Singular Information for Cancer Cluster Detection

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): When rare cancers, such as larynx or childhood leukemia, are encountered it is often of interest to assess whether clustering of cases has arisen. Typically large areas of zero-incidence are punctuated with small aggregations of cases. There is a need to be able to detect clusters of certain cancers as these can lead to important information concerning general etiology of the disease and localized environmental risk factors for the disease. In cancer surveillance the detection of clustering has become important due to the perceived residential environmental risks relating to certain industrial/commercial processes or activities (such as waste disposal, effluent dispersal, pesticide dispersal, mobile communication broadcasting). The types of disease outcome of concern have varied from respiratory cancers such as lung or larynx, to childhood leukemia, non-Hodgkin's lymphoma, and soft tissue sarcomas. The need for appropriate methods of detection of clusters is also further strengthened by the recent interest in surveillance for bioterrorism where clustering could be a vital clue to the existence of an attack. The study is aimed at 1) developing and evaluating singular information methods in multilevel semiparametric models for surface estimation of relative risk with sparse cancer data, 2) testing different distributional assumptions for random effects and 3) applying the methods to clusters detection in small area cancer studies. Multilevel semiparametric models are constructed to include multiple sets of random effects associated with nested partitions of the entire territory of interest that allow great flexibility for testing unusual rates within smaller parts of larger areas. Testing departures from the null value of no unusual rates is equivalent to testing the one or multiple variance components equal to zero and is approached as a singular information problem. Construction of confidence intervals is also studied. The random effects are assumed to follow a variety of different distributions, such as heavy-tailed, skewed and bimodal distributions. The expected result is a likelihood-based approach to fitting robust linear mixed models in a wider range of applications than was previously possible.