| Abstract: | High‐resolution spatial data have become increasingly available with modern data collection techniques and efforts. However, it is often inappropriate to use the default geographic units to perform spatial analysis due to unstable estimates with small areas (e.g. cancer rates for census blocks or tracts). Regionalization is aggregating small units into relatively larger areas while optimizing a homogeneity measure (such as the sum of squared differences). For exploratory spatial analysis, regionalization may help remove spurious data variation through aggregation and discover hidden patterns in data (such as areas of unusually high cancer rates). Towards this goal, this research introduces several improvements to a recent group of regionalization methods – REDCAP ( Guo 2008 ) and conducts evaluation experiments with synthetic data sets to assess and compare the capability of regionalization methods for exploratory spatial analysis. One of the major improvements is the integration of a local empirical Bayes smoother (EBS) with the regionalization methods. We generate a large number of synthetic data sets with controlled spatial patterns to evaluate the performance of both new and existing methods. Evaluation results show that the new methods (integrated with EBS) perform significantly better than their original versions and other methods (including the EBS method on its own) in terms of detecting the true patterns in the synthetic data sets. |
