RESUMO
BACKGROUND: Kulldorff's spatial scan statistic and its software implementation - SaTScan - are widely used for detecting and evaluating geographic clusters. However, two issues make using the method and interpreting its results non-trivial: (1) the method lacks cartographic support for understanding the clusters in geographic context and (2) results from the method are sensitive to parameter choices related to cluster scaling (abbreviated as scaling parameters), but the system provides no direct support for making these choices. We employ both established and novel geovisual analytics methods to address these issues and to enhance the interpretation of SaTScan results. We demonstrate our geovisual analytics approach in a case study analysis of cervical cancer mortality in the U.S. RESULTS: We address the first issue by providing an interactive visual interface to support the interpretation of SaTScan results. Our research to address the second issue prompted a broader discussion about the sensitivity of SaTScan results to parameter choices. Sensitivity has two components: (1) the method can identify clusters that, while being statistically significant, have heterogeneous contents comprised of both high-risk and low-risk locations and (2) the method can identify clusters that are unstable in location and size as the spatial scan scaling parameter is varied. To investigate cluster result stability, we conducted multiple SaTScan runs with systematically selected parameters. The results, when scanning a large spatial dataset (e.g., U.S. data aggregated by county), demonstrate that no single spatial scan scaling value is known to be optimal to identify clusters that exist at different scales; instead, multiple scans that vary the parameters are necessary. We introduce a novel method of measuring and visualizing reliability that facilitates identification of homogeneous clusters that are stable across analysis scales. Finally, we propose a logical approach to proceed through the analysis of SaTScan results. CONCLUSION: The geovisual analytics approach described in this manuscript facilitates the interpretation of spatial cluster detection methods by providing cartographic representation of SaTScan results and by providing visualization methods and tools that support selection of SaTScan parameters. Our methods distinguish between heterogeneous and homogeneous clusters and assess the stability of clusters across analytic scales. METHOD: We analyzed the cervical cancer mortality data for the United States aggregated by county between 2000 and 2004. We ran SaTScan on the dataset fifty times with different parameter choices. Our geovisual analytics approach couples SaTScan with our visual analytic platform, allowing users to interactively explore and compare SaTScan results produced by different parameter choices. The Standardized Mortality Ratio and reliability scores are visualized for all the counties to identify stable, homogeneous clusters. We evaluated our analysis result by comparing it to that produced by other independent techniques including the Empirical Bayes Smoothing and Kafadar spatial smoother methods. The geovisual analytics approach introduced here is developed and implemented in our Java-based Visual Inquiry Toolkit.
Assuntos
Análise por Conglomerados , Interpretação Estatística de Dados , Modelos Estatísticos , Neoplasias do Colo do Útero/mortalidade , Gráficos por Computador/tendências , Feminino , Humanos , Fatores de Risco , Estados Unidos/epidemiologiaRESUMO
Recent increases in deciduous shrub cover are a primary focus of terrestrial Arctic research. This study examined the historic spatial patterns of shrub expansion on the North Slope of Alaska to determine the potential for a phase transition from tundra to shrubland. We examined the historic variability of landscape-scale tall shrub expansion patterns on nine sites within river valleys in the Brooks Range and North Slope uplands (BRNS) between the 1950s and circa 2010 by calculating percent cover (PCTCOV), patch density (PADENS), patch size variability (CVSIZE), mean nearest neighbor distance (MEDIST) and the multi-scale information fractal dimension (d I ) to assess spatial homogeneity for shrub cover. We also devised conceptual models for trends in these metrics before, during, and after a phase transition, and compared these to our results. By developing a regression equation between PCTCOV and d I and using universal critical d I values, we derived the PCTCOV required for a phase transition to occur. All nine sites exhibited increases in PCTCOV. Five of the nine sites exhibited an increase in PADENS, seven exhibited an increase in CVSIZE, and five exhibited a decrease in MEDIST. The d I values for each site exceeded the requirements necessary for a phase transition. Although fine-scale heterogeneity is still present, landscape-scale patterns suggest our study areas are either currently in a state of phase transition from tundra to shrubland or are progressing towards spatial homogeneity for shrubland. Our results indicate that the shrub tundra in the river valleys of the north slope of Alaska has reached a tipping point. If climate trends observed in recent decades continue, the shrub tundra will continue towards homogeneity with regard to the cover of tall shrubs.