Hamilton, NJ, USA: Center for Health Care Strategies.
Center for Health Care Strategies, Hamilton, NJ, USA
Washington state began enrolling Medicaid SSI (low income, disabled) recipients into managed care in 1995 on a county-by-county basis. In 1996 the state decided to incorporate stop loss requirements and risk adjusted rates to help contracting health plans manage their risk. Implementation of risk adjusted capitation rates was scheduled for late 1998. The University of Washington (UW) risk adjustment research team obtained funding from the Center for Health Care Strategies to develop and test a method of risk assessment that could be used by Washington state for risk adjusting capitation rates for Medicaid SSI enrollees.
Risk adjustment of capitation is necessary if illness is not evenly distributed among contracting health plans for a given enrolled population. Risk assessment, the first step toward risk adjustment, is the process used to identify how risk is distributed among a population and to determine whether the risk can be statistically predicted. The UW team applied methods developed in previous research to test risk assessment methods for the Washington low-income, disabled population.
The UW team outlined five questions it sought to address through risk assessment:
The project used two years of claims data, 1994 and 1995, edited to reflect services that would be covered under Medicaid managed care. Medicaid SSI enrollees are largely adults and are evenly distributed by sex. More than 90 percent of the population used health services in each year, with over nine percent using more than $50,000 in services per year. Use rates were similar in both study years. Mean annual expense ranged from $3,610 to $4,672 among the eight geographic regions used by the state to subdivide the population.
The study applied three health status groupers, or proxy health status measures, to the Washington Medicaid SSI data: the Disability Payment System (DPS), Diagnostic Cost Groups (DCGs) and Ambulatory Care Groups (ACGs). Each grouper uses diagnostic codes and different algorithms to cluster individuals by clinical severity and, in the case of DCGs, by predicted expense. Each grouper was applied to the study data to group the population by clinical composition based on diagnoses.
The grouped data were applied to two statistical models to predict 1995 expenses from 1994 diagnoses. The response (dependent) variable was measured as the total 1995 charges for services that would be covered under Washington’s capitated rate. Covariates (independent variables) included health status (as measured by one of the health status groupers), age, and, where DPS was used as the health status variable, by a series of interaction terms. Early in the study, the Washington Medicaid administration selected DPS as the grouper of choice. Therefore, while the study performed overall group comparisons of the relative performance of DPS, DCGs and ACGs, most of its work focused on DPS as the measure of health status. Two functional forms of the model were compared: a generalized linear model (GLM) and an ordinary linear regression (OLS). The GLM model has two parts: the first part predicts whether enrollees will or will not use any service, and the second part predicts expenses among those who use services. The OLS model does not separate users from non-users in the prediction.
The study used both global measures, R2 and the Akaike Information Criterion (AIC), and subgroup analyses to compare alternative models. Subgroup analyses included comparison of predictions among geographic regions, among DCG groups and counts of DPS groups (proxies for illness groups).
Using global measures, all models that used health status groupers showed greater predictive ability than a model using age alone. Analyses of subgroup predictions show GLM models uniformly overpredict expenses by varying degrees, while the OLS models generally underestimate expenses; however, this result may be overstated due to bias toward OLS models of the prediction ratio which was used as the measure of performance.
There are several key policy implications of the empirical results of the study. Because of health plans’ likely concern about enrolling a disproportional number of the relatively large number of high cost individuals in this population, risk adjusting capitation payments will increase plans’ willingness to serve this population. The consistency from these two years of fee-for-service data appear adequate to expect that a risk assessment model based on these data would likely apply to later years of health care use for this population. However, this study cannot fully anticipate changes in health service use that may come from the move to managed care. The finding of a large amount of variation in annual expense across geographic regions suggests that it is likely the variation in per enrollee expense across health plans will be too great to justify a single capitation rate payment system, and that risk adjustment is necessary. The overall model results, as measured by R2 and AIC, indicate that DPS performs as well as DCGs and ACGs in predicting subsequent health care expenses. Any of the three health status measures adds a great deal of explanatory power over models with only age variables. While the choice between OLS and GLM models is not clear at the overall population level using global measures, the OLS model appears to perform better within subgroups. Finally, the results of this study suggest that risk adjustment is an important component of capitated payment systems for populations with disabilities. The robustness of the empirical models developed suggest that the fact of risk adjustment may be more important than its form.
Independent of the efforts to develop a health status-based risk adjustment method for Healthy Options SSI, Washington’s Medicaid agency decided in November 1997 that the managed care program would be terminated. The principal reason for the termination cited by the agency was the higher-than-anticipated short-term costs to the state of the move to managed care. The more than 10,000 clients enrolled in the program up to that date were returned to the fee-for-service program. The state is studying the factors that led to the unexpected high costs to ensure that the lessons learned from the experience are not lost.