Abstract
OBJECTIVE: To develop an empiric approach for evaluating the performance of physician peer groups based on patient-sharing in administrative claims data. DATA SOURCES: Surveillance, Epidemiology and End Results-Medicare linked dataset. STUDY DESIGN: Applying social network theory, we constructed physician peer groups for patients with breast cancer. Under different assumptions of key parameter values-minimum patient volume for physician inclusion and minimum number of patients shared between physicians for a connection-we compared agreement in group membership between split samples during 2004-2006 (T1) (reliability) and agreement in group membership between T1 and 2007-2009 (T2) (stability). We also compared the results with those derived from randomly generated groups and to hospital affiliation-based groups. PRINCIPAL FINDINGS: The sample included 142 098 patients treated by 43 174 physicians in T1 and 136 680 patients treated by 51 515 physicians in T2. We identified parameter values that resulted in a median peer group reliability of 85.2 percent (Interquartile range (IQR) [0 percent, 96.2 percent]) and median stability of 73.7 percent (IQR [0 percent, 91.0 percent]). In contrast, stability of randomly assigned peer groups was 6.2 percent (IQR [0 percent, 21.0 percent]). Median overlap of empirical groups with hospital groups was 32.2 percent (IQR [12.1 percent, 59.2 percent]). CONCLUSIONS: It is feasible to construct physician peer groups that are reliable, stable, and distinct from both randomly generated and hospital-based groups.