Evaluating the impacts of enriched information on crash prediction performance: case studies in Brazil and Flanders

Abstract

Increasing road fatalities have urged transportation safety professionals and researchers to pursue efficient strategies to promote road safety and to reduce the number of injuries and deaths. However, in some countries like Brazil, modeling road casualties is a challenging task due to the unavailability of essential information. The main objective of this study is to expound the improvements in the performance of spatial crash prediction models by enhancing potential explanatory variables. Such practice will further facilitate suitable safety countermeasures’ enforcement. The analyses were conducted by associating the casualty figures with exposure, road-network, socioeconomic and demographic data of the state of São Paulo in Brazil, and Flanders in Belgium. These models are developed within the framework of geographically weighted regression models. For both countries, casualties as the response variable was divided into two separate sets based on transport mode, named active (pedestrian and cyclists) and motorized transport (motorized vehicles occupants). In order to evaluate the statistical impacts of the enriched information on the models, the methodological procedure followed two main phases. Firstly, we modelled casualties for both countries based on the available corresponding Brazilian data. At this stage the best fitted models, for which we utilized the minimum available data, are developed and called basic models. In the second phase, another round of modelling exercise was carried out, this time only for Flanders and by considering all available variables in the Flemish dataset. These models are called improved models. The results of analysis show that the improved models outperformed the basic models, for both dependent variables. In the model developed for motorized transport, reductions of 25% and 30% were observed compared with the basic model, respectively for AICc and MSPE. For active model, 30% and 40% reductions were resulted for AICc and MSPE, respectively.