Accident Prediction Models for Bus Rapid Transit Systems: Generalized Linear Models Compared with a Neural Network

This research sought to model traffic accidents in the bus rapid transit (BRT) system in Bogotá, Colombia. For each BRT station, 35 variables related to system flows, infrastructure, service, surroundings, and socio-economic context were tested. After a selection process, a set of 11 explanatory variables was obtained and used in the development of generalized linear models (Poisson and negative binomial models) and a neural network model. The results showed that the neural network model had better predictability indicators than did those obtained by the Poisson and negative binomial models. Additionally, the negative binomial regression model did not produce better predictions than did the Poisson regression model. Finally, a scenario analysis was developed from the most relevant variables: bus flow, number of accesses, and proximity to at-grade vehicular intersections.