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.
2 Methods of Measuring Performance Chapter 1 identified three components essential to a program of in-service performance evaluation: the administrative and planning framework that determines the scope and objectives of evaluation, responsibilities for evalu- ation, and the applications of results; data systems, which constitute the infrastructure of evaluation; and evaluation methods, that is, specifications for analysis of crash and roadway data to measure performance. This chap- ter presents examples of evaluation methods. The examples also illustrate challenges regarding administration and planning of in-service evaluations. Alternative methods of measuring performance are illustrated with ex- amples from past road safety evaluations. The alternatives include the cat- egories identified in Chapter 1, retrospective and prospective comparative evaluation and descriptive evaluation. Chapter 3 proposes applications of these generic methods in special evaluation research studies, and Chapter 4 describes applications for routine evaluation needs of state highway agen- cies as part of their safety management and asset management programs. COMPARATIVE EVALUATIONS The objective of a comparative evaluation of two alternative types of road- side safety features is to determine which type provides greater protection (that is, the lower risk of injury or death) for road users. The same evalu- ation methods may be applied to determine whether risk is reduced by the use of a particular roadside safety treatment as compared with risk if no action is taken. Statistical techniques are used to control for factors other than the treatments being evaluated that may affect safety. The methods are 28
METHODS OF MEASURING PERFORMANCE 29 standard practice in research on the safety effects of roadway and vehicle design features and the benefits of safety interventions. In the examples of methods of evaluation of guardrail end treatments given below, the primary measure of performance usually is the distribution of the severity of crash outcomes (for example, the fraction of all collisions with an end treatment that result in a severe injury or fatality). The methods are applicable to evaluations of other kinds of roadside features, although for some features, such as rumble strips or traffic roundabouts, the measure of performance will be the frequency of collisions rather than (or in addi- tion to) the consequences of collisions. Prospective Evaluation In a prospective study, procedures are put in place for identifying and documenting all crashes of interest (e.g., collisions with guardrail end treat- ments) on specified sections of road within a specified future time period. Such a study proceeds in the following steps: 1. Definition of the hypotheses concerning safety performance to be tested; 2. Selection of the roadways and future time period for which crashes are to be observed; 3. Collection of collision data; 4. Assembly of a road features inventory, that is, data on the charac- teristics of the features being evaluated and on other characteristics of the environment that may influence the risk of crashes and ca- sualties; and 5. Analysis by means of statistical methods to determine whether the observed frequency and characteristics of crashes are consistent with the hypotheses. Meeting the study objectives also may require comparing the costs of alternative safety treatments and engineer- ing assessments of the sources of poor performance. These steps are described below. National Cooperative Highway Research Program (NCHRP) Report 490, In-Service Performance Evaluation of Traffic Barriers (Ray et al. 2003) and a report by the Texas Transportation Institute (TTI) for the Texas Department of Transportation (DOT) (van Schalkwyk et al. 2004) propose detailed procedures for each of the steps. (Both reports are summarized in the annex to Chapter 3.) Box 2-1 presents an example of a prospective comparative evaluation of end treatments.
30 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS Box 2-1 Prospective Evaluation Example: Washington State Evaluation of Guardrail End Treatment The Washington State Department of Transportation undertook an in-service evaluation of guardrail end treatments in use in the state to help determine state policy regarding replacement of terminals of older designs (Igharo et al. 2004). The project also served as a demonstration of the methods developed in NCHRP Report 490 (Ray et al. 2003). The same data collection and analysis effort also evaluated the performance of unrestrained precast concrete barriers and illustrated that cost-effectiveness can be increased by combining evaluations of multiple features. Procedure The steps in the end treatment evaluation were as follows: 1. Definition of hypotheses to be tested. Two hypotheses regarding end treatments were tested statistically: â The probability distribution of outcome severity, given that an end treatment is struck, is independent of the device type. â The probability of being struck is independent of the design type. (This can be regarded as a test of whether the end treatment design installed at a location depended on road characteristics.) 2. Selection of roadways and time period. Data were collected for 1 year for 752 miles of state roads in three highway agency maintenance districts with 6 billion annual vehicle miles of travel. 3. Road features inventory. As the state had no inventory of end treatments, a survey of the location and design type of end treatments on the study roads was conducted. The survey recorded 2,318 end treatments of the six designs to be evaluated. Geometric and traffic data for the study roads were obtained from existing state databases. 4. Collection of collision data. The departmentâs maintenance staff identified and inspected the site of each end treatment collision on the study roads within several days of the collision to determine device installation char- acteristics and damage to the device. A standard data form was used. These maintenance reports were matched with police accident reports. Collision scenarios were determined on the basis of the maintenance staff inspections and police reports. Thirty collisions were identified, of which 20 had accident reports. No crashes resulted in fatalities. Four crashes resulted in disabling injuries, eight resulted in lesser injuries, and 18 resulted in property damage only. 5. Analysis. Of the six end treatment types in the road inventory data, twoâthe breakaway cable terminal (BCT) and the slotted rail terminal (SRT)âexperienced a sufficient number of collisions to support statistical comparison of severity frequencies. The number of collisions by severity for the two types was as follows:
METHODS OF MEASURING PERFORMANCE 31 Type Disabling Injury Other Injury Property Damage Only BCT 3 4 11 SRT 1 2 6 A test of statistical significance of the difference in severity distribution between the two types shows that the hypothesis (that the probability of crash outcome is independent of end treatment type) cannot be rejected. The similarity of the distributions is illustrated in Figure 1, which shows the percentage of crashes in each severity category for the two types of devices. The analysis also tested the hypothesis that the probability of being struck is independent of the design type. SRT were struck four times more often than BCT per 100 million vehicles passing, as Figure 2 shows. The difference is statistically significant. No information in the report explains this difference. Limitations of the Method ⢠The scale of the study, in terms of the time period and extent of the roads included, allowed only 30 end treatment collisions to be observed. Most device types experienced too few collisions to support conclusions about relative performance. Any failure modes arising from flaws in design or installation of the devices probably would not have been observed unless they occurred in a high percentage of collisions. 1 0 20 40 60 80 Incapacita ng Injury Other Injury Property Damage Only Cr as he s ( % ) BCT SRT FIGURE 1 Percentage of BRT and SRT crashes by severity of injury. continued
32 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS ⢠The comparison of end treatment types with respect to crash severity does not adequately take into account factors other than the design of the devices that may affect severity. Road conditions and crash charac- teristics varied greatly: roads included Interstates and minor low-volume roads; ramps were the locations of 3 percent of end treatment installa- tions but 16 of the 30 crashes observed. Such differences are likely to be important in explaining severity distribution. The comparative analysis made no use of the device inventory data collected, although the device inventory was used in a descriptive analysis in the report. ⢠The study attempted to control for other factors using crash modification factors (Igharo et al. 2004, 26â28). However, the factors calculated for the SRT and BCT device types hardly differed from unity (1.008 and 1.012, respectively) and application of the factors had no effect on the large dif- ference in collision rates between the two device types. ⢠The difference in collision frequency between BCTs and SRTs might arise if the stateâs policies favored installation of either BCTs on roads with relatively low overall run-off-road crash rates or SRTs on ramps (where a large share of collisions occurred) or if one end treatment type presented a larger cross-section for collisions because of differences in design or installation. Any of these circumstances might invalidate the studyâs sta- tistical comparison of crash severity for the two types. ⢠As the study report noted, the rate of injury collisions per 100 million vehicles passing was much lower than in other published studies (e.g., 10 percent of the rate observed in a study that used Iowa data). The dif- ference is further indication that all factors that influence crash severity require full examination. 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 Incapacitang Injury Other Injury Property Damage Only Total Co lli si on s pe r 1 00 M ill io n Ve hi cl es P as si ng BCT SRT BOX 2-1 Continued FIGURE 2 Guardrail end treatment collision rate, by type of device and severity of injury.
METHODS OF MEASURING PERFORMANCE 33 Step 1. Safety Performance Hypotheses The essential first step of the evaluation is to define the hypotheses to be tested to ensure that the necessary data are obtained and that the planned sample size will be adequate and to avoid effort in collecting inessential information. In a comparative evaluation of end treatment types, the hy- potheses may include the following: ⢠For specified alternative roadside feature designs (e.g., two types of end treatment), the distribution of outcome severity (e.g., the fraction of all collisions that result in a fatality or serious injury), given that a feature is struck, is independent of the device type. ⢠The probability of being struck is independent of the design type. Testing this hypothesis will serve as a check for systematic errors in data collection for the importance of overlooked variables in the analysis. Evaluations of roadside devices have generally assumed that the probability of being struck is independent of device type. If this assumption is sound, any observed difference in the frequency of collisions (per vehicle passage) may arise from a systematic difference in the characteristics of the roads on which each of the types compared is installed. Conceivably, the design of some types of devices may render them more likely to be struck than other types. For example, one type may have greater overall dimensions than others or may require placement closer to the travel lane. ⢠The distribution of outcome severity is independent of the quality of installation, age, and condition of the struck feature. Comparison of the life-cycle cost or cost-effectiveness of alternative types of devices may be an additional objective of the evaluation, in which case a hypothesis involving comparison of repair costs would be included. Step 2. Selection of Roads and Study Duration Past evaluations of roadside features have used crash data from road net- works ranging from a segment of a single route (e.g., Bischoff and Battaglia 2007) to the entire state road system (e.g., Schrum 2014) or the systems of multiple states (e.g., Gabauer 2014) and have been conducted in time periods typically ranging from 1 year to several years. Considerations re- garding the selection of the locations and the time period for crash data collection are as follows: ⢠Sample size requirements. The final section of this chapter dis- cusses the determination of the number of observations of crashes
34 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS required to allow the hypotheses to be tested. The number of observations obtained will depend on the extent of the roads in- cluded in the study, the density of the roadside features of interest, traffic volume, and the frequency of crashes per vehicle passing the features. ⢠Representativeness of the sample of roads and crashes. Confining the study to roads with high traffic volumes will reduce the road mileage and time required to observe the target number of crashes; however, the study results may not be generalizable to low-volume roads, which generally differ from high-volume roads in geometric and operational characteristics that affect crash severity. ⢠Presence of roadside features to be evaluated (e.g., end treatments of selected design types). Similarly, selecting roads with a high den- sity of the features being evaluated will reduce mileage and time requirements, but the performance of the feature on these roads might differ from the performance on other kinds of roads. ⢠Cooperation of the necessary participants. A state-conducted evalu- ation will need to recruit the cooperation of road maintenance staff and contractors, police agencies, and, possibly, local road agencies. Some past evaluations have been confined to selected highway agency maintenance districts to simplify the involvement of main- tenance workers. National evaluation research studies will require the cooperation of multiple state highway agencies. ⢠Data availability. The timeliness and completeness of crash re- ports, roadside and roadway inventory data, and maintenance reports may vary within a state or from state to state in a national evaluation. Past evaluations that used data on all crashes within a defined study area and time period have thereby avoided the problems of statistical sampling; however, as noted above, selection of the study area or roads included by judgment rather than by a sampling process might bias results. Step 3. Collection of Collision Data Past evaluations have relied on three sources of information about crashes: ⢠Police accident reports, ⢠Maintenance department reports of damage to roadside devices, and ⢠Special inspections of crash sites, including inspection of the struck roadside devices involved.
METHODS OF MEASURING PERFORMANCE 35 Obtaining prompt, reliable reports of relevant crashes has been a difficult organizational challenge in past in-service evaluations. The procedure for a prospective in-service evaluation generally will call for inspection of crash sites as soon as possible after the event. The normal procedure for entering police crash reports in state crash data systems may be too slow to provide notification to the evaluators; in this case, special arrangements with police will be needed to identify relevant crashes. Consistent notification of crashes and data quality assurance require cooperative relationships between the agency staff responsible for the evaluation, police, and maintenance staff. Box 2-2 lists crash data elements that have been employed in past safety evaluations of roadside devices or that may be relevant to such evalua- tions. The data elements required will depend on the specific objectives Box 2-2 Crash-Related Data Elements Relevant to Evaluation of Roadside Devices Crash Scenario ⢠Collision type (e.g., run-off-road) ⢠Sequence of events (including first harmful event and most harmful event) ⢠Number and types of vehicles involved ⢠Contributing factors (e.g., speeding) ⢠Driver characteristics (age, sobriety, gender) ⢠Injury severity ⢠Number of occupants involved ⢠Number of injured occupants ⢠Restraint usage ⢠Severity of vehicle damage Roadside Device Information ⢠Characteristics of devices involved (design type, installation details, condition, age, maintenance history) ⢠Postcrash condition ⢠Repair cost Road and Environmental Conditions ⢠Road surface condition ⢠Roadway geometry ⢠Traffic characteristics ⢠Light condition ⢠Time of day ⢠Weather condition
36 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS of the evaluation. Box 2-3 shows the crash data elements that the Federal Highway Administration (FHWA) judged to be relevant for its recent inves- tigation of the in-service performance of the ET-Plus end treatment design. Data collection forms in NCHRP Report 490 (Ray et al. 2003), the TTI- proposed evaluation method for Texas (van Schalkwyk et al. 2004), and the FHWA end treatment in-service evaluation pilot study (FHWA n.d.) are described in the annex to Chapter 3. Box 2-3 FHWA Federal Register Data Request Regarding Crashes Involving the ET-Plus End Treatment The Federal Highway Administration (FHWA) judged the following crash data elements to be relevant for its recent investigation of the in-service performance of the ET-Plus end treatment design (Federal Register 2014): 1. ET-Plus Crash Data and Information. The FHWA is seeking data and information concerning vehicle crashes involving the ET-Plus. We are seeking crash reports, photographs of damaged ET-Plus devices at crash scenes, photographs of vehicles at crash scenes that impacted ET-Plus devices, and crash reconstruc- tion reports with corresponding data. The type of data and information we are requesting includes, but is not limited to, the following: â Crash narratives; â Crash diagrams; â Severity of the crash as noted in the crash report (Killed, A injury, B injury, etc.); â The approximate mass, speed, and angle of impact of the vehicle; â The orientation of the vehicle as it impacted the terminal (head-on, side- impact, front corner, etc.); â The location of the crash (state, route, county, mile marker); â The type of road on which the crash occurred; â The weather at the time of the crash; â The condition of the shoulder and/or roadside at the time of the crash; â The installation and maintenance history of the terminal; and â The condition of the terminal prior to the impact. 2. ET-Plus Dimensions as Installed. The FHWA is seeking data concerning the dimensions of the ET-Plus devices installed on highways. In particular, we are interested in a few key dimensions: channel width, exit gap, guide chute exit height, and outside guide channel length. We also are interested in any other dimensions that could be useful in determining the in-service performance of the ET-Plus. We are asking for existing data and information that the public may have and are not asking the public to undertake any activities that may risk the safety of themselves or others. NOTE: Severity as classified by the KABCO injury scale: K = fatal (killed), A = incapacitat- ing injury, B = nonincapacitating injury, C = possible injury, and O = property damage only.
METHODS OF MEASURING PERFORMANCE 37 Some past evaluations have restricted crashes included in the analysis to those in which collision with the roadside device to be evaluated was the most harmful event (as judged by the police officer completing the crash report form), reasoning that, in other crashes involving the device, the outcome severity was determined by events other than collision with the device. However, this restriction probably excludes some crashes that could provide information about the performance of the device and intro- duces a subjective judgment in the selection of crashes. Studies also have placed other restrictions on crashes included in the analysis, for example, the exclusion of motorcycle and truck crashes. The evaluation method developed in NCHRP Report 490 and applied in the Washington DOT study described in Box 2-1 (Igharo et al. 2004) calls for collection of data on roadside device collisions not reported to police by arranging for maintenance crews to report on damaged devices observed during regular or special roadside inspections. In contrast, the authors of the procedure developed at TTI for the Texas DOT concluded that the benefit of the information to be derived from investigation of non- police-reported crashes would not be great enough to justify the added cost and effort and therefore decided not to include special inspections to iden- tify unreported crashes in its recommended procedure (van Schalkwyk et al. 2004, 29â30). Large numbers of minor collisions are unreported. A study that monitored guardrail end treatments on a 22-mile segment of Interstate highway in Iowa over a 12-month period observed evidence of 69 colli- sions, of which four were reported to police. Most collisions involved no more than very minor damage to the device (Ray and Hopp 2000, 46â47). Including non-police-reported crashes in a comparative evaluation may be useful in several ways: ⢠If the likelihood of a crash of a given severity being reported to police depends on the location of the crash, comparisons of device types on the basis of the ratio of severe to total police-reported crashes may be misleading. For example, if minor crashes were less likely to be reported on rural roads than in urban areas, but severe crashes had similar rates of reporting, then any device type that was more common on rural roads would tend to have a higher ratio of severe to total police-reported crashes. ⢠The benefits of improving the design of a roadside safety device may include reduction in the costs of less severe crashes, including costs of crashes not reported to police. Such a benefit may not be measurable in an evaluation that uses only police-reported crashes. ⢠Investigation of barrier performance in minor crashes may yield information about mechanical performance or failure risks.
38 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS ⢠Information about damage to and repair of the devices involved in all crashes is needed if the evaluation is to compare the cost- effectiveness of alternative types of devices. More experience with the conduct of in-service evaluations will be neces- sary before the utility of data on non-police-reported crashes can be as- sessed. In evaluations in which available resources are very limited, use of only police-reported crashes may be a justifiable simplification. The ability to inspect the crash site within a short time of the event is the major advantage of a prospective study as compared with a historical study. The inspection may ⢠Determine the type of roadside feature involved (e.g., the manufac- turer and model of the end treatment struck) and the mechanical response of the feature in the collision, ⢠Observe crash site circumstances that may affect severity (e.g., road alignment and roadside conditions), and ⢠Collect evidence of the precrash condition of the feature and the crash scenario. A vehicle collision with a roadside device may obliterate details about the installation and precrash condition of the device. However, the experience of the American Association of State Highway and Transportation Officials (AASHTO)-FHWA Task Force in assembling data on crashes involving guardrail end treatments shows that relevant information on installation and precrash conditions is often obtainable from postcrash inspections (Joint AASHTO-FHWA Task Force on Guardrail Terminal Crash Analysis 2015, 111â115). Information also may be obtainable from precrash main- tenance records and photologs. Inspections serve to verify information in police crash reports, which may not consistently record roadside feature involvement. Instances of high rates of miscoding in crash data files were reported to the committee: 25 percent of crashes coded as involving a guardrail terminal in one state were found not to involve a terminal, and in another state, 20 percent of actual crashes involving guardrails were coded as involving bridge rails.1 Determining the sequence of events in barrier and end treatment crashes remains a difficult task even in a prospective evaluation. Techniques have been developed especially for reconstruction of these crashes (Coon and Reid 2006). 1 Dean Sicking presentation to the committee April 14, 2015.
METHODS OF MEASURING PERFORMANCE 39 Step 4. Assembly of Roadway Data Possible uses of information on roadway and roadside characteristics in the study area and at crash sites include the following: ⢠Data on the locations and design types of the roadside features to be evaluated are necessary for the selection of the study area. ⢠Comparison of the performance of alternative device types must take into account other factors that may influence crash severity, including roadway geometry and traffic characteristics. ⢠Evaluations that compare devices on the basis of the rate of severe collisions per vehicle passing the device require data on the num- bers and locations of all devices in the study area. ⢠Roadway geometry also may influence the probability that a road- side device is struck per vehicle passing the device. A comparison on the basis of the rate of severe collisions per vehicle passing should take this effect into account. ⢠If the evaluation is to measure the effects of device age and condi- tion on performance, historical records on installation and main- tenance will be necessary. ⢠If an objective of the evaluation is to guide selection of the best device type for particular locations, information will be needed on road geometry, traffic, and roadside characteristics. The sources of road data in past studies have been highway agency asset management, maintenance management, and traffic databases; photologs; special surveys conducted for the evaluation; and data from inspection of crash sites. Step 5. Safety Analysis The data analysis tests the evaluation hypotheses by computing measures of performance for each of the feature types being compared and by using statistical methods to determine whether the differences in the measures are great enough that they are unlikely to result solely from random variation in crash frequency. In addition, if an intended application of the evalua- tion is to improve the design of roadside devices or to validate or improve crash testing, then engineering analysis will be necessary to determine the mechanical sources of differences in performance between device types. The full distribution of outcome severityâthat is, the fractions of all crashes (police-reported and not reported) that result in property damage only, minor injuries only, incapacitating injuries, and fatalitiesâis the most complete measure of the performance of a roadside safety feature intended
40 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS to mitigate crash outcomes. Some evaluations have used simplified forms of severity distribution for comparing the safety performance of alternative types of roadside features, including ⢠The rate of severe crashes (crashes resulting in fatal or incapacitat- ing injuries) per vehicle passing the roadside feature and ⢠The ratio of severe crashes to all police-reported crashes involving the feature or the distribution of severity in police-reported crashes. Finally, in some studies, the performance measure is the failure rate of the roadside feature, that is, the fraction of all police-reported crashes in which the roadside feature does not perform as intended by its designers. The validity of these measures as indicators of the risk of a severe outcome given that a crash has occurred depends on assumptions that may not hold in all cases. If the likelihood of a crash of a given severity being re- ported to police depends on the location of the crash, comparisons of ratios of severe crashes to all police-reported crashes may be biased. Use of failure rate as the measure introduces judgment into the evaluation; moreover, two alternative device types with similar failure rates could have differing rates of severe outcomes if one type dissipated crash energy more effectively. The risk of severe injury in a collision with a roadside device is in- fluenced by numerous characteristics of the vehicles involved, the vehicle occupants, and the roadway and roadside, in addition to the design and condition of the roadside device struck. In a comparative evaluation of device types, if the locations of one device type are correlated with the pres- ence of another risk factor, and the analysis does not take the other factor into account, the results of the comparison will be misleading. For example, in a Connecticut maintenance district in 1994â1995, the fraction of all police-reported barrier collisions that resulted in fatal or incapacitating in- juries was 1 percent on Interstate highways and 5 percent on non-Interstates (see Figure 2-1). If one type of barrier was prevalent on Interstates and an- other type was more frequently used on non-Interstates (and assuming that the difference in severity between Interstates and non-Interstates was not due solely to the difference in barrier types), then the difference in severity distribution attributable to the difference in road class would obscure any difference attributable to barrier type. A correlation between external risk factors and the locations of a particular device type also might arise if the highway agency had a policy of using one particular device type as the replacement for any device dam- aged in a collision. Then the favored replacement type would become over- represented at locations with relatively high crash risk. Methods of controlling for confounding factors that have been applied in in-service evaluations of roadside devices include the following:
METHODS OF MEASURING PERFORMANCE 41 ⢠Assuming that the locations of a roadway safety device type are not correlated with roadway and traffic characteristics that affect crash risk (i.e., that a particular device type is not more likely to be found at a location with a high crash risk than at a low-risk location), ⢠Estimating a multivariate model of crash severity, ⢠Borrowing an existing model of crash frequency or severity and examining whether prediction error correlates with roadside device design type,2 and ⢠Using a case-control study design. The examples of in-service performance evaluation in this chapter illustrate these methods. The Washington DOT study described in Box 2-1 attempted to apply crash modification factors. The retrospective evaluation described in Box 2-4 used a multivariate model to control for several risk factors. A later section in this chapter describes examples of the case-control method. Because end treatment crashes other than minor crashes are infrequent 2 Crash modification factors are a kind of model that has been used for this purpose (Ray and Weir 2002, 80â83). The trial of this method in the Washington State evaluation described in Box 2-1 did not appear promising. However, research in progress will provide crash modifi- cation factors specific to run-off-road crashes (Persaud 2016, 5), which may provide a means of controlling for road and traffic characteristics in future evaluations. 0 10 20 30 40 50 60 70 Fatal or Incapacitang Injury Other Injury Property Damage Only Co lli si on s ( % ) Interstate Non-Interstate FIGURE 2-1 Severity of barrier collisions, Connecticut District 1, 1994â1995. SOURCE: Ray et al. 2003, 33, Table 6.
42 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS Box 2-4 Retrospective Evaluation Example: Factors Affecting Injury Risk in Crashes with Guardrails and Guardrail End Terminals In a study of injury risk in frontal crashes with guardrail and guardrail end termi- nals, Johnson and Gabler used historical crash data compiled by the U.S. Depart- ment of Transportation to identify the factors influencing the severity of crashes with guardrails and guardrail terminals (Johnson and Gabler 2014). The steps in the evaluation, which closely paralleled the steps listed in the text section on prospective evaluations, were as follows: 1. Hypotheses to be tested. The hypotheses concerned comparisons of injury risk, including - Comparison of risk in collisions with guardrail terminals of types that had passed crash tests specified in National Cooperative Highway Research Program (NCHRP) Report 350 (Ross et al. 1993) with risk in collisions with end treatment types that had not passed the tests. - Comparison of risk in collisions with the guardrail face with risk in col- lisions with guardrail terminals. 2. Roadway, time period, and collision data. The crash records analyzed were from the National Automotive Sampling SystemâCrashworthiness Data System (NASS-CDS), a sample of crashes drawn from all U.S. police-reported crashes in which a vehicle is towed from the scene. About 5,000 crashes per year are included. Data are from an investigation of each crash conducted by the National Highway Traffic Safety Adminis- tration (NHTSA) as well as from the police report. Therefore, the crash records are more complete and detailed than records that would be avail- able for a retrospective study that used state crash record files alone. The crashes analyzed were all those crashes in the database that occurred in the years 1997 to 2008 involving frontal impact of a car or a light truck with a guardrail or a guardrail end treatment and in which the impact, or a rollover initi- (see Chapter 1, Tables 1-1 and 1-2), prospective in-service evaluations usually have collected only small samples of crashes. Consequently, dis- tinguishing differences in performance among device types is difficult. For example, the Washington State DOT study described in Box 2-1 collected a sample of 30 crashes, and the Wisconsin DOT study described in Box 2-5 observed 20 crashes. The necessity of controlling for confounding factors in the data analysis compounds the problem of the small sample size. Statisti- cal analysis methods have been developed for making the best use of small samples, and these methods may be applicable in roadside device in-service evaluations. Bayesian statistical methods are a collection of techniques that
METHODS OF MEASURING PERFORMANCE 43 ated by the impact, was the most harmful event. The crashes that were identified as meeting these criteria involved a total of 711 vehicles. 3. Road features data. Photos taken in the course of the NHTSA crash investigations were used to determine whether the collision was with the guardrail face or with an end treatment and whether end treatments were of a type that had passed an NCHRP Report 350 crash test. The authors noted that âDetermining the exact end terminal system involved in a crash from NASS scene photos can be very difficultâ (Johnson and Gabler 2014, 3). 4. Analysis. Regression analysis was used to control for confounding fac- tors in comparisons of the risk of severe injury in crashes with NCHRP Report 350âcompliant terminals versus noncompliant terminals and in crashes involving a guardrail face versus an end treatment. The con- founding variables considered included - Vehicle type (car or light truck or van); - Airbag deployment; the study also investigated whether NCHRP Re- port 350 compliance affected airbag deployment; - Occurrence of rollover; the study also compared the risk of rollover in end treatment crashes with the risk in guardrail face crashes; and - Seat belt use. The authors do not report considering the effect on injury risk of road characteris- tics other than guardrail and guardrail end treatments in the analysis. The NASS- CDS crash records contain limited information on road characteristics. The study concluded that the odds of a serious injury in a crash with an end treatment that was not NCHRP Report 350âcompliant were between 4.2 and 5.3 times greater than in a collision with a compliant end treatment. (The range in the estimate reflects differing assumptions about the type of end treatment for end treatments whose type could not be determined with certainty.) The study also concluded that the odds of serious injury in a crash with an end treatment are 5.1 times greater than in crashes with the guardrail face. incorporate collision data collected in the evaluation with other sources of information (e.g., historical collision data or expert opinion) to provide a probabilistic comparison of the relative risk of different road safety treat- ments or conditions. Bayesian methods have been increasingly applied in highway safety studies (e.g., Hauer et al. 2002; Schultz et al. 2011). The methods can produce more precise estimates than traditional methods of the safety consequences of a treatment. Evaluators can minimize data analysis problems arising from nonran- dom selection of device types to be installed at particular locations by in- forming themselves of the highway agencyâs device selection practices. The
44 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS Box 2-5 Descriptive Evaluation Example: Trial Installation: Evaluation of Wisconsin ET-2000 The Wisconsin Department of Transportation undertook an evaluation of a par- ticular guardrail end treatment design (the ET-2000) (Bischoff and Battaglia 2007). The objectives were to evaluate safety performance and also ease of installa- tion and maintenance, weather resistance, and other cost-related performance features. Procedure The evaluation proceeded as follows: ⢠Forty-three ET-2000 end treatments were installed along a 17-mile sec- tion of urban Interstate in Milwaukee County. ⢠Installation followed state specifications, and a manufacturerâs represen- tative observed installations. ⢠Performance was observed for 5 years. The report does not state whether any formal procedures were in place for inspection of the devices or special crash investigations. ⢠Comments on frequency and ease of maintenance were solicited from county maintenance workers. ⢠Collisions with the test end treatments were identified from police acci- dent reports. Twenty collisions were documented, including two collisions involving nonfatal injuries and 18 property-damage-only collisions. ⢠No data on non-police-reported collisions were collected. evaluation can reduce problems from omission of explanatory variables by explicitly considering in the analysis all variables that reasonably may be expected to influence crash incidence and outcomes and for which data are available. Statistical methods are available to help avoid misleading conclu- sions caused by unrecognized omitted variables or by nonrandom device selection (Mannering and Bhat 2014; Mannering et al. 2016; Wolshon and Pande 2016, 13â15). Retrospective Evaluation A retrospective comparative evaluation uses existing data on historical crashes and on roadway and traffic characteristics to test hypotheses concerning the relative safety of roadside features. Such a study can be conducted in less time and at lower cost than a prospective evaluation;
METHODS OF MEASURING PERFORMANCE 45 Analysis and Conclusions The evaluators concluded that The ET-2000 systems installed as part of this research study have performed well. Vehicles that collided with the end terminals were safely brought to a stop, and very few injuries were reported. . . . Reduced repair costs and added safety features also justify the higher initial cost of the ET-2000. (Bischoff and Battaglia 2007) Inspections after collisions revealed that reflectors installed on the end treatments sometimes prevented guardrail separation from support posts as intended during a collision. The reflectors were removed. Limitations of the Method Some features of the study limit the strength of its conclusions: ⢠There was a lack of diversity of the road types included in the evaluation (a single 8-mile section of urban Interstate was evaluated). ⢠The sample was small. Infrequent failure modes that would be unlikely to appear in a sample of 20 crashes could nonetheless be important for safety. If a kind of failure occurs with a frequency of 1 in 30 collisions, the probability of observing it in 20 crashes is less than 50 percent. ⢠The special care employed in the installation of the end terminals may not represent typical practice. however, routinely collected data generally lack the detail needed to sup- port a conclusive comparative evaluation. Some retrospective studies have supplemented the data in crash and road inventory databases with limited special data collection, for example, a special inventory of roadside fea- tures in the study area. Box 2-4 presents an example of a retrospective comparative evaluation. A second example is the evaluation of guardrail end treatment performance in Missouri described in the section below on case-control studies (see Box 2-6). These two examples illustrate that a primary difficulty in a retrospec- tive evaluation is a lack of information on the characteristics of roadside devices involved in crashes. Historical crash databases typically contain little information on device type or precrash condition or on the response of the device to the collision, although information may be obtainable from examination of crash scene photos and photologs. As noted above, errors
46 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS Box 2-6 Case-Control Evaluation Examples Insurance Institute for Highway Safety Study of Effect of Truck Configura- tion on Crash Rates The Insurance Institute for Highway Safetyâs study to observe the effect of trac- tor trailer configuration on crash rates provides an illustration of the case-control method (Braver et al. 1997). The hypothesis of the part of the study described here was that the risk of involvement in a crash per mile of travel is the same for double trailer configurations as for tractor semitrailers operating under the same conditions. Procedure The cases were all crashes involving tractor trailers on Interstate highways in Indiana during a 5-month period; 2,033 crashes were identified: 102 double trailer crashes and 1,931 single trailer crashes. The controls were intended to be all trac- tor trailers observed passing the site of each crash on the same day of the week and within 30 minutes of the time of day of the crash during the 4 weeks following each crash. Controls matched to cases were observed in this way for about half of the crash cases; the researchers did not learn of the remaining crash cases in time to observe control vehicles. The researchers describe their original research design as a âprospective matched case-control study,â but because controls could not be observed for every case, âthe study changed from a matched to an unmatched case-control designâ (Braver et al. 1997, 83). The control vehicles observed were 3,059 double trailers and 59,860 single trailers. Analysis The following table summarizes the data: Type of Vehicle Cases (crash-involved vehicles) Controls (vehicle count at crash sites) Double trailer 102 3,059 Other configuration (single trailer) 1,931 59,860 The odds ratio is (crashes involving double trailers)/(crashes involving other configurations) (controls that are double trailers)/(controls that are other configurations)
METHODS OF MEASURING PERFORMANCE 47 that is, (102/1,931)/(3,059/59,860) = 1.03 The 95 percent confidence interval for the ratio is 0.84 to 1.27; therefore, the hypothesis that double trailers have the same crash involvement rate as other configurations cannot be rejected. The statistical analysis in the study was more careful than this calculation. A multivariate model was estimated to control for differences between cases and controls in day of week, time of day, urban or rural location, and specific road. The cases differed from the controls in the distributions of these characteristics because the cases were not fully matched to controls. Comparative Evaluation of End Treatment Types in Missouri Schrum (2014) compared in-service performance of various end treatment design types in Ohio and Missouri. The results from Missouri are described below. The hypothesis tested was that the frequency with which an ET-Plus termi- nal is involved in a crash with an incapacitating or fatal injury is the same as the frequency for other ET terminals. It was assumed that all ET terminals in the case and control groups had equal probability per unit of time of being struck in a crash. Under this assumption, if one end treatment type is significantly over-represented among severe crashes, then the risk of injury, given that a crash has occurred, must be greater for that end treatment type. Procedure The cases were 156 end treatments involved in single-vehicle run-off-road crashes in Missouri in 2005 to 2014. Crashes were identified in the stateâs police-reported crash file. A crash-involved end treatment was not included in the cases if it was not a design that met the crash test standard of NCHRP Report 350 (Ross et al. 1993), if the vehicle was a truck or motorcycle, if the downstream end treatment of the guardrail was struck, or if essential data were not available. The cases included 93 ET-Plus terminals and 49 other ET terminals (ET-2000 type). The controls were all end treatments in the 10-mile road segment prior to each end treatment crash site. These included 1,200 ET-Plus terminals and 961 other ET terminals. The analysis in effect assumed that traffic characteristics and roadway characteristics at the locations of the controls were sufficiently similar to the characteristics of the crash sites that any observed differences in crash frequency by end treatment type could not be attributable to differences in traffic or roadways at the terminalsâ locations. continued
48 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS Analysis The data were as follows: Treatment Cases (ET terminals involved in severe crashes) Controls (all ET terminals in 10-mile road segment prior to crash site) ET-Plus 93 1,200 Other ET (ET-2000) 49 961 The odds ratio is (cases involving ET-Plus)/(cases involving other ET) (controls that are ET-Plus)/(controls that are other ET) that is, (93/49)/(1,200/961) = 1.52 The 95 percent confidence interval for the odds ratio is 1.06 to 2.17; therefore, the hypothesis that ET-Plus terminals have the same frequency of severe crashes as other ET terminals is rejected. Limitations of the Method Similarity of traffic and road conditions at the crash site and in the 10 miles pre- ceding the crash is a critical assumption that requires testing. For example, if a highway agency had a policy of replacing crash-damaged ET-2000s (the older design type) with the newer ET-Plus terminals, ET-Plus terminals would become over-represented at locations with a high crash risk. The retrospective study design required 10 years of crash records. Deter- mining and verifying crash circumstances from past years is challenging. Other studies (e.g., Ray and Hopp 2000, 47; D. Sicking presentation to committee, April 14, 2015) have noted problems with reliably identifying end treatment crashes in police crash records. Box 2-6 Continued in coding information about roadside devices involved in crashes may be frequent in historical crash databases. Only a few states have inventory databases containing information on the types, locations, and maintenance history of roadside devices. Even where such inventories are available, assembling a database of crash re- cords linked with information about roadway and traffic characteristics at crash sites remains difficult. For example, a study of motorcycle collisions
METHODS OF MEASURING PERFORMANCE 49 with guardrails that used FHWAâs Highway Safety Information System (a multistate database of crash, roadway, and traffic data [FHWA 2012]) was able to include crashes from only two states in its analysis because road- way alignment data were too incomplete for other states in the database (Gabauer 2014; Gabauer and Li 2015). Such limitations in highway agen- ciesâ basic information systems hinder the conduct of in-service evaluations. The motorcycle study found that roadway geometry (specifically, whether the road is divided or undivided) influences the risk of motorcycle rider injury given that a crash has occurred and that alignment influences the risk that a collision will occur at a location. These results indicate the need to consider geometry as a confounding factor in an in-service evalua- tion of guardrail end treatments. The retrospective evaluation described in Box 2-4 used data on crashes across a period of 12 years, and the Missouri case-control study described in Box 2-6 used data across a period of 10 years. Data analysis in histori- cal studies that used data across a multiyear period should consider the possibility that cyclical patterns in crash frequency and crash rates may affect comparisons of devices. Crash frequency and crash rates tend to decline during recessions and to increase during economic recoveries. In a comparative evaluation of two alternative device types, if installation of the devices is occurring during the period of the study, omission of consider- ation of cyclical patterns in crash frequency and rate may lead to incorrect conclusions. The method of controlling for confounding factors by using regression analysis, as in the study described in Box 2-4, is applicable to prospec- tive as well as retrospective evaluations. The external factors considered in the example study were found to strongly influence risk, indicating the importance of taking them into account in comparing the performance of alternative roadside feature types. If important external factors are numer- ous, obtaining a crash sample large enough to allow estimating a model of injury risk will be challenging. Case-Control Studies A case-control study of the relationship of a road feature to crash frequency or severity proceeds in four steps: 1. Establishment of the hypothesis. It is hypothesized that a particular characteristic of a road feature (e.g., the type of design of guardrail end treatments) affects the severity of crashes with that feature or that the presence or absence of a treatment (e.g., shoulders with and without rumble strips) affects the risk that a crash will occur.
50 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS 2. Identification of cases. If the measure of performance is the crash rate (i.e., the purpose of the road feature is to prevent crashes), then the cases are all installations of the road feature under study that have experienced a crash (or a sample of all such installa- tions) within the study area and time period. If the measure of performance is crash severity (as in an evaluation of a guardrail end treatment, whose purpose is to mitigate the consequences of a crash that has occurred), then the cases are all installations of the feature that have experienced a crash exceeding some threshold severity. 3. Selection of controls. For each case, one or more controls are se- lected. These are road features selected from the same study area and time period that have not experienced a collision (or a collision above the threshold severity), and which are matched to the case in respects that are believed to affect the frequency or severity of crashes. 4. Analysis of data. The data are analyzed by comparing the propor- tion of cases with the particular characteristic hypothesized to be related to crash risk or severity with the proportion of controls possessing the characteristic. Box 2-6 describes examples of the application of the case-control method to a study comparing truck crash rates and to a safety evaluation of alternative types of guardrail end treatments. The examples illustrate the advantages of the case-control research design. For a study of compara- tive truck safety, the more common method of comparing crash risks is estimation of crash involvement rates per vehicle mile for each truck type. This method requires detailed data on vehicle miles of travel for each truck type over the entire road system included in the study, classified by traf- fic, roadway, and environmental characteristics. The case-control design requires truck traffic data only at crash sites. On the other hand, the odds ratio derived in a case-control study is a measure only of relative risk, not absolute risk. The method does not yield a model of how risk depends on the characteristics of interest and external factors. The case-control method is applicable in either prospective evaluations, in which data collection is planned for the purpose of the evaluation, or retrospective evaluations, which use historical data that may not have been collected with the needs of the evaluation in mind. The comparison of truck crash rates described in Box 2-6 was a prospective evaluation; the guardrail end treatment comparison in Box 2-6 was retrospective.
METHODS OF MEASURING PERFORMANCE 51 DESCRIPTIVE EVALUATIONS In a descriptive evaluation, the subject roadside feature is regularly ob- served for a period of time. At the end of the trial period, the evaluator judges whether the feature is performing as intended and whether a need is indicated for any changes in design or in installation or maintenance practices. In contrast with a comparative evaluation, no quantitative esti- mate is made of the effect of the feature on casualty risk as compared with alternative designs or treatments. Forms of descriptive evaluation that have been applied to roadside features include the following: ⢠A trial installation of a new device. Box 2-5 presents an example. The new feature evaluation proposed in the Manual for Assessing Safety Hardware (MASH) (see Chapter 1, Box 1-3) is a further example. ⢠A series of crash investigations as case studies. An example is the 2015 AASHTO-FHWA Task Force investigation of the perfor- mance of guardrail end treatments in crashes (Joint AASHTO- FHWA Task Force on Guardrail Terminal Crash Analysis 2015). The AASHTO-FHWA Task Force assembled a set of records of historical crashes involving end treatments. The cases were exam- ined by teams of expert reviewers who were to judge whether the devices performed as intended during crashes and to identify per- formance limitations that could affect the risk of injury in a crash. This analysis is described in Chapter 3. ⢠A special inventory and inspection of roadside features, to deter- mine the frequency of conditions believed to affect crash injury risk. An example is the AASHTO-FHWA Task Force inspection of guardrail end treatments to determine whether the design of devices installed on roads differs from the design of the devices subjected to crash testing (AASHTO-FHWA Measurement Task Force 2015, n.d.). The elements of such evaluations may be summarized as follows: ⢠Definition of objectives, typically observation of failures or of fea- tures of device design and environment that affect performance; ⢠A procedure for identifying cases, which may be retrospective, as in the AASHTO-FHWA Task Force investigations, or prospective, as in the Wisconsin trial described in Box 2-5; ⢠A standard data collection protocol (in a prospective study); and ⢠Analysis of the cases by experts to determine the incidence of the design or performance features of interest.
52 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS Statistical sampling principles are relevant to the plan of a descriptive evaluation, even if the design does not involve using statistical methods for hypothesis testing. First, the set of cases chosen for examination (e.g., collisions or device installations) must be representative of the population of interest. This can be achieved by including all known cases in the evalu- ation (e.g., all collisions with a device type in the study area and period) or random sampling from among all known cases. Choosing cases according to subjective criteria risks biasing the evaluation. If cases are all drawn from one road class or one region, the results of the evaluation may not be generalizable to other parts of the road system. Second, if the possibility of rare events is a concern, the number of cases must be large enough that rare events are likely to be observed. Descriptive evaluations have the advantages of simplicity and low cost compared with more rigorous methods. They may be capable of providing earlier warning of potential problems that might go undetected for longer if a new device or design was installed without formal evaluation. However, the method does not provide a quantitative estimate of relative risk com- pared with alternative designs or roadside treatments. Also, in a formal test such as the Wisconsin trial, the care taken in installation and maintenance and the locations selected for installation may not reflect practices that would prevail if the roadside feature under study came into general use. If these test conditions are not representative of general use, the evaluation results may be misleading. REFERENCES Abbreviations AASHTO American Association of State Highway and Transportation Officials FHWA Federal Highway Administration NCHRP National Cooperative Highway Research Program AASHTO-FHWA Measurement Task Force. 2015. AASHTO-FHWA Task Force on ET-Plus 4â Dimensions. March 11. https://www.fhwa.dot.gov/guardrailsafety/dimensionsreport. pdf. AASHTO-FHWA Measurement Task Force. n.d. FHWA Review of ET-Plus. https://www.fhwa. dot.gov/guardrailsafety/mtf.cfm. Bischoff, D., and I. Battaglia. 2007. ET-2000 End Treatment for Guardrail: Final Report. Wis- consin Department of Transportation. December. http://wisconsindot.gov/documents2/ research/wifep-03-07guardrailendtreatment.pdf.
METHODS OF MEASURING PERFORMANCE 53 Braver, E. R., P. L. Zador, D. Thum, E. L. Mitter, H. M. Baum, and F. J. Vilardo. 1997. Tractor- Trailer Crashes in Indiana: A Case-Control Study of the Role of Truck Configuration. Accident Analysis and Prevention, Vol. 29, No. 1, pp. 79â96. Coon, B. A., and J. D. Reid. 2006. Reconstruction Techniques for Energy-Absorbing Guardrail End Terminals. Accident Analysis and Prevention, Vol. 38, No. 1, January, pp. 1â13. Federal Register. 2014. ET-Plus Guardrail End Terminal. Vol. 79, No. 247, Dec. 24, p. 77595. https://www.federalregister.gov/documents/2014/12/24/2014-30081/et-plus-guardrail- end-terminal. FHWA. 2012. Highway Safety Information System: The Essential Analysis Tool for Mak- ing Informed Safety Decisions. November. http://www.hsisinfo.org/download/HSIS_ Brochure_2013.pdf. FHWA. n.d. Data Collection: In-Service Performance Evaluation of Guardrail End Terminals. Gabauer, D. J. 2014. Roadway Characteristics Associated with Motorcycle Crashes into Lon- gitudinal Barriers and the Influence on Rider Injury. Presented at 93rd Annual Meeting of the Transportation Research Board, Washington, D.C. Gabauer, D. J., and X. Li. 2015. Influence of Horizontally Curved Roadway Section Charac- teristics on Motorcycle-to-Barrier Crash Frequency. Accident Analysis and Prevention, Vol. 77, April, pp. 105â112. Hauer, E., D. W. Harwood, F. M. Council, and M. S. Griffith. 2002. Estimating Safety by the Empirical Bayes Method: A Tutorial. Transportation Research Record: Journal of the Transportation Research Board, No. 1784, pp. 126â131. Igharo, P. O., E. Munger, and R. W. Glad. 2004. In-Service Performance of Guardrail Terminals in Washington State. Report No. WA-RD 580.1. Washington State Depart- ment of Transportation, Olympia. June. https://www.wsdot.wa.gov/research/reports/ fullreports/580.1.pdf. Johnson, N. S., and H. C. Gabler. 2014. Injury Risk in Frontal Crashes with Guardrail and Guardrail End Terminals. Presented at 93rd Annual Meeting of the Transportation Re- search Board, Washington, D.C. Joint AASHTO-FHWA Task Force on Guardrail Terminal Crash Analysis. 2015. Safety Analysis of Extruding W-Beam Guardrail Terminal Crashes. https://www.fhwa.dot.gov/ guardrailsafety/safetyanalysis. Mannering, F., and C. Bhat. 2014. Analytic Methods in Accident Research: Methodological Frontier and Future Directions. Analytic Methods in Accident Research, Vol. 1, pp. 1â22. Mannering, F., V. Shankar, and C. Bhat. 2016. Unobserved Heterogeneity and the Statistical Analysis of Highway Accident Data. Analytic Methods in Accident Research, Vol. 11, pp. 1â16. Persaud, B. 2016. Critical Review of Methodologies for Evaluating In-Use Safety Performance of Guardrail End Treatments and Other Roadside Treatments. Background Paper for Special Report 323: In-Service Performance Evaluation of Guardrail End Treatments. Transportation Research Board, Washington, D.C. Ray, M. H., and J. Hopp. 2000. Performance of Breakaway Cable and Modified Eccentric Loader Terminals in Iowa and North Carolina: In-Service Evaluation. Transportation Research Record: Journal of the Transportation Research Board, Vol. 1720, pp. 44â51. Ray, M. H., and J. Weir. 2002. Appendix D: In-Service Performance Evaluation Procedures Manual. Contractorâs Final Draft, NCHRP Project 22-13(2). Transportation Research Board of the National Academies, Washington, D.C. http://onlinepubs.trb.org/online pubs/nchrp/nchrp_rpt_490aD.pdf. Ray, M. H., J. Weir, and J. Hopp. 2003. NCHRP Report 490: In-Service Performance Evalu- ation of Traffic Barriers. Transportation Research Board of the National Academies, Washington, D.C.
54 PERFORMANCE EVALUATION OF GUARDRAIL END TREATMENTS Ross, H. E., Jr., D. L. Sicking; R. A. Zimmer; and J. D. Michie. 1993. NCHRP Report 350: Recommended Procedures for the Safety Performance Evaluation of Highway Features. Transportation Research Board of the National Academies, Washington, D.C. Schrum, K. 2014. Relative Comparison of NCHRP 350 Accepted Guardrail Terminals. University of Alabama at Birmingham, October 28. http://www.thesafetyinstitute.org/ wp-content/uploads/2014/10/Relative-Comparison-of-NCHRP-350-Accepted-W-Beam- Guardrail-Terminals.pdf. Schultz, G. G., D. J. Thurgood, A. W. Olsen, and C. S. Reese. 2011. Analyzing Raised Median Safety Impacts Using Bayesian Methods. Transportation Research Record: Journal of the Transportation Research Board, No. 2223, pp. 96â103. van Schalkwyk, I., R. P. Bligh, D. C. Alberson, D. L. Bullard, Jr., D. Lord, and S.-P. Miaou. 2004. Developing an In-Service Performance Evaluation (ISPE) for Roadside Safety Features in Texas. Texas Transportation Institute, College Station. December. http:// d2dtl5nnlpfr0r.cloudfront.net/tti.tamu.edu/documents/0-4366-1.pdf. Wolshon, B., and A. Pande. 2016. Critical Review of Methodologies for Evaluating In-Use Safety Performance of Guardrail End Treatments and Other Roadside Treatments. http:// onlinepubs.trb.org/onlinepubs/sr/sr323paper3.pdf.
