Utilizing auxiliary covariate data in epidemiologic studies with binary outcome

2014-01-30In this dissertation we consider methods for using auxiliary information from a generalized “two-phase” design and analysis problem. A typical configuration of such studies involves subsampling from a large epidemiologic study, or the “first phase” sample. Outcomes, auxiliary covariates, and sometimes a component of exposures are measured on the “first phase” sample. The subsample, or the “second phase”, usually consists of detailed covariate and exposure assessment. When the “first phase” comprises a set of risk sets from a cohort and the subsampling is outcome-dependent, it is a nested case-control design. When the “first phase” is a case-control sample as in a standard case-control study or a simple random sample as in a cross sectional study, and the subsampling is outcome‐auxiliary‐dependent sampling, it is a two‐phase case-control design. ❧ The two specific lines of research that are covered in detail are a novel estimation framework for unmatched nested case-control studies and a design optimization problem with respect to odds ratio estimates in two-phase case‐control studies. The first line of research develops a family of method‐of‐moments estimators for the proportional odds model, motivated by the need for a valid variance estimator for the baseline odds parameter and subsequently confidence intervals on the absolute risk. This method matches true cohort totals to their unbiased estimates—functions of parameters of interest—based on the subsample, and solves these estimating equations to obtain estimators. The second line of research seeks to optimize two‐phase case‐control designs with respect to the cost-efficiency of odds ratio estimation, building upon an established semiparametric likelihood developed by Breslow and Holubkov. The optimization yields subsampling fractions that achieve the minimal asymptotic variance of odds ratios under cost constraint for each of the outcome‐auxiliary defined strata