Clustering to Strengthen a Categorical Instrument

• Activity Number: 368
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Douglas Lehmann*+ and Yun Li and Yi Li
• Companies: University of Michigan and University of Michigan and University of Michigan
• Keywords: Categorical Instrument ; Weak Instrument ; Clustering ; Unmeasured Confounding

Abstract

Instrumental variable (IV) analysis has been widely used to control for unmeasured confounding. Groups of patients (e.g., by clinical centers, physicians, or service areas) have often been proposed as categorical instruments based on the premise that differential medical practices attributable to groups, for example, those arising from differential group policies, mix of insurance coverage, and physicians' preferred drug prescriptions or medical knowledge, are independent of unmeasured confounders for the main outcome of interest. However, weak instruments, such as group-based categorical instruments, are sensitive to IV assumptions and can lead to biased results even with large sample sizes. We propose a clustering method to strengthen the categorical instrument. Through simulations we show that the instrument after clustering is more robust towards violations of the assumption that the instrument is randomly assigned. Compared with near/far matching, our method is data-adaptive in nature and more efficient. These methods are illustrated using Medicare data to study patient outcomes at kidney dialysis facilities.