Jacob Benjamin J, Krapp Fiorella, Ponce Mario, Gottuzzo Eduardo, Griffith Daniel A, Novak Robert J
Department of Medicine, William C. Gorgas Center for Geographic Medicine, University of Alabama, 845 19th Street South, Birmingham, AL 35294, USA.
Geospat Health. 2010 May;4(2):201-17. doi: 10.4081/gh.2010.201.
Spatial autocorrelation is problematic for classical hierarchical cluster detection tests commonly used in multi-drug resistant tuberculosis (MDR-TB) analyses as considerable random error can occur. Therefore, when MDRTB clusters are spatially autocorrelated the assumption that the clusters are independently random is invalid. In this research, a product moment correlation coefficient (i.e., the Moran's coefficient) was used to quantify local spatial variation in multiple clinical and environmental predictor variables sampled in San Juan de Lurigancho, Lima, Peru. Initially, QuickBird 0.61 m data, encompassing visible bands and the near infra-red bands, were selected to synthesize images of land cover attributes of the study site. Data of residential addresses of individual patients with smear-positive MDR-TB were geocoded, prevalence rates calculated and then digitally overlaid onto the satellite data within a 2 km buffer of 31 georeferenced health centers, using a 10 m2 grid-based algorithm. Geographical information system (GIS)-gridded measurements of each health center were generated based on preliminary base maps of the georeferenced data aggregated to block groups and census tracts within each buffered area. A three-dimensional model of the study site was constructed based on a digital elevation model (DEM) to determine terrain covariates associated with the sampled MDR-TB covariates. Pearson's correlation was used to evaluate the linear relationship between the DEM and the sampled MDR-TB data. A SAS/GIS(R) module was then used to calculate univariate statistics and to perform linear and non-linear regression analyses using the sampled predictor variables. The estimates generated from a global autocorrelation analyses were then spatially decomposed into empirical orthogonal bases using a negative binomial regression with a non-homogeneous mean. Results of the DEM analyses indicated a statistically non-significant, linear relationship between georeferenced health centers and the sampled covariate elevation. The data exhibited positive spatial autocorrelation and the decomposition of Moran's coefficient into uncorrelated, orthogonal map pattern components revealed global spatial heterogeneities necessary to capture latent autocorrelation in the MDR-TB model. It was thus shown that Poisson regression analyses and spatial eigenvector mapping can elucidate the mechanics of MDR-TB transmission by prioritizing clinical and environmental-sampled predictor variables for identifying high risk populations.
空间自相关对于多药耐药结核病（MDR-TB）分析中常用的经典分层聚类检测测试来说是个问题，因为可能会出现相当大的随机误差。因此，当耐多药结核病集群存在空间自相关时，集群是独立随机的这一假设就不成立了。在本研究中，使用乘积矩相关系数（即莫兰系数）来量化在秘鲁利马圣胡安德卢里甘乔采样的多个临床和环境预测变量的局部空间变化。最初，选择了包含可见光波段和近红外波段的QuickBird 0.61米数据，以合成研究地点土地覆盖属性的图像。对涂片阳性耐多药结核病个体患者的居住地址数据进行地理编码，计算患病率，然后使用基于10平方米网格的算法，将其数字叠加到31个地理参考卫生中心2公里缓冲区内的卫星数据上。基于地理参考数据的初步底图生成每个卫生中心的地理信息系统（GIS）网格化测量值，这些数据汇总到每个缓冲区的街区组和普查区。基于数字高程模型（DEM）构建研究地点的三维模型，以确定与采样的耐多药结核病协变量相关的地形协变量。使用皮尔逊相关性来评估DEM与采样的耐多药结核病数据之间的线性关系。然后使用SAS/GIS(R)模块计算单变量统计量，并使用采样的预测变量进行线性和非线性回归分析。然后，使用具有非齐次均值的负二项回归，将全局自相关分析生成的估计值在空间上分解为经验正交基。DEM分析结果表明，地理参考卫生中心与采样的协变量海拔之间存在统计上不显著的线性关系。数据呈现出正空间自相关，将莫兰系数分解为不相关的正交地图模式分量，揭示了在耐多药结核病模型中捕获潜在自相关所需的全局空间异质性。因此表明，泊松回归分析和空间特征向量映射可以通过优先考虑临床和环境采样的预测变量来识别高危人群，从而阐明耐多药结核病传播的机制。
