We present a comparison of the most commonly used methods for traffic accident hotspots identification.
There are in general three types of methods for hotspots identification. The most straightforward approach is based on aggregated counting of records. The sums are either used directly to rank the segments of roads, or the local spatial autocorrelation statistic (local Getis Ord statistic) is computed. The latter option seems better because it allows for setting an objective threshold for distinguishing significantly dangerous locations. These methods have, however, several drawbacks: segmentation of roads, not considering the regression to the mean and aggregation when the exact positions of animal-vehicle collisions are known.
Additionally, various regression models are often built to analyse crash-frequency data. They express the number of animal-vehicle collisions by the use of explanatory variables. There are a number of methodological issues which have to be addressed prior to the application of this approach (e. g. time-varying explanatory variables, under-reporting, the low sample mean and sample size, omitted variables, disunity in the choice of the functional form, segmentation of roads). Hence, regression analyses are time-consuming in the case of crash-frequency data and often produce biased results. The empirical Bayes method uses the results from a regression model as the prior estimate of crash-frequency counts. Afterwards, the prior information is combined with the actual data and a posterior estimate is produced. Although this is a brilliant idea, the accuracy of the empirical Bayes method depends on prior estimates produced by a regression model. Furthermore, this approach provides no objective threshold for distinguishing significantly dangerous locations.
Finally, clustering analysis can be used to find locations where animal-vehicle collisions occur more frequently than expected. Clustering methods can either testify to a general clustering tendency or identify the exact positions of hotspots (clusters). In our previous research, we introduced the KDE+ method which is based on kernel density estimation. The KDE+ method is able to objectively determine significant clusters and allows for the ranking of the clusters.
We compared the methods for spatial analysis of animal-vehicle collisions from both a theoretical and practical view. Based on our results, we recommend the use of clustering techniques, particularly the KDE+ method, because they are effective, robust and capable of distinguishing whether an animal-vehicle collision occurred due to local factors (clusters) or global factors (random distribution along a network).
road kill; traffic accidents; spatial analysis; methodology; clustering analysis; regression models