Sensitivity and specificity analysis of spatial cluster tests using simulated cancer risk surface

Lemke D, Mattauch V, Hense HW

Abstract

Background: In this study we evaluate the statistical performance of spatial cluster tests (adaptive kernel density, Kullendorf's spatial scan statistic, Besag&Newell's method, Turnbull's method, local Moran I and Getis-Ord local G test). As a reference standard we implemented a risk surface simulation where the underlying disease rates were known. Methods: We simulated cluster areas in the Regierungsbezirk Münster using a modeled high resolution map of the underlying population density based on ancillary land cover data (CORINE). We defined two cluster areas with a relative risk of 2.0 and 4.0, respectively. The observed cases were then sampled as a random Poisson process using this risk surface. We tested the five cluster tests on basis of the original population density data and on two different spatial scales (disaggregated land cover areas census districts). A sensitivity and specificity analysis were conducted with the test Results. Results & discussion: The results showed a general decrease in sensitivity (mean sensitivity = 44 vs. 29%) with increasing the spatial resolution of the underlying population data. The largest likelihood ratio for a positive result has Kullendorf's scan statistic (LR+ = 102.0) on level of census districts and the adaptive kernel density (LR+ = 26.2) on level of the disaggregated areas. The mean proportion of false positives (0.84 vs. 4.1%, respectively) was negligible. For the tests based on kernel density estimation, the Euclidian distance between cluster centroids takes into account the other cluster morphology of these tests compared to the area-based Methods: Here the Besag&Newell's method shows the shortest distances for the two clusters (d1 = 3513.6 m d2 = 1859.4 m and d1 = 1473.4 m d2 = 2719.4 m, resp.). This study shows that cluster simulations are a possibility to find the most powerful clustering method when statistical power and type 1 error are unknown.