Predicting road crashes using calendar data

Abstract

In road safety, macroscopic models are developed to support the quantitative targets in safety programmes. Targets are based on estimated numbers of fatalities and crashes that are typically derived from models. When constructing these models, typical problems are the lack of relevant data, the limited time horizon and the availability of future values for explanatory variables. As a solution to these restrictions, we suggest the use of calendar data. These include a trend, a trading day pattern, dummy variables for the months and a heavy traffic measure. In this paper, we test the relevance of calendar data for the explanation and prediction of road safety. ARIMA models and regression models with ARMA errors and calendar variables are built. Predictions are made by both models and the quality of the predictions is compared. We use Belgian monthly crash data (1990-2002) to develop models for the number of persons killed or seriously injured, the number of persons lightly injured and the corresponding number of crashes. The regression models fit better than the pure ARIMA models. The trend and trading day variables are significant for the outcomes related to killed or seriously injured persons, while the heavy traffic measure is significant in all models. The predictions made by the regression models are better than those from the ARIMA models, especially for the lightly injured outcomes.