| |
M. Alan Brookhart, Ph.D. and Til Stürmer, MD
Department of Epidemiology, Gillings School of Global Public Health, University
of North Carolina, Chapel Hill, NC
Epidemiologic studies are increasingly used to investigate the comparative safety and effectiveness of medical products. This workshop reviews issues related to study design and confounder control in non-experimental studies of medical interventions. We begin the workshop by discussing important sources of confounding bias in observational studies: including confounding by indication and the healthy user effect. We then discuss how the comparative new user study design can mitigate many of these sources of confounding bias. However, even after careful choice of comparator group, there still may exist baseline difference between treatment groups in a comparative new user study. We then discuss the use of propensity scores (PS) as a tool to balance residual differences between groups to control for measured confounders. We describe various approaches to using the PS, including matching, sub-classification, and inverse-probability weighting. One of the central issues with implementation of PS methods is selecting which covariates to include in the PS model. We discuss advantages and disadvantages of different approaches to this problem, including automated high-dimensional approaches versus theoretically-based a priori approaches. We briefly review some approaches that can be used to in the presence of unmeasured confounders, including sensitivity analyses, external adjustment, PS calibration, and instrumental variable methods. Finally, we consider statistical and epidemiologic approaches to handling events that occur during follow-up, such as treatment discontinuation and therapy changes.
|
|
