Online Program

Comparison of One-Part Models and a Two-Part Marginalized Model for the Analysis of Health Care Expenditures

*Valerie Smith, Department of Veterans Affairs; University of North Carolina at Chapel Hill 
Brian Neelon, Medical University of South Carolina 
John Preisser, University of North Carolina at Chapel Hill 
Matthew Maciejewski, Duke University Medical Center 

Keywords: semicontinuous data, two-part models, health care expenditures, marginalized models

In health services research, it is common to encounter semicontinuous data, such as medical expenditures, characterized by a point mass at zero and a continuous distribution with positive support. These data are often analyzed using two-part mixtures that separately model the probability of health services use and the distribution of positive expenditures among users. Because the second part conditions on a nonzero response, conventional two-part models do not provide a marginal interpretation of covariate effects on the overall population of users and nonusers, even though this is often of greatest interest. We propose a marginalized two-part (MTP) model that yields more interpretable effect estimates on the marginal mean. Using simulations, we compare the performance of the MTP model to one-part generalized linear models (GLMs) fit via quasilikelihood. We simulate data with varying distributions, sample sizes and percentages of zeros. Estimates of covariate effects and type I error rates under the MTP model remain robust. GLMs result in increased bias and inflated type I error rates compared to the MTP model, with bias and type I errors increasing with the proportion of zeros.