Latent class analysis of autonomous vehicle crashes

Abstract

Introduction: Since September 2014, the California Department of Motor Vehicles has requested autonomous vehicle (AV) manufacturers to report their accidents if they take field tests on public roadways in California. These collision reports are heterogeneous containing a variety of accident factors. Method: To describe the accident more elaborately, we add three new category variables: ‘traffic control and status,’ ‘speed/speed change,’ and ‘type of accident location,’ extracted from crash narratives. Combining with the existing variables as model inputs, we use Latent Class Analysis (LCA) to investigate the mixture types of traffic accidents. After using ‘Mplus’ (LCA tool), the data set with 308 cases has been segmented into three clusters, including ‘rear-end collisions after the speed change of AV,’ ‘sideswipe collisions at parking places,’ and ‘hit-object collisions in normal traffic road.’ Results: These three clusters are not highlighted in previous literature and Cluster 1 shows AV should not be designed too ethically. To follow the driving habits of traditional drivers, AVs should accelerate vehicles quickly when they start to move and delay stopping in front of stop lines, traffic lights, and yielding. The cluster-based analyses show that applying LCA as a preliminary analysis can reveal the interesting hierarchical patterns hidden in the dataset and help traffic safety researchers improve AV safety performances.