Binary logistic regression model applied to data on accidents occurred on federal highways in Brazil

• Autor(a) principal: Santos, Damião Flávio dos
• Data de Publicação: 2022
• Outros Autores: Souza, Yuri Machado de
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Research, Society and Development
• Texto Completo: https://rsdjournal.org/index.php/rsd/article/view/36833
• topic: Supervised analysis; Machine learning; Odds ratio; Lethality of accidents; Highway accidents.; Análisis supervisado; Aprendizaje automático; Razón de probabilidades; Letalidad de los accidentes; Accidentes de carretera.; Análise supervisionada; Aprendizado de máquina; Razão de chances; Letalidade dos acidentes; Acidentes rodoviários.

Accidents on federal highways in Brazil lead to social and economic impacts on the country. Data from the Federal Highway Police reveal that thousands of people lose their lives in these accidents year after year. This paper aims to examine the factors that influence the probability of death based on the occurrence of the accident. The estimation of a binary logistic regression model took place, in which the event of interest is the circumstance of death in an accident with data from 2021. Following variable selection procedures, it was possible to obtain the final model, which was later validated with data from 2022. The accuracy of the model for both 2021 and 2022 data was around 70%. Then, the odds ratio was calculated between some distinct categories, and how much of an increase in accident lethality it generates compared to the reference category. For example, in a crash, a pedestrian is 15.6 times more likely to die when compared to the driver, while a cyclist is 5.3 times more likely to die. Although most accidents have a human cause, some results show the need of public policies that can help reduce these tragedies. To explain the model, a dashboard was created in a way that the user is able to obtain the probability of death by selecting specific accident characteristics and those involved.