Enabling personalized, evidence-based medicine requires measuring the similarity between patients in the clinic and patient records in claims and electronic health record (EHR) databases. In this paper we evaluate the performance of a patient similarity measure in the context of distinguishing patients who experienced adverse drug events after taking a new drug from those who did not. We consider two ways of evaluating performance, one that estimates the separation of the two clusters and the other using the weighted subspaces that define the cohort. Our effort highlights the importance of integrative, multi-modal similarity measures and the methods for evaluating their performance.