predictors of medicare costs in elderly beneficiaries with breast, colorectal, lung, or prostate...

Health Care Management Science 2 (1999) 149–160 149

Predictors of Medicare costs in elderly beneficiaries with breast,colorectal, lung, or prostate cancer ∗

Lynne Penberthy a,b,c, Sheldon M. Retchin b, M. Kathleen McDonald a, Donna K. McClish c, Christopher E. Desch a,b,Gerald F. Riley e, Thomas J. Smith a, Bruce E. Hillner b and Craig J. Newschaffer d

a Massey Cancer Center, Medical College of Virginia, Virginia Commonwealth University, Richmond, VA 23298, USAb Department of Internal Medicine, Medical College of Virginia, Virginia Commonwealth University, Richmond, VA 23298, USA

c Department of Biostatistics, Medical College of Virginia, Virginia Commonwealth University, Richmond, VA 23298, USAd Thomas Jefferson University, Philadelphia, PA, USA

e Health Care Financing Administration

Received September 1998; revised March 1999

Background: Determining the apportionment of costs of cancer care and identifying factors that predict costs are important forplanning ethical resource allocation for cancer care, especially in markets where managed care has grown.

Design: This study linked tumor registry data with Medicare administrative claims to determine the costs of care for breast, colorectal,lung and prostate cancers during the initial year subsequent to diagnosis, and to develop models to identify factors predicting costs.

Subjects: Patients with a diagnosis of breast (n = 1,952), colorectal (n = 2,563), lung (n = 3,331) or prostate cancer (n = 3,179)diagnosed from 1985 through 1988.

Results: The average costs during the initial treatment period were $12,141 (s.d. = $10,434) for breast cancer, $24,910 (s.d. =$14,870) for colorectal cancer, $21,351 (s.d. = $14,813) for lung cancer, and $14,361 (s.d. = $11,216) for prostate cancer. Using leastsquares regression analysis, factors significantly associated with cost included comorbidity, hospital length of stay, type of therapy, andZIP level income for all four cancer sites. Access to health care resources was variably associated with costs of care. Total R2 rangedfrom 38% (prostate) to 49% (breast). The prediction error for the regression models ranged from <1% to 4%, by cancer site.

Conclusions: Linking administrative claims with state tumor registry data can accurately predict costs of cancer care during the firstyear subsequent to diagnosis for cancer patients. Regression models using both data sources may be useful to health plans and providersand in determining appropriate prospective reimbursement for cancer, particularly with increasing HMO penetration and decreased abilityto capture complete and accurate utilization and cost data on this population.

1. Background

Costs of cancer care have been estimated to account fornearly 12% of total health care expenditures [1], and ap-proximately $50 billion was spent on direct costs of cancercare in 1996 [2]. Determining how costs of cancer careare distributed and which factors are significant contribu-tors to the cost of cancer treatment may be important forfuture planning of health care resource allocation and forethical control of cancer care costs. These issues are espe-cially relevant with regard to the change in the financingof health care from fee-for-service to managed care meth-ods in which payment is capitated. As enrollment in man-aged care plans continues to increase, assessing the costsof care among enrollees with different medical conditions,comorbidities, and types of treatment is essential in orderto maintain equitable reimbursement rates for payment andto monitor the quality of care provided and for comparingthe cost effectiveness of treatment [3].

Recently, Riley and others used Medicare administrativeclaims linked with Surveillance Epidemiology and End Re-sults (SEER) data to estimate the average cost to Medicare

∗ Supported by a grant from the Agency for Health Care Policy and Re-search (AHCPR R01 HS0659-01A1).

for the five most prevalent cancers among the elderly in theUnited States in the year subsequent to diagnosis (breast,colorectal, lung, prostate and bladder cancers) [4]. Estab-lished by the National Cancer Act of 1971, the SEER pro-gram collects cancer data on a routine basis from desig-nated population-based cancer registries in geographic ar-eas that represent about 13.9% of the United States pop-ulation. Approximately 125,000 new cases of cancer areincluded annually in the SEER program in 9 geographicareas, including the states of Connecticut, Iowa, New Mex-ico, Utah, and Hawaii and the metropolitan areas of Detroit,San Francisco, Seattle-Puget Sound, and Atlanta. Trendsin cancer incidence, mortality and patient survival for theUnited States are estimated based on SEER data. Whilethese estimates published by Riley and others are an im-portant component for monitoring the national estimatesof the costs of cancer care, they may not be representa-tive of Medicare reimbursements for other geographic ar-eas and across other populations of cancer patients in theUnited States. For instance, it is well known that patternsof cancer care vary widely due to variations in practicepatterns, differences in patient characteristics such as stageat diagnosis, comorbidities and demographics as well asvariation in Medicare reimbursement rates [5–10]. There-

Baltzer Science Publishers BV

150 L. Penberthy et al. / Predictors of Medicare costs in elderly beneficiaries

fore, this study used data from the Virginia Cancer Reg-istry (VCR) to complement and build on the work by Rileyand others. In this study, costs for a non-SEER popula-tion for breast, colorectal, lung and prostate cancer dur-ing the initial year subsequent to diagnosis are provided,and factors that are associated with higher costs are iden-tified through a multivariate analysis. This analysis devel-ops models to identify factors that may explain some ofthe differences in costs of care in cancer patients as re-imbursed by Medicare. The model assesses the relativeimpact of each of these factors on the cost of care dur-ing the initial year subsequent to diagnosis; and it teststhe validity of these models by measuring the predictionerrors.

2. Methods

Data for the study were obtained from the VCR,Medicare Provider Analysis and Review (MEDPAR) file,Medicare Automated Data Retrieval System (MADRS),Medicare Health Insurance Master file, Medicare AnnualDemographic files, Area Resource File (ARF), and the 1990Census Data for Zip Code Level information (1990 Cen-sus Bureau Summary Tape 3b). A summary of variabledefinitions and the data sources from which these variableswere derived is shown in table 1. Incident cancer caseswere obtained from the VCR. These included all cases ofbreast, colorectal, lung and prostate cancer diagnosed inpersons aged 65 and older for 1985 through 1988 [11].The VCR is a statewide cancer registry that has collectedand maintained data on cancer patients since 1970. TheHealth Insurance Master file was used for the initial link-age of VCR cases to Medicare claims and to furnish infor-mation on beneficiary eligibility. The MEDPAR file con-tained data on hospital payments from Medicare, while theMADRS file was used for information on physician andother professional payments from Medicare. The ARF wasa source file for information on the availability of healthcare resources in the patient’s county of residence, whilethe census file was used to identify socioeconomic charac-teristics and included the proportion of residents with lessthan secondary education for the geographic area where thepatient resided. The data from the VCR were linked withMedicare files using an algorithm based on social secu-rity number or name and confirmed with date of birth andsex. Complete details on linking the files and methods forcost calculations can be obtained from prior publications[6,9,11].

The final sample for calculating the Medicare reimburse-ment in cancer patients was evaluated according to vary-ing levels of treatment, stage and other clinical and demo-graphic characteristics. The sample included patients con-tinuously eligible for both Medicare Part A and Part B witha diagnosis of incident breast, colorectal, lung or prostatecancer from 1985 through 1988. Criteria for exclusion werediagnosis at autopsy (<1%), no VCR-reported diagnosis

date (<0.5%), enrollment in an HMO during the initialtreatment period (<0.5%), and no claims during the ini-tial treatment window from either MEDPAR or MADRS(<2%). For persons who survived less than one year, initialtreatment period costs were calculated for the total survivaltime during that year. Persons diagnosed in 1989 wereexcluded since a full year of follow-up data were unavail-able.

The initial treatment period was defined as the year sub-sequent to diagnosis. This definition was established withthe recognition that in some cases, treatment during thefirst year subsequent to diagnosis may include treatmentfor recurrent disease. This time period was selected basedon three factors. First, initial data exploration suggestedthat the initial course of therapy, including adjuvant radia-tion therapy, was completed for the four cancer sites withina one year period. Second, this time period representeda consensus from a group of clinical oncologists. Third,one year was chosen for comparison purposes. This wasthe period of time used by Riley and others for their costestimates [4].

Initial treatment was categorized into four basic strata:(1) definitive surgery, (2) nonsurgical treatment, (3) surgeryplus nonsurgical therapy, and (4) no treatment. The ini-tial treatment reported by the VCR was used to catego-rize initial treatment with the exception of definitive sur-gical therapy. For definitive surgery, cases were excludedif there were no matching MEDPAR claims because de-finitive surgical therapy was unlikely to be provided in anoutpatient setting during this time period. The “definitivesurgery” treatment category included cancer-specific surgi-cal procedures that are generally aimed at tumor resectionand are unlikely to be used only for diagnosis or for symp-tom relief. Because nonsurgical therapy is frequently pro-vided on an outpatient basis it is not completely capturedthrough the inpatient-based MEDPAR file [11,12]. There-fore, if a nonsurgical treatment was reported by the VCRit was assumed that the associated cost was included inthe MADRS reimbursement data, even without an associ-ated inpatient claim. This category included chemotherapy,radiation therapy or hormonal therapy. The “surgery plusnonsurgical therapy” category included a definitive surgi-cal procedure plus at least one of the nonsurgical treat-ments described above. The “no treatment” category re-ferred to those persons who did not have any cancer treat-ment reported from either VCR or MEDPAR. Diagnosticprocedures such as transurethral resection of the prostateand a subset of breast biopsies, such as needle and in-cisional, were included in this category. These were in-cluded as “no therapy” since they do not typically consti-tute definitive surgical treatment. Patients categorized ashaving nonsurgical therapy and no treatment may have re-ceived treatment that was not reported to either the VCRor Medicare [12]. This underreporting is likely to bias theresults towards the null in estimating the effect of treatmenton cost.

L. Penberthy et al. / Predictors of Medicare costs in elderly beneficiaries 151Ta

ble

1V

aria

ble

defin

ition

san

dso

urce

s.


Stage at diagnosis was based on the summary stagingsystem. Three categories of stage (local, regional and dis-tant) were used in this analysis. This three level systemis useful because it is the staging method most consis-tently available to cancer registries [13,14]. Comorbiditywas calculated from the MEDPAR file using the Dartmouth-Manitoba ICD-9-CM conversion algorithm for modificationof the Charlson comorbidity index [15,16]. Comorbidity isadditional diseases or conditions that are concurrent withthe primary disease of interest and that might impact onoutcomes. For example diabetes is an important comorbidcondition typically taken into account. The Charlson scoreis a summary of the weighted scores for a set of prognos-tically significant medical conditions. The score for theDartmouth-Manitoba modification was based on selectedICD-9-CM codes occurring in hospital admissions in the365 days prior to the diagnosis date. Concurrent cancerICD-9-CM codes were excluded from the comorbidity cal-culations. Scores ranged from zero to six. For analyticpurposes, comorbidity scores were categorized as 0 (no co-morbidity), 1 (one comorbid condition with a lower weightsuggesting less severity or less likely to predict mortality)and >1 (cases with either more than one comorbid disease,or one comorbid disease with higher potential for serioussequelae).

Costs in this paper were defined as the amount the HealthCare Financing Administration (HCFA) reimburses to hos-pitals, physicians and other health care providers (e.g., hos-pice) for the care of Medicare beneficiaries. The amountreimbursed by HCFA does not capture all the direct costsof care borne by the providers and the patients. Nor do fig-ures based on HCFA reimbursement include indirect costssuch as loss of time to work.

Cost data were obtained from the MEDPAR file for PartA data, and the remainder of costs (physician outpatient,suppliers and other costs) was obtained from the MADRSfile. MEDPAR contains records of all hospital admissionsand includes five International Classification of Diseases9th Revision Clinical Modification (ICD-9-CM) diagnosticcodes and up to three ICD-9-CM procedure codes alongwith total charges and reimbursement for that hospital ad-mission. MADRS less consistently includes ICD-9-CDcodes, but includes charges and reimbursements for all in-patient, outpatient, home health and hospice claims. Costcalculations were performed using SAS for Windows. Allcosts were adjusted to 1990 dollars using the changes inPart A and Part B per capita Medicare payments to pro-vide comparability with others [4]. Cost estimates for uni-variate presentation were adjusted to 1997 dollars basedon the Expanded Health Cost Index from Ginsburg andGabel [17]. The Medicare reimbursement amount wasused to estimate cost rather than charges, since chargesdo not reflect actual payments. Because outpatient physi-cian office claims are summed and provided as an annualamount, the costs added for this category were weightedbased on the number of months included for the analy-sis. Categories for which Medicare costs were calculated

included the total Medicare reimbursements, and costs byage group, race, sex, comorbidity status, stage at diagno-sis and treatment category during the year subsequent todiagnosis.

3. Analysis

The initial step was a univariate analysis to calculatethe average costs across categories of clinically relevantvariables. Because multiple comparisons were done in thisunivariate analysis, Bonferroni’s method of adjustment wasused to establish statistical significance at p < 0.001 [18].In addition to calculating the average costs in Virginiaacross the four cancer sites, a univariate comparison ofMedicare costs in the SEER/Medicare population was per-formed. Cost data from the SEER Medicare linkage per-formed by Riley and others were obtained to provide com-parative information as to how cost to Medicare mightvary across geography and different patient populations.Comparisons were limited to total costs and stage by sur-vival status because these were the SEER data availablefor analyses from this period. Cost data were calculatedusing the same time period, and the same staging schemefor comparability.

The second step in the analysis was to identify vari-ables, independent of data source, that were significantlyassociated with total cost of care during the initial treat-ment window using multiple regression analysis. As thefirst step, all continuous variables were centered to reducethe variance, risk of collinearity and improve model fit [19].Prior to developing the model, correlations among the vari-ables in the model were assessed. Only two – number ofphysicians and distance to a cancer radiation treatment cen-ter – had substantial correlation. Therefore, the oncology-related physician specialty density (number of oncology-related specialists/ZIP) was used alone as the dependentvariable to represent access to health care.

A least squares regression was performed using the SASProcedure PROC REG. This was done as a two stage leastsquares analysis (2SLS) to predict cost, since it is similar toan analysis of market prices. For this study prices were theHCFA reimbursement amounts for cancer care during theinitial year following diagnosis. In such an analysis, thesupply factors related to access and other market factorswere separated from the demand factors, patient clinicaland sociodemographic characteristics [20].

The first stage of the two stage least squares model in-cluded predictors of the “supply”. For this study, supplywas defined as the physician specialty density. The ex-clusion restrictions, or variables excluded from the secondstage model, but included in the first stage were the percentof persons aged 65 and older in the ZIP code of residencefor each patient, the percent of nonwhite persons living inthe ZIP code of residence for each patient, as well as theMSA status of that ZIP code. The model used to predictthe number of physicians with oncology-related specialties

L. Penberthy et al. / Predictors of Medicare costs in elderly beneficiaries 153

in the ZIP code of residence for the patient was as fol-lows:

Predicted number of MDs

= b0 + b1 % Elderly + b2 % NonWhite + b3 MSA

+ b4 Age + b5 Race + b6 Income . . . .

The value for the predicted number of physicians for eachpatient was then included as a variable in the second stagemultiple regression. The second stage model included allvariables in the first stage model except the three exclu-sion restriction variables and added the predicted numberof MDs.

Because the distribution of costs was skewed, the naturallogarithm of total costs was used as the dependent variable.Therefore, the model to predict cost was based on

ln(Cost) = b0 + b1 Age + b2 Race + b3 MSA . . . ,

which becomes

$Cost = eb0 × eb1 Age × eb2 Race × eb3 MSA . . . .

From this, the unlogged or exponentiated estimate of thecoefficient provided the factor by which the cost increasedfor each level of increment in the parameter of interest. Re-gression diagnostics were performed to evaluate collinear-ity using an eigenvalue analysis as part of SAS Proc REG.There was no meaningful collinearity identified betweenany of the variables used in either the first or second stageof the least squares model based on eigenvalues, varianceinflation and variance proportion estimates.

In order to assess the accuracy of the model used to pre-dict cost for elderly patients with cancer, the percent predic-tion error was calculated based on two methods. First, thesample was randomly divided into two halves. One halfof the sample was used to build the model and calculatethe parameter estimates. The remaining half of the samplewas used to validate the method by calculating the pre-dicted cost based on the previously constructed model andcomparing those predicted costs with the actual costs ob-tained from the Medicare claims data. The second methodused was similar, but instead of randomly splitting the sam-ple, a temporal division was created, using the data fromthe earlier period, 1985–86, to calculate the parameter esti-mates and construct the model and data from the latter timeperiod, 1987–88, to assess the accuracy of the model forpredicting cost. The actual prediction error was calculatedfor each individual patient, and the mean and standard er-ror for the population prediction were calculated using SASProc Univariate. The Type II error or power of this studyto identify costs is estimated to be at least 80%.

4. Results

The final data set included 1,952 breast, 2,563 colorectal,3,331 lung and 3,179 prostate cancer cases. Overall costsand costs stratified by demographic and clinical characteris-tics for the initial treatment period after a cancer diagnosis

in Virginia are provided in table 2. The patterns of costsacross each of the cancer sites varied markedly by demo-graphic and clinical characteristics. The average cost byage group did not differ, with the exception of lung cancer,in which persons aged 75+ had nearly 20% lower coststhan younger persons did. The costs of care for black pa-tients were higher than for white patients for all cancer siteswith the exception of colorectal cancer.

Increasing comorbidity was consistently associated withincreased costs for all cancer sites. The pattern of costs as-sociated with increasing stage of disease varied across can-cer sites. The other factor that showed a consistent patternassociated with reimbursement on univariate analysis wastreatment. The costs for treatment significantly decreasedwith decreasing intensity of treatment. Patients receivingno treatment had the lowest costs for all sites.

Table 3 contains univariate cost data for the SEER pop-ulation compared with the Virginia Cancer Registry popu-lation. The costs in Virginia are consistently significantlylower across survival status and stage for three of the fourcancers studied. The difference in Medicare costs for breastcancer patients in Virginia who did not survive was the ex-ception. Although the sample size for the SEER popula-tion is much larger than for the Virginia cancer population,the variation in costs is wide, with standard deviations inthe two populations consistently of similar large magni-tude.

Table 4 contains the unlogged parameter estimate (costmultiplier) and level of statistical significance for each vari-able used in the model to predict costs of cancer, for each ofthe four cancer sites studied. Therapy (surgical plus non-surgical), increasing income and the interaction term forcomorbidity and length of stay were significantly associ-ated with reimbursement across all four cancer sites. Inaddition, increasing income was associated with increasedcosts. Measures of access or market forces, defined asthe number of specialty physicians/ZIP, was a significantpredictor of costs for breast and colorectal cancer. Increas-ing age was significantly associated with increased costs forcolorectal cancer and race was associated with cost for lungcancer. The independent association between stage at diag-nosis and cost varied by cancer site. Distant stage cancerat diagnosis significantly predicted increasing costs only forcolorectal cancer. However, regional disease at diagnosisincreased costs substantially compared to local disease forbreast and prostate cancer. There was a significant interac-tion between hospital LOS and comorbidity. The impact ofthe combined factors on cost was substantial, ranging froma 20% to 40% increment in cost for various combinationsof LOS and comorbidity.

The total R-squared variance ranged from 49% for breastcancer, to 46% for colorectal cancer, 39% for lung cancerand 38% for prostate cancer. The prediction error for thesetwo samples ranged from less than 1% to 4% by cancersite.


Tabl

e2

Tota

lav

erag

em

edic

are

reim

burs

emen

t∗du

ring

the

initi

altr

eatm

ent

peri

odfo

rel

derl

ypa

tient

sw

ithbr

east

,co

lore

ctal

,lu

ngor

pros

tate

canc

ers,

1985

–88.

∗C

ost

adju

sted

to19

97do

llars

base

don

the

Exp

ande

dH

ealth

Cos

tIn

dex

(fro

m[1

7]).


Tabl

e3

Com

pari

son

ofm

ean

cost

sfo

rM

edic

are

inV

irgi

nia

vers

usth

eSE

ER

Reg

istr

ies,

1985

–88.

All

diff

eren

ces

wer

est

atis

tical

lysi

gnifi

cant

lydi

ffer

ent

atp6

0.01

exce

ptw

here

desi

gnat

edw

ith∗

or∗∗

.∗

Indi

cate

sno

stat

istic

ally

sign

ifica

ntdi

ffer

ence

.∗∗

Indi

cate

sa

stat

istic

ally

sign

ifica

ntdi

ffer

ence

atp

=0.

05.

Dat

aba

sed

on(−

1,11

)m

onth

saf

ter

diag

nosi

sof

canc

erfo

rco

mpa

riso

nw

ithav

aila

ble

SEE

Rda

ta.

156 L. Penberthy et al. / Predictors of Medicare costs in elderly beneficiariesTa

ble

4Su

mm

ary

ofL

east

Squa

res

Ana

lysi

sus

edto

pred

ict

incr

easi

ngco

sts

toM

edic

are

duri

ngth

ein

itial

trea

tmen

tpe

riod

for

brea

st,

colo

rect

al,

lung

and

pros

tate

canc

erin

the

elde

rly,

1985

–88.

∗M

odel

led

Acc

ess

(sup

ply)

from

Firs

tSt

age

LS

pred

ictin

gnu

mbe

rof

MD

s.Sh

adin

gin

dica

tes

stat

istic

ally

sign

ifica

ntas

soci

atio

nac

ross

all

four

canc

ersi

tes

for

that

vari

able

.


5. Discussion

The costs of cancer are rising and since cancer occurswith disproportionate frequency in the elderly, this is anespecially important issue for the Medicare program. Stud-ies that examine the costs of cancer are vital to effortsaimed at controlling these expenditures. Earlier investi-gations that endeavored to link SEER data with Medicareclaims for evaluating the costs of cancer care have beenextremely successful, though limited to areas representedby the SEER program. This study revealed that importantvariables for predicting costs included some demographics,such as increasing income, and clinical characteristics suchas comorbidity level, stage at diagnosis and treatment. Thisstudy of Medicare beneficiaries with cancer in a specific,non-SEER, geographic region demonstrated that the amountreimbursed by Medicare for treatment of four major cancersvaried substantially and was significantly less for Virginiathan for the SEER elderly population. Further analysesfor comparison, such as using the SEER data in predictingcosts would be useful. Part of the lower costs in Virginiamay be related to differences in the Medicare Part A reim-bursement rate variation across the SEER regions. This isno average about 11% higher for SEER than for Virginia,generally within the range of percent agreement noted inour comparisons in table 3 (70–98% agreement).

In this study of the costs reimbursed by Medicare forthe initial treatment of cancer; increasing income, treat-ment, and the interaction between comorbidity and LOShad the most consistent pattern of increasing costs for allcancer sites in both univariate and multivariate analyses.Other factors that were associated with increased costs ofcare across different cancer sites were age, stage, and race.The lack of a consistent association of stage and cost maybe related to real differences or due to limitations in thedata such as the lack of more accurate staging available atthat time. Additional limitations related to consistency ofpatterns may be related to the time interval used to assesscosts. As noted earlier, the one year period after initialdiagnosis is likely to include all costs for initial treatmentbut may also include costs of recurrent disease for somecancers, especially lung.

The effect of comorbidity on cost, together with its re-lation to LOS in predicting costs, is important when con-sidering risk adjustment, not just for general health careproviders, but potentially for setting rates for specialty ser-vices, such as oncologic care. Comorbidity is an exogenousvariable, which can significantly influence the cost of care,but is often outside the control of the provider. It is, there-fore, essential to include comorbidity in any risk adjustmentmodel. From a clinical and provider perspective, the im-pact of comorbidity and stage on costs of care might bewithin the control of providers to some extent. Provisionof uniform coverage and access to appropriate preventivecare for comorbid conditions by an HMO might reduce theeffect of comorbidity on LOS and costs. Access to can-cer prevention services may reduce the stage at diagnosis,

which has an impact on cost for at least some of the cancersincluded in this study. An additional consideration is thepotential for differential coding or coding bias of comorbid-ity in inpatient admissions. Reimbursement by Medicare isin part based on underlying comorbid conditions. There-fore differential coding across providers might result in abiased estimate of cost across certain populations. Giventhe lack of association with certain geographic and demo-graphic characteristics and the association of increased costwith increased income, it is unlikely that coding bias is asignificant factor in this analysis.

Age and race, two other exogenous variables, wereinconsistently associated with increased cost. These arecommonly used in models for risk adjustment, but havebeen shown to be insufficient by themselves for risk-adjusting [3]. The inconsistency and relative unimportancein predicting cost are supported in this study as well. Al-though treatment is a significant predictor of cost, it isan endogenous variable, and therefore has the potential tobe manipulated by providers to maximize reimbursement.Thus, it is less useful in developing risk adjustment mod-els. Nonetheless, costs of care in cancer patients accordingto treatment have not been previously evaluated among theMedicare elderly. Although the effect of treatment on costsof care may not be useful for risk adjustment, informationon costs by treatment is important for assessing quality ofcare and for comparing cost effectiveness of various treat-ments for specific cancer sites. The effect of market factors(number of specialty MDs/ZIP) was surprisingly significantonly for those cancers, breast and colorectal, for whichthere is least variability in treatment patterns. This may bebecause of increased variation related to differences in pat-terns of care for prostate and lung, or may be truly related toan increased likelihood of market forces influencing costsfor breast and colorectal cancer differentially.

Increasing income at the group level (ZIP code), a mea-sure of social class, was consistent in predicting costs. In-creased social class, as measured by the ZIP level age-and race- adjusted income, may reflect increased access toMedigap type insurance, and may independently predict ad-ditional costs not represented by the general categories oftreatment assessed in this study. The increased ability topay may permit elderly beneficiaries to afford additionalutilization that requires copay and that also results in in-creased expenditure to Medicare.

The strong association with comorbidity was in contrastto findings by others who have evaluated the effects ofage, stage and comorbidity on costs of cancer care in anHMO population [21]. Except for the lack of agreement onincreasing costs with increasing comorbidity, the other find-ings in this study relating to stage were in agreement withthose reported elsewhere [4,21]. The contradictory find-ings for comorbidity may be a result of either real differ-ences in the comorbidity levels of the two populations, theresult of methodologic differences in comorbidity assess-ment, or the effect of different approaches in calculation ofcosts. Comorbidity in the HMO population was based on


the number of medications prescribed in a one year period[16,22] while this study used ICD-9 codes from Medicareinpatient claims data for a one year period prior to the diag-nosis date. Both methods are imperfect and probably repre-sent different levels and types of comorbidity. In addition,the costs captured by previous studies included pharmacycosts, while in this study only the amount reimbursed byMedicare was included. However, the costs of care fornon-chemotherapy-related pharmacy are likely to representonly a small proportion of the total costs of care for cancerpatients. It is also unlikely that there should be a differen-tial distribution of Medicare covered and uncovered coststhat is related to comorbidity. The lack of association ofcomorbidity and costs of care in the previous study con-ducted in an HMO may be related to actual differences inthe comorbidity levels of the two populations due to differ-ences in their age distributions. Finally, HMOs represent adelivery system that in themselves may account for differ-ences in cost, since the advantage of an HMO is to providecare at lower cost.

The data from this study raise important policy issues.First, there were sizable variations in costs by cancer site.The average costs of care for patients with breast andprostate cancer were similar during the initial treatmentperiod. However, the costs were approximately twofoldhigher for lung cancer and colorectal cancer. These dif-ferences are likely explained by the costs associated witha higher likelihood of additional therapies (e.g., radiation,chemotherapy) and more expensive hospitalizations (i.e.,longer lengths-of-stay from more invasive surgery). It isinteresting that these cost ratios among the four differ-ent cancers are relatively similar to those described byScheffler and Phillips [23]. An important policy impli-cation of these variations indicates the necessity of calcu-lating the costs of specific cancers when determining caserates for cancer. The latter is a frequent strategy for risk-sharing arrangements between managed care health plansand providers.

Another meaningful policy issue raised by these data isthe variation in costs within each cancer site. This differ-ence remained even when adjusting for stage of disease.For instance, the coefficient of variation for breast cancerand prostate cancer during the initial treatment period wasalmost unity. However, the coefficient of variation wasapproximately 0.6 for colorectal cancer and 0.69 for lungcancer. These differences may reflect the heterogeneity inseverity among cancer cases, both due to the cancer itselfand to other comorbid illnesses. Alternatively, these datamay reflect the diversity in practice styles among providersby cancer site and even by stage. These variations maybe especially likely for a disease like cancer, where thereis disagreement over appropriate use of expensive inter-ventions such as chemotherapy. The practice variations ofcostly technologies for the treatment of cancer have beenpreviously described [24].

The heterogeneity in costs among the cancer cases ob-served in this study has important implications for risk ad-

justment methodologies being considered for the Medicareprogram. Under the Balanced Budget Act of 1997, theMedicare risk program is required to link capitation pay-ments with the health status of their enrollees by January 1,2000. The data from this study reveal that conventional riskadjustment techniques need to take into consideration thewide variations in costs for cancer. Some risk adjustmentmethods that have already been proposed underscore thisissue [25]. Furthermore, data on costs from administrativeclaims such as presented in this study are especially im-portant to evaluate before the introduction of risk adjustorsin the Medicare program. With the expansion of diseasemanagement programs for cancer, and the organization ofspecialty physician groups (e.g., medical oncologists) intopractice management companies [26], case rates for cancerare likely to develop. However, once risk adjustors for theMedicare program have been instituted it will be extremelydifficult to establish a baseline in costs for some chronicillnesses, such as cancer, because of the likelihood of codecreep that plagued the prospective payment system in itsearly years [27].

The policy implications of the data from this study com-paring the costs for decedents and survivors of canceralso deserve special mention. The costs of care for dece-dents with cancer are important and accounts for 8% of allMedicare expenses [28]. Decedents with cancer averageapproximately 60% higher costs for medical care duringthe last year of life compared to decedents as a whole [29].In this study, the cost ratios were similarly higher betweennon-survivors and survivors for breast cancer and prostatecancer (2.2 times higher for non-survivors than survivorsof breast cancer; 1.8 times higher for non-survivors thansurvivors of prostate cancer). The discrepancy betweencosts for non-survivors and survivors was much smallerfor both lung cancer and colorectal cancer (1.2 for lungand 1.4 for colorectal cancer). This may suggest morerationing of care for patients with cancers that are lesslikely to be salvageable because of historically poor sur-vival rates (e.g., lung cancer). Conversely, it may reflectmore aggressive care of non-surviving patients with cancersthat are frequently amenable to heroic interventions (e.g.,breast cancer). These heroic measures have been shownto vary considerably among physicians, according to prac-tice style and beliefs [30]. We conclude that these patternsin the costs of decedent care for cancer merit further re-search.

Since this analysis of the costs of care for cancer in-cluded information from administrative claims, the datashould be interpreted with caution. Administrative claimscan be an unreliable data source [31], particularly for clin-ical variables such as comorbidity. And yet, claims datahave been used successfully to create models for risk ad-justment [3]. It is evident from the multivariate analysisthat comorbidity is a key predictor of costs in this pop-ulation with cancer. Since the magnitude of the effectof comorbidity on costs of care for cancer was large, itshould be included when setting risk adjusted capitation


rates. Another limitation of this study is that the data from1985 through 1989 do not accurately represent current pat-terns of care. The data from VCR and Medicare do notpermit the investigation to identify adjuvant versus neoad-juvant versus palliative therapy, nor to identify recurrentdisease. This limitation is likely to be most influentialfor lung cancer. However, the use of neoadjuvant ther-apy during the study period was not common. Analysesusing more current data would permit the inclusion of suchtreatments based on physician office claims. Using thesedata has limitations; however, the advantage is that the in-creasing penetration of Medicare HMOs into many mar-kets [32] makes analyses of more current administrativeclaims fraught with problems [31]. Therefore, until thedata from HMOs on utilization become consistently avail-able [33], studies such as this one may be increasinglydifficult to perform concurrently. These will, however, benecessary to accurately reflect current patterns of practiceand the associated costs.

Among the most useful exogenous variables in the modelfor predicting cancer costs were comorbidity, income, andtreatment. Cancer stage at diagnosis, age and race weresignificant predictors of costs of care, but varied acrosscancer sites. Furthermore, while the low prediction errorrates suggest that factors included in the model may beuseful for risk adjustment purposes, the predictive abilityof the model varied by cancer site. The model predictedbetween 38 and 49% of the total variation in costs duringthe initial treatment period. Although the total amount ofvariation explained by the model was relatively low, it wassimilar to that found in other models [3]. The split sampleapproach actually used the data to predict the cost, with aprediction error of less than one to four percent accordingto cancer site. Further development of such models shouldbe done to predict costs of care for cancer patients beyondthe first year subsequent to diagnosis as the initial treat-ment period that includes outpatient information. Becausemany of the costs of cancer are likely to occur in the initialyear subsequent to diagnosis, a mix of retrospective andprospective risk adjustment may be the optimal method forrate setting [3].

Using administrative claims data to evaluate the costsof cancer can accurately predict reasonable levels of thetotal variation in costs during the initial treatment period.In addition, these models can provide guidance to healthplans, providers and patients regarding the critical factorsessential in determining appropriate prospective reimburse-ment for elderly patients with cancer, and further using riskadjustments when measuring quality of care.

References

[1] D.P. Rice, T.A. Hodgson and F. Capell, The economic burden ofcancer, 1985: United States and California, in: Cancer Care andCost: DRGs and Beyond, eds. R.M. Scheffler and N.C. Andrews(Health Administration Press, Ann Arbor, MI, 1989), pp. 39–59.

[2] R. Rundle, Sellick pioneers selling cancer care in HMOs, WallStreet Journal 12 (August 1996) 81–82.

[3] J. Fowles, P. Weiner, D. Knutson, E. Fowler, A. Tucker andM. Ireland, Taking health status into account when setting capitationrates; a comparison of risk-adjustment methods, Journal of theAmerican Medical Association 276(16) (1996) 1316–1321.

[4] G.F. Riley, A.L. Potosky, J.D. Lubitz and L.G. Kessler, Medicarepayments from diagnosis to death for elderly cancer patients bystage at diagnosis, Medical Care 33 (1995) 828–841.

[5] G.L. Lu-Yao, D. McLerran, M. Wasson and J.E. Wennberg, Anassessment of radical prostatectomy time trends, geographic vari-ation and outcomes, Journal of the American Medical Association269(20) (1993) 2633–2636.

[6] C.E. Desch, L. Penberthy, C.J. Newschaffer, B.E. Hillner, M. Whit-temore, D. McClish, T.J. Smith and S.M. Retchin, Factors thatdetermine the treatment for local and regional prostate cancer,Medical Care 34 (1996) 152–162.

[7] A.B. Nattinger, M.S. Gottlieb, J. Veum, D. Yahnke and J.S.Goodwin, Geographic variation in the use of breast-conservingtreatment for breast cancer, New England Journal of Medicine 326(1992) 1102–1107.

[8] T.J. Smith, L. Penberthy, C.E. Desch, M. Whittemore, C. Newschaf-fer, B.E. Hillner, D. McClish and S.M. Retchin, Differences ininitial treatment patterns and outcomes of lung cancer in the elderly,Lung Cancer 13 (1995) 235–252.

[9] C.J. Newschaffer, L. Penberthy, C.E. Desch, S.M. Retchin andM. Whittemore, The effect of age and comorbidity of non-metastaticbreast cancer in elderly women, Archives of Internal Medicine 156(1996) 85–90.

[10] Anonymous, Revised standardized per capita rate of payment for1995 (Health Care Financing Administration, 1995).

[11] D.K. McClish, L. Penberthy, M. Whittemore, C. Newschaffer,C. Woolard, C.E. Desch and S.M. Retchin, Ability of Medicareclaims data and cancer registries to identify cancer cases andtreatment, American Journal of Epidemiology 145 (1997) 227–233.

[12] A.L. Potosky, G.F. Riley, J.D. Lubitz, R.M. Mentnech and L.G. Kes-sler, Potential for cancer related health services research using alinked Medicare-tumor registry database, Medical Care 31 (1993)732–748.

[13] M. Charlson and A. Feinstein, An analytic critique of existingsystems of staging for breast cancer, Surgery 73(4) (1973) 579–598.

[14] D.E. Henson, L. Ries and E.M. Shambaugh, Survival results dependon the staging system, Semin. Surgical Oncology 8 (1992) 57–61.

[15] M.E. Charlson, P. Pompei, K.L. Ales and C.R. MacKenzie, A newmethod of classifying prognostic comorbidity in longitudinalstudies: development and validation, Journal of Chronic Diseases40 (1987) 373–380.

[16] P.S. Romano, L.L. Roos and J.G. Jollis, Adapting a clinicalcomorbidity index for use with ICD-9-CM administrative data:differing perspectives, Journal of Clinical Epidemiology 46 (1993)1075–1079.

[17] P.B. Ginsburg and J.R. Gabel, Tracking health care costs: what’snew in 1998?, Health Affairs 17(3) (1998) 141–146.

[18] K. Godfrey, Medical Uses of Statistics, 2nd ed., Chapter 12 (NewEngland Journal of Medicine Books, Boston, MA, 1992).

[19] Kleinbaum, Kupper and Muller, Applied Regression Analysis andOther Multivariable Methods, Chapter 12: Regression Diagnos-tics, 2nd ed. (Kent Publishing Company, Boston, MA, 1988),pp. 197–209.

[20] J. Cohen and P. Cohen, Applied Multiple Regression/CorrelationAnalysis for the Behavioral Sciences, 2nd ed. (Lawrence Erlbaum,1983).

[21] S.H. Taplin, W. Barlow, N. Urban, M.T. Mandelson, D.J. Timlin,L. Ichikawa et al., Stage, age, comorbidity, and direct costs ofcolon, prostate, and breast cancer care, Journal of the NationalCancer Institute 87 (1995) 417–426.

[22] M. Von Korff, E. Wagner and K. Saunders, A chronic disease scorefrom automated pharmacy data, Journal of Clinical Epidemiology45 (1992) 197–203.


[23] R.M. Scheffler and K.A. Phillips, DRGs and the financing of cancercare in the United States: new estimates and new issues, in: CancerCare and Cost: DRGs and Beyond, eds. R.M. Scheffler and N.C.Andrews (Health Administration Press Perspectives, Ann Arbor,MI, 1989), pp. 3–23.

[24] J.Z. Ayanian and E. Guadagnoli, Variations in breast cancertreatment by patient and provider characteristics, Breast CancerRes. Treat. 40 (1996) 65–74.

[25] L.I. Iezzoni, J.Z. Ayanian, D.W. Bates and H.R. Burstin, Payingmore fairly for Medicare capitated care, N. Engl. J. Med. 339(1998) 1933–1938.

[26] J.C. Robinson, Consolidation of medical groups into physicianpractice management organizations, JAMA 279 (1998) 144–149.

[27] D.C. Hsai, W.M. Krushat, A.B. Fagan, J.A. Tebbutt and R.P. Kus-serow, Accuracy of diagnostic coding for Medicare patients under theprospective-payment system, N. Engl. J. Med. 318 (1988) 352–355.

[28] S. Long, J. Gibbs, J. Crozier et al., Medical expenditures of terminalcancer patients during the last year of life, Inquiry 21 (1984)315–327.

[29] G. Riley, J. Lubitz, R. Prihoda and E. Rabey, Use and costs ofMedicare services by cause of death, Inquiry 24 (1987) 233–244.

[30] L.C. Hanson, M. Danis, J.M. Garrett and E. Mutran, Who decides?Physicians’ willingness to use life-sustaining treatment, Arch.Intern. Med. 156 (1996) 785–789.

[31] S.M. Retchin and D.J. Ballard, Establishing standards for the utilityof administrative claims data, Health Services Research 32 (1998)861–866.

[32] S.M. Retchin and R.E. Hurley, The revision of government-sponsored health care, American Journal of Medicine 98 (1995)529–530.

[33] W.P. Welch and H.G. Welch, Fee-for-data: a strategy to open theHMO black box, Health Affairs 14(4) (1995) 104–116.

predictors of medicare costs in elderly beneficiaries with breast, colorectal, lung, or prostate...

Documents