Online Program

An Evaluation of Dynamic Linear Models in Predicting Monthly Medicare Payment per Capita (306769)

Lendie Follett, Drake University; College of Business & Public Administration 
*Maria L. Joseph-King, General Dynamics Information Technology 
Brian O'Donnell, General Dynamics Information Technology 
Yan Qian, General Dynamics Information Technology 
Ying Wei, General Dynamics Information Technology 
Xiaoyin Zhang, General Dynamics Information Technology 
Xiwen Zhu, General Dynamics Information Technology 

Keywords: dynamic linear model, prediction, Medicare

The final value of a claims-based metric is not observed until months after services are provided. We evaluate dynamic linear models (DLM) used to predict the final value of a claims-based metric, given a premature observation. Payment per capita (PPC) for Medicare fee-for-service (FFS) Part A enrollees was computed for months in 2008-2016 based on the FFS claims at 0-12 months after the service (lag), where the final value is observed at 12 months lag. FFS claims were acquired from the Centers for Medicaid & Medicare Services (CMS) Chronic Condition Warehouse (CCW). The mean absolute prediction error in PPC for the DLM with time-varying intercept and no covariate is 10.25±1.35, and 7.19±0.93 accounting for month-of-year. Among the DLMs with fixed slope on the 1-month lag metric, the fixed and time-varying intercept models with no month-of-year component were comparable in mean absolute prediction error, 4.28±2.88 and 4.14±2.68, resp.. The DLM with fixed intercept and fixed slope resulted in the most accurate and precise predictions at the earliest lags. The addition of a fixed component to account for month-of-year improved mean prediction, but increased prediction error variance.