Issues in the Current Practices of Spatial Cluster Detection and Exploring Alternative Methods

Abstract

Local Moran and local G-statistic are commonly used to identify high-value (hot spot) and low-value (cold spot) spatial clusters for various purposes. However, these popular tools are based on the concept of spatial autocorrelation or association (SA), but do not explicitly consider if values are high or low enough to deserve attention. Resultant clusters may not include areas with extreme values that practitioners often want to identify when using these tools. Additionally, these tools are based on statistics that assume observed values or estimates are highly accurate with error levels that can be ignored or are spatially uniform. In this article, problems associated with these popular SA-based cluster detection tools were illustrated. Alternative hot spot-cold spot detection methods considering estimate error were explored. The class separability classification method was demonstrated to produce useful results. A heuristic hot spot-cold spot identification method was also proposed. Based on user-determined threshold values, areas with estimates exceeding the thresholds were treated as seeds. These seeds and neighboring areas with estimates that were not statistically different from those in the seeds at a given confidence level constituted the hot spots and cold spots. Results from the heuristic method were intuitively meaningful and practically valuable.