Medicare Risk Adjustment Development by Johns Hopkins Chad Abrams, MA Cabrams@jhsph.edu Johns Hopkins University School of Hygiene and Public Health 624 N Broadway #600 Baltimore, Maryland 21205 June 6, 2004 San Diego CA Copyright 2003, Johns Hopkins University, 10/19/2003 Objectives • To provide an overview of JHU’s work on Medicare risk adjustment • To summarize what we have learned • To discuss recent findings and how the ACGPredictive Model is being refined for the elderly Copyright 2004, Johns Hopkins University, 5/20 2 Long History of Working with Medicare Data Final Reports Delivered to Center for Medicare & Medicaid Services (formerly HCFA) 1996 Risk-Adjusted Medicare Capitation Rates Using Ambulatory and Inpatient Diagnoses 2000 Updating & Calibrating the Johns Hopkins University ACG/ADG Risk Adjustment Method for Application to Medicare Risk Contracting 2003 Development and Evaluation of the Johns Hopkins Univeristy Risk Adjustment Models for Medicare+Choice Plan Payment Copyright 2004, Johns Hopkins University, 5/20 3 Better Modeling or Better Data Quality? 4 Project: Year-1 diagnoses used to predict year-2 Medicare expenditures Project 1: 1991-1992 Explanatory Power Of JHU Model Project 2: 1995-1996 8.4 Project 3: 1996-1997 9.1 Copyright 2004, Johns Hopkins University, 5/20 5.5 Components of the Basic Model Selected ADGs 13 ADGs demonstrated to have a significant impact on future resource use Hospital Dominant Marker A marker indicating high probability of a future admission Copyright 2004, Johns Hopkins University, 5/20 5 The HOSDOM Marker • Persons with a HOSDOM diagnosis have a high probability (usually greater than 50%) of being hospitalized in the subsequent time period. • Based on two-years of Medicare claims data and careful clinical review • A single concise list of 266 “settingneutral” diagnosis codes. Copyright 2004, Johns Hopkins University, 5/20 6 Examples of HOSDOM Diagnoses 491.21: Obstructive Chronic Bronchitis with Acute Exacerbation 518.81: Acute Respiratory Failure 584.9: Acute Renal Failure, Unspecified 198.5: Secondary Malignant Neoplasm, Bone 785.4: Gangrene 518.4: Acute Lung Edema, Unspecified 789.5: Ascites 571.5: Cirrhosis of Liver without mention of alcohol 403.91: Hypertensive Heart Disease with Renal Failure 284.8: Aplastic Anemia Copyright 2004, Johns Hopkins University, 5/20 7 Impact of HOSDOM on Resource Consumption 0 HOSDOMs Percent Of Pop. 90.7% 8 Year 2 Costs (relative weight) 0.82 1 HOSDOM 7.4% 2.45 2 HOSDOM 1.5% 4.25 3 HOSDOMs 0.3% 5.74 4 HOSDOMs 0.05% 6.59 5+ HOSDOMs 0.01% 7.86 Data Source: 1996-97 Medicare 5 Percent Sample Copyright 2004, Johns Hopkins University, 5/20 Other Variables Considered • Frailty Marker –A list of 75 codes that appear to clinically describe frail beneficiaries. –Divided into 11 “clusters” each representing a discrete condition consistent with frailty. • Selected Disease Conditions –Johns Hopkins Expanded Diagnosis Clusters (EDCs) Copyright 2004, Johns Hopkins University, 5/20 9 Percent of Beneficiaries with Frail Clusters Cluster Description 1 2 3 4 5 6 Malnutrition Dementia Impaired Vision Decubitus Ulcer Incontinence of Urine Loss of Weight 7 8 9 10 11 Incontinence of Feces Obesity (morbid) Poverty Barriers to Access of Care Difficulty in Walking Copyright 2004, Johns Hopkins University, 5/20 Percent of Percent of all Elderly DualEligibles 0.08% 0.20% 0.82% 2.64% 0.25% 0.65% 1.08% 3.05% 0.04% 0.05% 2.51% 4.40% 0.15% 0.03% 0.00% 0.02% 2.88% 0.20% 0.07% 0.01% 0.05% 4.37% 10 Impact of Frail on Resource Consumption Number of Frail Clusters Percent of all Elderly Year 2 Costs (relative weight) 0 93.8% 0.9 1 5.7% 1.9 2 0.5% 3.0 3 0.04% 4.0 4 0.005% 3.5 Copyright 2004, Johns Hopkins University, 5/20 11 Results: What Have We Learned? Copyright 2003, Johns Hopkins University, 10/19/2003 13 1) The frailty variable increases explanatory power AND provides greater predictive accuracy Data Source: 1996-97 Medicare 5 Percent Sample Copyright 2004, Johns Hopkins University, 5/20 14 2) Be careful. Higher R2 and improved accuracy for top quintiles may result in substantial overpayment for first quintile. Data Source: 1996-97 Medicare 5 Percent Sample Copyright 2004, Johns Hopkins University, 5/20 15 3) Sometimes the kitchen-sink approach works Data Source: 1996-97 Medicare 5 Percent Sample Copyright 2004, Johns Hopkins University, 5/20 Comparison to CMS 61-Disease Model and HCC Data Source: 1996-97 Medicare 5 Percent Sample *61-Disease Model the then “current” model as of Nov. 2001. ** HCC model results from Pope et all Dec 2000 Copyright 2004, Johns Hopkins University, 5/20 16 The Goal-Ideally, payment models should pay appropriately for sick individuals while at the same time removing or reducing traditional incentives for promoting biased selection Copyright 2004, Johns Hopkins University, 5/20 17 How are we doing? • Current technologies probably not adequate • Re-insurance and/or carve-outs are still necessary to assure adequate payment for treating high cost patients • R-squared is probably NOT the correct criteria for evaluating model performance Copyright 2004, Johns Hopkins University, 5/20 18 Conclusions 19 • The type of variables included matters • In general, disease specific markers –do not provide adequate payment for the sick, and –possibly lead to substantial overpayment for healthy individuals • Markers such as “hospital dominant” (likely to lead to a hospitalization) and “frail-symptoms” (a proxy for ADLs) successfully target the sick without falsely identifying healthy Copyright 2004, Johns Hopkins University, 5/20