With the growing need to develop appropriate mechanisms for reimbursing managed care plans in an equitable and efficient manner, the use of risk-adjustment to set capitation payments has been extensively studied. Researchers have found that risk-adjustment methodologies based on claims diagnoses, such as the Ambulatory Care Groups (ACGs) and Diagnostic Cost Groups (DCGs), are better at prospectively predicting total expenditures than simple age-sex adjustments (Anderson, Steinberg, Powe et al., 1990; Dunn et al., 1995; Newhouse, 1994; Newhouse et al., 1989). However, less is known about whether risk-adjustment methodologies that work in a particular context may successfully be used in other applications. One such question is the applicability of commercial risk-adjustment systems, which were developed to set capitation rates for integrated health plans (i.e., capitated plans that cover both medical and psychiatric care), to efforts to adjust payments for behavioral health care "carve-out" vendors.
Under a risk-based carve-out arrangement, the vendor receives a capitated payment per covered life in exchange for financing and managing all services within a given category, in this case mental health and substance abuse (MH/SA) services. A separate health plan is responsible for the medical services used by the same enrollees. Behavioral health care "carve-outs" have become increasingly common. The proportion of privately insured U.S. citizens whose behavioral health benefits were managed in some form or another is estimated to have increased from 44 percent in 1992 to 75 percent in 1997 (Oss et al., 1997); this figure includes risk-based network programs, utilization review programs, employee assistance programs, nonrisk-based network programs, integrated programs and internal management of behavioral health care services within health maintenance organizations (Oss et al., 1997). Health maintenance organizations (HMOs) are entities that provide and manage the coverage of health services provided to plan members in return for a fixed premium (Rognehaugh, 1996). Risk-based network programs, which are capitated plans responsible for behavioral health care only, increased from 16 percent of the population in 1993 to 22 percent in 1997 (Oss et al., 1997). Thus one potentially important use of risk-adjustment is to determine the appropriate capitation amount to pay behavioral health care carve-out vendors.
Risk-adjustment can be useful both when a single carve-out vendor serves the entire patient population and when multiple vendors compete to enroll patients. When purchasers (employers or government programs) contract with a single carve-out vendor, risk-adjustment is useful for modifying payments as the behavioral health care needs of the population change over time. For example, Medicaid, which is the primary U.S. insurance program for the indigent and is jointly financed through Federal and state funds, is a major purchaser of managed care services, including those provided by behavioral health care carve-out vendors. The typical duration of a Medicaid contract with a carve-out vendor is three years. The behavioral health care needs of the population could change substantially over such a long period, so the purchaser needs a methodology for purchaserscontract with multiple, competing carve-out vendors, risk-adjustment is critical, for adjusting payments over time so that the carve-out vendor is neither under nor overpaid. The same reasons exist with integrated health plans: unless capitation payments are adequately adjusted for the health care needs of individuals, plans have an incentive to try to attract healthy, low-cost patients ("cream-skimming") and avoid enrolling sick, high-cost patients ("dumping").
Despite the need to develop viable risk-adjustment methodologies for behavioral health care, most existing systems were designed only to predict total medical expenditures. Earlier research suggests that these systems tend not to perform particularly well in predicting expenditures on behavioral health care (Ettner et al., 1998; Ettner and Notman, 1997). This study focuses on the use of risk adjustment to set capitation payments for behavioral health care carve-out vendors serving populations that include people with psychiatric disabilities. We use data on privately insured adult employees and dependents to examine whether existing risk-adjustment models sufficiently adjust capitation payments to behavioral health care carve-outs for psychiatric disability.
In particular, the study addresses a number of related questions.
How well do the best-known commercial risk adjustment systems, the ACG models and the DCG models, account for psychiatric disability in prospectively predicting behavioral health care expenditures?
Can the performance of these methodologies be improved by taking psychiatric disability explicitly into account?
What is the potential for competing behavioral health care carve-outs to experience profits or losses under each risk-adjustment system, in the presence of patient self-selection into health plans?
In comparison with pure capitation, do systems that mix capitation and fee-for-service methods reduce the potential for behavioral health care carve-out vendors to be "winners" or "losers," depending on whether they enroll mentally healthy or sick patients?
1. Methods and Data
The analyses were based on 1994-1995 medical and behavioral health care claims from Employer B, a large employer in the northern United States. Claims for pharmaceutical services were not available. Employees chose between several types of health plans, including 11 HMOs, an indemnity plan, and a preferred provider organization (PPO).
All of the HMOs were similar in that enrollees were required to choose primary care physicians to manage their care and to choose providers from closed panels. Furthermore, all of the HMOs fully covered inpatient services after pre-certification. However, the HMOs differed slightly in their coverage of outpatient MH/SA benefits. Most of the HMOs charged $10 per visit up to a maximum of either 20 visits or $500 per year. Several HMOs increased cost-sharing requirements depending on the number of visits, (e.g. charged $10 for visits 1-8 and $35 for visits 9-20). Among the most generous HMOs was one that covered up to 30 visits per year with the first 10 visits free, and another that fully covered the first $500 per year and then imposed 50 percent cost-sharing for costs up to $800 or 20 visits per year, whichever came first. Nonetheless, none of the variations in mental health benefits among HMOs appeared to be substantial. Similarly, differences in substance abuse coverage were modest, with the exception of one HMO that offered unlimited visits at its health centers at a constant $10 per visit.
Employees who did not choose to enroll in an HMO received their MH/SA services through a single behavioral health care carve-out vendor. Carve-out enrollees were allowed a choice between in-network and out-of-network providers, with lower cost-sharing requirements for care received from in-network providers. Carve-out enrollees also had access to care managers and an Employee Assistance Program offering assistance with problems of daily living. Carve-out enrollees were not required to meet any deductibles and also had out-of-pocket maximum payments ($1,000 for in-network and $3,000 for out-of-network). The carve-out vendor completely reimbursed inpatient services for MH/SA problems and intermediate care provided by in-network hospitals (including general, psychiatric and substance abuse). Hospitalizations in out-of-network facilities were reimbursed at 80 percent, with deductibles and some limits. Copayments for outpatient visits to in-network providers were $0 for visits 1-4, $20 per visit for visits 5-25, and $40 per visit for visits 26 or higher; outpatient visits to out-of-network providers were reimbursed at 50 percent for up to 15 visits per year. To be reimbursed, services had to be authorized as medically necessary.
2. Definition of the Sample
The population on which the analyses were based included all non-elderly adult employees, early retirees (less than 2 percent of the population) and dependents who obtained insurance coverage through employees. Sixty-one percent of Americans receive their coverage through an employment-based plan (Custer, 1999), so our sample is representative of a large segment of the population. We excluded beneficiaries who were not continuously enrolled in one of the health plans for the two-year period or who were enrolled in one of four HMOs for which data could not be obtained. Analyses were performed separately for beneficiaries who obtained MH/SA services through an HMO vs. those in the behavioral health care carve-out (indemnity and PPO enrollees). Final sample sizes were 56,174 and 52,990, respectively.
The dependent variable in the analyses is 1995 MH/SA expenditures. MH/SA claims were identified as those having major diagnostic category (MDC) 19 or 20. MDC 19 corresponds to "mental diseases and disorders"; MDC 20 corresponds to "alcohol/drug use and alcohol/drug induced mental disorders." Charges from all claims meeting this criterion were aggregated to construct the expenditure measure.
Explanatory variables are based on 1994 data in order to test the ability of risk-adjusters to prospectively predict 1995 MH/SA expenditures. Each risk-adjustment model controlled for basic demographic characteristics (sex, age, and age squared) and an adjustment for the overall level of medical prices in the geographic location, proxiedby the Health Care Financing Administration (HCFA) wage index. All metropolitan statistical areas (MSAs), defined as core areas containing a large population nucleus in conjunction with adjacent communities having a high degree of economic and social integration with that core, were assigned a hospital wage index. Areas outside MSAs were assigned a state index. Labor expenditures make up the majority of hospital expenditures, which in turn make up a large portion of total health care expenditures, so the HCFA hospital wage index is often used to control for geographic variation in health care prices. Finally, sensitivity analyses that controlled separately for benefit design in the analyses based on HMO enrollees yielded estimates that were almost identical to the baseline estimates; these results are available from the authors upon request.
The baseline model controls for the demographic factors only. Seven alternative specifications adjust for the same demographics as well as the following sets of diagnostic measures, examined in turn:
- an indicator for whether the enrollee has any psychiatric disability;
- indicators for the type of psychiatric disability;
- Ambulatory Diagnostic Groups (ADGs);
- Hierarchical Coexisting Conditions (HCCs);
- ADGs plus indicators for the type of psychiatric disability; and
- HCCs plus indicators for the type of psychiatric disability.
To develop criteria to identify individuals likely to have a psychiatric disability, we studied the National Institute of Mental Health (NIMH) and Substance Abuse and Mental Health Services Administration (SAMHSA) definitions of severe mental illness, examined Centers for Disease Control (CDC) reports of disability days by ICD-9-CM diagnosis (Work-Loss Data Institute, 1996), and reviewed the literature on labor market impacts of specific mental disorders (Bartel and Taubman, 1986; Benham and Benham, 1982; Ettner et al., 1997; Miller and Kelman, 1992). NIMH and SAMHSA are Federal agencies that fund U.S. research and demonstration projects in mental health. Finally, we examined National ComorbiditySurvey data to identify combinations of psychiatric conditions leading to reduced activity levels (Kessler and Frank, 1997). In order to identify psychiatric disability using administrative data, we created a two part definition of severe mental illness, including a per se diagnosis component and a qualified diagnosis component. Virtually all of the sources of information we studied suggested that the following conditions are associated with very high rates of impairment in daily functioning and loss of productivity: schizophrenic or schizoaffective disorders, manic depression/bipolar illness, autism, and recurrent major depression. Thus people who had any inpatient or outpatient claims with these diagnoses were considered to be psychiatrically disabled per se.
Diagnoses of major depression (without chronicity specified) and other psychiatric conditions could not automatically be used to assess disability without further information on severity. We therefore focused on the presence of comorbid conditions to determine whether the patient's psychiatric diagnoses were likely to lead to impairment. Comorbidity combinations likely to have strong adverse effects on productivity (and which were therefore used to identify additional cases of psychiatric disability) included (1) major depression in conjunction with anxiety disorder, and (2) drug or alcohol abuse in conjunction with any other psychiatric disorder. We also explored the use of historical patterns of illness and utilization patterns as markers of severity; however, historical measures could not be constructed, due to data limitations. The use of treatment patterns suggesting severe illness resulted in the identification of very few additional cases of psychiatric disability.11 Furthermore, risk-adjustment based on the type of services received offers perverse incentives to health plans. Thus we did not use treatment patterns as markers for severity in the final analyses.
The model controlling separately for the type of psychiatric disability aggregates all diagnoses and diagnostic combinations defining disability into the following three categories:
- psychosis, including schizophrenic/schizoaffective disorders and manic depression/bipolar illness;
- recurrent major depression or depression in conjunction with anxiety disorder; and
- substance abuse in conjunction with any other psychiatric condition.
4. Analyses and Descriptive Statistics
We first calculated descriptive statistics. MH/SA expenditures were then estimated as a function of each set of risk-adjusters described above. Split-sample estimation was used to avoid overstating the predictive ability of the models. Regression coefficients were estimated with 70 percent of the sample and the measures of predictive ability were calculated with the remaining 30 percent of the sample. Split-sample methods are particularly advisable when overfitting is a potential problem, that is, when the sample sizes within particular cells defined by patient characteristics are relatively small. In the extreme case, a regression estimate might be based on a single observation, leading to a zero residual error for that individual and artificially inflating the measured predictive ability of the regression model when the same sample is used for both estimation and prediction (Duan et al., 1983; Gujarati, 1988; Manning et al., 1987).
5. Two Part Models for Estimating Expenditures
Expenditure measures are generally characterized by a disproportionate number of zero values in conjunction with a highly skewed distribution of expenditures among people using services. This arises because upwards of 20 percent of those in a typical insurance plan use no services during any given year, while many fewer will have extremely high expenditures. We therefore estimated MH/SA expenditures using the "two part model" (Duan et al., 1983). The expected level of total expenditures for an individual is modeled as the product of two parts: the probability of any expenditures, and the level of expenditures given that some expenditure occurred.
The probability of having any expenditure was estimated using logistic regression. The level of expenditures was estimated with linear regression, using the subsample of beneficiaries who had any expenditures. To reduce skewness in this subsample, we used the square root of expenditures in place of actual expenditures. In order to interpret the estimated coefficients in terms of dollars, we then retransformed them using a "smearing" procedure (Duan, 1983).12 We judged the goodness of fit for each risk-adjustment model using a synthetic R2 (Efron, 1978). Like the usual R2, it measures the proportion of variation in the dependent variable (expenditures) that is explained by variation in the independent variables (demographic measures and risk-adjustment variables like HCCs).
6. Prediction Errors for Assessing Payment Methods
The mean and standard deviation of the absolute values of the individual prediction errors (actual minus predicted MH/SA expenditures) are also presented. Due to the nonlinearity of the two part model, these measures (synthetic R2 and mean absolute prediction error) do not necessarily yield the same rank ordering of the models. Where these measures differed in their assessment of the relative performance of the risk-adjustment models, we tended to favor the mean absolute prediction errors as the standard rather than the proportion of variance explained. Because the purpose of risk-adjustment is to tailor payments to expected costs, minimizing prediction errors seems to be the most relevant performance criterion when choosing a risk-adjustment system.
The prediction errors are calculated based on the entire sample, including adults both with and without psychiatric disability. To examine the incentives to "dump" psychiatric patients that remain after risk-adjustment, we calculated predictive ratios (sum of predicted MH/SA expenditures divided by the sum of actual MH/SA expenditures) for two groups: adults with psychiatric disability and those with no psychiatric disability.
7. Profit and Loss by Payment Method
Finally, we compared the ability of full capitation, soft capitation and mixed payment methodologies (defined below) to account for natural selection into health plans. We used actual enrollment in the seven HMOs in the study to provide an example of the range of profits and losses that competing behavioral health care carve-outs experiencing similar selection might incur. For each of the seven health plans, we calculated the average per enrollee profit or loss, based on the plan's patient population. The range in profits/losses across all of the health plans under each of the three payment schemes is given for each risk-adjustment model.
Definitions of Payment Systems
In the full capitation model, health plans are paid 100 percent of the risk-adjusted capitation amount for each enrollee, regardless of what actual expenditures were. In the mixed payment system used for the simulations, the health plan would be paid 50 percent of the risk-adjusted capitation payment and 50 percent of actual costs. The soft capitation payments were calculated two ways: using plan averages and using individual-level data. Soft capitation contracts are generally based on plan averages. The risk-adjusted average payment for all health plan enrollees is specified as the target amount and payments are based on how far away the actual average spending per enrollee is from this target. For comparison purposes, however, we also examined soft capitation applied at the individual level. In that system the payment for each individual enrollee was based on that enrollee's expenditures, relative to the enrollee's target amount and the risk corridor around the target amount.
Two soft capitation models are studied. In the first, the risk corridor is defined to be between 90 and 110 percent of the target amount. Within this corridor, the purchaser and the plans share the risk equally. In other words, plans retain only $0.50 for every dollar saved by reducing expenditures below the target amount; conversely, plans incur only $0.50 of the cost of every dollar spent above the target amount. Outside of the risk corridor (i.e., below 90 percent or above 110 percent of the target amount), the purchaser (not the plan) bears all of the risk: it retains all of the profits or incurs all of the losses. The second soft capitation model is similar, except that the risk corridor is between 75 and 125 percent of the target amount, while the sharing of profits or losses within this corridor is 60 percent by the purchaser and 40 percent by the plan.
Payment based on full capitation is typically adjusted so that the purchaser achieves budget neutrality, meaning that total payments to the health plan by the purchaser equal the total expenditures for services by the health plan. Thus the full capitation amounts were adjusted to be budget neutral in this study. The capitated portion of the mixed payment was also adjusted to be budget neutral, although the overall payment was not necessarily budget neutral because of the actual cost portion. It was assumed that payers would be less likely to impose budget neutrality with a mixed payment system or when using soft capitation, because it would diminish the risk sharing function of these payment methodologies. Thus neither the mixed payment nor the soft capitation amounts were constrained to be budget neutral.