This study demonstrates that it is possible to identify primary care physicians (PCPs) who perform better or worse than expected in managing diabetes. Study subjects were 14 033 adult diabetics and their 133 PCPs. Logistic regression was used to predict the odds that a patient would have uncontrolled diabetes (defined as HbA1c ≥8%) based on patient-level characteristics alone. A second model predicted diabetes control from physician-level identity and characteristics alone. A third model combined the patient- and physician-level models using hierarchical logistic regression. Physician performance is calculated from the difference between the expected and observed proportions of patients with uncontrolled diabetes. After adjusting for important patient characteristics, PCPs were identified who performed better or worse than expected in managing diabetes. This strategy can be used to characterize physician performance in other chronic conditions. This approach may lead to new insights regarding effective and ineffective treatment strategies.