Vous êtes sur la page 1sur 14

MEDICAL CARE Volume 41, Number 1, pp 5669 2003 Lippincott Williams & Wilkins, Inc.

The Best of Both Worlds?


Potential of Hybrid Prospective/Concurrent Risk Adjustment

R. ADAMS DUDLEY, MD,* CAROL A. MEDLIN, PHD, LISA B. HAMMANN, MIRIAM G. CISTERNAS, RICHARD BRAND, PHD, DEBORAH J. RENNIE, AND HAROLD S. LUFT, PHD

BACKGROUND. There remains considerable uncertainty about whether prospective or concurrent risk adjustment (RA) is preferable. Although concurrent models have better predictive power than prospective models, the large payments associated with concurrent RA create incentives for fraudulent coding. A hybrid strategyin which prospective payments were used for patients with low expected costs and concurrent payments were available upon the diagnosis of a small number of common, expensive conditionsmight improve predictive performance while requiring less auditing than fully concurrent RA. In addition, withincondition RA (using clinical data) for the selected conditions could further improve predictive power. OBJECTIVES. To assess how such a hybrid strategy might perform, focusing on a small number of chronic, expensive conditions that are verifiable (hence auditable). SUBJECTS AND MEASURES. All patients from seven health plans who had two complete years of utilization data were considered. RA models were estimated among patients younger than 65 (n 319,209) using the Hier-

archical Coexisting Conditions (HCC) model with and without stratification of the sample based on the presence of one or more of 100 verifiable, expensive, predictive conditions (VEP100). R2 and predictive ratios were calculated for each model studied. RESULTS. Patients with a VEP100 condition (9.3% of the population) accounted for 84.3% of the variation in cost. R2 was 0.08 using a prospective HCC model on the entire population, but increased to 0.26 for a hybrid using prospective HCCs on the 90.7% of the sample without a VEP100 condition and a simple concurrent model consisting of dummy variables for each of the VEP100 conditions. CONCLUSION. Combined with targeted auditing, a hybrid approach to RA could improve our ability to match payments to costs. However, because this would require additional, costly data collection, more research is needed to determine whether this benefit justifies the data collection and auditing burden. Key words: Risk adjustment; risk assessment; health care financing; chronic disease; auditing. (Med Care 2003;41:56 69)

The potential for differential risk selection among health plans is widely recognized, and

many payers have adopted or are considering risk adjustment.13 Nonetheless, controversy remains

From the *Department of Medicine, the Institute for Health Policy Studies, the Institute for Global Health, and the Department of Epidemiology and Biostatistics, University of California, San Francisco (UCSF), California. From Genentech, South San Francisco, California. From MCG Data Services, San Mateo, California.

Supported by the Robert Wood Johnson Foundation (Grant 29661). Address correspondence and reprint requests to: R. Adams Dudley, MD, Department of Medicine and Institute for Health Policy Studies, University of California, San Francisco (UCSF), 3333 California Street, Suite 265, San Francisco, CA 94118. E-mail: adudley@itsa.ucsf.edu Received October 11, 2001; initial review January, 2002; accepted July 9, 2002.

56

Vol. 41, No. 1

HYBRID PROSPECTIVE/CONCURRENT RISK ADJUSTMENT ease that are larger than prospective RA payments for the same disease. For example, concurrent models give greater weight to you have heart failure nowthan prospective models give to you had heart failure last year. Critics of concurrent RA argue these large payments invite over utilization and upcoding (reporting diagnoses more severe than the patients true condition). They argue that concurrent RA is essentially fee-for-service (FFS) with a time lag (especially concurrent models that include as predictors the performance of medical or surgical procedures5). Although any RA creates incentives to upcode, the larger payments associated with concurrent RA create stronger incentives. Thus, concurrent RA would require additional auditing (as Medicare currently audits physician bills in FFS). Even if it focused primarily on outlier plans reporting a high proportion of their population to have expensive diseases, auditing all their encounter data still would be costly. Some observers have suggested blending prospective RA with FFS.17 This has the advantage of more closely matching payments to costs, but does not fully fix incentives to over utilize and requires the collection of data for both RA and FFS payment. Furthermore, plans that believe they do a good job of keeping patients with chronic diseases in optimal health do not support payment systems that are tightly linked to utilization.18 Another alternative is carve-outs of specific high cost conditions for either condition-specific capitation or FFS.19 However, these approaches require some model to estimate the appropriate capitation rate or FFS budget. Studies have shown that a single rate for each condition usually creates the opportunity for risk selection within a condition, and current RA software is insufficient to eliminate the incentive to avoid high risk patients.14,19 We propose an alternative based on the observation that a substantial fraction of health care costs arise from the treatment of a relatively small number of common, expensive chronic conditions, such as cardiovascular disease, cancer, and human immunodeficiency virus (HIV) infection. Developing better predictions of cost for patients with these conditions could greatly improve the predictive power of RA models. Translating this observation into a modeling strategy, we propose a hybrid system of RA that adjusts prospectively for most conditions, but provides additional concurrent payments for patients with selected condi-

about how to do the risk assessment (RA) that is needed to carry out risk adjustment.1,4 8 Competing approaches vary in their choice of predictors of cost (diagnoses,1,5 8 pharmacy data,9,10 health status questionnaires11) and whether they use hierarchical logic incorporating severity of illness.1,4,5 Also, RA can be done prospectively or concurrently. In this paper we focus primarily on the issue of timing, and confine our analyses to diagnosis-based RA, because this is the approach favored by policy makers.2,12 We argue that both prospective and concurrent RA have inherent problems that cannot be overcome by incremental changes to existing models. A hybrid approach, however, can enhance the strengths and minimize the weaknesses of each. Moreover, if additional clinical detail is added to the concurrent component of this hybrid, further improvement in RA is possible, and the data collected could also be used for quality measurement and improvement. Prospective RA models have low predictive power (usually measured by R2, which ranges from 0.08 to 0.15). This creates incentives for health plans to encourage the disenrollment of high risk patients within high risk categories, because plans achieving such selection can garner significant profits.13,14 In addition, prospective RA compensates plans poorly for some utilization driven by patients (eg, obstetrical utilization). Changes in enrollment status are also difficult to address prospectively. None of the extant claims-based RA models handle deaths well.15 In building their models, the developers either dropped patients who died from the population5,8which understates true costs or annualize, based on monthly expenditures up to the time of death6,7which overstates costs. This decision is important because a high proportion of total expenditures occur at the end of life.16 Furthermore, in some populations especially Medicaid and Medicarepatients frequently churnon and off a health plans rolls, so plans may enroll patients who are expensive, but have no prior data to submit for RA. Concurrent RA handles enrollment changes better. Costs associated with patients who die can be included in concurrent model building without annualization, and for new enrollees there is no need to have data from the previous insurer. Concurrent RA models also have higher predictive power. This is achieved primarily by allowing payments for the concurrent occurrence of a dis-

57

DUDLEY ET AL tions. The concurrent payments would be fixed amounts based on the condition and its severity. In this sense, the concurrent payments are like Medicares DRG payments one knows prospectively how much will be paid for a given diagnosis, but does not try to predict when a patient will develop the problem. However, they differ from DRGs in that they are tied to a clinical event (the patients diagnosis), rather than a utilization decision (a hospital admission). We would address upcoding through targeted auditing. To minimize both the data collection costs and the audit burden, we would offer concurrent payments only for a small number of common conditions that are clinically verifiable (that is, for which a specific pathologic or laboratory result or constellation of symptoms must be present before a clinician would accept the diagnosis). Hence the data for verification are the same as those already collected for clinical purposes. This strategy may be able to better match payments to costs than prospective RA and, by limiting the concurrent payments to a short list of conditions, may allow clinicians and policy makers to focus attention, resources, and audits where they matter most. In addition, a new area for improving RA is opened; one in which the goal is to perform adjustment among patients with a particular condition. Rather than being limited to data relevant to a broad mix of patients, such as diagnoses or drugs used, within-condition RA could use any type of information eg, lab results measuring severityrelevant to patients with a given diagnosis. This is an important difference between our proposal and that of Carter et al,4 who suggest blending prospective and concurrent RA using diagnosis codes only, because other researchers have shown that diagnosis-based RA is insufficient to capture the range of withincondition variation in concurrent costs.14,19 Prior research has shown that within-condition models of utilization can be developed with appropriate clinical input,20 22 and the ability to stratify patients at risk for high utilization within a disease category is a critical element of many disease management programs.23 The current performance of RA models is limited primarily by the failure to handle extremely expensive patients well,5,24 which suggests that a focus on withincondition RA for the most expensive conditions could have a large impact on both the predictive power of models and market incentives.

MEDICAL CARE In this paper, we investigate the potential of a hybrid RA approach. We created a list of conditions that were verifiable and chronic and expected to contribute significantly to total population costs. We then used health plan data to identify 100 of these100 being simply a round number of the right magnitudeto use as the relatively small (compared with the number of diagnoses in the International Classification of Diseases-9, Clinical Modification [ICD-9-CM] system) number of verifiable, expensive, predictive conditions (the VEP100). We then assessed the likely impact on market incentives of the blend of prospective RA plus concurrent payments for the VEP100.

Materials and Methods


Data and Patients We used data collected by the Society of Actuaries (SOA). This included enrollees diagnoses (ICD-9-CM codes) and costs during 1991 and 1992 from seven insurers, with each contributing data from one or more commercial (nonMedicare) products. Costs, measured by charges from submitted claims, were summarized for each patient annually. From the initial sample of 3,103,253 patients, we set aside those who: were not enrolled for the entire 2 years study period (n 2,662,918 patients or 85.81% of the original sample), and excluded those who (1) had only invalid diagnosis codes or only charges without diagnoses (n 45,867, 1.48%), (2) had multiple, conflicting entries for annual cost (n 73,982, 2.38%), (3) were older than 64 (n 1234, 0.04%), and (4) had missing gender information, or had inconsistent gender or age information between 1991 and 1992 (n 63, 0.00%). Our sample for model testing included 319,209 patients. There were 1,937,808 patients who were enrolled 1992 but not in 1991. These patients were used to select candidate conditions for inclusion in the VEP100.

Identifying the VEP100 Creating a list of conditions for which to provide concurrent payments involved subjective and objective evaluation of all ICD-9-CM codes. As a

58

Vol. 41, No. 1

HYBRID PROSPECTIVE/CONCURRENT RISK ADJUSTMENT from population mean cost weight and R2 were calculated for all models. To calculate R2 for the hybrid models, the total error sum of squares for the combined populations (VEP100 and REMAIN) was calculated as the error sum of squares from the VEP100 regression plus the error sum of squares from the REMAIN regression. The corrected total sum of squares was calculated as the sum of squares adjusted for the mean of the overall population. We then calculated R2 as one minus the ratio of the error sum of squares to the corrected total sum of squares.

first step, a physician (RAD) reviewed every code and identified a list of candidate conditions based on his perception of their verifiability, relative expense, and chronicity. Verifiability was judged by whether the diagnosis is typically based on specific clinical criteria, eg, review of pathologic specimens for cancer, absence of blood proteins for hemophilia. Relative expense and chronicity were defined by likelihood of creating ongoing expenditures through the course of the current year and subsequent periods. Expensive, acute conditions that do not require long-term expenditures (eg, cholecystitis) were not included in the VEP100. Using the patients with clean data for 1992 only, concurrent mean charges for each of the candidate conditions were based on mean total annual charges for all patients with that VEP condition (patients with multiple VEP conditions were excluded). The top 100 conditions in terms of mean concurrent annual charges were included in the VEP100.

Results
Patients Table 1 shows the demographic, clinical, and plan characteristics for our population. Only 9.3% had a VEP100 condition in 1992, and 13.8% had a VEP100 condition in 1991, 1992, or 1991 and 1992.

Developing and Applying Risk Assessment Models We used relative cost weights for 1992 (an individuals actual charges divided by the population mean, without truncation or transformation) as the outcome variable for modeling. Although there are several different RA packages available, we did not intend to compare them and used only DxCG software (release 4.2, DxCG, Boston, MA), which groups patients into Hierarchical Coexisting Conditions (HCCs). We used linear regression to predict cost weights based on demographic variables and binary variables representing the presence of one or more of a group of diagnosis codes. We estimated a demographic model and prospective and concurrent versions of HCCs. We also estimated a model in which patients were first stratified into those with and without a VEP100 condition in 1992. A concurrent model was created for patients with a VEP100 condition in which the predictor variables were age, gender, and a separate dummy variable for each VEP100 condition. The HCC prospective model was then recalibrated to patients without VEP100 conditions (the remainder population, or REMAIN). Combining the predictions generated from concurrent RA on patients with a VEP100 condition in 1992 and prospective RA on REMAIN patients, we created a hybrid model. The sum of squared deviations

Distribution of Charges Mean annual charges for all patients in 1992 was $1334, with a relative cost weight (by definition) of 1.00. Mean relative cost weight was 5.23 for patients with a VEP100 condition (Table 2) and 0.57 for patients without a VEP100 condition. Among VEP100 patients, those with only one VEP100 condition had a mean relative cost weight of 3.58, whereas those with more than one had a mean of 12.33. Almost all the variation in cost from the population average is attributable to VEP100 patients. This group accounts for nearly half of total expenditures and 84.3% of the variation in cost (Table 2). The 1.7% of individuals with more than one VEP100 condition account for 21.6% of expenditures and 62.5% of the variation.

Predictive Power of Models Table 3 shows the R2 for several models. The demographic model has little power. Using HCCs improves prospective predictive power (R2 0.08), but still leaves almost all variation in cost unexplained. The concurrent HCC model dramatically improves R2 (to 0.37). When the sample is stratified into those with a VEP100 condition in 1992 and those without

59

DUDLEY ET AL TABLE 1. Study Population Characteristics


Characteristic Age on January 1, 1992 (mean SD) Gender Male Female Plan Type Indemnity HMO PPO Patient total cost in 1991 (mean SD) Patient total cost in 1992 (mean SD) VEP100 Subgroup Status Had a code for a VEP100 Diagnosis in 1991 Did not have a code for a VEP100 Diagnosis in 1991 Had a code for a VEP100 Diagnosis in 1992 Did not have a code for a VEP100 Diagnosis in 1992 Had a code for a VEP100 Diagnosis in 1991 or 1992 Did not have a code for a VEP100 Diagnosis in 1991 or 1992

MEDICAL CARE

30.4 16.7 N (%) 160,895 (50.4) 158,314 (49.6) N (%) 224,609 (70.4) 28,913 (9.1) 65,687 (20.6) $1,082 4,652 $1,334 5,923 N (%) 26,050 (8.2) 293,159 (91.8) 29,594 (9.3) 289,615 (90.7) 44,138 (13.8) 275,071 (86.2)

(REMAIN), the model using dummy variables for each of the VEP100 conditions achieves an R2 of 0.21 within the VEP100 population. Combining these concurrent estimates for the VEP100 population with prospective HCC predictions for REMAIN yields an R2 of 0.26 (Model H1).

and concurrent HCCs (rather than VEP100 dummies) to patients in the VEP100, generates an R2 of 0.36. Thus, using concurrent HCCs on the 9.3% of the population with a VEP100 diagnosis achieves almost the same predictive power as applying HCCs concurrently to the entire population (0.37).

Impact of Stratification on Hierarchical Coexisting Conditions Models Table 3 also shows that Model H2, a hybrid model that applies prospective HCCs to REMAIN

Predictive Ratios Table 4 shows the predictive ratios for the various models when calculated for clinical or

TABLE 2. Relative Cost Weights and Contribution to Variation in Cost, by VEP100 St aus in 1992
Contribution to Variation in Cost (Sum of Squared Deviations from Mean Population Cost) SS (%TSS) 987,174 (15.7) 5,305,647 (84.3) 6,292,821 (100.0) 1,374,638 (21.8) 3,931,009 (62.5) % of the total sum

Number of VEP100 Conditions

Mean Annual Relative Cost Weight

Sum Patients N (%) 289,615 (90.7) 29,594 (9.3) 319,209 (100.0) 24,008 (7.5) 5,586 (1.7)

Sum Relative Cost Weight Sum (% of Total) 164,434 (51.5) 154,775 (48.5) 319,209 (100.0) 85,907 (26.9) 68,868 (21.6)

0 0.57 1 or more 5.23 All enrollees 1.00 Among patients with a VEP100 condition: 1 condition 3.58 1 condition 12.33

Note: SS sum of squared deviations from the population mean for the subgroup; %TSS of squared deviations from the population mean attributable to the subgroup.

60

Vol. 41, No. 1

HYBRID PROSPECTIVE/CONCURRENT RISK ADJUSTMENT TABLE 3. R2 for Various Risk Assessment Models

Model A1 A2 A3 V1 V2 R1 H1* H2

Predictors age and gender only HCCs HCCs VEP100 dummies HCCs HCCs HCCs VEP100 (mixed) HCCs HCCs V1. V2

Clinical Population All patients All patients All patients VEP100 patients VEP100 patients REMAIN patients All patients All patients

Percent of Patients 100% 100% 100% 9.3% 9.3% 90.7% 90.7% 9.3% 90.7% 9.3%

Timing Prospective Prospective Concurrent Concurrent Concurrent Prospective Hybrid Hybrid

R2 0.02 0.08 0.37 0.21 0.33 0.07 0.26 0.36

*Component models: R1 Component models: R1

cost subgroups. Among the clinical subgroups, the VEP100 hybrid Model H1 has excellent predictive ratios for those conditions that are in the VEP100, including breast cancer, depression that is not in remission or mild, and cerebral palsy. However, we have also shown for each of these conditions a related condition that is not included in the VEP100 (melanoma, depression in remission or mild, and mental retardation,

respectively), and the VEP100 model does not do as well for these as concurrent models that use all ICD-9-CM codes (Models A3 and H1). However, these conditions are either rare (melanoma, mental retardation) or difficult to verify (depression in remission or mild). When predictive ratios are calculated by quintile, the two hybrid models are similar, whereas concurrent HCCs applied to all models has better predictive ratios.

TABLE 4. Predictive Ratios for Alternative Risk Adjustment Models among Clinical or Cost Subgroups
Age-sex Prospective Concurrent HCC/VEP HCC Prospective HCC HCCs Hybrid Hybrid (Model A1) (Model A2) (Model A3) (Model H1) (Model H2) By Clinical Condition (Those that are among the VEP100 are in plain text, similar conditions that are not in the VEP100 are in italics): Breast Cancer Melanoma Cerebral Palsy Mental Retardation Depression (not in remission or mild) Depression (in remission or mild) By Quintile: Lowest quintile 1992 Expenditures* Second quintile 1992 Expenditures Third quintile 1992 Expenditures Fourth quintile 1992 Expenditures Highest quintile 1992 Expenditures

0.21 0.26 0.07 0.08 0.22 0.38 N/A 47.46 6.61 2.36 0.29

0.43 0.49 0.61 0.39 0.44 0.68 N/A 40.51 5.89 2.54 0.42

0.91 1.18 1.23 1.00 0.88 1.03 N/A 2.50 2.94 2.43 0.79

1.00 0.76 1.00 0.28 1.00 0.71 N/A 27.51 4.64 2.31 0.61

0.94 0.80 1.27 0.93 0.94 0.81 N/A 27.03 4.21 2.00 0.63

*The sum actual costs for this quintile is zero, so it is not possible to calculate this ratio.

61

DUDLEY ET AL

MEDICAL CARE

Discussion
RA is essential to establishing a health care market that functions well. Because the point of RA is to explain variation in cost and because clinicians often know which patients costs could be high, one might expect RA to be easy. Unfortunately, it is not. Although clinicians (and health plans) may know before expenditures are made which patients will be expensive, they usually do not know so far ahead of time that they can answer this question with much accuracy during the preceding annual enrollment period. Predicting costs a year in advance is difficult. In the absence of adequate risk adjustment, plans are likely to prefer to avoid cases that may be expensive.

Choosing the Timing of Risk Assessment A key question, then, is: Is it necessary to predict cost at annual enrollment (prospective RA) or can RA wait until a diagnosis is made (concurrent RA)? From the standpoint of preventing upcoding and over utilization, we should predict costs prospectively. If thousands of diagnoses can be associated with large concurrent payments, the auditing effort required to prevent gaming would be substantial. However, if one could limit auditing to a small number of important conditions that are clearly defined and can be verified relatively easily, then substantial payment for selected conditions is more plausible. Auditing costs could also be reduced by focusing on situations in which reported diagnoses are inconsistent with normal epidemiologic patterns. That is, a plan reporting a prevalence of VEP100 conditions that is consistent with its demographic mix could be audited less frequently, although a plan claiming a large proportion of such patients might be audited more often. Having concurrent payments for only a few conditions also will give health plans more budget stability than concurrent RA for all conditions would, as the prevalence of VEP100 conditions is less likely to vary annually than acute conditions such as trauma. The hybrid approach we propose would provide plans payments based on prospective HCC predictions for all patients at the beginning of the contract period, and then provide additional concurrent payments as patients with VEP100 conditions are reported by the plan. Thus,

the plan can estimate its expected revenues with reasonable accuracy by determining how many patients it expects to have with VEP100 conditions, but does not have to tie its revenue stream to rare or unpredictable events. Finally, there are advantages to not providing large concurrent payments for events that could be complications. Many infections such as pneumonia are complications of chronic diseases. By using prospective payments for chronic conditions and not including acute infections among the VEP100, our approach rewards plans that avoid complications. The VEP100 we describe is an initial effort to list conditions for this hybrid approach to RA. The list appears to have performed well in identifying patients who contribute to cost variations. However, it is not clear that these are the best 100 conditions or that 100 conditions is the right number. Further research will be needed to determine what the selection criteria should be for receipt of concurrent payments. These criteria should include the expected costs of auditing and the expected benefits in terms of reducing risk selection.

Improving the Predictive Power of Risk Adjustment The R2 values and predictive ratios reported in Tables 3 and 4 prompt the question: If concurrent HCCs do better than a hybrid model using prospective HCCs and concurrent VEP100 dummy variables, why bother developing the VEP100 concept? The answer lies in part in reducing auditing costs and focusing on priority conditions, but also in the potential of within-condition RA to improve predictive power. Prior research has shown that within-condition RA is possible. For example, Farley et al25 have generated models predicting costs among patients with end stage renal disease. Within-condition RA among VEP100 conditions would reduce predictive errors in the patient group that explains 84.3% of the total variation in cost in the population. This might significantly improve the ability of RA to match payments to actual costs. Within-condition RA seems feasible for specific VEP100 conditions. Many have numerical indices of severity. For example, the clinical status of patients with HIV is known to correlate with two quantitative markers: the CD4 count (more CD4 cells implies a more intact immune system) and

62

Vol. 41, No. 1

HYBRID PROSPECTIVE/CONCURRENT RISK ADJUSTMENT benefits of concurrent RA while generating fewer of the incentives to game by limiting concurrent payments to a small number of patients with predictably expensive, clinically defined, and verifiable conditions. We may not have chosen the best conditionsthere may be better choices among the high cost categories in HCCs or other models but this is a matter for further research. Our strategy of focusing on a small number of chronic conditions is consistent with the Institute of Medicines recent call for the identification of priority conditionsfor quality improvement, payment reform, and policy attention,26 and with the recent announcement by CMS that it is considering focusing RA in Medicare on a list of 25 to 100 conditions.12 The key issue, then, is: How much is it worth to improve predictive power? Our approach would add data collection and auditing costs. Nonetheless, it offers an intermediate solution between prospective and concurrent RA and may be superior to either, because it has greater predictive power than prospective and lower cost but similar predictive power with greater potential than concurrent RA using claims data. The final advantage of our approach is that the clinical data we collect could be used for other purposes, so that the costs incurred should be weighed against more benefits gained than just improved RA.

viral load. These markers also correlate with cost, because the prescription of anti-HIV drugs is based on changes in CD4 count and viral load and complications of HIV become more frequent as immune status worsens. Similar continuous measures of severity exist for heart, lung, and liver disease. Other conditions such as sickle cell disease and cancer have categorical/ordinal severity scales that may be amenable to within-condition RA. For instance, every cancer has a staging system, and treatment and cost often vary by stage. Thus, focusing RA on a short list of conditions like the VEP100 and pursuing within-condition RA may be a fruitful approach to developing the next generation of RA tools.

Applications and Limitations The failure of RA models to predict well withincondition has negative effects on several financial aspects of health care, such as carve-out rate setting14 and the ability to do specialty provider profiling, in addition to health plan premium adjustment. Our approach could help in each of these cases, because it would allow for more variation in carve-out rates (by using clinical detail to estimate within-condition severity) and would permit profiling of specialists utilization patterns that addresses case mix variation within diagnostic categories (as would be necessary, eg, in evaluating oncologists efficiency in managing lung cancer). There are limitations to our analysis. We required that patients have a full 2 years of eligibility, excluding all who died in the second year. (Concurrent models could include these people, but fully prospective ones cannot without annualization.) This could have created bias tending to overestimate (if patients who disenrolled had high utilization but did not have VEP100 conditions) or underestimate (especially if patients who died had high costs attributable to VEP100 conditions) the predictive potential of the VEP100 approach. Any attempt to implement the VEP100 approach should use models developed on data appropriate for the population to which they will be applied, using data that includes partial-year enrollees and is more recent than ours. Nonetheless, the limitations of our data are unlikely to have had much effect on our main point: that it may be possible to reap most of the

Conclusions
The VEP100 may not be exactly the right set of conditions, but there can certainly be a short list of conditions that are verifiable, predictably expensive, and account for a high percentage of total cost variation. The data collection and auditing burdens associated with concurrent RA for selected conditions is much less than it would be for concurrent RA for all conditions, so the predictive power is gained at lower cost. In addition, identifying these conditions suggests a direction for further RA research to generate additional predictive power and further reduce incentives to select risk. We cannot claim these benefits justify the additional costs. Further empirical efforts to develop within-condition models and determine data collection costs are needed, as is debate about the cost-benefit tradeoffs. There may be, however, an RA strategy that may be superior to those previously considered.

63

DUDLEY ET AL

MEDICAL CARE

References
1. Kronick R, Gilmer T, Dreyfus T, et al. Improving health-based payment for Medicaid beneficiaries: CDPS. Health Care Financ Rev 2000;21:2964. 2. Hash M. From the Health Care Financing Administration: Medicare Choice Plans. JAMA 2000;284:2988. 3. Knutson D. Case study: the Minneapolis Buyers Health Care Action Group. Inquiry 1998;35:171177. 4. Carter GM, Bell RM, Dubois RW, et al. A clinically detailed risk information system for cost. Health Care Financ Rev 2000;21:6591. 5. Ellis RP, Pope GC, Iezzoni L, et al. Diagnosisbased risk adjustment for Medicare capitation payments. Health Care Financ Rev 1996;17:101128. 6. Kronick R, Dreyfus T, Lee L, et al. Diagnostic risk adjustment for Medicaid: The disability payment system. Health Care Financ Rev 1996;17:733. 7. Meenan RT, OKeeffe-Rosetti C, Hornbrook MC, et al. The sensitivity and specificity of forecasting high-cost users of medical care. Med Care 1999;37:815823. 8. Weiner JP, Dobson A, Maxwell SL, et al. Risk-adjusted Medicare capitation rates using ambulatory and inpatient diagnoses. Health Care Financ Rev 1996;17:7799. 9. Roblin DW. Physician profiling using outpatient pharmacy data as a source for case mix measurement and risk adjustment. J Ambul Care Manage 1998;21:68 84. 10. Gilmer T, Kronick R, Fishman P, et al. The Medicaid Rx model: Pharmacy-based risk adjustment for public programs. Med Care 2001;39:1188 1202. 11. Parkerson GR, Jr., Harrell FE, Hammond WE, et al. Characteristics of adult primary care patients as predictors of future health services charges. Med Care 2001;39:1170 1181. 12. Centers for Medicare and Medicaid Services. Medicare Choice Risk Adjustment Public Meeting Meeting Materials. Available at: http://www.hcfa.gov/ medicare/rskadjmtrls.htm. March 15, 2002. 13. Newhouse JP, Manning WG, Keeler EB, et al. Adjusting capitation rates using objective health measures and prior utilization. Health Care Financ Rev 1989;10:4154.

14. Ettner SL, Frank RG, McGuire TG, et al. Risk adjustment alternatives in paying for behavioral health care under Medicaid. Health Serv Res 2001;36:793 811. 15. van Vliet RC, Lamers LM. The high costs of death: should health plans get higher payments when members die? Med Care 1998;36:14511460. 16. Gaumer GL, Stavins J. Medicare use in the last ninety days of life. Health Serv Res 1992;26:725742. 17. Newhouse JP. Patients at risk: Health reform and risk adjustment. Health Aff 1994;13:132146. 18. Porell FW, Gruenberg L. Discretionary hospital use and diagnostic risk adjustment of Medicare HMO capitation rates. Inquiry 2000;37:162172 19. Conviser R, Murray M, Lau D. Medicaid managed care reimbursement for HIV and its implications for access to care. Am J Manag Care 2000;6:990 999. 20. Ashton CM, Petersen NJ, Souchek J, et al. Geographic variations in utilization rates in Veterans Affairs hospitals and clinics. N Engl J Med 1999;340:32 39. 21. Berlowitz DR, Ash AS, Hickey EC, et al. Profiling outcomes of ambulatory care: Casemix affects perceived performance. Med Care 1998;36:928 933. 22. Thomas JW, Bates EW, Hofer T, et al. Interpreting risk-adjusted length of stay patterns for VA hospitals. Med Care 1998;36:1660 1675. 23. Rossiter LF, Whitehurst-Cook MY, Small RE, et al. The impact of disease management on outcomes and cost of care: a study of low-income asthma patients. Inquiry 2000;37:188 202. 24. Fowles JB, Weiner JP, Knutson D, et al. Taking health status into account when setting capitation rates: A comparison of risk-adjustment methods. JAMA 1996;276:1316 1321. 25. Farley DO, Carter GM, Kallich JD, et al. Modified capitation and treatment incentives for end stage renal disease. Health Care Financ Rev 1996;17:129 142. 26. Institute of Medicine. Crossing the quality chasm: A new health system for the 21st century. Washington, DC: National Academy Press; 2001:335.

64

Vol. 41, No. 1 APPENDIX A. TABLE 1.


VEP # 1 2 3 4 5 6 7 8 Short Description Anoxic brain damage Stomach cancer AMI Leukemias Peritonitis Chronic benign panc. dz Sickle cell anemia Lung/pleural cancer

HYBRID PROSPECTIVE/CONCURRENT RISK ADJUSTMENT VEP100 Conditions and Associated ICD-9-CM Codes
ICD-9-CM Codes 348.1 230.2, 151, 151.0, 151.1, 151.2, 151.3, 151.4, 151.5, 151.6, 151.7, 151.8, 151.9 410.xx, 411.0 204.xx, 205.xx, 206.xx, 207.xx, 208.xx 567, 567.0, 567.1, 567.2, 567.3, 567.4, 567.5, 567.6, 567.7, 567.8, 567.9 577.1, 577.2, 577.3, 577.4, 577.5, 577.6, 577.7, 577.8, 577.9, 251.4, 251.5, 251.6, 251.7, 251.8, 251.9 282.5, 282.6 195.1, 231.1, 231.2, 235.7, 235.8, 162, 162.0, 162.1, 162.2, 162.3, 162.4, 162.5, 162.6, 162.7, 162.8, 162.9, 163, 163.0, 163.1, 163.2, 163.3, 163.4, 163.5, 163.6, 163.7, 163.8, 163.9 806.xx, 851.xx, 852.xx, 853.xx 769, 770.7 (764.xx, 765.xx) and age (0, 1) 331.xx, 290.xx, 294, 294.1, 294.2, 294.3, 294.4, 294.5, 294.6, 294.7, 294.8, 294.9 576, 576.0, 576.1, 576.2, 576.3, 576.4, 576.5, 576.6, 576.7, 576.8, 576.9 730.xx 237.0, 237.1, 237.3, 237.5, 237.6, 237.7, 237.8, 237.9, 191, 191.0, 191.1, 191.2, 191.3, 191.4, 191.5, 191.6, 191.7, 191.8, 191.9, 192, 192.0, 192.1, 192.2, 192.3, 192.4, 192.5, 192.6, 192.7, 192.8, 192.9 585, 586, 588, 588.0, 588.1, 588.2, 588.3, 588.4, 588.5, 588.6, 588.7, 588.8, 588.9 164, 164.0, 164.1, 164.2, 164.3, 164.4, 164.5, 164.6, 164.7, 164.8, 164.9 236.2, 183, 183.0, 183.1, 183.2, 183.3, 183.4, 183.5, 183.6, 183.7, 183.8, 183.9 733.4x 157, 157.0, 157.1, 157.2, 157.3, 157.4, 157.5, 157.6, 157.7, 157.8, 157.9 233.4, 236.5, 185 159.0, 159.8, 159.9, 195.2, 230.3, 230.4, 230.5, 230.6, 230.7, 230.9, 235.2, 235.5, 239.0, 152, 152.0, 152.1, 152.2, 152.3, 152.4, 152.5, 152.6, 152.7, 152.8, 152.9, 153, 153.0, 153.1, 153.2, 153.3, 153.4, 153.5, 153.6, 153.7, 153.8, 153.9, 154, 154.0, 154.1, 154.2, 154.3, 154.4, 154.5, 154.6, 154.7, 154.8, 154.9 277.6 230.1, 150, 150.0, 150.1, 150.2, 150.3, 150.4, 150.5, 150.6, 150.7, 150.8, 150.9 (Continues)

9 10 11 12 13 14 15

CNS trauma Neonatal/chronic resp dstrss Premature/low bwt babies Alzheimers/cer. deg/dementia Chronic benign bil trct dz Osteomyelitis Cancer of nervous system

16 17 18 19 20 21 22

Chronic renal failure Mediastinal cancer Cancer of ovaries/adj strctrs Aseptic necrosis of bone Pancreatic cancer Prostate cancer Cancer: sm/lg intestine/anus

23 24

A1Antitrypsin df/hered angio Esophageal cancer

65

DUDLEY ET AL APPENDIX A. TABLE 1


VEP # 25

MEDICAL CARE (Continued)


ICD-9-CM Codes 431, 438, 432, 432.0, 432.1, 432.2, 432.3, 432.4, 432.5, 432.6, 432.7, 432.8, 432.89, 433, 433.0, 433.1, 433.2, 433.3, 433.4, 433.5, 433.6, 433.7, 433.8, 433.9, 434, 434.0, 434.1, 434.2, 434.3, 434.4, 434.5, 434.6, 434.7, 434.8, 434.9, 435, 435.0, 435.1, 435.2, 435.3, 435.4, 435.5, 435.6, 435.7, 435.8, 435.9, 436.xx, 437.xx 042.xx, 043, 043.0, 043.1, 043.2, 043.3, 043.4, 043.5, 043.6, 043.7, 043.8, 043.9, 440, 044.1, 044.2, 044.3, 044.4, 044.5, 044.6, 044.7, 044.8, 044.9, 795.8, 079.53, V08 236.4, 186, 186.0, 186.1, 186.2, 186.3, 186.4, 186.5, 186.6, 186.7, 186.8, 186.9 038.xx, 482.8x, 482.0, 482.1, 790.7 619, 619.0, 619.1, 619.2, 619.3, 619.4, 619.5, 619.6, 619.7, 619.8, 619.9 617, 617.0, 617.1, 617.2, 617.3, 617.4, 617.5, 617.6, 617.7, 617.8, 617.9 230.8, 235.3, 155, 155.0, 155.1, 155.2, 155.3, 155.4, 155.5, 155.6, 155.7, 155.8, 155.9, 156.0, 156.1, 156.2, 156.3, 156.4, 156.5, 156.6, 156.7, 156.8, 156.9 200.xx, 201.xx, 202.xx, 238.4, 238.5, 238.6, 238.7 286, 286.0, 286.1, 286.2, 286.3, 286.4, 286.5, 286.6, 286.7, 286.8, 286.9 508.1, 515, 516.3 233.0, 238.3, 239.3, 174, 174.0, 174.1, 174.2, 174.3, 174.4, 174.5, 174.6, 174.7, 174.8, 174.9, 175, 175.0, 175.1, 175.2, 175.3, 175.4, 175.5, 175.6, 175.7, 175.8, 175.9 (411, 411.1, 411.8, 412, 429.2, 413, 413.0, 413.1, 413.2, 413.3, 413.4, 413.5, 413.6, 413.7, 413.8, 413.9, 414, 414.0, 414.1, 414.2, 414.3, 414.4, 414.5, 414.6, 414.7, 414.8, 414.9) and not (410.x or 411.0) 287.3 140, 140.0, 140.1, 140.2, 140.3, 140.4, 140.5, 140.6, 140.7, 140.8, 140.9, 141, 141.0, 141.1, 141.2, 141.3, 141.4, 141.5, 141.6, 141.7, 141.8, 141.9, 142, 142.0, 142.1, 142.2, 142.3, 142.4, 142.5, 142.6, 142.7, 142.8, 142.9, 143, 143.0, 143.1, 143.2, 143.3, 143.4, 143.5, 143.6, 143.7, 143.8, 143.9, 144, 144.0, 144.1, 144.2, 144.3, 144.4, 144.5, 144.6, 144.7, 144.8, 144.9, 145, 145.0, 145.1, 145.2, 145.3, 145.4, 145.5, 145.6, 145.7, 145.8, 145.9, 146, 146.0, 146.1, 146.2, 146.3, 146.4, 146.5, 146.6, 146.7, 146.8, 146.9, 147, 147.0, 147.1, 147.2, 147.3, 147.4, 147.5, 147.6, 147.7, 147.8, 147.9, 148, 148.0, 148.1, 148.2, 148.3, 148.4, 148.5, 148.6, 148.7, 148.8, 148.9, 149, 149.0, 149.1, 149.2, 149.3, 149.4, 149.5, 149.6, 149.7, 149.8, 149.9, 160, 160.0, 160.1, 160.2, 160.3, 160.4, 160.5, 160.6, 160.7, 160.8, 160.9, 161, 161.0, 161.1, 161.2, 161.3, 161.4, 161.5, 161.6, 161.7, 161.8, 161.9, 165, 165.0, 165.1, 165.2, 165.3, 165.4, 165.5, 165.6, 165.7, 165.8, 165.9, 195.0, 230.0, 231.0, 231.8, 231.9, 235.0, 235.1, 235.6, 235.9, 239.1 344.xx, 342, 342.0, 342.1, 342.2, 342.3, 342.4, 342.5, 342.6, 342.7, 342.8, 342.9 (Continues)

Short Description Cerebrovasc dz and comps

26

HIV

27 28 29 30 31

Testicular cancer Sepsis, prob. nosocomial pneumonia Female genital fistula Endometriosis Hepatobiliary cancer

32 33 34 35

Lymph/sarcomas, otr lymph tmrs Coagulation defects Pulmonary fibrosis Breast cancer

36

Coronary artery dz (not AMI)

37 38

Primary thrombocytopenia Head and neck cancer

39

Paralysis

66

Vol. 41, No. 1 APPENDIX A. TABLE 1


VEP # 40 41 42 43

HYBRID PROSPECTIVE/CONCURRENT RISK ADJUSTMENT (Continued)


ICD-9-CM Codes 330, 330.0, 330.1, 330.2, 330.3, 330.4, 330.5, 330.6, 330.7, 330.8, 330.9 233.7, 236.7, 239.4, 188, 188.0, 188.1, 188.2, 188.3, 188.4, 188.5, 188.6, 188.7, 188.8, 188.9 203.0x 298.0, 301.12, 296.0, 296.1, 296.2, 296.3, 296.4, 296.5, 296.6, 296.00, 296.02, 296.03, 296.04, 296.10, 296.12, 296.13, 296.14, 296.20, 296.22, 296.23, 296.24, 296.30, 296.32, 296.33, 296.34, 296.40, 296.42, 296.43, 296.44, 296.50, 296.52, 296.53, 296.54, 296.60, 296.62, 296.63, 296.64, 296.7, 296.8, 296.9 443.8x, 441.2, 441.4, 441.7, 441.9, 443.9, 440, 440.0, 440.1, 440.2, 440.3, 440.4, 440.5, 440.6, 440.7, 440.8, 440.9, 557, 557.0, 557.1, 557.2, 557.3, 557.4, 557.5, 557.6, 557.7, 557.8, 557.9 277.1 385.3x 614.1, 614.2, 614.4, 614.6, 614.7, 614.8, 614.9, 615.1 292.xx, 304.xx, 305.2x, 305.3x, 305.4x, 305.5x, 305.6x, 305.7x, 305.8x, 305.9x 996.8x, V42, V42.0, V42.1, V42.2, V42.3, V42.4, V42.5, V42.6, V42.7, V42.8, V42.9 596.5x, 788.2x, 596.4, 591, 593, 593.xx 334, 334.0, 334.1, 334.2, 334.3, 334.4, 334.5, 334.6, 334.7, 334.8, 334.9 253, 253.0, 253.1, 253.2, 253.3, 253.4, 253.5, 253.6, 253.7, 253.8, 253.9 284, 284.0, 284.1, 284.2, 284.3, 284.4, 284.5, 284.6, 284.7, 284.8, 284.9 283, 283.0, 283.1x, 283.2, 283.3, 283.4, 283.5, 283.6, 283.7, 283.8, 283.9 556, 555, 555.0, 555.1, 555.2, 555.3, 555.4, 555.5, 555.6, 555.7, 555.8, 555.9 335.xx 571.xx, 456.0, 456.1, 456.2, 572.2, 572.3, 572.4, 572.5, 572.6, 572.7, 572.8, 789.5, 452, 453.0 343, 343.0, 343.1, 343.2, 343.3, 343.4, 343.5, 343.6, 343.7, 343.8, 343.9 233.1, 180, 180.0, 180.1, 180.2, 180.3, 180.4, 180.5, 180.6, 180.7, 180.8, 180.9 if 250.4x or 250.5x or 250.6x or 250.7x or 250.8x or 362.0x 295.xx 720.xx, 721.xx, 722.xx 429.3, 425, 425.0, 425.1, 425.2, 425.3, 425.4, 425.5, 425.6, 425.7, 425.8, 425.9, 428, 428.0, 428.1, 428.2, 428.3, 428.4, 428.5, 428.6, 428.7, 428.8, 428.9 279.0x 303.xx, 305.0x, 291, 291.0, 291.1, 291.2, 291.3, 291.4, 291.5, 291.6, 291.7, 291.8, 291.9 272.7 (Continues)

Short Description Cerebral degen of childhood Bladder cancer Multiple myeloma Depression: not remission or mild

44

Chronic athero periph vasc dz

45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63

Disorders of porphyrin metab Cholesteatoma Chronic pelvic inflam. dz Drug abuse Organ transplantation Other chronic GU disease Spinocerebellar dz Pituitary disorders Aplastic anemias Acquired hemolytic anemias Inflam bowel dz Ant. horn cell dz/ALS Chronic liver dz Cerebral palsy Cervical cancer Diab Mel w/chronic compl Schizophrenia Spondylopath/spondylos/disc CHF, cardiomyopathies

64 65 66

Chronic humoral deficiencies Alcohol abuse Lipidoses

67

DUDLEY ET AL APPENDIX A. TABLE 1


VEP # 67 68 69 70 71 72 73 74 75 76 77 78 79 80

MEDICAL CARE (Continued)


ICD-9-CM Codes 236.9x, 239.5, 189, 189.0, 189.1, 189.2, 189.3, 189.4, 189.5, 189.6, 189.7, 189.8, 189.9 233.3, 236.3, 184, 184.0, 184.1, 184.2, 184.3, 184.4, 184.5, 184.6, 184.7, 184.8, 184.9 234.8, 237.2, 237.4, 239.7, 194, 194.0, 194.1, 194.2, 194.3, 194.4, 194.5, 194.6, 194.7, 194.8, 194.9 366.xx 530.0, 530.3, 530.5, 530.6 277.0x 340, 341, 341.0, 341.1, 341.2, 341.3, 341.4, 341.5, 341.6, 341.7, 341.8, 341.9 952.xx 045.xx, 138 if (250.1x or 250.2x or 250.3x or 250.9x) and 250.4x and NOT (250.5x and 250.6x, 250.7x and 250.8x) 179, 181, 233.2, 236.0, 236.1, 182, 182.0, 182.1, 182.2, 182.3, 182.4, 182.5, 182.6, 182.7, 182.8, 182.9 345.xx 273, 273.0, 273.1, 273.2, 273.3, 273.4, 273.5, 273.6, 273.7, 273.8, 273.9, 275.xx (581.xx, 582.xx, 583.xx, 590.0x, 587) and not (585, 586, 588, 588.0, 588.1, 588.2, 588.3, 588.4, 588.5, 588.6, 588.7, 588.8, 588.9) 579, 579.0, 579.1, 579.2, 579.3, 579.4, 579.5, 579.6, 579.7, 579.8, 579.9 740.xx, 741.xx, 742.xx, 743.xx, 744.xx, 745.xx, 746.xx, 747.xx, 748.xx, 749.xx, 750.xx, 751.xx, 752.xx, 753.xx, 754.xx, 755.xx, 756.xx, 757.xx, 758.xx, 759.xx 279.1x 336, 336.0, 336.1, 336.2, 336.3, 336.4, 336.5, 336.6, 336.7, 336.8, 336.9 710, 710.0, 710.1, 710.2, 710.3, 710.4, 710.5, 710.6, 710.7, 710.8, 710.9 234.0, 190, 190.0, 190.1, 190.2, 190.3, 190.4, 190.5, 190.6, 190.7, 190.8, 190.9 426.xx, 427.0x, 427.1x, 427.2x, 427.3x, 427.4x, 427.5x, 427x, 427.8x, 427.9x 310, 310.0, 310.1, 310.2, 310.3, 310.4, 310.5, 310.6, 310.7, 310.8, 310.9 090.4x, 093.xx, 094.xx, 095.xx, 090.5, 090.6, 090.7, 090.8, 090.9 707, 707.0, 707.1, 707.2, 707.3, 707.4, 707.5, 707.6, 707.7, 707.8, 707.9 447.6, 446.xx 333.9x, 333.0, 332, 332.0, 332.1, 332.2, 332.3, 332.4, 332.5, 332.6, 332.7, 332.8, 332.9 714.xx 234.9, 199, 199.0, 199.1, 199.2, 199.3, 199.4, 199.5, 199.6, 199.7, 199.8, 199.9 307.1x, 307.5x (Continues)

Short Description Cancer: kidney/upper urin trct Other female genital cancers Other endocrine malignancies Cataract Chronic benign esoph dz Cystic fibrosis Multiple sclerosis, etc. Spinal cord injury Polio Diab mel w/acute compl Uterine cancer Seizure disorders Disorders of plsma prot & min metab Chronic renal dz w/o chrn ren failure

81 82

Malabsorption syndrome Congenital anomalies

83 84 85 86 87 88 89 90 91 92 93 94 95

Chronic cellular deficiencies Myelopathies Diffuse conn tissue dz Cancer of the eye Cardiac rhythm disturbances Non-psych org mental disorder Advanced syphilis Chronic skin ulcer Chronic arteritis/related conds Parkinsons/degen of bsl gag/etc Inflam polyarthropathies Cancer: unspec/uncertain site Anorexia/eating disorders

68

Vol. 41, No. 1 APPENDIX A. TABLE 1


VEP # 96 97 98 99

HYBRID PROSPECTIVE/CONCURRENT RISK ADJUSTMENT (Continued)


ICD-9-CM Codes 135 530.1, 530.2, 531.4x, 531.5x, 531.6x, 531.7x, 532.4x, 532.5x, 532.6x, 532.7x, 533.4x, 533.5x, 533.6x, 533.7x 255, 255.0, 255.1, 255.2, 255.3, 255.4, 255.5, 255.6, 255.7, 255.8, 255.9 195.3, 195.4, 195.5, 195.6, 195.7, 195.8, 235.4, 238.0, 238.1, 239.2, 170, 170.0, 170.1, 170.2, 170.3, 170.4, 170.5, 170.6, 170.7, 170.8, 170.9, 171, 171.0, 171.1, 171.2, 171.3, 171.4, 171.5, 171.6, 171.7, 171.8, 171.9 494

Short Description Sarcoidosis Chronic peptic ulcer dz Adrenal disorders Cancer of bone or soft tissue

100 x

Bronchiectasis any value from 0 to 9.

69