Influence of Infrastructure Design on Distracted Driving (2022)

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.

59 Chapter 6. Safety Framework for the Use of Crash Narratives to Assess the Impact of Infrastructure Design on Distraction 6.1 Introduction This Safety Framework assessed whether crash narratives can be utilized to identify IRDs. 6.1.1 Background State crash report forms typically include three main sections. The first identifies the people and vehicles involved in the incident and its location. The second consists of multiple choice questions about the roadway, drivers, vehicles, and crash circumstances. This is followed by a narrative section where responding officers can provide freeform comments and a sketch describing the incident. In many jurisdictions, completion of a narrative is optional. Most of the crash data available to analysts is derived from the multiple choice section of the crash report. These data are relatively easy to obtain, are usually in electronic form, and can provide general indicators of roadway features, vehicle characteristics, and driver characteristics. Distraction is usually included as a multiple choice question and is coded as either inside the vehicle or outside the vehicle. Additional discussion on how distractions are treated in crash forms is provided in Section 2.2. Under the current crash reporting schema, all types of OVD crashes are assigned the same reporting code: “External Distraction (People, Objects, or Events).” In most crash forms, no additional information beyond this designation is provided as to the source of the OVD. One of the concerns with the use of crash forms for analysis is that distraction is inferred from either interviews with drivers or witnesses (who may not be forthcoming) or circumstantial evidence. Another concern is that distraction is not consistently defined, and officers are not consistently trained in how to identify and report distraction, particularly distractions outside the vehicle. As a result, distraction is not consistently coded in crash reports. Not all crash reports include narratives. Frequently, even when available, information contained in narratives is duplicative of information entered into other fields. However, in some cases narratives provide insights and details that cannot be gleaned from the multiple choice questions. For instance, an officer might write, “The driver of Vehicle 1 said she was distracted by guide signs.” Regardless of the value of such details, a key disadvantage is that the use of crash narratives often entails resource-intensive manual review of hundreds (or thousands) of crash reports. Many state motor vehicle crash databases contain a field that allows for the rapid identification of crashes flagged by law enforcement as involving an external distraction. Some of these crashes are infrastructure related, but many involve other types of distractions such as pedestrians or animals in or near the roadway, other motor vehicles making unusual or unexpected maneuvers (e.g., skidding), previous crashes, or emergency vehicles in response mode.

60 To separate infrastructure from non-infrastructure external distractions, the facts and circumstances described in crash narratives can be reviewed manually. The reviewer can then assign the crash to one of several categories based on its primary cause. These could include a variety of infrastructure categories (roadside advertising, wind turbine, light rail transit station, roadside sculpture, etc.) and non-infrastructure categories (pedestrian, animal, previous crash, etc.). Although manual analysis requires a degree of judgment on the part of the reviewer, subjectivity can be reduced by establishing the categorization criteria in advance. If necessary, subjectivity can be further reduced by having each narrative analyzed by two reviewers, with any conflicts subject to discussion between the reviewers or an independent review by a third person. The analysis described in this section involved a review of crash narratives to determine whether additional information about crashes where OVDs are indicated can be gathered. This can help determine the extent to which the reported distractions are related to the built environment as opposed to people, animals, and other types of distraction. 6.1.2 Scope of Framework The first objective of this analysis was to determine whether crash narratives are likely to have sufficient information for the evaluation of IRDs. The second objective was to assess the efficacy of using crash narratives to develop relationships between infrastructure elements and distractions. A list of potential research questions that could be addressed using crash narratives was developed. These questions include but are not limited to the following: • What infrastructure characteristics are identified as being related to an OVD in crash narratives? • Are there differences between infrastructure characteristics identified as being related to an OVD noted in all crashes versus fatal and serious injury crashes? • Do crash narratives provide additional information about the source of an OVD other than that provided in the coded crash data? • What changes to crash forms or training programs are recommended to obtain more robust information about OVDs? • How do the details about OVD crashes vary among jurisdictions, and what are the implications for analysis of the relationship between IRD and crash risk? • To what extent is information about an OVD translated from crash narratives to the coded crash form? One research question was selected for further investigation to determine whether the use of crash narratives was feasible: What infrastructure characteristics are identified as being related to an OVD in crash narratives? This research question was intended to identify which infrastructure elements are commonly noted in narratives, which is an important first step in determining whether crash narratives are a reasonable data set to utilize.

61 6.2 Data Sets Utilized Many states provide public or reasonably straightforward access to crash data. However, due to the potential for exposing personal identifying information and the difficulty in coding narratives (especially if not entered electronically by attending officers), many states do not provide ready access to crash narratives. In some cases, access requires a memorandum of understanding or other protocols for this information to be accessed. Crash data for Michigan, which include crash narratives, are publicly available through the Michigan Traffic Crash Fact (MTCF) data query tool (https://www.michigantrafficcrashfacts.org/), as shown in Figure 13. As a result, this data set was utilized for this analysis. Source: MTCF © 2021 The Regents of the University of Michigan (https://www.michigantrafficcrashfacts.org/querytool#q1;0;2020;;). Figure 13. MTCF portal. 6.3 Data Request and Data Reduction External Distractions is a relatively new reporting field in motor vehicle crash databases. For instance, such crashes have been flagged in the Michigan database only since January 1, 2016. Crash narratives were obtained from the MTCF data query tool from 2017 through 2019. A total of 17,127 fatal and serious injury crashes were identified during this timeframe. Filters were set to only include crashes where OVDs were noted as having occurred. Although crashes may involve IRD and not be coded as involving an OVD, the use of this subset of crashes was the only way to narrow the number of narratives reviewed to a reasonable quantity. The driver distraction field was used to identify OVD crashes, and those crashes coded as “Activity Outside Vehicle” were selected and then filtered by crash severity (in this case, fatal [K] or suspected serious injury [A]). K+A crashes were utilized because officers are more likely to record details for serious crashes than for minor crashes. Crash reports that met these criteria

62 for each of the three years were downloaded. As shown in Table 7, a total of 428 crash narratives were available. Table 7. Crash narratives coded as involving an OVD. Year K A Total for Year 2017 20 109 129 2018 26 122 148 2019 18 133 151 Overall Total 64 364 428 A data reduction protocol was developed, and the data were reviewed by one data reductionist, with another data reductionist checking approximately 25% of the narratives. Each narrative was manually reviewed, and the following information was noted: • Date. • Severity. • OVD (yes or no). • Infrastructure related (yes or no). The type of OVD that was noted in the narrative (if available) was also included. For OVD crashes that were not infrastructure related, the general categories included the following: • Ambiguous (driver was noted as looking at something or being distracted by something but no further information was provided). • Animal. • Construction. • Debris. • Disabled vehicle. • Downed utility pole. • Emergency vehicle. • Other vehicle (but not specified as disabled or parked). • Other object. • Parked vehicle. • Person. • Prior crash. • Smoke. • Train. • Weather related (e.g., glare from sun, fog). • No distraction noted. A list of categories was also developed to indicate the type of infrastructure noted as a distractor. Although most of the categories were ultimately not used, they included the following: • Bridge. • Bridge abutment. • Other bridge features.

63 • Sign. • Signal. • Other intersection features. • Culvert. • Drainage structure. • Guard rail or other barrier. • Concrete or cable barrier median. • Pedestrian structures. • Transit structures. • Wind turbine. • Billboard or business sign. 6.4 Analysis The narratives fell into three categories: • No OVD noted (n = 240). • OVD noted but not infrastructure related (n = 183). • OVD and infrastructure related (n = 5). A description of the types of narratives that were included in each category is provided below. Although the first two categories are not infrastructure related, the description may be useful for stakeholders interested in reviewing crash narratives for other information. When the text from the crash narratives was provided, details identifying the person or the location were redacted. Irrelevant details (e.g., time) were also removed for ease of reading. 6.4.1 Crash Narratives Where No OVD Was Noted A total of 240 narratives coded as involving an OVD were determined to not actually involve an OVD or for none to be evident in the narrative. Based on their narratives, at least half of the reports appear to have been miscoded, in that they do not describe a situation that could be interpreted as a distraction. In other cases, the narrative described an inside the vehicle distraction. Examples include the following: • “Unit 1 was traveling WB and swerved onto the shoulder of the roadway striking a construction worker (Unit 2) and a maintenance vehicle. The driver of unit 1 advised he was texting on his cell phone at the time of the crash.” o The driver was focused on an internal distraction (cell phone). As a result, the distraction may have been miscoded as involving an OVD. • “Veh #1 traveled through the intersection at a high rate of speed and disregarded the "Stop" sign causing a collision with Veh #2.” o No obvious source of distraction was available. In many cases, it is likely that the officer was using the field to indicate that an activity outside the vehicle (but not necessarily a distraction) contributed to the crash. In some cases, the likely

64 source of outside activity was a traffic situation that the driver should have been attending to and, as such, should not have been interpreted as a distraction. Examples include the following: • “Driver 1 of the semi-truck stated she saw that traffic was slowing down but was unable to stop; she stated she applied the emergency brake.” o The officer likely interpreted traffic as the source of external distraction. • “Driver of truck entered the shoulder to apparently avoid an oncoming vehicle that had crossed the centerline. Driver of truck struck bicyclist.” o The officer likely viewed the oncoming vehicle as the source of distraction leading to a crash with a bicyclist. 6.4.2 Crash Narratives Where OVD Was Present but Not Infrastructure Related In other cases, a distraction was present outside the vehicle but was not related to infrastructure. This scenario accounted for 183 of the narratives coded. Examples include the following: • “#1 was traveling north and abruptly turned left in the path of #2. The listed witness stated #1 abruptly turned in front of #2 and appeared to be turning to look at a vehicle for sale on the west side of the road.” o The source of distraction was a vehicle for sale. • “Unit 1 was traveling NB. Driver 01 stated he was looking at a man attempting to flag a vehicle down on the side of the road.” o The source of distraction was a person outside the vehicle. • “Unit 1 was north bound. A vehicle stopped with its hazard lights on partly in the south bound lane divided the driver's attention. The driver also noticed a vehicle in the ditch from a prior accident.” o The source of distraction was a prior crash. • “Unit 1 was NB in eastern most lane approaching the intersection. Unit 2 was EB approaching the intersection with a red light. Unit 2 disregarded traffic control red light and drove into intersection. Driver of Unit 2 stated that he was looking for a McDonalds that GPS indicated was close by and he did not notice the traffic signal.” o The source of distraction was the driver looking for a business. Twenty-four of the narratives coded as non-infrastructure-related OVDs involved the driver paying attention to an animal. In most cases, the animal was on or near the roadway and the crash likely resulted from the driver attempting to avoid the animal. Depending on the definition of “outside-the-vehicle distraction,” these cases may not have involved OVDs in the strictest sense. For instance, a driver paying attention to an animal in the roadway may have been listed as involving an external distraction, and the animal may have taken the driver’s attention off the roadway. However, this may not have strictly been a distraction because the driver did need to attend to the activity. Since this distinction for non-infrastructure-related objects was beyond the scope of this project, no further assessments were made, and such crashes were included as involving an OVD but not an IRD. Examples include the following: • “Vehicle 1 was EB. Driver stated she was looking at horses in a field. Veh 1 ran off the south side of the roadway and collided with a tree.” • “Vehicle 1 was traveling S/B and swerved to avoid a squirrel, striking a tree.”

65 6.4.3 Crash Narratives Where OVD Was Present and Infrastructure Related Five narratives were noted as being infrastructure related based on the definition utilized for this research (see Section 1.4.1). Three were noted as involving a distraction related to outdoor advertising and include the following: • “Driver 1 was attempting to turn left into a driveway and was struck by Driver 2 who was distracted by a roadside food stand.” o The distraction was determined to be the advertising sign for a food stand. • “Driver 1 was seriously injured when he rear ended a semi that had stopped at a red traffic signal. The crash occurred near a fueling station. Driver 1 admitted to taking his eyes off the road to check gasoline prices.” o The distraction was determined to involve the driver looking at an advertising sign. • “Driver 1 rear ended Driver 2, who was preparing to turn left into a farm market/orchard. This forced the sedan into the path of the pickup, which was approaching from the opposite direction, resulting in a second collision.” o Although there is some ambiguity in the report, it appears that both the van driver and pickup truck driver were distracted by the orchard’s signage. Two narratives that indicated external distraction noted an infrastructure element. In one case, the narrative was not clear, but it was inferred that the driver was distracted by what appeared to be a nonfunctioning streetlight. The second case involved a fire hydrant: “Unit 1 was traveling w/b when they stopped at a fire hydrant that was running. They then proceeded through the water spray and struck UNIT 2.” Both were included as IRD because infrastructure was the main distractor. 6.4.4 Results The crash narrative coding yielded a disappointing number of cases where a distraction in relation to infrastructure could be inferred. Around 2.5% (n = 428) of the 17,127 fatal and serious injury crashes that occurred over the three-year analysis period (2017 through 2019) were coded as “Activity Outside Vehicle.” A review of the narratives for these crashes indicated that no evidence was available that the crash was related to an external distraction in 240 cases, which translates to 1.4% of fatal and injury crashes and 56.1% of those coded as involving an OVD. Around 1.1% of all fatal/serious injury crashes were determined to be OVDs, but there was no indication that the distraction was related to an infrastructure element. These accounted for 42.8% of crashes coded as involving an OVD. Finally, 0.03% of fatal/serious injury crashes involved both an OVD and an IRD, accounting for 1.2% of crashes coded as involving an OVD. Figure 14 shows a summary of the crash narrative coding results.

66 Figure 14. Summary of crash narrative coding results. Only five crashes were ultimately noted as involving an OVD related to an infrastructure element. As a result, no statistical methodologies could be applied. In order to demonstrate a potential type of analysis that could be utilized, data from narratives that were determined to be OVDs but not infrastructure related were assessed using text mining analyses. Text mining analyses are a collection of techniques to extract and analyze information from text. One such technique, a word cloud, depicts clusters of words in a document where the font size of each word is proportional to its usage in the document. These provide a graphical tool by which the relevance of each word can be inferred. The main category of distraction coded for the crash narratives was used to create a word cloud using those crashes determined to involve OVDs but not IRDs (n = 183). As shown in Figure 15, “Prior Crash” was the most common distraction noted in the narratives, followed by “Person” (referring to a person outside the vehicle), “Animal,” and “Weather.”

67 Figure 15. Word cloud created for crashes coded as involving an OVD but not IRD. 6.5 Outcomes and Discussion 6.5.1 Summary The purpose of this Safety Framework was to evaluate whether crash narratives could be utilized to identify IRDs. The analysis presented in this chapter used crash data and corresponding crash narratives for Michigan, which are publicly available through the MTCF data query tool. Crash narratives from 2017 to 2019 were filtered by crashes where “Activity Outside Vehicle” was noted in the crash report and a fatal or suspected serious injury was involved. This resulted in 428 crash narratives, which were manually reviewed to identify and code the source of an OVD. In over half of the crashes (56.1%), no source of distraction outside the vehicle could be identified. Around 42.8% of the crashes were determined to involve an OVD, but no infrastructure element was noted or could be identified as a distraction. This left only five crashes (1.2%) where infrastructure was determined to be a cause of distraction. 6.5.2 Performance Metrics The following assesses the utility of using the Safety Framework described in this chapter to evaluate IRDs through the evaluation of crash narratives. Feasibility of the Data Sets Utilized Crash narratives from Michigan were utilized for this analysis. They were publicly available and, as a result, were easy to access. While many states provide public or reasonably straightforward access to crash data, many states do not provide ready access to crash narratives due to the potential for exposing personal identifying information as well as the difficulty in coding narratives (if not available electronically). In some cases, access to this information requires a memorandum of understanding or other protocols. Another challenge in using crash narratives for analysis is the sheer volume of narratives that would need to be parsed in order to find sufficient information to evaluate IRDs.

68 In demonstrating this framework, only a subset of crashes coded as involving an OVD were used for evaluation. The evaluation suggested that officers do not readily consider infrastructure items (beyond outdoor advertising) as a cause for distraction. As a result, some indication of an IRD could be present in a crash narrative without the officer coding the crash as such in the crash form. For instance, an officer may note in the narrative, “Driver was viewing changeable message sign and changed lanes.” If the officer did not interpret viewing the sign as a potential distraction, he or she may not have coded the crash as involving an OVD. Consequently, a review of crash narratives not coded as involving an OVD may yield additional IRDs. For instance, crashes within 2 miles of wind turbines could be selected and the crash narratives reviewed to determine whether an indication was present that the turbines contributed to the crashes. Sample Size Only five out of 428 crash narratives that were flagged as involving an OVD were ultimately determined to be infrastructure related. As a result, obtaining a sample of 50 IRD crashes would require review of 4,280 crash narratives, while obtaining a sample size of 250 would require a review of 21,400 crash narratives. In the analysis described in this Safety Framework, crash reports were manually reviewed. The process averaged 1 to 2 minutes per narrative to review and code the desired information. While each narrative can be reviewed reasonably quickly, a tremendous amount of time would still be required to review a sufficient number of narratives to obtain a sample size of even 50 IRD crashes. Originally, the team had intended to review crash narratives for both Michigan and Iowa. When a review of three years of data for Michigan yielded only five viable samples, it was determined that an evaluation of the Iowa crash database was not a good use of further resources. Iowa has around one-third of the annual crashes of Michigan, and the state’s crash data need to be accessed at a secure terminal. As a result, further evaluation was unlikely to result in sufficient samples to conduct analyses. One method that could be utilized to facilitate the process of extracting information from crash narratives could be to conduct a document search to parse crash narratives in order to identify key words or phrases. However, the use of a keyword search does have some drawbacks. For instance, officers may use words or spellings of words that do not correspond to any of the search criteria used. Additionally, many states do not code crash narratives in text form. For instance, crash narratives are often captured and stored in PDFs, which are not text searchable. Performance of Statistical Models Due to the small sample size, no analyses could be performed on the data that included IRD. As a result, the performance of statistical models could not be assessed. However, given a sufficient sample size, statistical models similar to those used for crash analyses could be utilized. Various fields in the law enforcement reports can be used as explanatory variables, such as driver characteristics (e.g., age, gender), roadway type, crash severity, contributing factors, and weather/lighting conditions. Most states maintain crash data spatially, and, as a result, roadway information such as traffic volume, speed limit, type of shoulders, and so on could also be gathered.

69 6.5.3 Discussion The main objective of this analysis was to determine the feasibility of using crash narratives to evaluate the impact of IRDs on safety. The main advantage of using crash narratives is that crashes are a direct measure of safety, and an examination of crash narratives has the potential to pinpoint the actual cause of distraction. The review of crash narratives in this analysis suggests that it is not unusual for law enforcement reports to be erroneously flagged as involving OVDs or to lack the source of the distraction. For example, many of the crashes in the reviewed data set actually involved driver inattention or a distraction internal to the vehicle rather than an external distraction. Manual analysis allows these miscoded crashes to be removed from the data set to avoid misattribution. Many crash narratives mention infrastructure items to describe the crash location or circumstances. Manual analysis currently appears to be the most effective way to distinguish casual references to infrastructure from those where the infrastructure was a potential distractor. For example, if an officer writes, “Driver was distracted by a deer and struck a bridge abutment,” it is unlikely that the bridge was the main external distraction. Although the use of crash narratives has the potential to pinpoint actual distractions, several drawbacks to this method exist. First, IRD crashes were estimated to make up only a fraction of crashes identified as involving an OVD. In this evaluation, IRD crashes were identified in only five out of 428 narratives. The evaluation also suggests that while officers do code OVDs, they are not likely to code the contribution of infrastructure items. However, there was some evidence that officers note nonfunctioning infrastructure as a likely contributor (e.g., a broken fire hydrant or a nonfunctioning sign). The results of the crash narrative evaluation also suggest that it is unlikely that officers consider the presence of a normal infrastructure feature to have a negative impact on driver attention. If an agency wished to address this, further training for officers could be undertaken. Another drawback to the use of crash narratives is that significant resources are needed to obtain a sufficient sample size. In addition, many states do not maintain crash narratives in a searchable format (i.e. scanned pdf or image file). Additionally, crash narratives may be difficult to access due to the potential for exposing personal identifying information. As a result, the conclusion of this analysis is that use of crash narratives to identify and analyze IRD crashes is likely to be very resource-intensive and not likely to yield useful results. Rather than investigate only crashes coded as involving an OVD, an alternative strategy is to select crashes within a certain proximity to an infrastructure element of interest and review the narratives for those crashes. Officers may have provided some indication that the feature was distracting without having considered it an OVD. For instance, all crashes within 1 mile of an overhead changeable message sign could be reviewed.