Identifying the most informative variables to discriminate between fatal and non-fatal road accidents

Abstract

Road accidents account for millions of fatal and non-fatal victims worldwide each year, causing losses for victims and their families in many ways. Understanding the causes that differentiate fatal from non-fatal accidents is an essential topic in the area of traffic safety. In this way, this paper proposes a framework to identify the most informative variables for discriminating road accidents as fatal and non-fatal. The proposed framework is comprised of three operational steps: (i) Preprocess the data describing road accidents; (ii) Iteratively classify accident events into fatal or non-fatal and eliminate the less relevant variables using the omit-one-variable-at-a-time (O1AT) technique; and (iii) Qualitatively assess the selected variables. For that matter, data reporting accidents in rural and urban areas of Brazil (BR) and Great Britain (GB) in 2018 were analyzed. The proposed framework substantially increased the accuracy of event classification into fatal and non-fatal categories, and reduced the number of variables required for analysis. Frontal collision and pedestrian run over were deemed the most informative variables to describe fatal accidents in Brazil. As for events in Great Britain, variables associated to the period of the day the accident took place, number of vehicles involved, and whether the police attended the accident scene were deemed statistically significant in distinguishing fatal from non-fatal accidents.