Refinement of Accident Prediction Models for Spanish National Network

Abstract

This paper describes a research project conducted at Madrid Polytechnic University with the objective of refining the negative binomial accident prediction models that had been developed previously for two-lane rural roads in Spain. Cumulative scaled residuals plots were used to identify the value ranges of annual average daily traffic (AADT) where the model over- or underestimated accident frequencies. They were also used to determine whether the calibration sample contained redundant information for some values of the explanatory variables that was detrimental to the model-fitting process. On the basis of the results of these analyses, two approaches were explored to refine the models. First, a random reduction of the sample size was tried to mitigate the effect of redundancies in the statistical information. Second, the sample was stratified, and independent models were fitted for the regions of AADT values where the cumulative residuals plot showed moderate fluctuations. These processes reduced considerably the amount of over- and underestimation of the models, which indicates that in some cases they may be a valid tool to refine accident prediction models and to overcome the lack of flexibility in the functional forms commonly used in multivariate regression modeling.