Insurers, public health officials, providers, and health management professionals are interested in predicting who uses health care and why. The ability to predict who will use services is important for several practical applications, such as health care financing and targeting health promotion initiatives. Recent research suggests that better predictive estimates of individual level health care utilization are now possible. This research seeks to build a better model to predict the use of health services. Specifically, it seeks to predict the health care expenses for the full population of Georgia Medicaid children in 1999. This dissertation research combines the strengths of two leading cost forecasting models in order to address three primary goals: First, to build a better predictive model of health service utilization; Second, to address a particular health policy area, that of children with special health care needs (CSHCN), by providing information to help in their definition and categorization. Third, to build upon previous theory in health service utilization, such as the behavioral model of health service utilization and the principal agent model, using knowledge gained from new methodological approaches. To achieve these goals, a two-part model will be developed, with each portion of the model being treated as a finite mixture of at least two sub-populations. The model is estimated using maximum likelihood estimation, through use of the estimation/maximization (EM) algorithm. The new model is compared to published models to determine improvements in predictive power, using the BIC and the GOF criteria. Information from the new model can be used to design health care financing strategies for children in Medicaid; to identify different groups of CSHCN; and to expand the theory about who uses health services and why. The methodologies used to build this predictive model are directly applicable to several other areas of health services research.
