Incidence cost estimates or longitudinal estimates of medical

Transcription

1 CONDUCTING THE COST ANALYSIS Comparison of Approaches for Estimating Incidence Costs of Care for Colorectal Cancer Patients K. Robin Yabroff, PhD,* Joan L. Warren, PhD,* Deborah Schrag, MD, Angela Mariotto, PhD,* Angela Meekins, Marie Topor, and Martin L. Brown, PhD* Background: Estimates of the costs of medical care vary across patient populations, data sources, and methods. The objective of this study was to compare 3 approaches for estimating the incidence costs of colorectal cancer (CRC) care using similar patient populations, but different data sources and methods. Methods: We used 2 data sources, linked SEER-Medicare and Medicare claims alone, to identify newly diagnosed CRC patients aged 65 and older and estimated their healthcare costs during the observation period, 1998 to Controls were matched by sex, age-group, and geographic location. We compared mean net costs, measured as the difference in total cost between cases and controls, for: (1) a SEER-Medicare cohort, (2) a Medicare claims alone cohort, and (3) a modeled phase of care approach using linked SEER-Medicare data. The SEER-Medicare cohort approach was considered the reference. Results: We found considerable variability across approaches for estimating net costs of care in CRC patients. In the first year after diagnosis, mean net costs were $32,648 (95% CI: $31,826 and $33,470) in the SEER-Medicare cohort. The other approaches understated mean net costs in year 1 by about 16%. Mean net 5-year costs of care were $37,227 (95% CI: $35,711 and $38,744) in the SEER-Medicare cohort, and $30,310 (95% CI: $25,894 and $34,726) in the claims only approach, with the largest difference in the 65 to 69 age group. Mean net 5-year costs of care were more similar to the reference in the modeled phase of care approach ($37,701 range: $36,972 and $38,446 ). Differences from the SEER-Medicare cohort estimates reflect misclassification of prevalent cancer patients as newly diagnosed patients in the Medicare claims only approach, and differences in years of data and assumptions about comparison groups in the modeled phase of care approach. Conclusions: CRC incidence cost estimates vary substantially depending on the strategy and data source for identifying newly diagnosed cancer patients and methods for estimating longitudinal From the *Division of Cancer Control and Population Sciences, National Cancer Institute, Bethesda, Maryland; Massachusetts General Hospital Cancer Center, Harvard Medical School, Boston, Massachusetts; and Information Management Services, Inc. Rockville, Maryland. Reprints: K. Robin Yabroff, PhD, MBA, Health Services and Economics Branch/Applied Research Program, Division of Cancer Control and Population Sciences, National Cancer Institute, Executive Plaza North, Room 4005, 6130 Executive Blvd., MSC 7344 Bethesda, MD yabroffr@mail.nih.gov. Copyright 2009 by Lippincott Williams & Wilkins ISSN: /09/ costs. Our findings may inform estimation of costs for other cancers as well as other diseases. Key Words: health care costs, health services research, cost and cost analysis, neoplasms, medicare, SEER program (Med Care 2009;47: S56 S63) Incidence cost estimates or longitudinal estimates of medical costs after diagnosis with disease, can be useful inputs for policy makers in decision-making about resource allocation and coverage of specific treatments. Cost-effectiveness analyses of interventions to prevent or treat disease also rely on incidence cost estimates. Sources of cost data for these estimates in the United States include insurance claims, hospitals, billing systems, hospital discharge data, and surveys. 1 3 These sources of cost data generally do not have information about the date of diagnosis with disease, however, complicating estimation of incidence costs. To address this limitation, cancer researchers have linked cancer registry data that include date of diagnosis with health insurance claims or encounter data, to estimate costs of cancer care for incident cohorts 4 or to model long-term costs of care following diagnosis. 1,2,4 7 Modeling approaches typically divide care into clinically relevant periods or phases of care the initial period after diagnosis, the end of life, and the intervening or continuing period, and allow estimation of long-term costs when applied to survival life tables. 7 This phase of care approach is a more efficient use of data than a cohort approach and can lead to more robust estimates, particularly for less common tumor sites. In the absence of registry linkage, algorithms using diagnosis codes have been developed to identify newly diagnosed cancer patients in claims data, but the sensitivity and positive predictive value of these algorithm approaches are generally poor. 8,9 In addition to issues with identifying incident cases, variation in study design may also influence the estimates of the costs of cancer care. A recent systematic review of studies of the costs of cancer care in the United States found significant heterogeneity across the studies in care settings, populations studied, types of services included, measurement of costs, and study methods. 10 Because all of these factors can affect cost estimates, this heterogeneity makes comparisons across studies difficult, even when evaluating the same type of cancer. Without comparisons using the same time period, S56 Medical Care Volume 47, Number 7 Suppl 1, July 2009

2 Medical Care Volume 47, Number 7 Suppl 1, July 2009 Comparison of Incidence Cost Approaches settings, populations, services, and measurement of costs, it is difficult to assess and understand the relative impact of different methods and data sources on estimating costs of care. In this study, we compared longitudinal costs of care after diagnosis in colorectal cancer (CRC) patients using incident cohorts from linked SEER-Medicare data and Medicare claims only, and a modeled phase of care approach using linked SEER-Medicare data. Our goal was to understand the consequences of relying on different approaches for deriving cost estimates in descriptive cost studies and cost-effectiveness analyses and inform interpretation of other studies of cancer as well as other diseases where registry systems are less common. METHODS Overview We compared 3 approaches of estimating longitudinal costs of care from diagnosis through 5 years in CRC patients: (1) a SEER-Medicare cohort identified with a registry diagnosis of CRC, (2) a cohort identified from CRC diagnosis codes in Medicare claims data, and (3) a modeled phase of care approach in patients with a registry diagnosis of CRC from linked SEER-Medicare data. We used the SEER-Medicare cohort as the reference case for comparing cost estimates because newly diagnosed CRC patients are identified by a registry, and unlike the phase of care approach, it does not rely on modeling assumptions. Selection criteria for cancer patients and controls were as similar as possible across the approaches (Table 1). Data Sources We used data from several sources: cancer registry data from the SEER program maintained by the National Cancer Institute (NCI), the SEER data linked to Medicare claims data, and Medicare claims data alone. The SEER-13 registries collect information about all incident cancer patients from geographically defined areas, including 5 states (Connecticut, Hawaii, Iowa, New Mexico, and Utah) and 6 metropolitan areas (Atlanta, Detroit, Los Angeles, San-Francisco-Oakland, San Jose-Monterey, and Seattle). These areas included approximately 14% of the US population. 11 For each patient, the SEER data contain every occurrence of a primary incident cancer, month and year of diagnosis, cancer site, stage, histol- TABLE 1. Selection Criteria by Approach to Estimating Incidence Costs of Care in Colorectal Cancer Patients, Observation Period 1998 to 2002 CRC patient identification Year CRC patients diagnosed CRC patients exclusions CRC patient followup period Years controls identified Control exclusions Control follow-up period Survival SEER-Medicare Cohort 5% Medicare Claims Only Cohort Registry Claims Registry Phase of Care* Initial Continuing Last Year of Life ICD-O codes 18.0, , 19.9, 20.9, 26.0 ICD-9 codes , and 230.3, ICD-O codes 18.0, , 19.9, 20.9, Identified Prior cancers Death certificate/autopsy diagnosis Prior identified cancers Medicare managed care or Medicare fee-forservice Part A only Prior cancers Death certificate/autopsy diagnosis Unusual histology Unusual histology Medicare managed care Medicare managed care or Medicare fee-for-service Part A only or Medicare fee-forservice Part A only Up to 60 mo following Up to 60 mo following Up to 12 mo following Up to 60 mo Up to 12 mo prior diagnosis identification diagnosis to death Prior cancers per registry, Medicare managed care or Medicare fee-forservice Part A only Up to 60 mo following identification Measured directly for CRC patients and controls Prior cancers per claims, Medicare managed care or Medicare fee-forservice Part A only Up to 60 mo following identification Measured directly for CRC patients and controls Prior cancers per registry, Medicare managed care or Medicare fee-for-service Part A only Up to 60 mo during observation period Up to 12 mo prior to death Measured directly for CRC patients and controls for monthly phase-specific cost estimates Monthly crude survival probabilities calculated from separate cohort of newly diagnosed CRC patients *Initial phase of care is the first 12 mo following diagnosis, the last year of life is the final 12 mo of life, and the continuing phase is the period between the initial phase and last year of life. Patients surviving less than 24 mo contributed months of observation and costs of care first to the last year of life phase of care, and any remaining months and costs were then assigned to the initial phase of care. Colorectal cancer cases were identified as patients with any hospital claims with diagnosis codes for colorectal cancer or 2 physician or outpatient claims with claims for colorectal cancer 60 d apart, but within 1 year. CRC indicates colorectal cancer; SEER, Surveillance Epidemiology and End Results; ICD-O, International Classification of Diseases for Oncology; ICD-9, International Classification of Diseases-ninth revision Lippincott Williams & Wilkins S57

3 Yabroff et al Medical Care Volume 47, Number 7 Suppl 1, July 2009 ogy, and vital status, with cause of death for patients who died. Some registries started collecting information about newly diagnosed cancer patients in Individuals reported to SEER have been matched against Medicare s master enrollment file and Medicare claims have been extracted for eligible persons with fee-forservice coverage. Among individuals aged 65 and older with a cancer diagnosis appearing in the SEER data, 94% have been linked with Medicare enrollment data. 12 Additionally, the NCI has created a file that identifies a 5% random sample of all Medicare beneficiaries residing in SEER areas. Because some beneficiaries in this sample have been diagnosed with cancer, the file also contains an indicator of whether or not they have a cancer diagnosis reported in SEER. This 5% sample was used as a data source for evaluating the claimsonly approach for identifying incident cancer patients and estimating costs of care. Additionally, beneficiaries without cancer in the 5% sample were used as controls in comparisons with cancer patients. A more detailed description of the linked SEER-Medicare data is available at: gov/seermedicare/. For all approaches described in this article, patient demographic characteristics and vital status were obtained from Medicare enrollment data and payments were obtained from Medicare claims. All Medicare claims files, including inpatient (Medicare Provider Analysis and Review MEDPAR ), Hospital Outpatient, Carrier, Hospice, Home Health, and Durable Medical Equipment, were used to estimate costs of care for cancer patients and controls. Part D data for oral prescription drugs were not available at the time of this study. Chemotherapy and administration for drugs administered parenterally were covered by Medicare Part B during the study period, as were Prodrugs, the oral drug equivalent of drugs administered parenterally. Longitudinal data for covered healthcare services are available from the time of a individual s Medicare eligibility until death. Cohort Identified From Linked SEER-Medicare Data Newly diagnosed CRC patients aged 65 or older in 1998 were identified from linked SEER-Medicare data and all costs of care were measured from diagnosis for up to 5 years. Patients with a prior cancer diagnosis, unusual histology, identified as having cancer through a death certificate or autopsy, or who did not have Medicare fee-for-service Part A and Part B throughout the entire observation period were excluded. The final cohort included 6377 CRC cancer patients. Potential controls were Medicare beneficiaries in the 5% sample aged 65 or above in 1998 without any cancer diagnoses recorded by a SEER registry 1973 to A total of 31,885 controls were selected from the 5% sample of Medicare enrollees and frequency matched to cancer patients by gender, 5-year age strata (65 69, 70 74, 75 79, 80 ), and SEER registry area. Cohort Identified From Diagnosis Codes in Medicare Claims Data The claims only approach used the entire 5% sample of Medicare beneficiaries (both with and without cancer) in the SEER areas to identify CRC patients and controls and measured all costs of care for up to 5-years in both groups. CRC patients aged 65 and older were identified from 1998 to 1999 claims data as either having an inpatient claim with a CRC diagnosis code or 2 outpatient claims with a CRC diagnosis at least sixty days apart, but within 1 year. Similar algorithms have been used elsewhere with claims data. 13 We identified and excluded patients with CRC before 1998 using a similar algorithm with claims from 1986 to Patients who did not have both Medicare Part A and Part B throughout the entire observation period were excluded. The final cohort included 826 CRC patients. Potential controls were Medicare beneficiaries aged 65 or above in 1998 to 1999 without any cancer diagnoses identified in the claims data from 1986 to A total of 4130 controls were frequency matched to cancer patients by gender, 5-year age strata, and SEER registry area. Modeled Phase of Care Approach Using Linked SEER-Medicare Data The modeled phase of care approach identified both newly diagnosed as well as prevalent CRC patients during the observation period, 1998 to Months of observation and costs for cancer patients were divided into clinically relevant periods: the initial period following diagnosis, last year of life, and the intervening or continuing period. As shown in Figure 1, the longitudinal monthly cost patterns for patients diagnosed with CRC and died follow a u-shape across survival cohorts. 14 The consistency of this u-shape pattern and corresponding phases across survival cohorts is a central assumption of the phase of care approach. Monthly phasespecific costs were then applied to monthly survival probabilities from a separate cohort of newly diagnosed CRC patients over a 60 month period to estimate longitudinal costs of care. Thus, the resulting year 1 and 5-year cost estimates are analogous to those of an incident cohort. To assign CRC patients and their costs to phases of care, patients diagnosed between 1973 and 2002 and aged 65 or above during 1998 to 2002, the study observation period, were selected from the linked SEER-Medicare data. Patients $18,000 $16,000 $14,000 $12,000 $10,000 $8,000 $6,000 $4,000 $2,000 $ Month Survival 48 Month Survival 60 Month Survival 72 Month Survival 84 Month Survival 96 Month Survival SOURCE: Adapted from Brown et al. Medical Care 2002;40 (supp):iii-63 III-72. FIGURE 1. Monthly costs of care for colorectal cancer patients by length of survival. S Lippincott Williams & Wilkins

4 Medical Care Volume 47, Number 7 Suppl 1, July 2009 Comparison of Incidence Cost Approaches with more than one cancer diagnosis, unusual histology, identified on death certificate or autopsy, or did not have both Medicare Part A and Part B for the entire observation period were excluded. The sample consisted of 103,068 CRC patients. Potential controls were Medicare beneficiaries aged 65 or above during the observation period, without any cancer diagnoses recorded by SEER 1973 to A total of 171,146 controls were selected from the sample of Medicare enrollees and frequency matched to cancer patients by gender, 5-year age strata, and SEER registry area. Intermediate Monthly Phase-Specific Estimates Phase of care definitions were based on prior studies of direct medical costs. 1,2,7,14,15 For cancer patients, months of observation and cost of care between 1998 and 2002 were divided into phases of care based on the dates of service on the Medicare claims in relation to dates of death and diagnosis. Date of death (or its absence) in the Medicare enrollment file through 2003 was used to determine vital status. For cancer patients, cause of death (classified as cancer, noncancer) was identified from SEER based on death certificate information. Based on prior studies, 7 the initial phase was defined as the first 12 months after diagnosis, the last year of life phase was defined as the final 12 months of life, and the continuing phase was defined as all months between the initial and last year of life phases of care. Not all patients contributed to all phases of care, however. For patients surviving less than 24 months after diagnosis, the final 12 months of observation and costs of care were allocated first to the last year of life phase, consistent with other studies, 7,16 and the application of cost estimates to survival probabilities in this study. The remainder of months of observation and costs were allocated to the initial phase, with no contribution to the continuing phase. Patients diagnosed before 1997 who survived beyond 2003 contributed only to the continuing phase. Because controls did not have a cancer diagnosis, they were randomly assigned a pseudo-diagnosis date that corresponded to the diagnosis date of one of the pool of cancer patients. Months of observation and costs of care were assigned to either the last year of life phase or the continuing phase of care in the same manner used with cancer patients. In addition to frequency matching by gender, 5-year age group, and SEER area stratum, controls were also matched to cancer patients by phase of care in up to a 1:5 case: control ratio. To reflect costs associated with cancer care in the last year of life, cancer patients who died of cancer were matched to continuing controls, and cancer patients who died of other causes were matched to last year of life controls. Within each phase of care, we calculated total costs of care and total months of observation for cancer patients and controls. Mean monthly costs of care by phase of care were then estimated, and net costs were estimated as differences in costs between cancer patients and controls. Modeled Longitudinal Cost Estimates To calculate longitudinal costs, monthly cost estimates in the initial, continuing, and last year of life phases were applied to crude monthly survival probabilities calculated from a separate cohort of newly diagnosed CRC patients. Five-year crude survival probabilities were calculated among CRC patients aged 65 and older diagnosed 1998 to 2004 using SEER-Stat software version For example, the proportion of patients dying of cancer in month 30 after diagnosis was multiplied by 12 months of initial phase costs following diagnosis, then 6 months of continuing phase costs, and 12 months of cancer-death last year of life costs. Analyses In the SEER-Medicare and Medicare claims only cohort approaches, costs of care were calculated for each year after diagnosis or identification for cancer patients and controls, as well as for the 5-year period (up to 60 months) after diagnosis. Mean net costs were calculated as the difference in total costs between cases and controls. In the phase of care approaches, mean year 1- and 5-year total and net costs were modeled by combining monthly phase-specific cost estimates with crude monthly survival probabilities. Standard errors and 95% confidence intervals were calculated for cost estimates. In the phase of care approaches, we calculated a plausible range of costs for year 1 and the 5-year period using the upper and lower 95% confidence interval for phasespecific costs. We used Medicare payment variables, rather than billed charges, to reflect costs of care. Charges reflect price-setting rather than resource consumption, and as a result, are thought to be a poor proxy of the true economic cost of medical care. 18 Payments for Medicare Part A (inpatient services) and Part B (outpatient services) were calculated separately. The Hospital Input Price Index 19 and the Medicare Economic Index 20 were used to adjust for inflation in Medicare Parts A and B estimates, respectively, during 1998 to We also adjusted for geographic variability in costs of care across SEER registry sites, using the Capital Geographic Adjustment Factor for Part A and the Geographic Practice Cost Index for Part B. 14 Costs were discounted at 3% annually, and estimates are reported in 2004 dollars. Sensitivity Analysis of Phase of Care Approach We conducted sensitivity analysis to assess the impact of our phase definitions on longitudinal cost estimates. We calculated monthly costs in the 24 months following diagnosis and the 24 months before death among the 2 subgroups of our phase of care sample who were (1) diagnosed in 1998 to 2002 and (2) died in 1998 to 2002, respectively. The monthly costs were log-transformed to provide a more normal distribution and then Joinpoint analyses were conducted to identify the month where the slope of the costs of care changed, indicating the end of a phase of care. 21 Statistically significant changes in monthly cost were identified at month 8 after diagnosis and 8 months before death, corresponding the initial phase of care and the end of life phase. The continuing phase of care consisted of all the months between the initial and end of life phase, as in the base case analysis. Monthly cost estimates by phase of care were applied to the same crude monthly survival probabilities described above to model longitudinal costs of care Lippincott Williams & Wilkins S59

5 Yabroff et al Medical Care Volume 47, Number 7 Suppl 1, July 2009 TABLE 2. Sample Characteristics by Approaches to Estimating Incidence Costs For Colorectal Cancer Patients, Observation Period 1998 to 2002 SEER-Medicare Cohort 5% Medicare Claims Only* Cohort Phase of Care Last Year of Life Died of Other Causes Initial Continuing Died of Cancer No. CRC patients 6, ,689 81,538 18,185 22,034 No. controls 31,885 4, , ,614 90,925 44,068 Mean Age at diagnosis/identification (yr) Gender Male 45.2% 41.1% 45.8% 48.0% 48.9% 46.8% Female 54.8% 58.6% 54.2% 52.0% 51.2% 53.2% Stage distribution at diagnosis In situ 4.9% 5.8% 8.1% 2.6% 7.7% Localized 37.0% N/A 45.1% 48.6% 23.9% 47.7% Regional 35.5% 38.9% 35.6% 37.2% 33.9% Distant 16.6% 7.3% 2.7% 28.5% 3.7% Unstaged/missing 6.0% 2.9% 4.9% 1.0% 2.9% Mean/median survival months following diagnosis or identification during CRC patients 37.7/ / / / / /100.0 Controls 50.1/ / / / / /81.0 Mean months in phase CRC patients N/A N/A Controls *Colorectal cancer patients were identified as patients with any hospital claims with diagnosis codes for colorectal cancer or 2 physician or outpatient claims with claims for colorectal cancer 60 d apart, but within 1 yr. Initial phase of care is the first 12 mo following diagnosis among patients diagnosed, the last year of life is the final 12 mo of life among patients diagnosed, and the continuing phase is the period between the initial phase and last year of life among patients diagnosed. Patients surviving less than 24 mo contributed months of observation and costs of care first to the last year of life phase of care, and any remaining months and costs were then assigned to the initial phase of care. CRC indicate colorectal cancer; SEER, Surveillance Epidemiology, and end results. RESULTS Mean age at diagnosis was similar in the SEER-Medicare and Medicare claims only cohorts and in the initial phase of care sample (Table 2). Mean age at diagnosis was lower in the continuing phase of care, likely reflecting long-term survivors diagnosed at younger ages. The proportion of localized cases varied between the SEER-Medicare cohort (37.0%) and initial and continuing phases in the phase of care approaches (45.1% and 48.6%, respectively). Mean survival in cancer patients was lower than in controls in the SEER- Medicare and Medicare claims only cohorts and in the initial phase of care. Deaths in year 1 following diagnosis varied by approach: 25.6% of CRC patients died in the SEER-Medicare reference; 14.8%, in the Medicare claims only; and 24.3%, in the modeled phase of care (data not shown). Intermediate Monthly Phase-Specific Estimates In the initial and continuing phases of care for the base case analysis, mean net monthly costs were $2682 and $155, respectively (Table 3). In the last year of life, mean net monthly costs were $3760 for patients who died of cancer and $547 for patients who died of other causes. In the sensitivity analysis, mean net monthly costs were $3779 and $182 in the initial and continuing phases of care, respectively, and $4429 and $833 in the last year of life for patients who died of cancer and of other causes, respectively. Comparison of Total and Net Costs of Care Across Approaches Costs in the first year were the majority of the 5-year costs (Table 4). Mean year 1 net costs were $32,648, $27,310, and $27,575 in the SEER-Medicare cohort, Medicare claims only cohort, and the base case SEER-Medicare modeled phase of care approaches, respectively. Compared with the SEER-Medicare cohort estimate, the Medicare claims only cohort and both the base case and the sensitivity analysis modeled phase of care approach underestimated net costs in year 1 by about 16%. Mean net 5-year costs of care estimates were similar in the SEER-Medicare cohort and the base case modeled phase of care approach ($37,041 and $37,652, respectively), whereas the 5-year costs of care in the Medicare claims cohort were about 18% lower ($30,256). Mean year 1 and 5-year net costs of care in the base case and sensitivity analysis of the modeled phase of care approach were almost identical. Differences between the SEER-Medicare and Medicare claims only cohorts were mainly among patients in the youngest age group (Figs. 2A, B). Net costs in the first year S Lippincott Williams & Wilkins

6 Medical Care Volume 47, Number 7 Suppl 1, July 2009 Comparison of Incidence Cost Approaches TABLE 3. Intermediate* Monthly Phase-Specific Costs of Care in Colorectal Cancer Patients Total Costs Estimate ($) 95% CI ($) Net Costs Estimate ($) 95% CI ($) Phase of care base case Initial 3123 (3090,3155) 2682 (2649,2715) Continuing 583 (577,589) 155 (148,162) Last year of life Cancer death 4214 (4155,4273) 3760 (3701,3819) Other death 3319 (3262,3377) 547 (479,614) Phase of care sensitivity analysis Initial 4238 (4197,4280) 3779 (3737,3821) Continuing 625 (619,631) 182 (175,189) Last year of life Cancer death 4902 (4831,4974) 4429 (4358,4501) Other death 4100 (4027,4173) 833 (737,929) All estimates in 2004 dollars. *Monthly phase of care estimates applied to monthly survival probabilities to estimate longitudinal costs of care. following diagnosis in the age group were $32,971 in the SEER-Medicare cohort and $19,463 in the Medicare claims only cohort, a 41.0% difference. The net 5-year costs in the 65 to 69 age group were also about 40% higher in the SEER-Medicare reference cohort (data not shown). Differences in year 1 and 5-year costs were smaller in the other age groups. DISCUSSION In this study, we compared 3 approaches to estimating the incidence costs of care in elderly CRC patients. Although we used the same setting, observation period, and similar sample definitions, we found significant variation in mean year 1 and 5-year total and net cost estimates. Our findings suggest that different methods and data sources for estimating costs result in different descriptive cost estimates and variability in estimates of cost-effectiveness and their interpretation. For example, in 2004, approximately 87,000 Medicare beneficiaries aged 65 and older in the United States were diagnosed with CRC. 17 Applying the mean net five-year cost estimates would yield approximately $3223 million and $2632 million, respectively, with the SEER-Medicare and Medicare claims only approaches, a $591 million difference. The aggregate estimate developed with the SEER-Medicare modeled phase of care approach ($3276 million) was more similar to that developed with the SEER-Medicare reference cohort. Does a gold-standard data source and method for estimating incidence costs of cancer care exist? Although many dimensions of the underlying data sources and methods should be considered, including generalizability, completeness of cost data, and feasibility, the validity of the cancer diagnosis and diagnosis date is critical. The misclassification of prevalent cancer cases using claims algorithms reported elsewhere 8,9 also seems to have a significant impact on cost of care estimates, particularly in younger age-groups, where fewer years of claims are available to include and exclude prevalent CRC patients from an incident cohort. Our findings suggest that descriptive cost studies and cost-effectiveness analyses that rely on claims algorithms only to identify CRC patients will underestimate cancer-related incidence costs. Misclassification may be even greater in studies that use fewer years of data to identify and exclude prevalent CRC patients. Because the longitudinal costs of cancer care follow a u-shaped curve with the highest costs in the initial period following diagnosis and in the end of life period, with the lowest costs in the period in-between (Fig ), inclusion of prevalent cases whose costs are lowest (the bottom of the u) will underestimate the true incidence costs of cancer care. Thus, the utility of claims algorithms without linkage to registry data is limited for estimating incidence costs of cancer care. The utility of claims algorithms for identifying incidence cohorts in estimating costs of care for other diseases is unknown, but likely to vary based on the shape of the longitudinal cost curve following diagnosis. The relative impact of contamination by prevalent cases on cost estimates may be minimized in diseases with flatter cost curves. Use of the modeled phase of care approach will also result in different estimates of costs relative to the SEER- Medicare cohort in the first year following diagnosis, although 5-year net costs were remarkably similar with the 2 approaches ($37,652 versus $37,041). These approaches differ in the number of years used to identify cancer patients and in the calculation of net costs in cancer patients who died during the period. Cost estimates in the first year following diagnosis were based on patients diagnosed in 1998 in the cohort, and on patients diagnosed in 1997 to 2002 in the modeled phase of care approach. During the later part of this interval, CRC screening increased in the United States, 22,23 and any shifts towards earlier stage at diagnosis will make modeled phase estimates lower, because costs of care are lower in patients diagnosed at earlier compared with later stage of disease. 7 Additionally, CRC survival improved slightly over this period. 11 Differences in the methods for estimating net costs in cancer patients who died will also make modeled phase estimates lower than the cohort approach. In the SEER-Medicare cohort approach, costs in cancer patients are compared with costs in controls as they occur, with cancer patients being more likely to die during the 5-year period, with the largest difference in year 1 after diagnosis. In the phase of care of approach, costs in cancer patients dying of cancer are compared with continuing control costs. Costs in cancer patients dying of other causes are compared with non-cancer patients who were also dying. This will understate costs among those patients dying compared with the cohort approach, particularly in intervals where more patients die (ie, year 1 after diagnosis). Our findings suggest that the phase of care cost approach may best approximate incidence estimates over longer time periods, and may be most useful for estimating costs in rare cancers, cancers where treatment patterns are rapidly 2009 Lippincott Williams & Wilkins S61

7 Yabroff et al Medical Care Volume 47, Number 7 Suppl 1, July 2009 TABLE 4. Comparison of Approaches for Estimating Incidence Costs of Care in Colorectal Cancer Patients, Observation Period 1998 to 2002 Estimate Mean Total Costs 95% CI/Range* Mean Net Costs Difference From Reference Estimate 95% CI/Range* Difference From Reference Yr 1 costs of care SEER-Medicare incident cohort $39,381 ($38,578,$40,185) $32,648 ($31,826,$33,470) Medicare claims incident cohort $36,092 ($33,852,$38,333) $3289 $27,310 ($25,001,$29,619) $ % 16.4% SEER-Medicare modeled phase of $33,233 ($32,852,$33,605) $6148 $27,575 ($27,197,$27,968) $5073 care base case 15.6% 15.5% SEER-Medicare modeled phase of care sensitivity analysis $32,768 ($32,427,$33,126) $6613 $27,264 ($26,902,$27,628) $ % 16.5% 5-yr costs of care SEER-Medicare incident cohort $65,554 ($64,113,$66,994) $37,041 ($35,525,$38,558) Medicare claims incident cohort $70,029 ($65,901,$74,157) $4475 $30,256 ($25,841,$34,673) $6, % 18.3% SEER-Medicare modeled phase of $58,081 ($57,390,$58,765) $7473 $37,652 ($36,926,$38,394) $611 care base case 11.4% 1.7% SEER-Medicare modeled phase of care sensitivity analysis $57,965 ($57,350,$58,657) 7589 $37,846 ($37,133,$38,562) $ % 2.2% All estimates in 2004 dollars. Five-year estimates discounted at 3% annually. *Range based on upper and lower 95% CI of phase specific cost estimates applied to survival estimates. SEER-Medicare cohort approach defined as reference. A Mean Cost of Care (in 2004 Dollars) B Mean Cost of Care (in 2004 Dollars) 50,000 40,000 40,000 30,000 30,000 20,000 SEER-Medicare Medicare Claims Only 20,000 SEER-Medicare Medicare Claims Only 10,000 10, Age at diagnosis Age at diagnosis FIGURE 2. Year 1 mean total (A) and net (B) costs of care in colorectal cancer patients by age of diagnosis. changing, and for longer time periods. Notably, total and net costs in the sensitivity analysis of the phase of care approach, which based the definition of the initial and end-of-life phases on Joinpoint analyses of changes in monthly costs, were almost identical to those in the base case phase of care approach. The usefulness of the phase of care approach for other conditions relies on the shape of cost curve following diagnosis. Evaluation of additional phases (eg, preterminal phase) and comparison of the phase definition across cancer sites will be an important area for additional research. This approach has been applied to measurement of health limitations and quality of life, 24 but to our knowledge, the phase of care approach has not been applied to estimating longitudinal costs of care for other conditions. In a companion paper, we compared prevalence approaches to estimating mean total and net annual and 5 year costs of CRC using linked SEER-Medicare data and Medicare claims only. 25 Compared with our 5-year mean net incidence costs estimated from the SEER-Medicare reference cohort, 5 years of mean net annual prevalence costs from the SEER-Medicare reference were much lower. Differences between incidence and prevalence estimates may be smaller for other cancers with short survival duration, such as gastric, lung, and pancreas. There were several limitations that impact all of our cost estimates. Our Medicare cost estimates are based on Medicare payments, and do not include costs for services reimbursed by other sources, including Medicaid, private S Lippincott Williams & Wilkins

8 Medical Care Volume 47, Number 7 Suppl 1, July 2009 Comparison of Incidence Cost Approaches supplemental insurance, and the Veterans Administration. Estimates do not include out of pocket expenses or copayments. Medicare payments have been reported to represent approximately 51% to 65% of all health care costs, including other payors and out-of-pocket expenses Our estimates were based on the approximately 85% of Medicare enrollees in fee-for-service plans. CRC stage at diagnosis 29 and survival 30 have been reported elsewhere to vary between Medicare managed care and fee-for-service settings. Exploration of potential selection biases and differences in the costs of cancer care by delivery setting will be an important area for additional research. In summary, incidence approaches to the estimation of disease-specific medical costs can be useful for descriptive and analytical purposes. However, such estimates can vary substantially depending on the method and source of data that is used to identify newly diagnosed patients, date of diagnosis, and cost elements. We found that using Medicare claimsonly to identify incidence cases may result in substantial underestimations of both short- and longer-term costs. Longer-term, net estimates of cost from the SEER-Medicare modeled phase-of-care approach are comparable to estimates obtained directly from a SEER-Medicare cohort. REFERENCES 1. Riley GF, Potosky AL, Lubitz JD, et al. Medicare payments from diagnosis to death for elderly cancer patients by stage and diagnosis. Med Care. 1995;33: Taplin SH, Barlow W, Urban N, et al. Stage, age, comorbidity, and direct costs of colon, prostate, and breast cancer care. J Natl Cancer Inst. 1995;87: Howard DH, Molinari N-A, Thorpe KE. National estimates of medical costs incurred by nonelderly cancer patients. Cancer. 2004;100: Barlow WE, Taplin SH, Yoshida CK, et al. Cost comparison of mastectomy versus breast-conserving therapy for early stage breast cancer. J Natl Cancer Inst. 2001;93: Fishman P, Von Korff M, Lozano P, et al. Chronic care costs in managed care. Health Aff. 1997;16: Fireman BH, Quesenberry CP, Somkin CP, et al. Cost of care for cancer in a health maintenance organization. Health Care Financ Rev. 1997; 18: Yabroff KR, Lamont EB, Mariotto A, et al. Cost of care for elderly cancer patients in the United States. J Natl Cancer Inst. 2008;100: Warren JL, Feuer E, Potosky AL, et al. Use of Medicare hospital and physician data to assess breast cancer incidence. Med Care. 1999;37: Cooper GS, Yuan Z, Stange KC, et al. The sensitivity of Medicare claims data for case ascertainment of six common cancers. Med Care. 1999;37: Yabroff KR, Warren JL, Brown ML. Costs of cancer care in the USA: a descriptive review. Nat Clin Pract Oncol. 2007;4: Ries LAG, Harkins D, Krapcho M, et al. SEER Cancer Statistic Review, Bethesda Maryland: National Cancer Institute; Warren JL, Klabunde CN, Schrag D, et al. Overview of the SEER-Medcare data: content, research applications, and generalizability to the United States elderly population. Med Care. 2002;40 suppl :IV-3 IV CMS Chronic Conditions Data Warehouse. Available at: /about php Brown ML, Riley GF, Schussler N, et al. Estimating health care costs related to cancer treatment from SEER-Medicare data. Med Care. 2002;40(suppl): Warren JL, Brown ML, Fay MP, et al. Costs of treatment for elderly women with early-stage breast cancer in fee-for-service settings. J Clin Oncol. 2001;20: Brown ML, Riley GF, Potosky AL, et al. Obtaining long-term disease specific costs of care: application to Medicare enrollees diagnosed with colorectal cancer. Med Care. 1999;37: SEER-Stat. Available at: ; February 20, Finkler SA. The distinction between cost and charges. Ann Intern Med. 1982;96: The Boards of Trustees FHIaFSMIF. Hospital Wage Index. Available at: gov/acuteinpatientpps/wifn/list asp?listpage Medicare Economic Index. RatesStats/downloads/mktbskt-economic-index pdf Kim HJ, Fay MP, Feuer EJ, et al. Permutation tests for join point regression with application to cancer rates. Stat Med. 2000;19: (correction: 2001;20:655). 22. Meissner HI, Breen N, Klabunde CN, et al. Patterns of colorectal cancer screening uptake among men and women in the United States. Cancer Epidemiol Biomarkers Prev. 2006;15: Breen N, Wagener DK, Brown ML, et al. Progress in cancer screening over a decade: results of cancer screening from the 1987, 1992, and 1998 national health interview surveys. J Natl Cancer Inst. 2001;93: Yabroff KR, McNeel TS, Waldron WR, et al. Health limitations and quality of life associated with cancer and other chronic diseases by phase of care. Med Care. 2007;45: Yabroff KR, Warren JL, Banthin J, et al. Comparison of approaches for estimating prevalence costs of care for cancer patients: what is the impact of data source? Med Care. 2009;47(suppl):S64 S Hackbarth GM. Statement to the House Subcommittee on Health. Medicare cost-sharing and supplemental insurance. Available at: Last accessed April 30, Crystal S, Johnson RW, Harman J, et al. Out-of-pocket health care costs among older Americans. J Gerontol B Psychol Sci Soc Sci. 2000;55: S51 S AARP Public Policy Institute. What share of beneficiaries total health care costs does medicare pay? Riley GF, Potosky AL, Lubitz JD, et al. Stage of cancer at diagnosis for Medicare HMO and fee-for-service enrollees. Am J Public Health. 1994;84: Merrill RM, Brown ML, Potosky AL, et al. Survival and treatment for colorectal cancer Medicare patients in two group/staff health maintenance organizations and the fee-for-service setting. Med Care Res Rev. 1999;56: Lippincott Williams & Wilkins S63