JOURNAL OF PALLIATIVE MEDICINE
Volume 12, Number 2, 2009
© Mary Ann Liebert, Inc.
Cancer Care in the United States:
Identifying End-of-Life Cohorts
Ethan M. Berke, M.D., M.P.H.,1,2,3Tenbroeck Smith, M.A.,4Yunjie Song, Ph.D.,1
Michael T. Halpern, M.D., Ph.D., M.P.H.,4and David C. Goodman, M.D., M.S.1,2,5
Objectives: End-of-life care is increasingly recognized as an important part of cancer management for many
patients. Current methods to measure end-of-life care are limited by difficulties in identifying cancer cohorts
with administrative data. We examined several techniques of identifying end-of-life cancer cohorts with claims
data that is population-based, geographically scalable, and amenable to routine updating.
Methods: Using Medicare claims for patients 65 years of age and older, four techniques for identifying end-of-
life cancer cohorts were compared; one based on Part A data using a broad primary or narrow secondary di-
agnosis of cancer, two based on Part B data, and one combining the Part A and B methods. We tested the per-
formance of each definition to ascertain an appropriate end-of-life cancer population.
Results: The combined Part A and B definition using a primary or secondary diagnosis of cancer within a win-
dow of 180 days prior to death appears to be the most accurate and inclusive in ascertaining an end-of-life co-
hort (78.7% attainment).
Conclusion: Combining inpatient and outpatient claims data, and identifying cases based upon a broad pri-
mary or a narrow secondary cancer definition is the most accurate and inclusive in ascertaining an end-of-life
tients with cancer is an important and understudied dimen-
sion of geriatric care. Although standardized end-of-life co-
horts for Medicare beneficiaries have been defined and
extensively studied,2,3there has been little effort to develop
disease-specific cohorts, such as for cancer. One barrier to
studies of end-of-life cancer care has the difficulty in as-
signing the cause of death to cancer for large populations.
Medicare data is one approach to measuring end-of-life
cancer care with several advantages. Medicare claims files
are readily available to researchers and include the popula-
tion-based medical care experience of the elderly. Analyses
can include the full range of utilization and are also amenable
to routine updating over time.
The goal of this study was to develop a cohort of Medicare
end-of-life patients using claims files. We developed several
methods for identifying Medicare beneficiary decedents with
a diagnosis of cancer and then evaluated them for cohort size
ITH A GROWING POPULATION over 65 years old,1accurate
measurement of end-of-life medical services in pa-
and the likelihood that cancer was a major contributing cause
Four cancer cohorts were based on fee-for-service
Medicare beneficiaries Part A and B eligible age 65 years
and older for the period 2001 to 2005. We first identified
beneficiaries who died, and then searched within defined
periods prior to death for utilization events indicating that
the patient had cancer. Cohorts were then evaluated on the
basis of the number of beneficiaries assigned to the cohort
and the likelihood that patients receiving hospice care with
a principal diagnosis of cancer were also found within the
Duplicate claims and claims with an allowable charge of
$0 were excluded from the analyses. Patients were classified
as black or non-black, based on the work of others advocat-
ing this dichotomization, and availability of racial data in the
Medicare claims dataset.2
The first cohort included decedents whose last inpatient
1The Center for Healthcare Research and Reform, The Dartmouth Institute for Health Policy and Clinical Practice, Dartmouth Medical
School, Hanover, New Hampshire.
2Department of Community and Family Medicine, Dartmouth Medical School, Hanover, New Hampshire.
3The Norris Cotton Cancer Center, Dartmouth Hitchcock Medical Center, Lebanon, New Hampshire.
4Research and Evaluation Department, American Cancer Society, Atlanta, Georgia.
5Department of Pediatrics, Dartmouth Medical School, Hanover, New Hampshire.
stay from the MedPAR file had a principal diagnosis of can-
cer (Part A Definition) (ICD-9 codes: 140-208 or 239.0-239.9,
excluding V codes) or with a secondary diagnosis of cancer
considered to be metastatic or with a poor prognosis, based
on previous work by Iezzoni et al.4(see Appendix available
online at www.liebertpub.com/jpm). By using a secondary
diagnosis of cancer identified as severe, we were able to cap-
ture those subjects admitted to the hospital for a potentially
associated condition (e.g., pneumonia), even though the sec-
ondarily listed cancer diagnosis was the likely cause of sub-
In the next two cohorts, we used diagnoses on physician
and other clinician claims indicate the presence of cancer.
This method included patients with cancer without a recent
hospital discharge related to their terminal illness but who
received outpatient care. These cohort definitions used a 20%
beneficiary sample Carrier File of Medicare Part B physician
claims (Part B thereafter). These files contain dates of service
and associated diagnosis codes for each physician claim for
an individual enrolled in Medicare Part B.5The second co-
hort included patients who had claims with a primary di-
agnosis of cancer within the time frames specified above
(Broad Definition). A diagnosis meeting these criteria had to
occur two or more times, at least 6 days apart but not more
than 120 days apart, with the later of the two claims occur-
ring within the designated period from death. We then de-
fined a third cohort using Part B data, but added a secondary
diagnosis of a restricted set of severe cancers, similar to co-
hort 1 (Combined Definition).
Finally, we defined a fourth cohort based on the union of
Part A and Part B claims from the sources above (Combined
Part A Part B Definition). We selected a 20% beneficiary sam-
ple of the 100% Part A file comparable to the 20% Part B file.
This sample was joined with the subjects in the Part B co-
hort that had a primary or secondary diagnosis of cancer us-
ing the methods outlined above.
We evaluated these cohorts four ways. First, we compared
the number of deceased beneficiaries identified by each co-
hort with the U.S. vital records6estimates of the number of
cancer deaths in those over 65 for the target year. Second, if
the cohort members died of cancer, we would expect that a
reasonable proportion of them would receive hospice ser-
vices. We calculated rates of hospice utilization for the en-
tire cohort using the Medicare Hospice File. Hospice enroll-
ment was defined as the “admission” to hospice services,
regardless of location, nearest to death.
Third, we assumed that the principal diagnosis in a hos-
pice admissions was the likely cause of death7and then cal-
culated the proportion of hospice patients with a cancer di-
agnosis that was also identified by our defined cohorts.
Higher scores on this proportion indicate better “sensitiv-
ity” of a given cohort definition for detecting cases that died
Finally, we calculated the proportions of true-positives
within the cohorts. The denominator consisted of all mem-
bers of a given cohort admitted to hospice, and the nu-
merator the subset of those cases whose primary diagno-
sis for admission to hospice was cancer. Higher scores on
this statistic indicated a given cohort definition identified
a high proportion of true-positive cases in those receiving
hospice care, and excluded cases that died of noncancer di-
Based on a death year of 2005 and a 180-day predeath win-
dow, the Combined Part A Part B definition identified 58,978
patients representing 294,890 nationally when upweighted
to 100% (Table 1). Adjusting this number for risk bearing
HMO beneficiaries excluded from the analysis (6.1 million
of 43 million total enrollees in 2005),8this cohort was 88.5%
of the estimated number of deaths from cancer age older than
65 years in non-HMO patients according to 2005 U.S. vital
The proportion of cohort members in 2005 receiving hos-
pice benefits varied from 53% to 57% depending on the co-
hort definition. Hospice use increased monotonically from
2001–2005. These are rates are consistent with the upward
trend of hospice care in cancer patients reported by Mc-
Carthy et al.9
Comparing the proportion of patients admitted to hospice
with a primary diagnosis of cancer who were also found in
the study cohorts assessed the sensitivity of the various co-
hort definitions to identify patients who died of cancer (Table
2). The Combined Part A Part B Cohort identified 80.1% of
hospice patients with a primary diagnosis of cancer, sug-
gesting this cohort definition is the most sensitive definition
for identifying those who died of cancer. These results were
consistent for all 5 years of data and for all three inclusion
Examining solely the cohort members who were admitted
to hospice, Table 3 shows the proportion for which the prin-
cipal hospice diagnosis was cancer. The Combined Part A
cohort definition achieved slightly higher levels (around
95%) indicating the proportion of true positive to false pos-
itives is slightly higher in this cohort than the others.
Varying definitions of cancer and place of service affect
the number of patients identified for end of life cancer co-
horts. Our analyses show that a cohort definition based on
a union of Part A and Part B Medicare claims data, using a
180-day window prior to death, includes the largest number
of patients while still overlapping with hospice patients with
a principal diagnosis of cancer. The 180-day window corre-
sponds to the 6-month eligibility criteria established by
Medicare,10and is consistent with data showing that the ma-
jority of patients enroll in hospice in the 180 days prior to
By taking advantage of national Medicare claims data as
opposed to data sets limited to selected states or a health care
system, we are more confident that our cohort definition is
nationally representative of the elderly population. While no
indicator of cause of death is perfect, our claims-based mea-
sure may have advantages over methods based on death cer-
tificates given research questioning the accuracy of vital sta-
tistics recording.11–13These studies suggest that cancer is
often not listed as the underlying cause of death, leading to
underestimation of cancer cases. By including the entire fee-
for-service Medicare aged population, our cohort definition
has the additional value of applicability to analyses at na-
tional, regional, state, county, or health service area levels.
There are a number of limitations in this approach.
Medicare data cannot account for cancer patients who are
under 65 years old and who are not disabled. Our definition
IDENTIFYING END-OF-LIFE COHORTS129
END-OF-LIFE CANCER COHORT DEFINITIONS AND PROPORTION OF COHORT ENROLLED IN HOSPICE CARE, 2001–2005
with Part B
Part A claim
Part A or B
# with cancer
aTwenty percent sample, up-weighted to 100%.
bICD-9 definition: 140–208 or 239.0–239.9, excluding V codes.
cFirst diagnosis is ICD-9 codes 140–208 or 239.0–239.9, excluding V codes, or second diagnosis using Iezzoni definition, Appendix (available online at www.liebertpub.com/jpm).
dMaximum possible time span in days from death to admission date of last hospitalization.
Part A and B Combined Definitiona,c
IDENTIFYING END-OF-LIFE COHORTS131
OF PATIENTS ASCERTAINED IN VARIOUS COHORT DEFINITIONS WITH DIAGNOSIS OF CANCER
OF THOSE PATIENTS IN HOSPICE WITH A PRIMARY DIAGNOSIS OF CANCER, PROPORTION
Part B Definitionsa
Of those in
Of those in
Of those in
with Part A
Of those in
hospice, % within
combined Part A
and Part B
aTwenty percent sample, up-weighted to 100%.
bMaximum time span in days from death to admission date of last hospitalization.
cFirst diagnosis is ICD-9 codes 140–208 or 239.0–239.9, excluding V codes, or second diagnosis using Iezzoni definition, Appendix (avail-
able online at www.liebertpub.com/jpm).
dICD-9 definition: 140–208 or 239.0–239.9, excluding V code.
TABLE 3.OF COHORT MEMBERS WHO ENTER HOSPICE, PROPORTION THAT HAVE A PRIMARY HOSPICE DIAGNOSIS OF CANCER
Part B Definitionsb
Part A Definitiona
Death Year # beneficiaries# beneficiaries# beneficiaries# beneficiaries
aFirst diagnosis is ICD-9 codes 140–208 or 239.0–239.9, excluding V codes, or second diagnosis using Iezzoni definition, Appendix (avail-
able online at www.liebertpub.com/jpm).
bTwenty percent sample, up-weighted to 100%.
cICD-9 definition: 140–208 or 239.0–239.9, excluding V code.
dMaximum time span in days from death to admission date of last hospitalization.
Combined Part A or B
relies on claims data, and not chart abstraction or other mea- Download full-text
sures that might better indicate a diagnosis of cancer in the
last months of life. It is possible that some patients were ex-
cluded from our cohort, particularly from the Part A defini-
tion, if a cancer diagnosis was listed lower than a primary
or secondary diagnosis in the discharge records. This would
negatively impact the sensitivity of our measure, but should
do little to affect the specificity.
Using a cohort definition based on joined Medicare Part
A and B data and a primary or secondary diagnosis of can-
cer, with a more severe secondary diagnosis, appears to
yields the most appropriate nationally representative sam-
ple of cancer deaths for age older than 65. These cohorts will
allow for analysis of trends and regional variation in cancer
care near the end of life.
The authors are grateful to the American Cancer Society,
The Robert Wood Johnson Foundation, and the National In-
stitute on Aging Grant P01-AG-19783 for financial support.
Author Disclosure Statement
No competing financial interests exist.
1. U.S. Census Bureau: U.S. Interim Projections by Age, Sex,
Race, and Hispanic. 2004.
2. Wennberg JE, Cooper MM: The Dartmouth Atlas of Health Care
in the United States. Chicago: American Hospital Publishing,
3. Wennberg JE, Fisher E, Goodman DC, Skinner JS: Tracking
the Care of Patient with Severe Chronic Illness: The Dartmouth
Atlas of Health Care 2008. Lebanon: The Dartmouth Institute
for Health Policy and Clinical Practice, 2008.
4. Iezzoni LI, Heeren T, Foley SM, Daley J, Hughes J, Coffman
GA: Chronic conditions and risk of in-hospital death. Health
Serv Res 1994;29:435–460.
5. Research Data Assistance Center (ResDAC): www.resdac.
umn.edu/Index.asp (Last accessed October 16, 2007).
6. National Center for Health Statistics: Health, United States,
2006 With Chartbook on Trends in the Health of Americans. Hy-
attsville, MD: National Center for Health Statistics, 2006.
7. Abdel-Karim IA, Sammel RB, Prange MA: Causes of death
at autopsy in an inpatient hospice program. J Palliat Med
8. Henry J. Kaiser Family Foundation: Medicare: Medicare Ad-
vantage Fact Sheet. March 2007.
9. McCarthy EP, Burns RB, Ngo-Metzger Q, Davis RB, Phillips
RS: Hospice use among Medicare managed care and fee-for-
service patients dying with cancer. JAMA 2003;289:2238–
10. Medpac: Hospice care in Medicare: Recent trends and a re-
view of the issues. Report to the Congress: New Approaches
in Medicare. 2004, pp. 139–154.
11. Ravakhah K: Death certificates are not reliable: revivifica-
tion of the autopsy. South Med J 2006;99:728–733.
12. Sington JD, Cottrell BJ: Analysis of the sensitivity of death
certificates in 440 hospital deaths: A comparison with
necropsy findings. J Clin Pathol 2002;55:499–502.
13. Hoel DG, Ron E, Carter R, Mabuchi K: Influence of death
cetificate errors on cancer mortality trends. J Natl Cancer
Address reprint requests to:
Ethan M. Berke, M.D., M.P.H.
Department of Community and Family Medicine
The Dartmouth Institute for Health Policy and Clinical Practice
Dartmouth Medical School
35 Centerra Parkway, Room 206
Lebanon, NH 03766
BERKE ET AL. 132