Modeling Health Expenditures Using Generalized Linear Models

Abstract

Healthcare costs have surged dramatically: from $74.1 billion in 1970 to $1.4 trillion in 2000, and reaching $4.5 trillion by 2022(McGough et al. 2023). This escalation puts immense pressure on both systems and individuals, highlighting the need for advanced statistical models to inform policies, set insurance premiums, and guide financial planning. Our focus is on the 2003 Cohorts 7 and 8 from the Medical Expenditure Panel Survey (MEPS), where data shows many zeros and a long-tailed distribution for non-zero expenditures. Inspired by my mentor, Professor Yang’s work on generalized linear models (GLMs) for cost estimation, this study applies a Two-Part Model (TPM) and Tweedie GLM, incorporating AIC and Lasso for variable selection. Previous research by Frees, Gao, and Rosenberg (2011) extended the TPM to predict healthcare costs(Edward W. Frees and Rosenberg 2011), while other studies have compared GLMs with Tweedie GLMs for aggregate claims(Quijano Xacur and Garrido 2015).