Evaluating Local Non-Stationarity when Considering the Spatial Variation of Large-scale Autocorrelation

Multi-scale effects of spatial autocorrelation may be present in datasets. Given the importance of detecting local non-stationarity in many theoretical as well as applied studies, it is necessary to “remove” the impact of large-scale autocorrelation before common techniques for local pattern analysis are applied. It is proposed in this paper to employ the regionalized range to define spatially varying sub-regions within which the impact of large-scale autocorrelation is minimized and the local patterns can be investigated. A case study is conducted on crime data to detect crime hot spots and cold spots in San Antonio, Texas. The results confirm the necessity of treating the non-stationarity of large-scale spatial autocorrelation prior to any action aiming at detecting local autocorrelation.