Spatial Autocorrelation and Bayesian Spatial Statistical Method for Analyzing Intersections Prone to Injury Crashes

Identification of hazardous road locations (i.e., crash hot spots or black spots) is a standard practice in departments of transportation (DOTs) throughout the United States. Often, DOTs use relatively straightforward methods to detect high crash sites; these methods, however, may not be accurate. The objectives of this study are to develop two different methodologies: (a) a geographic information system (GIS)-based method to detect crash hot spots and (b) a spatial regression method with structured and unstructured random effects in continuous space to investigate the intersection-level factors that influence the concentration of fatal injury crashes. Using injury and property damage only (PDO) crash data from Tucson, Arizona, this study shows that spatial dependence plays a strong role during the analyses of road-traffic crashes. These spatial dependencies, accounted through spatial autocorrelation, help to detect statistically significant clusters of crashes involving fatalities or severe or minor injuries, in a GIS framework. This study also develops statistical models of severe- and minor-injury crashes, using classical negative binomial (NB) and Bayesian spatial statistical methods incorporating spatially structured and unstructured random effects. The coefficient estimates from a Bayesian framework are similar to those from NB estimation, but with better precision. The model that includes spatial correlations also indicates the potential to reduce the bias associated with model mis-specification by changing the estimate of the annual average daily traffic coefficients. This study's results also indicate that spatially structured correlation is quite significant in cases of crashes involving minor injury or PDO and that unstructured effects are somewhat significant at the intersection-level for cases of severe-injury crashes.