Using Bayesian Quantile Regression Model with Group LASSO to Identify Key Health Risk Assessment Variables and Evaluate Their Predictive Power in the Patient's Future Medical Costs

• Activity Number: 323
• Type: Contributed
• Date/Time: Tuesday, August 2, 2016 : 8:30 AM to 10:20 AM
• Sponsor: Health Policy Statistics Section
• Author(s): Hsiu-Ching Chang* and Hyokyoung (Grace) Hong and Yu Yue and Min Tao and Darline El Reda
• Companies: BlueCross BlueShield of MI and Michigan State University and Baruch College and BlueCross BlueShield of MI and Michigan State University
• Keywords: Health risk assessments (HRA) ; Claims ; Bayesian quantile regression ; Group LASSO ; Deviance Information Criterion (DIC) ; Conditional Predictive Ordinate (CPO)

Abstract

Each year in the United States chronic diseases cause 7 in 10 deaths and account for approximately 75% of the $2 trillion spent on medical care. In an effort to control the cost burden of chronic disease, increasingly more large Employer groups now offer disease management and wellness programs as a part of their employees' benefit package. Health risk assessments (HRA) are a common element of such care management programs and are often used to identify candidates for program outreach. In this study, we use Bayesian quantile regression with the group LASSO penalty to 1) identify the specific HRA variables that influence a patient's future health care costs at different quantiles of the outcome distribution and then 2) to identify significant risk factors that contribute to high costs. For the continuous variables, we assign second-order random walk priors to the smooth functions used to depict the nonlinear relationships with the outcome. DIC and CPO are used to assess model fit and predictive power gains compared to a model limited to administrative claims data.