Controlling for unobserved confounders in observational studies using large health care databases by means of instrumental variables in time-to-event analysis

Randomized controlled trials cannot provide all necessary information about drug reactions as they are limited by several factors. Observational studies in free-living populations are therefore necessary. Large health care databases are frequently used for this purpose. A common problem of analyses based on these databases is that confounding variables are often not recorded, so that effects are inconsistently estimated. Under certain conditions instrumental variables can eliminate confounding bias, so that IV estimators can consistently estimate treatment effects. Instrumental variable methods are well established for continuous outcomes using linear regression models, where two-stage least squares are typically used. However, in time-to-event analysis no such common instrumental variable method exists. Even if the proportional hazards model is used, two-stage estimators to account for instrumental variables are only justified for rare events. The aim of this thesis is therefore to explore two-stage instrumental variable estimation for time-to-event outcomes in large health care databases if the assumption of rare events does not hold true.