improving health-based payment for medicaid …cdps.ucsd.edu/cdps_hcfr.pdfdevelopment of cdps the...

This article describes the Chronic Illnessand Disability Payment System (CDPS), adiagnostic classification system thatMedicaid programs can use to make health-based capitated payments for TANF anddisabled Medicaid beneficiaries. Theauthors describe the diversity of diagnosesand dif ferent burdens of illness among dis-abled and AFDC Medicaid beneficiaries.Claims from seven States are analyzed, andpayment weights are provided that Statescan use when adjusting HMO payments.The authors also compare the taxonomyand statistical performance of CDPS toother leading diagnostic classification sys-tems and find that the new model performsbetter in a number of respects.

INTRODUCTION

In previous work, we argued that health-based payment for Medicaid beneficiarieswith disabilities is both important and fea-sible (Kronick, Zhou, and Dreyfus, 1995;Kronick et al., 1996). Among people withdisabilities, health expenditures arestrongly related to recent diagnoses, andhealth plans are well aware that attractingtoo many people with costly problems canlead to large financial losses. If a StateMedicaid program does not pay more tohealth plans whose members have above-

average levels of need, it will penalize plansattractive to people with greater needs andjeopardize quality of care. The greater pre-dictability of expenditures among peoplewith disabilities compared with a generalpopulation both increases the importanceof health-based payment and makes it easi-er to do well. The strong relationshipbetween diagnoses and future expendi-tures allows Medicaid programs to usediagnoses to make good predictions ofhealth care needs.

We can now report that Medicaid pro-grams have been leaders in the implemen-tation of health-based payment (Table 1).Using diagnoses from both ambulatoryand inpatient encounters, Maryland imple-mented risk adjustment in May 1997,Colorado in July 1997, Oregon in June1998, and Delaware in January 2000.Using inpatient data only, Utah implement-ed a limited version of health-based pay-ment in June 1998. Utah is planning onexpanding to full diagnostic risk adjust-ment if the encounter data supplied byhealth maintenance organizations (HMOs)are of sufficient quality. Minnesota imple-mented health-based payment in Januaryof 2000, adjusting 5 percent of the capita-tion based on diagnostic case mix, with theremaining 95 percent based on traditionaldemographic rate cells. Michigan is usingdiagnostic adjustment as part of its com-petitive procurement process; Michigandivides a plan’s bid by that plan’s case mixin order to compare bids against eachother on an equitable basis. New Jerseyhas announced its intentions to implement

HEALTH CARE FINANCING REVIEW/Spring 2000/Volume 21, Number 3 29

Richard Kronick, Todd Gilmer, and Lora Lee are with theDepartment of Family and Preventive Medicine, University ofCalifornia, San Diego. Tony Dreyfus is an independent consul-tant. This research was supported by grants from the RobertWood Johnson Foundation and from the Health Care FinancingAdministration (HCFA). The views expressed in this article arethose of the authors and do not necessarily reflect the views ofthe University of California or HCFA.

Improving Health-Based Payment for Medicaid Beneficiaries: CDPS

Richard Kronick, Ph.D., Todd Gilmer, Ph.D., Tony Dreyfus, M.C.P., and Lora Lee, M.S.

risk-adjusted payments in 2000 and hasdone substantial work in preparation forimplementation. Other State Medicaidprograms, including Massachusetts, NewYork, and Pennsylvania, are seriously eval-uating health-based payment options.

Medicaid programs are much moreactive than private employers in imple-menting health-based payment, and someare ahead of the Medicare program. HCFAbegan phasing in risk-adjusted paymentsto Medicare in January 2000, using onlyinpatient diagnoses, while most Medicaidprograms implementing health-based pay-ment are using or planning to use diag-noses from both ambulatory and inpatientencounters.

Concerns about health care for benefi-ciaries with disability account for much ofthe impetus for Medicaid health-based pay-ment, although some States have extendedhealth-based payment to beneficiaries ofTemporary Assistance to Needy Families(TANF) as well. As health-based paymentin Medicaid is implemented more widely,we expect increased interest in its use forTANF beneficiaries. Yet relatively little

information is available about the burdenof disease among TANF beneficiaries, norhas much analysis been presented of theability of diagnostic classification systemsto do a good job of fairly compensatingHMOs for this population (Weiner et al.,1998). In this article, we:• Describe a diagnostic classification sys-

tem, CDPS, that we have developed as atool for State Medicaid programs to usewhen paying HMOs for SupplementalSecurity Income (SSI) and TANF benefi-ciaries.

• Describe the burden of disease amongSSI and TANF beneficiaries and theextent to which expenditure effects ofdiagnoses are similar or different in thetwo groups.

• Compare the structure and statisticalperformance of CDPS with those ofother leading diagnostic classificationsystems.

• Describe the extent to which diagnosespersist from year to year in fee-for-serviceclaims data and discuss the implicationsfor health-based payment systems.

30 HEALTH CARE FINANCING REVIEW/Spring 2000/Volume 21, Number 3

Table 1

Medicaid Health-Based Payment Activity

Population Date Classification DataState Covered Implemented System Source

ImplementedMaryland SSI + TANF 5/97 ACGs Prior FFS ClaimsColorado SSI + TANF 7/97 DPS HMO Encounter DataOregon SSI 6/98 DPS HMO Encounter DataUtah SSI 6/98 Marker Diagnosis Inpatient Only EncountersMinnesota1 TANF 1/00 ACGs HMO Encounter DataDelaware SSI + TANF 21/00 CDPS HMO Encounter DataMichigan SSI 6/00 CDPS HMO Encounter Data

PlannedNew Jersey SSI 2000 DPS Prior FFSDelaware SSI 2000 CDPS HMO Encounter DataWashington TANF 2001 CDPS HMO Encounter DataUtah3 SSI 2001 CDPS HMO Encounter Data1 Affects 5 percent of total capitation.2 TANF on 7/00.3 Dependent upon quality of encounter data.

NOTES: SSI is Supplemental Security Income. TANF is Temporary Assistance to Needy Families. ACGs is Adjusted Clinical Groups. FFS is fee-for-service. DPS is Disability Payment System. HMO is health maintenance organization. CDPS is Chronic Illness and Disability Payment System.

SOURCE: Kronick, R., et al., San Diego, California, 2000.

DEVELOPMENT OF CDPS

The revision of the Disability PaymentSystem (DPS) was intended to make thesystem more complete and more effectivein its adjustment of payments for the TANFpopulation. We used a much larger database than previously, with claims recordsfor nearly 4 million Medicaid beneficiariesfrom seven States. Effects of diagnoses onfuture expenditures were analyzed for allthe 15,000 diagnosis codes in theInternational Classification of Diseases, 9thRevision, Clinical Modification (ICD-9-CM)(Public Health Service and Health CareFinancing Administration, 1980), andphysician specialists were consulted exten-sively to help determine the appropriate-ness and organization of diagnoses includ-ed in the new system.

The product of the revision, the CDPS,includes 20 major categories of diagnoses,which correspond to body systems or typeof diagnosis. (For prospective estimationof payment weights, we exclude the cate-gories for infants, leaving the model with19 major categories.) Most of the majorcategories are further divided into severalsubcategories according to the degree ofthe increased expenditures associated withthe diagnoses. For example, diagnoses ofthe nervous system are divided into threesubcategories for high-cost, medium-cost,and low-cost conditions. (Refer to Table 2for a list of the subcategories and samplediagnoses.)

CDPS includes many more diagnoses andworks better than the original DPS in pre-dicting expenditures for Medicaid beneficia-ries of all types.1 Medicaid programs can

use the new CDPS with greater confidencethat the system fully exploits diagnoses thatpredict significantly elevated future expendi-tures and that are sufficiently well definedfor payment purposes. Software to imple-ment CDPS is available at no charge athttp://www.medicine.ucsd.edu/fpm/cdps/.

Data

The initial selection of diagnoses forCDPS was based on analysis of expendi-ture data for approximately 600,000 dis-abled Medicaid beneficiaries and 3.3 mil-lion Aid to Families with DependentChildren (AFDC) and AFDC-relatedMedicaid beneficiaries (Table 3).2 Thedata base was substantially larger than thedata base for the original DPS, which usedonly 120,000 SSI Medicaid beneficiaries intwo States for identification of diagnosesand 400,000 beneficiaries in five States forfinal testing and determination of cate-gories. Our data contained information onservices and procedures, Medicaid pay-ments, and diagnoses, including usuallyone diagnosis code for ambulatory claimsand up to five diagnosis codes in mostStates for inpatient claims.

To focus on more complete diagnosticrecords, we included in the analysis onlythose beneficiaries with a full initial year ofMedicaid eligibility and at least 1 month ofsubsequent-year data. We excluded bene-ficiaries with Medicare coverage and thoseenrolled in health plans, for whomMedicaid has partial or no claims informa-tion. We also excluded from the analysisbeneficiaries in institutions and those in


1 The new name, Chronic Illness and Disability Payment System,instead of Disability Payment System, is intended to correct themistaken impression that the payment system consists primari-ly of disabilities or could be used only for a disabled population.The vast majority of the diagnoses in the old and new models arenot disabilities but diagnoses of disease. Some of these diag-noses are very serious and could be disabling, e.g., musculardystrophy, but many others, e.g., migraines or uncomplicatedadult-onset diabetes, are unlikely to be disabling conditions.

2 The disabled beneficiaries include people receiving SSI as wellas others who are eligible for Medicaid because of disability.The AFDC program has been transformed to TANF, but all ofour data predate the conversion. Welfare reform has resulted ina sharp decline in the Medicaid caseload, and there is specula-tion that the remaining caseload is sicker (Ellwood and Lewis,1999). However, we think that the relationships between diag-noses and expenditures that we find among AFDC beneficiariesin the early and mid-1990s will be applicable to current Medicaidbeneficiaries as well.


Table 2

Chronic Illness and Disability Payment System Categories with Sample Diagnoses

Diagnostic Category Sample Diagnoses

CardiovascularVery High Heart transplant status or complicationsMedium Congestive heart failure, cardiomyopathy, tricuspid and pulmonary valve diseaseLow Endocardial disease, myocardial infarction, angina, coronary atherosclerosis, dysrhythmiasExtra Low Hypertension

PsychiatricHigh SchizophreniaMedium Bipolar affective disorderLow Other depression, panic disorder, phobic disorder

Skeletal and ConnectiveMedium Chronic osteomyelitis, aseptic necrosis of boneLow Rheumatoid arthritis, osteomyelitis, systemic lupus, traumatic amputation of foot or legVery Low Osteoporosis, musculoskeletal anomalies, thoracic and lumbar disc degenerationExtra Low Osteoarthrosis, skull fractures, other disc and vertebral disorders

Nervous SystemHigh Quadriplegia, amyotrophic lateral sclerosis and other motor neuron diseaseMedium Paraplegia, muscular dystrophy, multiple sclerosisLow Epilepsy, Parkinson’s disease, cerebral palsy, migraine, cerebral degeneration

PulmonaryVery High Cystic fibrosis, lung transplant, tracheostomy status, respirator dependenceHigh Respiratory arrest or failure, primary pulmonary hypertension, selected bacterial pneumoniasMedium Other bacterial pneumonias, chronic obstructive asthma, adult respiratory distress syndromeLow Viral pneumonias, chronic bronchitis, asthma, COPD, emphysema

GastrointestinalHigh Peritonitis, hepatic coma, liver transplantMedium Regional enteritis and ulcerative colitis, chronic liver disease and cirrhosis, enterostomyLow Ulcer, hernia, GI hemorrhage, intestinal infectious disease, intestinal obstruction

DiabetesType 1 High Type 1 diabetes with renal manifestations or coma Type 1 Medium Type 1 diabetes without complications or with neurological or ophthalmic complicationsType 2 Medium Type 2 or unspecified diabetes with complications, proliferative diabetic retinopathyType 2 Low Type 2 or unspecified diabetes without complications

SkinHigh Decubitus ulcerLow Other chronic ulcer of skinVery Low Cellulitis, burn, lupus erythematosus

RenalVery High Chronic renal failure, kidney transplant status or complicationsMedium Acute renal failure, chronic nephritis, urinary incontinence, cystostomy or urinostomyLow Kidney infection, kidney stones, hematuria, urethral stricture, bladder disorders

Substance AbuseLow Opioid, barbiturate, cocaine, amphetamine abuse or dependence, drug psychosesVery Low Alcohol abuse, dependence, or psychosis

CancerHigh Lung cancer, ovarian cancer, secondary malignant neoplasms, leukemia, multiple myelomaMedium Mouth, breast or brain cancer, malignant melanoma, radiation or chemotherapyLow Colon, cervical, or prostate cancer, carcinomas in situ

Developmental DisabilityMedium Severe or profound mental retardationLow Mild or moderate mental retardation, Down’s syndrome

GenitalExtra Low Uterine and pelvic inflammatory disease, endometriosis, hyperplasia of prostate

See footnotes at end of table.


Table 2—Continued

Chronic Illness and Disability Payment System Categories with Sample Diagnoses

Diagnostic Category Sample Diagnoses

MetabolicHigh Panhypopituitarism, pituitary dwarfism, non-HIV immunity deficienciesMedium Kwashiorkor, marasmus, and other malnutrition, parathyroid, and adrenal gland disordersVery Low Other pituitary disorders, gout

PregnancyIncomplete Normal pregnancy, complications of pregnancyComplete Normal delivery, multiple delivery, delivery with complications

EyeLow Retinal detachment, choroidal disorders, vitreous hemorrhage Very Low Cataract, glaucoma, congenital eye anomaly, corneal ulcer

CerebrovascularLow Intracerebral hemorrhage, precerebral occlusion, hemiplegia, cerebrovascular accident

Infectious DiseaseAIDS, High AIDS, pneumocystis pneumonia, cryptococcosis, Kaposi’s sarcomaInfectious, High Staphylococcal or pseudomonas septicemia, cytomegaloviral diseaseHIV, Medium Asymptomatic HIV infectionInfectious, Medium Other septicemia, pulmonary or disseminated candida, toxoplasmosis, typhusInfectious, Low Poliomyelitis, oral candida, herpes zoster, parasitic intestinal infections

HematologicalExtra High Congenital factor VIII and factor IX coagulation defects (hemophilia)Very High Hemoglobin-S sickle-cell disease Medium Other hereditary hemolytic anemias, aplastic anemia, splenomegaly, agranulocytosisLow Other white blood cell disorders, purpura, other coagulation defects

NOTES: COPD is chronic obstructive pulmonary disease. GI is gastrointestinal. HIV is human immunodeficiency virus. AIDS is acquired immunodeficien-cy syndrome. CDPS is Chronic Illness and Disability Payment System. CDPS also includes categories for infants and a more detailed categorization forpregnancy. A complete description of CDPS diagnostic categories by ICD-9-CM codes is available at http://www.medicine.ucsd.edu/fpm/cdps/. ICD-9-CMis International Classification of Diseases, 9th Revision, Clinical Modification (Public Health Service and Health Care Financing Administration, 1980).


Table 3

Number of Observations in Regression Analysis, by State and Beneficiary Group:Selected States

Adults with Children with AFDC AFDCState Totals Disability Disability Adults Children

Total 6,280,443 960,760 130,324 1,548,488 3,640,871

California 3,415,068 402,987 39,427 905,474 2,067,180 Colorado 338,925 50,454 15,701 85,221 187,549 Georgia 667,424 88,538 16,848 148,709 413,329 Michigan 1,007,649 118,996 16,758 288,462 583,433 Missouri 72,270 67,886 4,384 — —Ohio 137,451 118,700 18,751 — — Tennessee 641,656 113,199 18,455 120,622 389,380

Unduplicated Countof Beneficiaries 3,936,626 549,595 80,646 1,001,775 2,304,610

NOTES: AFDC is Aid to Families with Dependent Children. Data from Michigan, Ohio, and Tennessee are for 1991-1993; California and Georgia arefor 1990-1992; Missouri, 1991-1994; Colorado, 1992-1996. Observations are included in the regression analyses for beneficiaries with 12 months ofeligibility in the base year and at least 1 month in the rate year. Beneficiaries were excluded if they had Medicare coverage, were institutionalized,enrolled in a health maintenance organization, or in a home and community-based waiver program. The unduplicated count of beneficiaries is lowerthan the total number of observations because beneficiaries who were continuously eligible for more than 24 months account for 2 or more observa-tions. The table lists the total number of observations; model development was performed on a 75-percent sample.


home and community-based waiver pro-grams, who are not often enrolled in man-aged care. We calculated expenditures forservices typically included in an acute careHMO benefit package, excluding dentalservices and long-term care services.(Approximately 50 percent of expendituresfor beneficiaries with disability are madefor long-term care, but these services arenot currently included under capitated con-tract.) We set aside one-quarter of the datafor use as a validation sample, which left uswith 3 million individuals for use in thedevelopment of the model.

In the regression analysis, we combinedmultiple years of data to increase the sta-bility of the estimates. As a result, thenumber of observations was larger thanthe unduplicated count of beneficiaries. Toobtain results that could be useful in a vari-ety of Medicaid programs, we modified thedata from the seven States in two ways.First, to minimize the effects of interstateand interyear variation in the level ofexpenditures, we normalized expendituresin each State-year to 1.0 by dividing theexpenditures per month by the meanexpenditures per month for the State-year.These normalized expenditures were thenused as the dependent variable for regres-sion against prior-year diagnoses. Second,we weighted the observations used in theregression, so that the set of observationsfrom each State received equal weight. Inthis way, we avoided results dominated bythe expenditure patterns in California,which had a large share of the total obser-vations. We also reduced the weight forbeneficiaries who were eligible for onlypart of the year in which expenditureswere observed. We analyzed the residualsfrom an unweighted regression, found thatthe standard deviation of the residuals wasapproximately four times larger for per-sons eligible for 1 month than persons eli-gible for 12 months and weighted each

observation by 1 - 0.067 x (12 - number ofeligible months). This method decreasesthe weight progressively more for benefi-ciaries with less eligibility in the secondyear and reduces the influence of theshorter period observations in rough pro-portion to their increasing variability.

Method of Analysis

The selection and grouping of diagnosesfor CDPS depended upon analysis of ourexpenditure data and on the advice of 15clinician consultants. The basic method ofanalysis was to use the presence of diag-noses recorded in the first year of individu-als’ claims as regression variables to pre-dict expenditures in individuals’ subse-quent year of claims. We empirically iden-tified diagnoses that are significantly asso-ciated with increased future health carecosts. These diagnoses, largely chronicconditions, can serve the aim of health-based payment to provide additionalresources to plans that enroll people withgreater ongoing needs.

An important challenge to any effort toconstruct a diagnosis-based payment sys-tem is the defining of diagnoses in terms ofICD-9-CM codes. These 15,000 codes areorganized under nearly 1,000 three-digitgeneral codes, nearly all with four-digit orfive-digit subcodes required for referenceto a more specific diagnosis. For example,the ICD-9-CM code 428 refers to heart fail-ure, but more specific four-digit codes are428.0 for congestive heart failure and 428.1for left heart failure. Some codes arevague, such as 429.2 for unspecified car-diovascular disease, while others are morespecific, such as 250.43 for uncontrolledType 1 diabetes with renal manifestations.

Creating the diagnostic classificationsystem required decisions about what levelof detail should be used in defining eachdiagnosis in the system. For example,


consider hemoglobin-S disease, a type ofsickle-cell anemia. Should hemoglobin-Sdisease with crisis (282.62) be considered itsown condition, separate from hemoglobin-Sdisease without mention of crisis (282.61)?Or should these two conditions be consid-ered a single entity? Or should hemoglo-bin-S disease be combined with other sick-le-cell anemias or with all the hereditaryhemolytic anemias?

Defining a diagnosis more narrowlyappears to give greater accuracy in predict-ing expenditures, but too narrow a definitioncould make it difficult for clinicians to agreewhether an individual’s condition justifies agiven diagnosis, and could lead to unstableexpenditure estimates. We were eager todistinguish among diagnoses that wereassociated with markedly different levels ofelevated future cost, but we did not want toseparate codes into different groups if ourconsulting clinicians believed that it mightbe difficult to distinguish between them. Ingeneral, we regarded groups of diagnosescoded with the same first three digits as like-ly candidates for the codes to define a singlediagnosis, and we broke up three-digitgroups only when both the data suggestedand clinician judgment concurred that thesewere distinct conditions that could have dif-ferent effects on future cost.

Even in cases where we found signifi-cantly different effects on expenditures fortwo separately coded diagnoses, we keptthem together in a single variable if ourclinical consultation suggested that the dis-tinction between the diagnoses could notbe easily made. We found, for example,that the diagnosis of paranoid schizophre-nia was associated with more elevatedfuture cost than catatonic schizophrenia,but our clinical consultants were not con-vinced that diagnoses of subtypes of schiz-ophrenia were made consistently, so wekept all diagnoses of schizophrenia togeth-er as a single variable.

Excluding Ill-Defined Diagnoses

Much of our consultation with clinicianswas intended to screen out diagnoses thatare clinically not well defined. We madespecial efforts to exclude ill-defined diag-noses from CDPS in order to make the sys-tem more reliable and reduce the chancesthat health plans, clinicians, and Medicaidprograms would find themselves question-ing diagnoses. Given that health-basedpayment will naturally cause plans to makegreater efforts to report diagnoses, a focuson well-defined diagnoses seems advisableto prevent difficult disagreements betweenpayers and plans. Ill-defined diagnoseswould make it difficult for States to auditplans and distinguish between accurateand inaccurate reporting. Our efforts toexclude ill-defined diagnoses parallel thework of other researchers in health-basedpayment. Notably, Ash and colleaguesexclude many diagnoses from theirHierarchical Condition Category (HCC)model as too vague to be used for adjust-ment of payments (Ash et al., 1998).

We considered a diagnosis to be welldefined if it had a clear, shared meaningamong clinicians. The diagnosis should bedistinctive enough that an auditing clini-cian could judge from a good medicalrecord whether the diagnosis was made onan adequate clinical basis. Laboratoryresults or diagnostic imaging help makemany diagnoses well defined, but manydiagnoses rely significantly on physicianobservation. As a result, diagnosis-basedpayment resembles traditional billing in itsreliance on honest physician observation,and States will need to audit diagnoses tokeep diagnostic reporting honest.

Although many diagnoses that wouldhelp predict future cost were judged as notwell defined and were excluded from themodel, the majority of diagnoses that arepredictive of elevated future costs were


judged as well defined. Many of theexcluded diagnoses came from sections inthe ICD-9-CM specifically set up for ill-defined descriptions of disease or forrecording symptoms that have not yetbeen tied to a specific disease.

Some of the conditions judged to be ill-defined were diagnoses that lack clear defi-nition for clinicians, such as chest pain ordyspnea. We also excluded many commonsymptoms that we judged to be too easilyelicited in patient histories: Symptoms suchas headache, backache, or joint pain mightbe recalled by any adult from some time inthe months previous to a physician visit.

In a few instances we bent our standardsto include diagnoses in CDPS that a clini-cian thought were not well defined. Forexample, because of our concern that someindividuals with dementia or traumaticbrain injury might not otherwise be recog-nized by CDPS, we included the diagnosesof Alzheimer’s disease and of non-psychoticmental disorders due to organic brain dam-age despite clinician concerns that theseconditions are not well defined. We alsoincluded asthma and chronic bronchitis,but not acute bronchitis, even though thedistinction between chronic bronchitis andasthma on the one hand and acute bronchi-tis on the other may not always be wellobserved. We did so because asthma canbe quite expensive, and for many people,asthma is likely their only CDPS diagnosis.We anticipate further research and contro-versy about the inclusion of more and lesswell-defined conditions in payment sys-tems, because many of the conditions weexclude for being ill-defined are useful pre-dictors of future cost.

The inclusion of ill-defined diagnosesmay increase predictive accuracy but willlikely reduce accuracy in implementation.In general, as more diagnoses are includedin a payment system, a greater volume ofdiagnoses needs to be reported and audit-

ed, and a higher proportion of variation inlevel of need observed among plans wouldresult from differences in plans’ abilities tomake and report diagnoses rather thanfrom actual differences in their enrollees.It seems likely that the inclusion of ill-defined diagnoses would particularly makethe payment system more vulnerable toaggressive plan efforts to increase report-ing. The modest improvement in accuracyon a given data set that is gained throughill-defined diagnoses seems far less impor-tant than having a system that is more eas-ily administered and probably more accu-rate in practice.

We believe that the exclusion fromCDPS of ill-defined conditions is a virtue,and we encourage others to consider thisissue more carefully. Our work is far fromsufficient to settle the question of whichdiagnoses are well defined and which arenot. For each diagnosis that our data indi-cated was predictive of elevated costs, ourapproach was to ask specialists directlyhow well defined they thought the diagno-sis was. We did not ask about the manydiagnoses that failed to show any associa-tion with elevated future cost. A muchmore intensive approach might involveasking clinicians to make diagnoses fromsample medical records and regardingdiagnoses as ill-defined where the clini-cians’ test diagnoses show weak agree-ment. It might be important to include bothspecialist and primary physicians becausesome diagnoses might be well defined forthe specialist but not the generalist.

Excluding Low-Cost Diagnoses

A related question is whether groups ofdiagnoses with high frequency and verylow-cost implications should be included ina payment system. For example, bladderand urethral infections were diagnosed for95,000, or 10 percent, of our sample of adults


with disability and more than 185,000, or 12percent, of AFDC adults. Estimated addi-tional monthly costs in the next year associ-ated with this diagnosis were only $12 foradults with disability and $11 for AFDCadults. A respiratory tract condition such assinusitis, pharyngitis, acute bronchitis, orcough was diagnosed for 1.5 million, or 42percent, of the AFDC children in the sampleand for 33 percent of AFDC adults.Estimated additional monthly costs in thenext year were $8 for the AFDC childrenand $13 for the AFDC adults.

These additional amounts are very smallin comparison with the additional amountsin the range of $200-800 for the more cost-ly diagnostic groups. The cost effects ofless than $30 are also small relative to theaverage monthly expenditures for peoplewith disability in our sample of $416. ForAFDC beneficiaries, however, these smallcost effects are more significant becauseaverage AFDC adult expenditures are $158per month, and average expenditures forAFDC children are $57 per month.

In theory, if the gathering and reportingof diagnoses were perfect, it would beadvantageous to include high-frequency,low-cost diagnoses because their presencewould increase accuracy and fairness bybringing more money to the plans thatserve people with greater needs. In prac-tice, however, plans’ ability and eagernessto make and report diagnoses might vary.As we argued previously, the more high-frequency, low-cost diagnoses are includ-ed, the more apparent variation in need islikely to result from differences in report-ing, not actual differences in need.

An important consideration in decidingwhether to include diagnoses for paymentpurposes is whether there is reason toexpect uneven distribution of enrolleeswith a certain diagnosis across plans. Formany of the extremely low-cost diagnoses,such as bladder infections, minor upper

respiratory conditions, or ear infections,there is little reason to expect that peoplewith these conditions will be distributedunevenly among plans. On the other hand,some other low-cost conditions, such ashypertension, migraines, or asthma, havesomewhat higher cost effects and might bedistributed unevenly among plans, andplans that have stronger specialist net-works or are located in poorer neighbor-hoods might attract a disproportionateshare. In addition, encouraging plans todiagnose these conditions has a value initself because attention to them can behighly beneficial for individual health. Theinclusion of hypertension is particularlyimportant for this reason and also because13 percent of disabled adults are codedwith hypertension but not more seriouscardiovascular disease.

As a result of all these considerations,we decided to eliminate many of the lowercost conditions from our recommendedpayment model. We recommend the use of56 diagnostic subcategories for paymentpurposes and an additional 15 subcate-gories of high-frequency diagnoses withvery small cost effects only for profilingpurposes.

Counting Diagnoses with the CDPSSubcategories

The organization of diagnostic cate-gories and the rules for counting diag-noses are somewhat different in CDPSthan they were in the original DPS. Themost obvious change is an increase in thenumber of diagnostic subcategories, from43 in DPS to 56 in CDPS, which resultspartly from the more comprehensive andlarger set of diagnoses included in CDPS.Some of the new subcategories result fromincreasing distinctions among diagnosesthat were in DPS, while other new subcate-gories are in new major areas. New major


areas include infectious disease, pregnan-cy, and infants. Two new subcategoriesresulted from creating separate subcate-gories for Type 1 and Type 2 diabetes. Andseveral other major categories gained asubcategory to reflect finer distinctionsamong cost levels.

A less obvious but equally significantchange is in the rules used for countingdiagnoses within major categories. In theoriginal DPS, 10 of the 18 major categorieswere designated as “hierarchic” categoriesin which only the single most severe diag-nosis within the major category was count-ed, while 8 were designated as “fully count-ed” categories in which multiple diagnosescould be counted. Our use of fully countedcategories had been intended to capturethe additional needs that arise from dis-tinct diseases, but in revisiting this issue,we placed a higher value on limiting incen-tives for proliferative coding and on consis-tency across major categories. We alsofound relatively little predictive benefit incounting multiple diagnoses within majorcategories. As a result, every one of themajor categories in CDPS is counted hier-archically. This change in the countingrules simplifies the model, strengthens itsresistance to additional coding, and pro-duces only small decreases in the accuracyof simulated payments.

Single counting within major categories isintended to avoid encouraging a prolifera-tion of different diagnoses reported for a sin-gle disease process just to increase payment.For example, if someone is diagnosed with asignificant cardiovascular disease such ascongestive heart failure, an additional diag-nosis of hypertension is probably not ofmuch additional significance for cost. Inother cases, additional coding would clearlyhave no implication for cost, for example, ifsomeone with Type 1 diabetes with compli-cations were subsequently coded withuncomplicated Type 1 diabetes.

Thus, if an individual has a medium-costinfectious disease and a low-cost infectiousdisease, he or she would be counted inCDPS as simply having medium-cost infec-tious disease. An individual with two dif-ferent medium-cost psychiatric illnesseswould be counted simply as having medium-cost psychiatric illness. As aresult of this approach, the expendituresassociated with people with multiple diag-noses in a single major category are loadedonto the single-highest category.

We experimented with various countingapproaches, stimulated in part by theexample of Ash and colleagues (1998), whodeveloped counting methods that mixcounting and hierarchy within diagnosticcategories. We tried an intermediateapproach between full counting and singlecounting: Major categories were subdivid-ed into different diagnostic areas and addi-tional counts were allowed for diagnoses inthe different areas. For example, we divid-ed cardiovascular diagnoses into areassuch as valvular, myocardial, dysrhythmic,and peripheral conditions. We found, how-ever, that such subdivisions added sub-stantially to the complexity of the modelbut yielded very little improvement in itsperformance.

Meanwhile, CDPS follows DPS in count-ing multiple diagnoses when they are fromdifferent major categories. Thus, if an indi-vidual had two medium-cost renal diag-noses and low-cost developmental disabili-ty, he or she would be counted as havingmedium-cost renal disease and low-costdevelopmental disability. By consideringnot only a person’s single most seriousdiagnosis but also diagnoses from othermajor categories, accuracy is substantiallyimproved. Using diagnoses from multiplemajor categories improves accuracybecause average expenditures are muchhigher for people with diagnoses fromgreater numbers of major categories.


In addition, the diagnosis of additionalconditions from different major categoriesoffers a greater potential benefit for individ-uals than the additional coding of highlyrelated diagnoses. For a male already diag-nosed with schizophrenia who receives anadditional diagnosis of paranoid state, theadditional diagnosis may not be associatedwith any new treatment he would other-wise not have received. But the diagnosisof a gastrointestinal or cardiovascular con-dition could bring him valuable additionalcare. By counting diagnoses from differentmajor categories, CDPS encourages identi-fication of additional chronic conditionsthat may deserve attention. Such attentionto the whole individual is important foreveryone but can be particularly importantfor people with serious mental illness ordevelopmental disability, whose medicalattention has often focused on their cogni-tive problems and neglected physicalhealth (Druss et al., 2000).

Estimating Payment Weights forSubgroups

The population of Medicaid beneficiariescan be treated as a single group or dividedinto subgroups when estimating paymentweights. If a single group for all Medicaidbeneficiaries is used, then variables for cat-egory of assistance and age and interac-tions of these variables with diagnosticgroups can be introduced to better fit themodel. Alternatively, the model can be esti-mated separately on subgroups such as theAFDC adults or the SSI children. If theeffects of diagnoses on expenditures aresimilar across groups or if there are rela-tively few people in a subgroup, then a sin-gle model with supplementary variableswill work well. On the other hand, if theeffects of diagnoses are dissimilar acrosssubgroups and the subgroups are large,then separate models are preferable.

We estimated separate weights for per-sons with disability and AFDC beneficiariesbecause: (1) there are very large differ-ences in the expenditure effects of CDPScategories between adults with disabilityand those enrolled in AFDC; (2) there arelarge differences in the average expendi-tures for the two groups; (3) we have largesamples of both AFDC adults and thosewith disability; and (4) State Medicaid pro-grams typically have separate base rates forSSI beneficiaries and TANF beneficiaries.

Among persons with disabilities, we esti-mated a combined regression for adultsand children because we had found thatparameter estimates from regressions esti-mated separately on adults and childrenwith disabilities were relatively similar. So,too, were average expenditures per monthand the average number of CDPS diag-noses. We also used the combined regres-sion because most States do not have a sep-arate base rate for non-Medicare adultsand children with disabilities. Althoughthe effects of CDPS diagnoses on expendi-tures are similar among adults and chil-dren with disabilities, they differ in someareas. We estimated a regression withinteractions of age (coded as a dichoto-mous variable for under or over 18) andeach of the CDPS categories, and retainedin the final model 11 interactions withCDPS categories that appeared useful.(We selected interactions that had substan-tial numbers of children, t-statistics withabsolute values above 2.0 [with the excep-tion of diabetes], and reasonably stableestimates across States.) For 9 of these 11categories, the effects of diagnoses arelarger among children with disabilitiesthan among adults with disabilities.

In contrast to our approach of using acombined model for adults and childrenwith disabilities, we estimated separateregressions for AFDC adults and AFDCchildren for several reasons. First, the


expenditure effects of CDPS diagnoses aresubstantially different for AFDC childrenand adults in a large number of CDPS sub-categories.3 Second, many States haveseparate base rates for AFDC children andadults, making separate payment weightsuseful. Third, we have a very large sampleof AFDC children and adults, supportingthe estimation of reasonably stable pay-ment weights for the two groups.

DISEASE BURDEN AND EXPENDITURE EFFECTS

CDPS provides a detailed diagnosticdescription of Medicaid beneficiaries interms of the frequency of different kinds ofdiagnoses and their expenditure effects.Our analysis shows that only a small pro-portion of the SSI disabled have suchsalient conditions as paraplegia, acquiredimmunodeficiency syndrome (AIDS), orcystic fibrosis, which policymakers or thepublic may commonly associate with dis-ability and Medicaid. Instead, Medicaidbeneficiaries with disability experience awide variety of serious and less seriousconditions to which policymakers andhealth plans might give more attention.

Most of the more serious diagnoses arefound much less frequently among AFDCadults than among adults with disability.Among children receiving Medicaid, chil-dren with disability have much greater rel-ative frequency of serious diagnoses, butthe absolute numbers of AFDC childrenenrolled in Medicaid are so much largerthan the number of SSI children that thenumber of AFDC children with a givenserious condition may be greater than thenumber of disabled children with that con-

dition. (When we refer to “children withdisability,” we mean those who areMedicaid beneficiaries because theyreceive SSI as well as other children whoare eligible for Medicaid because of dis-ability. As we show later, there are manyAFDC children who also have chronic ill-nesses and disabilities.) The diagnosticdescription details the diverse challengesof providing health care to Medicaid bene-ficiaries with disability and to AFDC bene-ficiaries.

Disease Burden for Adults

Among adults with disabilities, approxi-mately one-quarter of beneficiaries have apsychiatric or a cardiovascular diagnosisthat is included in CDPS, and approximately10-15 percent of beneficiaries have a diagno-sis in the skeletal, central nervous system,pulmonary, gastrointestinal, or diabetesCDPS categories (Figure 1 and Table 4).4Smaller numbers have a diagnosis in otherCDPS categories such as renal, substanceabuse, cancer, and metabolic. There are afew CDPS categories, such as infectiousdiseases (including AIDS) and hematologythat are recorded for only 2 percent of adultbeneficiaries with disabilities.

With the exceptions of pregnancy anddiagnoses in the genital category, theprevalence of major CDPS categories forAFDC beneficiaries is uniformly lowerthan for persons with disabilities. Formany major CDPS groups, the frequencyamong AFDC beneficiaries is approximate-ly one-third to one-half of the rate amongpersons with disabilities. Pregnancy, how-


3 We estimated a regression on AFDC adults and children com-bined, with interaction terms for age and CDPS categories.Using the decision rule that we would keep an interaction termin the model if the CDPS category had substantial numbers ofchildren, had a significant t-statistic, and was similar in regres-sions on individual States, we would have kept interaction termsfor slightly more than one-half of the 56 CDPS categories.

4 The frequencies in Table 4 are weighted frequencies, givingequal weight to the observations for each State. The weight foreach observation is proportional to (total number of observa-tions in a beneficiary group across all States)/(total number ofobservations in the State in the beneficiary group). The fre-quencies for most categories vary relatively little across States,so that the weighted and unweighted frequencies are, for themost part, similar. However, for a few CDPS categories,California is notably different, and the weighted frequenciesgive less weight to California than the unweighted.

ever, is far more common for AFDC adults:24 percent of AFDC adults who were con-tinuously eligible for at least 12 monthseither complete a pregnancy or were preg-nant during the year; in contrast, pregnan-cy is uncommon among persons with dis-ability. (Completed pregnancy includesmiscarriages, abortions, normal delivery,and delivery with complications; incom-plete pregnancy includes normal and com-plicated pregnancies.)

There are a number of high-cost andvery-high-cost CDPS categories that arerare among adults with disabilities but areclose to non-existent among AFDC adults

(Figure 2). For example, very-high-costcardiovascular problems (primarily hearttransplants) are coded for 0.23 percent ofpersons with disabilities but only for 0.02percent of AFDC adult beneficiaries; high-cost central nervous system problems(primarily quadriplegia) are diagnosed in0.34 percent of persons with disabilitiesand less than 0.01 percent of AFDC benefi-ciaries. Even more striking, high-cost psy-chiatric problems (primarily schizophre-nia) are relatively common among personswith disabilities (11.7 percent of beneficia-ries) but are diagnosed in only 0.3 percentof AFDC adults.


Hematological

Infectious

Cerebrovascular, Low

Eye

Pregnancy

Metabolic

Genital, Extra Low

Developmental Disability

Cancer

Substance Abuse

Renal

Skin

Diabetes

Gastrointestinal

Pulmonary

Central Nervous System

Skeletal

Psychiatric

Cardiovascular

Adults with Disability

Percent of Beneficiaries

CD

PS

Cat

ego

ry

AFDC Adults

0 105 15 20 25 30

NOTES: CDPS is Chronic Illness and Disability Payment System. AFDC is Aid toFamilies with Dependent Children. Claims and eligibility data from Michigan, Ohio, andTennessee, 1991-1993; California and Georgia, 1990-1991; Missouri, 1992-1994;Colorado, 1992-1996. (AFDC data were not available from Ohio or Missouri.)Beneficiaries are included if they have 12 months of eligibility in the base year and atleast 1 month in the subsequent year. Beneficiaries were excluded if they had Medicarecoverage or were institutionalized, enrolled in a health maintenance organization, orenrolled in a home and community-based waiver program. Frequencies are weighted;each State gets equal weight.


Figure 1

Frequency of Major CDPS Categories for Adults with Disability and AFDC Adults


Table 4

Frequency of CDPS Categories, by Beneficiary Group

Disabled AFDC Disabled AFDCAdults Adults Children Children

Category (n) = 960,760 (n) = 1,548,488 (n) = 130,324 (n) = 3,640,871

PercentCardiovascular 27.06 9.38 7.98 1.19

Very High 0.23 0.02 0.66 0.01Medium 3.52 0.46 0.04 0.00Low 11.13 3.80 6.53 1.00Extra Low 12.18 5.10 0.75 0.18

Psychiatric 22.67 6.83 11.64 3.32High 11.65 0.34 0.36 0.04Medium 1.63 0.27 0.40 0.06Low 9.39 6.22 10.88 3.22

Skeletal 16.81 8.23 12.01 3.08Medium 0.26 0.03 0.12 0.01Low 4.17 1.43 3.66 0.60Very Low 4.29 3.09 7.25 1.90Extra Low 8.09 3.68 0.98 0.57

Nervous System 16.65 5.87 31.02 2.78High 0.34 0.01 0.68 0.00Medium 1.86 0.27 5.37 0.10Low 14.45 5.59 24.97 2.67

Pulmonary 16.10 8.66 15.36 9.91Very High 0.21 (1) 1.56 (1)

High 0.94 0.21 0.61 0.22Medium 0.87 0.27 1.17 0.24Low 14.08 8.18 12.02 9.45

Gastrointestinal 12.59 6.96 7.48 3.98High 0.29 0.09 0.44 0.02Medium 2.11 0.68 1.31 0.15Low 10.19 6.19 5.73 3.81

Diabetes 11.25 4.23 0.94 0.45Type 1 High 0.11 0.01 (1) (1)

Type 1 Medium 2.61 0.45 (1) (1)

Type 2 Medium 0.63 0.10 (1) (1)

Type 2 Low 7.90 3.67 0.94 0.45

Skin 7.88 4.37 4.93 3.49High 0.48 0.02 0.28 0.01Low 0.97 0.21 0.17 0.04Very Low 6.43 4.14 4.48 3.44

Renal 5.67 3.33 4.96 1.32Very High 0.63 0.05 0.25 0.02Medium 1.70 0.42 0.24 0.06Low 3.34 2.86 4.47 1.24

Substance Abuse 4.92 2.27 0.25 0.18Low 1.75 1.25 0.11 0.07Very Low 3.17 1.02 0.14 0.11

Cancer 4.55 2.79 3.55 0.33High 1.15 0.26 1.33 0.06Medium 2.20 0.78 1.47 0.17Low 1.20 1.75 0.75 0.10

Developmental Disability 3.90 0.09 9.12 0.10Medium 0.76 (1) 1.86 0.01Low 3.14 0.09 7.26 0.09


To contrast the diagnostic profile ofadults with disabilities and AFDC adults,we selected the 32 highest cost subcate-gories. Among adults with disabilities, 33percent have a diagnosis in at least one ofthese relatively high-cost categories, a ratethat is almost five times the 7 percent rate among AFDC adults (Table 5).Considering all CDPS categories (exclud-ing pregnancy), adults with disabilitiesaverage approximately twice as manyCDPS diagnoses as AFDC adults. Thehigher proportion of adults with disabili-

ties who have CDPS diagnoses is asexpected: Adults with disabilities areMedicaid beneficiaries because of a signifi-cant physical or mental condition, often achronic physical or psychiatric illness.

One result that at first seems surprisingis the significant proportion of people withdisability who appear to have no diagnosisin a CDPS category: 29 percent for adultswith disability and 35 percent for childrenwith disability (Table 5). A few of theseindividuals may be receiving SSI becauseof conditions that were judged not suffi-


Table 4—Continued

Frequency of CDPS Categories, by Beneficiary Group

Disabled AFDC Disabled AFDCAdults Adults Children Children

Category (n) = 960,760 (n) = 1,548,488 (n) = 130,324 (n) = 3,640,871

PercentGenital, Extra Low 3.59 10.39 0.93 0.66

Metabolic 3.37 1.11 8.24 1.09High 0.79 0.21 1.32 0.11Medium 0.77 0.35 0.98 0.14Very Low 1.81 0.55 5.94 0.84

Pregnancy 3.53 24.12 0.75 0.87Incomplete 2.21 17.19 0.41 0.48Complete 1.32 6.93 0.34 0.39

Eye 3.20 0.53 1.43 0.21Low 0.46 0.13 (1) (1)

Very Low 2.74 0.40 1.43 0.21

Cerebrovascular, Low 2.39 0.43 1.89 0.15

Infectious 1.18 0.41 1.11 1.64AIDS, High 0.40 0.09 0.17 0.03Infectious, High 0.11 0.03 0.14 0.02HIV, Medium 0.12 0.06 0.03 0.01Infectious, Medium 0.55 0.23 0.77 0.24Infectious, Low 0.89 0.66 0.86 1.34

Hematological 1.74 0.65 2.93 0.43Extra High 0.06 0.01 0.29 0.01Very High 0.29 0.02 1.32 0.07Medium 0.53 0.26 0.90 0.20Low 0.86 0.36 0.42 0.15

With No CDPS Diagnosis 28.6 53.3 35.1 72.41 Subcategories were combined with the subcategory or subcategories below for the purposes of the regression, because the numbers of beneficia-ries in the category were too small to allow a reliable estimate of the expenditure effect. For example, the pulmonary very-high-cost subcategory wascombined into the pulmonary high-cost category for AFDC adults and AFDC children. For both disabled children and AFDC children, all subcate-gories of diabetes were collapsed into a single category.

NOTES: CDPS is Chronic Illness and Disability Payment System. AFDC is Aid to Families with Dependent Children. HMO is health maintenanceorganization. Individuals can be counted in more than one diagnostic category. Data from Michigan, Ohio, and Tennessee are for 1991-1993;California and Georgia are for 1990-1992; Missouri, 1991-1994; Colorado, 1992-1996. (AFDC data were not available from Ohio or Missouri.)Beneficiaries are included if they have 12 months of eligibility in the base year and at least 1 month in the subsequent year. Beneficiaries wereexcluded if they had Medicare coverage, were institutionalized, enrolled in an HMO, or in a home and community-based waiver program. Frequenciesare weighted; each State gets equal weight. AIDS is acquired immunodeficiency syndrome. HIV is human immunodeficiency virus.


ciently well defined to be included inCDPS. Many of these individuals, howev-er, probably had or could have had a CDPScondition diagnosed at some point, but thediagnosis was not recorded in claims dur-ing the year in which we counted diag-noses. (Analysis of disabled adults withcontinuous eligibility of more than 24months shows 20 percent with no CDPSdiagnosis; over a 36-month period, 15 per-cent with no CDPS diagnosis.) In particu-lar, we suspect that a significant number of beneficiaries with mental retardation are not coded as such. In the section“Adjusting to Change in Diagnostic

Reporting,” Medicaid fee-for-service (FFS)data show significant underreporting ofdiagnoses, with many individuals codedwith a significant chronic illness not codedwith it again in the following year.

Disease Burden for Children

Children are, on average, much healthi-er than adults and are much less likely tohave most of the conditions that are includ-ed in CDPS (Table 4 and Figure 3). Forexample, cardiovascular problems, dia-betes, skeletal and connective diagnoses,and psychiatric problems are all relatively


Hematological, Extra High

Diabetes, Type 1, High

Infectious, High

Pulmonary, Very High

Cardiovascular, Very High

Hematological, Very High

Gastrointestinal, High

Central Nervous System, High

Infectious, AIDS, High

Skin, High

Metabolic, High

Renal, Very High

Pulmonary, High

Adults with Disability


AFDC Adults

0 0.20.1 0.3 0.4 0.5 1.00.6 0.7 0.8 0.9

CD

PS

Cat

ego

ry

NOTES: CDPS is Chronic Illness and Disability Payment System. AFDC is Aid toFamilies with Dependent Children. Claims and eligibility data from Michigan, Ohio, andTennessee, 1991-1993; California and Georgia, 1990-1992; Missouri, 1991-1994;Colorado, 1992-1996. (AFDC data were not available from Ohio or Missouri.)Beneficiaries are included if they have 12 months of eligibility in the base year and atleast 1 month in the subsequent year. Beneficiaries were excluded if they hadMedicare coverage or were institutionalized, enrolled in a health maintenance organi-zation, or enrolled in a home and community-based waiver program. Frequencies areweighted; each State gets equal weight. AIDS is acquired immunodeficiency syndrome.


Figure 2

Frequency of Selected High-Cost CDPS Categories for Adults with Disability and AFDC Adults

common among adults with disabilities butare much less common among children.There are some areas, such as pulmonaryconditions, in which the prevalence ofCDPS diagnoses is similar among childrenand adults, and a few areas, such as devel-opmental disabilities and metabolic condi-tions, in which diagnoses are more fre-quent among children than adults. Low-cost central nervous system diagnoses aremade for 25 percent of children with dis-abilities—far higher than the rate of suchdiagnoses among adults—primarily reflect-ing high rates of cerebral palsy, spina bifi-da, and epilepsy.

AFDC children have a very low inci-dence of chronic illness and disability rela-tive either to AFDC adults or to childrenwith disability. Among AFDC children, 74percent have no CDPS diagnoses, and 98percent have no relatively high-cost CDPSdiagnoses (Table 5). Children with disabil-ities have approximately 75 percent asmany diagnoses as adults with disabilities,but AFDC children have only 44 percent asmany diagnoses as AFDC adults. The rel-atively healthy diagnostic picture of AFDCchildren is consistent with the relativeaverage monthly expenditures of the

groups: $57 per eligible month for AFDCchildren, $158 per month for AFDC adults,$363 for children with disabilities, and $425for adults with disabilities.

Although children with disabilities aremuch more likely than AFDC children tohave serious diagnoses, there are so manymore AFDC children than there are dis-abled children that in absolute numbersthere are more AFDC children with a vari-ety of severe problems than there are chil-dren with disabilities. For example, in oursample, just under 0.3 percent of childrenwith disabilities have extra-high-cost hema-tological problems (hemophilia with defi-ciencies in clotting factors VIII or IX), com-pared with 0.01 percent of AFDC children.However, because there are 30 times asmany AFDC children as there are childrenwith disabilities, there are slightly moreAFDC children with hemophilia than thereare children with disabilities. Similarly,high-cost metabolic problems are diag-nosed among approximately 1.3 percent ofchildren with disabilities and 0.11 percentof AFDC children but affect more thantwice the number of AFDC children as thenumber of children with disabilities.


Table 5

Number of CDPS Categories, by Recipient Group and Type of CDPS Diagnosis

All CDPS Categories, Relatively High-Excluding Pregnancy Cost Categories Only

Average AverageNumber of Number of

Average Categories Average CategoriesNumber of Percent for Recipients Number of Percent for Recipients

Type of Categories with No with at Least Categories with No with at LeastRecipient per Recipient Categories One Category per Recipient Categories One Category

Adults with Disability 1.66 29 2.34 0.44 67 1.31AFDC Adults 0.77 54 1.68 0.08 93 1.12Children with Disability 1.27 35 1.96 0.31 75 1.28AFDC Children 0.34 74 1.30 0.02 98 1.07

NOTES: CDPS is Chronic Illness and Disability Payment System. AFDC is Aid to Families with Dependent Children. Analysis of Medicaid claims andeligibility data from seven State Medicaid programs for persons with disability; five States for AFDC recipients. For disabled adults, n = 960,760; forAFDC adults, n = 1,548,480; for children with disability, n = 130,324; for AFDC children, n = 3,640,871. Relatively high-cost categories exclude 22low-cost CDPS categories.


Expenditure Effects of DiagnosticCategories

Among beneficiaries with disabilities,diagnoses made in a given year have strongassociations with expenditures in the subse-quent year (refer to Table 6, which showsparameter estimates from prospectiveregressions, in which diagnoses from year 1are used to predict expenditures in year 2).5

A variety of relatively rare, extra-high-costand very-high-cost conditions increase sub-sequent year expenditures by $10,000 peryear or more, for example, very-high-costcardiovascular conditions (primarily trans-plants), very-high-cost pulmonary condi-tions (primarily cystic fibrosis), renal fail-ure, AIDS, and certain extra-high and veryhigh-cost hematological conditions.Diagnoses in each of these very-high-costcategories are found in fewer than 1 percentof beneficiaries with disabilities.

A variety of slightly more frequent butstill relatively rare CDPS categories haveexpenditure effects of approximately$4,000 to $9,000 per year. By far the most


Hematological

Infectious

Cerebrovascular, Low

Eye

Pregnancy

Metabolic

Genital, Extra Low

Developmental Disability

Cancer

Substance Abuse

Renal

Skin

Diabetes

Gastrointestinal

Pulmonary

Central Nervous System

Skeletal

Psychiatric

Cardiovascular

Children with Disability


AFDC Children

0 105 15 20 25 30 35

CD

PS

Cat

ego

ry

NOTES: CDPS is Chronic Illness and Disability Payment System. AFDC is Aid to Familieswith Dependent Children. Claims and eligibility data from Michigan, Ohio, and Tennessee,1991-1993; California and Georgia, 1990-1992; Missouri, 1991-1994; Colorado, 1992-1996.(AFDC data were not available from Ohio or Missouri.) Beneficiaries are included if theyhave 12 months of eligibility in the base year and at least 1 month in the subsequent year.Beneficiaries were excluded if they had Medicare coverage or were institutionalized, wereenrolled in a health maintenance organization, or enrolled in a home and community-basedwaiver program. Frequencies are weighted; each State gets equal weight.


Figure 3

Frequency of Major CDPS Categories, Children with Disability and AFDC Children

5 The regressions reported in Table 6 are based on the full sam-ple—combining the development and validation sample. Forpresentation purposes, we converted the original regressionresults, which were ratios of expenditure effects to averagemonthly expenditures, into annual dollar amounts. To do so, wemultiplied the ratios by the weighted average annualized expen-ditures: $4,980 for persons with disability, $1,884 for AFDCadults, and $684 for AFDC children.


Table 6

Subsequent-Year Annual Expenditure Effects of CDPS Categories, by Beneficiary Group

Adults andChildren with AFDC AFDC

Category Disability Adults Children

CardiovascularVery High $14,939 $7,343 $9,459Medium 4,444 2,345 2,947Low 1,799 943 890Extra Low 708 701 489

PsychiatricHigh 4,841 22,477 6,037Medium 3,770 22,477 3,322Low 1,671 1,076 1,550

SkeletalMedium 5,313 3,822 1,365Low 1,886 1,027 587Very Low 1,233 2809 369Extra Low 545 2809 225

Nervous SystemHigh 9,726 2,699 10,518Medium 3,314 1,737 3,343Low 1,582 954 654

PulmonaryVery High 13,586 (1) (1)

High 7,548 1,991 2,422Medium 5,163 2,268 1,385Low 1,852 891 496

GastrointestinalHigh 8,677 22,021 3,231Medium 3,353 22,021 1,451Low 1,506 798 304

DiabetesType 1 High 9,911 10,312 (1)

Type 1 Medium 3,787 2,863 (1)

Type 2 Medium 3,111 2,514 (1)

Type 2 Low 1,452 664 729

SkinHigh 7,049 2,523 1,698Low 2,594 1,122 787Very Low 867 407 175

RenalVery High 14,741 8,387 2,270Medium 2,536 1,465 646Low 1,183 650 472

Substance AbuseLow 2,253 1,506 2,393Very Low 1,115 821 967

CancerHigh 5,114 3,080 4,661Medium 1,727 1,153 1,199Low 431 204 766

Developmental DisabilityMedium 5,314 (1) 5,328Low 1,642 412 2,118


common of these high-cost conditions arehigh-cost psychiatric conditions (primarilyschizophrenia), which are diagnosed innearly 12 percent of beneficiaries with dis-abilities, and medium-cost cardiovascularconditions (primarily congestive heart fail-ure), which are diagnosed in 4 percent ofbeneficiaries. Many of the more common-ly occurring CDPS categories have expen-diture effects of $1,000-$2,000 per year.These expenditure effects associated with

additional diagnoses are on top of a base-line amount of $1,382 for beneficiaries withdisability. For context, the average expen-diture for beneficiaries with disability inour sample is $4,980 per year. (The regres-sion includes dummy variables for age andgender. Coefficients for these variablesare listed in the Technical Note. The base-line amount of $1,382 per year is the aver-age value of the intercept plus the age-gender effects.)


Table 6—Continued

Subsequent-Year Annual Expenditure Effects of CDPS Categories, by Beneficiary Group

Adults andChildren with AFDC AFDC

Category Disability Adults Children

GenitalExtra Low $175 $464 $559

MetabolicHigh 4,946 1,670 3,550Medium 3,156 1,079 1,019Very Low 1,089 883 582

PregnancyIncomplete 560 492 951Complete 1,114 1,903 2,231

EyeLow 2,199 1,174 (1)

Very Low 1,018 707 686

CerebrovascularLow 1,109 1,066 688

InfectiousAIDS, High 211,477 23,125 21,282Infectious, High 211,477 23,125 21,282HIV, Medium 4,200 1,159 834Infectious, Medium 4,200 1,159 834Infectious, Low 1,369 285 145

HematologicalExtra High 62,576 7,821 12,137Very High 13,874 6,634 3,350Medium 3,972 1,047 854Low 1,967 982 500

Baseline 1,382 944 4291 Subcategories were combined with the subcategory or subcategories below for the purposes of the regression because the numbers of beneficia-ries in the category were too small to allow a reliable estimate of the expenditure effect.2 The coefficient of this subcategory was constrained to be equal to the coefficient of the subcategory above it because the unconstrained coefficientswere not different enough to justify the estimation of separate coefficients.

NOTES: CDPS is Chronic Illness and Disability Payment System. AFDC is Aid to Families with Dependent Children. AIDS is acquired immunodefi-ciency syndrome. HIV is human immunodeficiency virus. Refer to the Technical Note for an explanation of regression details. Claims and eligibilitydata from Michigan, Ohio, and Tennessee, 1991-1993; California and Georgia, 1990-1991; Missouri, 1991-1994; Colorado, 1992-1996. (AFDC datawere not available from Ohio or Missouri.) Beneficiaries are included if they have 12 months of eligibility in the base year and at least 1 month in thesubsequent year. Beneficiaries were excluded if they had Medicare coverage, were institutionalized, enrolled in a health maintenance organization,or enrolled in a home and community-based waiver program. Regressions are weighted; each State gets equal weight.


The expenditure effects of CDPS diag-noses are much smaller among AFDCadults than among adults with disabilities.For most CDPS categories, the diagnosesthat are expensive among beneficiarieswith disability also tend to be expensive forAFDC beneficiaries but with much lowerlevels of added expense. For example,medium-cost cardiovascular problems add$4,444 for beneficiaries with disability but$2,345 for AFDC adults. We suspect thatthe effects of CDPS diagnoses are muchlarger for persons with disabilities becauseany given chronic illness or disability ismore severe or more advanced, on aver-age, for persons with disabilities than forAFDC beneficiaries. In addition, the aver-age poorer health or functional status ofadults with disabilities may increase theirneed for medical services during an ill-ness. Comparing expenditure effects forAFDC adults and children, the picture ismixed, with only a small majority of cate-gories showing higher expenditures foradults. For example, the skeletal, renal,and infectious disease categories showhigher expenditures for adults, while thepsychiatric and cancer categories showhigher expenditure effects for children.

Maternity and Mental Health Services

An important element of health care forthe AFDC population is for pregnancy, deliv-ery, and neonates. There is little advantage,however, in using prospective health-basedpayment to cover maternity and neonatalcare. Most of the costs associated withpregnancy, delivery, and neonates areincurred within the year. Thus, when vari-ables for pregnancy, delivery, or prematurebirth are used in prospective regressions,they are associated with very small addi-tional expenditures in the following year, farless than the costs associated with even aroutine delivery. Use of prospective

weights for maternity and neonatal carewould cause States to significantly underpayplans with a disproportionately high shareof beneficiaries who give birth. Some Statesalready use supplemental payments on topof capitation to cover the cost of each deliv-ery, and we recommend that this approachbe continued under health-based payment.Alternatively, weights for maternity andneonatal care that we have calculated onconcurrent regressions could be substitut-ed in the prospective model. Similarly, thehigh costs of neonatal care might be paid foron a supplementary basis.

Many States do not make mental healthservices the responsibility of HMOs and carveexpenditures for such services out of thecapitation. The main effect of re-estimatingthe regression to reflect this policy is tosharply reduce the coefficients for the psy-chiatric and substance abuse categorieswith little change in coefficients in othercategories.

Predicted Expenditures UsingMultiple Diagnoses

The predicted total expenditure for anindividual is the sum of the baseline pay-ment and the additional expenditureamounts for all of the diagnostic subcate-gories in which an individual is counted.(The total would also include the expendi-tures associated with an age-sex variableand any interaction between age and diag-nostic categories.) Because many individ-uals with serious diagnoses are counted inmultiple categories, the average predictedexpenditures for individuals in most CDPScategories are substantially higher thanthe parameter estimates shown in Table 6.For example, the average predicted pay-ment for beneficiaries in the medium-costcardiovascular CDPS category is $13,044,although the additional expenditure associ-ated with this category is $4,444; similarly,


the average predicted expenditure for ben-eficiaries with medium-cost Type 1 dia-betes is $12,084, while the regression coef-ficient is only $3,787.

We have presented a model thatassumes that the effects of diagnoses indifferent major CDPS categories are addi-tive. We assume, for example, that theeffects of having cardiovascular and cen-tral nervous system problems simultane-ously are equal to the sum of the effects ofhaving each of these problems individually.But this assumption of additivity may beincorrect: In some cases, having two ormore problems might be more expensivethan the sum of the independent effects,while in other cases, having two or moreproblems might be less costly than thesum of the independent effects.

As a rough test of the additivity assump-tion, we examined the actual and predictedexpenditures for beneficiaries with disabil-ity categorized by the number of CDPSsubcategories in which they have a diagno-sis (Figure 4). On average, the additivityassumption appears reasonable: as thenumber of CDPS categories increases,both the actual and predicted expendituresincrease. For the few beneficiaries withfive to nine CDPS categories, actual expen-ditures are slightly higher than predicted,suggesting that the effects of having manydifferent diagnoses are slightly greaterthan the sum of the individual effects.Preliminary work with a variety of interac-tion terms of CDPS subcategories yieldedonly slight improvements in overall modelperformance.

Goodness of Fit

Diagnostic information does a better jobof predicting expenditures among personswith disabilities than among AFDC benefi-ciaries. As we have shown previously(Kronick et al., 1996), expenditures are

concentrated among a small number ofhigh-cost beneficiaries among personswith disability, just as they are for AFDCbeneficiaries: Among both AFDC benefi-ciaries and persons with disability, themost expensive 20 percent of beneficiariesaccount for 80 percent of expenditures.The difference between the two popula-tions is the extent to which diagnosticinformation helps identify those who willbe expensive. Diagnostic information doesa better job for persons with disabilitiesbecause both the prevalence and the sever-ity of diagnoses are much greater thanamong AFDC beneficiaries.

The statistical summary of this discussionis that the R 2 statistics are substantially high-er for regressions on persons with disabilitythan for AFDC beneficiaries. Estimated onour validation sample (the 25 percent of thedata that was reserved from use while wewere developing CDPS), the R 2 for the dis-abled is 0.18, compared with 0.08 for AFDCadults and 0.04 for AFDC children (Table 7).The difference in R 2 between AFDC adultsand children results primarily from thelower prevalence of chronic illness amongchildren and secondarily from smallereffects of diagnoses on expenditures for avariety of CDPS categories.

The R 2 for the validation sample is simi-lar to the R 2 for the entire sample, sug-gesting that we did not significantly overfitthe data.

State-Specific Versus MultistateWeights

A question of concern to a number ofStates is whether they should estimate pay-ment weights on their own data or use pay-ment weights already estimated on datafrom a number of States. To use the multi-state weights in Table 4, a State would needto adjust them to reflect average expendi-tures in the State in order to ensure budget


neutrality. The decision about State-specificor multistate weights depends in part uponwhether patterns of care in a given Stateare thought to be notably different fromnational patterns and in part on whetherthere are enough beneficiaries in the Stateto estimate payment weights reliably.

To explore this question, we use the 75-percent development sample to estimate sep-arate sets of State payment weights usingdata from individual States, and sets of multi-state weights that exclude each State in turn.We then used both the multistate weightsand the State-specific weights to predictexpenditures for beneficiaries from the vali-dation sample grouped into simulated healthplans (refer to the section “Comparison withOther Payment Systems” for a description of

the plans). In four of the seven States, themultistate weights led to better predictionsthan the State-specific weights; in one State,the predictions were very similar, and in twoStates, the State-specific weights were better.Only in California did the State-specificweights show a clear superiority, probablybecause of the very large number of benefi-ciaries and perhaps in part because of thesmall proportion of diagnoses for which thedetailed fifth digit of the ICD-9-CM code wasretained in California data. It is likely that allbut the largest States would do better using payment weights estimated from multistate data.

A Medicaid program need not view pay-ment weights as set in stone. Largechanges in treatment costs due to changes


Predicted Expenditures

Number of CDPS Categories

Do

llars

per

Yea

r

Actual Expenditures

0 41

$35,000

30,000

25,000

20,000

15,000

10,000

5,000

0

5 7 8 10 orMore

2 3 6 9

NOTES: CDPS is Chronic Illness and Disability Payment System. Claims and eligibility datafrom Michigan, Ohio, and Tennessee, 1991-1993; California and Georgia, 1990-1992;Missouri, 1991-1994; Colorado, 1992-1996. (AFDC data were not available from Ohio orMissouri.) Beneficiaries are included if they have 12 months of eligibility in the base yearand at least 1 month in the subsequent year. Beneficiaries were excluded if they hadMedicare coverage or were institutionalized, enrolled in a health maintenance organization,or enrolled in a home and community-based waiver program. Frequencies are weighted;each State gets equal weight. Predicted expenditures come from the regression in Table 6.


Figure 4

Actual and Predicted Expenditures for Beneficiaries with Disability, by Number of CDPS Categories

in drugs and other technology might justi-fy ad hoc modification of individual catego-ry weights.

COMPARISON WITH OTHER PAYMENT SYSTEMS

States and other payers implementinghealth-based payment could consider avariety of diagnostic classification systemsthat have been created for various paymentand profiling purposes, but only threeapproaches are now publicly available thatwere designed for adjustment of capitatedpayments to health plans.6 The twoapproaches other than CDPS are theDiagnostic Cost Groups (DCGs) and theAdjusted Clinical Groups (ACGs). (In pre-vious versions of the ACG models, ACGstood for Ambulatory Care Groups.) Thissection describes how CDPS differs fromthese other systems in method and thenext section presents differences in predic-tive performance based on regressions weran using the various models on ourMedicaid data.

The new CDPS and recent versions ofthe DCG and ACG models are similar inimportant respects, and all three modelscan be used to adjust payments to health

plans with diagnostic data far more effec-tively than traditional risk adjustmentthrough demographic data alone. Theadditive DCG model, known as theHierarchical Condition Category (HCC)model, and CDPS are similar. The ACGmodel is substantially different from CDPSand the HCCs both in how it classifies diag-noses and in its inclusion of many ill-defined and high-frequency, low-cost diag-noses. As we show later, for persons withdisability, CDPS performs better than theHCCs, which in turn perform better thanthe ACGs. For AFDC beneficiaries, theperformance of the ACGs and CDPS is fair-ly similar, with HCCs performing some-what less well.

We believe that CDPS is preferable tothe HCCs both because CDPS performsbetter for people with disabilities and forTANF beneficiaries and because it orga-nizes and distinguishes among diagnosesin a number of cases more appropriately.In addition, Medicaid payment weightsestimated from a variety of State Medicaidprograms are now available for the CDPScategories and not yet for the HCCs.7 Webelieve that both CDPS and the HCCmodel are preferable to the ACGs becausethe ACGs include many ill-defined andhigh-frequency, low-cost diagnoses that wethink should not be used in a payment sys-tem and because its performance for peo-ple with disabilities is less good. In thissection, we examine in detail some of thedifferences between HCCs and CDPS inorder to stimulate further discussion aboutthe variety of decisions involved in con-structing health-based payment models.We then look more briefly at the ACGmodel.


Table 7

R2 Statistics for CDPS, by Beneficiary Group

Persons with AFDC AFDCGroup Disability Adults Children

DevelopmentSample 0.191 0.087 0.032

ValidationSample 0.183 0.083 0.041

NOTES: CDPS is Chronic Illness and Disability Payment System.AFDC is Aid to Families with Dependent Children. The model develop-ment sample used 75 percent of the data. The validation sample used25 percent of the data.


6 A fourth, the Clinical Risk Groups, will soon be available from3M Health Information Systems. A fifth, the Clinically DetailedRisk Information System for Cost (CD-RISC), was developed byGrace Carter et al., (2000).

7 The Medicaid payment weights for HCCs provided in Ash et al.(1998) were estimated on data from one State. These paymentweights are estimates of the expenditure effects of diagnoses ona combined population of AFDC and SSI beneficiaries. Weargued previously that it is important to have separate paymentweights for the two population groups.

Comparison of CDPS with the HCCs

The HCCs are one of several DCG mod-els. The HCC model organizes ICD-9-CMcodes into 543 diagnosis definitions(“DxGROUPS”). Many diagnoses aredefined simply by a three-digit ICD-9-CMcode and all its subcodes, other diagnosesare defined by multiple three-digit codes,and still others by individual four-digitcodes or combinations of three-, four-, andfive-digit codes. Despite many differencesin detail, the HCC diagnosis definitions andthe CDPS single-diagnosis variables, suchas heart failure, schizophrenia, or hemo-globin-S disease, are parallel concepts.

The HCCs then cluster their DxGROUPsinto 118 “condition categories,” which cor-respond to the CDPS diagnostic subcate-gories. Despite differences in nomencla-ture and content, the HCC condition cate-gories and CDPS diagnostic categories arefairly similar. For example, the HCC has acategory for quadriplegia, which is verysimilar to the CDPS category central ner-vous system high-cost. Both the CDPSand the HCC categories function asdummy or zero-one variables. If an indi-vidual’s record contains a diagnosis code inone of the defined diagnoses in the catego-ry, the model initially sets the category toone for that individual; otherwise, the cate-gory is set to zero.

Both models count multiple diagnosesacross categories that are different fromeach other, for example, a cardiovasculardiagnosis and a psychiatric diagnosis.Thus, both models share the assumptionthat the cost effects of multiple differenttypes of diagnoses should be added togeth-er in order to produce an accurate predic-tion of total expenditures. (By contrast,the Principal Inpatient DCG model pre-dicts expenditures using only the singlemost serious diagnosis made in the inpa-tient setting.8)

Both models also limit the counting theydo within body systems or types of disease.Like CDPS, which counts only the mostserious subcategory in a major category,the HCC model uses single counting inmany areas. But in a few diagnostic areas,for example metabolic disorders, HCCsimpose no counting rules, so that a differ-ent diagnosis in each of several conditioncategories can be counted. For severalareas, the HCC model uses special count-ing rules that compromise between singlecounting and unrestricted counting. Forthe heart disorders, up to five differentcondition categories can each be countedseparately. With separate HCC categoriesfor vascular disease, a maximum of sevendiagnoses could be counted, where CDPSwould count only the single most seriouscardiovascular subcategory. We consid-ered multiple counting of this kind forCDPS in some areas but rejected itbecause it rewards increased coding whilegiving very little benefit in improved pre-dictive power.

We think that one advantage of CDPS isits more precise identification and catego-rization of a number of medium- and high-cost diagnoses. Distinguishing a moreexpensive diagnosis from other related butless expensive diagnoses can help makebetter cost predictions and get resourcesto plans that attract people with greaterneeds. For example, the HCCs combine allof the diagnoses of hereditary hemolyticanemias, including hemoglobin-S disease,which we found to have far higher cost


8 The Principal Inpatient DCG model has been selected byHCFA for its current implementation of health-based paymentfor Medicare’s managed care reimbursement. Although health-based payment with only inpatient data sharply reduces theamount of data required for implementation, almost all analysts,including the creators of the DCG models, support a promptmovement by Medicare to adjustment by diagnoses from bothinpatient and ambulatory settings. Using both sources of diag-nostic data brings much greater accuracy and, more important-ly, avoids creating inappropriate incentives to hospitalize andinappropriate penalties for plans that have successfully reducedhospital use (Iezzoni and Ayanian, 1998). We strongly concur(Dreyfus and Kronick, 1999).

implications than the other hereditaryhemolytic anemias. In CDPS, hemoglobin-Sdisease is put in our very-high-cost hema-tological category, which is associated for4,546 disabled beneficiaries with $13,874per year in additional cost. In the HCCmodel, the hereditary hemolytic anemiasare grouped with other blood and immunedisorders, which together are associated inour sample with only $6,504 of additionalcost. (The estimates cited in this sectionfor HCCs come from a regression we haveestimated using HCCs. Further informa-tion on the regression is provided later.)

Among pulmonary conditions, we foundcystic fibrosis associated with unusuallyhigh costs for the disabled and placed it inthe subcategory of very-high-cost pul-monary diagnoses, which predicts $13,586per year. For adults, the HCCs havegrouped cystic fibrosis with other fibrosesand other chronic lung disorders, a catego-ry that has an annual coefficient of only$2,844 when we ran the HCC model on ourMedicaid disabled sample. For children,the HCC model does recognize the highcosts of cystic fibrosis, putting it with otherdiagnoses in a separate category for veryhigh-cost pediatric disorders, but it is asso-ciated with only $4,440 per year. Otherexamples of conditions we categorized sep-arately from neighboring codes, where theHCCs do not, include decubitus ulcer, bac-terial endocarditis, primary pulmonaryhypertension, acute cor pulmonale, andcertain coagulation disorders (congenitalfactor VIII and factor IX disorders). Ourlarger data set with so many disabledMedicaid beneficiaries may have allowedus to pick up notable differences in costeffects that were missed in the develop-ment of the HCCs.

In addition, we see one area in which theCDPS organization of diagnoses seemsmuch more appropriate than that of theHCCs. The HCC model seems to have

erred in its counting for psychiatric andsubstance abuse diagnoses, which aregrouped together in the area of mental dis-orders: If the single category for drug andalcohol psychosis or dependence is count-ed, then none of the psychiatric disordersare counted. Both our data and the HCCregression indicate that schizophrenia ismuch more costly than alcohol depen-dence among Medicaid beneficiaries. But,as a result of the HCC hierarchy, an indi-vidual with both alcohol dependence andschizophrenia would have less money paidfor him or her than an individual with onlyschizophrenia. (Also, a plan that enrollssomeone with schizophrenia would receivethe same money as a plan enrolling some-one with bipolar, affective, and otherdepressive disorders, which our analysissuggests would not be appropriate.)

CDPS may also have a modest advan-tage over the HCC model in retainingsome helpful low-cost diagnoses. For theHCC payment model, more than 30 HCCcategories were excluded from the pay-ment model because of vagueness orbecause of suspected or observed lack ofsignificant cost effects (Ash et al., 1998).But the elimination of vague or no-costdiagnoses was made using entire cate-gories, and some useful diagnoses werelost from the HCCs. For CDPS, we evalu-ated diagnoses one at a time, based on themagnitude of their cost effects and the clar-ity of their definition. For example, theHCC Medicaid payment model excludescentral nervous system infection and otherinfectious disease, losing diagnoses ofpolio, bacterial meningitis, almost all tuber-culosis, herpes zoster, herpes simplex, andoral thrush—diagnoses that we included invarious low-cost categories. Other condi-tions lost to the HCCs that we thought wor-thy of use include heart conduction disor-ders, cellulitis, adrenal gland disorders,and kidney infections and stones.


Comparison of CDPS with ACGs

The ACG models use a differentapproach from CDPS and the HCCs to pre-dict expenditures from diagnostic data.Like CDPS and the HCCs, the ACG modelsassign diagnosis codes to various cate-gories, which can be used to predict expen-ditures for individuals. But the ACG mod-els differ sharply from the other two in themethod of categorizing diagnoses. We seetwo problems in the ACG approach: poorseparation of high-cost diagnoses fromother conditions and inappropriate use ofall ICD-9-CM codes, including many ill-defined and extremely low-cost diagnoses.

Nonetheless, ACG-based models haveseen considerable use and have helpedbring acceptance of case-mix analysis tohealth plans and payers. ACG models havebeen used by the State of Maryland andthe Minnesota Buyers’ Health Care ActionGroup for payment purposes and exten-sively by health plans for evaluating casemix. Maryland’s use of an ACG modelmade it the first State to set diagnosticallyadjusted rates for large numbers ofMedicaid beneficiaries.

The ACG approach assigns ICD-9-CMcodes to 32 diagnostic categories calledAmbulatory Diagnostic Groups (ADGs)based primarily on expectations about thecondition’s effect on individual health andresource needs (Johns Hopkins University,1998). Likelihood of persistence, disability,reduced life expectancy, and need for diag-nostic, specialist, therapeutic, and hospitalcare are considered to assign ICD-9-CMcodes to the ADGs. As a result, many ofthe ADGs include conditions that appearunrelated but are thought to have similareffects on future resource use. For exam-ple, cerebral thrombosis and acute pancre-atitis are both in the group for progressiveconditions that are likely to recur.

Other groups are used for discrete con-ditions that are likely to recur, unstablechronic medical conditions, and time-limited minor psychosocial conditions. Afew of the ADGs are based on more specif-ic types of diagnoses, for example, separategroups for asthma, dermatologic condi-tions, malignancy, and pregnancy. Somegroups are defined with combinations ofresource expectation and type of condition,such as stable orthopedic conditions,unstable eye conditions, or unstable recur-rent or persistent psychosocial conditions.

These diagnostic categories can be usedin two ways for predicting expenditures.One approach is to use the ADGs much asone would the diagnostic categories of theCDPS or HCC models, with each ADGoperating as a dummy or zero-one variableset to one if a diagnosis in the ADG isfound in an individual’s record. Byregressing individual expenditure dataagainst the counts of the ADGs for eachindividual, one can produce estimates ofthe additional future costs associated withthe diagnoses in each of the ADGs.

Another approach is the standard recom-mended ACG model, which uses an algo-rithm to convert the dummy variables of theADGs into more than 80 mutually exclusivecategories (from 82 to 93 categories,depending upon some choices by the modeluser). These mutually exclusive categories,or ACGs, are defined through combinationsof different classes of the ADGs and age vari-ables. Individuals of the same age with diag-noses in multiple ADGs are classified intodifferent ACGs depending upon how manyof their ADGs are considered major. Forexample, for adults, the ADG for progres-sive, likely-to-recur conditions is consideredmajor, while the ADG for allergies is not.

An advantage of the mutually exclusiveACGs is that they can be used in a tradi-tional ratesetting fashion, with average


costs for past members of each grouptrended forward to predict the future costsof people assigned to the group. The ADGregression model works better in predict-ing expenditures than the standard ACGmodel, but like CDPS and the HCCs, itrequires estimating a regression and set-ting up a system that can calculate an aver-age casemix for each health plan.

Both the ADG and ACG models have thedisadvantage of putting conditions withvery dissimilar cost implications togetherinto single groups. For example, the unsta-ble chronic medical conditions mix togeth-er in one category cystic fibrosis, multiplesclerosis, unspecified cardiac dysrhythmia,unspecified heart disease, ulcer of lowerlimbs, and degeneration of intervertebraldisc. Our analysis shows that these condi-tions have vastly different implications forfuture cost.

A more important problem is that theACG models use all the ICD-9-CM diag-nostic codes, many of which are ill-definedor so common that they could be elicitedfrom almost any patient. A straightforwardapplication of the ACG model would pro-vide substantially more resources for indi-viduals who have a medical visit for eventhe most minor problems. For example,for a disabled adult with a sore throat orheadache, the ACG model run on our sam-ple would add $348 in annual payment andfor a TANF adult, $228 in annual payment.For another example, the diagnosis of abackache would add $636 for a disabledadult and $276 for a TANF adult. Similarly,a swollen ankle would add $1,128 for thedisabled adult and $456 for the TANFadult. The combination of a backache anda swollen ankle would add $1,356 for a dis-abled adult and $636 for the TANF adult.

These additional amounts seem far in excess of what plans should receive for such relatively minor problems.Meanwhile, the increased payment for a

disabled adult or child with a diagnosis ofsickle-cell anemia or cystic fibrosis wouldbe only $3,828. The inclusion of the high-frequency, low-cost conditions helps themodel in simulated predictive accuracy butrewards plans for increased coding ofminor problems and does relatively little todirect resources to plans that enroll a needier-than-average membership. Theproblem of modest additional payments forhigh-cost diagnoses can be attenuated ifthe ACGs are implemented with separatepayments for specified high-cost diag-noses, as was done in Maryland and as isrecommended by the developers of ACGs(Weiner, 1998).

Predictive accuracy is improved by usinghigh-frequency, low-severity conditions,because past users of even small amountsof service are more likely than non-users toseek services in the future. AlthoughMedicaid programs might well want toencourage contact between physicians andbeneficiaries, attempting to accomplish thisthrough health-based payment seemsunwise. Using so many very-low-cost diag-noses in setting rates also requires a hugevolume of diagnoses and exacerbates prob-lems of auditing and of favoring plans thatreport better than others.

Comparison of StatisticalPerformance

We present comparisons of statisticalperformance for six models: our CDPSmodel, an ADG model, an ACG model, anHCC model, our original DPS model, and aCDPS “catch-all” model, in which weinclude variables for a set of extremely low-cost and ill-defined diagnoses excludedfrom our recommended payment model.

For each model we estimated threeregressions: one for beneficiaries with dis-abilities (adults and children combined),one for AFDC adults, and one for AFDC


children. We estimated these regressionson the validation sample—the 25 percentof the data set aside before CDPS modeldevelopment. For the HCC model, we usethe subset of 85 HCCs that Ash et al.(1998) recommend for the Medicaidprospective payment model.9 In the ADGmodel, we included a number of ADGs thathave statistically significant, negative para-meter estimates and that would likely beexcluded if an ADG payment model wereimplemented (it would be awkward toreduce plan payments because of addition-al diagnoses). Keeping these ADGs in themodel, however, improves statistical per-formance, and we were eager not to disad-vantage any of the models.10

For persons with disability, the CDPSregression has a significantly higher R2

than HCCs, 0.183 compared with 0.143, andalmost twice the predictive power of ACGs,0.183 compared with 0.098 (Table 8). Forpersons with disabilities, the revised CDPSand the original DPS have similar explana-tory power, and the catch-all CDPS modeldoes not increase the R2 statistic by muchcompared with the recommended CDPSmodel. Among AFDC adults, the R2 statis-tics from the various models are more sim-ilar than for persons with disability, but theCDPS R2 is slightly higher than the HCCstatistic, and higher than the ACG andADG R2 statistics as well. The revisions inCDPS to make it more sensitive for anAFDC population are evident in theimprovement in R2 from the original DPSmodel. Predictive performance can be

improved measurably by inclusion of avariety of ill-defined and high-frequency,low-severity diagnoses, as indicated in thecatch-all model. Overall R2 statistics aremuch lower among AFDC children thanfor AFDC adults, and the models show sim-ilar explanatory power.

A more important issue than R2 iswhether the models get the right amountof money to plans with biased enrollment.Does it matter which system is used, or dothe leading classification systems givemuch the same results when used for pay-ment? To explore this question, we simu-lated a universe of health plans by assum-ing that beneficiaries with expensive diag-noses will sort themselves disproportion-ately into different plans. To avoid favoringany one system, we identified high-costdiagnoses using three-digit ICD-9-CMcode groups as regression variables ratherthan the diagnoses in any one system’shigh-cost categories. We selected as high-cost diagnoses the 100 highest cost ICD-9-CM codes with more than 100 disabledbeneficiaries in the validation sample. Wealso identified the 100 most common ICD-9-CM codes with very moderate costeffects (between $30 and $100 per month).

We then constructed five hypotheticalhealth plans by assigning disproportionateshares of beneficiaries with the high-cost


9 In each model we include the same set of age-sex dummy vari-ables. When estimating HCCs for persons with disability, weinclude the set of child interactions suggested by Ash et al.,(1998). When estimating for AFDC adults and children sepa-rately, no interactions of age and HCC categories are needed.The HCC model estimated with our data set produces a betterresult in predicting expenditures than would the published HCCmodel parameter estimates.10 In the ADG regression for persons with disability, we includeinteractions of each ADG with age. When estimating the origi-nal DPS, we do not include interactions with age. When esti-mating the CDPS catch-all model, we include the same set of ageinteractions as is included for CDPS.

Table 8

R 2 from Alternative Models, by BeneficiaryGroup

Persons with AFDC AFDCModel Disability Adults Children

CDPS 0.183 0.083 0.041HCC 0.143 0.080 0.031ACG 0.098 0.069 0.031ADG 0.111 0.077 0.042DPS 0.175 0.065 0.037CDPS,

Catch-All 0.188 0.093 0.047

NOTES: AFDC is Aid to Families with Dependent Children. CDPS isChronic Illness and Disability Payment System. HCC is HierarchicalCondition Category. ACG is Adjusted Clinical Group. ADG isAmbulatory Diagnostic Group. DPS is Disability Payment System. Allregressions estimated on the 25-percent validation sample.


diagnoses and with the moderate cost diag-noses to the five plans. (The use of thevery moderate-cost diagnoses was intend-ed to favor systems such as the ACGs andthe CDPS catch-all model, which should doa good job of predicting costs for plans thatattract disproportionate shares of peoplewith light diagnoses.) For beneficiarieswith disability, the five simulated plansinclude one plan with actual expendituresthat are almost 1.6 times the averageexpenditures, one plan with expenditures1.3 times average, one plan very close tothe average, one plan slightly below aver-age, and one plan with very healthyenrollees (Table 9). Each of these simulat-ed plans has about 50,000 members. (ForAFDC adults, each of the 5 plans hasapproximately 80,000 members; for AFDCchildren, 180,000.)

For beneficiaries with disability, theCDPS model predicts expenditures thatare very close to the actual expenditures ineach simulated plan: It slightly underpre-dicts actual expenditures for the plan withsick enrollees and slightly overpredictsexpenditures for the health plan withhealthy enrollees. But the differencesbetween actual and predicted expendituresare relatively small. The HCC model isclose to the CDPS model in performancebut does less well for the two extremeplans. The ACG model would not be near-ly so satisfactory: It would underpay theplan with sick enrollees by 13 percent andoverpay the plan with healthy enrollees by19 percent. On average, the absolute valueof the payment error for CDPS is 3 per-cent, for the HCCs 5 percent, for ACGs 10percent. The ADGs would perform some-what better than the ACGs but not so wellas CDPS or the HCCs.

For AFDC adults, there is much less dif-ference among models in performance.CDPS would do a good job of paying plansappropriately, with payments within 3 per-

cent of actual expenditures for three of theplans, and a 5-percent overpayment for oneplan. HCCs perform similarly to CDPS butslightly worse for two of the five plans.ADGs perform very well for this set ofplans; ACGs are slightly closer than CDPSto actual for plan 2, but slightly furtheraway on plan 4 and the high-cost plan 1.The CDPS catch-all model has predictiveexpenditures very close to actual, indicat-ing that it is possible to predict very accu-rately, if one is unconcerned about the useof ill-defined and low-severity diagnoses.

Performance comparisons for AFDCchildren are similar to AFDC adults. CDPSdoes reasonably well, though ADGs, likethe CDPS catch-all model, get consistentlycloser to actual. HCCs appear to have aproblem for AFDC children, with signifi-cant underpayments for the plan with sickenrollees and significant overpayments forthe plan with healthy enrollees.

It appears from this work that there issome difference among the models in theextent to which they will pay appropriately,although the differences are not over-whelming. For beneficiaries with disability,CDPS appears to perform significantly bet-ter than the ACG-ADG models and some-what better than the HCC model. For AFDCchildren and adults, CDPS appears to per-form somewhat better than HCCs and hassimilar performance to ACGs. For bothAFDC children and adults, both the ADGsand the CDPS catch-all model perform bet-ter than the other models, but the inclusionof all diagnoses in these models makes themless appropriate for payment purposes.

The performance comparisons are notconclusive, but they support the proposi-tions that CDPS:• Can be used to make equitable payments

for both SSI and TANF beneficiaries.• Performs better than the similar HCC

model.• Is significantly better than the ACG


model for beneficiaries with disabilityand is similar to the ACGs for AFDC ben-eficiaries.

Given the much greater vulnerability of theACG model to proliferative coding and theavailability of CDPS payment weights esti-mated on multistate data, we see a clearoverall advantage to Medicaid programs inthe implementation of CDPS. In addition,CDPS is public-use software, and mostleading actuaries that work with Medicaidprograms have experience in the use of theclosely related DPS model and software.

ADJUSTING TO CHANGE IN DIAGNOSTIC REPORTING

States face a number of important opera-tional issues in implementing health-basedpayment: the timing of risk assessmentand payment adjustment, the use ofprospective or concurrent weights, evalu-

ating risk for newly eligible and dually eli-gible persons, and limiting profits and loss-es through risk-sharing or stoploss. Wehave addressed these issues in detail else-where (Kronick et al., 1996, Kronick andDreyfus, 1997). In this section, we brieflyaddress another implementation issue thatalso deserves attention.

One key challenge in implementinghealth-based payment is obtaining datathat can be used to equitably measurehealth status across plans (Dreyfus andKronick, 1999). We can aspire eventuallyto have data that accurately reflectenrollees’ diagnoses, but we know that, inthe short run, the data will be incomplete.

FFS data show significant underreport-ing of diagnoses. For many chronic diag-noses, between 20 and 60 percent of bene-ficiaries with these diagnoses appearing ona FFS claim in a given year do not have thatdiagnosis appear on a claim during the


Table 9

Predicted Expenditures for Simulated Health Plans, by Model and Beneficiary Group

ActualExpenditures CDPS

Group and Plan per Month CDPS HCCs ACGs ADGs DPS Catch-All

Beneficiaries with Disability1 1.58 1.52 1.49 1.38 1.43 1.51 1.522 1.31 1.28 1.27 1.21 1.24 1.27 1.283 0.96 0.97 0.97 0.97 0.97 0.97 0.964 0.82 0.83 0.83 0.91 0.89 0.83 0.845 0.59 0.63 0.66 0.70 0.66 0.65 0.62

Average Percentage Error — 0.03 0.05 0.10 0.07 0.04 0.03

AFDC Adults1 1.25 1.21 1.21 1.18 1.21 1.18 1.222 1.17 1.12 1.11 1.14 1.15 1.07 1.143 0.99 1.00 1.00 1.00 1.00 1.00 1.004 0.86 0.89 0.89 0.90 0.88 0.91 0.885 0.81 0.85 0.86 0.85 0.83 0.89 0.83


AFDC Children1 1.36 1.29 1.20 1.22 1.29 1.27 1.312 1.26 1.22 1.15 1.20 1.25 1.20 1.243 1.01 1.00 1.00 1.00 1.00 1.00 1.004 0.89 0.90 0.93 0.93 0.90 0.91 0.905 0.82 0.87 0.91 0.88 0.85 0.87 0.85


NOTES: CDPS is Chronic Illness and Disability Payment System. HCC is Hierarchical Condition Category. ACG is Adjusted Clinical Group. ADG isAmbulatory Diagnostic Group. DPS is Disability Payment System. AFDC is Aid to Families with Dependent Children.


subsequent 12-month period (Figure 5).Examining beneficiaries with disabilitieswho were continuously eligible forMedicaid for a 24-month period, we findthat 20 percent of beneficiaries with schiz-ophrenia on at least one claim during thefirst 12 months do not have a diagnosis ofschizophrenia during the second 12months. And schizophrenia is the mostpersistent of the major diagnoses we haveanalyzed. The Medicare PaymentAdvisory Commission has found an evengreater lack of persistence in Medicare fee-for-service data (Medicare PaymentAdvisory Commission, 1998).

As Medicaid, Medicare, and other pay-ers begin to base payment on diagnoses,we expect that this lack of persistence willchange. Plans being paid based on diag-

noses will certainly want to increase thenumber of diagnoses they report. One canimagine the enterprising health care con-sultant selling software that will maximizepersistence: Before each scheduled visit,the software might search through lastyear’s encounter records to identify thediagnoses that could lead to additional pay-ment and then print a customized medicalservice record to prompt the physician todetermine if any of these diagnoses are stillpresent and could reasonably be construedas contributing to the need for care.

There are several potential responses tothe problem of lack of persistence anderror in measuring diagnoses. A payermight simply assume that, at least in theshort run, plans will underreport more orless equally. In this case, payment will be


Percent

Dia

gn

osi

s

0 4010

Schizophrenia

Diabetes

Multiple Sclerosis

Quadriplegia

Ischemic Heart Disease

50 70 80 9020 30 60

NOTES: Figures are the percent of Medicaid beneficiaries with disabilities with the spec-ified diagnosis in year 1 who have the diagnosis appear on at least one claim in year 2.Beneficiaries are included in the analysis if they were continuously eligible in year 1 andyear 2. Data from California, 1990-1991; Colorado, 1992-1995; Georgia, 1990-1991;Michigan, 1991-1992; Missouri, 1991-1993; Ohio, 1991-1992; Tennessee, 1991-1992.


Figure 5

Persistence of Diagnosis from Year 1 to Year 2

equitable, even though based on lightreporting of diagnoses. Changes in codingpractices will only come slowly in physicianoffices, especially if Medicaid is the onlypayer using ambulatory diagnoses. On theother hand, it is reasonable for payers andplans to be concerned that coding practicesmight be different in plans in which physi-cians are on salary or receiving subcapita-tion. In such environments, more activestrategies by the payer may be needed.

One strategy is for payers to monitorcoding practices at competing plans bycomparing records for individuals overtime. If the enrollees at one plan seem tobe getting sicker significantly more quick-ly than enrollees at other plans, then it islikely that coding is changing. Similarly, ifenrollees appear to get sicker at a faster-than-expected rate as they switch from fee-for-service into a particular health plan,then closer scrutiny of health-plan codingis needed. If some plans are clearly codingdifferently than others, then payment canbe adjusted to offset these coding differ-ences. Colorado uses such a process—a“data-reporting adjustment”—that wasoriginally implemented to correct forunderreporting but that works just as wellto address increased intensity of diagnosticreporting (Tollen and Rothman, 1998). Inaddition, the auditing of diagnoses that isneeded to keep health-based payment hon-est will also help keep reporting rates moreconsistent.

Finally, States could explore usinglonger periods for accumulating diagnosticinformation. Once a diagnosis of quadri-plegia is made, for example, there is littleneed to have this diagnosis confirmed onan annual basis. We are beginning work onanalyzing models with multiyear periodsfor accumulating diagnoses.

Although health-based payment can bestarted using diagnoses from FFS claims,its ongoing use depends upon health plans

reporting their members’ diagnoses.Some plans are concerned that health-based payment will require them to expendadditional resources on gathering the data.Many plans can use their current systemsto report diagnoses without too much addi-tional effort, while others will need tomake large improvements. Some plansmay also be concerned that health-basedpayment will unfairly cause their revenuesto decline.

Despite plans’ concerns, we remain con-vinced that the effort to implement health-based payment is well justified by theimproved incentives to create equality forpeople with significant illness or disability.The experiences of the Colorado andOregon Medicaid programs and theWashington State Health Care Authority,using HMO-supplied encounter data toassess health status, reinforce our opinionthat gathering diagnostic data from inpa-tient and ambulatory encounters is feasi-ble. The Medicare, Medicaid, and SCHIPBudget Reconciliation Act of 1999 directsthe U.S. Department of Health and HumanServices to study how to make diagnosticreporting work better. Although this andother elements of the Act may slow downthe implementation of health-based pay-ment in the Medicare program, furtherstudy on the issue of diagnostic reportingshould, in the long run, be helpful.

CONCLUSION

We have shown that CDPS can be usedwith confidence by States to evaluatehealth status and pay HMOs equitably forboth SSI and TANF Medicaid beneficia-ries. We have argued previously thatStates contracting with HMOs for the careof persons with disabilities should imple-ment health-based payment. The growingnumber of States that have begun health-based payment indicate that it can be done.


The question of whether health-basedpayment for TANF beneficiaries is worththe effort is harder to settle. The sizableeffects of diagnoses on next-year expendi-tures for AFDC beneficiaries mean thathealth-based payment would be useful fordiscouraging risk selection and for moti-vating better services to TANF beneficia-ries. But the potential of health-based pay-ment to improve incentives for TANF ben-eficiaries is not so great as it is for SSI ben-eficiaries because a plan is less likely toattract a very biased selection of TANFbeneficiaries through its service designand provider network. If a State is notusing health-based payment for SSI benefi-ciaries, then the advantages of health-based payment for the TANF populationmay not be worth the administrative effort.But if a State is using health-based pay-ment for SSI beneficiaries, then the addi-tional work of implementation for TANFbeneficiaries is likely worthwhile.

We provide a set of payment weights thatStates can use. We also provide a diagnos-tic profile of the SSI and AFDC beneficia-ries that indicates the diversity of diag-noses and the greater level of illness anddisability among the SSI population.

Our comparison of the taxonomy andstatistical performance of the CDPS modelto other leading classification systems indi-cates that CDPS has advantages in bothstructure and predictive accuracy forMedicaid beneficiaries. CDPS shows mod-erate advantage over the HCC model andsignificant advantage over the ACG model,which includes many ill-defined orextremely frequent diagnoses that improvestatistical performance but reduce the reli-ability of health-based payment.

A key challenge in implementing health-based payment is getting data from healthplans that can equitably be used in adjust-ing payments. In the transition to adjust-ment by diagnoses, payers and plans willhave to make considerable efforts toensure that reporting is consistent acrossplans. As health status measured by diag-nosis is increasingly used to determinepayment levels, more efforts will be need-ed to understand the process of diagnosisand to improve diagnostic reporting.

ACKNOWLEDGMENTS

We wish to acknowledge the contribu-tions of our clinical consultants: LarryBerger, M.D., Stephen Bonfilio, Ph.D.,Donald Duncan, M.D., Jonathan Epstein,M.D., A. Stuart Hanson, M.D., Martin Lee,M.D., Richard Lentz, M.D., MichaelMacaulay, M.D., Hal Martin, M.D., MichaelShaw, M.D., Gregg Strathy, M.D., JanetteStrathy, M.D., Frederick Taylor, M.D.,Anton Willerscheidt, M.D., and especiallyTerry Wingert, M.D., and David Knutson,who helped arrange the clinical consulta-tions. The clinical consultants, however,bear no responsibility for the content of thearticle. Finally, David Strom of theUniversity of California at San Diego provid-ed excellent support in the preparation ofthe article’s figures and tables. We wouldalso like to thank those who provided valu-able comments on an earlier draft: ArleneAsh, John Bertko, Randall Ellis, ChristopherHogan, Sandi Hunt, Lisa Iezzoni, TonyTucker, Jonathan Weiner, Pete Welch, andthree anonymous reviewers.


REFERENCES

Ash, A.S., Ellis, R.P., Yu, W., et al.: Risk Adjustmentfor the Non-Elderly. Boston University, Departmentof Medicine. Report to the Health Care FinancingAdministration. Boston. June 1998.Carter, G.M., Bell, R.M., Dubois, R.W., et al.: AClinically Detailed Risk Information System forCost. Health Care Financing Review 21(3):65-91,Spring 2000.Dreyfus, T., and Kronick, R.: Paying Plans to Carefor People with Chronic Illness. In Kronick, R., andde Beyer, J., eds: Medicare HMOs: Making ThemWork for the Chronically Ill. Chicago. HealthAdministration Press, 1999.Druss, B.G., Bradford, D.W., Rosenheck, R.A., etal.: Mental Disorders and Use of CardiovascularProcedures After Myocardial Infarction. Journal ofthe American Medical Association 283(4):506-511,January 26, 2000.Ellwood, M.R., and Lewis, K.: The Ins and Outs ofMedicaid: Enrollment Patterns for California andFlorida in 1995. Abstracts. Chicago: Association ofHealth Services Researchers, 1999. Iezzoni, L., and Ayanian, J.: Paying More Fairly forMedicare Capitated Care. New England Journal ofMedicine 139(26):1933-1938, December 24, 1998.Johns Hopkins University School of Hygiene andPublic Health, Health Services Research andDevelopment Center.: The Johns Hopkins UniversityACG Case Mix Adjustment System. Documentationfor PC-DOS and Unix. Baltimore, Maryland.January 1998.

Kronick, R., and Dreyfus, T.: The Challenge of RiskAdjustment for People with Disabilities: Health-BasedPayment for Medicaid Programs. Princeton, NJ.Center for Health Care Strategies, 1997.Kronick, R., Dreyfus, T., Lee, L., and Zhou, Z.:Diagnostic Risk Adjustment for Medicaid: TheDisability Payment System. Health Care FinancingReview 17(3):7-33, 1996.Kronick, R., Zhou, Z., and Dreyfus, T.: Making RiskAdjustment Work for Everyone. Inquiry 32(1):41-55, 1995.Public Health Service and Health Care FinancingAdministration: International Classification ofDiseases, 9th Revision, Clinical Modification. U.S.Department of Health and Human Services.Washington, DC. U.S. Government Printing Office,September 1980.Medicare Payment Advisory Commission: Reportto the Congress, Medicare Payment Policy, Volume I.Washington, DC. Medicare Payment AdvisoryCommission, March 1998.Tollen, L., and Rothman, M.: Case Study: ColoradoMedicaid HMO Risk Adjustment. Inquiry 35(7):154-170, Summer 1998.Weiner, J.P., Tucker, A.M., Collina, A.M., et al.: TheDevelopment of a Risk-Adjusted Capitated PaymentSystem: The Maryland Medicaid Model. Journal ofAmbulatory Care Management 21(4):29-52, 1998.

Reprint Requests: Richard Kronick, Dept. of Family &Preventive Medicine, University of California, San Diego, 9500Gilman Drive, La Jolla, CA 92093-0622. E-mail:[email protected]


TECHNICAL NOTE

Regressions include an intercept termand age-sex dummy variables. Parameterestimates for the age-gender terms: In theregression for persons with disabilities:under 1 year of age, $697; age 1-4, -$630;age 5-15 male, -$433; age 5-14 female, -$387; age 15-24 male, -$124; age 15-24female, $168; age 25-44 male, $0 (referencecategory); age 25-44 female, $348; age 45-64 male, $395; age 45-64 female, $730;intercept, $1,126. For AFDC adults, age18-24 male, -$231; age 18-24 female, $478;age 25-44 male $0 (reference category);age 25-44 female, $259; age 45-64 male$496; age 45-64 female, $595; intercept,$624. For AFDC children, under 1 year ofage, $107; age 1-4, -$8; age 5-15 male $0(reference category); age 5-14 female, -$45;age 15-17 male, -$125; age 15-17 female,$537; intercept, $400. The “baseline”amount shown in the bottom of Table 6 isthe sum of the intercept plus the weightedaverage of the age-sex terms.

The regression for persons with disabili-ty includes interaction terms for age(coded as 1 for persons under age 19; 0otherwise) and selected CDPS categories.The values for these interaction terms are:very high cardiovascular, -$5,084; centralnervous system, medium, $1,201; very

high-cost pulmonary, $5,239; high-cost gas-trointestinal, $4,679; medium-cost gastroin-testinal, $3,738; low-cost gastrointestinal,$1,314; diabetes (all children are coded arecoded as Type 2, low), $666; metabolichigh, $7,844; metabolic medium, $947;infectious medium, $5,457; hematologicalvery high, -$5,645.

For AFDC adults and children, veryhigh and high-cost pulmonary conditionsare combined. For AFDC children andchildren with disability, all diabetes cate-gories are combined into a single categoryand all eye conditions are combined into asingle category.

The weighted average annual expendi-tures for the groups were $4,980 for per-sons with disability, $1,884 for AFDCadults, and $684 for AFDC children.

Using the significance statistics generat-ed by the weighted least-squares regres-sion, all coefficients shown in Table 6 aresignificant at P < 0.0001, except genitalextra low-cost at P < 0.01 for the disabled.All the coefficients of the variables forinteraction between age and diagnostic cat-egories for the disabled described abovewere also significant at P < 0.0001, exceptdiabetes low-cost at P < 0.05. However,because the expenditure data are not nor-mally distributed, the true standard errorsand P values are higher.


improving health-based payment for medicaid …cdps.ucsd.edu/cdps_hcfr.pdfdevelopment of cdps the...

Documents