Wednesday, December 26, 2007

Diagnosis-based risk adjustment for Medicare prescription drug plan payments

INTRODUCTION


The 2003 MMA created Medicare Part D, a voluntary prescription drug benefit program. The benefit is a government subsidized prescription drug benefit in Medicare and is administered by private sector plans. Such plans may be standalone prescription drug plans (PDPs) or Medicare Advantage prescription drug plans (MA-PDs). While there are numerous essential components determining how these plans are paid, this article focuses on the nouns of the prescription drug risk-adjustment model used to adjust payments to reflect the vigour status of plan enrollees. According to the MMA, payments are based on a standardized plan bid that represents the estimated cost for an enrollee next to average risk and a score of 1.0. payments for respectively enrollee are risk adjusted by multiplying the standardized bid by a person-level risk factor so that plan payments emulate the projected health of actual enrollees. Higher standardized bids result contained by higher per enrollee revenues, but also highly developed premiums in the competitive flea market. The process of developing the prescription drug risk-adjustment model, CMS prescription drug hierarachical condition categories (RxHCC) are also described surrounded by this article.


BACKGROUND


The basic Medicare prescription drug benefit structure somewhat covers the expenses of the majority of plan enrollees and has a catastrophic benefit for exceptionally high users. A Part D enrollee pays a premium, which be expected to be approximately $35 (1) a month. Enrollment is on a voluntary basis. There is a premium increase for those who enroll after their initial opportunity, as in attendance is in Medicare Part B. The structure of the standard benefit for 2006 is shown surrounded by Figure 1.


[FIGURE 1 OMITTED]


Enrollees are responsible for the first $250 in drug expenditures. The standard benefit bunch covers 75 percent of the next $2,000 surrounded by drug expenditures. Once total expenditures reach $2,250, the beneficiary is responsible for adjectives costs in what have become known as the "donut hole." The 100 percent coinsurance continues until total drug expenditures accomplish $5,100 ($1,500 plan liability plus $3,600 out-of pocket expenses). The catastrophic portion of the benefit covers 95 percent of any additional drug expenditures: 15 percent of the cost is the plan's responsibility; 80 percent is reinsurance compensated by Medicare. In the early years in that is also plan-Medicare risk sharing for the difference between Medicare payments and actual plan operational costs computed surrounded by a year-end reconciliation. The coverage thresholds are to be indexed for inflation in future years. PDPs and MA-PDs hold some flexibility in offering plans that differ from the standard benefit. In rider, formularies are set by the plans, subject to legislated requirements, and may vary across plans.


Payments to PDPs and MA-PDs are risk in synch, since payments are based on a standardized bid amount, which assumes an enrollee next to a risk factor of 1.0. Using a standardized bid to determine the beneficiary premiums insulates the beneficiary from the variation contained by health status of plan enrollees. Medicare pays the adjustment for risk. The starting point for the bid is the projected monthly revenue requirements to provide defined standard drug coverage for an enrollee near the plan's projected average risk factor. The standardized bid is computed by dividing monthly revenue requirements by the plan's projected average risk factor. Payment adjustments above the risk-adjusted rate are made for low-income and long-term institutionalized beneficiaries due to their greater expected utilization.


The risk factor is derived from the model presented in this article. The CMS-HCC model used for the MA program served as the foundation for our work here and is prospective. It uses diagnoses in a stub year to predict medical costs in the following year. The CMS-HCC model groups the approximately 15,000 International Classification of Diseases, Ninth Revision Clinical Modification (ICD-9-CM) codes into 178 disease groups (Centers for Disease Control and Prevention, 2006). The 70 disease groups that are most predictive of adjectives costs are included in the final 2005 payment model. Pope et al. (2004) discuss the primary criteria for grouping diseases together and for decide on which diseases comprise the final model.


There are several prescription drug risk-adjustment models that have be developed. Some are based on the prior use of drugs to predict adjectives medical costs or future prescription drug use. We could not use such a methodology to develop our model. In charge to implement the program, we needed to compute risk scores for adjectives Medicare beneficiaries. Since we lacked drug utilization information for most beneficiaries, we were incompetent to implement this type of model. Once the drug benefit is established, data on prior utilization will be available for use surrounded by calibration.


Gilmer et al. (2001) developed a model that predicts prospective Medicaid medical costs based on bottom year prescription drug utilization. Drug claims were analyzed, near national drug codes (NDCs) grouped together based on the disease they are typically used to treat. Thus, it is similar to other risk-adjustment models surrounded by that it uses diseases to predict future costs, but infers the diagnoses from prescription drug use, not ICD-9-CM codes.


Zhao et al. (2005) found that models using diagnoses and prior drug use predict adjectives prescription drug costs better than models using only diagnostic facts. Such research highlights the need to consider prior use within future model nouns. Inclusion of utilization measures among predictor variables must be done with circumspection in expenditure models, in contrast to analytical models, as perverse incentives to increase utilization or to favor a individual mode of treatment can be generated.


While prior drug use may predict adjectives drug use better than diagnostic data, added work was needed to determine whether diagnostic background sufficiently predict future drug use to produce the desired drug risk-adjustment model. Wrobel et al. (2003/2004) used the Medicare Current Beneficiary Survey (MCBS) to analyze the cleverness of the CMS-HCC model to predict prescription drug expenditures. Demographic variables only explain 5 percent of the ebb and flow in drug expenditures, while tallying diagnostic groups increases the explained variance to 10-24 percent. Adding lagged drug use increases the R2 to 55 percent. Overall, diagnoses are essential predictors of future drug use and the results of their study indicate the CMS-HCC model is an appropriate starting point for a model to predict drug expenditures.


DATA SOURCES AND MODEL OVERVIEW


Data Sources


Development of a risk-adjustment model for drug spending depends on have appropriate data from which to create diagnosis groups and cost estimates. As nearby were no Part D information available, CMS used drug expenditure data for Federal retirees beside Medicare in the Federal Employee Health Benefit plan run by Blue Cross [R] Blue Shield [R] (BCBS). The BCBS plan is national within scope, beside uniform benefits. The BCBS pharmacy benefit plan is an uncapped benefit with a coinsurance amount for retail purchases and two tiers of copayment for messages order purchases. Only those retirees age 65 be used from these data. For disabled beneficiaries underneath age 65, data on Medicare and Medicaid dually eligible beneficiaries from the Medicaid Statistical Information System (MSIS) be used. For each information set the development of the model used diagnoses from standard Medicare files and drug spending from respectively program's drug benefit. The BCBS plan spending year 2002 was used for calibration. For Medicaid, the up-to-the-minute available data associated to Medicare were for spending year 2000.


Next, we obtain information for these beneficiaries from the enrollee database (EDB). The EDB is the primary repository for Medicare current and historical enrollment and entitlement data. It was the source of demographic and Medicare Program information not available in the BCBS plan or Medicaid information. Critical data from the EDB includes Parts A and B coverage period, hospice coverage, and managed prudence coverage periods.


We used diagnostic information from the Medicare Provider Analysis and Review (MEDPAR), hospital outpatient, and physician claims from the platform years (2001 for the BCBS plan and 1999 for Medicaid). Diagnoses were standard from the following five source records: (1) principal hospital inpatient; (2) lower hospital inpatient; (3) a hospital outpatient; (4) physician; and the (5) clinically-trained non-physician (e.g., psychologist, podiatrist). The model does not distinguish among sources. These are the same notes sources for diagnoses used in the CMS-HCC model.


The BCBS plan notes provided to CMS contain annual prescription drug expenditures for each enrollee and annual copayments by enrollees. We converted the BCBS plan costs to total pharmacy costs for respectively beneficiary by adding the beneficiary's cost sharing amounts to the BCBS plan costs. The BCBS plan offered two different types of benefits in 2002: standard benefits and elemental. The standard pharmacy benefit included a 25 percent coinsurance on retail pharmacy purchases, while the mail decree benefit had a two-tiered copayment. The key benefit included a two-tiered copayment on retail purchases, and no mail directive benefit. Retail pharmacy costs for enrollees in the standard BCBS plan be imputed using the BCBS plan costs and the 25 percent coinsurance.


Medicaid be more difficult, however. The Medicaid Program is very complex, varying across States. To create a reliable notes file we removed individuals when diffident about the completeness of diagnostic or cost facts. We excluded individuals living in Arizona, Hawaii, and Tennessee due to high manage care access. We also removed managed diligence enrollees from other States, and individuals with other insurance coverage, since Medicaid is the payer of concluding resort. We also excluded individuals who did not have prescription drug coverage through their Medicaid Program. For example, some individuals eligible for Medicaid as qualified Medicare beneficiaries (QMBs), specified low-income Medicare beneficiaries (SLMBs), or qualify individuals (QIs) did not receive prescription drug coverage through Medicaid.


Additional modifications to the data be necessary to remove secure drug claims from the data because Part D specifically does not cover unshakable drugs. Only prescription drugs are included, but with Medicare Part B covered drugs removed. Drugs covered by Part B, such as immuno-suppressives, will verbs to be covered by Part B Medicare. Removal of the Part B drugs was straightforward within the Medicaid data as respectively claim has both an NDC and amount remunerated. Adjusting the BCBS plan data be more complex. We had individual total spending for each party, with no rewarded amount on the claims to be excluded. Using the Medicaid data we estimated the percentage money off in spending associated near removal of Part B drugs for beneficiaries with conditions associated near high use, such as cancer and transplants. We then reduced spending for similar beneficiaries in the BCBS plan files in matching proportion. Other non-covered drugs, benzodiazepines, and barbiturates, were intentionally disappeared in the database because their costs proxy for the costs of substitutes. This was deem preferable to removing the claims and costs altogether.


At the conclusion of the data compilation, for respectively beneficiary we had demographic, programmatic, and diagnostic information for the basis year along with prescription drug cost information for the transmittal year. Descriptive statistics for the BCBS plan and Medicaid samples are provided within Table 1. Given beneficiary cost sharing, a plan offering the standard benefit is liable for less than one-half total drug expenditures. The Medicaid example is younger on average than the BCBS plan sample because adjectives ages, including the disabled under age 65 can be dually eligible beneficiaries, while in attendance is no equivalent group in the BCBS plan background. Consequently, disease prevalence is different for the two samples.


We stratified respectively data set into two groups. The first group comprised those for whom we have sufficient information to include them in the risk-adjustment estimation model. For the purpose of calibrating a drug risk-adjustment model, we begin with the population of fee-for-service Medicare beneficiaries beside Medicare Parts A and B for the entire base calendar year. This allowed us to own a complete year of diagnostic information for these beneficiaries. We further required that individuals be enrolled contained by the BCBS plan or Medicaid Program for at least soon in the salary year. It is important to retain relations with smaller amount than full payment year eligibility to appropriation the potentially different drug use pattern of decedents. Weighting is applied to partial year enrollees.


The second group comprised those for whom we did not hold a year of complete diagnostic information, but for whom we had prescription drug costs contained by the following year. These beneficiaries could not be used for model estimation. Nevertheless, they represent one group of enrollees who must be given a score base on information other than diagnoses. A model for these different enrollees is also created.


The initial model developed (on the BCBS plan data) to predict spending, omitted two groups that received special treatment at the end of the process--those who would receive the low income subsidy (LIS) and the long-term institutionalized (LTI.


GROUPER


The model uses particular demographic characteristics and diagnoses to predict the following years expected costs for an individual. The ICD-9-CM diagnoses are clustered in groups homogeneous both clinically and in costs. Each included characteristic and condition present contributes to the total prediction for an individual through a formula that sums the incremental contributions. The groupings used to predict drug spending are variant of the groups used to predict Parts A and B spending.


We wanted to create a grouper that be similar to the grouper that was used to predict Parts A and B spending while person homogeneous for drug spending rather than non-pharmacy spending. We begin by estimating a prospective model regressing spending in the payment year on the dais year diagnosis grouping ([DXG.sup.2]) of diagnoses that are the basis of the CMS-HCC model. Results of this regression and some specific issues of the evaluation be: (1) whether there be DXGs that did not have implication for drug spending in the next year; (2) whether the grouping of DXGs into condition category used in the CMSHCC model be appropriate for a drug spending model; (3) whether the DXGs should be combined differently than in the CMS-HCC model; and (4) whether any CCs should not be included in the drug model. We re-estimated the model based on the received recommendation and had them reviewed by an interdisciplinary panel of clinicians. The clinicians reviewed the statistical results and assessed the groupings using one and the same criteria as previously mentioned. We re-estimated the model based on clinical input. Iterating this process next to the clinicians ultimately resulted in a grouper that changed few of the narrow DXG building blocks. However, the DXGs are assembled into larger condition disease category that often differ from the CMS-HCC groups. The relationship between diagnosis and costs is not like for Parts A and B spending as for drug spending.


In development of the model's grouper, drug spending in dollars be the dependent variable of a linear regression that estimated the incremental spending related to respectively of the explanatory variables in the model. It be easier for clinicians to evaluate a model that predicts the total cost of drugs needed for a condition than plan liability, which is the result of a complex formula. In May 2004, based on these preliminary results, CMS announced the 5,542 ICD-9-CM codes lower than consideration for inclusion in the drug risk- adjustment model.


The RxHCC diagnostic classification system groups the more than 15,000 ICD9-CM diagnosis codes into 197 condition categories, or RxCCs. As near the CMSHCC model, all ICD-9-CM codes are classified into disease groups despite the constrained number in the final model. RxCCs describe highest diseases and are broadly organized into body systems. As in the CMS-HCC model some of the disease groups are clustered in hierarchies. Clinical review found that drug regimens may get more intense, and more drugs may be added when a disease have a higher severity. In such a luggage, when the model has complex and lower severity categories, if the superior cost category of the related diseases is reported, coding of the lower cost category is ignored. Such is the overnight case with diabetes: diabetes beside complications overrides uncomplicated diabetes. If the drugs for diseases differ from one another, even if the diseases are related, the RxHCCs are not placed in duplicate hierarchy and remain chemical addition. Conditions not in one and the same hierarchy contribute independently to the total prediction. After the hierarchies are imposed, the RxCCs become RxHCCs. The category and hierarchies used in the model are presented within Tables 2 and 3.


Pooling BCBS Plan and Medicaid Data


While the grouper was formed by estimating a spending model using singular BCBS plan data, the final model be estimated using a pooled plan Medicaid data set. There be a number of problems within integrating the data sets: (1) the Medicaid group is low income and received drugs at out-of-pocket costs rather different from BCBS plan enrollees; (2) because of price differences, utilization would probably differ from that under the BCBS plan benefit, even for duplicate diseases; and (3) the cost data be from a different year and from many Medicaid Programs. In integrating the two information sets we converted the Medicaid data to spending pattern similar to that which would have occur, on average, under a BCBS plan benefit.


First, since the facts are for different years, inflation factors be applied to eliminate spending differences due to price inflation. The spending in both notes sets was multiplied by inflation factor calculated using the 2003 national health narrative prescription drug spending projections by CMS actuaries to project spending levels within 2006. We inflated to 2006 dollars because the cost-sharing ranges are defined in absolute dollar vocabulary for 2006; thus, spending had to be projected to level appropriate to 2006. Dollars from the year 2000 were multiplied by 2.039, while 2002 dollars be multiplied by 1.554.


Second, the model estimated with BCBS plan background for the aged, was applied to the dual eligible aged population to predict their spending as it would be lower than a BCBS plan benefit. This modeling incorporated the different demographic and disease profiles of the Medicaid population in the predictions. The actual spending in the Medicaid data be then compared to the predicted spending. The ratio of the predicted to the actual spending be used to convert the spending in the Medicaid files to levels compatible near BCBS plan. The conversion factor was analyzed across the age/sex groups appearing in both notes sets and, except for the sparse age group 95 or over was relatively stable. With the data sets merged it become possible to estimate a full model across all ages and include age-specific add-ons for some diseases. This example represents beneficiaries all of whom are presumed to own the BCBS plan benefit structure. The data within the two samples be weighted to make the facts representative of the Medicare population.


Computing Standard Benefit Plan Liability


The requirement of the risk-adjustment model was to predict the cost of drugs to the Part D plans, not the total spending that be modeled thus far. The decision to estimate a plan liability model base on the standard benefit was arrived at within consultation with industry actuaries after studying the difficulties, both hi-tech and operational, within modeling an unknown spectrum of possible benefit variations. Despite the discontinuous stencil of plan liability as spending varies, a linear model base on plan liability produces reasonable results. The plan liability model uses the grouper developed for the total spending model. The coefficients be estimated, however, on data altered to imitate plan liability.


Before applying the cost sharing to create plan liability, the spending data go through one additional adjustment. It is collectively observed that spending patterns are artificial by income and prices. The model described thus far incorporated the cost-sharing patterns of the plan benefit. The cost sharing in Part D is somewhat sophisticated than in plan for the non-[LIS.sup.3] population. CMS' Office of the Actuary estimated a 19-percent impact on spending from eminent the Part D benefit structure on these data. Thus, we reduced spending by 19 percent for non-institutionalized beneficiaries. Spending by institutionalized beneficiaries is assumed to be smaller quantity discretionary and invariant to the change contained by benefit structure.


We used the benefit structure rules applied to the adjusted spending to derive plan liability for respectively beneficiary. Payments were annualized by dividing by the fraction of the contribution year each beneficiary be eligible. In the regressions, the observations were weighted by alike eligibility fraction. Two models were estimated: (1) an overall spending model and (2) a plan liability model using the non-institutionalized beneficiaries.


MODELS


RxHCC


The RxHCC models enjoy the specification: [Cost.sub.it]=[beta.sub.0] + [beta.sub.1] Age/[Sex.sub.it] + [beta.sub.2] Org[Dis.sub. it] + [beta.sub.3] [RXHCC.sub.it-1] + [beta.sub.4] Disabled'[RXHCC.sub.it-1] + [epsilon.sub.it] where Age/Sex denotes 24 mutually exclusive age/sex cell, and OrigDis represents originally disabled status: those who are currently age 65 or over, but were first entitled to Medicare beforehand age 65 by disability. RxHCC is a vector of diagnostic categories; and Disabled RxHCC denotes three potential incremental payments for beneficiaries entitled by disability. The model is chemical addition across age/sex status, originally disabled status, and the RxHCC categories. The three disease groups beside additional payments for the disabled are schizophrenia, other crucial psychiatric disorders, and cystic fibrosis. These amounts are added to the main entry for the diagnosis. In the spending model, Cost denotes total prescription drug expenditures, while in the payoff model Cost denotes the plan liability.


Risk-Adjustment Spending Model


A risk-adjustment model predicting total drug spending at the person even is displayed in Table 2. The final spending model is comprised of 84 RxHCCs. Similar to the nouns of the CMS-HCC model, the final spending model excludes diagnostic categories when the diagnoses be vague/nonspecific, discretionary in medical treatment or coding, not significant predictors of drug use, or transitory or not admit of definitive treatment.


Because one cannot predict all of the subsequent year diseases and drug consequences from prior year diagnoses, the demographic coefficients are significant in vastness. The age/sex coefficients indicate that drug expenditures not directly associated next to the diseases in the model rise near age until they reach a fell for the age group 45-54. Older age groups tend to use fewer prescription drugs not accounted for by their agreed disease profile. The RxHCC coefficients reflect the average drug implication of different diseases to individuals. By far, the largest costs are associated with human immunodeficiency virus acquire immunodeficiency syndrome (HIV/MDS), but other disease groups also have substantial drug implication including diabetes, schizophrenia (especially among the disabled), multiple sclerosis, Parkinson's disease, and cystic fibrosis. Total costs of a disease to the Medicare Program, however, are driven by disease prevalence as well as the coefficient size.


Risk-Adjustment Plan Liability Model


Figure 1 illustrate that plan liability has a non-linear relationship to spending. If the coefficients from a spending model be applied to the plan liability amounts, the predictions would likely overestimate plan liability and be invalid. Consequently, we estimated the plan liability model using the in tune spending data. The plan liability coefficients are smaller than the coefficients for the spending model, and as would be expected, some changed more than others. For example, the HIV/AIDS coefficient fell from $12,314 to $2,028. The plan liability coefficient is substantially smaller than the corresponding spending coefficient when the disease imply drug use reaches the donut hole or above. Plans are not responsible for any of the costs between $2,250 and $5,100 contained by total and only 15 percent of the cost above $5,100. As such, diseases near high spending coefficients hold much lower coefficients in the plan liability model.


The model is ultimately expressed not surrounded by dollars, but as relative factors. The incremental dollars associated near each inconstant in the model are divided by the be going to predicted dollars to produce a relative costliness or risk factor. Summing the risk factors for an individual yield a total risk adjustment factor that, when multiplied by a base rate, yield an individualized capitation payment.


When the coefficients surrounded by the two models are expressed as relative factors, the differences are smaller. This is because the conversion to relative factor entails dividing respectively coefficient by the national mean for spending or liability, as appropriate. Dividing a life-size spending coefficient by a large spending imply produces results similar to dividing the smaller liability coefficient by the smaller liability mean. The proportionality is not uniform, however. Diseases characterizing beneficiaries who tend to own a large proportion of spending in the 100 percent cost sharing continuum, have their factor reduced by a greater proportion than others. Much of drug spending can have a nothing impact on plan liability.


Both the spending and the plan liability model have accurate predictive power. The Rz (i.e. the proportion of the total variation within the dependent variable that is to say explained by the model) exceeds 0.20. This is higher than the explanatory power for the models predicting the more erratic Parts A and B costs and comparable to other diagnosis based models for drugs within the literature.


New Enrollee Model


The new enrollee model is applied to beneficiaries for whom a year of complete diagnostic information does not exist. This includes not with the sole purpose those beneficiaries newly entitled to Medicare, it also includes those who be entitled to only Part A during the facts collection year or who were within an MA-PD plan during any part of the facts collection year.


The sample for the estimation of this model includes both those who are risk adjustable (i.e., those who be included in the prior regression) as well as those who want full diagnosis data, but enjoy eligible coverage and costs in the clearing year. The estimation is based solely on demographic characteristics.


The results of the foreign enrollee regression are shown in Table 4. All cell are mutually exclusive. For example, the predicted drug expenditures for a male, age 65, who is not originally disabled are $748.16, while predicted expenditures are $1,102.01 if he is originally disabled. The coefficients for both sexes indicate that beneficiaries originally entitled to Medicare due to disability hold much higher drug utilization than beneficiaries originally entitled due to age. Coefficients for females are also consistently greater than for males.


VALIDATION


Analyses hold been made of the predictive ratio (plan predicted liability in the facts divided by actual plan liability) for beneficiaries in deciles of predicted liability (Table 5). Predictive ratio above 1.0 indicate overprediction; ratios lower than 1.0 indicates underpredicfion. The model perform well for both the plan and Medicaid sample. The model over-predicts for the bottom and top deciles. Because a substantial portion of a person's risk factor is associated near age and sex, even when diseases are accounted for, the model tends to overpay for beneficiaries who are predicted to be contained by the lowest deciles of costs some of whom use no drugs. Unlike the case for Parts A and B, the model also overpredicts sum for the beneficiaries in the great decile of predicted costs. This is because the coefficients cannot fully reflect the flattening of plan liability for large spenders. In the middle deciles of predicted costs there is a small amount of underprediction.


Predictive ratios from an age/sex model are also presented for comparison. The age/sex model underperforms the RxHCC model for most of the deciles. The most worthy differences exist in the bottom and top deciles. The age/sex model overpredicts more contained by the low deciles and underpredicts rather than overpredicts contained by the highest decile.


Table 5 also reports predictive ratio for individuals who were hospitalized contained by the base year. The comparison between the age/sex model and risk-adjustment model is more than ever striking. The age/sex model overpredicts by 7 percent for individuals without hospitalizations, but underpredicts by 34 percent for individuals next to four or more hospitalizations. The risk-adjustment model predicts very accurately for beneficiaries next to fewer than four hospitalizations. Unlike the age/sex model, which underpredicts for the costliest enrollees, the risk model overpredicts for individuals next to the most hospitalizations.


SPECIAL ADJUSTMENTS


Medicare's LTI Subpopulations


It has be observed that the LTI (defined here as those in a nursing home for more than 90 days) are heavy users of drugs and that, to some extent, the pricing of their drugs is superior than pricing in the community. Many reasons related to pricing and utilization can be posited for the differences. Analysis of information from IMS, a leading collector of prescription drug sale data, have shown that for the most frequent drugs the mean price difference is roughly speaking 2 percent. The difference is larger for generic drug than brand name drug, but the brand label drug dominates when the data are expenditure weighted. To weigh up empirically the overall effect of being surrounded by an institution rather than within the community, the pooled plan/Medicaid data for the LTI population be analyzed to determine how much capitated payments should be changed from that which is predicted by the model.


In developing a model that predicts drug use from knowledge of prior year diagnoses the LTI populations be intentionally omitted because CMS and the Department of Health and Human Services wished to own a clear and separate adjustment for institutionalized status. Other modeling methods could have integrated the institutionalized into the model or structured a separate model for them. However, the LTI example size was relatively small. To derive the adjustment, the community model be used to predict spending and plan liability for the institutionalized enrollees. The actual spending and plan liability were later compared to the predicted to derive an adjustment factor.


Table 6 shows the predicted and actual means for spending by the LTI. The results indicate that actual spending by LTI beneficiaries exceeds predicted spending in the aged and disabled groups by 22 and 40 percent respectively. Increments of these amounts would be corrective for spending predictions. It is considerable to note that the tight-fisted predicted and actual spending for LTI patients falls into the 100 percent coinsurance range for the aged, and that the be set to actual spending for the disabled falls into the catastrophic range. Because the predicted suggest for the aged using the community model is one-third of the distance through the 100 percent coinsurance range; increments to spending related to institutionalization will also crash largely within the 100 percent coinsurance scope. The disabled model prediction is close to the catastrophic range and incremental spending related to institutionalization will tend to spill into the field for which plans have some liability. Spending change in the 100 percent coinsurance breadth result in no switch to plan liability.


Analysis of the effect of institutionalization on plan liability results in LTI adjustment factors consistent near the previous observations. The factors are smaller because 100 percent coinsurance reduce changes surrounded by plan liability. The aged liability increment multiplier is only 7.6 percent, down from the 22 percent for spending. The liability increment multiplier for the disabled is substantial at 21.1 percent, though one-half of specifically for spending. If an individual is both a low-income subsidy eligible beneficiary and is in long-term care, solitary the long-term care multiplier applies to that beneficiary.


Low-Income Subsidy


The populations eligible for the LIS subsidies are defined in the MMA. CMS' Office of the Actuary estimated multipliers for two groups spanning the LIS population (Table 7). They are 1.08 for Group 1 individuals and 1.05 for Group 2 individuals. Eligibility is defined on a concurrent idea. For example, if an individual is not defined as low income for January 2006, but is determined to be a Group I beneficiary for February 2006, the plan would receive the low income multiplier for February (and beyond), but not for January.


CONCLUSION


This article has presented the nouns of the CMS-RxHCC prescription drug risk-adjustment model implemented surrounded by 2006. A major brave to the work was finding and adapt data that would span the Medicare population and be justifiably geographically representative. Future work, using actual program data, is needed to evaluate the presentation of the model, to recalibrate on program data, and to develop subsequent generation models that may incorporate prior drug use. One of the issues for any model for drug spending is the adapt of available products over time. New high-priced drugs are mortal brought to market as elder drugs are becoming cheaper generics. How robust this type of model is in a dynamic market is a topic of great interest. The certainty that the model is used for only a portion of the total payments to plans make its absolute care less critical and allows time to develop potential improvements.


No comments: