Twopart random effects models for longitudinal cost data
View Presentation
*Matthew L Maciejewski, Durham VA HSR&D & UNC
Maren Olsen, Durham VA HSR&D & Duke
Keywords: expenditures, two part model, estimation
Certain types of health expenditures, such as inpatient expenditures, have a predominance of observations at zero and a continuous, often skewed, distribution among the remaining responses. In crosssectional analyses of such expenditures, twopart models can be estimated in a straightforward way because the two parts are separable. For longitudinal analyses of such expenditures, analysts may extend the twopart model to fit a generalized linear mixed model for the probability of a nonzero response over time and a conditional linear mixed model for the mean response given that it is nonzero. However, the zero and nonzero data generating processes may be serially correlated in longitudinal data and a more appropriate twopart model would include jointly distributed random effects for the zero and nonzero processes. We apply both types of longitudinal modeling to annual specialty expe
