Designs and analytical strategies to control for unmeasured confounding in studies based on administrative health care databases

Abstract

Studies based on routine data of statutory health insurances require the adequate consideration of unmeasured confounders. This thesis investigated two methods to cope with this problem: (i) Classic two-phase designs collect additional data for a stratified subset (phase 2) of all patients (phase 1). An extension was proposed that does not need a stratification of the data but a proper model for participation in phase 2. A simulation study comparing the extended method with multiple imputation (MI) as an alternative revealed that MI resulted in less biased and more precise estimators of the treatment effect. (ii) The high-dimensional propensity score (HDPS) algorithm automatically selects hundreds of empirical confounders from the underlying database, but was mainly applied in pharmacoepidemiology. This thesis investigated the HDPS algorithm in a study in health services research, where a shift in the effect estimates towards a more plausible result was achieved. The further development of these or similar methods is needed in the near future, since studies based on linking routine data with other data source are gaining in importance.