Is a Risk-Adjusted Bootstrap a Better Way to Find "Bad Apple" Providers?

• Activity Number: 352
• Type: Invited
• Date/Time: Wednesday, August 6, 2003 : 10:30 AM to 12:20 PM
• Sponsor: Section on Health Policy Statistics
• Author(s): Dina Alper*+ and Yang Zhao and John Haughton and Arlene S. Ash
• Companies: DxCG, Inc. and DxCG, Inc. and DxCG and Boston University School of Medicine
• Keywords: provider profiling ; DCG ; report cards

Abstract

Health care oversight requires comparing the outcomes that providers achieve with their expected outcomes due to patient health burden. Provider "profiling" involves calculating an observed and expected outcome (O and E) for each provider; summarizing the discrepancy as, say, O/E; determining which providers' O/Es are "unacceptably" different from 1.0. When a good risk adjuster is used to determine the Es and "chance" discrepancies are ignored, "outlier" O/E values can point to practices that really are worse (or better) than typical. We use a DCG model to predict annual costs for each patient from demographics and diagnoses (R2~ 0.50) and to calculate providers' Es in a large health insurance database. For each provider, we construct a bootstrapped distribution of O/E for patient panels with the same n and DCG distribution. We use these to construct "real" 95%-acceptance intervals, compare these intervals to analytically calculated ones, and summarize our findings as to how well the analytic intervals conform to reality. We also explore conceptual and practical issues in provider profiling.