Sophisticated computer programs known as geographic information systems (GISs) have revolutionized the analysis of spatially referenced datasets, through their ability to """"""""layer"""""""" multiple data sources over a common study area. However, methods for statistical inference on these complex and often spatially and temporally misaligned datasets are only now beginning to develop. In this proposal we develop spatial statistical methodology in seven specific aim areas related to cancer control and epidemiology. First, we consider hierarchical models for cancer control, developing both univariate and multivariate models for analyzing cancer mortality, incidence, staging, and screening data. Second, we propose common spatial factor models for explaining correlations among cancer mortality or incidence rates at different locations. Third, we develop enhanced spatial lattice models for exploring the relationship between various community factors (e.g. smoking levels, education, poverty, health care access, etc.) and cancer-related behaviors (e.g. frequency of breast exam). Fourth, we propose flexible spatial process models for modeling multivariate carcinogen data, using coregionalization. Fifth, we generalize the notion of the spatial CDF to covariate-weighted, conditional, and fully bivariate versions, and propose its use in analyzing possibly multivariate cancer-related exposures. Sixth, we develop spatial cure rate models, and suggest their application to spatially associated smoking quit rate data. Seventh, we propose spatial directional gradient methods that enable identification of and inference for spatial rates of change in carcinogen surfaces. We provide several cancer-related examples to illustrate the methods we propose, as well as an outline of our vision for linking the Markov chain Monte Carlo computing our methods require with existing GIS mapping and database tools.
Showing the most recent 10 out of 33 publications
