Medicare Risk Adjustment with Systematically Missing Data

• Activity Number: 421 - Contributed Poster Presentations: Health Policy Statistics Section
• Type: Contributed
• Date/Time: Tuesday, August 1, 2017 : 2:00 PM to 3:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Savannah Bergquist* and Tim Layton and Thomas G. McGuire and Sherri Rose
• Companies: Harvard University and Harvard Medical School and Harvard Medical School and Harvard Medical School
• Keywords: Ensemble learning ; Propensity score ; Prediction ; Risk adjustment ; Medicare

Abstract

Currently, CMS pays Medicare Advantage (MA) plans, the private option in Medicare, based on cost patterns generated by Traditional Medicare beneficiaries. Individuals who choose to enroll in MA differ from Traditional Medicare enrollees in terms of health status, race/ethnicity, age, and income. MA comprises 31% of total Medicare enrollment. This paper proposes a new empirical strategy to address the problem of systematically missing MA enrollee diagnoses data by drawing a sample of Traditional Medicare beneficiaries who resemble MA enrollees. We employ a two stage machine-learning-based propensity score matching approach that relies on a common source of survey data across MA and Traditional Medicare enrollees for the first stage and Medicare claims data for the second stage. In both stages we use an ensemble machine learner to estimate the propensity score. We then estimate the MA risk adjustment formula in the matched Traditional Medicare sample and full Traditional Medicare sample to predict costs and assess the impact of sample selection on risk adjustment. To date, propensity scores have almost exclusively been used for causal inference and not prediction.