Sensitivity of disease cluster detection to spatial scales: an analysis with the spatial scan statistic method

The spatial scan statistic method has been widely used for detecting disease clusters. Its results may be affected by scales, including the aggregation level of the input data and the population threshold used in the detection. Previous studies offered inconsistent findings, and few had considered both types of scales at the same time. Using 24 simulated datasets and two real disease datasets, we investigated the method’s sensitivity to the two types of scales. We aggregated the individual-level data into areal units of three levels, including county, town, and a 900 m grid. We detected clusters with three population thresholds, including 10%, 25%, and 50%. We used two measurements, distance between cluster centres and the Jaccard index, to quantify the consistency of clusters detected with different scale settings. We find: (1) the method is not greatly sensitive to the data aggregation level when the cluster is strong and in a place with high population density; (2) the method’s sensitivity to the population threshold is determined by the actual size of the true cluster; and (3) a regular grid with fine resolution is advantageous over the subjectively defined areal units. The process and findings may have broader meanings to similar spatial analyses.空间扫描统计方法（spatial scan statistic method）已被广泛应用于疾病聚集区检测。其检测结果易受两类尺度因素影响，即输入数据的聚合层级与检测过程中采用的人口阈值。既往相关研究结论并不统一，且鲜有研究同时考量这两类尺度因素。本研究依托24个模拟数据集与2个真实疾病数据集，探究了该方法对这两类尺度的敏感性。我们将个体级数据聚合为三级面状单元（areal units）：县级、乡镇级与900米网格单元。设置10%、25%、50%三类人口阈值开展聚集区检测，并采用聚类中心间距与雅卡尔指数（Jaccard index）两种量化指标，对不同尺度设置下检测得到的聚集区一致性进行量化。研究结果显示：（1）当聚集区特征显著且所处区域人口密度较高时，该方法对数据聚合层级的敏感性较低；（2）该方法对人口阈值的敏感性由真实聚集区的实际规模决定；（3）高分辨率规则网格相较于主观定义的面状单元更具优势。本研究的流程与结论可为同类空间分析提供更广泛的参考价值。