Extended prediction models for crashes at roundabouts

This paper builds upon the results of previously developed crash prediction models for roundabouts. The originally investigated sample was extended from 90 to 148 roundabouts. Poisson and gamma modelling techniques were used, the latter ones since underdispersion in the crash data was observed. Separate models were fit for crashes with six different types of road users: bicyclists, motorcyclists, passenger and heavy four-wheel vehicles, moped riders and pedestrians. A further distinction was made between single-vehicle and multiple-vehicle crashes. The results show that the overall number of crashes is more or less proportional to the number of motorized vehicles. The mean number of single-vehicle crashes per passing vehicle is lower on busier roundabouts. Confirmation is found for the existence of a 'safety in numbers' effect for different types of road users. Three-leg roundabouts tend to perform worse than roundabouts with four or more legs. More crashes seem to occur at roundabouts with bypasses for traffic in some direction. Larger central islands correlate with more single-vehicle crashes. Moped riders and motorcyclists are strongly overrepresented in both single-vehicle and multiple-vehicle crashes whereas bicyclists are clearly overrepresented in multiple-vehicle crashes. Roundabouts with cycle paths perform better than roundabouts with other types of cycle facilities, particularly in comparison with roundabouts with cycle lanes close to the roadway.