Bayesian Analysis for Semicontinuous Data Distribution: Forecasting Hospital Cost of Care Using Double Hurdle Model

• Activity Number: 82
• Type: Contributed
• Date/Time: Sunday, July 29, 2012 : 4:00 PM to 5:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Hsiu-Ching Chang*+
• Companies: BlueCross BlueShield of Michigan
• Keywords: Bayesian ; double hurdle model ; MCMC ; zero-inflated

Abstract

In the health care industry, future medical expense is usually predicted through parametric regression modeling (e.g., logistic regression or normality-based regression). However, more cares are needed under this type of regression modeling because the distribution of hospital costs tends to be skewed and zero-inflated. Although log transformation on the outcome is widely applied so that a symmetric error term can be used, the ad hoc log transformation might fail to establish suitable results. In this study, we implement Bayesian approach via Markov chain Monte Carlo (MCMC) to propose an appropriate Box-Cox transformation on the outcome under the assumption that zero values are generated from two different sources: (1) patients do not receive any treatment; the hospital costs are therefore zero, (2) patients do receive treatments with costs not precisely known but smaller than some value C. Using actual member-level claim data, this study characterizes the source of zero by computing the elasticities and marginal effects with a generalized double hurdle model.