The relationship between driving volatility in time to collision and crash-injury severity in a naturalistic driving environment
Introduction
The 2018 Traffic Safety Facts published by the National Highway Traffic Safety Administration (NHTSA) reported a total of 36,560 traffic fatalities and an additional 2,491,000 injuries in the U.S. (NHTSA, 2018). Of all the fatalities and injuries sustained by vehicle occupants, drivers sustained 75% of fatalities and 72% of injuries (NHTSA, 2018). As a result of extensive research over the decades (Mannering and Bhat, 2014), a broad spectrum of factors are known to be associated with the injury severity outcomes of drivers including drivers’ characteristics, crash and roadway factors, vehicle features, weather, and environment-related factors (Kockelman and Kweon, 2002, Quddus et al., 2002, Abdel-Aty, 2003, Zajac and Ivan, 2003, Khattak and Targa, 2004, Mooradian et al., 2013, Behnood and Mannering, 2015, Ahmad et al., 2019b). Driving behavior, however, is a leading contributor to the occurrence of crashes and the injuries/fatalities therein. A better understanding of driving behavior prior to involvement in a crash is fundamental to the design of behavioral countermeasures. Thus, an analysis of the behavioral correlates of crash-injury severity has been of interest (Abdel-Aty, 2003, Paleti et al., 2010, Zhu and Srinivasan, 2011). Primarily, the focus has been on what is referred to as “aggressive” driving (such as driver was speeding, tailgating, improper lane changes, making obscene gestures, and so on), and its correlation with injury outcomes1 (Nevarez et al., 2009, Richards and Cuerden, 2009, Paleti et al., 2010, Weiss et al., 2014). By using “aggressive” driving as a latent construct, Paleti et al. (2010) quantified the moderating effect of aggressive driving in increasing injury severity outcomes (Paleti et al., 2010). Likewise, as surrogates of driving behavior, higher speeds or speed limits are known to be correlated with higher injury severity outcomes (Duncan et al., 1998, Klop and Khattak, 1999, Renski et al., 1999, Abdel-Aty, 2003, Quddus et al., 2009, Weiss et al., 2014).
In the majority of the literature, crash causation studies or police crash reports have been used to gain an understanding of the relationships between crash-injury outcomes and driver-specific behavioral factors (Paleti et al., 2010, Savolainen et al., 2011, Mannering and Bhat, 2014). As acknowledged in the literature (Paleti et al., 2010, Mannering and Bhat, 2014, Imprialou and Quddus, 2019), classifying “aggressive” driving based on information (such as speeds, maneuvers, etc.) in police crash reports is a subjective process and there exists the possibility of misclassification. Also, the extent to which the speed information in police crash reports, typically used as a measure of driving behavior, is accurate is unclear. Importantly, while analysis of such a nature has helped to formulate actionable strategies for development of behavioral countermeasures, it does not capture the real-world instantaneous decisions or actual driving tasks immediately prior to a crash (Kim et al., 2016). Having said this, a deeper analysis is required to gain an understanding of how real-world microscopic driving decisions (e.g., speed, accelerations, vehicular jerk, etc.) in time to collision correlate with crash-injury outcomes. With the advent of naturalistic driving data, such an analysis is now possible.
Thanks to maniaturization and the rapid advancements in ambient sensing related technologies, the combination of sensing, computation, and communication now allows collection of great amounts of spatiotemporal behavioral data in an unintrusive manner. With the integration of communication technologies with video and radar surveillance, enormous magnitudes of real-world contextual driving data are now easily available (Campbell, 2012, Henclewood, 2014). The real-world large-scale driving data generated by advanced technologies are not informative to drivers in the raw form (Khattak and Wali, 2017). However, the raw information can be transformed into useful and actionable knowledge with appropriate data mining techniques – allowing a deeper and richer understanding of instantaneous driving decisions (Khattak and Wali, 2017). Important in this regard is the concept of “driving volatility” that captures the extent of variations in driving, especially hard accelerations/braking and jerky maneuvers (Liu et al., 2015, Liu et al., 2017, Wang et al., 2015b, Liu and Khattak, 2016). As a key measure of driving performance – it characterizes extreme behaviors and variations in real-world driving decisions (Liu and Khattak, 2016, Khattak and Wali, 2017). Compared to traditional surrogates of driving behavior (such as speed and driver demographics), the idea of driving volatility allows the development of proactive and more personalized warning and control assist systems2 (Liu and Khattak, 2016, Khattak and Wali, 2017).
An in-depth examination of short-term driving decisions in time to collision can shed light on the actual mechanism in which a vehicle is maneuvered or operated before a crash. While previous research used rigorous data mining techniques to characterize volatility in driving behaviors (Liu et al., 2015, Liu et al., 2017, Wang et al., 2015b, Liu and Khattak, 2016, Khattak and Wali, 2017), the real-world driving volatility metrics were not linked with unsafe outcomes. To this end, the concept of “driving volatility matrix” enables conceptualization of the extent of variations in driving behaviors at multiple levels of the transportation ecosystem and integrates real-world driving volatility with safety–critical events (e.g., crashes/near-crashes) (Wali et al., 2019b). Along these lines, previous studies extended driving volatility concept to specific locations and proposed a methodology for linking high frequency microscopic connected vehicles driving data with police-reported crashes (Kamrani et al., 2017, Wali et al., 2018b). Likewise, Kim et al. (2016) analyzed the links between micro-scale driving behavior and rear-end crash propensity in an exploratory manner (Kim et al., 2016). Extreme rates of deceleration were correlated with crashes occurring at signalized intersections (Wali et al., 2018b) and freeway ramp related rear-end crashes (Kim et al., 2016). Using these new insights, proactive and personalized strategies for enhancing safety were highlighted (Kim et al., 2016, Kamrani et al., 2017, Wali et al., 2018b). Methodologically, the importance of accounting for hierarchical heterogeneity in maximum-simulated and Bayesian framework was emphasized (Wali et al., 2018b, Arvin et al., 2019).
It seems reasonable to expect that the variations in microscopic driving behaviors immediately prior to crash involvement, termed as driving volatility, can be majorly correlated with crash outcomes, i.e., injury severity. The previous studies focused on crash frequency and not on the outcomes of crashes per se (Kim et al., 2016, Kamrani et al., 2017, Wali et al., 2018b). Also, previous studies analyzed aggregated data in the sense that location-specific behavioral data were linked with historic police-reported crashes at such locations. As such, insights regarding how driving volatility in time to collision relates to driver’s propensity of receiving injuries cannot be obtained. With the recent advent of Naturalistic Driving Study (NDS) data, driving decisions in time to collision can be analyzed vis-à-vis driving behavior in normal events. To this end, recent studies have shown that the variations in microscopic driving decisions in time to collision can be a leading indicator of safety (Wali et al., 2019b, Wali and Khattak, 2020). Using the concept of “event-based volatility”, the statistical relationships between instantaneous driving decisions in longitudinal and lateral directions and crash propensity (normal driving, crash, near-crash) are established (Wali et al., 2019b, Wali and Khattak, 2020). For example, greater “intentional” volatility in non-vulnerable and vulnerable locations (e.g., school zones) is reported to increase the likelihood of near-crashes/crashes (Wali et al., 2019b, Wali and Khattak, 2020). In line with the conceptual argument presented in the literature with relevance to the potential presence of heterogeneity in naturalistic driving data (see Table 1 in (Mannering et al., 2016)), both of the above studies concluded presence of substantial variations in the effects of naturalistic driving volatility on crash propensity due to systematic variations in factors unobserved in the data (Wali et al., 2019b, Wali and Khattak, 2020). Nonetheless, the previous studies did not focus on the injury severity component – i.e., little is known about how driving volatility in time to collision relates with injury outcomes.
This study focuses on the links between driving volatility and crash-injury severity. A tight observational study design is harnessed to quantify real-world driving volatility in time to collision, and how it relates to injury outcomes sustained by drivers. In particular, the study uses a unique Naturalistic Driving database of drivers involved in crash events. Large-scale data on real world instantaneous driving behaviors in time to collision are analyzed to create driving volatility indices using different driving performance measures. To quantify the links between driving volatility in time to collision and crash-injury severity outcomes, the volatility indices are then combined with individual-level data on injury severity, event-specific variables including pre-crash behaviors, maneuvers, fault status, secondary tasks and durations, roadway, traffic, and environmental factors. In doing so, the critical methodological concerns related to observed (systematic) and unobserved (random) heterogeneity are carefully addressed (Mannering et al., 2016). From a methodological standpoint, fixed and random parameter discrete outcome logit models are developed to account for both systematic and random heterogeneity. Possible heterogeneity-in-means and variances of the random parameters varying as a function of several observed explanatory factors is captured. By using advanced modeling methods, the study contributes by seeking a fundamental understanding of short-term microscopic driving volatility, and how can we map driving volatility to injury severities sustained by drivers in crashes. An analysis along these lines is indispensable for developing more personalized behavioral countermeasures to potentially reduce drivers’ injury outcomes.
Section snippets
Conceptual illustration
To understand the links between driving volatility in time to collision and injury outcomes (given a crash), detailed microscopic instantaneous driving data are needed. The SHRP2 NDS provides unique and relevant data (TRB, 2013). A key aspect of SHRP2 NDS is that it provides information on real-world driving decisions undertaken by drivers prior to involvement in a crash event. Crash is defined as “any contact that the subject instrumented vehicle has with an object (moving or fixed) at any
Data
The 2nd Strategic Highway Research Program (SHRP2) Naturalistic Driving Study (NDS) is the largest naturalistic driving environment till date (TRB, 2013), including 3400 participant drivers with over 4000 years of real-world naturalistic driving data collected between 2010 and 20135
Descriptive statistics
Table 1 presents the descriptive statistics of key variables used in this study. In the SHRP2 NDS database, the crash severity is coded into four categories: low-risk tire strike, minor crash [MC], police-reportable crash [PRC], and most severe crash [SC]. For detailed definitions of the different response outcome categories, see Hankey et al. (2016). As shown in Table 1, approximately 40% and 38% of crashes resulted in low-risk tire strike and minor crash respectively. Whereas 13.3% and 8.8%
Safety effects of driving volatility
The results and findings discussed here refer to the random-parameter models with heterogeneity-in-means and variances given its relatively best fit (Table 3, Table 4). However, in order to compare the performance of this model with that of commonly applied fixed and random parameter models (with no heterogeneity-in-means and variances), we also show the model estimation results for fixed parameter and mixed logit models in Table 3, Table 4. Overall, statistically significant and positive
Limitations/future work
The present study is based on a sample of ~9800 events (baseline, near-crash, and crash events), out of which 671 were identified as crash events. However, the SHRP2 NDS Event Detail Table (EDT) currently has 41,479 records, out of which 1877 are crash events (Wali and Khattak, 2020) (https://insight.shrp2nds.us/data/index). The authors used a subset of EDT due to lack of access to the entire SHRP2 NDS database. With regard to future work, there are several pathways for extending the proposed
Conclusions
As a key indicator of unsafe driving, driving volatility characterizes the variations in instantaneous driving decisions, especially capturing extreme driving behaviors. This study characterized longitudinal and lateral volatility in microscopic driving decisions and examined how driving volatility in time to collision relates to crash-injury severity. A rigorous data analytic methodology was proposed to extract critical information embedded in real-world naturalistic driving data related to
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
The data for this study were provided through a collaborative effort between Virginia Tech Transportation Institute, the U.S. Federal Highway Administration (FHWA), and Oak Ridge National Laboratory (ORNL). The timely assistance and guidance of the ORNL team about data elements is highly appreciated. The authors would also like to recognize the contribution of Alexandra Boggs in proof-reading the manuscript. This paper is based upon work supported by the US National Science Foundation under
References (96)
Analysis of driver injury severity levels at multiple locations using ordered probit models
Journal of Safety Research
(2003)- et al.
Temporal stability of driver injury severities in animal-vehicle collisions: A random parameters with heterogeneity in means (and variances) approach
Analytic Methods in Accident Research
(2020) - et al.
Crash modeling for intersections and segments along corridors: A Bayesian multilevel joint model with random parameters
Analytic Methods in Accident Research
(2017) - et al.
A statistical assessment of temporal instability in the factors determining motorcyclist injury severities
Analytic Methods in Accident Research
(2019) - et al.
A note on modeling vehicle accident frequencies with random-parameters count models
Accident Analysis and Prevention
(2009) - et al.
A study of factors affecting highway accident rates using the random-parameters tobit model
Accident Analysis and Prevention
(2012) - et al.
How instantaneous driving behavior contributes to crashes at intersections: extracting useful information from connected vehicle message data
Accident Analysis and Prevention
(2019) - et al.
Determinant of injury severities in large truck crashes: A weekly instability analysis
Safety Science
(2020) - et al.
Determinants of bicyclist injury severities in bicycle-vehicle crashes: A random parameters approach with heterogeneity in means and variances
Analytic Methods in Accident Research
(2017) - et al.
The effect of passengers on driver-injury severities in single-vehicle crashes: A random parameters heterogeneity-in-means approach
Analytic Methods in Accident Research
(2017)