Large-scale repositories of secondary-use patient data are emerging as a critical resource for both clinical and epidemiological research. Motivated by this opportunity, a variety of interactive visual analysis methods have been developed to make the use of this data more efficient and accessible. These techniques often combine interactive filters and on-demand computational analysis to allow ad hoc cohort exploration and refinement. This approach has indeed made it possible to quickly select and revise cohorts during analysis. However, the seemingly simple filters supported by these tools can produce dramatic—and often unseen—confounding effects on the makeup of the cohort across the thousands of variables often found in real-world medical data. This poster presents an approach to measuring and visually conveying to users the degree of drift in representation during iterative visual cohort selection.