Efficient use of phase 1 information in two-phase case-control studies based on administrative databases

Abstract

A common problem of studies based on administrative data is missing confounder information. Obtaining additional information from another data source for at least a subsample of the study population resolves this problem. Two-phase designs can be employed for the combined analysis of both data sources. In logistic two-phase studies a dichotomous outcome and some covariate information are available for the whole population whereas the full vector of covariates is only observed for a subsample in phase 2. Information of phase 1 is utilised in the analysis of the phase 2 sample via stratification. This approach allows for unbiased parameter estimation and can improve efficiency of the estimation of covariates. The aim of this thesis is the development of strategies for the efficient use of phase 1 information in logistic two-phase studies comprising a multitude of phase 1 covariates. An alternative stratification strategy is proposed, which is based on percentiles of a disease score. The core of the thesis is the development of a design criterion which allows the identification of efficient stratifications at the planning stage of the study.