Semi-Parametric Cluster Detection

A binary-based approach for detecting irregularly shaped clusters

Background

There are many applications for spatial cluster detection and more detection methods have been proposed in recent years. Most cluster detection methods are efficient in detecting circular (or circular-like) clusters, but the methods which can detect irregular-shaped clusters usually require a lot of computing time.

Methods

We propose a new spatial detection algorithm for lattice data. The proposed method can be separated into two stages: the first stage determines the significant cells with unusual occurrences (i.e., individual clustering) by applying the Choynowski’s test, and the second stage determines if there are clusters based on the information of the first stage by a binomial approximate method. We first use computer simulation to evaluate the performance of the proposed method and compare it with the scan statistics. Furthermore, we take the Taiwan Cancer data in 2000 to illustrate the detection results of the scan statistics and the proposed method.

Results

The simulation results support using the proposed method when the population sizes are large and the study regions are irregular. However, in general, the scan statistics still have better power in detecting clusters, especially when the population sizes are not large. For the analysis of cancer data, the scan statistics tend to spot more clusters, and the clusters’ shapes are close to circular (or elliptic). On the other hand, the proposed methods only find one cluster and cannot detect small-sized clusters.

Conclusions

In brief, the proposed methods can detect both circular and non-circular clusters well when the significant cells are correctly detected by the Choynowski’s method. In addition, the binomial-based method can handle the problem of multiple testing and save the computing time. On the other hand, both the circular and elliptical scan statistics have good power in detecting clusters, but tend to detect more clusters and have lower accuracy in detecting non-circular clusters.

Covariate adjusted weighted normal spatial scan statistics with applications to study geographic clustering of obesity and lung cancer mortality in the United States

In the field of cluster detection, a weighted normal model-based scan statistic was recently developed to analyze regional continuous data and to evaluate the clustering pattern of pre-defined cells (such as state, county, tract, school, hospital) that include many individuals. The continuous measures of interest are, for example, the survival rate, mortality rate, length of physical activity, or the obesity measure, namely, body mass index, at the cell level with an uncertainty measure for each cell. In this paper, we extend the method to search for clusters of the cells after adjusting for single/multiple categorical/continuous covariates. We apply the proposed method to 1999-2003 obesity data in the United States (US) collected by CDC's Behavioral Risk Factor Surveillance System with adjustment for age and race, and to 1999-2003 lung cancer age-adjusted mortality data by gender in the United States from the Surveillance Epidemiology and End Results (SEER Program) with adjustment for smoking and income.