single-Vehicle Multiple-Vehicle
MS Total Single-Vehicle Multiple-Vehicle
SC Total

Ejection Status

Not Ejected (%)

Ejected (%)

29

13

49

6

78

19

30

14

49

7

79

21

20

13

59

8

79

21

35

8

45

12

80

20

Others (%) 2 1 3 1 0 1 0 0 0 0 0 0

Figure 61: Single-Vehicle Fatal Crashes and Associated Occupant Ejection Status 86

Figure 62: Multiple-Vehicle Fatal Crashes and Associated Occupant Ejection Status
3.3.4.4 Occupant Protection System Use The occupant protection system variable primarily represents safety restraint (not used, shoulder and lap belt use, shoulder belt use only, lap belt use only, and others such as well as helmet use for motorcyclists). Most of vehicle occupants in single-vehicle fatal crashes did not use safety restrains (this percentage ranged from 23% to 32% accounting for about one-third of total involved occupants for all crashes) as shown in Table 33 and Figure 63. Figure 64 further demonstrates that for multiple-vehicle fatal crashes the proportion of vehicle occupants who wore safety restraints compared to those without restraints were similar at approximately 25%.
87

Table 33: Crash Type Compared to Occupant Protection System Use

State AL GA MS SC

Crash Type
Single-Vehicle Multiple-Vehicle
AL Total Single-Vehicle Multiple-Vehicle
GA Total Single-Vehicle Multiple-Vehicle
MS Total Single-Vehicle Multiple-Vehicle
SC Total

Not Used (%)
32 28 60 27 22 49 23 24 47 32 23 55

Occupant Protection System Use

Shoulder and Shoulder Belt Lap Belt Other

Lap Belt

Only

Only Situations

(%)

(%)

(%)

(%)

6

0

1

4

22

0

2

5

28

0

3

9

6

1

0

10

23

0

3

8

29

1

3

18

6

0

0

3

26

9

2

6

32

9

2

9

9

0

0

3

29

0

1

4

38

0

1

7

Figure 63: Single-Vehicle Fatal Crashes and Occupant Protection System Use 88

Figure 64: Multiple-Vehicle Fatal Crashes and Occupant Protection System Use
3.3.5 Vehicle Related Characteristics 3.3.5.1 Vehicle Configuration The vehicle configuration variable describes the type of motor vehicles involved in the fatal crashes. This variable included passenger car, light truck, truck/trailer/ tractor, motorcycle, bus, other types, or unknown vehicle types. This variable is not available in the database for Georgia. In general, the vehicle configuration distributions for the three states exhibit similar distributions (see Table 34, Figure 65, and Figure 66) with passenger cars and light trucks as the most common vehicle types for both single-vehicle and multiple-vehicle fatal crashes.
89

Table 34: Crash Type and Associated Vehicles

State

Crash Type

Vehicle Configuration

Truck/

Passenger Light Tractor

Car Truck Trailer Motorcycle Bus Others Unknown

(%) (%) (%)

(%) (%) (%) (%)

Single-Vehicle 28

17

0

0

1 0

0

AL Multiple-Vehicle 30

16

7

0

0 0

0

AL Total

58

33

7

0

1 0

0

Single-Vehicle 18

13

1

0

0 0

1

MS Multiple-Vehicle 37

18

9

1

1 1

1

AL Total

55

31 10

1

1 1

2

Single-Vehicle 26

13

2

1

0 0

0

SC Multiple-Vehicle 30

17

7

3

0 0

0

AL Total

56

30

9

4

0 0

0

Figure 65: Single-Vehicle Fatal Crashes and Associated Vehicles 90

Figure 66: Multiple-Vehicle Fatal Crashes and Associated Vehicles
3.3.5.2 Vehicle Travel Speed Travel speed is the estimated speed of the vehicle before a crash. This variable is prone to be based on the judgment of the reporting law enforcement officer instead of an actual known driving speed of a vehicle. While only about 42% of the travel speed information was available for Mississippi crashes, Georgia does not include vehicle speed on crash reports so the Georgia database does not include this vehicle speed variable. For Alabama and South Carolina, the travel speed distribution of single-vehicle fatal crashes skews towards high speeds with observed speeds often greater than 55 mph. Alternatively, the travel speed distribution skews is towards lower speeds with the 46 to 55mph range accounting for a large portion of the multiple-vehicle fatal crashes.
91

Table 35: Crash Type by Vehicle Travel Speed

State Crash Type
Single-Vehicle AL Multiple-Vehicle
AL Total Single-Vehicle GA Multiple-Vehicle
GA Total Single-Vehicle MS Multiple-Vehicle
MS Total Single-Vehicle SC Multiple-Vehicle
SC Total

<= 30 (%)
0 10 10 0 0 0 1 4 5 2 13 15

Travel Speed (mph) 35-45 46-55 56-65 > 65 (%) (%) (%) (%)

9

9

12

12

13

21

6

3

22

30

18

15

0

0

0

0

0

0

0

0

0

0

0

0

1

5

2

1

7

20

2

0

8

25

4

1

2

10

9

19

16

22

3

2

18

32

12

21

Unknown (%) 3 2 5 44 56 100 23 35 58 1 1 2

Figure 67: Single-Vehicle Fatal Crashes by Vehicle Travel Speed 92

Figure 68: Multiple-Vehicle Fatal Crashes by Vehicle Travel Speed

3.3.5.3 Vehicle Maneuver
The vehicle maneuver variable describes the action taken by the driver of a vehicle prior to the crash. As presented in Table 36, Figure 69, and Figure 70, more than 70% of all involved vehicles were traveling in a straight direction prior to the fatal crashes.

Table 36: Crash Type and Associated Vehicle Maneuver

Vehicle Maneuver

Lane Changing/

State Crash Type

Passing/ Turning Entering Slowing/ Others/

Straight Overtaking Movement Traffic Stopped Unknown

(%)

(%)

(%)

(%) (%)

(%)

Single-Vehicle 31

2

1

0

0

11

AL Multiple-Vehicle 39

2

3

2

4

3

AL Total

70

4

4

2

4

14

Single-Vehicle 35

3

0

2

1

3

GA Multiple-Vehicle 47

3

5

2

0

0

GA Total

82

6

5

4

1

3

Single-Vehicle 28

1

0

0

0

4

MS Multiple-Vehicle 57

3

4

1

1

2

MS Total

85

4

4

1

1

6

Single-Vehicle 42

0

0

0

0

1

SC Multiple-Vehicle 48

2

4

3

0

0

SC Total

90

2

4

3

0

1

93

Figure 69: Single-Vehicle Fatal Crashes and Associated Vehicle Maneuver
Figure 70: Multiple-Vehicle Fatal Crashes and Associated Vehicle Maneuver 94

3.3.5.4 Extent of Damage Towing Status
It is not surprising to observe that more than 90% of the vehicles involved in the fatal crashes were seriously damaged and towed from the crash scene (see Table 37). This depicts the severe vehicle destruction common to fatal crashes.

Table 37: Crash Type and Associated Vehicle Condition (Towing Status)

State AL GA MS SC

Crash Type
Single-Vehicle Multiple-Vehicle
AL Total Single-Vehicle Multiple-Vehicle
GA Total Single-Vehicle Multiple-Vehicle
MS Total Single-Vehicle Multiple-Vehicle
SC Total

Driven (%) 0 0 0 1 2 3 0 2 2 1 2 3

Vehicle Towing Status

Towed (%)

Abandoned/Unknown (%)

45

0

54

0

99

0

37

5

53

1

90

6

33

0

64

1

97

1

41

1

55

0

96

1

3.3.5.5 Vehicle Model Year
The age of a vehicle can be directly associated with crash severity, as newer vehicles should be equipped with advanced occupant protection features not common to earlier vehicle models. Less than 10% of the fatal crashes involved vehicles older than 20 years. Table 38, Figure 71, and Figure 72 demonstrate typical vehicle ages with most vehicles ten years old or newer (recall crashes are from the years 1997 through 1998 for the various states in this study). The vehicle age distribution was similar for the singlevehicle and multiple-vehicle crashes.

95

Table 38: Crash Type and Associated Vehicle Model Year

State Crash Type
Single-Vehicle AL Multiple-Vehicle
AL Total Single-Vehicle GA Multiple-Vehicle
GA Total Single-Vehicle MS Multiple-Vehicle
MS Total Single-Vehicle SC Multiple-Vehicle
SC Total

<= 1977 (%) 5 3 8 3 2 5 4 4 8 1 2 3

Vehicle Model Year

1978 1987 1988 1992

(%)

(%)

19

13

18

18

37

31

11

10

17

20

28

30

8

9

21

20

29

29

14

17

13

18

27

35

>= 1993 (%) 9 15 24 17 20 37 12 22 34 11 24 36

Figure 71: Single-Vehicle Fatal Crashes and Associated Vehicle Model Year 96

Figure 72: Multiple-Vehicle Fatal Crashes and Associated Vehicle Model Year 97

This page blank intentionally 98

4.0 Ten-Year Historic Fatal Crash Trends in the Southeastern United States
This chapter presents an examination of fatal crashes that occurred during the past decade (1997 to 2006) in four southeastern US states (i.e., Alabama, Georgia, Mississippi, and South Carolina). The objective is to identify possible longer terms trends, cycles, or anomalies in fatal crash occurrence. Section 4.1 presents an analysis of trends in fatal crashes, while Section 4.2 examines possible factors associated with changes in fatal crashes during the past decade. Specifically, fatal crashes are categorized by a variety of known contributing crash factors (e.g. use of alcohol and safety restraints) and assessed statistically. All data used in this chapter were obtained from the FARS (Fatality Analysis Reporting System) at http://www-fars.nhtsa.dot.gov. A discussion in Section 4.3 highlights the important findings.
4.1 Fatal Crash Trends 4.1.1 Fatal Crash Frequency The number of fatal crashes that occurred during the past decade in the southeastern United States (the four previously identified states) is summarized in Table 39. In all states, fatal crashes increased during the period from 1997 to 2006 (see Figure 73), although there is significant fluctuation from year to year. During the past decade, fatal crash frequencies in Georgia are higher than those in the remaining three states, while fatal crash frequencies in Mississippi are lowest among the four states. A consistent `up' or `down' trend among the states might suggest a national effect--such as an effect of gasoline prices--however the research team did not observe this in any given year, suggesting that yearly fluctuations are largely random or caused by `within-state' effects.
99

Table 39: Fatal Crash Frequency from 1997 to 2006 in the Southeastern US

Year 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

Alabama 1050 958 992 910 900 931 902 1032 1021 1073

Georgia 1405 1414 1314 1380 1471 1362 1463 1463 1582 1557

Mississippi 741 842 832 846 704 769 786 786 840 812

South Carolina 798 912 944 948 962 949 905 946 981 973

Fatal crash frequency/year 600 800 1000 1200 1400 1600

1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 Year

Alabama Mississippi

Georgia South Carolina

Figure 73: Fatal Crash Trend in the Southeastern US (1997 to 2006)

4.1.2 Vehicle Miles Traveled
While total fatal crash frequency differences are insightful, they do not account for aggregate exposure to risk. Table 40 shows the estimated total vehicle miles traveled (VMT) from 1997 to 2006 in the four states. The statistics show that VMT in Georgia is the greatest among the four states, while VMT in Mississippi is the smallest. It is

100

noteworthy that VMTs in Alabama, Georgia, and Mississippi have been monotonically increasing (see Figure 74). In South Carolina, the VMT in 2005 had slightly decreased from 2004, but the VMT in 2006 increased by 22% when comparing to the VMT in 1997. These statistics suggest that increases in fatal crashes are associated with increases in VMT (see Figure 75). Note that the relationship between VMT and fatal crashes illustrated in Figure 75 is presented using a smoothed curve, which reflects a possible nonlinear relationship (see http://en.wikipedia.org/wiki/Lowess).

Table 40: Vehicle Miles Traveled (Billions) in the Southeastern US (1997 to 2006)

Year 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

Alabama 53 55 56 57 57 58 59 59 60 61

Georgia 94 96 99 105 108 108 109 114 114 114

Mississippi 32 34 35 36 36 36 37 39 40 41

South Carolina 41 43 44 46 47 47 48 50 49 50

100 120

80

Vehicle Miles Traveled (Billions)

60

40

1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 Year

Alabama Mississippi

Georgia South Carolina

Figure 74: VMT (Billions) Trend in the Southeastern US (1997 to 2006)

101

Examination of this curve suggests that fatal crashes per year on aggregate increase somewhat linearly with VMT. Note that the variability in VMT in Alabama is quite small, so the variation seen in fatal crashes per year is due to both random and systematic influences within the state.
Figure 75: The Relationship between VMT and Fatal Crashes (Lowess Curve)
4.1.3 Fatality Rates Table 41 shows the fatality rate per 100 Million VMT by state and year. Such a value is widely used to compare the fatal crashes among states by controlling for the exposure. Overall, the fatality rates of the four states appear to decrease during the period from 1997 to 2006, although some fluctuations exist (see Figure 76). The statistics show that the fatality rate in Georgia is lowest among the four states shown in Figure 76, although the fatal crash frequency in Georgia is consistently highest among them (see Figure 73). In contrast, the fatality rate in Mississippi is highest among the four states, while the fatal crash frequency in Mississippi is consistently lowest among them during the period from
102

1997 to 2006. On average these four states have fatality rates above the national average for all years with few exceptions. Georgia's rate is lowest, followed by Alabama, then South Carolina, then Mississippi. These differences may reflect the influence of a large number of potential factors between states, including but not limited to weather, driver behavior, impaired driving, restraint use, speeding and speed limits, roadside hazards, and level of urbanization. For example, using level of urbanization as a possible and likely contributor to the observed fatality rate differences, more urbanized states will have a higher proportion of VMT during times of congestion when speeds are constrained and the likelihood of a fatal crash is reduced. As a result, while the differences across states reflect meaningful safety differences, it is not known to what factors the differences are attributable.

Table 41: Fatality Rate per 100 Million VMT by State and Year

Year

Alabama

Georgia

1997

2.23

1.69

1998

1.94

1.63

1999

2.03

1.52

2000

1.76

1.47

2001

1.75

1.53

2002

1.8

1.41

2003

1.71

1.47

2004

1.95

1.44

2005

1.92

1.52

2006

1.99

1.49

* USA Average fatality rate per 100 Million VMT

Mississippi 2.73 2.77 2.66 2.67 2.18 2.43 2.33 2.28 2.32 2.2

South Carolina 2.18 2.34 2.41 2.34 2.27 2.23 2.01 2.11 2.21 2.08

USA* 1.64 1.58 1.55 1.53 1.51 1.51 1.48 1.44 1.46 1.42

103

3

2.5

2

Fatality rate per 100 Million VMT

1.5

1

1997

1998

1999

2000 2001 2002 2003 2004 2005 Year

Alabama Mississippi USA

Georgia South Carolina

2006

Figure 76: Fatality Rate per 100 Million VMT by State and Year
The fatal crash rates shown in Figure 76 reflect a decreasing trend from 1997 to 2006 for each of the states. Table 42 shows that the fatal crash rate in Mississippi decreased by 24%, while the rate in South Carolina decreased by about 5% over that same period. Figure 77 shows the trend over the decade by state compared to the national trend which has a lower crash rate value than all four states. The rate of reduction in crash rate was highest in Mississippi and lowest in South Carolina, while the change in Alabama and Georgia was similar to the national average. Again, the reduction in crash rates spanning from 1997 to 2006 can be caused by a number of factors, including but not limited to investments in safety improvements, increased urbanization, improved vehicle safety features in the vehicle fleet, reduced impaired driving, increased restraint use, etc. It is not known which combinations of factors are responsible for the decrease across states.

104

Table 42: Percent Change in Fatality Rate per 100 Million VMT (1997 to 2006)

Year
1997 2006 Percent Change

Alabama 2.23 1.99
-12.1%

Georgia 1.69 1.49
-13.4%

Mississippi 2.73 2.2
-24.1%

South Carolina 2.18 2.08
-4.8%

USA* 1.64 1.42
-15.5%

Figure 77: Change in Fatality Rate per 100 Million VMT from 1997 to 2006 To summarize the major crash trends in the four southeastern US states we note that crash frequencies have increased, VMT has increased at an even greater rate than crash frequencies, and their ratios--frequencies divided by VMT--has decreased. Of course the causes of these changes are not known at this stage, and in fact cannot be known with certainty through observational studies such as this one. Instead, we examine possible contributing factors and draw some conclusions on which factors may serve to help explain some of the trends in fatal crashes observed over the recent decade.
105

4.2 Potential Contributing Factors Affecting Fatal Crash Trends This section focuses on the change in the portion of fatal crashes between 1997 and 2006 and attempts to associate these changes with a variety of potential contributing factors. The following factors are examined:
z Day of Week z Time of Day z Roadway Functional Classification z Traffic Control Devices z Urban/Rural z Alcohol Involvement z Restraint Usage The factors above are categorized by temporal, spatial, and behavioral factors, and the analysis results for each factor are described in the subsequent subsections.
4.2.1 Temporal Factors 4.2.1.1 Day of Week
During the period from 1997 to 2006 about 50% of fatal crashes occurred from Friday through Sunday in the Southeastern US states (see Table 43). The relative fatal crash frequencies during weekdays (Mondays to Thursdays) are less than the relative fatal crash frequency (0.14=1/7) expected in the absence of a relationship between day of week and fatal crashes. In contrast, the relative fatal crash frequencies during Fridays and weekends are higher than the expected relative fatal crash frequency of 0.14. This result is not unexpected, as impaired driving is increased and congestion is reduced during weekends.
To statistically assess the variation in fatal crashes among states, the research team conducted a homogeneity test using the Pearson's 2 test. In tests of homogeneity, each proportion of the fatal crashes for each day of week is assumed to be homogenous for all
106

states. Thus, the expected frequencies are computed by assuming the homogeneity, and the Pearson's chi-squared statistics is obtained using:

r
( ) X 2 =

c

Oij - Eij

2
,

(2)

ij

Eij

Where:

Oij is the observed fatal crashes in each cell in Table 43,

Eij is the expected fatal crashes when assuming the homogeneity, and r and c are the number of rows and columns respectively.

Note that the expected fatal crash frequency is obtained as:

Eij

=

OiiOi j rc

,

(3)

Where: Oii and Oi j are the marginals of the ith row and the jth column respectively.

Equation (3) is easily obtained using the concept of the independency between two events (i.e. the distribution of fatal crashes of each day of week is the same among states). The results of this test are shown in the Table 43 footnote. The critical value of chi-square with 18 degrees of freedom and alpha set to 0.05 is 9.39, while the test statistic value (shown in the table) is 50.906 and reflects a p-value of 0.001. The correct interpretation (and similar interpretations henceforth) of this finding is that there is less than a 0.001 chance of observing differences this large across days of the week and state if in fact day of week and state are independent. We conclude, therefore, that day of week and specific state influence the probability of a fatal crash.

107

Table 43: Fatal Crashes by State and Day of Week

State Alabama

Monday 1213* (12.42)**

Tuesday Wednesday Thursday

1127

1186

1281

(11.54) (12.14) (13.11)

Georgia

1864 (12.94)

1825 (12.67)

1851 (12.85)

1896 (13.16)

Mississippi

995 (12.5)

955

937

988

(12) (11.77) (12.42)

South

1116

1095

1134

1175

Carolina (11.98) (11.75) (12.17) (12.61)

Total

5188

5002

(12.52) (12.07)

Pearson chi2(18) = 50.906, Pr <0.001

* Fatal crash frequencies

** Within-row relative frequencies (%)

5108 (12.32)

5340 (12.88)

Friday 1604 (16.42) 2302 (15.98) 1278 (16.06) 1503 (16.13) 6687 (16.13)

Saturday 1837 (18.8) 2486 (17.26) 1565 (19.67) 1865 (20.02) 7753 (18.70)

Sunday 1521 (15.57) 2180 (15.13) 1240 (15.58) 1430 (15.35) 6371 (15.37)

Total 9769 (100) 14404 (100) 7958 (100) 9318 (100) 41449 (100)

Figure 78 shows a graph of the expected frequency of crashes by day (if all days had equal proportions) compared to the observed frequencies by day of the week for the four southeastern states examined.

.2

.15

Relative fatal crashes

.1

.05

0

Moday Tuesday WednesdayThursday Friday Saturday Sunday
Figure 78: Fatal Crashes by State and Day of Week
108

Figure 79 depicts the proportion of fatal crashes by time of day and state. The proportion of fatal crashes in Georgia that occurred during weekends and on Fridays is slightly higher than those in the remaining states. Thus, when conducting the homogeneity test excluding Georgia, the homogeneity assumption is not rejected at =0.05, i.e. the proportions of fatal crashes by day of week are homogeneous across the remaining three states.
Figure 79: Fatal Crashes by State and Day of Week (1997 to 2006) As previously shown, the aggregate fatal crash data can be examined with respect to state and day of week. We now turn to the change in fatal crashes between 1997 and 2006 by day of week and state. Table 44 shows the homogeneity test results, which indicates that the variation in fatal crashes between 1997 and 2006 is not affected by the day of week (p-values are greater than 0.05 for all tests). In other words, the proportions of fatal crashes by day of week between 1997 and 2006 did not increase or decrease significantly during this decade, even though the fatal crash frequency increased from 1997 to 2006. This result suggests that factors that influence weekend crashes have not significantly changed during this decade. The most likely factors that could influence day of week are impaired driving and increased differences in congestion from weekday to weekend travel. There is no evidence that these factors changed during the decade of observation.
109

Table 44: Change in Fatal Crashes between 1997 and 2006 by Day of Week and State

State

Year 1997

Monday Tuesday Wednesday Thursday

121*

116

116

151

(11.52)** (11.05) (11.05) (14.38)

Friday 183
(17.43)

Saturday 201
(19.14)

Sunday 162
(15.43)

Alabama

2006

129

119

(12.02) (11.09)

126 (11.74)

137

177

(12.77) (16.5)

206

179

(19.2) (16.68)

Pearson chi2(6) = 2.148 Pr =0.906; * Fatal crash frequencies; ** Within-row relative frequencies (%)

1997

152

183

(10.84) (13.05)

184 (13.12)

194

231

242

216

(13.84) (16.48) (17.26) (15.41)

Georgia

2006

212

195

(13.62) (12.52)

200 (12.85)

203

239

280

228

(13.04) (15.35) (17.98) (14.64)

Pearson chi2(6) = 6.266 Pr=0.394

1997

91

75

(12.28) (10.12)

95 (12.82)

86

128

147

119

(11.61) (17.27) (19.84) (16.06)

Mississippi 2006

113

101

(13.92) (12.44)

83 (10.22)

105

133

151

126

(12.93) (16.38) (18.6) (15.52)

Pearson chi2(6) = 6.029 Pr=0.420

South Carolina

1997 2006

95 (11.9) 108 (11.1)

100 (12.53)
111 (11.41)

105 (13.16)
112 (11.51)

104 (13.03)
126 (12.95)

122 (15.29)
164 (16.86)

152 (19.05)
198 (20.35)

120 (15.04)
154 (15.83)

Pearson chi2(6) = 2.905 Pr=0.821

4.2.1.2 Time of Day
Table 45 shows the cross-classification table of fatal crashes by state and time of day (TOD). About 17% of fatal crashes occurred during the time period between 3:00 p.m. and 5:59 p.m., while about 8% of fatal crashes occurred during the time period between 3:00 a.m. and 5:59 a.m. These statistics also reveal that fatal crashes in the four states did not occur at the same rate for each TOD.

110

Table 45: Fatal Crashes by State and Time of Day (1997 to 2006)

State

Midnight to 3:00 am to 6:00 am to 9:00 am to 2:59 am 5:59 am 8:59 am 11:59 am

Alabama

1013* (10.37)**

818 (8.38)

1031

1017

(10.56) (10.41)

Noon to 2:59 pm
1359 (13.92)

3:00 pm to 6:00 pm to 9:00 pm to

5:59 pm 8:59 pm 11:59 pm

1671

1577

1280

(17.11) (16.15) (13.11)

Georgia

1445 (10.1)

1055 (7.37)

1602

1510

1942

2458

2313

(11.19) (10.55) (13.57) (17.17) (16.16)

1989 (13.9)

Mississippi

857 (10.82)

609 (7.69)

931

813

1069

1351

1253

1041

(11.75) (10.26) (13.49) (17.05) (15.81) (13.14)

South 1209

902

919

793

1034

1452

1506

1500

Carolina (12.98) (9.68) (9.87) (8.51) (11.1) (15.59) (16.17) (16.1)

Total

4524 (10.95)

3384 (8.19)

4483 (10.85)

4133 (10)

5404

6932

6649

5810

(13.08) (16.78) (16.09) (14.06)

Pearson chi2(21) = 217.75 Pr<0.001; * Fatal crash frequencies; ** Within-row relative frequencies (%)

Table 45 also shows that the proportions of fatal crashes across states are statistically different at =0.05 (p-value < 0.001). This result is explained largely by the relatively high proportions of fatal crashes in South Carolina occurring between midnight and 5:59 a.m. (about 22% compared to 19% on average elsewhere).

The research team further analyzed the change in the proportions of fatal crashes between 1997 and 2006 by time of day as well as state. Table 46 shows the change in fatal crashes between 1997 and 2006 when controlling for time of day and state. Again, the homogeneity test results show that the proportions of fatal crashes in each state are statistically the same between 1997 and 2006. The analysis results in this section show that temporal factors (day of week and time of day) do not significantly affect the proportions of fatal crashes in the southeastern states between 1997 and 2006, although fatal crash frequency increased from 1997 to 2006.

In summary temporal factors are associated with the occurrence of fatal crashes in the four states examined. These factors differ slightly across the four states, but have remained constant within the states during the decade of analysis.

111

Table 46: Change in Fatal Crashes between 1997 and 2006 by Time of Day and State

State

Year

Midnight 3:00 am to 6:00 am to 9:00 am to to 2:59 am 5:59 am 8:59 am 11:59 am

Noon to 2:59 pm

3:00 pm to6:00 pm to9:00 pm to 5:59 pm 8:59 pm 11:59 pm

1997

96* (9.15)**

87 (8.29)

110

118

(10.49) (11.25)

149 (14.2)

157

183

(14.97) (17.45)

149 (14.2)

Alabama

2006

117 (10.9)

89

113

107

168

194

164

121

(8.29) (10.53) (9.97) (15.66) (18.08) (15.28) (11.28)

Pearson chi2(7) = 11.384 Pr=0.123; * Fatal crash frequencies; ** Within-row relative frequencies (%)

1997

129 (9.23)

95

147

135

192

253

249

197

(6.8) (10.52) (9.66) (13.74) (18.11) (17.82) (14.1)

Georgia

2006

169 (10.92)

131 (8.46)

181 (11.69)

144 (9.3)

193

264

252

214

(12.47) (17.05) (16.28) (13.82)

Pearson chi2(7) = 8.155 Pr=0.319

1997

72 (9.77)

78

96

(10.58) (13.03)

67 (9.09)

97

128

110

89

(13.16) (17.37) (14.93) (12.08)

Mississippi 2006

85 (10.47)

67 (8.25)

90 (11.08)

63 (7.76)

108

136

136

127

(13.3) (16.75) (16.75) (15.64)

Pearson chi2(7) = 8.883 Pr=0.261

South Carolina

1997 2006

122 (15.31)
139 (14.29)

61 (7.65) 105 (10.79)

85 (10.66)
102 (10.48)

77 (9.66)
69 (7.09)

85 (10.66)
104 (10.69)

102 (12.8) 139 (14.29)

130 (16.31)
147 (15.11)

135 (16.94)
168 (17.27)

Pearson chi2(7) = 9.576 Pr=0.214

4.2.2 Spatial Factors
This section reviews an analysis of the proportions of fatal crashes between 1997 and 2006 by several spatial factors such as roadway functional class, traffic control, and speed limit.
4.2.2.1 Roadway Function Class
Before analyzing the change in fatal crashes by roadway function class, it was necessary to simplify the roadway functional categories shown in Table 47. Table 48 shows the simplified roadway functional classes used in the following analyses. Table 49 shows the cross-classification table for fatal crashes by state and roadway functional class. The statistics show that about 35% of fatal crashes occur on collectors. The spatial distributions of fatal crashes by functional class are different across states. For example, about 34% of fatal crashes in Mississippi occurred on local roads or streets, compared to a range from 7% to 12% for the same roadway class in the remaining states.

112

Consequently, the homogeneity test results show that the distributions of fatal crashes by roadway class are statistically different across states at =0.05. Figure 80 shows the percent of crashes by functional class and state for the 10 year period.

Table 47: Roadway Functional Class Code (used in FARS)

Code 1 2 3 4 5 6 9 11 12 13 14 15 16 19 99

Definition Rural Principal Arterial-Interstate Rural Principal Arterial-Other Rural Minor Arterial Rural Major Collector Rural Minor Collector Rural Local Road Or Street Unknown Rural Urban Principal Arterial-Interstate Urban Principal Arterial-Other Freeways Or Expressways Urban Other Principal Arterial Urban Minor Arterial Urban Collector Urban Local Road Or Street Unknown Urban Unknown

Table 48: Simplified Roadway Functional Class

Code 1 2 3 4 5

Definition Interstate and Freeways Arterial Collector Local Road or Street Unknown

Table 49: Fatal Crashes by State and Roadway Functional Class (1997 to 2006)

State

Interstate and Freeways

Arterials

Collectors

Local Roads and Streets

Unknown

Alabama

1663* (17.82)**

3396 (36.38)

3124 (33.47)

1107 (11.86)

44 (0.47)

Georgia

2056 (15.24)

4773 (35.38)

4590 (34.03)

1413 (10.48)

657 (4.87)

Mississippi

960 (14.39)

1546 (23.17)

1847 (27.68)

2292 (34.35)

27 (0.4)

South Carolina

1097 (11.08)

3232 (32.66)

4355 (44)

661 (6.68)

552 (5.58)

Total

5776 (14.66)

12947 (32.87)

13916 (35.33)

5473 (13.89)

1280 (3.25)

Pearson chi2(12) = 3191.301; p-value < 0.001; * Fatal crash frequencies; ** Within-row relative frequencies (%)

113

Figure 80: Percent of Fatal Crashes by State (1997 to 2006)
An interesting question is whether the spatial distribution of crashes has changed over time in any of the states. The spatial distributions of fatal crashes as a function of roadway classification between 1997 and 2006 are shown in Table 50. It shows that the distribution of fatal crashes by roadway class in all four states is significantly altered from 1997 to 2006 (i.e., the p-values of the homogeneity tests are less than 0.05). Most notably in Alabama the proportion of fatal crashes that occurred on interstates and freeways increased from 14% to 23%, while the proportion of fatal crashes that occurred on arterials decreased from 45% to 30%. In Georgia, the proportion of fatal crashes that occurred on arterials decreased from 46% to 39%, while the proportion of fatal crashes without roadway functional class information increased from 0.5% to 13.94%. In Mississippi, the proportion of fatal crashes that occurred on collectors remarkably increased from 30% to 50%, while the proportion of fatal crashes that occurred on local roads and streets reduced from 35% to 16%. In South Carolina, the proportion of fatal crashes that occurred on local roads and streets reduced from 13% to 0%, while a nearly 6% rise in unknown functional class was also observed.
114

The analysis results show that the spatial distributions of fatal crashes by the roadway function class between 1997 and 2006 are significantly different. Figure 81 shows the change in the proportions of fatal crashes between 1997 and 2006 by the roadway class and state.

Table 50: Change in Fatal Crashes by Roadway Class and State (1997 to 2006)

State/Year

Interstate and Freeways

Arterials

Collectors

1997

149 * (14.19)**

472 (44.95)

291 (27.71)

Alabama

2006

246 (22.93)

322 (30.01)

314 (29.26)

Pearson chi2(4) =68.571; p-value < 0.001;
* Fatal crash frequencies; ** Within-row relative frequencies (%)

1997

181 (12.88)

652 (46.41)

333 (23.7)

Georgia

2006

190 (12.2)

612 (39.31)

301 (19.33)

Pearson chi2(4) =192.735; p-value < 0.001

1997

88 (11.88)

169 (22.81)

222 (29.96)

Mississippi 2006

98 (12.07)

177 (21.8)

408 (50.25)

Pearson chi2(4) =98.065; p-value < 0.001

South Carolina

1997 2006

83 (10.4) 114 (11.72)

336 (42.11)
460 (47.28)

241 (30.2) 320 (32.89)

Pearson chi2(4) =135.558; p-value < 0.001

Local Roads and Streets
136 (12.95)
175 (16.31)
232 (16.51)
237 (15.22)
261 (35.22)
128 (15.76)
101 (12.66)
0 (0)

Unknown
2 (0.19)
16 (1.49)
7 (0.5) 217 (13.94)
1 (0.13)
1 (0.12)
37 (4.64)
79 (8.12)

The observed differences over time and across states by functional class raise some interesting questions. Regarding temporal shifts within states, why would substantial shifts in fatal crash proportions across functional classifications occur? Certainly there are a number of possible explanations. Among them are shifts in reporting accuracy, changes in facility design over time, and significant shifts in exposure by functional class over time. These within-state shifts point to opportunities for further exploration as to what are causing or explaining these shifts in fatal crash proportions.

115

Figure 81: Change in Fatal Crashes by Roadway Class and State (1997 to 2006) 116

Differences across states are also interesting. Do states define functional classes consistently? Are there significant differences in exposure by functional class across states? Finally, are transportation system development patterns sufficiently different so as to introduce significantly different proportions of functional class roadways? 4.2.2.2 Traffic Control In addition to the roadway functional class, the proportions of fatal crashes by traffic control are analyzed. Table 51 shows traffic control descriptions summarized in FARS Encyclopedia (http://www-fars.nhtsa.dot.gov/Common/IdDefs.aspx), where the four classifications are presented (None, Traffic Signal, Stop, and Other/Unknown). Using the general traffic control classifications, the proportions of fatal crashes between 1997 and 2006 are summarized by state and traffic control. The statistics in Table 52 show that about 82% of fatal crashes occurred at locations without any of the designated traffic controls. As before, the proportions of fatal crashes that occurred at locations without traffic control are heterogeneous across states (p-value < 0.001). In Alabama, the proportion of fatal crashes that occurred at locations without traffic control was about 78%, compared to the average proportion of 82%. In Mississippi, the proportion of fatal crashes that occurred at locations without traffic control was about 87% (see Figure 82).
117

Table 51: Traffic Control Description

Traffic Control

Traffic Control Description

None

No Controls

Traffic control signal (on colors) without pedestrian signal

Traffic control (on colors) with pedestrian signal

Traffic control signal(on colors) not known whether or not pedestrian signal

Flashing traffic control signal

Traffic Signal Flashing beacon

Flashing highway traffic signal, type unknown or other than traffic control or beacon

Lane use control signal

Other highway traffic signal

Unknown highway traffic signal

Stop Sign Stop Sign

Yield Sign

Other regulatory sign

Unknown type regulatory sign

School speed limit sign

School advance or crossing sign

Other school related sign

Unknown type school zone sign

Warning Sign

Officer, crossing guard, flagman, etc.

Gates

Flashing Lights

Traffic Control Signal

Other/Unknown Wigwags

Bells

Other train activated device

Active device, type unknown

Cross bucks

Stop sign

Other railroad crossing sign

Special warning device-watchman, flagged by crew

Other passive device

Passive device, type unknown

Grade crossing controlled, type unknown

Other

Unknown

118

Table 52: Fatal Crashes by State and Traffic Control (1997 to 2006)

State

None

Traffic Signal Strop Sign Other/Unknown

Total

Alabama

7649* (78.3)**

576 (5.9)

838 (8.58)

706 (7.23)

9769 (100)

Georgia

12014 (83.37)

823 (5.71)

1457 (10.11)

117 (0.81)

14411 (100)

Mississippi

6884 (86.5)

118 (1.48)

790 (9.93)

166 (2.09)

7958 (100)

South Carolina

7539 (80.91)

438 (4.7)

1009 (10.83)

332 (3.56)

9318 (100)

Total

34086 (82.22)

1955 (4.72)

4094 (9.88)

1321 (3.19)

41456 (100)

Pearson chi2(9) = 1091.74; p-value < 0.001; * Fatal crash frequencies; ** Within-row relative frequencies (%)

Figure 82: Portion of Fatal Crashes by State and Traffic Control (1997 to 2006) The proportion of fatal crashes occurring at traffic signals in Mississippi is about 1% compared to about 5% in each of the other states.
The research team then examined the changes in the proportions of fatal crashes between 1997 and 2006 by state as well as traffic control. The results summarized in Table 53
119

show that the spatial distributions of fatal crashes in terms of traffic control between 1997 and 2006 did not statistically vary over time in Georgia and Mississippi (p-values are 0.313 and 0.189 respectively). However, in Alabama and South Carolina, the proportions of fatal crashes that occurred at locations without traffic control appeared to have increased (see Figure 83). It also turns out that these two states also had the largest proportion of unknown traffic control classifications, which could explain part or most of these differences over time.

Table 53: Change in Fatal Crashes by Traffic Control and State (1997 to 2006)

State/Year

None Traffic Signal Strop Sign Other/Unknown Total

1997

885* (84.29)**

53 (5.05)

94 (8.95)

18 (1.71)

1050 (100)

Alabama

2006

838 (78.1)

46 (4.29)

86 (8.01)

103 (9.6)

1073 (100)

Pearson chi2(3) =61.601; p-value<0.001; * Fatal crash frequencies; ** Within-row relative frequencies (%)

1997

1164 (82.85)

82 (5.84)

144 (10.25)

15 (1.07)

1405 (100)

Georgia

2006

1319 (84.71)

79 (5.07)

150 (9.63)

9 (0.58)

1557 (100)

Pearson chi2(3) =3.563; p-value=0.313

1997

629 (84.89)

7 (0.94)

82 (11.07)

23 (3.1)

741 (100)

Mississippi 2006

709 (87.32)

13 (1.6)

73 (8.99)

17 (2.09)

812 (100)

Pearson chi2(3) =4.770; p-value=0.189

South Carolina

1997 2006

616 (77.19)
786 (80.78)

32 (4.01)
48 (4.93)

99 (12.41)
108 (11.1)

51 (6.39)
31 (3.19)

798 (100) 973 (100)

Pearson chi2(3) =11.907; p-value=0.008

120

Figure 83: Change in Fatal Crashes by Traffic Control and State (1997 to 2006) 121

4.2.3 Behavioral Factors In Sections 4.2.1 and 4.2.2 we observed that the distributions of the fatal crashes in the four examined Southeastern states are affected by the spatial factors. In contrast, the distributions of the fatal crashes between 1997 and 2006 are homogenous when comparing by day of week and time of day. This section investigates whether or not fatal crashes from 1997 to 2006 are associated with behavioral factors such as alcohol involvement and safety restraint use.
4.2.3.1 Alcohol Involvement Alcohol-impaired crashes are the crashes that involve at least one driver or motorcycle rider (operator) with a Blood Alcohol Concentration (BAC) of .08 grams per deciliter (g/dL) or higher (http://www-fars.nhtsa.dot.gov/Crashes/CrashesAlcohol.aspx). It is commonly known that the impairment status of drivers involved in crashes has been notoriously poorly or under-reported in past studies focused on impaired driving--so the numbers examined in this section should be treated with caution. Figure 84 shows fatal crashes by state and impairment status. Alcohol impaired fatal crashes are lowest in Georgia and highest in South Carolina.
Table 54 shows fatal crashes by state and alcohol involvement, and shows that alcoholimpaired crashes represent about 31% of all fatal crashes that occurred in the southeastern states during the decade under scrutiny.
As was found for other previously examined factors, the proportion of alcohol-impaired fatal crashes is different across states, as reflected by the results of the homogeneity chisquare tests (see Table 54). In Georgia, about 24% of fatal crashes were alcoholimpaired, which is less than the average (31%). In South Carolina, about 36% of fatal crashes were alcohol-impaired (see Figure 84).
122

Figure 84: Fatal Crashes by State and Alcohol-involvement

Table 54: Fatal Crashes by State and Alcohol Involvement (1997 to 2006)

State

Non-impaired crashes

Alcohol-impaired crashes

Total

Alabama

6,532* (66.86)**

3,237 (33.14)

9,769 (100)

Georgia

10,817 (75.06)

3,594 (24.94)

14,411 (100)

Mississippi

5,327 (66.94)

2,631 (33.06)

7,958 (100)

South Carolina

5,934 (63.92)

3,349 (36.08)

9,283 (100)

Total

28,610 (69.07)

12,811 (30.93)

41,421 (100)

Pearson chi2(3) = 396.36; p-value < 0.001; * Fatal crash frequencies; ** Within-row relative frequencies (%)

Table 55 shows the change in fatal crashes between 1997 and 2006 by alcoholinvolvement and state. The homogeneity test results show that the proportions of the alcohol-impaired fatal crashes did not change from 1997 to 2006 at =0.05 in Alabama, Georgia, and Mississippi (see Figure 85). However, the proportion of the alcoholimpaired fatal crashes significantly increased from 30% to 40% (p-value < 0.001) in South Carolina. It is possible that impaired driving was reported more accurately over

123

time in South Carolina, that impaired driving actually increased in this state, or nonimpaired driving related crashes were significantly reduced in South Carolina over this same period.

Table 55: Change in Fatal Crashes by Alcohol-involvement and State (1997 to 2006)

State Year Non-impaired crashes

Alcohol-impaired crashes

Total

1997

690* (65.71)**

360 (34.29)

1,050 (100)

Alabama 2006

737 (68.69)

336 (31.31)

1,073 (100)

Pearson chi2(1) =2.217; p-value=0.145; * Fatal crash frequencies; ** Within-row relative frequencies (%)

1997

1045 (74.38)

360 (25.62)

1,405 (100)

Georgia 2006

1144 (73.47)

413 (26.53)

1,557 (100)

Pearson chi2(1) =0.312; p-value=0.576

1997

480 (64.78)

261 (35.22)

741 (100)

Mississippi 2006

521 (64.16)

291 (35.84)

812 (100)

Pearson chi2(1) =0.064; p-value=0.800

1997 South Carolina 2006

554 (69.42)
582 (59.82)

244 (30.58)
391 (40.18)

798 (100) 973 (100)

Pearson chi2(1) =17.599; p-value<0.001

124

Figure 85: Change in the Proportion of Alcohol-impaired Fatal Crashes (1997 to 2006)
4.2.3.2 Restraint Usage In addition to alcohol-involvement we also examined the distribution of fatal crashes by safety restraint use. Table 56 shows the restraint usage description used in this study as provided by FARS (http://www-fars.nhtsa.dot.gov/Common/IdDefs.aspx). It should be noted again that restraint use for fatal crashes can often be hard to discern and so can be reported incorrectly or could even be unknown in a non-trivial number of cases. Also note that restraint uses concerns motorcycle riders as well as adults and children in motor vehicles. Finally, increased use of safety restraints often results in reduced crash severity--thus `converting' an otherwise fatal crash into an injury crash.
125

Restraint Usage Restraint Used
Restraint Not Used Restraint Use Unknown

Table 56: Restraint Usage Description
Description Shoulder Belt Lap Belt Lap and Shoulder Belt Child Safety Seat Motorcycle Helmet Restraint Used - Type Unknown Safety Belt Used Improperly Child Safety Seat Used Improperly Non Used - Vehicle Occupant; Not Applicable Bicycle Helmet Helmets Used Improperly Blank Unknown

Table 57 shows the cross-classification table for the drivers in fatal crashes by state and restraint usage. The statistics show that about 42% of drivers in fatal crashes that occurred in the southeastern states during the period 1997 to 2006 were not wearing safety restraints. In Mississippi, about 41% of drivers in the fatal crashes were wearing safety restraints, which is significantly lower than restraint usage rates in the remaining states. Again, the homogeneity test results suggest that the restraint usage rates are different across states at =0.05.

Table 58 shows the change in proportion of drivers in fatal crashes by state, year, and restraint use. Safety restraint usage rates between 1997 and 2006 changed significantly in all southeastern states (=0.05). The proportions of unrestrained drivers in fatal crashes in the four states significantly decreased (see Figure 87). The magnitude of the reduction in the proportion of unrestrained drivers in fatal crashes in South Carolina between 1997 and 2006 is smallest among states (i.e. from 40% in 1997 to 38% in 2006). This small reduction may help explain the relatively small drop in the fatality rates of South Carolina between 1997 and 2006 discussed in Section 4.1.3.

126

Table 57: Drivers in Fatal Crashes by State and Restraint Usage (1997 to 2006)

State

Restraint Used

Restraint Not Used

Restraint Use Unknown

Alabama

6,728* (53.69)**

5,351 (42.7)

453 (3.61)

Georgia

10,439 (55.03)

6,062 (31.96)

2,467 (13.01)

Mississippi

4,118 (40.68)

5,866 (57.94)

140 (1.38)

South Carolina

6,193 (54.2)

4,864 (42.57)

370 (3.24)

Total

27,478 (51.8)

22,143 (41.74)

3,430 (6.47)

Pearson chi2(6) = 3378.73; p-value < 0.001; * Fatal crash frequencies; ** Within-row relative frequencies (%)

Figure 86: Drivers in Fatal Crashes by State and Restraint Usage (1997 to 2006) 127

Table 58: Change in Proportion of Drivers in Fatal Crashes (1997 and 2006)

State Year

Restraint Used

Restraint Not Used

Restraint Use Unknown

1997

652.00 (45.72)

700.00 (49.09)

74.00 (5.19)

Alabama 2006

747.00 (56.94)

525.00 (40.02)

40.00 (3.05)

Pearson chi2(2) =36.909; p-value<0.001; * Fatal crash frequencies; ** Within-row relative frequencies (%)

1997

828.00 (44.81)

730.00 (39.5)

290.00 (15.69)

Georgia 2006

1135.00 (57.32)

616.00 (31.11)

229.00 (11.57)

Pearson chi2(2) =60.358; p-value<0.001

1997

366.00 (37.46)

596.00 (61)

15.00 (1.54)

Mississippi 2006

453.00 (45.3)

546.00 (54.6)

1.00 (0.1)

Pearson chi2(2) =23.416; p-value<0.001

1997 South Carolina 2006

592.00 (58.85) 643.00 (55.38)

406.00 (40.36) 438.00 (37.73)

8.00 (0.8) 80.00 (6.89)

Pearson chi2(2) =51.405; p-value<0.001

The statistics show that the proportion of the restrained drivers in South Carolina was reduced from 59% to 55%. However, the proportion of unknown restraint status increased significantly (6%), which alone could explain the difference.

128

Figure 87: Change in the Portion of Drivers in Fatal Crashes by Restraint Usage and Year 129

4.3 Discussion of Results Overall, this chapter examines fatal crash trends in four Southeastern US states during the period from 1997 to 2006 (Alabama, Georgia, Mississippi, and South Carolina). The intent of the analysis was to illuminate possible trends in fatal crashes that have occurred during a decade of analysis, pointing hopefully to possible areas for further exploration and in-field examinations. The following are highlights of the analyses conducted in this chapter:
The total number of fatal crashes increased in all four states, although there were significant fluctuations in fatal crashes from year to year. Georgia had the highest frequencies (1405 and 1557 in 1997 to 2006 respectively) while Mississippi had the lowest frequencies (741 and 812 in 1997 to 2006 respectively).
Vehicle miles traveled also increased in each of the four states, with Georgia having both the highest VMT (114 billion in 2006) and the largest increase in VMT (20 billion) from 1997 to 2006.
Fatal crash rates--fatal crashes per million VMT--reduced in all states, although there was considerable fluctuation from year to year. Mississippi had the largest reduction in crash rate (24%) over the observation period, while South Carolina had the smallest reduction (4.8%). Overall crash rates in the four states in 2006 were 1.99 (Alabama), 1.49 (Georgia), 2.2 (Mississippi), and 2.08 (South Carolina).
Temporal factors including day of the week and time of day are associated with fatal crashes. About 50% of all fatal crashes occurred on weekends, about 17% of fatal crashes occurred between 3:00 and 5:59 p.m., and about another 20% occur between midnight and 5:59 a.m.
Fatal crashes occur on different roadway functional classifications in different proportions across states. Most fatal crashes in South Carolina occurred on collectors (44%), while most fatal crashes in Alabama occurred on arterials (36%). Mississippi had an unusually high proportion of fatal crashes on local roads and streets (34%) compared to about 11% in Alabama, 10% in Georgia, and 7% in South Carolina.
130

There was a significant change in where crashes occurred across states and across the period spanning 1997 to 2006. Significant changes of note include a reduction in fatal crashes occurring on arterials in Alabama (45% to 30% from 1997 to 2006), a significant increase on collectors in Mississippi (30% to 50% from 1997 to 2006) and a significant drop on local roads and streets in Mississippi (35% to 16% from 1997 to 2006). In Georgia and South Carolina there were only minor changes in the proportion of crashes by functional class of road.
Most fatal crashes occur at locations without traffic controls, ranging from 78% in Alabama to 87% in Mississippi. About 10% of fatal crashes occur at stop signs in all four states, while about 5% occur at signalized intersections. There were only slight changes in these proportions in the four states during the period spanning 1997 to 2006.
Alcohol impairment was associated with 24% of fatal crashes in Georgia to 36% in South Carolina. Alcohol impairment was fairly consistent for all states except South Carolina, which saw an increase from 30% to 40% of fatal crashes involving impaired drivers from 1997 to 2006.
Restraint use in fatal crashes varied considerably across the four states. Mississippi had the lowest rate of restraint use (41%), while Georgia had the highest (55%). Georgia also had a significantly higher proportion of unknown restraint use status (13%) compared to Alabama (4%) which had the next highest.
Restraint use trends seem to exist in all four states during the period spanning 1997 to 2006. Restraint use increased in Alabama (11%), Georgia (12%), and Mississippi (8%), while it decreased in South Carolina (4%) during the ten year span of analysis.
Examination of these statistics is insightful and points to opportunities for more in-depth analysis and field visits. For comparison purposes, the research team ranked the performance of the four states in 2006 relative to one another on a number of criteria, where 1 = BEST and 4 = WORST on several criteria previous discussed. Table 59 shows that Georgia has the lowest fatal crash rate and is also lowest on the proportion of weekend crashes, alcohol impaired crashes, and unrestrained drivers in crashes. Conversely, South Carolina is ranked 4th on crash rate and is also 4th on 2 other measures. While this table of rankings is not meant to establish the importance of
131

possible contributing factors on fatal crash rates, a pattern clearly exists and is deserving of further exploration.

Table 59: Relative Rank of Safety Performance in 2006 of four Southeastern States

State
Alabama Georgia Mississippi South Carolina

Crash Rate
2 1 3 4

Weekend Crashes
3 1 2 4

Alcohol Impairment
2 1 3 4

Unrestrained Drivers 3 1 4 2

132

5.0 Modeling Methodology and Strategy
5.1 Regression Model for Crash Type Prediction
5.1.1 Safety Predictive Models
Efficient and effective safety predictive models can vary based on specific objectives such as what to predict, at which level to predict, and which method to use. The vast majority of safety prediction models attempt to predict crash frequency (number of crashes that occurred during a period of time) or crash rate (crash frequency over the traffic exposure). These safety measures are treated as continuous variables. For example, the width of a lane could be any decimal value between 8 feet and 16 feet instead of sorted into discrete variable categories such as 11 feet, 11.5 feet, or 12 feet, for example. Common models often used for these assessments include Poisson regression models, negative binomial regression models, or variations of these models. Systemic-level safety measures, such as number of crashes that occurred over a time period for specific road segments, require analysts to aggregate (or sort) crash counts into categories and extract road geometric and roadside information for each individual road segment. In this case, a variable only represents an average condition of the corresponding road segment rather than reflecting a unique feature of a crash site.
Safety performance prediction at an aggregate level is important for roadway network screening and facility evaluation as it can help identify problematic areas. However, these systemic models do not permit evaluation at the individual crash level. In some cases, road or crash characteristics may have unique associations with safety measures at a disaggregate level that can be different than at the category or aggregate level. This phenomenon is known as ecological fallacy, an error that occurs when falsely assuming individuals in a group have the average attributes of the group as a whole. In other cases, some crash level variables are inappropriate for aggregation.
Among a range of safety measures, the most frequently evaluated measures of performance are crash frequency and crash rate. Meanwhile, as discussed in Chapter 2, a considerable number of researchers have also developed models to predict crash injury severity levels and their
133

associated crash cost. Crash type, on the other hand, has only been minimally investigated. Most researchers have historically addressed crash types by predicting crash frequency for a specific type of crash alone. This method can increase homogeneity of crash data since crash records would include only one specific type of crash, such as head-on crashes on two-lane rural highways. A homogenous dataset will be more likely to improve how well the resulting model fits, but this kind of crash type analysis may exclude potential connections between different types of crashes. Researchers and safety engineers can gain insight into how to reduce the frequency of crash types by considering potential contributing factors and their association with various crash types.
For this effort, the research team has elected to focus on crash types for fatal crash analysis at two-lane rural highways. It is expected, for example, that the distribution of crash types is different for crashes with all injury severities versus fatal crashes alone. Some types of crashes may be more likely to result in one or more fatalities while others would be more directly associated with less severe injuries. Therefore, the evaluation of fatal crash types may help reveal crash type associations that will be different from what can be determined when studying all crashes. Secondly, certain crash types tend to be over-represented for a specific type of highway facility. For example, a single-vehicle run-off-road crash type makes up about 60% of overall fatal crashes on two-lane rural highways; while the single-vehicle crash is less common and less severe at urban locations.
5.1.2 Crash Type Prediction Model Application
Fatal crash type prediction models can serve as an analytical assessment tool for projects from the Highway Safety Improvement Program (HSIP) as well as other related rural roadway safety improvements. In the process of identifying locations with the greatest safety needs, most current methods do not directly consider specific crash types. Crash type prediction models provide an approach to quantify safety performance measures by taking into account roadway design characteristics, road environmental features, as well as traffic conditions. A random sample of fatal crash records is available for this analysis. These records include detailed crashspecific roadway geometric characteristics that were collected through site investigation or video log inspection during a previous GDOT study. This comprehensive data set will allow the
134

research team to better understand the relationships between roadway design features, fatal crash types, and expected associations with prospective crash countermeasures.
Since the models described in this report were developed based on fatal crashes, the research team recommends using the models to assess expected reductions for fatality and severe injury crashes. It is generally recognized that fatal crashes and crashes with severe injuries may share common traits and are similar compared to crashes with minor injuries or with Property Damage Only (PDO). Most crash countermeasures are specific to a crash type; therefore, knowing how the road geometric design features and roadside environment conditions relate to fatal crash types will help safety engineers select more effective remedies for these hazardous crash conditions.
For this study, the research team's goal was to develop models to predict fatal crash type outcomes. Since crash type is a categorical variable, the research team performed categorical data analysis with a logistic regression model to predict categorical response variables with both continuous and categorical predictors. The next two sections introduce the definition of each crash type and the modeling approach used for this study.
5.1.3 Crash Types This study defines crash types based on the definition of "Manner of Collision" in the "First Harmful Events" category used by FARS (U.S. DOT and NHTSA, 2007). Figure 88 presents the fatal crash type classification structure. FARS describes the first harmful event as either the first property damage or injury-producing event of a crash occurrence. The single-vehicle run-offroad crash identifier is applied when the first harmful event is a non-collision (for example: driving off a cliff, rollover), a collision with an object that is not fixed (for example: pedestrians or animals), or a collision with a fixed-object (for example: trees, utility poles). Since there are only a few crashes striking objects that are not fixed, a simplified representative description for this study is a single-vehicle crash where the vehicle exited the roadway and either struck a fixed object or overturned.
135

For crashes that involved more than one vehicle, the two major fatal crash types observed for this data set were head-on collisions and angle crashes. A head-on collision is defined as two vehicles colliding with their front ends while traveling in opposite directions. An angle crash is categorized as the front end of one vehicle making contact with the broad side of another one.
Fatal Crash

Single-Vehicle Fatal Crash

Multiple-Vehicle Fatal Crash

Fixed-object Fatal Crash

Overturn Fatal Crash

Head-on Fatal Crash

Non-Head-on Fatal Crash

Figure 88: Fatal Crash Type Classification

5.2 Binary Logit Models
As previously discussed, one of the objectives of this study is to develop crash type models based on the premise that roadway design characteristics, roadside environment features, and traffic conditions each directly influence crash type occurrence. The research team applied a statistical model called a binary logit model to help identify influential factors that can differentiate two crash types at a time. These models reveal the association between crash type outcomes and contributing variables. Initial efforts resulted in the creation of three binary logit models based on the crash type classifications indicated with dashed lines in Figure 88. The research team examined several sets of potential factors that could significantly differentiate two types of crashes. These crash type pairs included:
Single-vehicle fatal crash vs. Multiple-vehicle fatal crash Fixed object fatal crash vs. Overturn fatal crash Head-on fatal crash vs. Other fatal crash (not head-on)
136

The use of the binary logit model to differentiate a fixed-object fatal crash from a rollover fatal crash proved unsuccessful. This outcome may be because these two types of single-vehicle crashes have similar potential causes initially as the vehicle will generally depart the travel lane prior to the crash. The likelihood of a vehicle rollover following lane departure may be due to steep side slope, narrow clear zone, pavement edge drop-offs, or similar conditions. Most of these variables, however, are not available individually in the crash database and are generally represented by a single subjective variable known as the roadside hazard rating. This binary logit model approach, therefore, may be suitable to differentiate single-vehicle fatal crashes from multiple-vehicle fatal crashes, but does not appear to have the power to significantly determine the likelihood of a fixed-object versus an overturn fatal crash. As a result, this research presents two crash type prediction models as presented in the following section:
Single-vehicle fatal crash vs. Multiple-vehicle fatal crash o Based on fatal crash history for rural two-lane highways, predict the probability of a single-vehicle fatal crash
Head-on fatal crash vs. Other fatal crash (not a head-on) o Based on multiple-vehicle fatal crash history for rural two-lane highways, predict the probability of a head-on fatal crash
5.2.1 Single-vehicle vs. Multiple-vehicle Crash
To compare conditions using a binary logit model, the first step is to assign a value of either zero or on for the crash type of interest. For this effort, the research team assigned the following values for variable Y for each fatal crash:
Y= 1, if the crash type is single-vehicle run-off-road fatal crash; Y= 0, otherwise.
The binary logit model then estimates the probability while Y has the value of 1 with independent variables as X1 ,....., Xk , that represent features associated with the crash conditions. The logistic function form estimates what the probability would be of observing a single-vehicle run-off-road crash in the event of a fatal crash, as shown below:
137

i=k

exp(0 + i Xi )

Pr(Y = 1) = Pr(Single - veh - runoff ) =

i =1 i=k

(4)

1+ exp(0 + i Xi )

i =1

Given: Pr(Single-veh-runoff ): the probability of observing a single-vehicle run-off-road fatal crash occurrence in the event a fatal crash occurred, this will be a value between 0 and 1; 0: estimated intercept; i: estimated coefficient for the corresponding independent variable Xi; Xi: the ith independent variable.

This model can be used to predict the likelihood that if a fatal crash should occur that it will be a single-vehicle run-off-road fatal crash. This is defined by a set of independent variables that have significant impacts on the specific crash type. Initially, the research team estimated this type of model based on the four-state (AL, GA, MS, and SC) combined fatal crash database. This database includes approximately 550 randomly selected fatal crashes that contain data for crash site conditions, environmental features, and general crash information. For variables related to drivers and vehicles, the analysis focused on the at-fault drivers and their vehicles. This assessment incorporated the driver's gender, age, safety belt usage, vehicle's type, model, year, etc. This study defines at-fault drivers as those drivers identified as the responsible parties whose actions directly contributed to crashes. As a result, each crash includes driver and vehicle information for only one driver. At-fault driver information could not be obtained for Mississippi and South Carolina fatal crash samples. Therefore, the combined-state model could not examine effects from these variables.

This analysis also included the development of crash type models for each individual state crash database. Where available, this individual state analysis includes the at-fault driver and vehicle information. The state specific models offer the opportunity to investigate differences as well as similarities for features that may significantly influence fatal crashes in the various states.

138

5.2.2 Head-on vs. Other Multiple-Vehicle Fatal Crashes
As previously indicated, the two most common multiple vehicle fatal crashes were head-on and angle crashes. Other observed crash types such as rear-end and sideswipe crashes account for approximately six percent of the total number of fatal crashes. This study focuses on reducing fatality collisions at rural road segments rather than the less common rural intersection crashes. Therefore, the second crash type model will focus on predicting the probability that when a multiple vehicle fatal crash occurs, that it is a head-on collision. By evaluating multiple vehicle crashes separate from single-vehicle crashes, the crash data has more homogenous characteristics resulting in better fitting models. The probability that if a multiple vehicle fatal crash occurs, that the crash type is a head-on is represented as:

i=k

exp(0 + i Xi )

Pr(Head - on) =

i =1 i=k

(5)

1+ exp(0 + i Xi )

i =1

Given: Pr(Head-on ): probability of observing a head-on crash occurrence in the event that a multiple-vehicle fatal crash occurred, this will be a value between 0 and 1; 0: estimated intercept; i: estimated coefficient for the corresponding independent variable Xi; Xi: the ith independent variable.

In a manner similar to the development of the single-vehicle fatal crash model, the research team estimated a head-on crash model for the combined-state database as well as for the individual states. Potential contributing factors include roadway geometric characteristics, roadside features, environment conditions, as well as at-fault drivers and their vehicle information. The final models and example use and interpretation of the models are presented later in this report.

139

5.3 Model Selection Common model verification procedures frequently used for ordinary regression models and some types of logistic regression models are not suitable for evaluating a binary logit model (Agresti, 2002; Ramsey and Schafer, 2002). The model checking and selection for binary logit model mainly relies on examining the significance of extra terms in the model including squared terms or possible interactions between variables. Influential variables that may significantly help differentiate crash types for the final model (with a p-value less than 0.20) were retained in the model during the examination of potential contributing factors. The ultimate goal for developing safety predictive models was to identify valuable information and quantify relationships between highway design characteristics and associated safety performance. While the statistical significance and model goodness-of-fit are very important considerations in this process, the research team members also made decisions based on their knowledge of how highway design characteristics relate to crash types. This known relationship between design characteristics and safety performance is why the ultimate final models may include higher-than-typical p-values of 0.20 or larger.
140

6.0 Crash Type Model Evaluation
6.1 Single-Vehicle Run-off-Road Fatal Crash
For this study the research team developed fatal crash type prediction models to estimate the probability of a single-vehicle run-off-road fatal crash occurrence in the event of a fatal crash based on the four-state combined fatal crash database (AL, GA, MS, SC), three-state combined database (AL, GA, SC), and state specific database. The combined-state model benefits from a larger sample size, with approximately 530 fatal crashes. This larger sample size enabled the research team to investigate more potential contributing factors at various significance levels. Meanwhile, the research team also developed state-specific models in order to investigate the opportunity of examining spatial transferability as well as to explore unique variables that may only have impacts on fatal crash outcomes in one or two states.
6.1.1 Combined-State Models 6.1.1.1 Four-State Model (AL, GA, MS, SC) Table 60 presents independent variables that this research determined to be significant for differentiating single-vehicle run-off-road fatal crashes from multiple-vehicle fatal crashes. These variables include: presence of road junction (intersection), lane width, paved shoulder width, graded shoulder width, horizontal curve direction, crest presence, road hazard rating, ADT, presence of commercial driveways, dark without supplemental lighting, and time of crash. In addition, Table 61 illustrates descriptive statistics for these independent variables. Table 62 summarizes the distribution of categorical variables. These variables are represented by values of either zero or one as defined in Table 60. Table 63 summarizes the actual model estimation as well as goodness-of-fit test results for the four-state combined (AL, GA, MS, SC) model.
141

Table 60: Variable Description (Combined-State Models, Single-Vehicle)

Types Location Indicator Road Junction type
Geometric design features
Roadside condition Traffic volume Land use type

Variables AL MS SC JUNCTION LW PSW GSW
LCURV
CREST RHR67 ADT
LU_C

Lighting condition DARKUNLIT

Circadian biological clock

HR_DEEPSLEEP

Descriptions
1 if in Alabama, 0 otherwise 1 if in Mississippi, 0 otherwise 1 if in South Carolina, 0 otherwise 1 if a road junction, 0 if road segment Lane width (ft) Paved shoulder width (ft) Graded shoulder width (ft) 1 if curve to the left, 0 otherwise (curve to the right or straight alignment) 1 if vertical crest curve, 0 otherwise 1 if road hazard rating is 6 or 7, 0 otherwise Average daily traffic (103 veh/day) 1 if commercial driveways in the proximity of the crash location, 0 otherwise 1 if dark with no supplemental street lights, 0 otherwise 1 if crash time between 1a.m. and 3a.m., 0 otherwise

Table 61: Continuous Variable Descriptive Statistics (Four-State Model, Single-Vehicle)

Variable LW (ft) PSW (ft) GSW (ft) ADT (vpd)

Mean 10.8 0.6 5.0 2,871

Std Dev 1.1 1.7 3.4 2,891

Minimum 7 0 0 75

Maximum 12 12 16
17,960

The value for lane width at crash locations ranges from 7 ft to 12 ft with an average of 10.8 ft (approximately 11 ft). Paved shoulder widths ranged from 0 ft to 12 ft, but the average paved shoulder width was near the minimum value at 0.6 ft with a 1.7 ft standard deviation. This low value for the paved shoulder width indicates that most fatal crash locations did not have paved shoulders. The graded shoulder width was an average of 5 ft wide with a standard deviation of 3.4 ft, and a range from 0 ft to 16 ft. Graded shoulder widths varied dramatically for the four states. The average ADT value was 2,871 vehicles per day, with a range from 75 up to 17,960

142

vehicles per day). While the distribution of the ADT was skewed to the right or the higher traffic exposure end, most of crashes occurred at low to moderate traffic locations.

Table 62: Distribution of Categorical Variables (Four-State Model, Single-Vehicle)

Variable JUNCTION LCURV CREST RHR67 LU_C DARKUNLIT HR_DEEPSLEEP

Status

Percent (%)

0 (Segment)

77

1 (Intersection)

23

0 (Curve to Right or Straight)

75

1 (Curve to Left)

25

0 (Not a Crest Vertical Curve)

88

1 (Crest Vertical Curve)

12

0 (Roadside Hazard Rating < 6)

88

1 (Roadside Hazard Rating of 6 or 7)

12

0 (No Commercial Driveways)

94

1 (Near Commercial Driveways)

6

0 (Daylight, Dark with Lights, Dusk, or Dawn)

55

1 (Dark without Supplemental Lights)

45

0 (Not between 1 a.m. and 3 a.m.)

95

1 (From 1 a.m. until 3 a.m.)

5

Among the 527 fatal crashes available for the four-state combined model, 309 crashes were single-vehicle run-off-road fatal crashes. The Hosmer and Lemeshow Goodness-of-Fit test that measures how well the model fits the observed fatal crash data reports an acceptable result (pvalue = 0.2873>0.05) as shown in Table 63. Based on the model estimation results shown in Equation (6) and Table 63, the variables that can significantly differentiate single-vehicle runoff-road fatal crash from multiple-vehicle fatal crash include junction type, lane width, paved shoulder width, graded shoulder width, horizontal curve direction, crest vertical curve, roadside hazard rating, ADT, driveway land use type, lighting condition, and crash time. In the event of a fatal crash, the probability of a single-vehicle run-off-road fatal crash occurrence increases at road segments with horizontal curvature to the left compared to locations with a straight alignment or a curve to the right (with or without the presence of a crest vertical curve). Meanwhile, variables that also significantly increase the likelihood of single-vehicle run-off-road fatal crash occurrence include a roadside hazard rating of 6 or 7, dark driving conditions without

143

supplemental lighting, as well as a time of crash that occurs between 1a.m. to 3 a.m. In addition, if a fatal crash occurred, the likelihood of the fatal crash involving only a single-vehicle tends to have a lower chance to occur with increased lane width a values and ADT values. The presence of road intersections (junctions), crest vertical curves, and commercial driveways in the vicinity of the crash locations are more likely associated with the less common multiple-vehicle crash at these rural locations.
This model also identifies a significant interaction effect between paved and graded shoulder width. This indicates that the influence from the paved shoulder width on the probability of a single-vehicle run-off-road fatal crash occurrence in the event of a fatal crash also depends on the graded shoulder width. Similarly, the horizontal curve direction and the presences of a crest vertical curve each present significant main effects as well as an interaction effect. Section 6.1.4 includes detailed discussions regarding these interaction effects.
The three location indicator variables, AL, MS, and SC, will have the value of 1 only if the crash occurred in the corresponding state, otherwise they will have a value of 0. Therefore, a crash location will be Georgia when all three indicator variables are equal to zero. As shown in Table 63, the estimation results are not significant for AL and SC, but the MS p-value is significant. This observation indicates that fatal crash types in Alabama and South Carolina are more likely to follow a similar pattern to those in Georgia as compared to the disparate findings for similar crash types in Mississippi. Since one objective of this study is to identify rural two-lane highway fatal crash models that can help analysts better understand crash trends in Georgia, the research team also investigated the combined-state model based on the fatal crash database from AL, GA, and SC separated (the three similar states). Model estimation results and discussion are presented in the following section.
144

Table 63: Model Estimation (Four-State Model, Single-Vehicle)

Variable

Estimate

P- value

Intercept

5.906

<.0001

AL

-0.1984

0.532

MS

-1.3453

0.0004

SC

-0.0836

0.8091

JUNCTION

-0.9922

0.0003

LW

-0.463

0.0006

PSW

-0.1087

0.2325

GSW PSW*GSW 1

-0.0463 -0.0622

0.304 0.0995

LCURV

0.7255

0.0132

CREST LCURV*CREST 1

-1.5389 2.2686

0.0002 0.0142

RHR67

1.3314

0.0022

ADT

-0.1078

0.025

LU_C

-1.4298

0.02

DARKUNLIT

1.3135

<.0001

HR_DEEPSLEEP

1.9744

0.0081

Observations (Single-Vehicle/Others) AIC

527 (309/218)
530.16

SC -2 Log L

602.703 496.16

R-Square

0.3396

Hosmer and Lemeshow Goodness-of-Fit Test

Chi-Square

DF

Pr > ChiSq

9.692

8

0.2873

1 Note: LCURV*CREST and PSW*GSW indicate these variable pairs interact.

The resulting single-vehicle run-off-road fatal crash prediction model as depicted in Table 63 is presented as follows:

145

Let :

4-state = 5.906 - 0.1984AL -1.3453MS - 0.0836SC - 0.9922JUNCTION - 0.463LW - 0.1087PSW - 0.0463GSW -0.0622(PSW *GSW) + 0.7225LCURV -1.5389CREST + 2.2686(LCURV *CREST) +1.3314RHR67 - 0.1078ADT -1.4298LU _C +1.3135DARKUNLIT +1.9744HR _ DEEPSLEEP

The probability of a single-vehicle run-off-road fatal crash can then be predicted for a given set

of road and environment conditions as:

Pr(Single - veh - runoff

)4-state

=

exp(4-state ) 1 + exp(4-state )

(6)

6.1.1.2 Three-State Model (AL, GA, SC)
As determined in the previous section, rural two-lane highway fatal crashes for three of the four study states exhibited appeared to be influenced by common characteristics, while Mississippi were atypical. For this reason, a three-state model can be developed to further assess the crashes and their influences in the three similar states. The three-state combined model includes the same set of independent variables as identified for the four-state model and as indicated in Table 60. Table 64 and Table 65 summarize descriptive statistics for these continuous and categorical variables, respectively. Table 66 presents the resulting model estimation and its goodness-of-fit test results.

Table 64: Continuous Variable Descriptive Statistics (Three-State Model, Single-Vehicle)

Variable LW (ft) PSW (ft) GSW (ft) ADT (veh/day)

Mean 10.8 0.6 5.2 2,896

Std Dev 1.1 1.6 3.5 2,941

Minimum 8 0 0 75

Maximum 12 12 16
17,960

146

Table 65: Distribution of Categorical Variables (Three-State Model, Single-Vehicle)

Variable JUNCTION LCURV CREST RHR67 LU_C DARKUNLIT HR_DEEPSLEEP

Status
0 (Segment) 1 (Intersection) 0 (Curve to Right or Straight) 1 (Curve to Left) 0 (Not a Crest Vertical Curve) 1 (Crest Vertical Curve) 0 (Roadside Hazard Rating < 6) 1 (Roadside Hazard Rating of 6 or 7) 0 (No Commercial Driveways) 1 (Near Commercial Driveways) 0 (Daylight, Dark with Lights, Dusk, or Dawn) 1 (Dark without Supplemental Lights) 0 (Not between 1 a.m. and 3 a.m.) 1 (From 1 a.m. until 3 a.m.)

Percent (%)
75 25 75 25 89 11 92 8 94 6 55 45 96 4

Among the 428 fatal crashes available for the three-state combined model, 259 crashes were single-vehicle run-off-road fatal crashes. As presented in Table 66, the Hosmer and Lemeshow Goodness-of-Fit test showed an acceptable indication that the model fits the data well (p-value = 0.4849 >0.05). As shown in Equation (7), variables that can significantly differentiate singlevehicle fatal crash from multiple-vehicle fatal crash include the presence of a road intersection (junction), lane width, paved shoulder width, graded shoulder width, horizontal curve direction, presence of a crest vertical curve, a roadside hazard rating 6 or 7, ADT, driveway land use type, lighting condition, and time of crash.

147

Table 66: Model Estimation (Three-State Model, Single-Vehicle)

Variable

Estimate

P- value

Intercept

6.6717

<.0001

AL

-0.1855

0.5614

SC

-0.1167

0.7404

JUNCTION

-0.8078

0.0051

LW

-0.5407

0.0003

PSW

-0.0542

0.5873

GSW PSW*GSW 1

-0.0475 -0.0676

0.3202 0.0929

LCURV

0.788

0.0156

CREST LCURV*CREST 1

-1.7264 2.5199

0.0002 0.0223

RHR67

1.1581

0.0716

ADT

-0.0965

0.0558

LU_C

-1.3722

0.0313

DARKUNLIT

1.3101

<.0001

HR_DEEPSLEEP

1.8318

0.0926

Observations (Single-Vehicle/Others) AIC

428 (259/169) 440.976

SC

505.922

-2 Log L

408.976

R-Square

0.3204

Hosmer and Lemeshow Goodness-of-Fit Test

Chi-Square

DF

Pr > ChiSq

7.4889

8

0.4849

1 Note: LCURV*CREST and PSW*GSW indicate these variable pairs interact.

Let:

3-state = 6.6717 - 0.1855AL - 0.1167SC - 0.8078JUNCTION - 0.5407LW - 0.0542PSW - 0.0475GSW -0.0676(PSW * GSW ) + 0.788LCURV -1.7264CREST + 2.5199(LCURV * CREST ) + 1.1581RHR67 - 0.0965ADT -1.3722LU _ C + 1.3101DARKUNLIT + 1.8318HR _ DEEPSLEEP

Then the probability of a single-vehicle run-off-road fatal crash can be predicted under a given set of road and environment conditions as:

148

Pr(Single

-

veh

-

runoff

)3- state

=

exp(3-state ) 1 + exp(3-state )

(7)

This model indicates, for example, that in the event a fatal crash occurs, the probability of a single-vehicle run-off-road fatal crash will increase at road segments with horizontal curves to the left. Variables that also significantly increase the likelihood of single-vehicle run-off-road fatal crashes include the road hazard rating of 6 or 7, dark conditions without supplemental lighting, and time of day between 1 a.m. to 3 a.m. Single-vehicle run-off-road fatal crashes tend to occur less frequently as both the lane width and traffic volume increase. Variables also not strongly associated with the single-vehicle rural fatal crash include the presence of an intersection, crest vertical curve locations, and proximity to commercial driveways.

The statistically significant interaction effect between the paved and graded shoulder width indicates that the influence of the paved shoulder width on the probability of single-vehicle fatal crashes is also dependent on the graded shoulder width. Similarly, the horizontal curve to the left and the presence of a crest vertical curve exhibit a similar interaction effect. For detailed variable analysis information, please refer to Section 6.1.4.

As observed previously in the four-state model, the state indicator variables AL and SC in the three-state model are again shown with insignificant estimation results. One of the major purposes of this research study is to develop analytical models that can be used as evaluation tools for safety investment by the Georgia Department of Transportation. Since the most common fatal crash observed on rural two-lane highways was the single-vehicle crash and the three-state model does a better job of predicting these crashes for Georgia-specific conditions, it would be a reasonable approach to apply the three-state combined model for these assessments. This would result in more reliable and stable crash estimates for Georgia.

6.1.2 Models by State
This section presents four model estimates based on the fatal crash sample data for the four individual states of Alabama, Georgia, Mississippi, and South Carolina. In addition to the modeling effort for a regional level as summarized by the four-state combined model (AL, GA,

149

MS, SC) and the three-state combined model (AL, GA, SC), this individual state assessment can be used as an indicator for identifying potential state-specific significant influential factors and their corresponding effects on the probability of single-vehicle run-off-road fatal crash occurrence. This state-level modeling effort can also help determine the suitability of model transferability for other state applications.
6.1.2.1 Alabama
The fatal crash database for the state of Alabama included 155 random fatal crashes for rural two-lane highways. Table 67 presents the independent variables determined to have significant influences on the fatal crash type outcomes. These variables include the presences of a road intersection (junction), lane width, horizontal curve direction, grade direction, roadside hazard rating, roadside lighting condition, and time of crash. The descriptive statistics for the continuous variable (lane width) and the categorical variables are presented in Table 68 and Table 69, respectively. For the Alabama fatal crash database, the lane widths ranged from 8 ft to 12 ft with an average lane width of 10.8 ft (approximately 11 ft). As summarized in Table 70, there were 99 observed single-vehicle run-off-road fatal crashes and 56 multiple-vehicle crashes, so approximately two-thirds of the total number of fatal crashes were single-vehicle. The Hosmer and Lemeshow Goodness-of-Fit test showed evidence that the model provides a good fit for the observed data (p-value=0.3439 >0.05).
As shown in Table 70 and Equation (8), there is a higher probability that if a fatal crash occurs it will be a single-vehicle crash at road segment locations with horizontal curves to the left. Other variables that also significantly increase the likelihood of single-vehicle fatal crash include road segments with downhill grades, roadside hazard ratings of 6 or 7, dark conditions without supplemental street lights, and time of day between midnight to 6 a.m. Single-vehicle fatal crashes are also less likely to occur at locations with increasing lane width. The Alabama-only model does not appear to have any significant variable interactions as observed for the multistate models.
150

Table 67: Variable Description (AL only Model, Single-Vehicle)

Types Road Junction type

Variables JUNCTION LW

Geometric design features

LCURV DOWN

Roadside condition RHR67

Lighting condition DARKUNLIT

Circadian biological clock

HR_EM

Descriptions
1 if a road junction, 0 if road segment
Lane width (ft) 1 if curve to the left, 0 otherwise (curve to the right or straight alignment) 1 if direction of slope is down (negative), 0 otherwise 1 if road hazard rating is 6 or 7, 0 otherwise 1 if dark with no supplemental street lights, 0 otherwise 1 if crash occurred between 12a.m. 6 a.m., 0 otherwise

Table 68: Continuous Variable Descriptive Statistics (AL only Model, Single-Vehicle)

Variable LW (ft)

Mean 10.8

Std Dev 1.1

Minimum 8

Maximum 12

Table 69: Distribution of Categorical Variables (AL only, Single-Vehicle)

Variable JUNCTION LCURV DOWN RHR67 DARKUNLIT HR_EM

Status
0 (Segment) 1 (Intersection) 0 (Curve to Right or Straight) 1 (Curve to Left) 0 (Vertical grade positive or level) 1 (Vertical grade negative) 0 (Roadside Hazard Rating < 6) 1 (Roadside Hazard Rating of 6 or 7) 0 (Daylight, Dark with Lights, Dusk, or Dawn) 1 (Dark without Supplemental Lights) 0 (Not between 6 a.m. and 12 midnight) 1 (From 12 midnight until 6 a.m.)

Percent (%)
83 17 73 27 56 44 88 12 54 46 84 16

151

Table 70: Model Estimation (AL only Model, Single-Vehicle)

Variable

Estimate

P- value

Intercept

4.7369

JUNCTION

-1.2158

LW

-0.5111

LCURV

1.5906

DOWN

0.8342

RHR67

1.8195

DARKUNLIT

1.3682

HR_EM

2.8888

Observations (Single-Vehicle/Others) AIC SC -2 Log L R-Square

155 (99/56) 149.483 173.83 133.483 0.3605

Hosmer and Lemeshow Goodness-of-Fit

Chi-Square

DF

8.9815

8

0.0446 0.039 0.0158 0.0048 0.0603 0.1117 0.0032 0.0089
Pr > ChiSq 0.3439

Let:
AL = 4.7369 -1.2158JUNCTION - 0.5111LW +1.5906LCURV + 0.8342DOWN +1.8195RHR67 +1.3682DARKUNLIT + 2.888HR _ EM

The probability of single-vehicle run-off-road fatal crash can then be predicted under a given set

of conditions as:

Pr(Single

-

veh

-

runoff

) AL

=

exp(AL ) 1 + exp(AL )

(8)

6.1.2.2 Georgia As presented in Table 71, independent variables which have significant impacts on the fatal crash type outcomes in the GA only model include intersection (junction) type, lane width, paved shoulder width, horizontal curve direction, horizontal alignment type, roadside lighting
152

condition, and safety restraint system usage for at-fault drivers. Table 72 and Table 73 summarize the descriptive statistics for those contributing factors. In the Georgia fatal crash database, lane widths ranged from 8 ft to 12 ft with average lane width at 10.7 ft (approximately 11 ft). Average width of paved shoulders was 0.6 ft but for a total range from 0 ft to 6 ft width.
As summarized in Table 74, 85 out of 146 fatal crashes were single-vehicle run-off-road fatal crashes. The recommended goodness-of-fit test reports a p-value as 0.3439, which indicates that the estimated model fits the observed data well. For the state of Georgia, if a fatal crash occurred, the likelihood of a single-vehicle run-off-road fatal crash would increase at a location with a horizontal curve to the left. Other significant variables for the Georgia single-vehicle fatal crash include roadside hazard ratings of 6 or 7, dark without supplemental lighting, and at-fault drivers not utilizing safety restraints. Single-vehicle fatal crashes are less likely to occur at locations with increasing of lane and paved shoulder widths. The maximum lane width for the Georgia highway crash sites was 12 ft and the maximum paved shoulder width was 6 ft so this observation should not be extrapolated outside these upper boundaries.

Table 71: Variable Description (GA only Model, Single-Vehicle)

Types Road Junction type
Geometric design features
Lighting condition Safety Protection

Variables JUNCTION LW PSW
LCURV
STRAIGHT
DARKUNLIT
RESTRAINT

Descriptions
1 if a road junction, 0 if road segment Lane width (ft) Paved shoulder width (ft) 1 if curve to the left, 0 otherwise (curve to the right or straight alignment) 1 if tangent horizontal alignment, 0 otherwise 1 if dark with no supplemental street lights, 0 otherwise 1 if driver wore safety restraint, 0 otherwise

Table 72: Continuous Variable Descriptive Statistics (GA only Model, Single-Vehicle)

Variable LW (ft) PSW (ft)

Mean 10.7 0.6

Std Dev 1.1 1.2

Minimum 8 0

Maximum 12 6

153

Table 73: Distribution of Categorical Variables (GA only Model, Single-Vehicle)

Variable LCURV JUNCTION STRAIGHT DARKUNLIT RESTRAINT

Status
0 (Curve to Right or Straight) 1 (Curve to Left) 0 (Segment) 1 (Intersection) 0 (Curved location) 1 (Tangent) 0 (Daylight, Dark with Lights, Dusk, or Dawn) 1 (Dark without Supplemental Lights) 0 (Safety Restraint not Used) 1 (Safety Restraint Used)

Percent (%)
77 23 75 25 49 51 57 43 71 29

Table 74: Model Estimation (GA only Model, Single-Vehicle)

Variable

Estimate

P- value

Intercept

8.9011

0.0003

JUNCTION LW PSW LCURV STRAIGHT DARKUNLIT RESTRAINT

-2.1473 -0.835 -0.3506 1.7437 1.5662 1.1195 -1.1604

<.0001 0.0003 0.0647 0.0112 0.0046 0.0124 0.0151

Observations (Single-Vehicle/Others) AIC SC -2 Log L R-Square

146 (85/61) 151.893 175.762 135.893 0.3484

Hosmer and Lemeshow Goodness-of-Fit Test

Chi-Square 4.983

DF

Pr > ChiSq

7

0.662

Let:
GA = 8.9011- 2.1473JUNCTION - 0.835LW - 0.3506PSW +1.7437LCURV +1.5662STRAIGHT +1.1195DARKUNLIT -1.1604RESTRAINT

154

The probability of a single-vehicle run-off-road fatal crash for Georgia can then be predicted as

follows:

Pr(Single

-

veh

-

runoff

)GA

=

exp(GA ) 1 + exp(GA )

(9)

6.1.2.3 Mississippi

As shown in Table 75, independent variables which have significant impacts on the fatal crash type outcomes in Mississippi include lane width, paved shoulder width, horizontal curve direction, roadside hazard rating 6 or 7, dark without supplemental lighting, and time of crash. Table 76 and Table 77 summarize the descriptive statistics for continuous and categorical variables, respectively. In the fatal crash database of Mississippi, lane widths ranged from 7 ft to 12 ft with an average lane width at 10.9 ft (approximately 11 ft). The average width of the paved shoulders was 0.8 ft with an overall range from 0 ft to 10 ft. As summarized in Table 78, 50 of the 99 fatal crashes were single-vehicle run-off-road fatal crashes. This model also provides a reasonable fit to the observed data based on the goodness-of-fit test (p-value=0.5212 >0.05).

Table 75: Variable Description (MS only Model, Single-Vehicle)

Types
Geometric design features

Variables LW PSW
LCURV

Roadside condition RHR67

Lighting condition DARKUNLIT

Circadian biological clock

HR_DEEPSLEEP

Descriptions
Lane width (ft)
Paved shoulder width (ft) 1 if curve to the left, 0 otherwise (curve to the right or straight alignment) 1 if roadside hazard rating is 6 or 7, 0 otherwise 1 if dark with no supplemental street lights, 0 otherwise 1 if crash time is between 1a.m. and 3a.m., 0 otherwise

Table 76: Continuous Variable Descriptive Statistics (MS only Model, Single-Vehicle)

Variable LW (ft) PSW (ft)

Mean 10.9 0.8

Std Dev 1.1 2.3

Minimum 7 0

Maximum 12 10

155

Table 77: Distribution of Categorical Variables (MS only Model, Single-Vehicle)

Variable LCURV RHR67 DARKUNLIT HR_DEEPSLEEP

Status
0 (Curve to Right or Straight) 1 (Curve to Left) 0 (Roadside Hazard Rating < 6) 1 (Roadside Hazard Rating of 6 or 7) 0 (Daylight, Dark with Lights, Dusk, or Dawn) 1 (Dark without Supplemental Lights) 0 (Not between 1 a.m. and 3 a.m.) 1 (From 1 a.m. until 3 a.m.)

Percent (%)
77 23 68 32 54 46 89 11

In the event of a fatal crash, the probability of single-vehicle run-off-road fatal crash increases at road segments with horizontal curves to the left. Other variables that significantly increase the likelihood of single-vehicle fatal crash include a roadside hazard rating 6 or 7, dark conditions without supplemental lighting, and a time of crash between 1 a.m. and 3 a.m. Single-vehicle fatal crashes also tend to occur less often at sites with increased lane and paved shoulder widths, though this interpretation should not be extrapolated to values greater than those depicted in Table 76. The Mississippi model did not include any significant interaction effects between paved and graded shoulder width or between the horizontal and vertical alignments. In fact, the sample data did not support the graded shoulder width as a significant influential variable in the model. The smaller sample size for the Mississippi fatal crash database may have contributed to reduced predictive capabilities as compared to the larger sample sizes for the other individual states and combined-state models.

156

Table 78: Model Estimation (MS only Model, Single-Vehicle)

Variable

Estimate

Intercept

3.2578

LW

-0.4282

PSW

-0.7045

LCURV

0.8562

RHR67

1.6633

DARKUNLIT

1.626

HR_DEEPSLEEP

2.0741

Observations (Single-Vehicle/Others)

99 (50/49)

AIC

105.39

SC

123.556

-2 Log L

91.39

R-Square

0.3706

Hosmer and Lemeshow Goodness-of-Fit

Chi-Square

DF

7.1436

8

P- value 0.2469 0.092 0.1156 0.1516 0.0043 0.0027 0.088
Pr > ChiSq 0.5212

Let:
MS = 3.2578 - 0.4282LW - 0.7045PSW + 0.8562LCURV + 1.6633RHR67 + 1.626DARKUNLIT +2.0741HR _ DEEPSLEEP

The probability of a single-vehicle run-off-road fatal crash in Mississippi can be predicted as

follows:

Pr(Single - veh

- runoff

)MS

=

exp(MS ) 1 + exp(MS )

(10)

6.1.2.4 South Carolina
Table 79 depicts the independent variables determined to have significant impacts on the fatal crash type outcomes in the South Carolina model. These critical variables include the lane width, graded shoulder width, horizontal curve direction, crest vertical curvature, proximity to

157

commercial driveways, road lighting condition, and time of crash. Table 80 and Table 81 summarize the descriptive statistics for continuous and categorical variables respectively.

Table 79: Variable Description (SC only Model, Single-Vehicle)

Types
Geometric design features
Land use type

Variables LW GSW LCURV CREST LU_C

Lighting condition DARKUNLIT

Circadian biological clock

HR_DEEPSLEEP

Descriptions
Lane width (ft)
Graded shoulder width (ft) 1 if curve to the left, 0 otherwise (curve to the right or straight alignment) 1 if crest, 0 otherwise 1 if commercial driveways in the proximity of the crash location, 0 otherwise 1 if dark with no supplemental street lights, 0 otherwise 1 if crash time is from 1a.m. till 3 a.m., 0 otherwise

Table 80: Continuous Variable Descriptive Statistics (SC only Model, Single-Vehicle)

Variable LW (ft) GSW (ft)

Mean 10.7 7.4

Std Dev 1.0 3.2

Minimum 8 0

Maximum 12 15

Table 81: Distribution of Categorical Variables (SC only Model, Single-Vehicle)

Variable LCURV CREST LU_C DARKUNLIT HR_DEEPSLEEP

Status
0 (Curve to Right or Straight) 1 (Curve to Left) 0 (Not a Crest Vertical Curve) 1 (Crest Vertical Curve) 0 (No Commercial Driveways) 1 (Near Commercial Driveways) 0 (Daylight, Dark with Lights, Dusk, or Dawn) 1 (Dark without Supplemental Lights) 0 (Not between 1 a.m. and 3 a.m.) 1 (From 1 a.m. until 3 a.m.)

Percent (%)
76 24 91 9 91 9 53 47 95 5

158

For the fatal crash database for South Carolina, lane widths ranged from 8 ft to 12 ft with average lane width of 10.7 ft (approximately 11 ft). The average width of graded shoulders was 7.4 ft with a range from 0 ft to 15 ft. Among the four state-specific models, the South Carolina model is the only one that identified the graded shoulder width as having a significant effect in terms of differentiating between single-vehicle and multiple-vehicle fatal crashes. As summarized in Table 82, of the 155 fatal crashes in the database there were 94 single-vehicle run-off-road fatal crashes. The recommended goodness-of-fit for binary logit models reports a pvalue as 0.1706, which suggests an acceptable model.

Table 82: Model Estimation: Single-Vehicle Fatal Crash Model (SC only Model, SingleVehicle)

Variable

Estimate

Intercept

12.1999

LW

-1.0576

GSW

-0.1247

LCURV

1.1555

CREST

-1.5341

LU_C

-2.5984

DARKUNLIT

1.3523

HR_DEEPSLEEP

2.0821

Observations (Single-Vehicle/Others) AIC
SC

155 (94/61) 159.735
184.082

-2 Log L

143.735

R-Square

0.3385

Hosmer and Lemeshow Goodness-of-Fit

Chi-Square

DF

11.5871

8

P- value <.0001 <.0001 0.0813 0.0293 0.0254 0.006 0.0017 0.1114
Pr > ChiSq 0.1706

Let: SC = 12.1999 -1.0576LW - 0.1247GSW +1.1555LCURV -1.5341CREST - 2.5984LU _ C +1.3523DARKUNLIT + 2.0821HR _ DEEPSLEEP

159

The probability of a single-vehicle run-off-road fatal crash can then be predicted as follows:

Pr(Single

-

veh

-

runoff

)SC

=

exp(SC ) 1 + exp(SC )

(11)

For the state of South Carolina, the probability that if a fatal crash occurs it will be a singlevehicle run-off-road fatal crash will increase at road segments with horizontal curves to the left. Other significant variables for the South Carolina single-vehicle run-off-road fatal crash model include dark without supplemental lighting and time of crash. Single-vehicle run-off-road fatal crashes tend to have a lower chance of occurring with increased lane and graded shoulder widths, but this finding is restricted to the ranges indicated in Table 80. In addition, fatal crashes at locations with crest vertical curves or in the proximity of commercial driveways are less likely to be associated with a single-vehicle fatal crash.

6.1.3 Summary of Single-Vehicle Fatal Crash Models
Table 83 and Table 84 summarize the model estimation results for the four individual-state models, the three-state combined model (AL, GA, SC), and the four-state combined model (AL, GA, MS, SC). The four individual-state models do not all contain the same set of independent variables. The two combined-state models include a collection of independent variables included in all four individual-state models. The effort of fitting four individual-state models with a same set of independent variables is not supported by the data. One of the requirements of testing model spatial transferability is to fit models with the same set of predictors. This condition can only be achieved if all four individual-state models include a limited collection of independent variables such as ADT only. These larger models would then have less accurate predictive power since critical contributing factors may be excluded.
The different model specifications for the individual-state models may indicate that a model developed for one state is unlikely to apply to another state; however, insufficient sample size may also explain poor fit or transferability. Both of the combined state models provided similar estimates for the categorical location indicator variables for Alabama and South Carolina. As previously discussed, this observation indicates that the fatal crash type outcome prediction
160

might be similar across at least three states: Alabama, South Carolina, and the base state Georgia.

Table 83: Model Comparison (Single-Vehicle)

Variables

AL only Model

GA only Model

AL

MS

SC

INTERSECTION -1.2158**

LW

-0.5111**

PSW

GSW

PSW*GSW

LCURV

1.5906**

STRAIGHT

VCREST

DOWN

0.8342*

LCURV*VCREST

RHR67

1.8195

ADTSCALE

LU_C

DARKUNLIT

1.3682**

HR_DEEPSLEEP

HR_EM

2.8888**

RESTRAINT

** Significant level < 0.05 * Significant level < 0.1

-2.1473** -0.835** -0.3506* 1.7437** 1.5662**
1.1195** -1.1604**

MS only Model
-0.4282* -0.7045 0.8562
1.6633** 1.626** 2.0741*

SC only Model

Three-State Model (AL, GA, SC)
-0.1855

-1.0576** -0.1247* 1.1555**

-0.1167 -0.8078** -0.5407** -0.0542 -0.0475 -0.0676* 0.788**

-1.5341** -1.7264**

-2.5984** 1.3523** 2.0821

2.5199** 1.1581* -0.0965* -1.3722** 1.3101** 1.8318*

Four-State Model
(AL, GA, MS, SC) -0.1984 -1.3453** -0.0836 -0.9922** -0.463** -0.1087 -0.0463 -0.0622* 0.7255**
-1.5389**
2.2686** 1.3314** -0.1078** -1.4298** 1.3135** 1.9744**

161

Table 84: Illustration of Effects (Single-Vehicle)

Variables

AL only Model

GA only Model

MS only Model

SC only Model

Three-State Model (AL, GA, SC)

Four-State Model
(AL, GA, MS, SC)

MS

- **

AL

-

-

SC

-

-

INTERSECTION

-**

- **

- **

- **

LW

-**

- **

- *

- **

- **

- **

PSW

- *

-

-

-

GSW

- *

-

-

PSW*GSW

- *

- *

LCURV

+**

+ **

+

+ **

+ **

+ **

STRAIGHT

+ **

VCREST

- **

- **

- **

DOWN

+*

LCURV*VCREST

+ **

+ **

RHR67

+

+ **

+ *

+ **

ADTSCALE

- *

- **

LU_C

- **

- **

- **

DARKUNLIT

+ **

+ **

+ **

+ **

+ **

+ **

HR_DEEPSLEEP

+ *

+

+ *

+ **

HR_EM

+ **

RESTRAINT

- **

+ : Increase the probability of single-vehicle fatal crash, when the continuous variable increases or the indicator has the value of 1 versus 0.
- : Decrease the probability of single-vehicle fatal crash, when the continuous variable increases or the indicator has the value of 1 versus 0.
** Significant level <0.05 * Significant level <0.1

The two combined-state models present very similar modeling results when the same set of independent variables is retained. As discussed previously, the research team proposed to use the three-state (AL, GA, and SC) combined model for safety evaluation of two-lane rural highways in Georgia.

162

As shown in Table 83 and Table 84, despite the differences among the individual-state models and the combined-state models, there are three independent variables (lane width, horizontal curve direction, and lighting conditions) that are significant predictors with similar effects for all six models. Roadway segments with narrower lanes have a greater likelihood of single-vehicle run-off-road fatal crashes than their wider lane counterparts. The location with a curve to the left tends to be more frequently associated with a single-vehicle run-off-road crash than are locations with either curves to the right or straight alignment. Similarly, the location that is dark without supplemental lighting is more likely to be a site for a single-vehicle run-off-road crash than locations with better lighting conditions or daytime conditions. These findings suggest that the lane width, curve direction, and lighting condition are strongly associated with the probability of a fatal crash type, with direct associations for single-vehicle fatal crashes, for the state level and the regional level.
6.1.4 Analysis of Variables for Single-Vehicle Run-off-Road Crashes
This section presents a sensitivity analysis for the probability of single-vehicle run-off-road fatal crash occurrence for a variety of previously identified contributing variables, including:
Lane width Paved shoulder width Graded shoulder width Average Daily Traffic Junction (intersection) versus segment Horizontal and vertical alignment Roadside hazard rating Land use type of driveways Lighting condition Time of day for crash occurrence
The research team performed an analysis based on the recommended three-state combined model (AL, GA, and SC), as shown in equation 5 in section 6.1.1.2. Meanwhile, the analysis also extended to the GA only model, as shown in equation 7 in section 6.1.2.2. In order to assess changes of predicted crash type outcome probabilities at different levels for an independent
163

variable, all other independent variables are held constant while the candidate variable's value is modified. Table 85 presents values that can be used to define a nominal condition for a typical study road segment for the crash sites. Most of the variables were assigned a value similar to their average condition in the sample data (e.g. lane width of 11 ft and graded shoulder width of 5 ft). Since approximately 80% of the crash locations did not have paved shoulders, the paved shoulder width is assigned a value of zero feet for the nominal condition. For a state indicator value of zero for AL and SC, a road segment defined by the nominal condition represents Georgia conditions. Since lighting conditions consistently had a significant influence on fatal crash type outcomes, the research team evaluated the influence from each predictor under day light and dark without supplemental lighting conditions separately so as to identify unique patterns.

Table 85: Description of Road Nominal Condition for Evaluating Single-Vehicle Models

Variables
AL SC JUNCTION LW PSW GSW LCURV CREST RHR67 ADT LU_C DARKUNLIT HR_DEEPSLEEP

Conditions
0 0 0 (a road segment) 11 ft (lane width = 11 ft) 0 ft (paved shoulder width = 0 ft) 5 ft (graded shoulder width = 5 ft) 0 (road horizontal alignment is not a curve to the left) 0 (road vertical alignment is not a crest vertical curve) 0 (roadside hazard rating 1 through 5) 3000 veh/day (average daily traffic estimated as 3000 veh/day) 0 (not in the proximity of a commercial driveway) 1 (dark without supplemental lighting) or 0 (other) 0 (crash did not occurred between 1 a.m. 3 a.m.)

6.1.4.1 Lane Width
Table 86, Figure 89, and Figure 90 demonstrate the probability of a single-vehicle fatal crash and how it varies with lane width (from 8 to 12 ft) at four daily traffic volume levels ranging from low to high volumes (400 veh/day, 2000 veh/day, 5000 veh/day, and 10,000 veh/day). All other variables in the model used the nominal condition values (see Table 85).

164

Table 86: Crash Probabilities based on Lane Width and ADT (Three-State Model, SingleVehicle)

ADT

(veh/day)

8

Lane Width (ft)

9

10

11

12

400

0.97 0.89 0.94 0.82 0.91 0.73 0.85 0.61 0.77 0.48

2000

0.96 0.87 0.94 0.80 0.90 0.70 0.83 0.57 0.74 0.44

5000

0.95 0.84 0.92 0.75 0.86 0.63 0.79 0.50 0.68 0.37

10,000

0.92 0.76 0.87 0.65 0.80 0.52 0.70 0.38 0.57 0.27

Note: The values shown in the shaded cells represent the probability of a crash during dark conditions where supplemental lighting is not present. The values in the cells that are not shaded represent crash probability for daylight conditions, dark conditions with supplemental lighting, dusk, and dawn.

Figure 89: Dark without Street Lights -- Lane width by ADT (Three-State Model, SingleVehicle)
165

Figure 90: Daylight, Dark with Lighting, Dusk, or Dawn -- Lane width by ADT (ThreeState Model, Single-Vehicle)
The likelihood of a single-vehicle fatal crash occurring reduces with an increase in the lane width despite lighting conditions and traffic volume levels. Alternatively, as the lane width increases, the single-vehicle run-off-road likelihood decreases more rapidly during daylight conditions than when it is dark without supplement lighting. This trend may imply that lane widening is a more effective countermeasure to help prevent daytime single-vehicle fatal crashes. More than half of the single-vehicle fatal crashes in this study, however, occurred during dark conditions without any supplemental lighting.
Single-vehicle fatal crash occurrence is sensitive to the various daily traffic volume levels for all lane width values. Interestingly, a higher traffic exposure can be associated with a lower likelihood of a single-vehicle fatal crash. As found in the project data, the higher traffic volume locations were more likely to occur at roads with the wider lane widths of 11 or 12 ft. For crashes that occur during nighttime conditions and where there is no supplement lighting, the probability of single-vehicle fatal crash is less sensitive to the various ADT levels. This lack of sensitivity is particularly evident for roads with narrow lane widths (8 to 10 ft).
166

6.1.4.2 Paved and Graded Shoulder Width
As previously indicated, the three-state model includes a statistically significant interaction effect between the paved and graded shoulder width. As a result, this analysis illustrates how sensitive the probability or a single-vehicle crash is based on varying graded shoulder widths (0 ft, 2ft, 4ft, 6ft, and 8ft) at various levels of paved shoulder width. As shown in Table 87, Figure 91, and Figure 92, the probability of a single-vehicle fatal crash when the graded shoulder width is increased does not vary substantially if there is no companion paved shoulder (paved shoulder width = 0 ft). Alternatively if there is a paved shoulder present, the probability of a singlevehicle fatal crash drops significantly when the graded shoulder width is increased. This relationship suggests that the combination of paved and graded shoulders collectively helps to enhance safety and reduce the likelihood of single-vehicle fatal crashes.

Table 87: Crash Probability for Paved and Graded Shoulder Width (Three-State Model, Single-Vehicle)

Paved Shoulder

Graded Shoulder Width (ft)

width (ft)

0

2

4

6

8

0

0.85 0.61 0.84 0.58 0.83 0.56 0.81 0.54 0.80 0.51

2

0.84 0.58 0.78 0.49 0.71 0.40 0.63 0.32 0.54 0.24

4

0.82 0.55 0.71 0.40 0.56 0.26 0.41 0.16 0.27 0.09

6

0.81 0.53 0.63 0.31 0.40 0.15 0.21 0.07 0.10 0.03

Note: The values shown in the shaded cells represent the probability of a crash during dark conditions where supplemental lighting is not present. The values in the cells that are not shaded represent crash probability for daylight conditions, dark conditions with supplemental lighting, dusk, and dawn.

For daylight conditions or locations with supplemental lighting, the probability of a singlevehicle fatal crash occurring decreases at a slower rate for wider graded shoulders than when the lighting conditions are dark with no supplemental lighting. This observation is particularly true at locations with a wider paved shoulder.

167

Figure 91: Dark without Street Lights -- Graded and Paved Shoulder Width (Three-State Model, Single-Vehicle)
Figure 92: Daylight, Dark with Lights, Dusk, or Dawn -- Graded and Paved Shoulder Width (Three-State Model, Single-Vehicle) 168

While the three-state model includes the interaction effect between paved and graded shoulder width, the Georgia model only includes the paved shoulder width as a significant shoulder-type variable. To enable a comparison of modeling results between the Georgia model and the threestate combined model, the variable analysis for the Georgia model is also based on the same road nominal conditions as described in Table 85. The Georgia model also includes fatal crash type probabilities as they are associated with safety restraint usage by at-fault drivers at the time of the crash.
Overall, the Georgia model predicts a reduction in the probability of a single-vehicle fatal crash occurring as the lane width increases for various paved shoulder widths. This observation is consistent with the findings for the three-state model. For at-fault drivers who use safety restraints, the likelihood of being involved in a single-vehicle fatal crash for various lanes widths and paved shoulder widths is depicted in Table 88. The observed relationship between paved shoulder widths (0 ft, 2 ft, 4 ft, and 6 ft) and lane widths (9 ft to 12ft) appears to vary in a manner similar to that observed for the three-state model for dark conditions without supplementation lighting (see Figure 93 and Figure 97) and for daylight or dark with supplemental lighting conditions (see Figure 94 and Figure 98) respectively.

Table 88: Crash Probability for Paved Shoulder and Restraint System Usage (GA only Model, Single-Vehicle)

Paved

Shoulder

Width (ft)

9

Lane Width (ft)

10

11

12

Driver Used Safety Restraints (RESTRAINT=1)

0 2 4 6

0.95 0.86 0.89 0.72 0.78 0.53 0.60 0.33 0.90 0.75 0.80 0.56 0.63 0.36 0.43 0.20 0.82 0.60 0.66 0.39 0.46 0.22 0.27 0.11 0.69 0.42 0.49 0.24 0.30 0.12 0.15 0.06

Driver Did Not

0

0.98 0.95 0.96 0.89 0.92 0.78 0.83 0.61

Use Safety

2

0.97 0.90 0.93 0.80 0.85 0.64 0.70 0.44

Restraints

4

0.94 0.82 0.86 0.67 0.73 0.47 0.54 0.28

(RESTRAINT=0)

6

0.88 0.70 0.76 0.50 0.57 0.31 0.37 0.16

Note: The values shown in the shaded cells represent the probability of a crash during dark conditions where supplemental lighting is not present. The values in the cells that are not shaded represent crash probability for daylight conditions, dark conditions with supplemental lighting, dusk, and dawn.

169

Figure 93: Dark without Street Lights -- Paved Shoulder Width and Safety Restraint Used (GA only Model, Single-Vehicle)
Figure 94: Daylight, Dark with Lights, Dusk, or Dawn -- Paved Shoulder Width and Safety Restraint Used (GA only Model, Single-Vehicle) 170

For at-fault drivers who do not utilize safety restraints (see Figure 95 and Figure 96) the Georgia model predicts a slightly different pattern for the single-vehicle fatal crash occurrence. In general, the likelihood that an at-fault driver who does not use safety restraints will be involved in a single-vehicle fatal crash is greater for all lighting conditions.
Figure 95: Dark without Street Lights -- Paved Shoulder Width and Safety Restraint Not Used (GA only Model, Single-Vehicle)
171

Figure 96: Daylight, Dark with Lights, Dusk, or Dawn -- Paved Shoulder Width and Safety Restraint Not Used (GA only Model, Single-Vehicle)
Figure 97: Dark without Street Lights -- Paved Shoulder Width (Three-State Model, Single-Vehicle) 172

Figure 98: Daylight, Dark with Lights, Dusk, or Dawn -- Paved Shoulder Width (ThreeState Model, Single-Vehicle)
6.1.4.3 Roadside Condition
Table 89 and Figure 99 demonstrates the sensitivity of the roadside hazard rating on the likelihood of a single-vehicle fatal crash occurring if all other conditions are nominal (as shown in Table 85). The perceived hazardous roadside environment (RHR67=1) increases the risk of a single-vehicle fatal crash for all lighting conditions. For example, sites with a roadside hazard rating 6 or 7 have an increased likelihood of a single-vehicle fatal crash ranging from 9 to 28% for daylight, dark with supplemental lighting, dusk, and dawn conditions when compared to the more traversable roadside condition (represented by RHR67=0). For dark conditions without supplemental lighting, the steeper, less traversable roadside condition also increases the likelihood of a fatal, single-vehicle crash from 3 to 17% for lanes widths of 8 ft to 12 ft. Values shown in Table 89 also indicate that the probability of a fatal single-vehicle crash for all roadside hazard rating categories is substantially greater at locations with narrower lane widths.
173

Table 89: Roadside Hazard Rating (Three-State Model, Single-Vehicle)

RHR67

8

LANE WIDTH (ft)

9

10

11

12

Flatter, MoreTraversable 0.99 0.95 0.98 0.92 0.96 0.87 0.93 0.79 0.89 0.69 Roadside (RHR67=1) Steeper, NonTraversable 0.96 0.86 0.93 0.78 0.89 0.68 0.82 0.55 0.72 0.41 Roadside (RHR67=0)

Note: The values shown in the shaded cells represent the probability of a crash during dark conditions where supplemental lighting is not present. The values in the cells that are not shaded represent crash probability for daylight conditions, dark conditions with supplemental lighting, dusk, and dawn.

Figure 99: Roadside Hazard Rating (Three-State Model, Single-Vehicle) 174

6.1.4.4 Horizontal and Vertical Alignment
Table 90, Figure 100, and Figure 101 depict the influence of horizontal and vertical curvature for the two general lighting conditions and the influence of this geometry on the probability of a single-vehicle fatal crash. The three-state model previously discussed included an interaction effect between horizontal curve direction and crest vertical curvature. It is likely that this interaction is due to deteriorated sight distance conditions resulting from the combined effects of the horizontal and vertical geometry. For dark conditions without supplemental lighting, the presence of a crest vertical curve combined with a horizontal curve to the left will increase the chance of a single-vehicle fatal crash substantially. This crash probability is marginally reduced with improved lighting conditions. For locations that do not have overlapping horizontal and vertical geometry, the locations with curves to the left have a lower probability of the singlevehicle fatal crash than the combined geometry conditions but also have a higher probability of this crash type when compared to roads that did not have horizontal curves to the left.

Table 90: Sensitivity to Curve Direction and Vertical Alignment (Three-State Model, Single-Vehicle)

Curve to the Left

CREST

8

LANE WIDTH (ft)

9

10

11

12

Yes

Yes

(LCURV (CREST 0.99 0.97 0.98 0.95 0.97 0.91 0.96 0.86 0.93 0.78

=1)

=1)

Yes

No

(LCURV (CREST 0.98 0.93 0.97 0.89 0.94 0.82 0.91 0.73 0.85 0.61

=1)

=0)

No

Yes

(LCURV (CREST 0.80 0.52 0.70 0.39 0.58 0.27 0.45 0.18 0.32 0.11

=0)

=1)

No

No

(LCURV (CREST 0.96 0.86 0.93 0.78 0.89 0.68 0.82 0.55 0.72 0.41

=0)

=0)

Note: The values shown in the shaded cells represent the probability of a crash during dark conditions where supplemental lighting is not present. The values in the cells that are not shaded represent crash probability for daylight conditions, dark conditions with supplemental lighting, dusk, and dawn.

175

Figure 100: Dark without Street Lights -- Road Alignment (Three-State Model, SingleVehicle)
Figure 101: Daylight, Dark with Lights, Dusk, or Dawn -- Road Alignment (Three-State Model, Single-Vehicle) 176

6.1.4.5 Road Junction/Intersection
As previously indicated, single-vehicle run-off-road and head-on crashes are primarily associated with road segment locations rather than intersections, while angle crashes are more likely to be associated with road junctions/intersections. Table 91 and Figure 102 present the sensitivity analysis for evaluation of the probability of a single-vehicle fatal crash occurring at an intersection or at a road segment. The three-state model predicts that single-vehicle fatal crashes have approximately a 13 to 20% greater chance or occurring at road segments than at intersection locations. This analysis assumes all other conditions are held constant at the nominal values (see Table 85). As can be expected, angle crashes are more likely to occur at intersection locations since they involve two vehicles that cross paths, whereas single-vehicle crashes occur more frequently at segment locations but narrow lane widths at intersection locations will tend to increase the likelihood of a single-vehicle crash at that location, too.

Table 91: Sensitivity to Road Junction/Intersection (Three-State Model, Single-Vehicle)

Location

Lane Width (ft)

8

9

10

11

12

Intersection (JUNCTION=1)

0.73

0.62

0.48

0.35

0.24

Segment (JUNCTION=0)

0.86

0.78

0.68

0.55

0.41

177

Figure 102: Road Junction (Three-State Model, Single-Vehicle Crash) 6.1.4.6 Land Use Type As shown in Table 92 and Figure 103, single-vehicle fatal crashes are approximately 25 to 33% less likely to occur in the vicinity of a commercial driveway than at other locations (including near residential and industrial driveway locations). At commercial driveway locations, heavier traffic can be expected resulting in a greater opportunity for vehicle interactions. The observed reduction in single-vehicle crash probability at these locations may be due to a more alert driver in the vicinity of a commercial driveway, but this trend could also simply be due to the greater opportunity for a multiple-vehicle crash due to the higher traffic volumes.
178

Table 92: Sensitivity to Land Use Types (Three-State Model, Single-Vehicle)

Dark Conditions/ No

Lighting

Land Width (ft)

Proximity to Driveways

8

9

10

11

12

Near Commercial Driveway (LU_C=1)

0.61

0.48

0.35

0.24

0.15

Not in proximity of a

Commercial

0.86

0.78

0.68

0.55

0.41

Driveway(LU_C=0)

Note: The data in this table represents daylight, dark with supplemental lighting, dusk, and dawn conditions.

Figure 103: Land Use Type (Three-State Model, Single-Vehicle)
6.1.4.7 Time of Day Single-vehicle fatal crashes occurred during all times of the day; however, the model development process identified an increased probability for single-vehicle fatal crashes on rural two-lane roads between the hours of 1 a.m. and 3 a.m. This time period coincides with the conventional deep sleep stage of the circadian biological clock for the drivers. Table 93 and Figure 104 demonstrate that the probability of a single-vehicle fatal crash (based on various lane
179

widths) does not fluctuate considerably during this "deep sleep" time period. Alternatively, this time period can be contrasted to the remainder of the day and evaluated based on lane width to identify a decreasing probability of a single-vehicle crash with an increasing lane width. In general terms, these finding reinforce the perception that the driver's performance during the "deep sleep" time period can be impaired due to increasing fatigue or similar biological reactions. It is also important to note that the toxicology information in the crash database was incomplete (due to a delay in testing and reporting), but it is also likely that some drivers may be under the influence of drugs or alcohol following an evening of entertainment and this could indirectly be represented by these findings.

Table 93: Sensitivity to Time of Crash (Three-State Model, Single-Vehicle)

Time of Crash

Lane Width (ft)

8

9

10

11

12

1 a.m. to 3 a.m. (HR_DEEPSLEEP=1)

0.99

0.99

0.98

0.97

0.94

All hours excluding

1 a.m. to 3 a.m.

0.96

0.93

0.89

0.82

0.72

(HR_DEEPSLEEP=0)

Figure 104: Time of Crash (Three-State Model, Single-Vehicle) 180

6.1.4.8 Lighting Condition
As summarized in the previous sensitivity evaluations, this research determined that there is a greater chance of a single-vehicle fatal crash during dark driving conditions where supplemental street lights are not present. As shown in Figure 89 and Figure 90, the narrower width of the travel lane has a greater influence on the increased crash probability during all lighting conditions (with dark conditions resulting in the greatest crash probability). This observation may also suggest that lane widths that are narrower than common standard widths introduce a much greater safety risk for all lighting conditions. Improving visibility with enhanced lane marking or supplemental delineation at night may help to reduce single-vehicle run-off-road fatal crash at these locations. Since horizontal curves appear to be one of the most common locations for single-vehicle fatal crashes on rural two-lane highways, the use of supplemental devices such as reflective pavement markers may prove beneficial at these select locations.
6.2 Head-on Fatal Crash A commonly observed multiple-vehicle fatal crash observed for the rural two-lane highway locations was a head-on crash. Though other multiple-vehicle crashes are more common, the frequently occur in the vicinity of driveways or intersections and often have much lower casualty rates that the more severe head-on crash. As a result, this section presents binary logit models developed for the combined four-state (AL, GA, MS, and SC) database. In addition, this study also presents three state-specific models for Alabama, Georgia, and Mississippi. The similar state-only South Carolina model for head-on fatal crashes and other types of multiple-vehicle fatal crashes did not produce meaningful results.
6.2.1 Combined-State Models (AL, GA, MS, and SC)
Table 94 presents independent variables that this research determined to be significant for differentiating head-on fatal crashes from other multiple-vehicle fatal crashes. These variables include: presence of road segment (not an intersection), lane width, horizontal curve direction, ADT, and number of driveways. In addition, Table 95 illustrates descriptive statistics for these independent variables. As shown, the average lane width was 11.2 ft with a range from 8 ft to 12 ft. The observed daily traffic volume at crash locations ranged from 150 to 16,550 vehicles per
181

day, with an average value of 3,990 vehicles per day. The number of driveways within 250 ft upstream and downstream of the crash locations was, on average, a value of two. In addition to location indicators (for the individual states), there were two categorical variables: road junction type (segment) and horizontal curve direction. The proportional distribution of these two variables is presented in Table 96.

Table 94: Variable Description (Four-State Model, Head-on)

Types

Variables

Descriptions

AL

1 if in Alabama, 0 otherwise

Location Indicator MS

1 if in Mississippi, 0 otherwise

SC

1 if in South Carolina, 0 otherwise

Road Junction type SEGMENT

1 if a road segment, 0 otherwise

Geometric design features
Traffic volume Roadside Conditions

LW RCURV ADT NUMDRVWAY

Lane width (ft)
1 if curve to the right, 0 otherwise (curve to the left or straight alignment) Average daily traffic (103 veh/day)
Number of driveways within 250 ft upstream and downstream of the crash location

Table 95: Continuous Variables Descriptive Statistics (Four-State Model, Head-on)

Variable LW (ft) ADT (veh/day) NUMDRVWAY

Mean 11.2 3,990
2

Std Dev 0.9 3,139 2

Minimum 8 150 0

Maximum 12
16500 13

Table 97 illustrates the model estimation results. Among the 219 multiple-vehicle fatal crashes available for the four-state combined model, 95 of the crashes were head-on fatal crashes. The Hosmer and Lemeshow goodness-of-fit test that measures how well a binary logit model fits the observed data reports and acceptable result (p-value = 0.2969 > 0.05) as shown in Table 97.

182

Table 96: Distribution of Categorical Variables (Four-State Model, Head-on)

Variable AL MS SC SEGMENT RCURV

Status
0 (Not Alabama) 1 (Alabama) 0 (Not Mississippi) 1 (Mississippi) 0 (Not South Carolina) 1 (South Carolina) 0 (Intersection) 1 (Segment) 0 (Curve to Left or Straight) 1 (Curve to Right)

Percent (%)
73 27 82 18 72 28 30 70 78 21

Table 97: Model Estimation (Four-State Model, Head-on)

Variable Intercept AL MS SC SEGMENT RCURV LW ADT NUMDRVWAY Observations (Head-on/Other) AIC SC -2 Log L R-Square
Chi-Square 9.5647

Estimate

P- value

4.5746

0.0421

0.6566

0.1591

0.0817

0.8607

0.1125

0.8004

1.8649

<.0001

1.1569

0.0027

-0.6344

0.0024

0.2293

0.0003

-0.1715

0.0517

219 (95/124) 253.849

284.351

235.849

0.2531

Hosmer and Lemeshow Goodness-of-Fit Test

DF

Pr > ChiSq

8

0.2969

183

The resulting multiple-vehicle head-on fatal crash prediction model is then presented as follows:
4-state = 4.5746 + 0.6566AL + 0.0817MS + 0.1125SC +1.8649SEGMENT +1.1569RCURV -0.6344LW + 0.2293ADT - 0.1715NUMDRVWAY

The probability of multiple-vehicle head-on fatal crash can then be predicted under a given set of

conditions as:

Pr(Head

-

on) 4 - state

=

exp(4-state ) 1 + exp(4-state )

(12)

As shown in Equation (12), the variables that are significantly associated with a head-on fatal crash occurrence include road junction type, lane width, horizontal curve direction, ADT, and the number of driveways. The probability of a head-on fatal crash occurrence increases at road segments with horizontal curves to the right when contrasted to other horizontal geometric configurations (straight alignments or curves to the left). Other variables that significantly increase the likelihood of head-on fatal crashes include road segment locations and locations with higher traffic volumes (ADT). Head-on fatal crashes tend to have lower likelihood of occurring at locations with wider lane widths and more frequent driveways. The driveway density influence could be a surrogate for other factors not directly measured such as a change in land use (so more driver awareness) or an increase in other types of multiple-vehicle crashes at these locations. The application of this head-on model should be limited to the range of data available in the dataset and presented in Table 95 for the continuous variables in the model.
The three location indicator variables, AL, MS, and SC, were statistically insignificant. This result suggests that head-on fatal crashes for all four states (Georgia was the base state for this model) are similar and do not have significant differences. As discussed previously, the singlevehicle fatal crash model identified crashes in Mississippi as unique when compared to the other states. Since this relationship does not extend to the head-on fatal crash model, the final combined-state model for the head-on fatal crashes will retain all four states and a three-state model is not necessary.

184

6.2.2 Models by State
This section presents results for state-specific head-on fatal crash prediction models for Alabama, Georgia, and Mississippi. As previously indicated, the individual state head-on crash modeling effort for South Carolina did not result in a meaningful fatal crash prediction model.

6.2.2.1 Alabama
For the state of Alabama, the head-on fatal crash probability model included the following four significant independent variables: road segment locations, lane width, average daily traffic, and number of driveways within 250 ft upstream and downstream of crash locations (see Table 98). Table 99 lists the descriptive statistics for the continuous variables of lane width, average daily traffic, and the number of driveways. Similarly, Table 100 depicts the distribution of the road junction type, the only significant categorical variable for this model. As shown in Table 99, the estimated daily traffic volume for the crash locations ranged from 150 up to 14,036 vehicles per day, with an average value of 3,924. Lane widths ranged from 8 to 12 ft with an average value of 11.2 ft. Crashes occurred at locations with an average of two driveways in the immediate vicinity of the crash.

Table 98: Variable Description (AL only Model, Head-on)

Variables SEGMENT LW ADT
NUMDRVWAY

Descriptions
1 if crash location is road segment, 0 otherwise Lane width (ft) Average Daily Traffic (veh/day) Number of driveways within 250ft upstream and downstream of the crash location

Table 99: Continuous Variable Descriptive Statistics (AL only Model, Head-on)

Variable LW (ft) ADT (veh/day) NUMDRVWAY

Mean 11.2 3,924
2

Std Dev 1.0 3,471 2.4

Minimum 8 150 0

Maximum 12
14,036 13

185

Table 100: Distribution of Categorical Variables (AL only Model, Head-on)

Variable SEGMENT

Status
0 (Intersection) 1 (Segment)

Percent (%)
37 63

As shown in Table 101 and Equation (13), for crashes that occur in Alabama there is a probability that if a multiple-vehicle fatal crash occurs it will be a head-on crash at road segments with higher traffic volumes. Head-on multiple-vehicle crashes are less likely to occur at locations with wider lane widths or frequent driveways. Out of the 56 multiple-vehicle fatal crashes studied for Alabama, 29 crashes were head-on fatal crashes. As presented in Table 101, the Hosmer and Lemeshow goodness-of-fit test resulted in a p-value as 0.2344. This represents an acceptable model fit for the observed data.

Table 101: Model Estimation (AL only Model, Head-on)

Variable

Estimate

P- value

Intercept

3.9412

0.4129

SEGMENT

3.8084

0.0003

LW (ft) ADT (103 veh/day)

-0.6225 0.2973

0.172 0.0696

NUMDRVWAY

-0.3525

0.0591

Observations (Head-on/Other) AIC

56 (29/27) 54.425

SC

64.552

-2 Log L

44.425

R-Square

0.4466

Hosmer and Lemeshow Goodness-of-Fit Test

Chi-Square 9.2624

DF

Pr > ChiSq

7

0.2344

186

Let:
AL = 3.9412 + 3.8084SEGMENT - 0.6225LW + 0.2973ADT -0.3525NUMDRVWAY

The probability of a multiple-vehicle head-on fatal crash can then be predicted as:

Pr(Head

-

on) AL

=

exp(AL ) 1 + exp(AL )

(13)

6.2.2.2 Georgia
Table 102 presents summarizes a description of the key (independent) variables determined to be significantly associated with the head-on fatal crash probability model for the state of Georgia. These variables include: road junction type (segment), horizontal curve direction, lane width, average daily traffic, and safety-restraint use by the at-fault drivers. Table 103 further defines the ranges for the continuous variables with an average lane width of 11.2 ft (range from 8.5 up to 12.0 ft) and an average observed ADT of 4,000 vehicles per day (ranging from 1,500 up to 16,500 vehicles per day). Table 104 depicts the distribution of the observed categorical variables (RCURV, SEGMENT, and RESTRAINT).

Table 102: Variable Description (GA only Model, Head-on)

Variables SEGMENT LW
RCURV
ADT RESTRAINT

Descriptions
1 if crash location is road segment, 0 otherwise Lane width (ft) 1 if curve to the right, 0 otherwise (curve to the left or straight alignment) Average Daily Traffic (veh/day) 1 if driver used safety restraint, 0 otherwise

Table 103: Continuous Variable Descriptive Statistics (GA only Model, Head-on)

Variable LW ADT (veh/day)

Mean 11.2 4,000

Std Dev 1.0 3,080

Minimum 8.5 1500

Maximum 12.0 16,500

187

Table 104: Distribution of Categorical Variables (GA only Model, Head-on)

Variable SEGMENT RCURV RESTRAINT

Status
0 (Intersection) 1 (Segment) 0 (Curve to Left or Straight) 1 (Curve to Right) 0 (Safety Restraint not Used) 1 (Safety Restraint Used)

Percent (%)
25 75 73 27 71 29

Table 105 and Equation (14) depicts the head-on fatal crash probability model for the state of Georgia. Out of the 58 multiple-vehicle fatal crashes studied for Georgia, 25 were head-on fatal crashes. The Hosmer and Lemeshow goodness-of-fit test resulted in an acceptable model fit (pvalue = 0.7549 > 0.05).

Table 105: Model Estimation (GA only Model, Head-on)

Variable

Estimate

P- value

Intercept

10.251

0.0776

SEGMENT

2.5699

0.0028

RCURV

1.975

0.0372

LW (ft) ADT (103 veh/day)

-1.2112 0.3684

0.0314 0.0368

RESTRAINT

-1.9335

0.0403

Observations (Head-on/Other) AIC

58 (25/33) 59.817

SC

72.18

-2 Log L

47.817

R-Square

0.4189

Hosmer and Lemeshow Goodness-of-Fit Test

Chi-Square

DF

Pr > ChiSq

5.0253

8

0.7549

188

Let:
GA = 10.251+ 2.5699SEGMENT +1.975RCURV -1.2112LW + 0.3684ADT -1.9335RESTRAINT

The probability of a multiple-vehicle head-on fatal crash can then be represented as:

Pr(Head

-

on)GA

=

exp(GA ) 1 + exp(GA )

(14)

This Georgia model indicates that the probability of a head-on fatal crash increases at road segment locations with horizontal curves to the right and higher traffic volumes. Locations with wider lane widths will result in a reduction in the number of head-on fatal crashes. Finally, the use of safety restraints by at-fault drivers is associated with a lower likelihood of a head-on fatal crash.

6.2.2.3 Mississippi
Table 106 depicts that in the event a multiple-vehicle crash should occur in Mississippi, the significant independent variables associated with the probability that the crash would be a headon include road junction type, horizontal curve direction, and average daily traffic. The observed daily traffic at the crash sites varied from 230 to 12,000 vehicles per day with an average daily traffic volume of 3,823 (see Table 107). Table 108 demonstrates the distribution of the categorical variables (RCURV and SEGMENT) with approximately 25% of the observed multiple-vehicle fatal crashes occurred at locations with a horizontal curve to the right, and about two-thirds occurring at road segment locations.

Table 106: Variable Description (MS only Model, Head-on)

Variables SEGMENT RCURV ADT

Descriptions
1 if crash location is road segment, 0 otherwise 1 if curve to the right, 0 otherwise (curve to the left or straight alignment) Average Daily Traffic (veh/day)

189

Table 107: Continuous Variable Descriptive Statistics (MS only Model, Head-on)

Variable ADT (veh/day)

Mean 3,823

Std Dev 3,118

Minimum 230

Maximum 12,000

Table 108: Distribution of Categorical Variables (MS only Model, Head-on)

Variable SEGMENT RCURV

Status
0 (Intersection) 1 (Segment) 0 (Curve to Left or Straight) 1 (Curve to Right)

Percent (%)
36 64 76 24

Table 109 and Equation (15) illustrates the head-on fatal crash prediction model for Mississippi. For the 50 multiple-vehicle fatal crashes studied for Mississippi, 21were head-on fatal crashes. The variables significantly associated with a head-on fatal crash in Mississippi included the road junction type, horizontal curve direction, and average daily traffic. More specifically, the probability of a head-on fatal crash increases at road segments with horizontal curves to the right and higher traffic volumes. The goodness-of-fit test resulted in p-value as 0.2917, indicating a good fit to the observed data.

190

Table 109: Model Estimation (MS only Model, Head-on)

Variable

Estimate

P- value

Intercept

-3.2673

0.0018

SEGMENT

2.3899

0.0079

RCURV ADT (103 veh/day)

1.2984 0.2255

0.1146 0.0544

Observations (Head-on/Other) AIC

50 (21/29) 57.921

SC

65.57

-2 Log L

49.921

R-Square

0.3038

Hosmer and Lemeshow Goodness-of-Fit Test

Chi-Square

DF

Pr > ChiSq

9.6335

8

0.2917

Let:
MS = -3.2673+ 2.3899SEGMENT +1.2984RCURV + 0.2255ADT

Then the probability of single-vehicle run-off-road fatal crash can be predicted under a given set

of conditions as:

Pr(Head

- on)MS

=

exp(MS ) 1 + exp(MS )

(15)

6.2.3 Summary of Head-on Fatal Crash Models
Table 110 summarizes the model estimation results for the state-specific models (AL, GA, and MS) as well as a combined-state model for head-on fatal crashes. In general, the head-on probability models include fewer influential variables than the models developed for the singlevehicle fatal crashes. Though it may be possible that there are simply fewer influential factors on the head-on crashes, this difference may also be due to the smaller sample size for head-on fatal crashes. The variables associated with the head-on fatal crashes include road segment location, horizontal curvature to the right, lane width, average daily traffic, driveway density, and at-fault

191

drivers' safety restraint use. The individual state models for Alabama, Georgia, and Mississippi generally contain a subset of the variables identified for the combined-state model. The use of safety-restraints, however, was only determined to be critical in the Georgia model. As previously indicated, there was no meaningful model available for head-on fatal crashes in South Carolina. The use of the state-only models for estimating crash probabilities in other states could not be tested due to the limited data sample size.

Table 110: Model Comparison (Head-on)

Variables

AL only GA only MS only Model Model Model

AL

MS

SC

SEGMENT

3.8084**

RCURV

LW (ft) ADT (x103 veh/day)

-0.6225 0.2973*

NUMDRVWAY

-0.3525*

RESTRAINT

** Significant level <0.05 * Significant level <0.1

2.5699** 1.975** -1.2112** 0.3684**
-1.9335**

2.3899** 1.2984
0.2255*

Four-State Model (AL, GA, MS, SC)
0.6566 0.0817 0.1125 1.8649** 1.1569** -0.6344** 0.2293** -0.1715*

As discussed previously, the research team recommends the use of the four-state (AL, GA, MS, and SC) combined model for evaluating the probability of head-on fatal crashes for two-lane rural highways. Despite the observed differences between the state-specific models and combined-state model, the crash location (road segment) and traffic volume (ADT) were determined to be significant independent variables with similar effects for all six models. In other words, in the event a fatal crash occurred, locations at road segments (not intersections) are more likely to experience a higher probability of a head-on fatal crash than at intersection locations (see Table 111). Increasing traffic volumes are also associated with a higher probability of a head-on fatal crash when a multiple-vehicle fatal crash occurs.

192

Table 111: Illustration of Effects (Head-On)

Variables

AL only Model

GA only Model

MS only Model

Four-State Model (AL, GA, MS, SC)

AL

+

MS

+

SC

+

SEGMENT

+ **

+ **

+ **

+ **

RCURV

+ **

+

+ **

LW (ft)

-

- **

- **

ADT (x103 veh/day)

+ *

+ **

+ *

+ **

NUMDRVWAY

- *

- *

RESTRAINT

- **

+ : Increase the probability of single-vehicle fatal crash, when the continuous variable increase or the indicator has the value of 1 versus 0.
- : Decrease the probability of single-vehicle fatal crash, when the continuous variable increase or the indicator has the value of 1 versus 0.
** Significant level <0.05 * Significant level <0.1

6.2.4 Variable Analysis This section presents a sensitivity analysis for the probability of head-on fatal crashes for a variety of previously identified contributing variables, including:
Lane width Average Daily Traffic Junction type (intersection) versus segment Lane width Horizontal alignment Number of driveways Safety restraint use

The research team performed an analysis based on the recommended four-state model (AL, GA, MS, and SC), as shown in Equation (12) in the Section 6.2.1. The analysis also includes the Georgia only model that was presented in Section 6.2.2.2 and represented by Equation 14. In order to assess changes of predicted crash type outcome probabilities at different levels of an independent variable, all other independent variables are held constant while the candidate
193

variable's value is modified. Table 112 presents values that can be used to define the nominal condition for a typical study road segment for the crash sites. Most of the variables were assigned a value similar to their average condition in the sample data (e.g. lane width of 11 ft and ADT of 4,000 vehicles per day). For a state indicator value of zero for AL, MS, and SC, a road segment defined by the nominal condition represents Georgia conditions. The research team also investigated the impact on the head-on fatal crash type occurrence for safety restraint usage by at-fault drivers in Georgia.

Table 112: Description of Road Nominal Conditions (Head-on)

Variables AL MS SC SEGMENT LW RCURV ADT NUMDRVWAY
RESTRAINT

Conditions
0 0 0 1 (a road segment) 11 ft (lane width = 11 ft) 0 (road horizontal alignment is not a curve to the right) 4,000 vehicles/day (average daily traffic estimated as 4,000) 2 (on average 2 driveways located within 250 ft of crash site) 1 (at-fault driver used safety restraint) or 0 (safety restraint not used)

6.2.4.1 Lane Width
Table 113 and Figure 105 demonstrated the probability of a multiple-vehicle, head-on fatal crash and how it varies with lane width (from 9 to 12 ft) at four daily traffic volume levels ranging from low to high volumes (400, 2000, 5000, and 10,000 vehicles per day). All other variables in the model used the nominal condition values (see Table 112). The head-on fatal crash occurrence is clearly sensitive to the various traffic volume levels for all lane width values. In particular for ADT values of 400, 2000, and 5000 vehicles per day, the three lines shown in Figure 105 appear to be relatively parallel indicating that the chance of a head-on fatal crash due to lane width changes at these various ADT levels is similar. In addition, the wider lane width associates have a lower chance of head-on fatal crashes, though as traffic volume increases this difference appears to diminish.

194

Table 113: Crash Probability for Lane Width (Four-State Model, Head-on)

ADT (veh/day)

9

Lane Width (ft)

10

11

12

400

0.62

0.46

0.31

0.19

2,000

0.70

0.55

0.40

0.26

5,000

0.82

0.71

0.57

0.41

10,000

0.94

0.89

0.80

0.68

Figure 105: Lane Width Crash Probability by ADT (Four-State Model, Head-on)
Table 114, Figure 106, and Figure 107 present the sensitivity of head-on fatal crash occurrence to lane width based on the Georgia-only model. The results are shown based on at-fault driver safety restraint usage. For conditions where at-fault drivers did not use their safety restraints, Figure 107 shows a limited sensitivity for the probability of a head-on fatal crash based on lane width at high traffic volume (10,000 vehicles per day) conditions, while stronger sensitivities are
195

associated with medium to low volumes. The probability of a head-on fatal crash is less sensitive to ADT levels at lane width of 9 or 10 ft. For example, the probability changes from 0.70 to 0.99 for lanes widths of 10 ft and ADT values of 400 to 10,000 vehicles per day, respectively. The probability varies from 0.17 to 0.88 for 12 ft lanes under similar traffic exposure conditions. For all traffic volume conditions, the head-on crash probabilities are sensitive to lane width and traffic volume when the at-fault driver was using safety restraints (see Figure 106).
Overall, the Georgia model indicates that appropriate use of safety restraints by at-fault drivers is more likely to reduce the likelihood of a head-on fatal crash for all levels of lane width and traffic exposure. When the at-fault driver does not appropriate use safety restraints, there is a higher probability that the vehicle will be involved in a head-on fatal crashes at all ADT levels across the various lane widths.

Table 114: Crash Probability for Lane Width (GA Only Model, Head-on)

Safety Restraints

ADT (veh/day)

9

Lane Width (ft)

10

11

12

400

0.53

0.25

0.09

0.03

Driver Used Safety Restraints (RESTRAINT=1)

2,000 5,000

0.67 0.86

0.38 0.65

0.15 0.36

0.05 0.14

10,000

0.98

0.92

0.78

0.51

Driver Did Not

400

0.89

0.70

0.41

0.17

Use Safety

2,000

0.93

0.81

0.56

0.27

Restraints

5,000

0.98

0.93

0.79

0.53

(RESTRAINT=0) 10,000

0.99

0.99

0.96

0.88

196

Figure 106: Lane Width by ADT and Used Safety Restraints (GA only Model, Head-on)
Figure 107: Lane Width by ADT and No Safety Restraints (GA only Model, Head-on) 197

6.2.4.2 Curve Direction
Table 115 and Figure 108 illustrate the influence of a curve to the right on the probability that a head-on fatal crash occurs based on varying lane widths. While the likelihood of a head-on fatal crash is substantial for lane widths of 9 ft and horizontal curves to the right, this probability reduces as lane widths increase. The likelihood that a road with a horizontal curve to the right will have a head-on fatal crash (in the event of a multiple fatal crash) is greater than for other horizontal geometry configurations (straight or curves to the left) by as much as 13 to 29% for 9 and 12 ft lanes respectively.

Table 115: Crash Probability for Curve Direction (Four-State Model, Head-on)

Curve to the Right?

9

Lane Width (ft)

10

11

12

Yes (RCURVE=1)

0.92

0.86

0.77

0.64

No (RCURVE=0)

0.79

0.66

0.51

0.35

Figure 108: Curve Direction (Four-State Model, Head-on) 198

The previous sensitivity analysis assessing Georgia head-on crashes for lane widths and horizontal curvature can be extended to the use of safety restraints by the at-fault drivers. For crashes where safety restraints were used the relationship of the lane width and the horizontal curvature appears similar to that previously reviewed (see Figure 109); however, as shown in Table 116 and Figure 110, the probability of the multiple-vehicle crash being a head-on crash is considerably greater when the at-fault driver does not utilize safety restraints.

Table 116: Crash Probability for Curve Direction (GA only Model, Head-on)

Safety Restraints Curve to the Right?

9

Lane Width (ft)

10

11

12

Driver Used

Yes (RCURV=1) 0.97

0.90

0.73

0.45

Safety Restraints

(RESTRAINT=1) No (RCURV=0)

0.81

0.56

0.28

0.10

Driver Did Not Yes (RCURV=1) 0.99

0.98

0.95

0.85

Use Safety

Restraints

No (RCURV=0)

0.97

0.90

0.73

0.44

(RESTRAINT=0)

Figure 109: Curve Direction and Used Safety Restraints (GA only Model, Head-on) 199

Figure 110: Curve Direction and No Safety Restraints (GA only Model, Head-on)
6.2.4.3 Road Segment According to the three-state model, road segments are more likely to be the location where a head-on multiple-vehicle fatal crash will occur (see Table 117 and Figure 111). For the Georgiaonly model, the use of safety restraints by at-fault drivers consistently results in a greater chance that a head-on fatal crash will occur at road segment locations with curves to the right more often than when the driver appropriately used the safety restraints( see Table 118, Figure 112, and Figure 113). Interpretation of this observation does not imply that the at-fault driver who does not use safety restraints has a more difficult time navigating the curve to the right, particularly at narrow lane locations. This observation provides possible insight that should a crash occur at this location, the crash with an at-fault driver who is not using safety restraints is more likely to result in a fatality.
200

Table 117: Crash Probability for Road Junction (Four-State Model, Head-on)

Location

9

Lane Width (ft)

10

11

12

Segment (SEGMENT=1)

0.79

0.66

0.51

0.35

Intersection (SEGMENT=0)

0.36

0.23

0.14

0.08

Figure 111: Road Junction (Four-State Model, Head-on) 201

Table 118: Crash Probability for Road Junction (GA only Model, Head-on)

Safety Restraints

Location

Lane Width (ft)

9

10

11

12

Driver Used Safety

Segment

Restraints

(SEGMENT=1)

0.81

0.56

0.28

0.10

(RESTRAINT=1)

Intersection

(SEGMENT=0)

0.25

0.09

0.03

0.01

Driver Did Not Use

Segment

Safety Restraints (SEGMENT=1)

0.97

0.90

0.73

0.44

(RESTRAINT=0)

Intersection

(SEGMENT=0)

0.70

0.40

0.17

0.06

Figure 112: Road Junction and Used Safety Restraints (GA only Model, Head-on) 202

Figure 113: Road Junction and No Safety Restraints (GA only Model, Head-on)
6.2.4.4 Number of Driveways
Table 119 and Figure 114 demonstrate the influence the number of driveways has on the probability that a multiple-vehicle fatal crash will be a head-on. As the lane width increases and the number of driveways increase the probability decreases. This observed trend likely indicates that with an increased driveway density, the multiple-vehicle fatal crashes are more likely to be crash types other than the target head-on crashes (such as angular crashes).

Table 119: Number of Driveways (Four-State Model, Head-on)

Number of

Driveways

9

Lane Width (ft)

10

11

12

0

0.84

0.73

0.59

0.44

2

0.79

0.66

0.51

0.35

4

0.72

0.58

0.42

0.28

203

Figure 114: Number of Driveways (Four-State Model, Head-on) 6.2.4.5 Restraint System (GA-only Model) As shown in Table 120 and Figure 115, if a multiple-vehicle fatal crash occurs, at-fault drivers who use safety restraints are less likely to be involved in a head-on fatal crash than at-fault drivers who do not use their safety restraints. This relation is consistent for a variety of lane widths (9 to 12 ft) but with an increasing probability as the lane width increases. This observation is similar to the relationship observed for the single-vehicle run-off-road fatal crash model.
204

Table 120: Crash Probability for Restraint System Use (GA only Model, Head-on)

Safety Restraints

9

Lane Width (ft)

10

11

12

Driver Used Safety

Restraints

0.81

0.56

0.28

0.10

(RESTRAINT=1)

Driver Did Not Use

Safety Restraints

0.97

0.90

0.73

0.44

(RESTRAINT=0)

Figure 115: Restraint System (GA only Model, Head-on) 205

This page blank intentionally 206

7.0 Practical Applications of Crash Type Prediction Models
7.1 Application Methodology The Highway Safety Improvement Program (HSIP), a Federal-aid funding program initiated in 2006, has a goal to achieve a significant reduction in fatalities and severe injuries on all public roads nationwide. The Georgia Department of Transportation is facing the challenge of how to effectively reduce fatal and serious injury crashes for rural two-lane highways since this road type is disproportionately representative of fatal crashes. Ultimately, an effective safety project provides the greatest benefit by reducing these serious crashes and, where possible, improving road conditions that may contribute to conditions conducive to such crashes.
In 2002, the authors of a final report for the investigation of fatal motor vehicle crashes on twolane rural highways in Georgia recommended five countermeasure categories that could be effective in terms of reducing fatal crashes on two-lane rural highways statewide (Washington, et al., 2002). As shown in Table 121, four of these general countermeasure categories previously recommended in the 2002 study align with variables identified as significant in this study (see Chapter 6). Both studies focused on fatal crash analysis but used different analysis approaches (the 2002 study evaluated crash causes while this study developed crash probabilities). Nevertheless, the general results for the probability models are consistent with those from the earlier study even though the analysis extended to neighboring states. This result suggests that the safety prediction models developed for this research effort can be applied to quantify crash type outcomes at specific sites if the detailed site geometry is known. As a result, these models can also be used to assist in potential countermeasure evaluations.
207

Table 121: Countermeasure Categories vs. Model Significant Effects

2002 Georgia Report
Geometric alignment improvements
Widening of lanes/pavement widths
Adding and/or widening graded/stabilized shoulders
Widening/improvement of clear zones

Significant Effects of Crash Type Study

Single-Vehicle

Head-on

Fatal Crash

Fatal Crash

Curve direction

Vertical alignment Interaction effect between curve

Curve direction

to the left and crest

Lane width

Lane width

Paved and graded shoulder

width Interaction effect of paved and

NA

graded shoulder width

Roadside hazard rating

NA

Safety engineers can apply fatal crash type prediction models as a unique tool for safety improvement projects, such as HSIP projects, and can use the models to specifically focus on reducing fatalities and serious injuries. This analytical tool also responds to the focus of the Georgia Strategic Highway Safety Plan (SHSP). As addressed in the Georgia SHSP, reducing serious injury crashes is one of the key emphasis areas, and lane departure and head-on crashes are two major crash types that often lead to severe injuries and fatalities. One of the major types of lane departure crashes on rural highways is the single-vehicle run-off road crash. Since current assessment techniques such as the rate quality control and equivalent property damage only methods commonly used in analysis do not include crash types, the use of predictive models can complement current procedures and help to identify locations where crash injury can be reduced with select countermeasure application. One example where crash type assessment can help clarify the expected crash conditions occurs when a median is constructed on a road so as to help prevent or reduce head-on collisions. If this road segment is over-represented by head-on collisions prior to median construction, then the number of severe injuries and fatalities should decrease following construction. If, on the other hand, the crashes were not associated with head-on crashes the median construction could conceivably have no influence on these severe crashes and may, in fact, contribute to increased speeds and more crashes. It is therefore helpful for safety engineers to know whether a candidate improvement location tends to have higher

208

likelihood of a major fatal crash type based on the existing road design characteristics. This assessment can occur on newer roads that do not have substantial crash history if these roads are built in a manner consistent with others in the region.
As discussed in the previous chapters, Georgia state agencies can evaluate the probability of a single-vehicle fatal crash type by using the three-state combined model as depicted by Equation (7) or by using the Georgia-only single-vehicle crash model as presented in Equation (9). Figure 116 outlines the proposed application procedure. In addition, Section 7.2 presents a sample problem that demonstrates how to apply the single-vehicle fatal crash prediction models to help engineers evaluate road safety based on existing conditions and proposed improvements.
Step 1: Identify candidate projects for evaluation: For example, high ranking HSIP projects for two-lane rural roads with higher proportion of single-vehicle crashes.
Step 2: Collect candidate crash site information as required by the model, see Equation 7.
Step 3: Calculate the single-vehicle fatal crash probability for existing conditions by applying Equation 7 (Three-State Model) or Equation 9 (Georgia only Model).
Step 4: Propose various safety improvement plans.
Step 5: Calculate the single-vehicle fatal crash outcomes for the proposed plans by applying the three-state single-vehicle model, see Equation 7.
Step 6: Evaluate / interpret the safety benefit of proposed plans.
Figure 116: Single-Vehicle Fatal Crash Type Model Application Six-Step Procedure
209

7.2 Application Example
This section provides an example of how to apply the six-step application process (See Figure 116). This analysis will specifically target reducing single-vehicle fatal crashes for two-way rural highway locations.
Step 1 and Step 2: Assume that a high-crash location has been previously identified using regional analysis procedures and that the road is a rural two-lane highway. This two-lane rural road segment has a known history of single-vehicle crashes that result in fatalities or serious injury. One specific location on this road has the existing characteristics as shown in Table 122.

Table 122: Sample Problem -- Existing Road Conditions for Georgia Site

Existing Condition Road segment? Alabama? South Carolina? Lane width Paved shoulder width Graded shoulder width Roadside hazard rating ADT
Land use
Driving during 1am to 3am? Curve to the left? Crest? Daylight, dark with lighting, dusk or dawn conditions Dark without supplemental street lights

Status Yes No No 11 ft 0 ft 8 ft 5 3,000 vehicles per day Driveways not for commercial use No Yes No
-
-

Variables JUNCTION = 0 AL = 0 SC = 0 LW=11 PSW = 0 GSW = 8 RHR67 = 0 ADT = 3
LU_C = 0
HR_DEEPSLEEP = 0 LCURV = 1 CREST = 0
DARKUNLIT = 0
DARKUNLIT = 1

Step 3: With the specific site information as depicted in Table 122, the probability of a single-vehicle fatal crash based on existing conditions during either the daylight, dark with lighting, dusk or dawn condition or the dark without supplemental street lights condition can be computed as follows:

210

Using the Three-State Combined Model (AL, GA, SC) as presented in Equation 5:
Let:
3-state = 6.6717 - 0.1855AL - 0.1167SC - 0.8078JUNCTION - 0.5407LW - 0.0542PSW - 0.0475GSW -0.0676(PSW * GSW ) + 0.788LCURV -1.7264CREST + 2.5199(LCURV * CREST ) + 1.1581RHR67 - 0.0965ADT -1.3722LU _ C + 1.3101DARKUNLIT + 1.8318HR _ DEEPSLEEP

For daylight, dark with lights, dusk, or dawn conditions (DARKUNLIT = 0):

3-state = 6.6717 - (0.1855 0) - (0.1167 0) - (0.8078 0) - (0.5407 11) - (0.0542 0) - (0.0475 8) -(0.0676 0 8) + (0.788 1) - (1.7264 0) + (2.5199 0 0) + (1.1581 0) - (0.0965 3) -(1.3722 0) + (1.3101 0) + (1.8318 0) = 0.8425

Pr(Single - veh - runoff )3-state

= exp(3-state ) 1 + exp(3-state )

=

1

e0.8425 + e0.8425

= 0.70

For dark without supplemental street lights (DARKUNLIT = 1):
3-state = 6.6717 - (0.1855 0) - (0.1167 0) - (0.8078 0) - (0.5407 11) - (0.0542 0) - (0.0475 8) -(0.0676 0 8) + (0.788 1) - (1.7264 0) + (2.5199 0 0) + (1.1581 0) - (0.0965 3) -(1.3722 0) + (1.31011) + (1.8318 0) = 2.1526

Pr(Single - veh - runoff

)3-state

= exp(3-state ) 1 + exp(3-state )

=

1

e2.1526 + e2.1526

= 0.90

Using the Georgia-only Model as presented in Equation 7: Let:
GA = 8.9011- 2.1473JUNCTION - 0.835LW - 0.3506PSW +1.7437LCURV +1.5662STRAIGHT +1.1195DARKUNLIT -1.1604RESTRAINT
For at-fault driver using safety restraints (RESTRAINT = 1) and under daylight, dark with lights, dusk, or dawn conditions (DARKUNLIT = 0):
GA = 8.9011- (2.1473 0) - (0.83511) - (0.3506 0) + (1.7437 1) + (1.5662 0) + (1.1195 0) -(1.1604 1) = 0.2994

211

Pr(Single - veh - runoff

)GA

= exp(GA ) 1+ exp(GA )

= e0.2994 1 + e0.2994

=

0.57

For at-fault driver using safety restraints (RESTRAINT = 1) but during dark conditions with no

supplemental street lights (DARKUNLIT = 1):

GA = 8.9011- (2.1473 0) - (0.83511) - (0.3506 0) + (1.7437 1) + (1.5662 0) + (1.11951) -(1.1604 1) = 1.4189

Pr(Single - veh - runoff

)GA

= exp(GA ) 1+ exp(GA )

= e1.4189 1 + e1.4189

=

0.81

For at-fault driver not using safety restraints (RESTRAINT = 0) and under daylight, dark with

lights, dusk, or dawn conditions (DARKUNLIT = 0):

GA = 8.9011- (2.1473 0) - (0.83511) - (0.3506 0) + (1.7437 1) + (1.5662 0) + (1.1195 0) -(1.1604 0) = 1.4598

Pr(Single - veh - runoff

)GA

= exp(GA ) 1+ exp(GA )

= e1.4598 1 + e1.4598

= 0.81

For at-fault driver not using safety restraints (RESTRAINT = 0) and during dark conditions with no supplemental street lights (DARKUNLIT = 1):

GA = 8.9011- (2.1473 0) - (0.83511) - (0.3506 0) + (1.7437 1) + (1.5662 0) + (1.11951) -(1.1604 0) = 2.5793

Pr(Single - veh - runoff

)GA

= exp(GA ) 1+ exp(GA )

=

1

e 2.5793 + e2.5793

=

0.93

The probability of single-vehicle fatal crash occurrence at this location based on the three-state combined model, Equation (7), is estimated as 0.70 at daylight, dark with light, dusk, or dawn conditions, and 0.90 for dark conditions without supplemental street lights. Similarly this example includes a comparison for the Georgia-only model values based on at-fault drivers with and without safety restraints and the varying lighting conditions. For at-fault drivers who used safety restraints, the likelihood that a single-vehicle fatal crash will occur during daylight, dark

212

with lights, dusk or dawn conditions is 0.57. For dark conditions without supplemental street lights the crash probability increases to 0.81. For at-fault drivers not using safety restraints, the Georgia-only model predicts much higher probabilities of 0.81 for the daylight, dark with lights, dusk and dawn and 0.93 for dark conditions without supplemental street lights.
The values predicted using the three-state combined model are higher than those predicted using the Georgia-only model for at-fault drivers using safety restraints and lower than those predicted for at-fault drivers not using safety restraints. Both models indicate that dark conditions without supplemental street lights are associated with higher probabilities of single-vehicle fatal crashes.
Step 4: Based on existing road conditions, proposed improvement plans may help to reduce singlevehicle fatal crashes by considering candidate improvement Plans B1 and B2 as depicted in Table 123.

Table 123: Existing Condition and Proposed Improvement Plan

Status
Existing Condition: A Proposed Improvement
Plan: B1 Proposed Improvement
Plan: B2

Lane Width (ft)
11 12
11

Paved Shoulder Width (ft)
0 0
3

Graded Shoulder Width (ft)
8 8
5

Countermeasures
--Lane Widening
Shoulder Enhancement

Step 5: Following a process similar to that demonstrated for Step 3, the probability of a single-vehicle fatal crash can be estimated using the three-state combined model (Equation 7) and the Georgiaonly model (Equation 9) respectively.

213

Table 124, Table 125, Table 126, and Figure 117 present the estimated results for the probability of a single-vehicle fatal crash at the example study location based on the existing conditions, and the two proposed improvement conditions, B1 and B2.

Table 124: Safety Evaluation (Three-State Model, Single-Vehicle)

Existing Plan: B1 Plan: B2

Lane Width
(ft)
11 12 11

Paved Shoulder Width (ft)
0 0 3

Graded Shoulder Width (ft)
8 8 5

Lighting Condition

Daylight

Dark No Lighting

0.70

0.90

0.57

0.83

0.45

0.75

Table 125: Safety Evaluation Used Safety Restraint (GA only Model, Single-Vehicle)

Existing Plan: B1 Plan: B2

Lane Width
(ft)
11 12 11

Paved Shoulder Width (ft)
0 0 3

Graded Shoulder Width (ft)
8 8 5

Lighting Condition

Daylight

Dark No Lighting

0.57

0.81

0.37

0.64

0.32

0.59

Table 126: Safety Evaluation No Safety Restraint (GA only Model, Single-Vehicle)

Existing Plan: B1 Plan: B2

Lane Width
(ft)
11 12 11

Paved Shoulder Width (ft)
0 0 3

Graded Shoulder Width (ft)
8 8 5

Lighting Condition

Daylight

Dark No Lighting

0.81

0.93

0.65

0.85

0.60

0.82

214

Figure 117: Safety Evaluation for Plan B1 and B2
Step 6: The three-state combined model requires more variables than the Georgia single-vehicle fatal crash model. The combined-state model benefits from evaluating more influential factors and can potentially identify confounding influences that may not be determined by the simpler-form Georgia model. The Georgia-only model provides the analyst with the potential to further assess the impact of an at-fault drivers' safety restraint use.
The three-state combined model and the Georgia-only model provide similar predictive results from two major aspects: Proposed plan B1 with lane widening and B2 with shoulder enhancement both would lower
the chance of a single-vehicle fatal crash when compared to the existing condition for both daylight, dark with lights, dusk, and dawn conditions as well as dark condition.
215

The proposed plan B2 with shoulder enhancements, based on both the three-state model and the Georgia model, results in a lower predicted number of single-vehicle fatal crashes when compared to estimated crashes resulting from the lane widening improvements proposed for plan B1.
Even though all models indicate a reduction in the single-vehicle fatal crashes for both day light and dark conditions, the effectiveness of countermeasures under different lighting conditions are different. For instance, the three-state model predicts that shoulder enhancement (Plan B2) reduce the probability of single-vehicle fatal crashes by 0.25 for daylight, dark with lights, dusk and dawn conditions. The same treatment only achieves a 0.15 probability reduction for dark conditions without supplemental street lights. Plan B1 and B2 improvements are less effective during dark conditions without supplemental street lights. Therefore, countermeasures that can improve the visibility of the travel lane may also be desirable, particularly at horizontal curve locations. Effective treatments may include the installation of delineators, raised pavement markers, and rumble strips at the selected locations. Safety engineers should make decisions based on overall considerations from predictive modeling results, previous documented countermeasure recommendations, economic impact, and engineering judgment.
Evaluation conclusion: Proposed plan B2, shoulder improvement, is the recommended countermeasure under the
context of single-vehicle fatal crash outcome reduction. Since the physical improvements have less influence on single-vehicle crashes during dark
conditions, it may be appropriate to enhance the location (particularly at horizontal curve locations) with other countermeasures that specifically increase safety during dark conditions. Due to the increased risk due to the lack of safety restraint use by at-fault drivers, it is recommended that the use of safety restraints be promoted in rural areas. Though the probability models provide indications regarding the effectiveness of improvements, the final improvement decisions should be based on cost/benefit analysis, as well as other potential conditions not available for assessment in the model development.
216

7.3 Application Limitation This study focuses on fatal crashes where at least one person was fatally injured. It is not appropriate to generalize the modeling results to crashes at all injury levels. Some crash types, for example, are more commonly associated with fatalities, while others may be more likely to result in crashes common to less severe injuries. For instance, a rear-end crash is more likely to be associated with minor injury crashes instead of fatal crashes. Head-on collisions, on the other hand, are more prone to contribute to fatalities. In addition, road geometric design features and other potential contributing factors can impact the crash type differently. The models developed for this study include a limited number of contributing factors. There are other potential factors that could influence fatal crashes, but these variables are not included in the model due to a variety of reasons. For example, the random fatal crash database may have some variables that are not well populated and therefore do not provide significant effects. It is also possible that there may be influence variables that are not available in the standard crash database or the supplement database used for this study. As previously indicated, the model development was based on a cross-sectional dataset that included fatal crashes that occurred primarily in 1997. Though this report includes an evaluation of historic crash trends in the southeast, it is not possible to determine that the 1997 crashes suitably represent current or future crash conditions.
217

This page blank intentionally 218

8.0 Expert Panel Findings: Rural Road Fatal Crash Countermeasures
As described previously in this report, fatal motor vehicle crash rates in the southeastern US are relatively high in comparison to the national average. Numerous possible explanations exist for the relatively high rates and include relatively lower safety restraint use, relatively higher rates of impaired driving, relatively higher proportions of driving on high speed rural roads, and perhaps less forgiving roadside environments on average. Other factors such as systematic differences in driver behavior, weather, etc., may also play a role.
Safety countermeasures typically considered to improve safety include a range of behavioral and engineering investments, ranging from enforcement and driver training to cable median barriers, improved culvert designs, and crash attenuation devices.
In some cases, however, there is insufficient prior evidence and/or information to determine exactly what impacts an investment may have on safety. This is unfortunate, as safety engineers need an estimate of the effectiveness of countermeasures prior to implementation, with emphasis on the region-specific transportation environment. In some cases the use of expert panels can be used to obtain estimates of the effectiveness of safety countermeasures.
The remainder of this chapter reports on an expert evaluation conducted to examine the effectiveness of countermeasures aimed to reduce fatal crashes in the southeastern US.
8.1 Description of Expert Evaluation Experts in transportation engineering and transportation safety were asked to independently evaluate the effectiveness of common countermeasures for a random set of 150 fatal crashes on rural roads using fundamental engineering principles of crash causation and knowledge of countermeasure functional mechanisms. The responses were then analyzed statistically to assess if the information culled from their combined expertise was repeatable. While the detailed
219

results of this experiment and similar evaluations of expert panels can be found in Melcher et al. (2001), Washington and Oh (2006), and Washington et al. (2009), the net result from all expert panel evaluations is that experts can be relied upon--especially in carefully controlled evaluation settings--to reach agreement across experts, to reflect uncertainty appropriately, and to provide useful information for selecting potentially useful countermeasures.
The experts in the Southeastern Fatal Crash study were asked to evaluate the countermeasures shown in Table 127 for 150 randomly selected fatal crashes. The safety countermeasures represent a set of commonly accepted strategies for improving safety on rural roads.

Table 127: Countermeasures examined in Southeastern US Fatal Crash Study

Category Pavement Markings
Lighting
Roadside Improvements
Reconstruct Roadway

Countermeasure (by No.)
1. Add edgeline, or upgrade existing edgeline 2. Add centerline, or upgrade existing centerline 3. Add no passing zone lines 4. Add segment lighting 5. Install guardrail 6. Improve clear zone 7. Relocate fixed object 8. Remove fixed object 9. Flatten side slope 10. Geometric realignment
11. Improve sight distance without geometric realignment
12. Install paved shoulder, or improve existing paved shoulder

Experts systematically evaluated these 12 countermeasures to determine their potential effectiveness on mitigating fatal crashes on rural roads in the southeastern US. Experts were provided details of the crashes, police reports, etc., to examine all of the relevant conditions associated with the random sample of crashes.

During the expert review process a number of countermeasure effectiveness measures were determined, specifically, how effective experts thought the countermeasures would be, how uncertain the effectiveness was estimate to be, and how often was a countermeasure determined to be inappropriate or not applicable to fatal crashes. The details of this evaluation are included in the following section.

220

8.2 Cumulative Advice from the Experts Experts evaluated the effectiveness of the 12 countermeasures for mitigating fatal crashes on rural roads in the southeastern US, as shown in Table 128. The effectiveness measure is the well known Accident Modification Factor (AMF), or Crash Modification Factor (CMF), which is an estimate of a countermeasures' impact on the number of crashes from before to after implementation of a countermeasure.
For example, the experts estimated that the addition of edge lines, when appropriate, has an average AMF of about 0.94, suggesting that one would expect about a 6% reduction in fatal crashes as a result of adding edge lines. Improving the clear zone (CM 6), in contrast, had an estimated effectiveness of 0.78--amounting to an estimated 22% reduction in fatal crashes when installed in appropriate instances.
Table 129 shows the expert evaluation results of fatal crash countermeasures ranked from the most effective to least effective. It also identifies the `typical' situations under which the countermeasures were found to be appropriate, and what proportion of crashes the countermeasures were recommended by the experts as effective. The table reveals a number of important and interesting findings.
The most effective countermeasure, improving clear zones, had an estimated CRF of 0.78 (would eliminate about 22% of fatal crashes) and was thought to be appropriate at about 6% of the rural road crash locations examined.
The second most effective countermeasure, with an estimated CRF of 0.80, was deemed appropriate in about 3.6% of rural road crash locations, where a gravel shoulder could be widened or added. Paving existing or widening and paving existing shoulders was thought to be effective in a combined 6.1% of rural road fatal crash locations.
221

Table 128: Expert Opinions on the Effectiveness of 12 Fatal Crash Countermeasures

CM 1 EDGELINE

Participant Mean Std Dev

1 2 3 4 5 TEAM

0.93 0.14 0.93 0.14 0.98 0.07 0.97 0.10 0.88 0.16 0.94 0.12

CM 2 - CENTERLINE

Participant Mean Std Dev

1 2 3 4 5 TEAM

1.00 0.00 0.98 0.07 0.98 0.07 1.00 0.00 0.92 0.15 0.98 0.08

CM 3 - NO PASSING ZONE LINES

Participant Mean Std Dev

1 2 3 4 5 TEAM

1.00 0.00 0.98 0.07 1.00 0.00 0.98 0.07 0.98 0.07 0.99 0.06

CM 4 ADD LIGHTING

Participant Mean Std Dev

1 2 3 4 5 TEAM

0.87 0.17 0.80 0.17 0.97 0.10 0.95 0.12 0.98 0.07 0.91 0.14

CM 5 - INSTALL GUARDRAIL

Participant Mean Std Dev

1 2 3 4 5 TEAM

0.67 0.29 0.75 0.26 0.97 0.10 0.95 0.12 0.67 0.00 0.83 0.24

CM 6 - IMPROVE CZ

Participant Mean Std Dev

1 2 3 4 5 TEAM

0.80 0.20 0.68 0.32 0.82 0.17 0.83 0.25 0.77 0.22 0.78 0.23

CM 7 RELOCATE OBJECT

Participant Mean Std Dev

1 2 3 4 5 TEAM

0.87 0.17 0.93 0.17 0.85 0.17 0.98 0.09 1.00 0.19 0.91 0.16

CM 8 - REMOVE OBJECT

Participant Mean Std Dev

1 2 3 4 5 TEAM

0.82 0.25 0.82 0.28 0.77 0.29 0.85 0.26 0.82 0.24 0.81 0.25

CM 9 - FLATTEN SIDE SLOPE

Participant Mean Std Dev

1 2 3 4 5 TEAM

0.72 0.16 0.87 0.25 0.93 0.17 0.91 0.15 0.97 0.10 0.88 0.19

CM 10 GEOM REALIGN

Participant Mean Std Dev

1 2 3 4 5 TEAM

0.87 0.17 0.93 0.14 0.88 0.16 0.85 0.25 0.98 0.07 0.90 0.17

CM 11 - SIGHT DISTANCE

Participant Mean Std Dev

1 2 3 4 5 TEAM

0.97 0.10 1.00 0.00 0.90 0.16 1.00 0.00 0.98 0.07 0.97 0.09

CM 12 - IMPROVE SHOULDER

Participant Mean Std Dev

1 2 3 4 5 TEAM

0.74 0.17 0.68 0.35 0.80 0.17 0.90 0.19 0.88 0.16 0.80 0.23

Installing guardrail at locations with unforgiving roadside environments was estimated to eliminate 17% of fatal crashes and was thought to be appropriate in 3.8% of the random sample of fatal crashes examined.

222

Perhaps surprisingly, the addition of lighting had an estimated CRF of 0.91 (9% reduction of fatal crashes) and was deemed most appropriate for critical road segments and not intersections. About 5% of crash locations were estimated to benefit from some form of lighting improvement.
Rigid objects on the roadside were also determined to be a fairly systematic problem, with about 3.7% of crash locations requiring object removal and 3% requiring object relocation. About 19% and 9% of fatal crashes were estimated to be reduced by such countermeasures respectively.
A fairly effective countermeasure is flattening the side-slope (CRF=0.88), which was determined to be an effective strategy in 4.4% of randomly sampled crash locations.
Geometric realignment--perhaps the most expensive of countermeasures (depends on extent of changes)--was ranked as the 6th most effective. It was estimated to be appropriate at about 4.8% of rural road fatal crash locations with a CRF of 0.90.
The experts independently evaluated 150 randomly sampled fatal crashes and determined that improving the roadside represented the most promising set of countermeasures. The most promising countermeasure--improving the clear zone--was thought to be appropriate in nearly 40% of rural road fatal crash locations. However, it is well known that rural roads are by their very nature located in remote areas with relatively low traffic volumes, relatively high speeds, and as such may in general be subject to such a general recommendation. As such, the finding should be taken in the context that the specific crash locations should be examined to determine if specific sites had sub-par roadside clearance zones.
223

Table 129: Statistics on the Expert Derived Countermeasure Effectiveness

Safety countermeasure Edge line
Center line
No passing zones Add lighting
Install guardrail
Improve clear zone
Relocate object
Remove object
Flatten side slope Geometric realignment Improve sight distance Improve shoulder

Effectiveness Rank
(effectiveness) 9
(0.94)
11 (0.98)
12 (0.99)
7 (0.91)
4 (0.83)
1 (0.78)
8 (0.91)
3 (0.81)
5 (0.88)
6 (0.90)
10 (0.97)
2 (0.80)

Typical Conditions for
Installation Worn, non-reflective or missing pavement
edge lines Worn, non-reflective or missing pavement
center lines Passing involved in
crashes Poor visibility for conditions; illumination thought to increase
safety Horizontal curve with
hazardous or unforgiving sideslope Obstacles in clear zone factor in crash; clear
zone insufficient Obstacle involved in crash can be relocated
to improve safety Obstacle involved in crash can be removed
to improve safety Side slope sufficient to worsen crash severity Geometric alignment major factor in crash Sight distance issues
contributed to crash
Shoulder surface or width contributed to crash

Proportion of Crash Locations Recommended
0.021
0.012
0.004 segments: 0.043 intersections: 0.013 upgrade: 0.002
0.038
0.061
0.030
0.037
0.044 0.048 0.021
add or widen gravel: 0.036 pave existing: 0.050
widen and pave existing: 0.011

Table 129 may leave the impression that the experts believed that a small proportion of crashes are preventable with the set of listed safety countermeasures. This table, however, does not convey information on what percentage of crash locations are believed to be correctable by one or more countermeasures. The accumulated opinions of the experts suggested that on average about 17% of all rural road fatal crash locations were correctable by at least one of the countermeasures listed, with a standard deviation of this estimate of 7.4%. If we apply the mean +- 2 standard deviations rule, this suggests that a 90% confidence interval includes the range

224

2.2% to 31.8%. In other words, for 90% of the crash locations examined the range of `correctability' is within the range 2.2% to 31.8%.
8.3 Summary of Expert Opinion Findings Experts were recruited to participate in a formal countermeasure evaluation process, with focus on fatal crashes on rural roads, mainly highways. The methodology was tested and evaluated with two different groups of experts in two different states on two independent occasions with favorable results. Experts are also being used extensively in the development of the Highway Safety Manual. In short, experts can contribute to understanding the safety performance of facilities, especially when little prior or relevant knowledge is available.
Experts who participated in the southeastern US fatal crash study contributed to the understanding of rural road safety in a number of interesting and insightful ways. Following are the highlights of what can be gleaned from these experts:
The most effective countermeasure, improving clear zones, had an estimated CRF of 0.78 (would eliminate about 22% of fatal crashes) and was thought to be applicable (useful) at about 6% of rural road crash locations examined.
The least effective countermeasure--installing "NO PASSING" zones, was thought to be practically ineffective, with a 1% reduction in fatal crashes. It was also found to be appropriate at only 0.4% of all crash locations examined.
The accumulated opinions of the experts suggested that on average about 17% of all rural road fatal crash locations were subject to improvement using at least one safety countermeasure, leaving about 83% as "uncorrectable" with the use of engineering countermeasures
The largest proportion of crash locations subject to improvement was 6%, associated with improving clear zones. The next most widely applicable countermeasures were improved geometric alignment (applicable to 4.8% of crash locations), flattening side slopes (4.4%), installing guard rail (3.8%), and removing objects from the roadside (3.7%).
225

The use of experts, combined with knowledge obtained through analysis and modeling--can drastically improve the collective understanding of the safety performance of rural roads (and other transportation system locations). The lessons learned from such evaluations should be combined with other sources of knowledge to gain a more complete understanding of rural road safety.
226

9.0 Conclusions
The rural two-lane highway in the southeastern United States is frequently associated with a disproportionate number of serious and fatal crashes. This study used physical site data from an earlier research effort to assess potential ways to address perceived safety hazards for these locations.
Prior to initiating an analysis, the research team conducted a literature review to determine what available safety models are published that apply to the rural road environment. Since road characteristics and the policies that establish the design of roads vary across jurisdictions, the published literature is limited to assessment of physical road features between jurisdictions and generally focuses on crashes within individual jurisdictions. The literature review also demonstrated that safety analysis includes five primary causal influences: vehicle occupant/driver, vehicle characteristic, road and roadside, crash characteristics, and environmental conditions. Though there are indications that alcohol and drug abuse are a primary contributor to the vehicle occupant/driver influence, these speculations cannot be quantitatively confirmed due to the quality of data for this variable.
This report also includes summary statistics for the available fatal crash databases. Since only four states provided databases that were populated with the variables needed for analysis, the summary statistics review data from Alabama, Georgia, Mississippi, and South Carolina. The summary statistics (see Chapter 3) evaluated how well the data set matched the larger fatal crash set from FARS for the year of crash using a chi-square test. The research team confirmed that the data adequately represented the larger fatal crash data set. This analysis then supported use of the data for development of probability models. The summary statistics also provided a good overview of the conditions for the crash data in each of the four states.
Since the use of crash data that is not recent may raise a question as to whether models developed using that data are applicable to current conditions, the research team performed a tenyear historic fatal crash trend analysis by contrasting the four study states with the FARS data
227

(see Chapter 4). Results from this trend analysis include several observations consistent with what one would expect. For example, the total number of fatal crashes has increased along with the vehicle miles traveled over time, but the crash rates have actually reduced. Actual causation factors could not be comprehensively evaluated, though, due to the limited number of variables available in the larger FARS dataset; however, over time the crash trends remained generally consistent. Two notable differences are variations in the use of safety restraints and crash site functional classifications. Both of these controls may be due to specific regional or policy changes. For example, the use of a primary seat belt law in the south is relatively recent. Georgia was the first of the four states to enact a primary seat belt law in 1996 (though the law includes an unusual exclusion for the driver of a pickup). In 1999, Alabama enacted a primary seat belt law while the South Carolina and Mississippi laws were not activated until 2005 and 2006 respectively. All four of these occurred during the ten-year crash trend and may be one explanation for a change in safety restraint usage.
The use of statistical models to help predict how a candidate countermeasure can help to reduce a specific type of crash can be valuable. The research team evaluated a wide variety of potential statistical models and determined that a logit model would be a powerful tool for determining the probability of a crash and by doing so helping to determine how to reduce the crash probability. Many options are available to estimate crashes (severity, frequency, crash type), but after considerable analysis the research team determined that the most meaningful models should be based on crash type. Crashes can include single-vehicle or multiple-vehicle collisions, but for rural two-lane roads the number of single-vehicle crashes is quite high so one of the crash types evaluated extensively in this research was a single-vehicle run-off-road crash. For the singlevehicle crashes, the following observations were identified:
Single-vehicle fatal crashes in Mississippi did not have similar causes than similar crashes in the other three states, so the cross-section model for single-vehicle crashes only applies to Alabama, Georgia, and South Carolina.
There are a wide variety of variables that influence a single-vehicle fatal crash in the three states. These include location, lane width, shoulder width and type, horizontal curve direction, crest vertical curves present, horizontal and vertical geometric
228

interactions, roadside hazard rating, traffic volume (ADT), driveway type, lighting conditions, and crash time. Individual single-vehicle models for the four states have similar influences, and consistently critical influences for fatal crashes in all four states include lane width, horizontal curve direction, and lighting conditions. For the Georgia only model, the use of safety restraints and lighting conditions were critical factors associated with single-vehicle fatal crashes.
For multiple-vehicle crashes the most common crash type associated with rural two-lane highways was a head-on crash. As a result, the second modeling approach predicts head-on fatal crashes versus the likelihood of alternative multiple-vehicle crash types. For this crash type model, the characteristics were consistent between all four states so the research team developed a composite model that extends to all four states. For the multiple-vehicle head-on crashes, the following observations were identified:
Multiple-vehicle head-on fatal crashes are influenced by similar characteristics in all four states, so the research team developed a four-state probability model.
There are a certain variables that influence a head-on fatal crash for all four states; however, there are considerably fewer variables than for the single-vehicle fatal crash condition. These influential factors include location, lane width, horizontal curve direction, traffic volume (ADT), and driveway frequency.
Individual head-on models for three of the four states provide meaningful results, but the research team was not able to develop an adequately performing head-on model for the state of South Carolina.
The three individual state models (Alabama, Georgia, and Mississippi) have similar influences including crash location and traffic volume (ADT). In each state, additional road characteristics can influence head-on crashes for that state.
For the Georgia only model, the use of safety restraints was a critical factor in the headon fatal crash predictive model.
229

The correct use of logit model results can be problematic if it is not clear how the models should be used and what limitations should be applied to use of the models, so this research also included a practical example that demonstrated how to implement the models to help make countermeasure decisions (see Chapter 7). Finally, this effort evaluated expert opinions recommendations for the Georgia crashes and developed crash modification factors based on their recommendations. In general, their recommendations centered based on the improvement of conditions at the edge of the road. Their most effective recommended countermeasure was improving clear zones. In addition they recommended improving geometric alignment, flattening side slopes, installing guard rail, and removing objects from the roadside. These independent recommendations closely align with the roadside hazard rating, the graded/paved shoulder condition, and the horizontal and vertical curve variables (each contributing to crashes as determined through the use of the statistical models). The corroboration of these two independent assessment techniques solidifies the accuracy of both the expert opinions and the statistical models.
230

10.0 Reference List
Abdel-Aty, M.A. (2003). "Analysis of Driver Injury Severity Levels at Multiple Locations Using Ordered Probit Models." Journal of Safety Research, Vol.34, No.5, pp. 597-603.
Abdel-Aty, M.A., Chen, C.L., and J.R. Schott (1998). "An Assessment Of The Effect Of Driver Age On Traffic Accident Involvement Using Log-Linear Models." Accident Analysis and Prevention, Vol. 30(6), pp. 851-861.
Abdel-Aty, M.A., and H. Abdelwahab (2004a). "Analysis and Prediction of Traffic Fatalities Resulting From Angle Collisions Including the Effect of Vehicles' Configuration and Compatibility." Accident Analysis and Prevention, Vol. 36(3), pp.457-469.
Abdel-Aty, M.A., and H. Abdelwahab (2004b). "Modeling Rear-End Collisions Including the Role of Driver's Visibility and Light Truck Vehicles Using a Nested Logit Structure." Accident Analysis and Prevention, Vol. 36(3), pp. 447-456.
Abdelwahab, H.T., and M.A. Abdel-Aty (2001). "Development Of Artificial Neural Network Models To Predict Driver Injury Severity In Traffic Accidents At Signalized Intersections." Transportation Research Record 1746. Transportation Research Board, National Research Council, Washington, D.C., pp. 6-13.
Agresti, A. (2002). "Categorical Data Analysis." 2nd edition, published by John Wiley & Sons, Inc. (ISBN: 978-0-471-36093-3).
Balkin, S., and J.K. Ord (2001). "Assessing the Impact of Speed-Limit Increases on Fatal Interstate Crashes." Journal of Transportation and Statistics, Vol. 4(1), pp. 1-26.
Bedard, M., Guyatt, G.H., Stones, M.J., and J.P. Hirdes (2002). "The Independent Contribution of Driver, Crash, And Vehicle Characteristics To Driver Fatalities." Accident Analysis and Prevention, Vol. 34 (6), pp. 717727.
Chang, L.Y., and F. Mannering (1998). "Predicting Vehicle Occupancies from Accident Data: An Accident Severity Approach." Transportation Research Record 1635. Transportation Research Board, National Research Council, Washington, D.C., pp. 93-104.
Chang, L.Y., and F. Mannering (1999). "Analysis of Injury Severity and Vehicle Occupancy in Truck- and Non-Truck-Involved Accidents." Accident Analysis and Prevention, Vol. 31(5), pp. 579592.
Derrig, A.A., Segui-Gomez, M., Abtahi, A., and L.L. Liu (2002). "The Effect of Population Safety Belt Usage Rates on Motor Vehicle-Related Fatalities." Accident Analysis and Prevention, Vol. 34(1), pp. 101-110.
231

Dissanayake, S., and J.J. Lu (2002). "Factors Influential in Making An Injury Severity Difference To Older Drivers Involved In Fixed Object-Passenger Car Crashes." Accident Analysis and Prevention, Vol. 34 (5), pp. 609618.
Dixon, K. K. (2005). "Final Summary Report Southeastern United States Fatal Crash Study." Report No. FHWA-GA-05-9815. Georgia Department of Transportation and Federal Highway Administration.
Duncan, C.S., Khattak, A.J., and F.M. Council (1998). " Applying the Ordered Probit Model To Injury Severity in Truck-Passenger Car Rear-End Collisions." Transportation Research Record 1635. Transportation Research Board, National Research Council, Washington, D.C., pp. 63-71.
Edwards, J.B. (1998). "The Relationship Between Road Accident Severity and Recorded Weather." Journal of Safety Research, Vol. 29(4), pp. 249262.
Eluru, N., and C. Bhat (2007). "A Joint Econometric Analysis of Seat Belt Use and CrashRelated Injury Severity." Accident Analysis and Prevention, Vol. 39(5), pp. 1037-1049.
Evans, L., and M.C. Frick (1993). "Mass Ratio and Relative Driver Fatality Risk in Two-Vehicle Crashes." Accident Analysis and Prevention, Vol. 25 (2), pp. 213224.
Evans, L. (1996). "Safety-Belt Effectiveness: The Influence of Crash Severity and Selective Recruitment." Accident Analysis and Prevention, Vol. 28(4), pp. 423433.
Farmer, C.M., Braver, E.R., and E.R. Mitter (1997). "Two-vehicle Side Impact Crashes: the Relationship of Vehicle and Crash Characteristics to Injury Severity." Accident Analysis and Prevention, Vol. 29(3), pp. 399406.
Farmer, C.M., and A.K. Lund (2002). "Rollover Risk of Cars and Light Trucks after Accounting for Driver and Environmental Factors." Accident Analysis and Prevention, Vol. 34(2), pp. 163173.
Foody, T.J., and M.D. Long (1974). "The Identification of Relationships Between Safety And Roadway Obstructions." Ohio Department of Transportation, Columbus, Ohio.
Griffin, L.I., and K.K. Mak (1989). "Benefits to Be Achieved from Widening Rural, Two-Lane, Farm-To-Market Roads in Texas." Presented at the 1989 Annual Meeting Of The Transportation Research Board, Washington, D.C.
Golob, T.F., and W.W. Recker (2003). "Relationships among Urban Freeway Accidents, Traffic Flow, Weather, and Lighting Conditions." Journal of Transportation Engineering, Vol. 129(4), pp. 342-353.
Haynes, R., Lake, I.R., Kingham, S., Sabel, C.E., Pearce, J., and R. Barnett (2008). "The Influence of Road Curvature on Fatal Crashes in New Zealand." Accident Analysis and Prevention, Vol. 40(3), pp. 843-850.
232

Heimbach, C.L., Hunter, W.W., and G.C. Chao (1974). "Paved Highway Shoulders and Accident Experience." Transportation Engineering Journal Proceedings, ASCE, Vol. 100, No. TE4, pp. 887-908.
Huelke, D. F., and C.P. Compton (1995). "The Effects of Seat Belts on Injury Severity of Front and Rear Seat Occupants in the Same Frontal Crash." Accident Analysis and Prevention, Vol. 27 (6), pp. 835-838.
Iowa Safety Management System Task Force on Speed Limits (1998). "Update Report on Speed Limits in Iowa." Web Accessed on July 24, 2008 at http://www.dot.state.ia.us/speedrpt.pdf.
Ivan, J.N., Pasupathy, R.K., and P. J. Ossenbruggen (1999). "Differences in Causality Factors for Single and Multi-Vehicle Crashes on Two-Lane Roads." Accident Analysis and Prevention, Vol. 31(6), pp. 695-704.
Joksch, H.C. (1993). "Velocity Change and Fatality Risk in a Crash A Rule of Thumb." Accident Analysis and Prevention, Vol. 25(1), pp. 103-104.
Jones, A.P., and S.H. Jorgensen (2003). "The Use of Multilevel Models for the Prediction of Road Accident Outcomes." Accident Analysis and Prevention, Vol. 35(1), pp. 59-61.
Jorgensen, R., and Associates (1978). "NCHRP Report No. 197: Cost and Safety Effectiveness of Highway Design Elements." Transportation Research Board, National Research Council, Washington, D.C.
Keall, M.D., Frith, W.J., and T.L. Patterson (2004). "The Influence of Alcohol, Age, and Number of Passengers on the Night-time risk of Driver Fatal Injury in New Zealand." Accident Analysis and Prevention, Vol. 36(1), pp. 49-61.
Khattak, A.J, Kantor, P., and F.M. Council. (1998), "Role of Adverse Weather in Key Crash Types on Limited-Access Roadways: Implications for Advance Weather Systems." Transportation Research Record 1621. Transportation Research Board, National Research Council, Washington, D.C., pp. 10-19.
Khattak, A.J., Pawlovich, M.D., Souleyrette, R.R., and S.L. Hallmark (2002). "Factors Related To More Severe Older Driver Traffic Crash Injuries." American Society of Civil Engineers, Journal of Transportation Engineering, Vol. 128 (3), pp. 243249.
Khattak, A.J., and M. Rocha. (2003). "Are SUVs `Supremely Unsafe Vehicles'?" Transportation Research Record 1840. Transportation Research Board, National Research Council, Washington, D.C., pp. 167-177.
Kim, D.G., Lee, Y., Washington, S., and K. Choi (2007). "Modeling Crash Outcome Probabilities at Rural Intersections: Application of Hierarchical Binomial Logistic Models." Accident Analysis and Prevention, Vol. 39(1), pp. 125-134.
233

Kim, K., Nitz, L., Richardson, J., and L. Li (1995). "Personal and Behavioral Predictors of Automobile Crash and Injury Severity." Accident Analysis and Prevention, Vol. 27(4), pp. 469481.
Krull, K.A., Khattak, A.J., and F.M. Council (2000). "Injury Effects of Rollovers and Events Sequence in Single-Vehicle Crashes." Transportation Research Record 1717, Transportation Research Board, National Research Council, Washington, D.C., pp. 46-54.
Kweon, Y.J., and K.M. Kockelman (2003). "Overall Injury Risk to Different Drivers: Combining Exposure, Frequency, and Severity Models." Accident Analysis and Prevention, Vol. 35(4), pp. 441450.
Lee, J., and F. Mannering (2002). "Impact of Roadside Features on the Frequency and Severity of Run-Off-Roadway Accidents: An Empirical Analysis." Accident Analysis and Prevention, Vol. 34(2), pp. 149-161.
Mao, Y. Zhang, J., Robbins, G., Clarke, K., Lam, M., and W. Pickett (1997). "Factors Affecting the Severity of Motor Vehicle Traffic Crashes Involving Young Drivers in Ontario." Injury Prevention, Vol. 3(3), pp. 183-189.
Melcher, D., Dixon, K., Washington, S., and C.H. Wu (2001). "Feasibility of `Subjective' Engineering Assessments of Road Safety Improvements: Bayesian Development." TRB, National Research Council, Washington, D.C. Journal of the Transportation Research Board, No. 1758. pp. 36-43.
Nakahara, S., Kawamura, T., Ichikawa, M., and S. Wakai (2006). "Mathematical Models Assuming Selective Recruitment Fitted to Data for Driver Mortality and Seat Belt Use in Japan." Accident Analysis and Prevention Vol. 38(6), pp. 175-184.
O'Donnell, C.J., and D.H. Connor (1996). "Predicting the Severity of Motor Vehicle Accident Injuries Using Models of Ordered Multiple Choice." Accident Analysis and Prevention, Vol. 28 (6), pp. 739-753.
Ossiander, E.M., and P. Cummings (2002). "Freeway Speed Limits and Traffic Fatalities in Washington State." Accident Analysis and Prevention, Vol. 34(1), pp. 13-18.
Persaud, B.N, Retting, R.A., and C.A. Lyon (2004). " Crash Reduction Following Installation of Centerline Rumble Strips on Rural Two-Lane Roads." Accident Analysis and Prevention, Vol. 36(6), pp. 1073-1079.
Preusser, D.F, Williams, A.F., and R.G. Ulmer (1995). "Analysis of Fatal Motorcycle Crashes: Crash Typing." Accident Analysis and Prevention, Vol. 27(6), pp. 845-851.
234

Ramsey, F.M., and D.W. Schafer (2002). "The Statistical Sleuth: a Course in Methods of Data Analysis." 2nd edition, published by Duxbury (ISBN 0-534-38670-9).
Renski, H, Khattak, A.J., and F.M. Council (1999). "Effect of Speed Limit Increases on Crash Injury Severity: Analysis of Single-Vehicle Crashes on North Carolina Interstate Highways." Transportation Research Record 1665, National Research Council, Washington, D.C., pp. 100108.
Retting, R.A., Williams, A.F., Preusser, D.F, and H.B. Weinstein (1994). " Classifying Urban Crashes For Countermeasure Development." Accident Analysis and Prevention, Vol. 27(3), pp. 283-294.
Robertson, L.S. (2002). "Bias in Estimates of Seat Belt Effectiveness (editorial)." Injury Prevention, Vol. 8(4), p. 263.
Ryan, G.A., Legge, M., and D. Rosman (1998). "Age related Changes in Drivers' Crash Risk and Crash Type." Accident Analysis and Prevention, Vol. 30(3), pp. 379-387.
Savolainen, P., and F. Mannering (2007). "Probabilistic Models of Motorcyclists' Injury Severities in Single- and Multi-vehicle Crashes." Accident Analysis and Prevention, Vol. 39(5), pp. 955-963.
Schiff, M.A., and P. Cummings (2004). "Comparison of Reporting of Seat Belt Use by Police and Crash Investigators: Variation in Agreement by Injury Severity." Accident Analysis and Prevention, Vol. 36(6), pp. 961965.
Shankar, V., Mannering, F., and W. Barfield (1996). "Statistical Analysis of Accident Severity on Rural Freeways." Accident Analysis and Prevention, Vol. 28(3), pp. 391-401.
Shannon, P., and A. Stanley (1976). "Pavement Width Standards For Rural, Two-Lane Highways." The Pacific Northwest Regional Commission, Idaho Transportation Department Bois, Idaho.
Shinar, D. (1998). "Speed and Crashes: A Controversial Topic and an Elusive Relationship. In Special Report 254: Managing Speed: Review of Current Practice for Setting and Enforcing Speed Limits." Transportation Research Board, National Research Council, Washington, D.C.
Shibata, A. and K. Fukuda (1994). "Risk Factors of Fatality in Motor Vehicle Traffic Accidents." Accident Analysis and Prevention, Vol. 26(3), pp. 391-397.
Srinivasan, K.K. (2002). "Injury Severity Analysis with Variables and Correlated Thresholds: Oreded Mixed Logit Formulation." Transportation Research Record 1784. Transportation Research Board, National Research Council, Washington, D.C., pp. 132-142.
Toy, E.L., and J.K. Hammitt (2003). "Safety Impacts of SUVs, Vans, and Pickup Trucks in Two-Vehicle Crashes." Risk Analysis, Vol. 23(4), pp. 641-650.
235

Traynor, L. (2005). "The Impact of Driver Alcohol Use on Crash Severity: A Crash Specific Analysis." Transportation Research. Part E: Logits & Transportation Review, Vol. 41(5), pp. 421-437.
Ulfarsson, G.F., and F. Mannering (2004). "Differences in Male and Female Injury Severities in Sport-Utility Vehicle, Minivan, Pickup and Passenger Car Accidents." Accident Analysis and Prevention, Vol. 36(2), pp. 135-147.
U.S. Department of Transportation and National Highway Traffic Safety Administration (2007). "2007 FARS Coding and Validation Manual." Web address: http://wwwnrd.nhtsa.dot.gov/Pubs/FARS07CVMAN.PDF (Accessed Jan 29, 2009).
Wang, X.K., and K.M. Kockelman (2005). "Use of Heteroscedastic Ordered Logit Model to Study Severity of Occupant Injury." Transportation Research Record 1908, National Research Council, Washington, D.C., pp. 195-204.
Washington, S., Metarko, J., Fomunung, I., Ross, R., Julian, F., and E. Moran (1999). "An Interregoinal Comparison: Fatal Crashes in the Southeastern and Non-southeastern United States: Preliminary Findings." Accident Analysis and Prevention, Vol. 31(1), pp. 135-146.
Washington, S., Dixon, K.K., White, D., and Chi-Hung E. Wu (2002). "Final Report: Investigation of Fatal Motor Vehicle Crashes on Two-Lane Rural Highways in Georgia." Report No. FHWA-GA-02-9905, Georgia Department of Transportation.
Washington, S., Lord, D., and B. Persaud (2009). The Use of Expert Panels in Highway Safety: A Critique. Transportation Research Record: Journal of the Transportation Research Board, National Academy Press, Washington D.C., In Press.
Washington, S., and J. Oh (2006). Bayesian Methodology Incorporating Expert Judgment for Ranking Countermeasure Effectiveness under Uncertainty: Example Applied to At Grade Railroad Crossings in Korea. Accident Analysis and Prevention. Elsevier. Vol. 38 No. 2, pp. 234-247.
Yau, K.K.W., Lo, H.P., and S.H.H. Fung (2006). "Multiple-Vehicle Traffic Accidents in Hong Kong." Accident Analysis and Prevention, Vol. 38(6), pp.11571161.
Zegeer, C.V., Hummer, J., Herf, L., Reinfurt, D., Herf, L., and W. Hunter (1988). "Safety Effects of Cross-Section Design for Two-Lane Roads." Transportation Research Record 1195. National Research Council, Washington, D.C., pp. 20-32.
Zegeer, C.V., Mayes, J., and R. Deen (1979). "Effect Of Lane And Shoulder Width On Accident Reduction On Rural, Two-Lane Roads." Transportation Research Record 806. National Research Council, Washington, D.C., pp. 33-43.
236

Zhang J., Lindsay, J., Clarke, K., Robbins, G., and Y. Mao (2000). "Factors Affecting the Severity of Motor Vehicle Traffic Crashes Involving Elderly Drivers in Ontario." Accident Analysis and Prevention, Vol. 32(1), pp. 117125.
237

This page blank intentionally 238

11.0 Appendix
239

Acronym Definitions

Acronym AADT ADT AL AMF BAC CMF FARS GA GDOT HSIP mph MS RHR SC SHSP SUV TOD US VMT vpd

Definition Average Annual Daily Traffic Average Daily Traffic Alabama Accident Modification Factor Blood Alcohol Concentration Crash Modification Factor Fatal Analysis Reporting System Georgia Georgia Department of Transportation Highway Safety Improvement Program Miles per hour Mississippi Roadside Hazard Rating South Carolina Strategic Highway Safety Plan Sports Utility Vehicle Time of Day United States Vehicle Miles Traveled Vehicles per day

240