Accounting for previous events to model and predict traffic accidents at the road segment level: A study in Valencia (Spain)

Abstract

Predicting the occurrence of traffic accidents is essential for establishing preventive measures and reducing the impact of traffic accidents. In particular, it is fundamental to make predictions using fine spatio-temporal units. In this paper, the daily risk of traffic accident occurrence across the road network of Valencia (Spain) is modeled through logistic regression models. The spatio-temporal dependence between the observations is accounted for through the inclusion of lagged binary covariates representing the previous occurrence of a traffic accident within a spatio-temporal window centered at each combination of day and segment of the network. A temporal distance of 28 days and a fifth-order spatial distance are set as the limits of such dependence. Furthermore, the models include fixed effects in terms of several socio-demographic, network-related, and weather-related covariates. Temporal (month and day of the week) and spatial (borough-level) effects are also considered. The predictive quality of the models is examined through the Matthews correlation coefficient and the prediction accuracy index. The results indicate that the incorporation of spatio-temporal dependence improves the predictive ability of the models. However, while the inclusion of temporally-lagged covariates representing short-and mid-term temporal dependence yields more accurate predictions, the higher-order spatial lags barely alter model performance.