Identifying hotspots or crash clusters is an important problem for detecting high-risk locations at which vehicle accidents frequently occur. Several hotspot identification methods have been developed in the literature; however, there are often large differences between the spatial distributions of hotspots obtained by these methods, and spatial weights such as free flow travel times and congested flow travel times have not been fully examined as alternatives to the weight of distance. This paper compares the following commonly implemented network-based hotspot detection methods: Getis-Ord Gi*, Local Moran's I, KLINCS (K-function local indicators of network-constrained clusters), and KLINCS-IC (Inverse Cost) in order to provide insight into understanding the similarities and differences between selected hotspot detection methods when used with different spatial weights. This assessment is performed through using different spatial weights as part of the statistical analysis and comparing them to a prediction accuracy index. Moreover, a sensitivity analysis was conducted based on alternative spatial weights and different parameters to test the effect of bandwidth on the identified hotspots. The findings on the success of alternative spatial weights has a potential to improve the accuracy of the hotspot detection. Results indicate distinct differences in the spatial distributions of hotspots obtained through the considered methods that are based on alternative spatial weights. An interesting finding is that all alternative approaches (spatial weights and bandwidths) cluster when certain bandwidth values are exceeded. From a practical perspective, the CPAI results show that using network-constrained Local Moran's I statistics with a distance-based spatial weights may provide the most feasible approach while implementing safety-focused efforts.
|Number of pages||9|
|Journal||Computers, environment and urban systems|
|Early online date||30 Aug 2019|
|Publication status||E-pub ahead of print/First online - 30 Aug 2019|