Abstract

“In this talk I will share a decade of experience and excitement of being involved with the Michigan Genomics Initiative, a longitudinal biorepository at the University of Michigan Health System, containing administrative healthcare data linked with genetic markers and many other external data sources on nearly 100,000 participants. This is a rich dataset with large p and large n. However, a perioperative recruitment strategy induces selection bias in the analytic sample. In addition, we have an incomplete capture of disease history leading to phenotype misclassification. I will present strategies to understand and reduce bias when these two sources are at play. Examples from COVID -19 and Cancer will be used to illustrate that foundational principles of study design and sampling are critical to turn big data into reliable knowledge.”