| A key policy issue debated in the context of introducing price competition in Medicare HMO market, is how to improve the demographic risk classification used by HCFA to adjust its capitation premiums to competing plans. This dissertation contributes to the debate at two levels. At the theoretical level, in Chapter 2 we show that improved risk classification system could reduce HCFA's program costs in two ways: reduce selection costs arising from plans' preferred risk selection strategies, and reduce risk premium of plans' bids through lowering of within-variance of the risk classes. At the empirical level, in Chapter 4, we apply our theoretical framework to develop a model of risk classification using prescription drug profiles (based on drug therapeutic classes) as risk-adjusters and test its effectiveness on a 3-year study sample drawn from a large HMO in California. Using base-year information, we apply several statistical model selection criteria to compare the econometric properties of alternative prediction functions of medical cost risk based on four classes of risk adjusters---demographic, diagnostic cost groups, survey scales, and prescription drug profiles (PDP). The predictive performance was tested at three levels: within-sample prediction of next year's costs, out-of-sample prediction of next year's costs, and of two-year future costs. The PDP model emerges as the preferred one at each of the three levels. The ranking of models was robust to alternative econometric specifications of the prediction functions. Chapter 3 reviews the administrative properties of prescription drug data. We find that it enjoys advantages over survey and diagnosis data, in terms of cost and timeliness of data availability, and objectiveness of data with respect to actions by providers. In Chapter 5, we show how modeling of medical cost risk could be improved by incorporating frequency of hospitalization as a binary variable or a Poisson frequency variable in a simultaneous equation framework. Although our findings are based on data from a single HMO population, our comparative study demonstrates that prescription drug data could be a promising candidate as a health risk-adjuster. |
