Developing Updatable Crash Prediction Model for Network Screening: Case Study of Czech Two-Lane Rural Road Segments

Abstract

This paper focuses on application of crash prediction models in network screening. The two main questions were (a) What variables should be involved in the model? and (b) What length should the modeled period be? Answers to these questions should provide guidelines for developing an updatable crash prediction model (i.e., a model that is both reliable and simple so that its updating for periodical network screening is not highly demanding). Data on approximately 1,000 km (600 mi) of a two-lane rural-road network from South Moravia, Czech Republic, were used. On the basis of 8 years of annual crash frequencies, together with exposure and geometrical variables, several variants of prediction models were developed. To study the quality of the models, a series of consistency tests was applied relative to comparison of the models themselves as well as to their diagnostic performance. As a result, simple crash prediction models (that included traffic volume, segment length, and curvature change rate) were found sufficient for network screening. If one supposes that length and curvature are unlikely to change often, only traffic volume data need to be periodically updated. Consistency analyses indicate that this period should be 4 years. Under these conditions, models are being applied in the studied region. Further planned activities include extensions to intersections and also to other Czech regions.