Investigating transferability and goodness of fit of two different approaches of segmentation and model form for estimating safety performance of motorways

The analysis of a road network could be carried out using different techniques, the most advanced of which for the identification of hazardous sites are based on a statistical tool known as a "Safety Performance Function" (SPF). SPFs are fundamental to procedures in the Highway Safety Manual (HSM) and Safety Analyst, both of which require calibration of the SPFs to local conditions. In the case of infrastructure in which the traffic flow is physically divided, such as motorways, there are two different methods to calibrate an SPF; one is based on a consideration of two directional segments, the other one is based on monodirectional segments with a correction factor that takes into account the presence of two directional traffic, as proposed by Safety Analyst. In the first case, traffic flows are the sum of traffic flows for each direction. The two methods give a different calibration factor for local condition, i.e., different performance in terms of goodness of fit and SPF transferability. This paper seeks to address these differences by investigating the two approaches to estimate crash count using a motorway sample data from Italy and focusing on a comparison in terms of model transferability. Furthermore two different SPFs were calibrated on the same data using the two different functional equation form, to evaluate the goodness of fit of the two different traffic based segmentation. The goodness of fit of the models is investigated using CURE plots and the R2 values. The four sets of models give good results in terms of goodness of fit but they have different values for the calibration factor. The calibration factor of the calibrated monodirectional model is closer to a value of one, suggesting a better goodness of fit on the contrary the HSM model gives the better performance in terms of transfer ability where crash data are not available. (Less)