What do surrogate safety metrics measure? Understanding driving safety as a continuum

Road safety is an important public health issue; technology, policy, and educational interventions to prevent crashes are of significant interest to researchers and policymakers. In particular, there is significant ongoing research to proactively evaluate the safety of new technologies, including autonomous vehicles, before enough crashes occur to directly measure their impact. We analyze the distributional form of five diverse datasets that approximate motor vehicle safety incident severity, including one dataset of hard braking events that characterizes the severity of non-crash incidents. Our empirical analysis finds that all five datasets closely fit a lognormal distribution (Kolmogorov-Smirnov distance < 0.013; significance of loglikelihood ratio with other distributions < 0.000029). We demonstrate a linkage between two well-known but largely qualitative safety frameworks and the severity distributions observed in the data. We create a formal model of the Swiss Cheese Model (SCM) and show through analysis and simulations that this formalization leads to a lognormal distribution of the severity continuum of safety-critical incidents. This finding is not only consistent with the empirical data we examine, but represents a quantitative restatement of Heinrich's Triangle, another heretofore largely qualitative framework that hypothesizes that safety events of increasing severity have decreasing frequency. Our results support the use of more frequent, low-severity events to rapidly assess safety in the absence of less frequent, high-severity events for any system consistent with our formalization of SCM. This includes any complex system designed for robustness to single-point failures, including autonomous vehicles.