A major objective of transportation agencies is to improve road safety. Accordingly, data on driver behavior and vehicle performance are a vital tool in formulating policies and designing systems to achieve that objective. The purpose of this chapter is to provide an understanding of the wide range of data sources that are potentially available to researchers for identifying factors that can reduce risk and enhance safety in the transportation network. For each source, the strengths and limitations of the data are considered, especially with respect to answering key research questions about fatigue among commercial motor vehicle (CMV) drivers. The sources reviewed include crash databases, naturalistic driving studies, and driving simulator studies. Both publicly available and private resources are described; the latter would require interaction with the data holders to be used for research purposes. Also discussed are promising data sources in development that have the potential to better inform the policy-making process of transportation agencies.
Research analyzing the relationships among hours driven, driver fatigue, and highway safety has relied on the paradigm of the Haddon Matrix (the most commonly used paradigm in the injury prevention field) to recognize factors that lead to different safety outcomes (Haddon, 1972). Safety outcomes include crashes and noncrashes, while predictors can be broadly categorized into driver characteristics, vehicle characteristics, car-
rier characteristics, and environmental factors. The data sources described below collect information on various aspects of safety outcomes and many of their predictors, but there is no single repository of these databases. Also, it is unknown whether there are identifiers in these databases that can facilitate linking across different files. This linking is essential since one must have information available for each unit of analysis, and these units are travel segments for a given vehicle. For each segment, therefore, one must have the relevant information on the predictors and the outcomes, often crashes, so that the causality of individual factors can be analyzed. Having marginal information on a potential causal factor does little good since one must know how it varies according to the other potential causal factors.
Databases containing information on CMV crashes are maintained primarily by the National Highway Traffic Safety Administration (NHTSA) and the Federal Motor Carrier Safety Administration (FMCSA). They are used to create national- and state-level estimates of CMV crash rates disaggregated by various characteristics of interest, such as fatigue. A key starting point for most of these databases is police accident reports. The following subsections describe these databases.
Fatality Analysis Reporting System (FARS)
FARS is a census file of motor vehicles involved in fatal traffic crashes in the United States. FARS data are compiled and maintained by the National Center for Statistics and Analysis within NHTSA. Data for the FARS file are collected by analysts within each state from police accident reports, death certificates, medical examiner reports, hospital reports, emergency medical services reports, state vehicle registration files, state driver licensing files, state highway department data, and other information from any follow-up investigations. Analysts use the information from these sources to code 100 data elements on certain events leading up to a crash, including crash characteristics, environmental conditions, driver distractions, circumstances obscuring the driver’s vision, and description of persons (e.g., age, gender, restraint use, injury severity) and vehicles (e.g., type, make/model, model year, cargo body for trucks) involved in the crash. Driver fatigue is coded, if specifically identified on the crash report, and is included in the data as part of a field that records driver impairments.
Trucks Involved in Fatal Accidents (TIFA) and Buses Involved in Fatal Accidents (BIFA)
TIFA was a census of medium and heavy trucks involved in fatal crashes, while BIFA was a census of buses involved in such crashes. TIFA and BIFA data were collected by the University of Michigan Transportation Research Institute (UMTRI) with support from FMCSA. Both of these data sets were based on the FARS file, supplemented by data collected by UMTRI researchers. The researchers used telephone interviews with drivers, police officers, dispatchers, emergency personnel, witnesses, and others with knowledge of the trucks or buses involved in the fatal crashes. Data collected included a detailed description of the truck or bus, the carrier and carrier operations, driver hours and compensation, and details about the initiating events of the crashes. This information was significantly more detailed than the general-purpose data collected by FARS. TIFA included data on hours driven prior to a crash and the intended trip length (which could serve as a surrogate for intended time on task). Both the TIFA and BIFA surveys were discontinued after 2010.
National Automotive Sampling System (NASS)
NASS has two components—the General Estimates System (GES) and the Crashworthiness Data System (CDS). GES is a general-purpose crash data file of motor vehicle crashes in the United States. CDS is focused on light vehicles and is not discussed further here. Unlike FARS, CDS includes all crash severities; unlike TIFA or BIFA, it includes all motor vehicle types.
The GES database is based on a hierarchical stratified sample of police-reported crashes that involved at least one motor vehicle traveling on a roadway with resulting property damage, injury, or death. GES data are used to provide national-level estimates of a comprehensive set of descriptors on crashes. GES data collectors make weekly visits to approximately 400 police jurisdictions in 60 Primary Sampling Units across the United States. Approximately 50,000 police accident reports are sampled each year. Data items in GES are coded entirely from those reports. Given the nature of the sampling structure, standard errors can be relatively large for small subsets of the data, although a significant strength is the consistency and comprehensiveness of the data.
Motor Carrier Management Information System (MCMIS)
FMCSA maintains the MCMIS crash file, a census of all trucks and buses involved in a crash that included a fatality, an injury transported
for immediate medical attention, or at least one vehicle towed because of disabling damage. Qualifying vehicles include trucks with a gross vehicle weight rating (GVWR) greater than 10,000 lb and buses designed to transport more than eight people, including the driver. States are required to extract and submit a standard set of data elements for each qualifying vehicle involved in a traffic crash that meets the crash severity threshold. The MCMIS crash file, carrier file (registration data on all qualifying motor carriers), and inspection file (data from all inspections of motor carrier vehicles and drivers) make up the MCMIS. The carrier file includes for each carrier, along with other information, the estimated vehicle-miles traveled (VMT); the number of trucks operated, disaggregated by ownership (owned, leased, trip leased); power unit type (tractor, straight truck); counts of certain trailer types; and counts of drivers, employed and leased. This information is provided by the carriers. The carriers are required to update the information every 2 years; carrier information also is updated as part of safety audits. These three different files can be linked using the carriers’ U.S. Department of Transportation (DOT) numbers.
FMCSA uses the MCMIS to monitor safety levels and identify unsafe carriers for interventions. The data are used primarily to evaluate carriers in terms of their crash performance; safety performance; and compliance with regulations, including hours-of-service (HOS) regulations. The data are less useful for detailed scientific evaluation of specific safety questions. Data elements on the individual driver and the truck/bus involved in a crash are limited; therefore, controlling for confounding factors (such as operational conditions) often is not possible in an empirical analysis. The exposure data in the carrier file are aggregated at the carrier level, so one cannot use these data to compute crash rates by environmental, road, or driver conditions. A researcher can theoretically obtain information on the driver’s condition (especially fatigue) from the police report, but doing so often is not possible in practice because MCMIS and state crash files typically include no common identifier. Carriers can challenge data on specific crashes through the DataQ Program, run by FMCSA, which allows a review of the data issued by FMCSA. Through this process, the system automatically forwards a request for data review to the appropriate office for resolution and collects updates and responses for current requests. Carrier data are checked during audits and some other enforcement activities, but only a small number of carriers are subject to these processes. In addition, FMCSA runs the Safety Data Improvement Program to help states improve their reporting. However, no systematic program is currently in place with which to evaluate the comprehensiveness and consistency of the data.
Given the diversity of the trucking and bus industries, it would be useful for FMCSA to support research that could provide a better under-
TABLE 5-1 Coverage of Crash Severity by Commercial Motor Vehicle Crash Databases
|Injury||No||Yes||If injured person(s) transported for medical attention or if any towed/disabled vehicle|
|Property damage only||No||Yes||If any towed/disabled vehicle|
NOTE: FARS = Fatality Analysis Reporting System; GES = General Estimates System; MCMIS = Motor Carrier Management Information System; TIFA = Trucks Involved in Fatal Accidents.
standing of which information is and is not included in the MCMIS. One way of doing so would be to compare the MCMIS with the sampling frame used by the Bureau of Labor Statistics’ Survey of Occupational Employment Statistics.
The coverage of crash severity in FARS/TIFA, GES, and the MCMIS is summarized in Table 5-1.
Limitations of CMV Crash Databases
The data in the FARS, TIFA, GES, and MCMIS crash databases all begin with police crash reports. When using these data, it is important to bear in mind that police officers’ primary function is protecting the public and enforcing the law, not collecting data for scientific studies. Thus driver fatigue at the time of a crash is identified either by driver admission, witness observation, or inference based on vehicle crash characteristics and driver work-rest schedule prior to the crash. Currently, there are no objective tests for fatigue that can feasibly be administered when crashes occur, analogous to tests for drugs or alcohol. Because the identification of a fatigued driver in these databases depends on the reporting police officer’s identifying and recording fatigue,1 CMV driver fatigue likely is underreported in existing crash data. An independent assessment of the reporting of fatigue in GES suggests that 1.4 to 3.1 times more crashes involve fatigue than are reported in this database (Knipling and Wang, 1995). Moreover, crash data in these databases include no information
1 The Federal Aviation Administration has developed a Fatigue Risk Assessment Tool, which is freely available on the agency’s website and is to be used by aviation maintenance workers to capture sleep- and fatigue-related inputs relevant to a maintenance incident. This tool ensures that all incident investigations follow a uniform definition of fatigue. A similar tool might be useful for the CMV industry.
about recent hours of sleep and quality of sleep obtained by the driver, and with the termination of the TIFA and BIFA projects, no data are available on hours driving since the last break.
Even with their limitations, these crash databases have their utility. First, they directly address the safety issue—that is, crashes. Second, while it is unlikely that all fatigue-related crashes are identified, these databases provide the most direct measure of the effects of driver fatigue on safety relative to hours of service.
The LTCCS was a collaborative project of FMCSA and NHTSA, aimed at collecting detailed information on crash events and associated causal factors. For this study, 963 large-truck crashes were selected using a clustered, stratified sampling procedure. The sampled crashes involved a total of 1,123 trucks and 932 other vehicles and took place between 2001 and 2003. Sampled crashes involved a least one truck with a GVWR of more than 10,000 lb and at least one fatality, incapacitating injury, or non-incapacitating but evident injury. Cases were sampled from 24 sites across the country to produce a nationally representative crash file. It should be noted that, because of the relatively small number of crashes and vehicles included in this study, standard errors for population estimates are large. Each case was investigated by a trained researcher, many experienced in the NASS CDS. In addition, a Commercial Vehicle Safety Alliance (CVSA)-trained truck inspector performed a CVSA Level 1 inspection2 of the trucks involved in most of the crashes for the study. The researchers conducted in-depth investigation for each of the crashes included in the study, encompassing scene diagrams; photographs; and extensive information on the vehicles, environment, and drivers.
The crash data collection was structured around precrash maneuvers, the critical event itself, the critical reason for the critical event, and associated factors. For details, see Blower and Campbell (2002), Findley et al. (2000), and Hedland and Blower (2006). These factors (especially the definition of critical event and critical reason, along with the nature of the associated factors) are sometimes misunderstood. The “critical event” was defined as the event that made the crash or collision unavoidable. The “critical reason” was the reason for the critical event, that is, the last failure or error that precipitated the crash. For example, in a crash in which a light vehicle turned across the path of a truck at an intersection and the truck could not evade it, the critical event would be the light vehicle’s
turning. The critical reason would be assigned to the light vehicle and would be attributed to whatever failure was responsible—for example, the light vehicle driver did not see the truck or misjudged the gap.
Fatigue was captured both as a critical reason, when determined to be the reason for the precipitating event, and as an “associated factor,” when present regardless of its contribution to the crash. In addition, the researchers coded data on hours spent driving prior to the crash and hours on duty, typically obtained from driver log books. They attempted to validate such driver log books using any other evidence available, such as the time stamps on fuel and food receipts and tolls. They also tried to obtain data on the hours of last sleep prior to the crash and the driver’s schedule for the 7 days prior to the crash. In some cases, the researchers were on the crash scene and could directly evaluate the driver’s condition. The assessment of fatigue was based on a reasonable judgment of the totality of available evidence.
Even though the data are more than 10 years old, the LTCCS is still the most comprehensive and detailed truck crash investigation data set available. (For more information on this study, see National Highway Traffic Safety Administration and Federal Motor Carrier Safety Administration, 2006a, 2006b, 2012.) The LTCCS data set has significant advantages over other crash data sets. Overall, it is much richer than the others. Data were captured for a comprehensive list of factors that could have contributed to the crashes. Also, use of the critical event/critical reason paradigm supports a flexible analytic approach (see Hedlund and Blower, 2006). Further, the researchers’ narratives provide a rich source of information. For purposes of the linkage between fatigue and crash risk, the researchers’ assessment of fatigue was better than any available crash data based on police reports (such as FARS, GES, or state crash reports).
However, it should be recognized that some of the information related to fatigue is self-reported, particularly a driver’s hours of sleep prior to the crash. Researchers had little reasonable opportunity to evaluate whether drivers actually had slept the hours claimed. In addition, since duty hours were collected from log books, even given that investigators used common techniques for verification, there was the potential for misreporting. Also, while the assessments of fatigue (i.e., driver sleep and alertness) were almost certainly better than those in other crash data sets, they were not as valid and reliable as the information that could be obtained using objective measures of sleep by wrist actigraphy and of alertness by performance on a psychomotor vigilance test. The primary disadvantage of the LTCCS data set is that the last crash in the data set occurred more than 10 years ago. Since then, many circumstances have changed, such as the HOS regulations, the ability to collect telematics information, the introduction of crash avoidance technologies, and distraction due to cell
phone usage. Further, there are no associated exposure data, so it is not possible for analysts to control for many potentially confounding factors (a shortcoming that is common to all crash data).
Crashes are the product of risk times exposure. Risk is a function of characteristics of the driver, the carrier, the environment, and the vehicle. However, crashes are also a function of exposure—that is, time or distance driven. Thus, measures of exposure are needed to determine crash risk, but valid and useful exposure data are typically lacking. The main source of national exposure data is provided by the Federal Highway Administration (FHWA). Table VM-1 in Highway Statistics provides VMT statistics by vehicle and roadway type. Trucks are classified as a single unit or as a combination (pulling a trailer). The tabulations are produced at an aggregate level (at the national and state levels by type of vehicle and roadway); they currently are not broken down by carrier type or even by time of day. This deficiency in sources of exposure data is exacerbated by the fact that the Bureau of the Census discontinued the Vehicle Inventory and Use Survey (VIUS) in 2002. VIUS was a survey of truck owners that provided aggregate travel and use data over the course of a year. FHWA is considering filling this information gap by working with the Bureau of Transportation Statistics to carry out the necessary data gathering. As part of an effort now under way, FHWA is planning to produce statistics on VMT by hour of day, day of month, and month of year by major vehicle groups.3
It is important to understand that, with the exception of the LTCCS, the available crash data are collected as part of administrative systems that exist for purposes other than answering scientific questions. Crash data are collected to monitor overall levels of safety, to enforce the law, and to allocate public resources for reducing the toll of motor vehicle crashes. Similarly, the limited exposure data available, such as VMT, are collected to monitor overall trends in highway usage and to allocate highway funds to the states. These systems were not designed by safety scientists to answer scientific questions related to HOS, CMV driver fatigue, and safety.
3 Presentation by Tianjia Tang from the Federal Highway Administration to Panel on Research Methodologies and Statistical Approaches to Understanding Driver Fatigue Factors in Motor Carrier Safety and Driver Health, September 3, 2014, Washington D.C.
Research- or study-based data sets include naturalistic driving studies (NDS) and driving simulator studies.
Naturalistic Driving Studies
Observational or naturalistic data collection is in situ, that is, “in the natural or original position or place.” In NDS, data on driver behavior and performance are collected in the normal operating environment. For truck and bus NDS, the operational environmental reflects revenue-producing operations, and study participants operate their vehicle as part of their regular job duties. Study participants volunteer to have their vehicles instrumented while they go about whatever driving they would have performed otherwise. Typically, NDS are designed such that study drivers use an instrumented vehicle for an extended period (often several weeks or longer); data are collected continuously via the instruments until the engine is turned off. Instruments such as video cameras and other sensors collect data on the position and performance of the vehicle relative to the road and other vehicles in the vicinity, as well as driver behavior. The final data set provides an “instant replay” of the entire driving trip, including any incidents, allowing researchers to focus on event factors including driver behavior and crash precursors. NDS have been conducted with light vehicles (e.g., the 100-car study [see Dingus et al., 2006]), long-haul trucks (e.g., Hanowski et al., 2009), local/short-haul trucks (e.g., Hanowski et al., 2003), buses (ongoing), motorcycles (Williams and McLaughlin, 2013), and bicycles (Dozza et al., 2013).
Early NDS, such as the 100-car study (Dingus et al., 2006) and local/short-haul study (Hanowski et al., 2003), involved a relatively modest number of instrumented vehicles—100 light vehicles in the 100-car study and 4 trucks in the local/short-haul study. However, as instrumentation costs have decreased and the robustness, capacity, and ease of installation of data collection systems have improved, newer studies have included much larger numbers of vehicles. For example, the Transportation Research Board’s Strategic Highway Research Program (SHRP) 2 involved the instrumentation of 3,353 vehicles and data collection from 3,546 drivers (McClafferty et al., 2015). In total, the SHRP 2 NDS resulted in approximately 50 million miles of continuous driving data.4 Recent and ongoing NDS with trucks and buses have similarly increased in scope relative to early NDS. For example, an NDS involving commercial trucks
4 SHRP 2 Products Chart. Available: http://onlinepubs.trb.org/onlinepubs/shrp2/SHRP2ProductsChart.pdf [April 2016].
and buses funded by FMCSA, soon to be concluded, has been collecting continuous naturalistic data from 161 trucks (169 drivers) and 43 buses (68 drivers). These large NDS data sets have resulted in tremendous research opportunities and have led to the development of new online databases in which the NDS data can be accessed. The most recent of these databases that have come online include data from the SHRP 2 study;5 at the InSight Data Access Website, some of the SHRP 2 data can be accessed for analyses.
Operational definitions of driving segments can vary across studies, and the trigger criteria used to identify them also can differ based on the research question of interest. In this way, NDS that use continuously recorded data offer the flexibility to customize search criteria to enable identification of events of interest to include in the analyses. With respect to the duration of interest, for example, analyses outlined in the 100-car study (Dingus et al., 2006) focused on a time frame of 6 seconds. By contrast, more recent analyses of light vehicle data focused on assessing driver behavior prior to an event trigger have reviewed video and other driving data 12 seconds prior to the event (Victor et al., 2014). Analyses of metrics associated with driver fatigue (e.g., PERCLOS, or the percentage of closure of the driver’s eyelids over some time period) have used much longer time frames for video review; for example, Hanowski and colleagues (2013) reviewed 3 minutes of video prior to an event trigger to assess PERCLOS.
As defined for the 100-car study and the SHRP 2 study (Guo et al., 2010; Victor et al., 2015), a crash event is “any contact that the subject vehicle has with an object, either moving or fixed, at any speed in which kinetic energy is measurably transferred or dissipated,”6 while a near-crash event is “any circumstance that requires a rapid, evasive maneuver by the subject vehicle, or any other vehicle, pedestrian, cyclist, or animal to avoid a crash.”7 Trigger criteria include a host of kinematic variables that act as filters by which a driving segment is selected from the video data for further review. Within a study, the objective is to select trigger criteria that reduce the chances of false alarms. Using different trigger criteria, various sets of events of interest can be captured.
Analysis of these large data sets has required new, and evolving, analytical approaches. Often, the data are characterized by events of interest (e.g., “crashes”) and compared with baseline, or normative, driving.
6 Examples of objects include other vehicles, roadside barriers, and objects on or off of the roadway, pedestrians, cyclists, or animals.
7 A rapid, evasive maneuver is defined as steering, braking, accelerating, or any combination of control inputs that approaches the limits of the vehicle capabilities.
In this way, researchers can study the pre-event behaviors of the driver during the time of interest leading up to the event. In a similar way, the analysis can use nonevent data for comparison. In the SHRP 2 data set, 1,541 crashes were recorded, and although these severe outcomes can serve as one event type of interest to study, less severe events that may involve underlying factors of interest (e.g., driver distraction, speeding, rural roads) also can be studied by examining other safety events with outcomes less severe than crashes. For example, many NDS studies analyze incidents often referred to as “safety-critical events” (SCEs), which may include near-crashes and other driver errors (e.g., unintended lane deviations). These event types, in addition to matched nonevents that serve as a baseline for comparison, can be a focal point for video and data review. Further, as the flexibility of the above definition of triggers suggests, one can and should examine the robustness of any inference to the definition of the trigger one is using (which in turn defines the length of the period prior to an event on which one collects data).
The strengths of NDS as an approach for data collection are as follows:
- Ecological validity: Data are collected on driver performance and behavior, as well as vehicle behavior, including the factors that can jointly cause a traffic crash. NDS can provide a way of observing all of the relevant factors acting jointly as they might be when the vehicle is not instrumented. Assuming that the instrumentation has no impact on driver behavior—there is some evidence to suggest this is the case after only a few days of use of the instrumented truck—one gets to observe that behavior in a true operational environment. This NDS advantage provides very high ecological validity. There is no pressure for a driver to behave unnaturally since the experimenter is not present physically, and has given no instructions on how to drive, which roads to travel on, how to deal with excess traffic, or how to handle the environmental conditions encountered. In addition, in NDS, one can observe the results of typical driver pressures to, for example, make appointments; driver choice behavior (e.g., where to purchase fuel, whether to use caffeinated or tobacco products or eat or drink alcohol or other beverages shortly before or while driving whether to drive in adverse weather conditions); the time of day driving starts; the timing of waking and sleep (onset-offset times, duration) in advance of driving; and the like. All such decisions are up to the drivers under study, and drivers are allowed to behave as they would routinely.
- Reduced errors in data collection: Since the vehicles are instrumented, NDS provide precise, detailed information on driver
behavior and driving performance for occasions that both do and do not lead to crashes or SCEs. Review of video and other data allows for “instant replay” capabilities that make it possible to better understand the genesis of crashes and SCEs. Data are collected not only on crashes but also on events that under some circumstances are likely to be associated with a higher chance of having a collision. Such SCEs may include excessive lane changes, hard braking, tailgating, and speeding. Data also are collected on the behavior of other vehicles in the vicinity.
- Case control and exposure: One can easily carry out case-control studies using data from NDS. An important benefit of NDS data collection is that data exists for travel times when crashes or SCEs did not happen.8 Since data are available for both crashes and SCEs, one can match a situation in which a crash or SCE occurred with an analogous situation in which one did not occur, and then examine the frequency of various possible causal factors to see whether it differs between the two situations. The data contain precrash and pre-SCE information on driver behaviors including the presence of fatigue or distraction and interactions with passengers, devices, and the vehicle.
- Utility: Finally, while NDS are expensive to conduct, they typically involve collecting data for months or longer, which allows for analyses investigating many research questions.
NDS also have limitations, although some of these can be addressed:
- Events of interest: In NDS the investigator has very little control over the driver, and therefore over the driving situations encountered. Thus one cannot study specific problems as intensively as is possible with simulators since those problems may not occur sufficiently often during the course of the study. This limitation is a standard by-product of high ecological validity. Also, since crashes are rare events, analysts tend to rely on SCEs. However, the validity of treating such events as near-crashes is uncertain. This is clearly less of a limitation with larger NDS such as SHRP 2, which included 701 crashes as of June 30, 2014 (Antin et al., 2015, Table 4.2). Also, relying on some SCEs may not be a serious problem in every situation, as analyses linking crashes to near-crashes have been conducted (see Guo et al.  for an example). NDS also can be used to conduct field operational tests for technology evaluation (e.g., Hickman and Hanowski, 2011).
8 SCEs can also provide information on how drivers successfully avoided crashes.
- Feature extraction: One of the reasons SCEs are defined as high-g-force events is that one can then focus attention on a small fraction of the data collected. The problem of feature extraction, if one suspects other causal factors in addition to these kinematic events, either requires going carefully over thousands of hours of video data capture or finding some other way to identify those moments. However, this may not be an issue for certain analyses that may look for events of interest other than SCEs. For example, if speeding over a set threshold (e.g., 80 mph) were of interest for analysis, search criteria could be set to identify all instances of high-speed events.
- Causality: NDS are observational studies and as such are subject to traditional limitations in their capability to address causality. In most cases, researchers do not have detailed information about driver state (fatigue, alcohol or drug use, and medical conditions that impair driving), which could introduce confounding factors.
- Generalizability: Because little information is available to identify differences in the characteristics of drivers of instrumented and noninstrumented vehicles, it may not be possible to determine whether such differences might confound any observed relationships. A related issue is that the subjects are volunteers, and it is reasonable to expect that those who engage in activities that affect their driving negatively would be less likely to participate. However, being a volunteer is an inherent component of almost all research involving human subjects, so the same limitation is inherent in all empirical studies. Therefore, extrapolating the results of NDS to the general trucking population is not entirely feasible, especially in the absence of baseline estimates, although NDS drivers can be compared demographically with other populations of drivers.
- Sample size: As noted previously, early NDS often had small sample sizes, although it is important to note that the same is true of many other empirical studies. However, the number of drivers in a small sample limits the variability one can observe in terms of fleet size, operation types, and vehicles. Again, this limitation is not unique to NDS; surveys and studies based on paper logs, for example, are restricted to certain segments of the CMV population. Achieving a study sample that is representative of the population is hampered by the fact that estimates of characteristics of all CMV drivers are not available. On the other hand, NDS provide an opportunity to study the same drivers over many months of driving.
In some cases, these limitations of NDS can be addressed through the use of additional data sets or modeling approaches. An example is the inability to assess driver sleep hours. Drivers in the Hanowski et al. (2007) study wore actigraphy devices that recorded their sleep quantity (timing and duration). These data were then merged with the driving data to better understand the impact of sleep on driver behavior and risk. Other data streams and data sets—such as data from geographic information systems, including data on weather and traffic density—can be combined with naturalistic driving data (see, e.g., Cannon and Sudweeks, 2011).
Driving Simulator Studies
Driving simulators are useful for research purposes as they allow a researcher to observe driver behavior under certain conditions while controlling for others. They are especially useful for the study of high-risk conditions, such as fatigue and other types of impairments, because performance errors do not result in injuries or fatalities. However, that fact also can limit the realism of the simulator experience. Simulators are used as well to test potential road and vehicle improvements and other technologies before they are deployed and obtain practical feedback from drivers. For example, simulators guided the design of advanced driver support systems by making it possible to estimate the marginal effect of their deployment on driver performance. Professional driving simulators also can serve as a training tool for individuals learning to drive trucks and buses. Depending on their fidelity to an actual truck or bus, they can provide approximate on-the-job training for a driving student and possibly reduce crash risk. Simulators are useful as well for investigating specific issues, such as whether fatigue is associated with unintended lane changes on curved highways, and for recognizing factors that need to be examined in broader studies. One can look at a variety of outcomes by repeating the experiment with a study participant, thereby controlling for individual differences. In addition, a researcher can investigate low-frequency/high-severity events such as a crash by repeating the same scenario without adding substantial costs. Not only are driving simulators useful for testing technologies but they also may have utility for testing various types of schedules to identify which is optimal for truck and bus drivers—although the lack of fidelity to the real world may limit the utility of this application. Finally, simulators give researchers the flexibility to gather additional information such as physiological data (e.g., heart rate), although more miniaturized portable physiological monitoring with data uplink is rapidly becoming possible, so this advantage is becoming less clear.
Simulator studies are limited in the sense that their results are difficult to generalize, and the estimate of an impact of a factor will shift upward or downward depending on how it interacts with other factors in the real world (that were fixed or controlled for in a simulator study). Finally, as the same experiment is repeated with a participant in a simulator study, the effect of learning becomes a confounding factor.
Proprietary data include data collected by the American Transportation Research Institute (ATRI) and by large truck carriers.
Data Collected by the American Transportation Research Institute
ATRI is a member of the American Trucking Associations (ATA) and is a not-for-profit research organization. ATA is organized as a federation of independent state motor carrier associations, councils, and committees, and ATRI was established as a separate organization to maintain its independence from ATA. In essence it is the research arm of ATA, with its own management (a separate board of directors consisting mainly of ATA members) and funding, which comes from the trucking industry. ATRI conducts studies in various topical areas relevant to the trucking industry and therefore collects data on drivers and motor carriers. It conducted surveys on fleet managers and drivers and also made use of logbook data from participating carriers to carry out a study on operational and economic impacts of the 2011 HOS regulations. ATRI has data on technology penetration among carriers, as well as video feeds from trucks and other ad hoc data for its own or government research purposes. It also collects insurance data on insurers or carriers who provide fleet-wide insurance, and this data set includes information on carriers’ crash involvement and costs associated with each crash. Along with cross-sectional information, ATRI has undertaken initiatives to collect real-time data that offer the potential to address questions related to violations, scheduling, and traffic patterns.
Since 2002 ATRI, working closely with FHWA, has led the Freight Performance Measures (FPM) Program, which evaluates the effectiveness of the highway system in facilitating fast, efficient movement of goods. Performance measurements are produced for this program through the use of real, anonymous, private-sector truck data sourced through unique industry partnerships. ATRI’s FPM database currently contains billions of truck data points from several hundred thousand unique vehicles spanning more than 7 years. Currently the program collects nearly 100 million data points per day, and it exceeded 1 billion points per week in late
2014. The data, which include periodic time, location, speed, and anonymous unique identification information, are collected and used by ATRI researchers to produce various indicators on truck movement, highway bottlenecks, crossing time and delay, and demand for truck routes and facilities on highways. Knowing the location of a truck or bus prior to a crash, one can estimate the number of hours driven. Also, the FPM database is a valuable source of exposure data; one can use the data to arrive at an estimate of the number of trucks on the road by time of day. However, a key limitation is that the data are collected under a strict confidentiality agreement and so are not currently available to researchers. Also, the data are collected only for a modest fraction of all trucks and buses, which may not be representative of the industries as a whole.
Data Collected by Large Truck Carriers
Most truck carriers collect information on their drivers for bookkeeping and operations management purposes. However, large carriers also often collect information on their drivers’ health, wellness, crash rates, and the like. Some carriers use such data to conduct studies for purposes of improving their safety performance.
Carrier-based data can include both events and exposure, addressing the exposure problem that exists for all trucks collectively, although the data are restricted to drivers on the carrier’s payroll. Also, crash data can be linked to personnel/work records, as well as to equipment manifests. Some carriers have relatively sophisticated data collection programs with respect to loss events, similar in construction to public crash files. These loss files must be used with care because they include incidents beyond police-reportable or MCMIS-reportable traffic crashes.
As part of its information-gathering process, the panel heard from Schneider National (a multinational trucking company) about the types of data it collects as part of its safety and health and wellness initiatives. This information illuminates the data elements that large truck carriers may collect on their employed drivers. Schneider collects information on crashes, which includes the time of day, the number of hours since the driver’s last break, and the location of the crash and the roadway type where it occurred. The company also has electronic log data from the trucks in its fleet that track shift variability and the number of days since the truck was last at the home terminal. Data on critical events such as hard braking, roll stability control, and collision mitigation also are collected. The company prescreens its drivers for sleep apnea and treats those who test positive. As a result, the company has data not only on the safety performance of its drivers but also on their health. One outcome that potentially could be
studied with these internal company data is the impact of proposed health and wellness programs on driver fatigue and safety.
Other data sources include driver paper logs, inspection reports, and surveys.
Driver Paper Logs
Researchers exploring the relationship between driving time and the probability of crashes have collected driver logs from private fleets. In these cases, the carrier typically supplies the driver logs for a particular time period (e.g., 2 weeks), which contain information on driving patterns. This method of data collection relies on establishing collaborative agreements with private fleets. The analyses rest on valid and reliable reporting by truck drivers. A researcher obtains access to exposure data; therefore, studies based on driver logs can have a case-control formulation. As driver logs provide information on sleeper berth time and the number of rest breaks taken by a driver, researchers can investigate the impact of change in various provisions of HOS regulations provided they have data on the same set of drivers before and after such a change (Jovanis et al., 2011). The resulting analysis will be difficult to generalize given that representativeness is an issue, but the association between different driving patterns and crash risk can be estimated. Even if the study sample based on paper logs will not be random or fully representative of all truck and bus drivers, the relationship between hours of service and crash frequency may be generalizable to the CMV driver population.
Another limitation of paper driver logs is that the data are self-reported. Falsification of log books is a possibility (Moses and Savage, 1996). Monaco and Williams (2000) found that 57.8 percent of drivers in the University of Michigan Trucking Industry Program (UMTIP) 1997 data set reported driving more hours than they entered in their logbooks in the last 30 days; and a large proportion of all drivers (82.58%) said that, in general, they thought logbooks were inaccurate. Although there are no known national estimates of the prevalence of the practice of falsification of logbooks, better technology, such as electronic on-board recorders and electronic logging devices, could help address some of these quality concerns.
A state inspection system nationwide conducts more than 3 million roadside inspections of commercial motor vehicles annually to ensure
that trucks and buses are operating safely. The selection of vehicles for inspection typically is not random; enforcement officers must have probable cause to inspect a vehicle. Trained inspectors in each state inspect vehicles using criteria developed by CVSA. Inspection involves an examination of the vehicle and/or driver to determine compliance with FMCSA regulations.9 As part of the most comprehensive level I inspection, a driver’s certificate from his or her medical examiner is checked, as is the driver’s record of duty status and hours of service. Drivers also are checked for visible signs of fatigue. If the vehicle and/or the driver is in violation of FMCSA regulations, the vehicle and/or driver may be placed “out of service.” An example of a vehicle violation is “oil and/or grease leak,” while an example of a driver violation is “failing to use seat belt.”
There were 3,497,937 roadside inspections in 2013 (Federal Motor Carrier Safety Administration, 2014, Table 2-5). The total numbers of vehicle and driver violations in that year were 4,118,869 and 1,047,496, respectively (as a result of some vehicles having multiple violations). Among the driver violations, 51,911 were related to driving beyond the 14-hour duty limit and 28,207 to exceeding the 11-hour driving limit. Approximately 44 percent of drivers involved in either of these types of violations were placed out of service (Federal Motor Carrier Safety Administration, 2014, Table 2-7).
Inspection reports and the MCMIS inspection file are used to identify trucking firms that are performing poorly on safety parameters. For each trucking firm, information is available on total roadside inspections, how many of its trucks were placed out of service, and the categories of the violations. This information can be used to identify sets of risk factors (at the trucking firm level) likely to characterize violators in a certain category. Moses and Savage (1996) used MCMIS roadside inspection data on 20,000 trucking firms to predict the firms’ accident rates. Their empirical analysis did not control for driver characteristics. As mentioned above, data elements in the MCMIS database on the individual driver and truck/bus involved in a crash are limited; therefore, controlling for many confounding characteristics is not possible. Another limitation of inspection reports is that they include no direct measurement of driver fatigue. Nor do they provide data on vehicles that are not subject to inspection. Although not based on inspection reports, attempts have been made to identify driver- and firm-level risk factors for HOS violations using survey data. These risk factors include scheduling of irregular routes, trip lengths, compensation schemes, availability of rest areas, and type of load (Beilock, 1995; Braver et al., 1992).
Useful surveys have occasionally been conducted among truck and bus drivers to gather data on their work schedules, fatigue levels, health status, access to health care services, and participation in health promotion programs. Among the surveys conducted on truck and bus drivers, two are highlighted here—one on truck drivers and the other on bus drivers—as they were most comprehensive in terms of the data elements collected.
National Survey of Long-Haul Truck Driver Health and Injury
The National Institute for Occupational Safety and Health (NIOSH) conducted a survey of long-haul truck drivers to gather baseline data on their health and safety, including the prevalence of selected health conditions and risk factors. Data were collected from 1,670 long-haul truck drivers at 32 truck stops in 20 states. Survey teams were present at each truck stop for 3 days, at varying times each day. The survey results produced estimates of the prevalence of obesity, cigarette smoking, and diabetes. The questionnaire included items on health insurance coverage and drivers’ self-perception of health status. As the survey was conducted at truck stops, the truck driver population that was interviewed for the survey comprised only long-haul drivers, and excluded drivers who deliver goods locally. It is difficult to know whether the survey was strongly unrepresentative of all drivers given that there are no baseline health data on all truck drivers in the United States. The approach of recruiting truck drivers at truck stops was reasonable since it enabled the survey to reach a quasi-random population of long-haul drivers. This may be the best option available for surveying this population given its high mobility, and is preferable to interviewing drivers belonging to a specific trucking company/companies given the selection bias implicit in that approach. Mail questionnaires, phone interviews, and web-based surveys can be impractical given that CMV drivers often are away from home and have unpredictable work-rest hours.
Bus Driver Fatigue Study
The Sleep and Performance Research Center at Washington State University conducted a month-long survey from August 2010 through August 2011 of 84 commercial bus drivers (middle-aged, overweight, and predominantly male) to determine whether these drivers were working within the HOS limits set by FMCSA (for details, see Belenky et al., 2012). The bus driver population is easier to survey than the truck driver popula-
tion given that most of them have regular schedules. The survey collected information on the length of various components of duty cycles. Drivers were told to keep a sleep-wake diary, and additional information on sleep-wake history was collected via actigraph. Drivers also were administered a psychomotor vigilance test (PVT) of behavioral alertness (Basner and Dinges, 2011) when they started and ended their work duty and when they went on and off on breaks during duty. Information was collected simultaneously on subjective fatigue using spontaneously perceived sleepiness and on sleepiness using the Karolinska Sleepiness Scale. The participating bus drivers represented charter (18 drivers), tour (13), regular route (25), or commuter express (24) operations; it is not known whether this percentage distribution was representative of the bus driver population.
In Chapter 10, the panel argues for the need to account for all major causal factors in analysis of the causal relationship among HOS regulations, fatigue, and crash frequency. Causal factors are categorized into characteristics of the driver, the vehicle, the carrier, and the environment. This section suggests a number of important factors related to the operational characteristics of trucking for which high-quality data are not regularly collected. All these factors are important for studying driver fatigue. Collecting this information would help researchers control for confounding factors when analyzing the relationship among fatigue, hours of service, and crash frequency:
- Exposure per hour of day: Even though the MCMIS provides a comprehensive list of registered trucks in a particular year, the number of trucks on the nation’s roads during a particular hour on a single day is unknown. This information is vital to provide a normalizing factor. Without this information, one cannot compute crash rates by time of day.
- Trip length and driving hours: Driving time and time on task (which encompasses driving time plus loading and unloading) are important predictors of crash risk and driver fatigue. An ATRI report on the safety impacts of HOS regulations, which consists of analysis based on TIFA data, states that 80 percent of fatal truck collisions in 2007 occurred within the first 8 hours of driving (American Transportation Research Institute, 2010, Figure 2). The importance of this percentage is difficult to assess without knowing, from exposure data, the percentage of total driving this figure represents. Figure 5-1 shows the distribution of all medium and
- Diversity of route and load: Loads range from various types of freight, to liquids, to agricultural products (e.g., livestock, produce), to all sizes and types of machinery, to construction materials and equipment, to hazardous materials, to intermodal
heavy trucks involved in fatal crashes from 2004 through 2009 by driver hours behind the wheel at the time of the crash. Although the data are restricted to fatal crashes, and driver hours behind the wheel are unknown for a large percentage of fatal crashes, the figure shows a wide distribution of hours driven by truck drivers involved in fatal crashes. It is difficult to generate a distribution of hours driven by the whole population of truck drivers for comparison as that information is not available. The lack of such exposure data makes it difficult to calculate crash rates by hours driven.
containers, to small packages, to food products, and many more. These various commodities are driven long-haul, regionally, or short-haul in rural, suburban, and urban settings. The road types and conditions vary widely across the country, and from local streets and arterials to Interstate highways.10 Useful information about specific routes and environments might include the occurrence of various kinds of precipitation, the degree of visibility, the road surface, the roadway geometry, the degree of congestion, whether the route includes work zones, and the like.
- Regular versus irregular schedules: As described in Chapter 2, depending on the type of operation, drivers may have regular or irregular schedules. Irregular schedules are more likely to hamper sleep patterns and in turn affect productivity at work or safe operation of the vehicle.
- Loading and unloading: As discussed in Chapter 2, some drivers have loading and unloading as part of their duty, while others do not. A physically demanding task, it exposes drivers to injuries and may cut into their sleeping time. It also may increase fatigue and thus the probability of fatigued driving.
- Operations at night (i.e., against circadian rhythm): HOS regulations include restart and sleeper berth provisions whose purpose is to enable the driver to sleep during the nighttime (see Chapter 4). As described in greater detail in Chapter 3, human physiology makes individuals inclined to sleep during night hours, and drivers are no exception. Nighttime driving heightens the risk of fatigue.
- Sleep quality: Information is needed not only on hours of sleep received by the driver but also on the quality of that sleep.11 Did the driver sleep in his or her berth or own bed at home? Sleeping in a berth may not be as comfortable and relaxing as sleeping in a bed. Darwent and colleagues (2012) investigated sleep obtained by long-haul truck drivers in Australia and found minor differences in the quality of sleep obtained in a sleeper berth versus at home.
10 This report does not specifically consider off-road operations.
11 The term “quality of sleep” is used here in the same way it is used in sleep medicine and sleep research—as the basis for a subjective complaint related to symptom reports of difficulty with sleep initiation, duration, and consolidation and daytime impairment (Buysee et al., 1989). The panel believes that the relevant scientific and medical professional sleep societies are in the best position to develop consensus definitions of the important sleep-related measures for the public. This report relies on those professional definitions and the methods for assessing them in CMV drivers relative to research, policy making, enforcement, and accident investigation.
- Method of compensation: Information is needed on how a driver is paid, as compensation methods can be a confounding factor. How methods and levels of compensation affect safe operations is unknown, although there is some evidence (Belzer et al., 2006) that higher levels of compensation attract and retain safer drivers.
The future of motor vehicle transportation in general and commercial truck and bus transportation in particular is in flux. With the growing use of unobtrusive on-board cameras and sensors, motor vehicles soon will provide much greater assistance to the driver, and fully autonomous vehicles are now being tested. On-board technologies available today can generate extensive amounts of data related to many aspects of a vehicle, its driver, and the surrounding environment. Although these devices have the potential to identify fatigued drivers and possibly prevent them from continuing to operate their vehicle while they are impaired, liability, privacy, and security concerns will shape the future use of these technologies, as well as of the extensive real-time data they generate. Setting aside for now the privacy and confidentiality aspects of these technologies, this section briefly describes potential new data sources and the potential capabilities such data may provide to researchers in this area.
Electronic On-Board Recorders or Electronic Logging Devices
In the next few years, many carriers and owner-operators will either decide on their own or have imposed on them the obligation to carry electronic on-board recorders or electronic logging devices (ELDs) to measure, at a minimum, when and where a truck was in operation and for what duration. From a research perspective, electronic data on hours driven by a driver prior to a crash represent valuable information. Also, the expectation is that ELDs are likely to increase compliance with the HOS regulations because compared with paper logs, they are likely to be more tamper-resistant (albeit not entirely). One may speculate that they therefore could serve to reduce the extent of fatigued driving.
In 2010, FMCSA published a final rule on mandatory installation of ELDs on commercial motor vehicles manufactured after June 4, 2012. In August 2013, the Seventh Circuit Court rendered judgment that the agency could not proceed with the rule as it failed to consider driver harassment. In 2014, FMCSA proposed amendments to the rule that included requirements for the mandatory use of these devices by drivers currently required to prepare HOS records of duty status, as well as measures designed to address concerns about harassment resulting
from the mandatory use of ELDs. This proposed rule is still in the comment stage. Therefore, ELDs are not currently mandatory, and the fact that industry-wide adoption of these devices will take time means that researchers using such data now will be analyzing only a selected subset of the driver population.
Telematics, On-Board Safety Systems, and Other Monitoring of Drivers
Trucks and buses increasingly are being wired for purposes of vehicle control and monitoring, as well as supervision and management of truck fleets. Telematics includes technologies for locating and tracking the vehicles (GPS) and for communicating with the driver and monitoring vehicle performance remotely. Sensors transmit a continuous stream of data from the vehicle to a central data warehouse. Depending on the kind of device installed in a truck, the vehicle’s performance and condition may be captured. As the data are collected in real time, the dispatcher can warn the driver of potential problems. Vehicle data potentially available for monitoring unsafe driving include hard-braking events, sudden accelerations, and speeding. These events can be recorded for later review or communicated in real time to dispatchers. Installing such technologies is currently cost-effective for big carriers and companies that manage large fleets. The resulting data are proprietary. Data captured by telematics devices also will differ among carriers depending on their requirements.
A host of on-board safety systems are available, such as electronic stability control, roll stability control, lane departure warning, blind spot warning, forward collision warning, adaptive cruise control, and collision mitigation braking systems. These systems warn the driver of dangerous conditions, and some can be programmed to take corrective actions automatically.
On-board safety systems that use sensors and video recording systems record events outside and inside the truck. These data can be used to identify unsafe driving practices, and potentially driver fatigue. Carriers can use these data to identify at-risk drivers and develop coaching and training programs. Carriers can choose among many on-board safety systems. A recent survey conducted by UMTRI of a random sample survey of the entire fleet of trucking companies (drawn from the MCMIS) asked companies the factors that determined their choice from among a list of on-board safety systems. Companies considered the proven safety benefits of the technologies, positive feedback from drivers, driver improvement, improved safety culture, reduced cost of accidents, and insurance benefits (Belzowski et al., 2007).
Despite the lack of uniformity in the data elements trucking companies obtain from these systems, it is safe to conclude that, in combination with electronic on-board recorders, they offer the potential for a dramatic increase in the amount of driving information available, including information on hours spent driving and driving behavior.
The data sources described in this chapter have their advantages and limitations (as summarized in Table 5-2). The appropriate data source depends on the research question being pursued. No one source collects
TABLE 5-2 Advantages and Limitations of Different Data Sources
Provide details on driver characteristics, vehicle characteristics, road conditions, and weather conditions
Generate aggregate crash statistics
Restricted to U.S. Department of Transportation (DOT)-reported crash events; lack exposure data
Difficulty of identifying driver fatigue from crash reports since the data are collected by nonresearchers
|Naturalistic Driving Studies||
Assess driver and vehicle performance under actual road conditions
Provide exposure data
Crashes are a relatively rare event; aspects of data reduction are done manually
Replicate experimental road conditions, which enables testing various scenarios
Can be used to quantify performance profile of drivers who suffer from various medical conditions
Enable assessment of relative validity but not absolute validity
|Electronic On-Board Recorders, Electronic Logging Devices, On-Board Safety Systems||
Identify unsafe driving practices and at-risk drivers
Different set of technologies oriented toward different factors related to safety
|Real-Time GPS Data||
Provide exposure data
SOURCE: Adapted from Rizzo (2011, Table 2).
comprehensive information on outcomes (crashes, noncrashes) and predictors (driver characteristics, vehicle characteristics, company characteristics, and environmental factors). Therefore, combining data from these various sources may be advantageous. The challenges entailed in doing so include the following: (1) some data sources are proprietary and would require collaboration of the data holders; (2) each data source has its own (potentially proprietary) set of definitions and taxonomies for outcomes and predictors, which would result in multiple measures of the same variable; and (3) identifiers or linking variables are very likely not available.