Case-control studies are dominant analytic tools in epidemiologic researchfor identifying potential risk factors of a disease. We explore three atypical data situations under a case-control sampling framework.We adopt the Bayesian paradigm as the inferential strategy.The first problem we consider is modeling disease subtypes in matched case-controlstudies using the stereotype regression model. The stereotype regression model (Anderson, 1984), is a relatively unexplored class ofmodels for categorical outcomes, and can be adapted to model both ordered and unorderedcategorical outcomes. Classical frequentist inference for this model is problematicdue to non-linearity and lack of identifiability in the parameter space. Wepropose a general Bayesian analyses and then extend it to deal withnon-ignorable missingness in the covariates. We illustrate our methods by modeling cancerstages in studies of prostate and colorectal cancer.The second problem involves modeling gene-environment interactions under a two-phasesampling design. We consider the situation where the Phase I samplecontains basic demographic and environmental covariates. The Phase II sampleis selected by stratified sampling conditional on case-control status andenvironmental exposures. Genotype data with potential non-monotone missingnessis available only on Phase II samples. We build a semi-parametric Bayesian model thatdata adaptively relaxes both the gene-gene and gene-environment independence assumptions.We introduce a variable selection strategy that can simultaneously handle multiple genes, environmentalcovariates and their interactions. We compare the Bayesian methodology withweighted and pseudo likelihood approaches.The third problem is motivated by a serial case-control study on diarrheal diseaseincidence in Ecuador where disease outcomes were recorded with geographicalcoordinates in a sample of 21 villages over a period of six years. We propose a Bayesiantwo stage spatio-temporal point process model to explain variation in diarrheal casepatterns by using a log Gaussian Cox process with spatial and temporal components.Beyond estimation of model parameters, we also consider the problem of predictingthe number of diarrheal cases at unsampled communities and compare ourprediction with that obtained by a standard Kriging approach.
【 预 览 】
附件列表
Files
Size
Format
View
Bayesian Modeling of Epidemiologic Data under Complex Sampling Schemes.