Selection bias in naturalistic driving studies
Holger Rootzén – Chalmers University of Technology

In a “naturalistic driving study” ordinary cars or trucks with ordinary drivers are equipped with cameras which film driver behavior and the surrounding traffic; radars which measure the distance to road edges and other cars; GPS instruments; and sensors which measure things like brake and gas pedal actions. The vehicles are then operated just as if the instrumentation wasn’t there, but if accidents occur, the instrumentation documents what happened in extensive detail. Naturalistic driving studies are believed to contribute to the steadily falling number of fatal traffic accidents in countries like Sweden or the US. Such studies which involve thousands of cars and several years of driving per car are underway. However, even for large studies, the number of crashes is low, and one additionally tries to exploit information also from “near-crashes”. In this talk I discuss how extreme value statistics can be used to 1) help verify the underlying assumption that preventing near-crashes will also prevent real crashes, 2) find variables, in particular related to inattention, which influence accident risk, 3) estimate risks for, and distributions of severity of crashes, in different traffic situations and for different drivers, and 4) how to handle the censoring which occurs when visual behavior only is measured in short (6-second) time intervals. This talk is based on joint work with Jenny Jonasson, AstraZeneca, and Jonas Bärgman and Dmitrii Zholud, Chalmers.

Fellow IMS, Member ISI, co-editor Scandinavian Journal of Statistics, associate editor Annals of Statistics, former editor Extremes and Bernoulli, elected member of the Royal Swedish Academy of Science and of Kungliga Fysiografiska Sällskapet, leader of the Wallenberg project “Big Data and Big Systems – bridging local and global”, senior researcher the Foundation for Strategic Research project Material Structures Seen Through Microscopes and Statistics. Research: In finance, on risk management, credit risk and discrete hedging. General, on martingales and extreme value theory with applications to medicine, wind storm modeling, metal fatigue, pit corrosion analysis, internet traffic, and traffic safety. Has published about 80 papers in international journals, one book which continues to be a highly cited classic, one edited book, and one textbook.
Relation to theme of conference: As suggested by the organizers, this talk gives another perspective on how extreme value statistics can be used to handle risks. The talk studies the risk is to be killed in a car accident. To know more…