Estimates of the cost of treating common illnesses are important for allocating prevention and research funds optimally,1 for understanding the impact of each disease on federal and state budgets, and for quantifying the costs of disease-specific insurance expansions and programs. Existing estimates of the cost of cancer care in the U.S. rely on several different approaches. Although total spending estimates are similar, each approach entails strong assumptions and may produce inaccurate estimates for subpopulations of interest.

Brown and Fintor2 calculated the quantity of services consumed by cancer patients using site-specific surveys, such as the National Hospital Discharge Survey, then multiplied the share of services (e.g., hospital days, office visits) attributable to cancer by the corresponding expenditure total for the service category in the National Health Accounts. Summing over service types produces an estimate of the total cost of cancer care: $27 billion in 19902 and $41 billion in 1995.3 This method assumes that per-day or per-visit reimbursements for cancer services do not differ from payments for other types of care. It also assumes that all services related to cancer care are coded as such, although cancer patients experience symptoms and side effects that may require treatment outside of visits for curative care.

A second method of estimating the cost of cancer care is described by Brown et al.4 First, using Medicare claims data, by cancer type, those authors estimate the cost of three distinct phases of cancer care: the initial care phase, the continuing care phase, and the terminal care phase. Second, they classify patients in the SEER data according to cancer phase and cancer type by month. Finally, they impute phase-specific costs to each patient; sum over patients; and, after a number of adjustments, arrive at a national, annual estimate of about $40 billion (1996 U.S. dollars). An important assumption embedded in this final step is that Medicare expenditures and reimbursements for cancer care are a good proxy for spending by and payments from private payers. However, private payers generally offer more generous benefits than Medicare,5 and reimbursement rates tend to be higher.6, 7

We undertook the current study to estimate the costs of medical care for nonelderly cancer patients. Our interest in this population stems from a previous study by two of us (D.H.H. and K.T.). In that study, we found that ≈ 11% of cancer patients age < 65 years were uninsured at the time of diagnosis and that spending and use of services were lower for those patients.8 Estimates of spending in the nonelderly population are necessary for estimating the cost of programs to extend insurance coverage to uninsured cancer patients and the resulting cost savings to private plans (public insurance programs usually cover many individuals who had private insurance prior to enrollment9). Although the close agreement between the cancer-attributable spending estimates of Brown and coauthors that were produced using two very different methods increases confidence in their validity, it is unclear whether they produce reliable estimates of the cancer expenditures incurred by the nonelderly population and by employer-sponsored health insurance plans. In addition, insurance programs for cancer patients necessarily must pay for some or all services unrelated to cancer treatment; therefore, it is important to have estimates of total spending by cancer patients, not just cancer-attributable spending.

The current study also provided an opportunity to evaluate different methodologies and data sources for estimating medical costs. We compared our estimates with those of Brown and coauthors, and we compared estimates based on methods that make use of the diagnostic information on claims and encounters with estimates based on differences in spending between individuals with and without cancer. We assessed the sensitivity of cost estimates to the underlying data source by applying each method for producing estimates to Medical Expenditure Panel Survey (MEPS)10 data and, separately, to the Medstat Marketscan data base. The MEPS is a nationally representative survey that contains detailed information on participants' utilization and medical costs. The Medstat Marketscan data base is a large, national insurance claims file that is compiled from large employers who typically offer very generous insurance coverage.


  1. Top of page


The MEPS collects information on insurance status and on use of and spending on health care services for all members of surveyed households. Every year, the Agency for Healthcare Research and Quality and the National Center for Health Statistics select the MEPS panel by drawing a sample of households from those that participated in the prior year's National Health Interview Survey (NHIS). The sample is representative of the civilian, noninstitutionalized U.S. population, except that minorities are over sampled (note that this population does include veterans but does not include active-duty military personnel). Respondents are followed for 2 years. Our data include 2 years of observations for the 1996, 1997, and 1998 panels and 1 year of observations for the 1999 panel. Summarizing response rates is somewhat difficult, because there are four panels, each drawn from respondents to the NHIS, and multiple rounds within each panel. Taking account of the 1995 NHIS response rate, 70.2% of MEPS-eligible households provided data for all of the first three rounds of 1996.11

MEPS data on utilization are self-reported. Reimbursement amounts and diagnosis codes for these self-reported encounters (“encounters” is a catch-all term that includes provider visits and prescriptions) are collected directly from providers, presenting an opportunity to verify that reported encounters actually occurred. Comparisons of MEPS data with encounter totals from other national surveys do not indicate an under-reporting problem.12 Comparisons of spending totals indicate a potential under-reporting problem for expenditures by Medicaid and Medicare, but not for private health insurance.12

The MEPS data may not fully capture costs related to end-of-life care. Individuals in institutions (for example, a hospice) are excluded from the survey. In addition, MEPS interviewers may have trouble obtaining utilization data for deceased respondents, who obviously cannot self-report this information. The MEPS interviewers attempt to collect end-of-life care data from other members of the respondent's household. Unlike claims data bases, MEPS includes information on out-of-pocket spending and encounters that did not generate insurance claims. Like claims data bases, MEPS does not record uncompensated care or unpaid bills (i.e., “bad debt”).

The Medstat Marketscan 1999 data base is a large national insurance claims file compiled from 45 large, self-insured employers. It does not include Medicare-eligible individuals, Medicaid recipients, or the uninsured, whose costs and utilization may differ substantially from those of individuals with employer-based coverage. The Medstat Marketscan data base, as an insurance claims data base, includes all claims approved and paid by the insurer for covered services, including hospital admissions, outpatient visits, and prescription drugs. The Medstat data base also includes costs for end-of-life care, such as nursing home care and hospice care. It does not include costs for encounters for which no insurance claims were submitted. Each claim includes the total amount paid to the provider, any copayment or deductible paid by the insured, and the net amount paid by the insurer. Inpatient admission claims, which include any hospital visit that involves an overnight stay, include principal diagnosis, principal procedure, and diagnostic-related group. Outpatient claims include primary and secondary diagnoses and procedure performed. Pharmaceutical claims do not include any diagnostic information.

Sample Selection and Variable Construction

From the MEPS data, we extracted records for individuals with at least 1 health care encounter with a clinical classification code for cancer (codes 11–43) who were age < 65 years upon enrollment in the MEPS. Clinical classification codes are groups of similar International Classification of Diseases, Ninth Revision (ICD-9) codes and are used in the MEPS to make it easier for researchers to identify episodes of care and patients with specific diseases. For each individual, we computed total spending (including both cancer encounters and noncancer encounters) and spending by site of care and payer type. We also computed spending on only those encounters with at least one clinical classification code for cancer.

The study sample from the Medstat Marketscan data base is comprised of individuals age < 65 years with at least 1 claim that includes a cancer diagnosis in their outpatient claims file. The ICD-9 codes 140.0–208.91, indicating primary or secondary malignant neoplasm, Hodgkin disease, sarcoma, lymphoma, or leukemia, were used to identify cancer diagnoses. Carcinoma in situ and neoplasms of uncertain behavior were not included. (Note that benign neoplasms of the brain were not included despite the inclusion of these tumors in National Program of Cancer Registries in 2002.) Once cancer patients were identified, all of these patients' claims generated in 1999 were compiled using patient identification numbers.

Statistical Methods

For each individual, we computed total annual spending and total spending on encounters with cancer diagnosis codes. We also computed spending by type of service (inpatient, outpatient, and prescription drug) and by payment source. These variables were used to describe individual-level spending distributions. Cost estimates from MEPS were weighted to the national level using the population weights provided in the MEPS. Cost estimates from the Medstat Marketscan data base were weighted to the national level using weights constructed from Census Bureau population totals by 5-year age group and gender. All cost figures were inflated to 2001 levels using the private health insurance component of the National Health Accounts.13

To calculate national spending estimates, we summed the weighted individual-level measures. Confidence intervals for the MEPS estimates, which account for fact that the probability of being included in the MEPS sample varies across individuals, were produced using the Svytotal command in Stata (Stata Corporation, College Station, TX). Confidence intervals from the Medstat Marketscan estimates were produced similarly with correction of standard errors to account for the probability of inclusion in the data base.

The weighted sum of spending on encounters recording a cancer diagnosis code is an estimate of cancer-attributable spending. However, estimates of cancer-attributable spending based on the identification of cancer-related encounters are subject to coding errors. For example, the cost of a physician visit to treat the side effects of chemotherapy should be counted as a cancer cost but may be coded as a visit for nausea. Alternatively, the cost of a physician visit during which the physician treats both cancer and diseases or conditions unrelated to cancer should not be counted entirely as a cancer expenditure. In the absence of a detailed claims audit, it is impossible to determine whether miscoding biases cancer cost estimates upward or downward. For this reason, statistical models that do not rely on diagnosis or procedure coding and, instead, compare costs between diseased and nondiseased individuals often are used to estimate disease-attributable costs.14

Using the MEPS, we reestimated cancer-attributable spending using a two-part model of medical costs. The sample for this analysis included all nonelderly respondents to the MEPS panels from 1996 through 1999, for a total of 56,021 observations. We adjusted for individuals who died during the study period (n = 219 patients; 0.4%) or who failed to complete all interviews (n = 4104 patients; 7.3%), consistent with an intent-to-treat approach to evaluating the impact of cancer on costs.

The two-part model is comprised of a first-stage logistic regression model in which the dependent variable equals 1 if the individual had strictly positive cost levels for the relevant category (inpatient, outpatient, and prescription drug), and the independent variables include a dummy variable for cancer (equal to 1 if the individual had cancer during their inclusion in the MEPS survey period and 0 otherwise) and controls for age group, race, and region. Brown et al.4 used age and race to match cancer cases with controls. The second stage is an ordinary least-squares regression on the subsample with strictly positive cost levels, in which the dependent variable is the logarithm of costs for the relevant category, and the control variables are the same as in the first stage. We estimated the two-part model for total medical expenditures over the entire MEPS observation period, inflated to 2001 levels, and total expenditures for inpatient care, outpatient care, and prescription drugs. Because, not surprisingly, all cancer patients incurred some medical costs, we omitted the dummy variable for cancer from the first stage for total expenditures.

We calculated two predicted values for each individual: For the first, the cancer dummy variable was set equal to 1; and, for the second, the cancer dummy variable was set equal to 0. We then transformed these predicted values to constant dollars (the second-stage dependent variable is the logarithm of spending) using the Duan smearing estimator.15 We estimated the per-person increase in medical spending attributable to cancer by subtracting the second predicted value from the first and then calculating the weighted average across the entire sample (using the sample weights in MEPS). This procedure nets out the impact of the covariates (age, gender, region) on medical spending. To calculate total cancer costs, we estimated the total number of cancer patients under treatment nationally by summing the weights for all individuals with cancer in the sample and multiplied this figure by the estimated per-person increase in medical spending attributable to cancer.

Predictions from regression models with logged dependent variables can produce misleading inferences when the error term is heteroskedastic.16, 17 A modified two-part model of spending, proposed by Mullahy,17 was estimated to determine the sensitivity of results to the specification of the second-stage regression. The second stage of the modified two-part model is a nonlinear regression in which spending equals the exponentiated value of a linear combination of parameters and independent variables. Predicted values from the modified two-part model are computed directly, without the necessity of retransforming log dollars to constant dollars.

Confidence intervals for predicted values from the standard and modified two-part models were computed by bootstrapping: The first-stage and second-stage coefficients were drawn from their respective multivariate normal distributions, and predicted values were computed following the steps outlined above. Repeating this routine 1000 times produced distributions of predicted values, and the lower and upper bounds of the confidence intervals were set equal to the 2.5th percentiles and the 97.5th percentiles of the distributions, respectively.

The Medstat Marketscan data base contains claims from such large numbers of nonelderly individuals (n = 3.5 million) that it is not necessary to use regression-based methods to compare costs between cancer patients and individuals without detected cancer. Instead, we used a “deviations from the mean” or fixed-effects approach to compute cancer-attributable spending. Using the entire nonelderly claims data base for 1999, which represented approximately 3.5 million covered lives, average expenditures per covered life were calculated for each age and gender category. Cancer-attributable spending for each cancer patient was defined as the individual's expenditures minus mean expenditures per covered life for the individual's age and gender category. Confidence intervals were calculated with appropriate corrections to standard errors.


  1. Top of page

The MEPS sample includes 842 nonelderly cancer patients receiving treatment for cancer for a total of 1042 patient years (counting only those years in which patients incurred positive levels of cancer-related expenditures). Hereafter, we refer to the samples from each data base of patients receiving treatment for cancer as the “cancer” samples and patients receiving treatment for cancer as “cancer” patients, recognizing that some of these patients are cured and are receiving follow-up care. The sample represents 10 million patients over 4 years. Some patients received cancer treatment in both years of the MEPS observation period; taking this into account, we estimate that 2.9 million individuals receive some treatment for cancer annually. According to the Cancer Prevalence Database maintained by the Statistical Research and Applications Branch of the National Cancer Institute, 3.1 million nonelderly individuals were alive on January 1, 1999 who had been diagnosed with cancer in the previous 20 years.18 The 1999 Medstat Marketscan data base includes 41,756 individuals who are receiving some cancer treatment, representing 2.2 million individuals annually.

Table 1 displays raw and weighted summary statistics of the cancer samples. The average age of individuals in the MEPS cancer sample is 47 years. Of these, 80% of individuals are non-Hispanic whites, 75% have private insurance, and 32% are poor (income < 200% of the Federal Poverty Level). The most common type of cancer in the MEPS cancer sample is breast cancer (11%), followed by melanoma (5%), and prostate cancer (5%). Oversampling of minorities by MEPS is reflected in the differences between the second column and the third column, which adjusts for differences in the probability of being included in the MEPS sample.

Table 1. Sample Characteristics of Nonelderly Cancer Patients, Ages Birth–64 Years, in the Medical Expenditure Panel Survey, 1996–1999, and in the Medstat Marketscan Data Base, 1999
CharacteristicMEPS, 1996–1999 (%)Medstat Data Base, 1999 (%)
Unweighted meanWeighted meanUnweighted meanWeighted mean
  • MEPS: Medical Expenditure Panel Survey; NA: not available.

  • a

    Poor was defined as a family income below 200% of the Federal Poverty Level.

Age (yrs)46.9746.2850.5448.04
Unknown region3388
Privately insured7578100100
Colon cancer3376
Lung cancer4498
Breast cancer11101817
Prostate cancer5422
Total no.84210,039,29641,7562,212,392

The average age of individuals in the cancer sample from the Medstat Marketscan data base is 50.5 years. By definition, all individuals in the sample have private insurance. Compared with the MEPS cancer sample, the Medstat cancer sample has a higher proportion of males (43% vs. 34%) and a lower proportion of individuals who live in the West census region (4% vs. 21%). The proportion of individuals who live in urban areas in each sample is nearly equal (75%).

Table 2 shows individual-level distributions for total annual expenditures, out-of-pocket expenditures, and out-of-pocket shares of total expenditures. These distributions reflect spending by all patients undergoing treatment for cancer. Spending distributions for patients with newly diagnosed cancer most likely would entail a higher mean and variance. We present distributions for individuals incurring total annual expenditures in excess of $1000, thereby eliminating individuals from the distributions whose only cancer-related care consisted of routine follow-up visits, to approximate the spending distributions for patients undergoing active treatment for cancer. According to the MEPS estimates, the median annual expenditure among all cancer patients is $1908 and $5692 for those incurring expenditures of at least $1000. At least 10% of patients face out-of-pocket expenditures in excess of $1000. However, given the high total spending by this population, out-of-pocket spending levels are relatively low. Compared with spending by the MEPS cancer sample, spending by the Medstat Marketscan cancer sample is higher at all points in the distribution. This result suggests that differences in the total spending estimates produced using MEPS and Medstat Marketscan are not driven by a few high-cost patients in the Medstat Marketscan sample. It is interesting to note that out-of-pocket expenses for cancer patients in the Medstat Marketscan sample, all of whom have relatively generous employer-sponsored health insurance, are generally higher in absolute terms compared with out-of-pocket expenses for patients in the MEPS sample.

Table 2. Distributions of Total Expenditures and Out-Of-Pocket Expenditures of Individual Cancer Patients Ages Birth–64 Years in the Medical Expenditure Panel Survey, 1996–1999, and the Medstat Marketscan Data Base, 1999
PercentileMEPS, 1996–1999Medstat Data Base, 1999
TotalCancer encounters onlyTotalCancer encounters only
  • MEPS: Medical Expenditure Panel Survey.

  • a

    Patients who spent > $1000 for the category.

Total spending        
Out-of-pocket spending        
Out-of-pocket share        

Table 3 displays MEPS-based estimates of aggregate annual medical spending by nonelderly patients undergoing treatment for cancer. The second column shows estimates of total medical spending, including spending on care unrelated to cancer treatment. According to these estimates, cancer patients incur medical expenditures of > $30 billion annually, including approximately $20 billion from private insurance sources and approximately $2.3 billion paid out-of-pocket. The third column shows cancer-attributable spending, which was computed by summing costs for encounters that listed a clinical classification code for cancer. Cancer-attributable spending computed by this method is just over $20 billion, including $13 billion from private sources and nearly $1.3 billion paid out-of-pocket. The fourth and fifth columns display cancer-attributable spending estimates from the standard two-part model and the modified two-part model, respectively. The cancer-attributable spending estimates from these models are approximately $20 billion and $19 billion, respectively. Each estimate is within the confidence interval of the other and within the confidence interval of the encounter-based estimate. Estimates of cancer-attributable inpatient expenditures differ between each method, with the encounter-based method producing the highest estimate, approximately $12 billion, and the standard two-part model yielding the lowest estimate, just above $8 billion.

Table 3. Cancer-Attributable Spending Estimates in Billions (2001 U.S. Dollars) for Patients Ages Birth–64 Years from the Medical Expenditure Panel Survey, 1996–1999
CategoryAll encounters (95% CI)aCancer-attributable spending (95% CI)
Cancer encounters onlyStandard two-part model predictionsModified two-part model predictions
  • 95% CI: 95% confidence interval.

  • a

    Includes care unrelated to cancer treatment.

Total$30.35 ($25.75–35.00)$20.08 ($15.95–24.20)$20.55 ($18.01–23.18)$18.97 ($15.92–21.82)
Private$19.46 ($16.43–22.50)$13.08 ($10.63–15.53)  
Out-of-pocket$2.32 ($1.97–2.68)$1.27 ($0.97–1.57)  
Medicare$2.23 ($1.63–2.83)$1.24 ($0.74–1.75)  
Medicaid$2.94 ($1.93–3.95)$1.93 ($0.99–2.88)  
Inpatient$17.37 ($13.13–21.60)$12.27 ($8.40–16.13)$8.32 ($6.22–10.37)$10.62 ($8.71–13.27)
Outpatient$12.58 ($11.85–13.30)$7.61 ($6.98–8.25)$8.30 ($7.31–9.32)$7.12 ($6.22–8.08)
Drug$0.41 ($0.36–0.46)$0.20 ($0.15–0.25)$1.51 ($1.29–1.76)$1.04 ($0.87–1.24)
Colon$0.80 ($0.27–1.33)$0.72 ($0.20–1.25)  
Lung$3.36 ($2.47–4.25)$2.66 ($1.87–3.45)  
Skin$0.98 ($0.62–1.34)$0.59 ($0.25–0.93)  
Breast$3.60 ($2.98–4.23)$2.44 ($1.89–3.00)  
Prostate$1.80 ($1.39–2.21)$1.15 ($0.83–1.48)  

Table 4 shows total and cancer-attributable spending estimates based on the Medstat Marketscan data base. Estimated total spending on care for cancer patients is nearly $46 billion, much higher compared with the MEPS-based estimate of approximately $30 billion. Estimated cancer-attributable spending from the fixed-effects method is nearly $33 billion compared with only approximately $17 billion from the encounter-based method. The nearly two-fold difference in the estimates appears to be attributable mainly to differences in outpatient spending; cancer-attributable outpatient spending according to the encounter based method is just over $9 billion compared with > $21 billion according to the fixed-effects method. To gain further insight into the difference in estimates, we identified the most common diagnoses associated with encounters that did not list a cancer diagnosis. Most diagnoses, as expected, were for the treatment of symptoms that were related plausibly to cancer and cancer treatment, suggesting that many visits associated with cancer care may not be identified as such by diagnostic codes.

Table 4. Cancer-Attributable Spending Estimates in Billions (2001 U.S. Dollars) for Patients Ages Birth–64 Years from the Medstat Data Base, 1999
 All encounters (95% CI)aCancer-attributable spending (95% CI)
Cancer events onlyCase–control analysis
  • 95% CI: 95% confidence interval.

  • a

    Includes care unrelated to cancer treatment.

  • b

    Outpatient includes spending on physician office and outpatient hospital care.

Total$46.72 ($35.97–49.95)$17.19 ($13.39–20.98)$32.82 ($26.26–39.37)
Private$44.47 ($32.34–46.16)$16.49 ($12.74–20.23)$31.71 ($25.23–38.2)
Out-of-pocket$2.25 ($3.37–4.05)$0.70 ($0.46–0.94)$1.10 ($0.78–1.43)
Inpatient$17.29 ($12.23–22.36)$5.95 ($3.19–8.72)$9.33 ($4.59–14.07)
Outpatientb$25.38 ($22.1–28.65)$9.19 ($7.25–11.12)$21.21 ($17.94–24.49)
Drug$4.05 ($3.59–4.51)$2.05 ($1.71–2.38)$2.27 ($1.81–2.73)
Colon$4.24 ($2.8–5.69)$1.28 ($0.58–1.97)$2.76 ($1.44–4.07)
Lung$7.57 ($5.4–9.74)$4.09 ($3.96–4.23)$5.50 ($3.47–7.53)
Skin$1.17 ($0.28–2.07)$0.10 ($0.02–0.18)$0.86 ($0.04–1.68)
Breast$10.35 ($7.82–12.89)$2.22 ($1.34–3.10)$7.79 ($5.40–10.19)
Prostate$1.12 ($0.31–1.93)$0.10 ($0.07–0.13)$0.74 ($0.67–0.82)

The MEPS and Medstat participants differ in terms of residential location, insurance coverage, and socioeconomic status. To control for the differences, we examined spending by cancer patients and cancer-attributable spending as a share of the total spending in each data base. If the shares differ between data bases, then the differences in total spending displayed in Tables 3 and 4 most likely are the result of some other factor, such as under-reporting of utilization in MEPS. Spending shares also are useful to payers, employers, and policymakers for computing spending by cancer patients and cancer-attributable spending from any set expenditure base.

Spending shares are shown in Table 5. The second and third columns show total spending by cancer patients and the standard two-part model estimate of cancer-attributable expenditures from MEPS, respectively, as shown in Table 3. The fourth column shows estimated aggregate medical spending by nonelderly individuals in the U.S., also from MEPS. The fifth and sixth columns show total spending by cancer patients and the fixed-effects estimate of cancer-attributable expenditures from Medstat, respectively, as shown in Table 4. The last column shows the estimate of aggregate medical spending by the nonelderly from the Medstat data base. When comparing the estimates, the shares of private spending are most relevant. In each case, the Medstat estimate exceeds the MEP estimate by ≥ 2 percentage points. For example, the share of cancer-attributable spending by private payers is 4.7% according to MEPS and 6.7% according to the Medstat estimates.

Table 5. National Estimates of Cancer Spending in Billions (2001 U.S. Dollars) as a Percent of Total Spending from the Medical Expenditures Panel Survey, 1996–1999, and the Medstat Marketscan Data Base, 1999
SpendingMEPS, 1996–1999Medstat Data Base, 1999
All encountersCancer-attributableTotalAll encountersCancer-attributableTotal
  1. MEPS: Medical Expenditures Panel Survey.



  1. Top of page

Traditionally, government insurance programs have operated on a disease-neutral basis, with the exception of Medicare's End Stage Renal Disease Program. A recent exception to this rule is the Breast and Cervical Cancer Prevention and Treatment Act of 2000 (the Act), which allows states to extend Medicaid benefits to women who are diagnosed with breast or cervical cancer under the National Breast and Cervical Cancer Early Detection Program. To the extent that the Act portends a shift to disease-based eligibility criteria, cost-of-treatment estimates are necessary to project the budgetary implications of insurance expansions. Our estimates show that the cost to the government of a universal program to provide insurance coverage to nonelderly cancer patients would be at least $30 billion, and possibly as much as $47 billion, annually (including current spending by the Medicaid and Medicare programs for this population). A program with eligibility criteria to discourage enrollment by patients with private insurance could cost substantially less, depending on the number of cancer patients who would drop private coverage to gain eligibility for a government program.

Various methods have been employed to estimate the medical costs attributable to specific diseases, including cancer. When applied to the MEPS, these methods produce surprisingly similar cancer-attributable spending estimates. Our results imply that the clinical classification codes associated with each encounter in the MEPS and the linkages of those codes with episodes of care are fairly accurate. Conversely, claims-based and fixed-effects estimates of cancer-attributable spending from the Medstat data base differ greatly, suggesting that cancer diagnosis codes are under-reported on claims for encounters in which patients received cancer care. Investigators should proceed cautiously when using claims data bases to characterize the utilization and costs of cancer patients and should attempt to identify encounters for treatment of cancer-related symptoms and side effects.

In absolute terms, spending estimates from MEPS and the Medstat data base differ substantially. Based on differences in spending between insured and uninsured cancer patients,8 only a small proportion of the total difference can be attributed to the fact that about 10% of the MEPS cancer sample is uninsured. Differences in insurance generosity among individuals with insurance may be more relevant; claims in Medstat are compiled from large employers, which tend to offer more extensive coverage compared with small employers and individually purchased policies. Another important source of spending differences is the exclusion by MEPS of patients residing in institutions. End-of-life care, which comprises a large share of total cancer spending,4 often is provided in institutions, such as hospice.

In percentage terms, the cancer-attributable spending estimate for private insurance from MEPS is 2 percentage points below the comparable figure from Medstat. Differences in the samples may explain the disparity. The exclusion of care provided in institutional settings may lower the numerator in the share calculations for MEPS relative to Medstat. In addition, Medstat includes mainly individuals who are employed. Individuals with chronic conditions who are out of the labor force contribute to the denominator in the spending share calculations for MEPS, but not necessarily to the numerator, lowering the shares relative to Medstat. The confidence intervals around the cancer-attributable spending estimates for private health insurance are fairly wide, and the differences in spending shares between data bases may be the result of sampling error.

Brown et al.4 estimated that, in 1996, cancer-attributable spending by the nonelderly population was $18.6 billion, or $26 billion in 2001 dollars, after adjusting for trends in private health insurance spending. Although it impossible to put this figure in percentage terms, because the National Health Accounts do not break out spending by age group, their estimate of the share of cancer-attributable spending for all age groups is 5%.

Which estimate is correct? Ultimately, it depends on the question asked. When assessing the burden of cancer or ranking diseases by cost, the estimate by Brown et al.4 that spending on cancer care is $40 billion annually (1996 U.S. dollars), or 5% of total spending, appears reasonable. When considering changes to cancer patients' insurance coverage that apply to all types of care (and not just cancer treatment), then the applicable shares of total spending are 5.5% for the nonelderly population overall and 9% for populations with generous employer-sponsored health insurance.


  1. Top of page
  • 1
    Gross CP, Anderson GF, Powe NR. The relation between funding by the National Institutes of Health and the burden of disease. N Engl J Med. 1999; 340: 18811887.
  • 2
    Brown ML, Fintor L. The economic burden of cancer. In: GreenwaldP, KramerBS, WeedDL, editors. Cancer prevention and control. New York: Marcel Dekker, 1995: 6981.
  • 3
    Brown ML, Lipscomb J, Snyder C. The burden of illness of cancer: economic cost and quality of life. Annu Rev Public Health. 2001; 22: 91113.
  • 4
    Brown ML, Riley GF, Schussler N, Etzioni R. Estimating health care costs related to cancer treatment from SEER-Medicare data. Med Care. 2002; 40(8 Suppl ): IV-104IV-117.
  • 5
    Antos JR, Goyburu A. Comparing Medicare and private health insurance spending. Washington, DC: Heritage Foundation, 2003.
  • 6
    Moon M. Beneath the averages: an analysis of Medicare and private expenditures. Washington, DC: Kaiser Family Foundation, 1999.
  • 7
    Medicare Payment Advisory Commission. Assessing payment adequacy and updating payments in fee-for-service Medicare. In: Medicare Payment Advisory Commission. Report to the Congress: Medicare payment policy (March 2003). Washington, DC: Medicare Payment Advisory Commission, 2003: 2752.
  • 8
    Thorpe K, Howard DH. Health insurance and spending among cancer patients. Health Aff (Millwood). 2003; W3: 189198.
  • 9
    Marquis MS, Long SH. Public insurance expansions and crowd out of private coverage. Med Care. 2003; 41: 344356.
  • 10
    Cohen JW, Monheit AC, Beauregard KM, et al. The Medical Expenditure Panel Survey: a national health information resource. Inquiry. 1996; 33: 373389.
  • 11
    Cohen SB. Sample design of the 1997 Medical Expenditure Panel Survey household component. Methodology Report No. 11. Rockville, MD: Agency for Healthcare Research and Quality, 2000.
  • 12
    Selden TM, Levit KR, Cohen JW, et al. Reconciling medical expenditure estimates from the MEPS and the NHA, 1996. Health Care Finance Rev. 2001; 23: 161177.
  • 13
    Levit K, Smith C, Cowan C, Lazenby H, Sensenig A, Catlin A. Trends in U.S. health care spending, 2001. Health Aff (Millwood). 2003; 22: 154164.
  • 14
    Miller VP, Ernst C, Collen F. Smoking-attributable medical care costs in the USA. Soc Sci Med. 1999; 48: 375391.
  • 15
    Duan N. Smearing estimate: a nonparametric retransformation method. J Am Stat Assoc. 1983; 78: 605610.
  • 16
    Manning WG. The logged dependent variable, heteroscedasticity, and the retransformation problem. J Health Econ. 1998; 17: 283296.
  • 17
    Mullahy J. Much ado about two: reconsidering retransformation and the two-part model in health econometrics. J Health Econ. 1998; 17: 247282.
  • 18
    Statistical Research and Applications Branch, National Cancer Institute. Cancer Prevalence Database [data base online]. Available from URL: [accessed February 27, 2003].