Guidelines for Cost-Effective Safety Treatments of Roadside Ditches (2021)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

46 CHAPTER 4. ANALYSIS OF EXISTING CRASH DATA INTRODUCTION Run-off-road crashes and associated roadway data from several existing databases were evaluated and analyzed to achieve two objectives. The first objective was to analyze the data to identify any trends in the type and severity of ditch-related crashes and their relationship to ditch geometry, roadway characteristics, vehicle type, presence of appurtenances, and other relevant characteristics. The second objective was to extract potentially useful data for supporting the development and conduct of a BCA procedure to develop ditch design guidelines and treatment strategies. For the second objective, the focus was on obtaining vehicle roadside encroachment characteristics for ROR crashes and estimating injury severity and economic cost for ditch- initiated crashes. Encroachment characteristics data of interest included vehicle type, encroachment speed, encroachment angle, and driver control input (i.e., braking and steering input at the point of departure [POD]). Ditch-initiated injury severity data of interest included those results that could be used to develop impact-severity relationships for rollover and non- rollover crashes. Economic costs of interest included total annual cost and per-crash cost for ditch-initiated rollover and non-rollover crashes, Databases in Consideration The following databases were evaluated for their capability to address the two objectives above: FARS (6, 7), administered and maintained by NHTSA. NASS CDS (28, 29), administered and maintained by NHTSA. NASS General Estimates System (NASS GES) (30, 31), administered and maintained by NHTSA. NCHRP Project 17-22 database (32), developed by the University of Nebraska at Lincoln (UNL). Highway Safety Information System (HSIS) (33), a nine-state database administered by FHWA. Potentially useful variables were first identified from the technical manuals of these databases, and the associated data were subsequently acquired, processed, and analyzed. The FARS database contains a census of all crashes of motor vehicles traveling on public roadways in which a person died within 30 days of the crash. Data in NASS GES and NASS CDS come from a probability sample of police-reported motor vehicle crashes of all severity types, from property damage only (PDO) crashes to fatal crashes. NASS GES covers all vehicle types, while CDS focuses on light-duty vehicles. FARS, NASS CDS, and NASS GES are all collected annually. For over 30 years, the FARS has been the most referenced source for U.S. fatal crash data. Since 1988, the NASS GES has been an essential source for nonfatal injury crash data. The NCHRP Project 17-22 database contains 890 SVROR crashes, which are a small subset of the NASS CDS samples, mainly from 1997 to 2001 (with 16 cases for 2004). For each crash case, the 17-22 database contains more detailed roadway and roadside geometric data than those contained in the original NASS CDS database. These additional roadway and roadside geometric data were collected retrospectively from several supplemental field data collection

47 efforts undertaken by NCHRP 17-22, NCHRP 17-11, and FHWA Rollover Causation Study projects. The NCHRP 17-22 database also provides estimates of vehicle roadside encroachment and impact characteristics data—for example, encroachment speed and angle distributions at the POD—using crash reconstruction techniques. NCHRP Project 17-43, Long-Term Roadside Crash Data Collection Program, is an ongoing project aimed at updating and expanding the 17- 22 database. HSIS assembles crash and roadway inventory data from nine selected states for which the location of crashes can be linked to the roadway inventory. The data are stored on an annual basis. File structures, available variables, and number of years of data vary from state to state. Relative Strengths and Limitations The evaluation of each database started with a review of analytical, coding, and user’s manuals; relevant literature; and, when accessible, a small sample of crash cases extracted from the database. The relative strengths and limitations of these databases, in terms of their general capability to address the research objectives, were identified. Variables of potential interest to this study were also identified from each database. An overview of the relative strengths and limitations of the five databases considered are presented in Table 4.1. More detailed discussions about each database are provided when the analysis results are presented in the following sections of the report. Crash Data Screening Based on the needs of this project, the data analysis focused on crashes involving passenger cars, SUVs, vans, and pickup trucks. Crashes involving medium and large trucks, large buses, motorcycles, and unknown vehicle types were excluded. The design guidelines developed in this project are intended to reduce the frequency and severity of crashes involving light-duty vehicles. The analysis was further constrained to those crashes occurring on highways with a posted speed limit (PSL) between 45 mph and 75 mph (72 km/h and 121 km/h), inclusive. The rationale to target crashes occurring on highways with a PSL of 45 mph or higher is that lower- speed highways, including many urban and suburban 4-lane and 2-lane undivided highways, tend to have a rather limited ROW for considering an open ditch design, and therefore a closed drainage system is often utilized. Unless indicated otherwise, the targeted crashes involve roadside ditches. Crashes involving median ditches were excluded from the analysis. Design of median ditches for unprotected, divided highways requires different considerations, including the need to reduce the frequency of cross-median crashes.

48 Table 4.1. Relative strengths and limitations of databases to address research objectives. Database Strengths Limitations FARS • Nationwide data. • Census—no statistical sampling errors. • Good data quality (in terms of its timeliness, accuracy, and completeness). • All vehicle types (including cars, light trucks, SUVs, vans, and medium and large trucks). • Unlike other databases, ditch is coded as a separate fixed object struck by ROR vehicles. (In most other databases considered, ditches and other fixed objects, such as culverts, embankments, and even curbs, are coded as single types of roadside objects and are therefore not distinguishable.) • Fatal crashes only. • As in all nationwide crash databases, it focuses on crashes/vehicles/occupants, with very limited roadway and roadside data where crashes occurred. NASS GES • Nationwide data. • All vehicle types (unlike NASS CDS, which focuses on light- duty vehicles). • Sample size is about 10 times that of the NASS CDS data (and thus estimates have smaller sampling errors in general). • A probability sample of about 55,000 crash cases per year. • Subject to statistical sampling errors. • A biased sample by design; data need to be analyzed with specialized (survey- based) statistical method. • Police-level crash data (not in-depth* crash data). • Limited roadway and roadside data. • Ditches and culverts are reported as single types of roadside objects and are therefore not distinguishable. NASS CDS • Nationwide data. • An in-depth crash data.* • A probability sample of about 5,000 crash cases per year and is subject to statistical sampling errors. • A biased sample by design; data need to be analyzed with specialized (survey- based) statistical method. • Sample crash cases have a wide variation of sampling weights, some cases have very large sampling weights, and estimates are likely to have large statistical uncertainties.

49 Table 4.1. Relative strengths and limitations of databases to address research objectives (continued). Database Strengths Limitations NASS CDS • Nationwide data. • An in-depth crash data.* • Light-duty vehicles only—passenger cars, light trucks, and vans (i.e., no medium and heavy trucks). • Limited roadway and roadside data. • Ditches and culverts are reported as single types of roadside objects and are therefore not distinguishable. (However, narratives of the investigators and scene diagrams and photos are available for each crash and can be manually reviewed to distinguish the two types of objects.) NCHRP 17-22 • The only database with retrospectively field-collected data on roadway and roadside geometry. This is the only database in which relationships between crash severity and ditch configurations can potentially be examined. (Discussed later in report; the sample size is, however, way too small to develop anything meaningful.) • Provides more vehicle roadside encroachment and impact characteristics data (based on crash reconstructions) than those provided in NASS CDS, including encroachment speed and angle at the POD. • A small subsample of the NASS CDS data consisting of 890 cases mainly from 1997 to 2001 (with 16 cases from 2004). • Inherit all the statistical limitations associated with NASS CDS. • A biased sample by design; data need to be analyzed with specialized (survey- based) statistical method. • Most estimates will have significant sampling errors. • Only for single-vehicle ROR crashes occurring on highways with a PSL of 45 mph or higher. • Light-duty vehicles only—passenger cars, light trucks, and vans (i.e., no medium and heavy trucks). • Serious coding errors were identified in earlier versions. It is still being checked for coding errors. HSIS— Data from nine States (CA, MN, MI, ME, UT, NC, IL, WA, and OH) • Statewide Data. • Census—no statistical sampling errors. • All vehicle types. • Geographical limitation. • Out of the nine states in HSIS, only Washington State data allow ditches to be distinguished from other roadside features. Other states report ditches and other fixed objects, e.g., embankment, culvert, and curb, as a single object type. • Limited roadway and roadside data. *In-depth crash data are commonly referred to as data that are acquired from studies where trained investigators examine vehicles, injuries, and the crash site in some detail to reconstruct the event and to estimate impact speed and the resulting velocity change, either while the vehicles were still present or after their removal.

50 The remainder of this chapter is organized as follows: FARS. NASS GES. NASS CDS. NCHRP 17-22 database. HSIS. Severity and Cost of Ditch-Initiated Crashes. Each section discusses the findings from the evaluation and analysis of a selected database (or data system). The last section summarizes the results of an analysis that combined FARS and NASS GES data to: Understand the magnitude and severity of ditch-initiated crashes, which the design guidelines developed in this project are intended to influence. Obtain good cost estimates of ditch-initiated crashes that can be used to evaluate ditch configurations through a BCA. FARS This section documents the results of an analysis of the FARS database. Run-off-road crashes during 1991–2009 that satisfy the following two initial screening criteria were extracted from the database for analysis: (a) crashes that struck ditches as the FHE, and (b) crashes occurring on highways with a 45 mph or higher PSL. Focusing on crashes involving ditches as the FHE (i.e., the first property-damaging or injury-producing event in the crash) ensures that the injury to occupants and damage to vehicles did not occur before the vehicle struck the ditch. The rationale to target crashes occurring on highways with a PSL of 45 mph (72 km/h) or higher is that lower-speed highways often use closed rather than open drainage systems. As indicated in Table 4.1, FARS is a nationwide database, and thus there is no geographical bias as in state-based databases. It is a census of all fatal crashes in the United States, so there are no statistical sampling errors as in the NASS CDS and GES and 17-22 databases. The database covers only fatal crashes, the most severe type of crashes. The population distribution of crash severities involving ditches cannot be studied with the FARS database alone. It contains limited roadway and roadside data on the crash site. Similar to other nationwide vehicle crash databases, it is a crash-based (or crash-centered) database. That is, the focus is on recording the nature of crashes and the details of vehicles and occupants involved in the crash, not on roadway and roadside design or operational features. For each extracted crash case, variables of interest were retrieved from the database for analysis. The following types of analyses were conducted: Time trend in terms of number of fatal crashes per year involving ditches and other roadside features as the FHE from 1991 to 2009. Presence of appurtenances, such as culverts, trees, and utility poles, and top five or six MHEs after a vehicle impacted a ditch or culvert as the FHE. Roadway characteristics, including functional class, highway type, number of lanes, median type, PSL, horizontal alignment, vertical profile, and lighting condition. Vehicle characteristics, including vehicle body type and deployment of driver side airbag.

51 Driver characteristics, including driver alcohol use, occupant seat belt usage, and injury severity. Distribution of vehicle travel speeds. Note that the MHE is the single impact that causes the greatest trauma and damage in each crash. In this section, unless indicated otherwise, the ROR crashes involved both roadside and median ditches, and all vehicle types, including large buses, large trucks, and motorcycles. The last section, Severity and Cost of Ditch-Initiated Crashes, reports the results of an analysis focused on crashes involving only roadside ditches and light-duty vehicles, which the design guidelines of this project are intended to address. For comparison purposes, crashes involving culverts, guardrails, concrete traffic barriers, and utility poles as the FHE were analyzed in most cases. Highlights of the analysis are summarized below. Time Trend Figure 4.1 presents the number of fatal crashes involving ditches, culverts, guardrails, concrete traffic barriers, and utility poles as the FHE from 1991 to 2009. The number of fatal crashes involving a ditch as the FHE has generally trended up from 1991 to 2006, with a sharp increase in the 2004–2006 period. As is the case with other roadside features, some decreases were experienced in the years 2007–2009. When compared to other roadside features, the annual rate of increase in crash frequency involving ditches is quite significant over the period, which illustrates the relative importance of this project.

52 Figure 4.1. Number of fatal crashes involving ditches and other roadside features as FHE from 1991 to 2009. Presence of Appurtenances and Most Harmful Events Using FARS data from 2004 to 2008, Table 4.2 presents the top five or six MHEs after vehicles strike a ditch, culvert, guardrail, concrete traffic barrier, or utility pole as the FHE. The following observations can be made: After impacting a ditch as the FHE, the consequent events, including the FHE itself, that generated the most serious damage to the vehicle (i.e., the MHE) are rollover (53%), the ditch itself (16%), and a standing tree (15%). After striking a culvert as the FHE, the consequential MHE events are rollover (50%), the culvert itself (30%), and a standing tree (9%). After striking a guardrail as the FHE, the consequential MHE events are the guardrail itself (35%) and rollover (34%). After striking a concrete traffic barrier as the FHE, the consequential MHE events are the concrete traffic barrier itself (44%) and rollover (29%). After striking a utility pole as the FHE, the consequential MHE events are the utility pole itself (71%) and rollover (18%). Note that the observations above are generally consistent over the studied period from 1991 to 2009. Rollover is obviously the most consequential event after striking a ditch or culvert as the FHE. As far as other features, standing trees, utility poles, and culverts (ranked in order of importance) were the roadside features that produced the most damage to the vehicle after the encroaching vehicle struck a ditch as the FHE. Fatal Crashes Involving Several Roadside Features as First Harmful Events 0 200 400 600 800 1000 1200 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 Year N um be r o f F at al c ra sh es Ditch Culvert Guardrail Face Concrete Traf Barrier Utility Pole

53 Table 4.2. Top five or six MHEs (FARS: 2004 to 2008). MHE FHE Ditch Culvert Guardrail Face Concrete Traffic Barrier Utility Pole Ditch 16.3% Culvert 2.4% 29.5% Guardrail Face 35.2% Concrete Traffic Barrier 44.4% Utility Pole 4.1% 2.7% 70.6% Rollover 52.9% 50.3% 33.5% 28.5% 17.7% Standing Tree 15.4% 9.0% 6.4% 6.0% Fire/Explosion 2.7% 2.3% 2.5% 2.0% Motor Vehicle on Same Road 7.3% 8.7% 0.4% Motor Vehicle on Other Road 3.4% Bridge Pier 1.6% Immersion 0.5% All Others Combined 8.9% 5.8% 13.7% 12.5% 2.8% Highway Functional Class Table 4.3 shows the frequency and relative percentage of fatal crashes by highway functional class for crashes involving a ditch, culvert, guardrail, concrete traffic barrier, and utility pole as the FHE. Fatal crashes involving a ditch and culvert as the FHE occurred primarily on rural highways (about 86%), especially rural major collectors, local roads, and minor arterials. On the other hand, crashes involving concrete traffic barriers as the FHE occurred primarily on urban highways (about 80%), especially urban interstate and other principal arterials. Highway Type: Number of Lanes and Median Type Table 4.4 presents the frequency and relative percentage of fatal crashes by FHE and highway type, which is classified by number of lanes, median type, or whether it is a ramp. Fatal ditch and culvert crashes occurred predominantly on two-lane undivided highways (84% and 82%, respectively). The four-lane undivided highways without positive barriers accounted for another 8 to 9% of the crashes. On the other hand, crashes involving concrete traffic barriers as the FHE were mainly on 4-lane, 6-lane, and 8-lane highways with a protective barrier. Roadside Versus Median Ditches/Culverts Out of the 4,960 fatal crashes involving ditches as the FHE, about 212 cases involved median ditches; in 87 cases, the location of the ditch was unknown or not clear; and the rest of the 4,661 cases involved roadside ditches (about 95.6%). Out of the 2,183 fatal crashes involving culverts as the FHE, about 103 cases involved median culverts; in 47 cases, the location of the culvert was unknown or not clear; and the rest of 2,033 cases involved roadside culverts (about 95.2%).

54 Posted Speed Limit Table 4.5 presents the frequency and relative percentage of fatal crashes by FHE and PSL. The majority of the crashes involving ditches, culverts, and utility poles as the FHE occurred on highways with a PSL of 55 and 45 mph (89 and 72 km/h), while those involving guardrail and concrete traffic barriers occurred on highways with a PSL of 55 and 65 mph (89 and 105 km/h). Horizontal and Vertical Alignment Table 4.6 presents the frequency and relative percentage of fatal crashes by FHE and horizontal alignment. Over 40% of the crashes involving ditches, culverts, guardrail faces, and utility poles as the FHE occurred on curved sections. Table 4.7 presents frequency and relative percentage of fatal crashes by FHE and vertical alignment. About 30% of the crashes involving ditches and culverts as the FHE occurred on a downhill or uphill grade. Lighting Condition Table 4.8 presents the frequency and relative percentage of fatal crashes by FHE and the lighting condition under which crashes occurred. About 57% of the crashes involving ditches and culverts occurred during dark, dawn, or dusk conditions. Vehicle Body Type Table 4.9 presents the frequency and relative percentage of fatal crashes by FHE and vehicle body type. Light-duty vehicles, including passenger cars, utility vehicles, pickups, and vans, constitute about 86% and 90% of the crashes involving ditches and culverts as the FHE, respectively.

55 Table 4.3. Frequency and relative percentage of fatal crashes by highway functional class (FARS: 2004 to 2008). Highway Functional Class FHE Ditch Culvert Guardrail Face Concrete Traffic Barrier Utility Pole Rural Interstate 230 4.6% 84 3.8% 739 18.0% 155 11.4% 9 0.4% Rural Other Principal Arterial 404 8.1% 261 12.0% 581 14.2% 72 5.3% 172 7.3% Rural Minor Arterial 636 12.8% 397 18.2% 399 9.7% 24 1.8% 288 12.3% Rural Major Collector 1,373 27.7% 652 29.9% 378 9.2% 7 0.5% 533 22.7% Rural Minor Collector 502 10.1% 184 8.4% 102 2.5% 4 0.3% 186 7.9% Rural Local Road 1,092 22.0% 286 13.1% 125 3.0% 12 0.9% 355 15.1% Rural Unknown 50 1.0% 9 0.4% 6 0.1% 1 0.1% 10 0.4% Urban Interstate 118 2.4% 38 1.7% 889 21.7% 635 46.6% 63 2.7% Urban Other Principal Arterial 64 1.3% 25 1.1% 498 12.1% 256 18.8% 53 2.3% Urban Minor Arterial 105 2.1% 62 2.8% 180 4.4% 114 8.4% 232 9.9% Urban Major Collector 145 2.9% 76 3.5% 102 2.5% 33 2.4% 195 8.3% Urban Minor Collector 84 1.7% 41 1.9% 32 0.8% 2 0.1% 87 3.7% Urban Local Road 130 2.6% 47 2.2% 50 1.2% 35 2.6% 146 6.2% Urban Unknown 1 0.0% 0 0.0% 3 0.1% 1 0.1% 3 0.1% Unknown 27 0.5% 21 1.0% 20 0.5% 12 0.9% 13 0.6% All 4,960 100.0% 2,183 100.0% 4,104 100.0% 1,363 100.0% 2,345 100.0%

56 Table 4.4. Frequency and relative percentage of fatal crashes by FHE and highway type (FARS: 2004 to 2008). Highway Type FHE Ditch Culvert Guardrail Face Concrete Traffic Barrier Utility Pole 2-Ln Undivided 4,169 84.1% 1,788 81.9% 1,321 32.2% 47 3.4% 1,783 76.0% 2-Ln w/ 2-Way Continuous Left Turn Ln 3 0.1% 5 0.2% 4 0.1% 0 0.0% 10 0.4% 4-Ln Undivided 31 0.6% 34 1.6% 43 1.0% 7 0.5% 79 3.4% 4-Ln Divided w/o Barrier 417 8.4% 194 8.9% 651 15.9% 98 7.2% 150 6.4% 4-Ln Divided w/ Barrier 59 1.2% 23 1.1% 517 12.6% 273 20.0% 35 1.5% 4-Ln w/ 2-way Continuous Left Turn Ln 4 0.1% 4 0.2% 2 0.0% 1 0.1% 27 1.2% 6-Ln Divided w/o Barrier 38 0.8% 12 0.5% 172 4.2% 46 3.4% 44 1.9% 6-Ln Divided w/ Barrier 19 0.4% 9 0.4% 374 9.1% 331 24.3% 38 1.6% 8-Ln Divided w/o Barrier 103 2.1% 45 2.1% 177 4.3% 30 2.2% 44 1.9% 8-Ln Divided w/ Barrier 31 0.6% 28 1.3% 345 8.4% 204 15.0% 18 0.8% 10-Ln Divided w/o Barrier 3 0.1% 3 0.1% 11 0.3% 8 0.6% 4 0.2% 10-Ln Divided w/ Barrier 3 0.1% 1 0.0% 52 1.3% 85 6.2% 6 0.3% 12-Ln Divided w/o Barrier 2 0.0% 1 0.0% 17 0.4% 6 0.4% 0 0.0% 12-Ln Divided w/ Barrier 10 0.2% 0 0.0% 55 1.3% 51 3.7% 3 0.1% 14-Ln Divided w/o Barrier 0 0.0% 0 0.0% 0 0.0% 1 0.1% 0 0.0% 14-Ln Divided w/ Barrier 2 0.0% 2 0.1% 24 0.6% 29 2.1% 4 0.2% 1-Ln Entrance/Exit Ramp 26 0.5% 11 0.5% 181 4.4% 70 5.1% 14 0.6% 2-Ln Entrance/Exit Ramp 1 0.0% 1 0.0% 51 1.2% 41 3.0% 3 0.1% Others 39 0.8% 22 1.0% 107 2.6% 35 2.6% 83 3.5% All 4,960 100.0% 2,183 100.0% 4,104 100.0% 1,363 100.0% 2,345 100.0%

57 Table 4.5. Frequency and relative percentage of fatal crashes by FHE and posted speed limit (FARS: 2004 to 2008). Posted Speed Limit FHE Ditch Culvert Guardrail Face Concrete Traffic Barrier Utility Pole 45 mph 930 18.8% 438 20.1% 408 9.9% 134 9.8% 762 32.5% 50 mph 202 4.1% 97 4.4% 222 5.4% 90 6.6% 254 10.8% 55 mph 3,052 61.5% 1,184 54.2% 1,492 36.4% 382 28.0% 1,044 44.5% 60 mph 127 2.6% 82 3.8% 274 6.7% 190 13.9% 63 2.7% 65 mph 316 6.4% 228 10.4% 1,027 25.0% 421 30.9% 73 3.1% 70 mph 215 4.3% 121 5.5% 490 11.9% 82 6.0% 15 0.6% 75 mph 14 0.3% 9 0.4% 93 2.3% 22 1.6% 1 0.0% Others 1 0.0% 0 0.0% 2 0.0% 0 0.0% 2 0.1% Unknown 103 2.1% 24 1.1% 96 2.3% 42 3.1% 131 5.6% All 4,960 100.0% 2,183 100.0% 4,104 100.0% 1,363 100.0% 2,345 100.0% Table 4.6. Frequency and relative percentage of fatal crashes by FHE and horizontal alignment (FARS: 2004 to 2008). Horizontal Alignment FHE Ditch Culvert Guardrail Face Concrete Traffic Barrier Utility Pole Straight 2,652 53.5% 1,294 59.3% 2,273 55.4% 889 65.2% 1,385 59.1% Curved 2,301 46.4% 886 40.6% 1,821 44.4% 468 34.3% 944 40.2 % Unknown 7 0.1% 3 0.1% 10 0.2% 6 0.4% 16 0.7% All 4,960 100.0% 2,183 100.0% 4,104 100.0% 1,363 100.0% 2,345 100.0% Table 4.7. Frequency and relative percentage of fatal crashes by FHE and vertical profile of highways (FARS: 2004–2008). Vertical Profile FHE Ditch Culvert Guardrail Face Concrete Traffic Barrier Utility Pole Level 3,334 67.2% 1,426 65.3% 2,469 60.2% 934 68.5% 1,678 71.6% Grade 1,418 28.6% 691 31.7% 1,488 36.3% 366 26.9% 575 24.5% Hill Crest 119 2.4% 45 2.1% 74 1.8% 28 2.1% 59 2.5% Sag 32 0.6% 6 0.3% 18 0.4% 3 0.2% 5 0.2% Unknown 57 1.1% 15 0.7% 55 1.3% 32 2.3% 28 1.2% All 4,960 100.0% 2,183 100.0% 4,104 100.0% 1,363 100.0% 2,345 100.0%

58 Table 4.8. Frequency and relative percentage of fatal crashes by FHE and lighting condition (FARS: 2004–2008). Lighting Condition FHE Ditch Culvert Guardrail Face Concrete Traffic Barrier Utility Pole Daylight 2,112 42.6% 938 43.0% 1,879 45.8% 502 36.8% 872 37.2% Dark 2,433 49.1% 1,062 48.6% 1,455 35.5% 322 23.6% 1,024 43.7% Dark But Lighted 163 3.3% 86 3.9% 577 14.1% 492 36.1% 360 15.4% Dawn 98 2.0% 82 3.8% 102 2.5% 25 1.8% 38 1.6% Dusk 123 2.5% 52 2.4% 67 1.6% 16 1.2% 41 1.7% Unknown 31 0.6% 40 1.8% 24 0.6% 6 0.4% 10 0.4% All 4,960 100.0% 2,183 100.0% 4,104 100.0% 1,363 100.0% 2,345 100.0% Table 4.9. Frequency and relative percentage of fatal crashes by FHE and vehicle body type (FARS: 2004–2008). Vehicle Body Type FHE Ditch Culvert Guardrail Face Concrete Traffic Barrier Utility Pole Passenger Cars 2,171 43.8% 1,055 48.3% 1,468 35.8% 587 43.1% 1,398 59.6% Utility Vehicles 731 14.7% 300 13.7% 652 15.9% 230 16.9% 233 9.9% Pickups 1,130 22.8% 523 24.0% 693 16.9% 166 12.2% 414 17.7% Vans 217 4.4% 76 3.5% 183 4.5% 68 5.0% 83 3.5% Medium Trucks 24 0.5% 8 0.4% 27 0.7% 11 0.8% 4 0.2% Heavy Trucks 93 1.9% 36 1.6% 269 6.6% 68 5.0% 20 0.9% Motorcycles 521 10.5% 162 7.4% 778 19.0% 216 15.8% 165 7.0% Buses 1 0.0% 2 0.1% 5 0.1% 7 0.5% 0 0.0% Others 72 1.5% 21 1.0% 29 0.7% 10 0.7% 28 1.2% All 4,960 100.0% 2,183 100.0% 4,104 100.0% 1,363 100.0% 2,345 100.0% Deployment of Driver Side Airbags Table 4.10 presents the frequency and relative percentage of fatal crashes by FHE and deployment of driver side airbags. About 31% and 39% of the crashes involving ditches and culverts, respectively, as the FHE were known to have the driver side airbag deployed during the crash.

59 Table 4.10. Frequency and relative percentage of fatal crashes by FHE and the deployment of driver side air bags (FARS: 2004–2008). Driver Side Air Bag FHE Ditch Culvert Guardrail Face Concrete Traffic Barrier Utility Pole Deployed 1,537 31.0% 845 38.7% 1,155 28.1% 463 34.0% 865 36.9% Equipped but Not Deployed 1,271 25.6% 432 19.8% 816 19.9% 269 19.7% 436 18.6% Not Equipped or Deployment Unknown 2,152 43.4% 906 41.5% 2,133 52.0% 631 46.3% 1,044 44.5% All 4,960 100.0% 2,183 100.0% 4,104 100.0% 1,363 100.0% 2,345 100.0% Driver Alcohol Use Table 4.11 presents the frequency and relative percentage of fatal crashes by FHE and driver drinking status. Over 48% of the crashes involving ditches as the FHE involved drunk drivers. This percentage is quite high relative to the overall rate of alcohol-impaired drivers involved in fatal crashes in general, which was about 32% in 2008–2009. Table 4.11. Frequency and relative percentage of fatal crashes by FHE and driver alcohol use (FARS: 2004–2008). Driver Drinking Status FHE Ditch Culvert Guardrail Face Concrete Traffic Barrier Utility Pole Involving Drunk Drivers 2,397 48.3% 998 45.7% 1,596 38.9% 560 41.1% 1,213 51.7% Not Involving Drunk Drivers 2,563 51.7% 1,185 54.3% 2,508 61.1% 803 58.9% 1,132 48.3% All 4,960 100.0% 2,183 100.0% 4,104 100.0% 1,363 100.0% 2,345 100.0% Seat Belt Use Table 4.12 shows the frequency and relative percentage of fatal crashes by FHE and seat belt use. Close to 71% of the fatal ditch crashes involved at least one occupant in the crash who did not wear a seat belt and was seriously injured. Seriously injured was defined as those occupants who sustained a fatal, incapacitating, or non-incapacitating injury (i.e., KAB crashes).

60 Table 4.12. Frequency and relative percentage of fatal crashes by FHE and seat belt use. Seat Belt Usage and Injury Level FHE Ditch Culvert Guardrail Face Concrete Traffic Barrier Utility Pole No Seriously Injured Occupants w/o Wearing Seat Belt 1,221 24.6% 588 26.9% 1,471 35.8% 475 34.8% 785 33.5% At Least One Occupant Seriously Injured w/o Wearing Seat Belt 3,514 70.8% 1,469 67.3% 2,340 57.0% 780 57.2% 1,389 59.2% Seat Belt Usage and/or Injury Level Unknown 225 4.5% 126 5.8% 293 7.1% 108 7.9% 171 7.3% All 4,960 100.0% 2,183 100.0% 4,104 100.0% 1,363 100.0% 2,345 100.0% Note: Seriously injured was defined as those occupants who sustained a fatal, incapacitating, or non-incapacitating injury (i.e., KAB crashes). Travel Speed Distribution by Posted Speed Limit Table 4.13 shows some characteristics (e.g., mean, standard deviation, and coefficient of skewness) of the travel speed distribution of SVROR fatal crashes by PSL from 45 to 75 mph. Qualified fatal crashes were grouped by the PSL of the highways where crashes occurred. For each PSL, gamma and normal distribution functions were used to fit the travel speed distribution. The model with a statistically better goodness-of-fit was selected. The selected probability density function for each PSL is indicated in the table. The number of crash cases used to estimate the parameters in the density function is also presented. These fitted travel speed distributions can potentially be used to estimate the upper bounds of encroachment speed conditions in a sensitivity study of a BCA. As indicated in the user’s manual, travel speed in FARS is “an estimate of the speed of the vehicle involved in the crash” and “is often an estimate of the actual speed by the investigating officers.” It also points out that it is an estimate of the travel speed after the crash, and it is often a judgment rather than a measurement. Crashes occurring in the median area of divided highways were excluded. In addition, travel speed was not provided by the investigating officer in about 50% of the crash cases.

61 Table 4.13. Some characteristics of travel speed distribution and fitted probability density function by posted speed limit (FARS: 2004–2009; only SVROR crashes on the roadside area). Speed Limit (mph) Sample Distribution Fitted Probability Density Function Number of Fatal Crashes Mean Std. Dev. Coeff. of Skewness 45 60.8 (±0.3) 16.2 0.62 Gamma(14.12932,0.232318) 3,742 50 62.0 (±0.6) 17.3 0.81 Gamma(12.90426,0.208204) 756 55 62.9 (±0.2) 14.2 0.77 Gamma(19.73254, 0.313549) 8,411 60 65.4 (±0.5) 13.6 0.86 Gamma(22.97505,0.351359) 658 65 70.3 (±0.3) 12.3 0.78 Gamma(32.8932,0.46770) 2,327 70 72.6 (±0.3) 9.4 0.91 Gamma(59.85556,0.824696) 1,159 75 72.7 (±0.6) 10.0 −0.23 Normal(72.7, (10.0)2) 262 All 64.3 (±0.1) 14.7 0.53 Gamma(19.14623,0.297693) 17,315 Notes: (1) A small number of crash cases with an estimated travel speed that was less than one-half of the PSL of the highway where the crash occurred were eliminated from the analysis. If the reporting and coding of the travel speeds were correct, these vehicles were likely to be operating abnormally. (2) Travel speed distribution is assumed to be either gamma or normal distributed. The particular formulation of the gamma density function used is as follows: X ~ Gamma(𝛼𝛼,𝛽𝛽) , and     ≥ = −− otherwise, ,0 0, )()( 1 xex Γxf x X βα α α β where βα , = constants, and 0,0 >> βα . (3) The ± values in the parentheses under the estimated means are standard errors of the estimates above.

62 NASS GES NASS GES data are obtained from a probability sample of police-reported traffic crashes covering all vehicle types and injury severity levels. About 55,000 crash cases per year have been sampled in recent years (from an estimated crash population of about 5.8 million per year). The GES survey is designed to sample from a predetermined set of geographical areas and crash strata. GES divides the United States into 1,195 geographical areas, called primary sampling units (PSUs). It is a multistage survey with a sample of 60 PSUs across the United States in the first stage, from which approximately 400 police agencies are surveyed in the second stage. The third and final stage is a sampling of crashes from several groups of crashes, called crash strata, which are determined by severity of the injured, type of vehicles involved, and tow status of the vehicles involved. Crashes have been grouped into six strata since 2002. Before 2002, 4 strata were used from 1990 to 2001, and three strata were used before 1990. Within each of these crash strata, a systematic sample of crashes is selected. That is, every k-th crash listed in the stratum is selected, where the integer k is dependent on a predetermined sampling ratio that varies by stratum. The design is a composite design typically described as a stratified multistage cluster- sampling design. Under the design, some subpopulations are intentionally sampled more intensively than others to reduce data collection cost and achieve several other objectives. The resulting sample can, however, be very different from the overall population from which it is drawn and about which researchers wish to make inferences. To begin with, it is a biased sample by design in which crashes are selected with unequal probabilities. Moreover, sample cases are not statistically independent because of the sampling from geographical clusters. Conventional statistical methods, which are based on simple random sampling (SRS) or an equal selection (inclusion) probability sampling assumption, are not appropriate for analyzing such data. As is discussed shortly, specialized survey-based statistical methods are needed. For GES and CDS, NASS statisticians have developed sampling (or case) weights to correct for selection bias as well as for biases incurred during implementation. The sample weight is the number of crashes in the target population that the sampled crash represents. These sampling weights make it possible for users to compute unbiased estimates of population statistics for variables of interest and to develop estimates that are representative of the crash population in the country. In addition, these sampling weights and the survey design variables provided in the database, such as the PSU and crash stratum associated with each sample case, allow the uncertainty of an estimate due to sampling to be statistically assessed. The results of a statistical analysis of sampled data from complex surveys such as GES and CDS are meaningless if sampling weights are not properly accounted for in the analysis. In addition, the uncertainty of an estimate cannot be properly determined if the clustering and stratified nature of the data are not factored into consideration. This element has been discussed extensively in the statistical literature since the 1950s. Example textbooks on sampling survey design and data analysis include Skinner et al. (34), Lohr (35), Korn and Graubard (36), Chambers and Skinner (37), and Bethlehem (38). Considering sampling weights and the clustering and stratified nature of the data in a statistical analysis requires specialized statistical methods. Several statistical software tools have built-in survey-based data analysis modules for analyzing such data, including STATA, SAS/STAT, R Project, SUDAAN, and WesVar.

63 Because the estimates produced from the GES and CDS are based on a probability sample of crashes, not a census of all crashes (as is the case with FARS), the estimates are subject to sampling errors and may differ from the true values. When available and deemed appropriate from a subject matter standpoint, the use of a larger sample size of crashes, such as using more years of data, is preferable in most analyses since the estimates so produced will have smaller sampling errors. In addition, since the sample size of GES per year is about 10 times that of CDS, GES data are preferred if the variables of interest are available in both GES and CDS. The statistical uncertainty of estimates associated with small domains in the population, such as estimates of totals, means, proportions, or probability distributions associated with fatal crashes, tends to be large when estimated from a probability sample such as GES and CDS. If small domain estimates from census data are available, such as the number of fatal crashes from FARS, they are preferred over estimates from a probability sample. As mentioned, the section entitled Severity and Cost of Ditch-Initiated Crashes in this chapter will report the results of a study that combined FARS and NASS GES data to understand the magnitude and severity of ditch-initiated crashes and obtain good cost estimates of ditch- initiated crashes. In that analysis, ditch-initiated crashes were defined as those SVROR crashes that involved roadside ditches as the FHE and in which the MHE was either the ditch or a rollover. The same definition was used for culvert-initiated crashes; that is, culverts were the FHE, and the MHE was either the culvert or a rollover. An estimate of 4,690 fatal ditch/culvert- initiated crashes was obtained from FARS for 2004 to 2009. Based on the same definitions and selection criteria, an estimate of 3,781 fatal crashes were obtained for the same period from GES, which underestimated the number of fatal crashes by 19%. Thus, the discrepancy between small domain estimates from a probability sample and true values can be significant. This small domain estimation problem is expected to be more pronounced for the CDS data given that its sample size is only one-tenth that of GES, and it will be even worse for the NCHRP 17-22 data, which is a subsample of CDS. As mentioned, GES and CDS contain limited roadway and roadside data related to the crash site. As in other nationwide vehicle crash databases, it is a crash-based (or crash-centered) database. In other words, the focus is on recording the nature of crashes and the details of vehicles and occupants involved in the crash, not on roadway and roadside design features. Furthermore, when coding fixed objects struck by the involved vehicles, GES and CDS have used the same code for ditch and culvert. It is therefore not possible to tell from the coded data whether a vehicle involved in a collision with “ditches/culverts” actually struck a ditch or a culvert. FARS data, on the other hand, allow this distinction to be made. With this limitation in mind, some of the analysis results using GES data are presented below. Note that the crash data screening criteria described in the introduction of this chapter were applied prior to the analysis. Frequency and Severity Distributions of Ditch-Initiated Crashes When estimating crash costs, an important piece of data to have is how many fatalities and injuries of various severities are involved on average in each crash. Using 6 years of FARS and GES data (2004–2009), the average numbers of fatalities and injuries involved in each ditch/culvert-initiated crash were estimated. Table 4.14 presents these estimates for each police- reported crash severity level. Given that the police-reported severity level is classified according to the known maximum injury severity incurred in a crash, a fatal crash should involve at least one fatality and may include other occupants with varying degrees of injury. Since the SV crash

64 is in consideration, only one vehicle is damaged. As shown in Table 4.14, on average, a fatal ditch/culvert-initiated crash involved 1.056 fatalities, 0.192 incapacitating injuries, 0.216 non- incapacitating injuries, 0.095 possible injuries, and one damaged vehicle. An incapacitating injury crash, on the other hand, involved 1.237 incapacitating injuries, 0.078 non-incapacitating injuries, 0.031 possible injuries, and one damaged vehicle. Table 4.14. Average numbers of injuries and damages involved in one ditch/culvert- initiated crash (FARS and GES databases, 2004–2009). Occupant Injury Severity Police-Reported Injury Severity of a Crash Fatal (K) Incapacitating Injury (A) Non- incapacitating Injury (B) Possible Injury (C) Property Damage Only (PDO) Fatalities 1.056 0 0 0 0 Incapacitating Injuries 0.192 1.237 0 0 0 Non- incapacitating Injuries 0.216 0.078 1.194 0 0 Possible Injuries 0.095 0.031 0.107 1.173 0 PDO (Vehicle) 1 1 1 1 1 Table 4.15 shows the estimated numbers of crashes by police-reported severity level and their distribution in percentage: Column 3 for rollover crashes, Column 5 for non-rollover crashes, and Column 7 for all crashes. Over 75% of the non-rollover crashes did not incur any injury, while only about 42% of the rollover crashes did not have injuries. About 3.03% of the rollover crashes resulted in fatalities, while only about 0.28% of the non-rollover crashes did. For each severity level, Table 4.15 also shows the portion of rollover and non-rollover crashes (in parentheses). Overall, after striking a ditch/culvert as the FHE, about 34% of the ditch-initiated crashes rolled over, and 66% did not. However, the rollover percentage varies significantly between injury severity levels. For example, about 85% of the fatal crashes rolled over in the end after striking a ditch/culvert as the FHE, while about 22% of the PDO crashes rolled over. Posted Speed Limit Table 4.16 shows the estimated distribution by PSL for SVROR crashes that involved a roadside ditch/culvert as the FHE. The estimate was generated using GES data from 2002 to 2009. The majority of the crashes (about 81%) occurred on highways with a PSL of 45 and 55 mph (72 and 89 km/h), which is consistent with the distribution for the fatal crashes in Table 4.5.

65 Table 4.15. Estimated numbers and distribution of crashes by injury severity (FARS and GES databases, 2004–2009). Police-Reported Injury Severity Rollover Crashes Non-Rollover Crashes All Crashes Frequency Severity % Frequency Severity % Frequency Severity % K: Fatal 3,966 (84.6%) 3.03 724 (15.4%) 0.28 4,690 (100%) 1.20 A: Incapacitating Injury 18,531 (59.0%) 14.14 12,880 (41.0%) 4.97 31,411 (100%) 8.05 B: Non- incapacitating Injury 28,279 (55.6%) 21.58 22,554 (44.4%) 8.70 50,833 (100%) 13.02 C: Possible Injury 25,393 (47.4%) 19.37 28,166 (52.6%) 10.87 53,559 (100%) 13.72 PDO 54,898 (22.0%) 41.88 194,885 (78.0%) 75.18 249,783 (100%) 64.00 All 131,067 (33.6%) 100 259,209 (66.4%) 100 390,276 (100%) 100 Notes: An additional criterion used to select crashes for generating this table is that the rollover status of the vehicle was available in the database. The percentage values in parentheses under the estimated frequency are portions of rollover and non-rollover crashes. Table 4.16. Distribution of SVROR crashes that involved a roadside ditch/culvert as the FHE by posted speed limit. Posted Speed Limit (mph) Relative Frequency (%) Sample Size Number of Crashes Represented 45 23.9 (±3.2) 1,597 143,840 50 4.9 (±0.8) 298 29,375 55 57.1 (±4.4) 3,186 344,093 60 1.2 (±0.4) 113 7,439 65 6.3 (±0.7) 383 38,064 70 6.3 (±1.6) 426 38,187 75 0.3 (±0.2) 19 1,960 All 100 6,022 602,958 Travel Speed Distribution by Posted Speed Limit Table 4.17 shows some characteristics (e.g., mean, standard deviation, and coefficient of skewness of the travel speed distribution) of SVROR crashes by PSL from 45 to 75 mph. Only those crashes involving a vehicle running off onto the roadside were selected. Qualified crashes were grouped by the PSL of the highways where the crashes occurred. For each PSL, gamma and normal distribution functions were used to fit the travel speed distribution. The model with a statistically better goodness-of-fit was selected. The selected probability density function for each PSL is indicated in the table. Available sample size for estimating the parameters in the density function, as well as the number of real-world crashes represented by the sample, is also

66 presented. These fitted travel speed distributions are used in conjunction with other data to estimate encroachment speed distributions. As in FARS, travel speed is an estimate of the speed of the vehicle involved in the crash and is often an estimate of the actual speed by the investigating officers. The GES also points out that travel speed is an estimate of the travel speed after the crash, and it is often a judgment rather than a measurement. Crashes occurring in the median area of divided highways were excluded. Table 4.17. Some characteristics of travel speed distribution and fitted probability density function by posted speed limit (GES: 2002–2009; only SVROR crashes on the roadside area). Speed Limit (mph) Sample Distribution (Weighted) Fitted Probability Density Function Sample Size Number of Crashes Represented Mean Std. Dev. Coeff. of Skewness 45 45.4 (±0.7) 11.1 0.86 Gamma(16.69045, 0.367972) 4,609 377,517 50 48.2 (±0.8) 10.7 1.46 Gamma(20.39273,0.422843) 689 57,379 55 49.0 (±1.0) 10.6 0.45 Gamma(21.38918,0.4366) 7,445 721,707 60 58.2 (±1.1) 10.8 0.94 Gamma(28.80123,0.494663) 334 18,523 65 59.5 (±0.5) 10.4 −0.25 Normal(59.5, (10.4)2) 2,375 210,730 70 67.6 (±0.6) 8.7 −0.67 Normal(67.6, (8.7)2) 1,167 72,446 75 71.8 (±1.5) 8.7 −0.70 Normal(71.8, (8.7)2) 152 13,100 All 50.8 (±0.9) 12.3 0.42 Gamma(17.01982, 0.335204) 16,771 1,471,402 Note: See Table 4.13. NASS CDS The crashes investigated in the NASS CDS are a probability sample of police-reported crashes in the United States. The data have been collected and published annually since 1979. A CDS crash must fulfill the following requirements: (a) must be police-reported, (b) must involve a harmful event (property damage and/or personal injury) resulting from a crash, and (c) must involve at least one towed passenger car or light truck or van in transport on a trafficway. Every crash that meets these conditions has a chance of being selected. About 5,000 crash cases per year have been sampled in recent years. As in the GES, the CDS survey is designed to sample from a predetermined set of geographical areas and crash strata. Just like the GES, the CDS divides the United States into 1,195 geographical areas called PSUs. It is a multistage survey with a sample of 24 PSUs across the United States in the first stage, from which a number of police jurisdictions that process reports of crashes occurring within each PSU’s boundaries are surveyed in the second stage. Note that 27 PSUs were selected

67 from 2002 to 2007. The final stage is a sampling of crashes from 10 crash strata that are determined by severity of the injured, type and model year of the vehicle involved, disposition and hospitalization of the injured, and tow status of the vehicles involved. As in GES, the design is typically described as a stratified multistage cluster-sampling design. As described, under such design, some subpopulations are intentionally sampled more intensively than others. It is a biased sample by design in which crashes with unequal probabilities are selected. Moreover, sample cases are not statistically independent because of the sampling from geographical clusters. Conventional statistical methods, which are based on SRS or equal selection (inclusion) probability sampling assumption, are not appropriate for analyzing such data. All the discussions in the last section regarding GES sampling limitations and the need to use specialized survey-based statistical methods to analyze the data apply to the CDS data as well. For example, the results of any statistical analysis of sampled data from a complex survey, such as the NASS GES and CDS, are meaningless if sampling weights are not properly accounted for in the analysis. Also, the uncertainty of an estimate cannot be properly determined if the clustering and stratified nature of the data are not factored into consideration. The large statistical uncertainty problem discussed in the last section regarding the use of GES data to conduct small domain estimations is expected to be even more pronounced when using the CDS data because of their smaller sample size and statistical design issue, which is discussed next. The CDS sampling design produces sample crash cases with a wide variation of sampling weights. Some cases have very large sampling weights, which results in estimates with large statistical uncertainties. Using CDS 2008 data as an example, there are 5,160 regular crash cases, with sampling weights varying from 1.216 to 73,136.18 and an average weight of 408.4. Thus, some sample cases represent close to one case in the real world, while some represent over 73,000 real-world cases. In contrast, GES is more efficient in its sampling design. For the 55,946 sampled cases in the 2008 GES, the sampling weights vary from 1.664 to 1,935.849. The statistical consequence is that a large disparity in sampling weights usually results in large variances in estimated sample statistics. Using sample statistics calculated from such data to infer population parameters can be highly unreliable and should be used with extra caution. As in GES, using more years of data to reduce the sampling errors of estimates is preferable when appropriate and feasible. Unlike GES data, which are police-level data, CDS data are considered in-depth crash data in that crashes are acquired and studied by trained investigators who examine the vehicles, injuries, and crash site in some detail to reconstruct the event and to estimate impact speed and the resulting velocity change (delta-V), either while the vehicles are still present or after their removal. As discussed, GES and CDS contain limited roadway and roadside data on the crash site. CDS do have some roadway data, including median type, number of travel lanes, horizontal alignment (straight, curve right, curve left), and vertical alignment (level, uphill grade, hill crest, downhill grade, sag). Very importantly, however, the information contained in the scene diagrams, narratives, and scene and vehicle photographs prepared by the trained NASS investigators are often useful sources of data to perform some quality checks on the coded data and obtain additional data (e.g., median type and lane configuration).

68 Like GES, when coding fixed objects struck by the involved vehicles, CDS uses the same code for ditch and culvert. It is therefore not possible to tell from the coded data whether a vehicle involved in a collision with “ditches/culverts” actually struck a ditch or a culvert. However, although time consuming, the CDS data, narratives, scene diagrams, and photos from the investigators can be manually reviewed to distinguish the two types of objects. This study manually reviewed 12 years of SVROR crashes involving ditch/culvert as the FHE from 1997 to 2008. The number of sampled cases involving ditches as the FHE and the number of crashes they represented in the population are listed in Table 4.18. There are 321 cases in total, 263 of which involved roadside ditches and 58 of which involved median ditches. Of the 263 cases involving a roadside ditch as the FHE, 109 cases (41%) rolled over. Though not a statistically valid comparison, it is interesting to note that 16 (or 28%) of the 58 cases involving median ditches rolled over. The bottom line is that the available number of crash cases involving roadside ditches as the FHE is quite small, even with 12 years of CDS data. Note that the crash data screening criteria described in the introduction to this chapter were applied. Table 4.18. Number of sampled cases involving ditches as the FHE (CDS: 1997–2008). Vehicle Stability Ditch Type Total Roadside Ditch Median Ditch Non-Rollover 154 (113,003) 42 (83,493) 196 (196,496) Rollover 109 (57,645) 16 (2,503) 125 (60,148) Total 263 (170,649) 58 (85,996) 321 (256,645) Note: Values in parentheses are the number of crashes represented by the sample. The object that initiated rollover was determined by the CDS investigators primarily based on scene and vehicle inspections. Secondary information sources included the police report and interviews. As part of the investigation, the investigators identified the source of the force that acted upon the vehicle to initiate rollover. For example, if a curb tripped the vehicle, the curb was selected as the source of the force. Similarly, if a vehicle vaulted (or climbed over) a longitudinal barrier and then rolled over, the barrier was the roll initiation object. If a yawing vehicle rolled as a result of centrifugal forces caused by normal surface friction or as a result of a tire furrowing into soft soil, then “ground” was selected as the source because the ground applied the force that acted as the tripping mechanism for the rollover. In general, ground applies anytime the rollover resulted from tires digging into soft soil, tripping over an accumulation of dirt or gravel, or gouging into the pavement. On the other hand, when a vehicle entered a ditch or culvert and the vehicle rolled over as a result of the slope of the ditch/culvert, the rollover initiation type equaled “turn-over or fall-over,” and the slope was selected as the source. The distinction between turn-over/fall-over and soft soil trip-over is that no furrowing, gouging, or the like occurs on the surface at the point of roll initiation during a turn-over/fall-over. For the 109 rollover cases involving ditches as the FHE, the relative percentage of rollover initiation objects contacted by the vehicle, as determined by the investigators, are listed in Table 4.19. The statistical uncertainty of the estimated percentages is high due to the small

69 sample size and large sampling weights discussed earlier. Nonetheless, close to 69% of the rollovers were estimated as being initiated by soft soil and 15% by ditch slope, indicating that soft soil and slope are the main initiation objects in ditch rollover crashes. Table 4.19. Top rollover initiation objects contacted after striking the ditch as the FHE (CDS: 1997–2008). Rollover Initiation Object Contacted Relative Frequency (%) Sample Size Number of Crashes Represented Ground (soft soil) 68.7 (±23.8) 60 39,574 Ditch Slope 15.3 (±12.9) 15 8,838 Other Fixed Objects (not typical objects, such as trees, poles, posts, barriers, wall, building, curbs, bridge, etc.) 2.4 (±2.5) 8 1,410 Turn-over or Fall-over 2.0 (±1.7) 7 1,130 Tree (>10 cm in diameter) 2.0 (±2.1) 3 1,129 Shrubbery or bush 1.9 (±2.3) 1 1,114 All Others Combined (including those that initiation object was not identified by the investigator) 7.7 15 4,450 All 100 109 57,645 Note: The ± values next to the estimated percentages are standard errors of the estimates (in percentage points) due to statistical sampling, which are computed using the Jackknife resampling technique developed specifically for data collected under a complex survey design—multistage clustered and stratified sampling with unequal inclusion/selection probabilities—as is the case in NASS CDS. NCHRP PROJECT 17-22 DATABASE The NCHRP Project 17-22 database contains 890 single-vehicle ROR crashes that occurred on highways with a PSL of 45 mph or higher. These cases are a small subset of the NASS CDS, mainly from the years 1997 to 2001 (with 16 cases from 2004). For each crash case, the 17-22 database contains more detailed roadway and roadside geometric data than the original NASS CDS database. These additional roadway and roadside data were collected retrospectively from several supplemental field data collection efforts undertaken by NCHRP Project 17-22, NCHRP Project 17-11, and the FHWA Rollover Causation Study projects (32, 13, 39). The NCHRP Project 17-22 database also provides estimates of vehicle roadside encroachment and impact conditions (e.g., encroachment speed and angle distributions at the POD). These elements were reconstructed from the clinical review of crash cases using NASS investigator’s scene diagrams, narratives, photographs of the scene and vehicle, and roadway and roadside data (e.g., surface type, surface condition, and sideslope ratios). Since the crash cases in the NCHRP Project 17-22 database are a subsample of NASS CDS, they inherit all the limitations of the NASS CDS data related to sampling as previously

70 discussed. In summary, it is a biased sample by design in which crashes are selected with unequal probabilities. Moreover, sample cases are not statistically independent because of the sampling from geographical clusters. Conventional statistical methods, which are based on SRS or equal selection probability sampling assumption, are not appropriate for analyzing such data. Specialized survey-based statistical methods are needed to analyze the data. The results of any statistical analysis are meaningless if sampling weights are not properly accounted for in the analysis. Also, the uncertainty of an estimate cannot be properly determined if the clustering and stratified nature of the data are not factored into consideration. The large statistical uncertainty problem discussed in the last two sections regarding the use of GES and CDS data to make small domain estimations is likely even more pronounced with 17-22 data because of a smaller sample size. The sample cases in the database are not particularly representative of real-world SVROR crashes on high-speed highways when compared to those from a larger database, such as NASS GES. For example, out of the 890 cases in the 17-22 database, 479 cases (54%) are rollover cases. With proper weighting, the database suggests 41% of SVROR crashes rolled over, which is too high when compared to the estimate from NASS GES. Out of the 890 cases, 48 cases involved a ditch as a harmful event during the crash, of which 29 cases involved the ditch as their FHE. Out of these 29 cases, 23 cases involved roadside ditches, and the other six cases involved median ditches. Out of the 23 cases involving roadside ditches, nine rolled over (one of which involved an end-to-end rollover). Along with the field-collected roadside data, the 17-22 database is the only database in consideration that could potentially be used to study the relationship between crash severity and ditch configuration. However, with only 23 cases available for study and given the complexity of the roadside encroachment conditions, no meaningful relationship can be developed. HSIS HSIS is a multistate database that contains crash, roadway inventory, and traffic volume data for a selective group of states. It is currently operated by the Highway Safety Research Center at the University of North Carolina and LENDIS Corporation under contract with FHWA. HSIS currently contains data from nine states: California, Minnesota, Michigan, Maine, Utah, North Carolina, Illinois, Washington, and Ohio. The participating states were selected based on the quality of data, the ability to merge various files, and particularly the linkage between the crash file and roadway inventory file. Guidebooks are available online for each participating state. These guidebooks provide detailed descriptions of the variables contained in the data files and how various data files, such as crash and road inventory files, should be linked. A review of the guidebooks indicated that only Washington provides a separate object code for ditches. For other states, the same code value is used for ditches and other roadside features, such as embankments, culverts, and even curbs. Four years of Washington data from 1999 to 2002 were analyzed. Unlike the other databases examined in this study, Washington data contain very limited information regarding the sequence of events of a crash. The first two objects struck by the involved vehicle were recorded in the crash file. The focus of the analysis was on crashes that involved ditches as the first object struck. Crashes on rural two-lane highways with a PSL of 45 mph (72 km/h) or

71 higher were selected for studying crash severity. The maximum PSL where crashes occurred was only 60 mph (97 km/h). Table 4.20 shows the distribution of crash severity and the number of crashes for the analysis period. Comparing these data with those presented in Table 4.15 (for all crashes), which was estimated from FARS and GES data, suggests that there was a lower probability for a crash to result in a fatal or incapacitating injury but a higher probability to result in a non- incapacitating injury, which is due, in part, to the lower PSLs of highways included in the Washington analysis. Table 4.20. Distribution of crash severities for rural two-lane highways (WA State: 1999– 2002). Police-Reported Injury Severity Level Relative Frequency (%) Number of Crashes O: No Injury (PDO) 65.00 390 C: Possible Injury 14.33 86 B: Non-incapacitating Injury 17.50 105 A: Incapacitating Injury 2.83 17 K: Fatal 0.34 2 Total 100.00 600 SEVERITY AND COST OF DITCH-INITIATED CRASHES This section documents a study aimed at estimating the frequency, severity, and economic cost of SVROR crashes. The main purpose of the study was twofold: 1. Understand the magnitude and severity of ditch-initiated crashes, which the design guidelines developed in this project are intended to influence. 2. Obtain good cost estimates of ditch-initiated crashes that can be used to evaluate ditch configurations through a BCA. Each SVROR crash may involve multiple occupants with varying severities of injury. As will be discussed shortly, current crash cost data are provided on a per-injury basis. Therefore, the first objective of the study was to estimate the average crash cost on a per crash basis from the available crash and crash cost data. The estimate was obtained for each of the five police- reported crash severity levels: fatal (Type K), incapacitating injury (Type A), non-incapacitating injury (Type B), possible injury (Type C), and PDO. Rollover crashes are known to be more severe than non-rollover crashes on average. The second objective was, therefore, to quantify the differential cost between rollover and non- rollover crashes. The final objective was to quantify the overall societal costs incurred as a result of ditch-initiated crashes. To meet this objective, the average annual crash cost was estimated for ditch-initiated SVROR crashes. Most SVROR crashes are complex events involving multiple collision and non-collision subevents in sequence during the crash. These subevents could occur both before and after the vehicle encroached onto the roadside. Harmful collisions include colliding with another vehicle on the road and striking fixed objects, such as trees, utility poles, and traffic barriers, on the roadside. Harmful non-collisions before impacting a ditch include rollover, fire, and explosion.

72 Focusing on crashes involving ditches as the FHE (i.e., the first property-damaging or injury-producing event in the crash) ensured that the injury to occupants and damage to vehicles did not occur before the vehicle struck the ditch. After impacting the ditch as the FHE, a significant portion of the involved vehicles struck fixed objects, especially trees and utility poles, which resulted in severe injury and damage. The severity outcomes of these crashes usually did not reflect the safety performance of the ditch itself. To exclude these crashes, some studies have suggested using the MHE—the single impact that causes the greatest trauma and damage in each crash—to further restrict the crashes for analysis. Based on crash data screening criterion advocated by Viner et al. (40) for studying roadside safety, ditch-initiated crashes are defined in this study as “those SV ROR crashes that involved roadside ditches as the FHE and the MHE was either the ditch or a rollover.” FHE was used in the screening to ensure that the ditch was the first subevent that produced injury or damage, and the MHE was used to further limit crash cases to those in which the cause of greatest harm was most likely the ditch. It is expected that the injury severity distribution of this subset of ditch-related crashes would be more indicative of the safety performance of the ditch design. Crash Data and Screening Criteria NHTSA FARS and NASS GES crash data from 2004 to 2009 were used (6, 7, 30, 31, 41). FARS contains a census of all fatal crashes and covers all vehicle types. It was used to extract relevant data on fatal crashes. Although the GES covers all injury severity levels, it was used to provide data for nonfatal crashes only. As discussed earlier, GES data are obtained from a probability sample of police-reported traffic crashes covering all severity levels and vehicle types. In order to calculate estimates of national crash characteristics, each crash case must be weighted using the national weight variable provided in the file. Because the estimates produced from the GES are based on a probability sample of crashes and not a census of all crashes, the estimates are subject to sampling errors and may differ from the true values. Over the years, the GES sampling design and coding methods have changed, and these changes have to be considered when selecting crash cases for analysis. For example, some variables were dropped, new ones were added, and coding of individual variables were changed in some years. When the FHE and MHE of each crash case are coded, the GES has used the same code for ditch and culvert. It is, therefore, not possible to tell from the data whether a crash involving ditches/culverts actually struck a ditch or a culvert. FARS data, on the other hand, allow this distinction to be made. With this GES data limitation in mind and understanding that culverts are an integral part of the drainage system, this study analyzed both ditch-initiated and culvert- initiated crashes and estimated crash costs for both. The portion of crash costs due to ditches was further estimated based on the FARS data. Not all ditch-initiated crashes were selected for analysis. Based on the objectives of this project, three additional constraints on vehicle type, PSL, and type of ditches/culverts were imposed to qualify the crash data. Thus, data had to meet the following criteria: Those involving passenger cars, SUVs, vans, and pickup trucks. Those occurring on highways with a PSL between 45 mph (72 km/h) and 75 mph (121 km/h), inclusive.

73 Those involving roadside ditches/culverts as the FHE, and the MHE was either the ditch/culvert or a rollover. This study limited its scope to roadside ditches/culverts. Design of median ditches for unprotected, divided highways requires different considerations, including the need to reduce the frequency of cross-median crashes. Ditch-initiated crashes occurred on both sides of the undivided highways, and those occurring on the right side based on travel direction on divided highways were selected. Note that those crashes occurring on entrance and exit ramps and those on undivided highways with a two-way continuous left turn lane were not considered. The rationale to target crashes occurring on highways with a PSL of 45 mph (72 km/h) or higher is that lower-speed highways, including many of the urban and suburban four-lane and two-lane undivided highways, tend to limited ROW for considering an open ditch design, and a closed drainage system is often the preferred choice. Crash Cost Data The recommended values of life and injuries for preparing economic evaluations of highway improvement projects have been issued and updated over the years by the Office of the Secretary of Transportation (OST), U.S. Department of Transportation, since 1993 (42). The relative value of an injury of a certain severity or property damage has been set as a percentage of the economic value of life since the first estimate was issued in 1993. Since then, the same percentage has been adopted in all updates. In recent years, the majority of state DOTs use the OST-recommended estimates in some fashion to prioritize proposed safety projects (43). Table 4.21 presents the OST-recommended crash cost per injury and per property damage for various police-reported severity levels in 1993 and 2008 (Columns 2 and 3). The 2008 recommended costs, in 2007 dollars, were adjusted to 2010 dollars using the Consumer Price Index (CPI) published by the Bureau of Labor Statistics. The adjusted costs are shown in the last column of the table. The cost of a fatality, in 2010 dollars, is estimated to be over $6 million, which is about 1,300 times the cost of a crash involving PDO. Table 4.21. OST-recommended crash cost per injury and per property damage for various police-reported severity levels. Occupant Injury Severity 1993 Recommended Cost in 1994 $ 2008 Recommended Cost in 2007 $ 2008 Recommended Cost in 2010 $* K: Fatal 2,600,000 5,800,000 6,052,478 A: Incapacitating Injury 180,000 401,538 419,018 B: Non-incapacitating Injury 36,000 80,308 83,804 C: Possible Injury 19,000 42,385 44,230 PDO 2,000 4,462 4,656 *Adjusted CPI (December to December index) for all urban customers on all items, published by the Bureau of Labor Statistics, U.S. Department of Labor, 2011. Cost per Crash by Severity Level The estimates provided by OST are on a per fatality, injury, and property damage basis. Since each SVROR crash may involve more than one fatality and injury, these estimates need to

74 be multiplied by an average number of fatalities and injuries involved in a crash and summed to obtain the total per-crash cost. At the nationwide level, data from FARS and GES are usually used for this purpose (41). By using 6 years of FARS and GES data, the average numbers of fatalities and injuries involved in each ditch/culvert-initiated crashes were estimated. Table 4.14 in the section on NASS GES presented these estimates for each police-reported crash severity level. As discussed, given that the police-reported severity level is classified according to the known maximum injury severity that occurred in a crash, a fatal crash will involve at least one fatality and may include other occupants with varying degrees of injury. Since the SV crash is in consideration, only one vehicle is damaged. By using the recommended per-injury cost in Table 4.21 and the average numbers of fatalities, injuries, and property damage per crash in Table 4.14, the average cost per crash by police-reported severity level was calculated and is presented in Table 4.22. On average, a police-reported fatal crash of interest costs about $6.5 million, and an incapacitating injury crash costs more than $0.5 million (in 2010 dollars). The relative cost is also shown in the table. The cost of a fatal crash is equal to the cost of about 1,396 PDO crashes, while the cost of an incapacitating injury crash is equal to the cost of about 114 PDO crashes. This great disparity in the crash cost between severity levels, especially between a fatal and a nonfatal crash, makes the reduction in fatal and incapacitating injury crashes a consequential objective in developing ditch design guidelines. Table 4.22. Average cost per crash by police-reported injury severity level for ditch/culvert-initiated crashes. Police-Reported Injury Severity Cost per Crash in 2010 $ Relative Cost (w/ PDO=1.0) K: Fatal 6,498,827 1,396 A: Incapacitating Injury 530,888 114 B: Non-incapacitating Injury 109,450 24 C: Possible Injury 56,537 12 PDO 4,656 1 Since FARS differentiates between ditches and culverts, the average numbers of injuries per fatal crash in Table 4.14 can be regenerated for ditches and culverts separately, and the per- crash cost for a fatal crash in Table 4.22 can be recomputed accordingly. The difference in per- crash cost between a fatal crash involving ditches and that involving culverts is practically insignificant (less than 1.0% of the cost involving ditches). In addition, since there were more ditch-initiated crashes than culvert-initiated crashes (about a 7:3 ratio in fatal crashes), the per- crash cost for a ditch-initiated fatal crash is less than the cost shown in Table 4.22 by only 0.3%. This result suggests that even though the per-crash costs in Table 4.22 were calculated for crashes involving ditches and culverts together, these per-crash costs should be quite representative of the cost associated with the ditch crash alone.

75 Costs of Rollover and Non-rollover Crashes As presented earlier for fatal crashes, after striking a ditch as the FHE, the MHE includes rollover (54%), standing tree (16%), ditch itself (14%), utility pole (4%), and culvert (3%). It is obvious that rollover is the most consequential subevent after striking the ditch. The second objective of this study was to estimate the average per-crash cost separately for rollover and non-rollover crashes. The analysis started by regenerating the estimates in Tables 4.14 and 4.22 for rollover and non-rollover crashes separately. The average per-crash costs by severity level are shown in Table 4.23. As shown, except for the incapacitating injury crashes, the difference in crash cost for each crash severity level was quite small between rollover and non-rollover crashes. For example, the cost of a “fatal rollover” crash is basically the same as the cost of a “fatal non-rollover” crash. Note that the same property damage cost, $4,656, was used both for rollover and non-rollover crashes, which is most likely an underestimation for rollover crashes. Table 4.23. Average cost per crash by police-reported severity level for rollover and non- rollover crashes. Police-Reported Injury Severity Rollover Crashes (in 2010 $) Non-Rollover Crashes (in 2010 $) All Crashes (in 2010 $) K: Fatal 6,497,023 6,512,387 6,498,827 A: Incapacitating Injury 552,885 498,668 530,888 B: Non-incapacitating Injury 112,821 105,090 109,450 C: Possible Injury 54,901 57,997 56,537 PDO 4,656 4,656 4,656 Table 4.15 in the section on NASS GES shows the estimated numbers of crashes by police-reported severity level and their distribution in percentage: Column 3 for rollover crashes, Column 5 for non-rollover crashes, and Column 7 for all crashes. Over 75% of the non-rollover crashes did not incur any injury, while only about 42% of the rollover crashes did not have injuries. About 3.03% of the rollover crashes resulted in fatalities, while only about 0.28% of the non-rollover crashes did. For each severity level, the table also shows the portion of rollover and non-rollover crashes (in parentheses). Overall, after striking a ditch/culvert as the FHE, about 34% of the ditch-initiated crashes rolled over and 66% did not. However, the rollover percentage varies significantly between injury severity levels. For example, about 85% of the fatal crashes rolled over after striking ditches/culverts as the FHE, while about 22% of the PDO crashes rolled over. Based on the per-crash cost in Table 4.23 and the severity distribution of crashes in Table 4.15, the estimated per-crash cost for rollover, non-rollover, and all crashes are presented in Table 4.24. On average, a ditch/culvert-initiated crash costs about $153,000, 79% of which is attributable to the occurrence of fatal (K) and incapacitating injury (A) crashes. If the vehicle rolled over after striking the ditch as the FHE, the expected crash cost is about $312,000, while the cost is about $62,000 for a non-rollover crash. For rollover crashes, 88% of the per-crash cost is attributable to the occurrence of fatal (K) and incapacitating injury (A) crashes, while the share is about 69% for non-rollover crashes.

76 Table 4.24. Estimated per-crash cost for rollover, non-rollover, and all crashes. Type of Crashes Cost per Crash in 2010 $ Rollover 311,694 Non-Rollover 61,915 All 153,042 For rollover and non-rollover crashes, the great disparity in the share of per-crash cost between severity levels indicates, again, the importance of reducing fatal and incapacitating injury crashes if the developed design guidelines from this project are to make a significant overall reduction in ditch-initiated crash costs. In addition, the cost disparity between rollover and non-rollover crashes suggests that reducing the rollover propensity of encroached vehicles is a consequential goal of the guideline development. Underreporting of Minor Crashes Police-reported crash data are known to underreport minor crashes involving roadside safety devices and hazards, and the rates of underreporting are difficult to determine. In addition, the underreporting rates vary significantly among safety hardware systems. Based on a literature review, Council and Stewart (44) estimated underreporting rates for some roadside features as follows: Approximately 30 to 40% of impacts with guardrails and median barriers might be unreported in some databases, and 10 to 15% of utility pole impacts, 4 to 8% of impacts with non-breakaway traffic signals and luminaire supports, 50% of crash cushion impacts, and 30% of impacts with other breakaway devices might be unreported. They cautioned that their estimates of underreporting were based on studies that were quite old (1970s and early 1980s), and those estimates based on maintenance data could not clearly separate below reporting-threshold impacts from those that should, in actuality, be in the police files. The underreporting rate estimate for the guardrails and median barriers of 30 to 40% were mainly for rigid and semi-rigid types of barriers. In another study, Sicking et al. (45) used an underreporting rate of 26% for cable median barriers in a BCA of installing cable barriers to reduce cross-median crashes. The number of unreported cable barrier crashes was estimated by comparing the number of cable barrier repairs to the number of cable barrier crash reports. Their further analysis indicated that this estimate of 26% was a high estimate. This study certainly raises a question as to whether the earlier 30-40% estimate for rigid and semi-rigid barriers was too high. No underreporting rate was available specifically for ditch-related crashes in the literature. However, it was expected that the underreporting should mainly reside in non-rollover PDO crashes. Under this project, an assumption was made that 30% of PDO non-rollover crashes were unreported. Based on this assumption, the number of non-rollover PDO crashes in Table 4.15 was increased from 194,885 to 278,407. Table 4.25 shows the adjusted per-crash cost under this assumption. As shown, after adjusting for the underreporting, a crash involving a ditch/culvert as the FHE costs about $127,000 on average, 77% of which is attributable to the occurrence of fatal (K) and incapacitating injury (A) crashes. If a vehicle rolled over after striking the ditch as the FHE, the expected crash cost is still about $312,000, while the cost is about $48,000 for a non-rollover crash—a 6.5:1 ratio. For rollover crashes, 85% of the per-crash

77 cost is attributable to the occurrence of fatal (K) and incapacitating injury (A) crashes, while this percentage is reduced to about 68% for non-rollover crashes. Table 4.25. Per-crash cost adjusted for underreporting of non-rollover PDO crashes. Type of Crashes Unadjusted Cost per Crash in 2010 $ Adjusted Cost per Crash in 2010 $ (Assuming 30% Underreporting of PDO Non-Rollover Crashes) Rollover 311,694 311,694 Non- Rollover 61,915 47,961 All 153,042 126,885 Average Annual Crash Cost and Conclusions To get a sense of the overall societal cost associated with the ditch-initiated crashes analyzed in this study, the final objective of this crash data study was to estimate the average annual crash cost of these crashes. Based on the crash frequencies provided in Table 4.15 with an adjustment of the assumed underreporting and the adjusted per-crash cost in Table 4.25, the total cost of ditch/culvert-initiated crashes from 2004 to 2009 was estimated to be about $60 billion, in which about $39 billion was ditch-related and $21 billion culvert related. The average annual crash cost was thus about $10 billion, with about $6.6 billion related to ditches and $3.4 billion to culverts (about a 2:1 ratio). About 71% of the annual crash cost was associated with rollover crashes, and 29% was associated with non-rollover crashes. In conclusion, ditch-initiated crashes, which the developed design guidelines in this project are expected to address, are currently costing society about $6.6 billion annually and, on average, a ditch-initiated crash costs about $127,000 per crash under the current ditch configurations in the field, in which a rollover crash costs about $312,000 and a non-rollover crash costs about $48,000.