Autonomous vehicle crashes: Insights from a nonlinear mixed logit model of occupant severities and a latent effects logit-gamma model of vehicle exposure selectivity

This dissertation explores and discusses findings with respect to reported crashes involving autonomous vehicle systems. The dissertation is comprised of five major sections with two different analyses taking place. The first section serves as an introduction to the research background, objectives, questions, and scope. The second section presents the modeling methodologies used in the two studies. The third section presents a nonlinear mixed logit model of occupant severities in crashes involving autonomous vehicles. The fourth section presents a latent effects logit-gamma model to identify endogeneity in total vehicle count with respect to four distinct autonomous treatments. Lastly, the fifth section serves as the conclusions and recommendations associated with this research. The nonlinear mixed logit model captures heterogeneous effects of contributing factors on autonomous involved occupant severities. Autonomous level information to this point has been quite sparse in the context of real-world crash scenarios and police reporting. However, the Texas Department of Transportation (TxDOT) began reporting autonomous involvement in April of 2023. With reporting still in its early stages, this analysis incorporated three distinct vehicle technologies: non-autonomous internal combustion engine (ICE) vehicles; ICE and hybrid electric autonomous vehicles; and fully electric autonomous vehicles. Crash data included any crash in Texas from April to December of 2023 that involved at least one autonomous-indicated vehicle (either the second or third distinct vehicle technology). Random parameters were found with respect to: an indicator for occupant involvement in the first harmful crash sequence event, with that event being collision with a fixed object, for no injury; proportion of autonomous vehicles for no injury; an intersection-related indicator for possible injury; total occupant count for possible injury; and total vehicle count for injury. The count and proportion variables were expressed as nonlinear relationships, for which random parameters improved prediction accuracy for possible injury and injury outcomes by 37.50 % and 30.00 % respectively, as compared to the conventional mixed logit estimates. The findings in this study highlight the applicability of the nonlinear mixed logit for severity analysis with respect to complex autonomous interactions in crashes. The latent effects logit-gamma model captures endogeneity with respect to total number of vehicles involved in the crash and proportion of total vehicles having autonomous capabilities. In this study, there are four distinct treatments (autonomous vehicle proportions) for the outcome (total vehicle count): strictly non-autonomous crashes (baseline treatment); non-autonomous vehicle majority (less than 50 % autonomous vehicles); autonomous vehicle majority (greater than 50 % autonomous vehicles but less than 100 %); and autonomous vehicle only crashes (100 % autonomous vehicles). Each of the treatments (excluding the baseline) were found to be significant and endogenous with total vehicle count. Additionally, the latent factors introduced in the methodology are significant for each of the treatments (excluding the baseline), pointing to unobserved factors common to selection and exposure. Findings from this study provide more clarity as to driving and crash contexts which influence selection in autonomous vehicle crashes. For example, manner of collision being entering at angle may be more selective to higher exposure of autonomous capabilities (excluding 100 % autonomous). This may be due to lack of driver adaptation in contexts that promote entering at angle collisions (intersections in many cases). Findings from these studies point to the necessity of understanding the nonlinear and heterogeneous nature of autonomous crashes, as well as their inherent selection bias. With more reported autonomous crashes, analysts can begin to target more specific research objectives (such as autonomous only crashes, the influence of specific intersection types, crashes involving a collision with a fixed object, sequencing of crash events, etc.) to continue to gain insights on autonomous crashes and inform policy makers as to the realities of the industry.