Glances That Matter: Applying Quantile Regression to Assess Driver Distraction from Off-Road Glances

Abstract

This study assessed whether quantile regression can identify design specifications that lead to particularly long glances, which might go unnoticed with traditional analyses focusing on conditional means of off-road glances. Although substantial research indicates that long glances contribute disproportionately to crash risk, few studies have directly assessed the tails of the distribution. Failing to examine the distribution tails might underestimate the disproportionate risk on long glances imposed by secondary tasks. We applied quantile regression to assess the effects of secondary task type (reading or entry), system delay (delay or no delay), and text length (long or short) on off-road glance duration at 15 th , 50 th , and 85 th quantiles. The results show that entry task, long text, and some combinations of variables led to longer glances than that would be expected given the central tendency of glance distributions. Quantile regression identifies secondary task features that produce long glances, which might be neglected by traditional analyses with conditional means.

Characterizing Driver Trust in Vehicle Control Algorithm Parameters

“…This is particularly important for the rapid development of automated vehicles because research with small samples or different research questions can be supported by more robust studies using prior distributions. Bayesian methods also encourage researchers to consider distributions rather than just measures of central tendency because it is the tails of the distributions that often undermine safety (Liu, Lee, Lee, & Venkatraman, 2018).…”

Patterns of Sequential Off-Road Glances Indicate Levels of Distraction in Automated Driving

“…Moreover, ANOVAs were used to compare several features aggregated from sequential off-road glances under the different levels of distraction in automated driving. The glance features included the 85 th quantile of off-road glance duration (Q85) which indicates the long off-road glance (e.g., Liu et al, 2018), the number of off-road glances per second (NoG/s) which represents the frequency of sequential glances, the standard deviation of the off-road glances (StdG) which represents the consistency of sequential glances, and the percentage of time of eyes being off-road (EoR%) which represents the overall degree of eyes being off-road over the data segment.…”

“…Previous research based on manual driving has found that distraction significantly increases drivers' off-road glance duration (e.g., Liu, Lee, Lee, & Venkatraman, 2018;Sodhi, Reimer, & Llamazares, 2002). During the period of an offroad glance, especially a longer glance (> 2 s), a driver may mismatch perceived traffic situations with actual vehicle kinematics, in the end getting involved in a crash when the safety margin is insufficient (Victor et al, 2015).…”

Designing for the Extremes: Modeling Drivers’ Response Time to Take Back Control From Automation Using Bayesian Quantile Regression

“…Other studies have found that longer-than-average responses are associated with greater safety implications for unexpected events (i.e., the left side door of a parked car suddenly opening, requiring drivers' intervention) (Summala, 1981). Another study used quantile regression analysis to show that text length, system delay, and secondary task type affected the tail of the glance distribution differently than the mean of the glance distribution (Liu et al, 2018).…”

“…In the context of driver behavior, studying the tail of the distribution using quantile regression could encourage safer designs by considering a wider range of driver behaviors. However, Liu et al performed an ISI Web of Science search on the application of quantile regression in the field of Human Factors, finding only six related research papers and only one published in Human Factors, demonstrating that this method is rarely used (Liu, Lee, Lee, & Venkatraman, 2018). The routine emphasis on mean response times could have a profound negative effect on the design of safety-critical systems, where more extreme parts of the distribution have an outsized influence on risk (Horrey & Wickens, 2007; Porter, 2015).…”