Removing Biases from Crash Odds Ratio Estimates of Secondary Tasks: A New Analysis of the SHRP 2 Naturalistic Driving Study Data 2017-01-1380
Dingus and colleagues recently estimated the crash odds ratios (ORs) for secondary tasks in the Strategic Highway Research Program Phase 2 (SHRP 2) naturalistic driving study. Their OR estimate for hand-held cell phone conversation (Talk) was 2.2, with a 95% confidence interval (CI) from 1.6 to 3.1. This Talk OR estimate is above 1, contrary to previous estimates below 1. A replication discovered two upward biases in their analysis methods. First, for video clips with exposure to a particular secondary task, Dingus and colleagues selected clips not only with exposure to that task, but often with concurrent exposure to other secondary tasks. However, for video clips without exposure to that task, Dingus and colleagues selected video clips without other secondary tasks. Hence, the OR estimate was elevated simply because of an imbalanced selection of video clips, not because of risk from a particular secondary task. After removing this “video clip selection bias,” the Talk OR estimate declined 42%, to 1.2. Second, Dingus and colleagues did not filter out driver behavior errors (that is, illegal or improper driving maneuvers), whose extremely high OR estimates also upwardly biased the secondary task OR estimates. After removing this “driver behavior error bias,” the Talk OR estimate fell by another 23% to 0.92 (CI 0.29 to 2.2), now consistent with previous studies. All other Dingus et al. secondary task OR estimates also substantially declined after removing these biases, usually reversing from causal to protective. After bias removal, it was discovered that, in the aggregate, crash risks do not increase from single secondary tasks, but rather from interactions between multiple secondary tasks, or one secondary task concurrent with one or more driver behavior errors.