A semiparametric cluster detection method — a comprehensive power comparison with Kulldorff's method

Abstract

BackgroundA semiparametric density ratio method which borrows strength from two or more samples can be applied to moving window of variable size in cluster detection. The method requires neither the prior knowledge of the underlying distribution nor the number of cases before scanning. In this paper, the semiparametric cluster detection procedure is combined with Storey's q-value, a type of controlling false discovery rate (FDR) method, to take into account the multiple testing problem induced by the overlapping scanning windows.ResultsIt is shown by simulations that for binary data, using Kulldorff's Northeastern benchmark data, the semiparametric method and Kulldorff's method performs similarly well. When the data are not binary, the semiparametric methodology still works in many cases, but Kulldorff's method requires the choices of a correct probability model, namely the correct scan statistic, in order to achieve comparable power as the semiparametric method achieves. Kulldorff's method with an inappropriate probability model may lose power.ConclusionsThe semiparametric method proposed in the paper can achieve good power when detecting localized cluster. The method does not require a specific distributional assumption other than the tilt function. In addition, it is possible to adapt other scan schemes (e.g., elliptic spatial scan, flexible shape scan) to search for clusters as well.