Visit these other HIMSS sites:
Skip Top Navigation

Professional Development

JHIM: Journal of Healthcare Information Management

Health-Risk-Assessment Tools Used to Predict Costs in Defined Populations
Fern FitzHenry, PhD, RN; Edward K. Shultz, MD

Download PDF version
Health-Risk-Assessment Tools Used to Predict Costs in Defined Populations

Fern FitzHenry, PhD, RN; Edward K. Shultz, MD
 

ABSTRACT

With the Balanced Budget Act of 1997 mandating that the Health Care Financing Administration (HCFA) implement risk-adjusted payment mechanisms for Medicare managed care plans (Medicare+Choice) by January 2000, risk-adjustment tools will play an important role in future capitated reimbursement. This is because there is growing evidence that healthier-than-average beneficiaries select Medicare+Choice. The risk adjustment that HCFA has adopted is initially based on primary inpatient diagnosis from hospitalizations in the previous year. Other payers are likely to adopt similar payment mechanisms. This article reviews nineteen risk-adjustment research papers, including the tool adopted for Medicare+Choice, some of which are likely to form the basis for subsequent HCFA risk-adjustment methods. In general, claims-based models are more powerful in predicting total costs than survey-based or demographics-based models. Survey-based models, although expensive and not as powerful claims-based models, can be used when claims data are unavailable. One of the most popular survey-based tools, SF-36, is likely to become increasingly important because HCFA will be using it to measure quality outcomes from Medicare+Choice plans and will make the results public. All of the models reviewed have limitations, but can be expected to be building blocks for future risk-based capitated reimbursement.
KEYWORDS
  • Risk adjustment
  • Risk assessment
  • Capitation
  • Reimbursement
  • Total costs
  • Selection bias


Introduction

Healthcare information management professionals will increasingly be asked by government and other payers to collect and analyze variables for predicting illness in order to risk-adjust capitated payment contracts. This is because, although managed care risk-based contracts have been effective in controlling costs, current tools for risk adjustment frequently result in inadequate compensation for the treatment of chronically ill patients, threatening the financial survival of providers with a high proportion of chronically ill patients. In addition, inadequate compensation encourages both "skimming" or "cherry picking" (selectively marketing healthcare plans to the healthiest potential health-plan enrollees) and "dumping" (excluding enrollees who have or are likely to have chronic problems) in order to improve profits.

Furthermore, if capitated reimbursement for treatment of persons with cancer or other high-cost chronic diseases is the same as for treatment of persons without them, then providers of good, cost-effective treatments for these conditions will disappear from the market.1,2 Managed care providers will not want to be the area's best cancer center if it attracts poorly compensated high- cost enrollees. Without adequate compensation, cost-effective disease-management programs and chronic-disease centers may vanish.

Many of the attributes that affect costs are recorded in clinical and administrative databases managed by health information management professionals. Providers will look to them to analyze data to measure the effects of variables such as chronic disease. This article provides a review of nineteen published studies that identify and examine attributes that are likely to affect costs, providing methodologies for applying attributes or variables of covered patients in order to assess risk and adjust payments based on level of risk.


Risk Adjustment and Regulation

The Balanced Budget Act of 1997 mandates that the Health Care Financing Administration (HCFA) implement a risk-adjusted payment mechanism for Medicare+Choice plans by January 1, 2000.3 In part, this is because the consensus is that the health-plan market is skimming and dumping to increase profits, leaving less healthy beneficiaries in fee-for-service Medicare.3,4,5 Medicare plans have been reimbursed on Adjusted Average Per Capita Cost (AAPCC), which are set by county and incorporate age, gender, Medicaid eligibility status, and institutional status (e.g., beneficiaries in long-term care facilities) adjustments.6

Reimbursement based on demographics has fostered the selective enrollment of beneficiaries. For example, to attract healthy low-cost seniors, a plan can focus recruiting efforts on events (such as dances or golf outings) likely to attract physically active seniors. A plan that is successful in such selective recruitment will be overpaid and rewarded for skimming because capitation rates are set on average costs that cover a population that includes people with chronic disease. This leaves a disproportionate number of less-healthy seniors in fee-for-service Medicare. HCFA estimates that reimbursement to health plans has been 9-18 percent higher than plan payments would be if risk were adjusted by a finer risk-adjustment tool.3 The fact that nearly one hundred plans have decided to withdraw from the program may be a response to the approach of implementation of a revised risk-adjustment model for Medicare+Choice.7

Based on a model developed by federally sponsored research started in the 1980s, between January 2000 and 2004 HCFA will gradually phase-in risk adjustments based on a fifteen-category Principal Inpatient Diagnostic Cost Group (PIP-DCG) plus adjustments for age, sex, original reason for Medicare eligibility (disability), and eligibility for Medicaid.3,8,9,10 Prior legislation restricted HCFA to a model requiring only inpatient data, so for the time being, the diagnostic groupings will be based on inpatient principal diagnosis. However, health plans have been mandated to submit hospital outpatient, skilled nursing facility, home health, and physician data on or after October 1999 to permit the use of additional data for more sophisticated risk adjusters in the future.3,10 The models under consideration, including both hierarchical coexisting conditions (HCCs) and ambulatory care groups (ACGs), incorporate outpatient-encounter data into risk adjustment.3,10

The federal government is not the only government agency mandating risk-adjusted capitation. State employee-health programs in some states and all state-capitated Medicaid-waiver programs use some form of risk adjustment including stop-loss insurance, adjuster for low enrollment size, service carve outs (such as for births), 13-category DCGs, Adjusted Clinical Groups AdCGs with Resource Adjusted Categories (RACs), and rare and expensive disease carve outs.3,10,11


Risk Assessment and Risk-Adjustment Tools

Risk assessment is the process of estimating the cost coefficient associated with a measured variable to predict risk. Risk adjustment is the process of applying the measured variable to severity-adjust reimbursement based on expected illness for future cost periods. The task of risk assessment is to define a measure that can be collected from a group that is predictive of the cost of healthcare for that group in the future.

There are several common tasks that must be completed to assess risk for a group. The first is collecting a measure (such as age and sex) of the population group of interest. The second task is analyzing the relationship of the measure to historical cost in the group, usually resulting in a linear coefficient of the risk-assessment measure(s) and costs. Finally, this coefficient is used in a risk-adjustment model to predict costs for a future period. The value of the risk-adjustment model is in how well predicted costs (projected on the basis of the risk-assessment measures) match actual costs. A common statistic used to express the value of risk-adjustment tools is percent of variance explained, or R-squared (R2).

Risk-adjustment tools are not limited to those applied to predicting total costs in populations. Risk-adjustment and case mix tools, such as All Patient Related Diagnosis Related Groups (APR-DRGs), are also used to predict costs and mortality of inpatients. This discussion, however, is limited to those being used to predict the total annual costs for whole populations and for individuals.


Explanatory Potential of Risk Adjusters

If the decision to join a managed care plan were random, then no risk adjustment would be needed. Mounting evidence suggests this is not the case. No current risk-assessment measure predicts costs perfectly.

The value of a health-risk adjuster is in its ability to account for variance in healthcare expenditures as expressed in the percent of variance the measure explains. If a perfect risk-assessment measure existed, we would be able predict accidents, burst appendices, breast cancers, and the progression of disease. Whether an individual will experience these types of healthcare conditions is not predicted well using currently known variables. Because we know no measure will be 100 percent accurate, what is a reasonable portion of total healthcare costs that is predictable?

From a health economist's perspective, health-risk adjusters are divided into three categories.4 First are fixed effects, which alter the cost of health for an individual indefinitely, as would a chronic disease such as diabetes. The consensus opinion is that fixed factors can explain approximately 15-20 percent of the variance in spending. Second are time-varying effects, which alter the cost of health for an individual for a time but not indefinitely, such as stopping smoking or other personal health practices, similar to those described by Edington and others.12 These factors explain another 3-5 percent of the variance in spending. Third are random effects, such as being hit by a car, which are by definition unpredictable. Conservatively, then, the upper boundary for variation in actual spending that can be explained by risk adjusters is about 20 percent of the variance.4,13 The remaining 80 percent of costs are considered to be random or unpredictable by currently collected prediction information. These costs are accounted for in capitated rates by an average base rate applied to all enrollees regardless of risk-assessment measures. If we accept that current science and technology limit the maximum explainable risk to 20 percent, then a tool that explains 5-15 percent of the total costs may be explaining 25-75 percent of the predictable variance.


High Level Determinants of Cost

Historically, the following types of information are related to risk (potential adverse outcomes) in healthcare:14 age, sex, race and ethnicity; acute clinical stabilit; principal diagnoses (case mix); severity of principal diagnosis; extent and severity of co-morbidities; physical functional status; psychological, cognitive, and psychosocial functioning; cultural and socioeconomic attributes and behaviors; health status and quality of life; and patient attitudes and preferences for outcomes. Rate setting adds some additional variables,15 including industry (of commercial populations), time period trend, benefit level, geographic service area, and medical management (as related to affecting use and average charge of services). In the absence of data on use, demographic characteristics (especially sex and age), eligibility or financial status (Medicaid), and geographic adjusters (for Medicare risk) have been used to predict costs.16 These demographic characteristics explain a relatively small amount of the variation in annual per-capita expense.9,17,18 Where use data are available, then prior use, functional status, prescriptions, procedures, and morbidities improve the ability to predict variations. Several risk-adjustment models the include use data are reported below.


Methods

Selection of Studies. Studies were identified through a literature search in Online Computer Library Center, PubMed, Ovid, and ProQuest, using the terms "risk assessment" or "risk adjustment plus costs." Additional references were found in the bibliographies of the identified articles. Cost was explicitly included as a search term to find models that have been used to associate dollars (versus mortality or morbidity) with the risk measure. In addition to the literature search, models being considered by the HCFA were also included. Although this review of risk-assessment tools is not comprehensive, it is representative.

Statistics. R2 is the primary statistic used to gauge the strength of models for risk adjustment. Based on ordinary least squares regression, it is the proportion of the variability in actual costs that is explained by the model. Another statistic that is reported in this review as a measure of the strength of the risk-adjustment model is the cost ratio. The cost ratio uses costs from a referent or control group as the denominator and the costs of the at-risk group as the numerator. Values greater than 1 indicate the at-risk group has a greater probability of higher total costs. Although this statistic is not directly comparable with R2, it offers a measure of the power of the risk-assessment model for predicting costs. Most of the studies also reported some measure of error (predictive ratio, the ratio of actual to predicted expenses for enrollees such as those with the lowest quintile expenditures and highest quintile expenditures).3,19 The level of predictive error varies with the population of interest and is a key issue in selecting a risk-adjustment tool. Although these measures are not included in the review, they are available in the references.

Prospective versus Concurrent Risk-Adjustment Statistics.For some studies, statistics were reported for both concurrent and future periods. Proposals for concurrent risk-adjustment models, in part, would pay providers retrospectively based on the presence of certain high-cost conditions. These models approach cost-based reimbursement and have greater explanatory power than the same risk-adjustment model applied to a future period17,20 (see Table 3). Since, by policy, the intention of capitated risk-based methodologies is to provide reimbursement incentives to reduce costs by managing care, adjusting payment based on treatment that is in progress, for example, concurrent risk adjustment, would muddle incentives to reduce costs.3 They are provided when available, but are unlikely to be adopted by capitated payers.

Cost Caps and Carve Outs. Setting a cost cap on liability per person or carving out liability for certain conditions can also reduce provider financial risk. In essence, the payer re-insures the provider against risk for these costs, so the provider is protected from some portion of financial risk. However, because the provider is protected from more of this risk, capitation payment rates are usually reduced, an alternative with little appeal for health plans.3 Although cost caps and carve outs are not personal attributes that can predict costs, when they are part of the capitation payment model, as the 1996 study by Weiner and others aptly demonstrates, they increase the explanatory power of the risk predictor statistic.17


Results

The risk-adjustment tools reviewed are divided into three tables based on the category of variables used to predict costs. The models in Table 1 use demographics-based attributes as predictors. The models in Table 2 use survey-based attributes. The models in Table 3 use claims-based attributes. Although some of the models combine tools using both survey-based and claims-based attributes and could legitimately fall into more than one of these categories, in the interest of conserving space, each model is listed only once. If a model has any survey-based variables, it is listed in Table 2.

Demographics-Based Risk Adjustment. The most prevalent basis for risk adjustment is demographic variables. Demographic variables, called AAPCC, have been the basis of Medicare risk-adjustment managed care since 1985.3 AAPCC considers age, gender, Medicaid-eligibility status, and institutional status adjustments. Demographic variables are widely available in enrollment and personnel or other administrative databases and have been favored by actuaries.

Table 1 provides a summary of the performance of demographic risk-adjustment models.8, 9, 10, 11, 17, 18, 20, 21, 22, 23, 24, 25 The highest explanatory power reported in the demographic models accounts for 5.8 percent of the variance in costs in a mixed Medicare and commercial population20 and the lowest, a seemingly implausible negative, -1.4 percent, in which the validation population had shifted demographically from the estimation data set.26 Overall, it appears that when the study population was primarily sixty-five years old or older, the variance explained clusters approximately 1 percent. In addition, supplementing the two most common demographic variables, age and sex, with additional demographic variables had little effect in the study populations. In the study by Fowles and others,20 the explanatory power was an outlier (5.8 percent), but perhaps this is due to a wider age range combined with a cost cap of $25,000. The two studies in which age and sex are supplemented with additional demographic models (marital status, employment longevity, salary, residence, education, and occupation) show little increase in explanatory power.23,24 In addition, applying demographic risk adjustment to concurrent costs versus future costs does little to increase R2. In the study by Weiner and others,17 the explained variance in one model is virtually unchanged. Similarly, in the studies by Clark and others,22 and Fowles and others20 the explained variance decreases.

Survey-Based Risk Adjustment. Self-reported survey measures for risk adjustment are appealing because unlike claims-based adjustments, they can be collected on every beneficiary at enrollment and do not require a database of prior claims to set rates on each enrollee. Although claims data include transitory and random costs, survey questions can focus on future predictable costs. Also, when a claims-based risk adjustment is limited to inpatient claims, survey risk adjustment may be better at predicting bad health risks.21

 However, surveys have disadvantages as well. They are expensive to administer and subject to greater error than claims-based adjustment tools. In addition, surveys, like claims, can provide incentives for "gaming" and "upcoding." Finally, surveys are subject to nonresponse bias. Nonresponders may have higher costs, more disabilities, and a greater likelihood of unemployment than responders do.20,21,26

Table 2 provides a summary of the performance of survey-based risk adjustment models.12,18,20,21,24,25,27,28,29 There are four classifications of risk-assessment measures that cross these studies: prior use; functional health, including the SF-36 and RAND-36, activities of daily living (ADLs), and instrumental activities of daily living (IADLs); known diseases; and body mass index (BMI) or obesity. The SF-36 and RAND-36 use the same thirty-six-item survey but slightly different scoring.30 This thirty-six-item survey includes scales for physical functioning, role-physical, bodily pain, general health, vitality, social functioning, role-emotional, and mental health.31 It may be the most widely used instrument with arguably the greatest reliability and validity for assessing functional status, appearing in over 950 articles in 100 journals.31

The table also includes intervention studies, which did not report R2, but did report an alternative statistic related to total costs.12,28,29 These were included because managing care by definition means intervening to decrease costs. Costs would presumably be reduced by weight-loss programs,28 health-risk-reduction programs,12 and geriatric evaluation and management.32 The highest explanatory power reported in the survey-based models was 13.39 percent in a Medicare beneficiary sample.25 The lowest explanatory power was 0.3 percent for the only model limited to consumer-oriented measures of healthcare providers (customer satisfaction, access, and integration) in a commercial HMO sample.24 This is consistent with hypothesis that the traditional patient measures of quality are not related to health risk or total costs. In the survey-based models, the study by Gruenberg and others,25 using measures of prior use, known disease, and functional health, was the only study with a population exclusively 65 years old or older.

In each of the four classifications of survey-based risk assessment measures, a measure was often modeled alone, with demographics, and with another classification of risk-adjustment measure. When the measure was used alone or with demographics, the explanatory power ranged as follows: (1) prior use, 3.9 percent24 to 4.3 percent;24 (2) functional health, 0.6 percent24 to 11.1 percent,20 but more consistently around 4 percent; (3) known disease, 0.9 percent24 to 11.1 percent;20 and (4) BMI, obese persons had costs 1.25 to 1.44 percent higher than the nonobese (R2 not available).28 When two or more of the four classifications of survey-based risk assessment measures were used in combination, explanatory power ranged from 1.8 percent24 to 13.38 percent,25 but more consistently around 4.6 percent, indicating that combination measures probably have some overlap in the costs they predict. In the model of Fowles and others,20 in which demographics alone explained approximately 5.8 percent of the variation, functional health status added approximately 5 percent more in explained costs, which was approximately the same as the variation explained by chronic disease (self-reported from survey) or Ambulatory Diagnostic Groups (from claims and encounter data). Lamers's study21 combined a claims-based risk assessment measure with a survey-based risk assessment measure. It used primary inpatient diagnosis (using a method closely approximating the one selected for reimbursement of Medicare+Choice plans in 2000) and increased the explanatory power of the survey-based measures by only 1.78 percent.

Claims-Based Risk Adjusters. Prior use as recorded in claim databases may be the best predictor of costs.10 Unfortunately, basing future reimbursement on past claims provides perverse incentives and could increase use. Diagnoses, especially chronic diagnoses, are probably the most valuable predictors of risk, but they do not capture use-propensity variables such as the number of outpatient visits. Also, the most prevalent source of claims-based diagnoses is inpatient claims. Many enrollees have not been inpatients, so their diagnosis histories will be unknown. Lamers found only 7 percent of a low-income Netherlands sample that included the elderly had had a hospitalization in the previous year and only 17 percent in the previous three years.21 Weiner and others found about 20 percent of Medicare beneficiaries were hospitalized in the previous year, but 85 percent had an outpatient encounter.17 Outpatient models will undoubtedly improve measures of enrollee diagnoses.

Overall, diagnosis models may be biased by the method or source of collection (provider systems), the absence of collection systems in ambulatory care sites, the failure to note multiple encounters on a single diagnosis, the inability to distinguish between well-controlled and progressive illness, individual differences in responses to illness or compliance to treatment, and the inability to capture from selected settings (nursing homes and HMOs).10,18 Also, projecting costs from diagnoses assumes that there is a consensus on the appropriate treatment and relative costs for a diagnosis.4 For example, Milliman and Robertson, an actuarial firm, projects costs for diagnoses assuming the medical management will follow optimal treatment guidelines.33 Increases in managed-care-type reimbursement itself can alter medical management for an area, resulting in significant decreases in the number of inpatient days per one thousand patients. Thus, projecting costs by diagnosis from the actual historical costs with claims datasets will be error prone, especially when the datasets are collected under fee-for-service reimbursement, but used for setting capitation rates. The degree of reduction or increase (as for preventive screening) would be expected to vary by diagnosis.

Table 3 summarizes the performance of the claims-based risk adjustment models.8,9,10,11,17,20,21,22,23,34,35,36 The highest explanatory power reported in the claims-based models was 15.12 percent in a Medicare beneficiary sample.34 The lowest was 1.2 percent for a grosser use model limited to the presence of an inpatient admission in the previous year.8 In general, the claims-based models that only included quantitative expense measures performed less well than models that also considered clinical measures included in claims, such as diagnoses, procedures, and prescriptions. Models based on claims data prior to the implementation of Diagnosis Related Groups (DRGs) in 1983 had lower explanatory power than models using claims data after 1983, when DRGs became more comprehensive as the result of being the basis of payment.19 The model with the highest explanatory power considered not only diagnoses, but also procedures, dates of service, and sites of service, actually creating episode diagnostic categories (EDSs).34 EDSs were then collapsed to risk-adjustment categories (RAdCs), such as healthy, moderate acute, single chronic, multiple chronic, three or more dominant chronics, metastatic malignancies, and catastrophic illnesses or conditions. In addition to its high explanatory power for costs, the model is clinically appealing.

There are two major subgroups for diagnoses-based risk adjustment from claims data.2,4 Diagnostic Cost Groups (DCGs) and Ambulatory Diagnostic Groups (ADGs).

Diagnostic Cost Groups. Because HCFA will be using DCGs to risk-adjust Medicare+Choice enrollees in 2000, in the form of principal inpatient (PIP) DCGs, they are likely to receive widespread general attention.37 DCGs were developed from Medicare claims datasets in the 1980s.3,8 The base model is predicated on principal inpatient diagnosis from the preceding year. However, several extensions to the model were also developed including secondary diagnoses from inpatient or ambulatory bills, multiple medical conditions (a framework the authors named Hierarchical Coexisting Conditions, HCCs); and life-sustaining medical procedures.9 Assigning PIP-DCGs requires only principal inpatient diagnosis. The more recent HCC variations require all inpatient and outpatient diagnoses and procedures.

Even if the HCC variations are excluded, DCGs have evolved adjusted categories since their creation, so in the purest sense the DCGs reviewed in Table 3 are similar, but not equal. Nevertheless, when DCGs are used with demographics, the explanatory power ranges from 3.8 percent in the original DCGs8 to 9.01 percent for the DCG variant that creates HCCs with measures for life-sustaining procedures and inpatient conditions with high incremental subsequent costs.9

ADG-MDC and ADG-Hosdom Models. Johns Hopkins University developed ADG-Major Diagnostic Category (MDC) and ADG-Hospital Dominant (Hosdom) models.17 Both models include ambulatory diagnosis codes as major predictors of risk. A subset of thirteen ADGs, the basic morbidity classification of ACGs were included from the original thirty-four groups. This excluded ADGs that were relatively poor predictors of risk, such as time-limited minor infections. The ADG-MDC model added another variable, MDC, to incorporate inpatient diagnoses as a predictor category. In contrast, ADG-Hosdom included inpatient costs by incorporating diagnoses that are usually but not always or necessarily treated in an inpatient setting. There were 843 hospital-dominant diagnoses based on the criteria that at least 50 percent of the sampled patients with that diagnosis were hospitalized.

Table 3 summarizes the explanatory power of ADGs and it variations. When ADGs were used with demographics, the explanatory power ranged from 5.5 percent for ADG-Hosdom in a Medicare population17 to 12.4 percent for ADGs in a commercial HMO that included patients 65 years old or older.20 The study of Clark and others22 achieved similar explanatory powers, pairing ADGs with a measure called Chronic Disease Score (CDS), based on prescription claims and physician ratings of associated disease severity. Risk Adjustment Categories (RAdC) in Goldfield and others34 and Weiner and others11 are independent, the former based on EDSs and the latter based on Ambulatory Case Groups (ACGs), an ambulatory case mix measure developed at Hopkins. Also included is one study using an alternative statistics to R2. It was a comparison of total costs for diabetics with total costs of matched controls.36 They found the diabetic group had costs 2.4 times greater than the controls. These results indicate the practice of placing diabetics into disease management programs has a basis in costs.

All the models predict the current year's expense better than the following year's, with ADG-MDC at 64.4 percent as the best predictor. Predicting the current year's expense versus next year's expense results in an average increase in predictive power of approximately 31 percent. In addition, models performed better when expenses were capped at fixed amounts ($100,000 and $50,000), as would be the case if reimbursement included reinsurance or carve outs for high-cost cases. When predicting next year's expenses with caps at $50,000, ADG-MDC accounted for 9 percent of the variance versus 6.3 percent with no cost cap.17


Conclusion

As healthcare information management professionals incorporate the development of databases suitable for risk adjustment into their strategic plans, they should be cognizant of the benefits and limitations of existing models.

In general, claims-based adjusters were more powerful than survey-based or demographics-based adjusters were. The best model explained 15.12 percent of total costs.34 Claims-based models use readily measurable variables and could offer providers a useful means of negotiating risk-adjusted capitation rates. Unfortunately, claims data may only be available for prior enrollees, not new enrollees (such as newly eligible Medicare beneficiaries). In developing models based on claims-based administrative data, however, it should be noted that, although data may be available on prior enrollees, error rates based on claims data have been from 2 to 20 percent,16 some of which is attributable to "coding creep." However, these types of error rates are likely to decline over time. Inaccuracies in claims were operationally redefined from errors to fraud in 1997. The Department of Justice and the Office of Inspector General have made prosecuting healthcare fraud second only to prosecuting violent crime,38 and estimated billing errors (including coding errors) declined 45 percent from $23 billion in 1996 to $12.6 billion in 1998, also generating approximately $1.2 billion in fines and restitution.39,40

Although survey-based models were not better predictors of risk than claims-based models, they can be useful if a claims history does not exist. Moreover, as the HCFA has announced it will measure and make public quality outcomes from Medicare+Choice plans using the SF-36,41 organizations may want to incorporate survey instruments such as the RAND/SF-36 into outpatient assessments in order to provide interim feedback to providers and to prepare clarifying data in advance of the planned public release of survey data by HCFA. This type information might be helpful in avoiding the negative public perceptions similar to those that were generated when mortality rates were released by HCFA without disclosing exculpatory information (such as differences in severity of illness or case mix). In addition, such surveys can be useful for internal planning and future marketing strategies.

Moreover, as the SF-36 survey has been used in over 950 research studies, it provides a useful benchmark for other research or outcomes studies. It should be noted when using survey-based models, however, that all of the studies that included survey-based variables, including the Medicare samples, are subject to nonresponse bias (response rates ranged from 44.4 to 87 percent), and that there is evidence that health status of nonresponders is different than responders.42

Demographic models, although the most common, had the lowest explanatory power of the studies reviewed. They appear to have greater explanatory power in populations that are younger than 65 years or mixed than in elderly or disabled populations. Concurrent models did not improve the predictive power of these models. Nevertheless, data for demographic models have the advantage of being universally available and inexpensive.

A common limitation of all of the models is that, with the exception of Medicare beneficiaries, the group most frequently studied, the generalizability of these studies is limited. The populations studied are different (disabled, commercial, working-age, low income, Medicaid, Dutch, American), and comparing the explanatory power of the risk-adjustment measures across populations is problematic. The explanatory power of the same risk-adjustment measure may change because of differences in the population, methods for collecting data, and standards of care.

It should also be noted that concurrent measurements of predictors tend to improve the nondemographic model's explanatory power. This is particularly evident with claims-based models, in which improvements of 30 percent or more are common, but also, to a lesser extent, in one survey-based model. However, because concurrent claims models closely approximate cost-based models, they are not likely to be adopted as a tool in negotiating risk contracts with payers. They would, in essence, return risk to the payer, failing to supply incentives to the provider for efficiency. Some of the models may also be useful for identifying candidates for interventions that might reduce total costs, such as weight reduction, safety, or geriatric evaluation and management programs. Unfortunately, interventions are not free. For the geriatric evaluation and management intervention, members would have had to stay with the plan at least 18 months and have 1.5 fewer inpatient days to recover the $1540 per person cost of the program.32

There is little doubt that more complex risk-adjustment models will appear in the future. The HCFA has already announced plans to move to even more comprehensive risk-adjustment models in the next three to four years.37 Understanding risk-adjustment tools will be important to all the players in managed care, because other payers are likely to follow HCFA's lead, as was the case with DRGs.


Acknowledgments

We acknowledge, with thanks, the editorial assistance of Susan FitzHenry.


References

  1. Rogal, D. L., and Gauthier, A. K. "Are Health-Based Payments a Feasible Tool for Addressing Risk Segmentation?" Inquiry, 1998, 35(2), 115-121.
  2. Dunn, D. L. "Applications of Health Risk Adjustment: What Can be Learned from Experience to Date?" Inquiry, 1998, 35(2), 132-147.
  3. Health Care Financing Administration. "Report to Congress: Proposed Method of Incorporating Health Status Risk Adjusters into Medicare+Choice Payments." Mar. 1, 1999, http://www.hcfa.gov/ord/rpt2cong.pdf.
  4. Newhouse, J. P., Buntin, M. B., and Chapman, J. D. "Risk adjustment and Medicare: Taking a Closer Look." Health Affairs, 1997, 16(5), 26-43.
  5. Buntin, M. B., and Newhouse, J. P. "Paying Medicare Managed Care Plans." Generations, 1998, 22(2), 37-40.
  6. Health Care Financing Administration. "Medicare Program: Request for Public Comments on Implementation of Risk Adjusted Payment for the Medicare+Choice Program." Federal Register, 1998, 63(173), 47506-47513.
  7. Fisher, M. J. "Bill Would Undo Medicare+Choice Rate Caps." National Underwriter 103(32), Aug. 9, 1999, p. 3-3.
  8. Ash, A. S., Porell, F., Gruenberg, L., Sawitz, E., and Beiser, A. "Adjusting Medicare Capitation Using Prior Hospitalization Data." Health Care Financing Review, 1989, 10(4), 17-29.
  9. Ellis, R. P., Pope, G. C., Iezzoni, L. I., Ayanian, J. Z., Bates, D. W., Burstin, H., and others. "Diagnosis-Based Risk Adjustment for Medicare Capitation Payments." Health Care Financing Review, 1996, 17(3), 101-128.
  10. Ellis, R. P., and Ash, A. S. "Refinements to the Diagnostic Cost Group (DCG) Model." Inquiry, 1995, 32, 418-429.
  11. Weiner, J. P., Tucker, A. M., Collins, A. M., Fakhraei, H., Lieberman, R., Abrams, C., and others. "The Development of a Risk-Adjusted Capitation Payment System: The Maryland Medicaid Model." Journal of Ambulatory Care Management, 1998, 21(4), 29-52.
  12. Edington, D. W., Yen, L. T., and Witting, P. "The Financial Impact of Changes in Personal Health Practices." Journal of Occupational and Environmental Medicine, 1997, 39(11), 1037-1046.
  13. Newhouse, J. P., Manning, W. G., Keeler, E. B., and Sloss, E. M. "Adjusting Capitation Rates Using Objective Health Measures and Prior Utilization." Health Care Financing Review, 1989, 10(3), 41-54.
  14. Iezzoni, L. I. "Dimensions of Risk." In L. I. Iezzon (ed.) Risk Adjustment for Measuring Healthcare Outcomes. Chicago: Health Administration Press, 1997, pp. 43-165.
  15. Cigich, S. M. "Rating and Underwriting." In P. R. Kongstvei (ed.), dt (ed.), Managed Health Care Handbook. Gaithersburg, Md.: Aspen, 1996, pp. 667-678.
  16. Trauner, J. B., and Thilgen, C. "Using Risk Assessment to Evaluate Adverse Selection Under Capitated Contracts." Journal of Ambulatory Care Management, 1998, 18(1), 1-11.
  17. Weiner, J. P., Dobson, A., Maxwell, S. L., Coleman, K., Starfield, B. H., and Anderson, G. F. "Risk-Adjusted Medicare Capitation Rates Using Ambulatory and Inpatient Diagnoses." Health Care Financing Review, 1996, 17(3), 77-99.
  18. Hornbrook, M. C., and Goodman, M. J. "Assessing Relative Health Plan Risk with the RAND-36 Health Survey." Inquiry, 1995, 32(1), 56-74.
  19. Greenwald, L. M., Esposito, A., Ingber, M. J., and Levy, J. M. "Risk Adjustment for the Medicare Program: Lessons Learned from Research and Demonstrations." Inquiry, 1998, 35(2), 193-209.
  20. Fowles, J. B., Weiner, J. P., Knutson, D., Fowler, E., Tucker, A. M., and Ireland, M. "Taking Health Status into Account When Setting Capitation Rates." Journal of the American Medical Association, 1996, 276(16), 1316-1321.
  21. Lamers, L. M. "Risk-Adjusted Capitation Based on the Diagnostic Cost Group Model: An Empirical Evaluation with Health Survey Information." Health Services Research, 1999, 33(6), 1727-1744.
  22. Clark, D. O., von Korff, M., Sanders, W. M., Baluch, W. M., and Simon, G. E. "A Chronic Disease Score with Empirically Derived Weights." Medical Care, 1995, 33(8), 783-795.
  23. Rosenkranz, S. L., and Luft, H. S. "Expenditure Models for Prospective Risk Adjustment: Choosing the Measure Appropriate for the Problem." Medical Care Research Review, 1997, 54(2), 123-143.
  24. Hornbrook, M. C., Goodman, M. J., Bennett, M. D., and Greenlick, M. R. "Assessing Health Plan Case Mix in Employed Populations: Self-Reported Health Status Models." In M. C. Hornbrook (ed.), Advances in Health Economics and Health Services Research. Greenwich, Conn.: JAI Press, 1991, pp. 233-272.
  25. Gruenberg, L., Kaganova, E., and Hornbrook, M. C. "Improving the AAPCC with Health-Status Measures from the MCBS." Health Care Financing Administration, 1996, 17(3), 59-75.
  26. Fillit, H. M., Hill, J., Picariello, G., and Warburton, S. "How the Principles of Geriatric Assessment are Shaping Managed Care." Geriatrics, 1998, 53(4), 76-89.
  27. Hornbrook, M. C., and Goodman, M. J. "Chronic Disease, Functional Health Status, and Demographics: A Multidimensional Approach to Risk Adjustment." Health Services Research, 1996, 31(3), 283-307.
  28. Quesenberry, C. P., Caan, B., and Jacobson, A. "Obesity, Health Services Use, and Health Care Costs Among Members of a Health Maintenance Organization." Archives of Internal Medicine, 1998, 158(5), 466-472.
  29. Pacala, J. T., Boult, C., Reed, R. L., and Aliberti, E. "Predictive Validity of the Pra Instrument among Older Recipients of Managed Care." Journal of the American Geriatrics Society, 1997, 45(5), 614-617.
  30. Ingber, M. J. "The Current State of Risk Adjustment Technology for Capitation." Journal of Ambulatory Care Management, 1998, 21(4), 1-28.
  31. The SF-36 Health Survey. Boston: The Health Institute, New England Medical Center. Available from http://www.sf-36.com.: 1998.
  32. Reuben, D. B., Boult, C., Boult, L., Morishita, L., Smith, S. L., and Kane, R. L. "Outpatient Geriatric Evaluation and Management." Journal of the American Geriatrics Society, 1998, 46(3), 296-302.
  33. Milliman and Robertson, I. Ambulatory, Primary, and Pharmaceutical Care. Seattle, Wash.: Milliman and Robertson, 1997.
  34. Goldfield, N., Averill, R., Eisenhandler, J., Hughes, J. S., Muldoon, J., Steinbeck, B., and others. "The Prospective Risk Adjustment System." Journal of Ambulatory Care Management, 1999, 22(2), 41-52.
  35. Anderson, G. F., Steinberg, E. P., Powe, N. R., Antebi, S., Whittle, J., Horn, S., and others. "Setting Payment Rates for Capitated Systems: A Comparison of Various Alternatives." Inquiry, 1990, 27, 225-233.
  36. Selby, J. V., Ray, G. T., Zhang, D., and Colby, C. J. "Excess Costs of Medical Care for Patients with Diabetes in a Managed Care Population." Diabetes Care, 1997, 20(9), 1396-1402.
  37. Health Care Financing Administration. "Announcement of Calendar Year (CY) 2000 Medicare+Choice Payment Rates." Mar. 1, 1999, http://www.hcfa.gov/stats/hmorates/ cover00/default.htm.
  38. Kleiner, S. G., Philp, E. A., and Yokoyama, J. "Health Care Fraud." American Criminal Law Review, 1999, 36(773), 773-807.
  39. U. S. Department of Health and Human Services. "A Comprehensive Strategy to Fight Health Care Waste, Fraud and Abuse." Feb. 10, 1999, http://www.hhs.gov/news/press/1999pres/ 990210.html. Accessed Oct. 4, 1999.
  40. Taylor, M. "When He Speaks, Healthcare Listens: HHS' Inspector General's Longtime Chief Counsel Elicits Praise-Along with Some Criticism." Modern Healthcare, 1999, 29(40), 42.
  41. "Reporting Requirements for Medicare Health Plans in 1998: Health Plan Employer Data Information Set (Hedis® 3.0/98)"; and "The Medicare Consumer Assessment of Health Plans Study (CAHPS)-Operational Policy Letter #59," Health Care Financing Administration, Dec. 11, 1997, http://www.hcfa.gov/medicare/opl059.htm. Accessed Oct. 8, 1999.
  42. Ingber, M. J. "The Current State of Risk Adjustment Technology for Capitation." Physician Profiling and Risk Adjustment. Gaithersburg, Md.: Aspen, 1999, pp. 483-522.


About the Authors

Fern FitzHenry is an internal process and systems consultant for Vanderbilt University Medical Center in Nashville, Tennessee.

Edward K. Shultz is associate professor of pathology, associate professor of biomedical informatics, and director of the information technology integration at Vanderbilt University Medical Center in Nashville, Tennessee.

JOURNAL OF HEALTHCARE INFORMATION MANAGEMENT®, vol. 14, no. 2, Summer 2000
© Healthcare Information and Management Systems Society and Jossey-Bass Inc., Publishers

HIMSS Resources: