Although the problem of adverse selection into more generous health insurance plans has been the focus of decades of work, risk-adjustment systems have only recently begun to be implemented to blunt its effect. The Health Care Financing Administration (HCFA) is employing Diagnostic Care Groups (DCGs) in the Medicare program. Maryland Medicaid is implementing Adjusted Clinical Groups (Wiener et al., 1998), and a number of case studies have recently been published describing the experiences of private sector employers with various risk-adjustment systems (Wilson et al., 1998; Bertkoet al., 1998; Dunn et al., 1998; Knutson et al., 1998; Tollenet al., 1998).
Risk-adjustment systems attempt to predict enrollees' expected future health care expenditures given their current characteristics, most notably their current health status. The predictions are then used to adjust payments to health plans. The intended result is that plans offering generous benefits or services that attract sicker enrollees will not be penalized financially for doing so. Thus, a plan's incentive to avoid high-cost patients will be minimized.
Risk-adjustment is most important for conditions that can be identified and avoided by health plans, especially those associated with a high probability of future service use and costly health care. People with potentially disabling chronic conditions are particularly vulnerable. Health plans may seek to avoid enrollees with chronic conditions through a number of methods. They can gear their marketing materials toward healthier populations, they can limit access to certain types of specialists, they can offer less generous insurance benefits (e.g., they can offer limited prescription drug coverage, high copayments, or minimal coverage for services like psychiatric care), and they can explicitly exclude such people by disallowing payment for specified treatments.
Because of the particular vulnerability of people with chronic illness, it is critical that we understand the ability of different types of risk-adjustment systems to predict their health care expenditures. Risk-adjustment may be better at predicting expenditures of people with certain conditions than others. Furthermore, some types of risk-adjusters may work better for some conditions (e.g., asthma) than for others (e.g., psychiatric disorders).
1. Risk-Adjustment Grouper Systems
Two of the leading risk-adjustment systems for predicting inpatient and outpatient expenditures are Adjusted Clinical Groups (ACGs, formerly called Ambulatory Care Groups) (Wiener et al., 1991) and DCGs (Ellis et al., 1996). These risk-adjustment systems group people into distinct categories based on their diagnoses and secondarily on their age and gender. These categories can then be used to explain why some people have higher health care expenditures than other people.
Adjusted Clinical Groups
The version of ACGs used in this study, Release 4.1, groups people into categories through several steps. First, the ACG software assigns each person's inpatient and outpatient claims to one of 32 Adjusted Diagnosis Groups (ADGs) based on the ICD-9-CM diagnosis code on the claim and five clinical/resource consumption dimensions of that diagnosis code. These clinical/resource consumption dimensions include the following:
"Duration (acute, recurrent or chronic): How long will health care resources be required for the management of this condition?
Severity (minor/stable versus major/unstable: How intensely must health care resources be applied to manage the condition?
Diagnostic Certainty (symptoms versus diseases): Will a diagnostic evaluation be needed (symptoms) or will services for treatment be the primary focus (diseases/diagnoses)?
Etiology (infectious, injury or other): What types of health care services will be used?
Specialty Care (medical, surgical, obstetric, hematology, etc.): To what degree will specialty care services be required?" (Johns Hopkins University, 1998).
|FIGURE 7-1. Creation of ACGs|
|SOURCE: Johns Hopkins University (1999).|
Diagnostic Cost Groups
DCGs use ICD-9-CM diagnostic codes to classify patients based on the clinical similarity of the conditions that are being treated. Patients with similar medical problems are assigned to similar groups. In this study we use a particular version of DCGs (release version 3.0) called Hierarchical Coexisting Conditions (HCCs). The HCC models use ICD-9-CM codes from all clinical encounters except laboratory and other ancillary tests and services provided by non-clinically trained personnel.9 The HCC grouper software first assigns claims to one of 543 DxGroups based on the person's ICD-9-CM diagnosis code, and in a few cases, on the person's age. Each ICD-9-CM code maps to one and only one of 543 DxGroups, although one person can be assigned to several DxGroups. The 543 DxGroups are then organized into 118 Condition Categories (CCs) on the basis of clinical similarity and resource use. Each person may have multiple CCs depending on the variety of his or her diagnoses. For example, there are 8 Neoplasm CCs, ranging from the most serious and costly, Neoplasm 1 (metastatic cancers), to the least, Neoplasm 8 (benign neoplasms). Other CC clusters include, among others, infections, diabetes, heart, and mental conditions. CCs are then organized hierarchically in terms of costliness. A CC designation for a person after this hierarchical pruning process has been applied is called an HCC. The process of creating the HCCs is summarized in Figure 7-2.
The HCCs reflect the clinical relationship between specific diseases as well as expected resource use. Hierarchies are imposed so that credit is given (in terms of predicted expenses) for only the most costly of clinically related conditions. For example, within the cancer hierarchy, each person is assigned only to the single highest cost category that applies. The cost category that remains after this hierarchical pruning process is called an HCC. The set of HCCs for a person forms the basis for predicting his or her resource use.
|FIGURE 7-2. Creation of HCCs|
|SOURCE: DxCG, Inc (1999).|
In contrast to the ACG software, the DCG software also provides weights for each HCC. The weights indicate the relative expected spending (or costliness) of patients across the HCCs. These weights are derived by calculating the expenditures for each group compared to the average for the developmental sample. The developmental sample used to generate the weights included privately insured, Medicare, and Medicaid data. Weights were calculated separately for each of these three populations. The privately insured population weights were developed using 1992 and 1993 data on 27 "clients" and about 1.7 million covered lives obtained from Mercer, Inc. (Ash et al., 1997). Clients were mainly large employers. Weights for the Medicare and Medicaid populations were derived from medical claims of the respective programs. In this study, we compare the predictive ability of the HCC weights included in the DCG software with the weights that we calculated based on data from each employer in our sample.
2. Previous Research on Predictive Power
Previous studies have used varied techniques to compare health risk-adjustment to simple age-and-sex adjustment. The studies conclude that health risk-adjustment significantly improves predictive ability (Dunn et al., 1996; Wiener et al., 1998). Few papers have examined the predictive ability of health risk adjustors for specific chronic conditions. Wiener and colleagues (1996) examined the predictive ratios of two types of models based on ADGs for the Medicare population in comparison to the Average Annual Per Capita Cost model. The ADG models performed better for 15 of 17 chronic conditions. For 9 of the 15 conditions, the predicted expenditures were within 10 percent of actual expenditures.
Ash and colleagues (1997) examined the predictive ratios of DCGs models for 26 different chronic conditions using data from a large private employer. They found that the predictive ratios for 13 disorders were within 10 percent of the actual costs. Seven chronic disorders had predicted costs that were more than 25 percent lower than the actual costs: breast cancer, hip fracture, rheumatoid arthritis, HIV/AIDS, cystic fibrosis, sexually transmitted diseases, and lymphoma.