Instrumental Variable Methods for Accounting for Selection and Survival Bias in Observational Studies
*Therese A. Stukel, Institute for Clinical Evaluative Sciences and University of Toronto 

Keywords:

Confounding frequently occurs in observational studies of the effects of treatments or exposures on health outcomes. While standard statistical methods can remove bias due to measured confounding, non-standard methods are required to remove bias due to unmeasured confounding. This workshop will address several statistical issues in estimating treatment effects when key confounders are unobserved or unobservable. Issues in the design and analysis of observational studies when estimating treatment effects using observational data will be highlighted. We will give an overview of analysis methods for removing confounding, including standard regression and propensity-based methods. We will introduce instrumental variable (IV) methods, providing an overview, properties, strength and validity of a proposed instrument, interpretation and analysis techniques. We will review examples of good and poor IV analyses in the health services literature, with an in-depth review of a study of the effects of invasive cardiac care on AMI mortality. Finally, we will assess which types of studies are more amenable to which techniques and will design a study of antipsychotic medications on patient mortality using varying techniques.