To use an empirical ayesian approach, blending practice, and group quality data with physician results to increase the accuracy of quality of care measures.
Performance data on diabetes glycemic screening for 8,357 physicians collected from multiple payers as part of a statewide physician performance reporting initiative.
A variance components analysis assessed the strength of group, practice, and physician effects compared with random error. We derived formulas to describe reliability and measurement error variances and calculated the optimal blend of physician, practice, and group data. We constructed a simulation to show what various methods can achieve. The value of blending strategies was assessed by simulating a common pay‐for‐performance criterion—performance in the top 25 percent. We estimated the proportion of physicians whose true percentage would place them in the top 20 percent but who would not receive payment based on the observed success rate.
Blending reduced the error rate from 29.7 to 22.7 percent. Simpler empirical Bayes estimates using shrinkage alone produced no gains over simple doctor percentages.
When good structural data about physician groups and practices exist, gains from blending can be substantial.