Conditional survival in gastric cancer: a SEER database analysis.

Gastric Cancer (2007) 10: 153–158
DOI 10.1007/s10120-007-0424-9
Offprint requests to: S. J. Wang
Received: December 19, 2006 / Accepted: May 2, 2007
 2007 by
International and
Japanese Gastric
Cancer AssociationsOriginal article
Conditional survival in gastric cancer: a SEER database analysis
Samuel J. Wang1,2, Rachel Emery3, Clifton D. Fuller4,5, Jong-Sung Kim6, Dean F. Sittig2,7,
and Charles R. Thomas Jr.1
1 Department of Radiation Medicine, MC KPV4, Oregon Health and Science University, 3181 SW Sam Jackson Park Road, Portland,
OR 97239-3098, USA
2 Department of Medical Informatics and Clinical Epidemiology, Oregon Health and Science University, Portland, OR, USA
3 School of Medicine, Oregon Health and Science University, Portland, OR, USA
4 Department of Radiation Oncology, University of Texas Health Science Center at San Antonio, San Antonio, TX, USA
5 Graduate Division of Radiological Sciences, University of Texas Health Science Center at San Antonio, San Antonio, TX, USA
6 Department of Mathematics and Statistics, Portland State University, Portland, OR, USA
7 Applied Research in Medical Informatics, Northwest Permanente, PC, Portland, OR, USA
in the United States, and 11 430 gastric cancer deaths
[1]. The incidence of gastric cancer increases with age,
with a median age at diagnosis of 70 years for males and
74 for females. Gastric cancer is almost twice as common
in men than in women, and incidence rates are 2.2 times
higher in blacks compared to whites [2]. While the
overall incidence of gastric cancer has been decreasing
over time since the 1930s, the incidence rates for black
males has not changed signifi cantly, and rates actually
increased for white females between 1974 and 1994
[2].
Gastric cancer is an aggressive malignancy that is dif-
fi cult to detect at an early stage. Survival estimates for
gastric cancer patients are usually only reported in the
literature as survival from the time of diagnosis, and 5-
year overall survival rates from diagnosis are reported
as about 22% [2,3]. Survival probability changes,
however, for gastric cancer patients who survive for a
period of time after diagnosis, and their prognosis is
more accurately described using conditional survival
(CS) [4]. CS is taken from the concept of conditional
probability, and is a more accurate estimate of survival
probability for patients who have survived for 1 or more
years after diagnosis. CS accounts for the fact that
hazard rates can change over time. Similar to many
other cancer types, hazard rates for gastric cancer are
relatively higher in the fi rst few years after diagnosis,
signifying greater risk, but then the rates decrease mark-
edly after the fi rst few years. For those patients who are
fortunate to survive beyond this initial period of time,
prognosis can be substantially improved, and the initial
estimates of survival made at the time of diagnosis no
longer apply. CS is a more relevant measure of survival
probability for these cancer survivors.
The concept of CS has important practical clinical
value for patients, providers, and researchers [4,5].
Cancer patients who are seen in follow-up clinics several
years after their diagnosis may wish to know how their
Abstract
Background. Gastric cancer survival is typically reported in
terms of survival from the time of diagnosis. Conditional sur-
vival is a more relevant measure of prognosis for patients who
have already survived 1 or more years since diagnosis.
Methods. Using the Surveillance, Epidemiology, and End
Results (SEER 17) database from the National Cancer Insti-
tute, we analyzed data from 20 018 gastric cancer patients
diagnosed between 1988 and 1998. Using the life-table method,
we computed 5-year relative conditional survival, grouped by
summary stage, age, sex, and ethnicity, for patients who had
already survived up to 5 years from diagnosis.
Results. Relative conditional survival improves over time for
all groups of gastric cancer patients who survive a period of
time after diagnosis. The largest gains in conditional survival
were seen in patients with advanced stage disease. In general,
females showed better survival than males. When grouped by
ethnicity, Asians continued to have improved survival com-
pared to other ethnic categories, even at 5 years out from
diagnosis.
Conclusion. For gastric cancer patients who survive a period
of time after diagnosis, the largest increases in conditional
survival are seen for patients with advanced stage disease and
for those less than 65 years old. Conditional survival can
provide more relevant prognostic information than survival
from the time of diagnosis for gastric cancer patients who
survive a period of time after diagnosis.
Key words Survival analysis · Epidemiological methods ·
Stomach neoplasms
Introduction
In 2006, the American Cancer Society estimated that
there were 22 280 new cases of gastric cancer diagnosed

154 S.J. Wang et al.: Conditional survival in gastric cancer
prognosis is changing over time. If a cancer patient’s CS
has risen high enough to essentially match the expected
survival of the general population, it allows a more
objective basis to deem a patient “cured” of their
disease. Providers can also make use of CS information
to more objectively determine an appropriate frequency
of follow-up visits and aggressiveness of surveillance
testing, based on the patient’s current risk profi le. When
designing clinical trials, clinical researchers may also
fi nd CS useful in helping to determine suffi cient follow-
up times for trial endpoints.
Several authors have previously published studies on
CS for various disease sites, including breast [6], colon
[7], central nervous system (CNS) [8,9], lung [10,11],
and other advanced carcinomas [12]. We have also pre-
viously presented our work on CS for various disease
sites [13–15]. However, to our knowledge, no formal
literature has explored the CS patterns of gastric carci-
nomas. The specifi c aim of this study was to investigate
the prognostic factors that affect the CS of gastric cancer
patients, using the Surveillance, Epidemiology, and End
Results (SEER) database [16].
Methods
Defi nition of conditional survival
Conditional survival (CS) is derived from the concept
of conditional probability in biostatistics. CS can be
calculated from traditional Kaplan-Meier or actuarial
life-table survival data. The mathematical defi nition of
CS can be expressed as follows: CS, CS(y|x), is the prob-
ability of surviving an additional y years, given that the
person has already survived x years. Let S(t) be the
traditional actuarial life-table survival at time t. CS can
be expressed as:
CS y x
S x y
S x
( | )
( )
( )
=
+
For example, to compute the 5-year CS for a patient
who has already survived 2 years, the survival at 5 + 2
years, S(7), is divided by the survival at 2 years, S(2).
When a survival curve has a changing hazard rate over
time, this will be refl ected as a change in CS as more
time elapses from the time of diagnosis.
Data selection criteria
The SEER Program [16] from the National Cancer
Institute is a population-based cancer registry covering
approximately 26% of the United States population
across several disparate geographic regions, the largest
publicly available domestic cancer dataset. SEER
program registries collect data on patient demograph-
ics, cancer type and site, stage, and fi rst course of treat-
ment, and they follow up vital status.
Using the November 2005 release of the SEER 17
database [16] with the SEER*Stat 6.2.4 software [17]
we analyzed survival data from all patients diagnosed
with gastric cancer between 1988 and 1998, with follow-
up to December 2003, to ensure a minimum of 5 years
of follow-up data. Patients were selected who had a
“site recode” data fi eld of “Stomach” with ICD-O-3
histology codes for adenocarcinoma (or unspecifi ed car-
cinoma) (8000-8001, 8010, 8020-8021, 8050, 8140-8221,
8255-8560, 8570-8576).
Case selection options in SEER*Stat were used to
restrict cases to “actively followed”, “malignant behav-
ior”, “exclude all death certifi cate only and autopsy
only”, and “exclude second and later primaries”.
Because American Joint Committee on Cancer
(AJCC) staging information was not available in SEER
for gastric cancer, we used the SEER summary stage
(1977) [18]. Using a combination of both the clinical and
the pathological documentation of the extent of disease,
summary staging categorizes how far a cancer has spread
from its point of origin. The stages included: “localized
only”, “regional by direct extension only”, “regional by
lymph node involvement only”, “regional by both direct
extension and lymph node(s) involvement”, and “distant
site/node involvement”.
Data analysis
Conditional survival probabilities were calculated using
SEER*Stat 6.2.4 [17], using the actuarial life-table
method. All CS rates are reported as relative survival,
defi ned as the observed CS rate divided by the expected
CS rate for similar individuals matched for age, sex,
ethnicity, and date at which the age was coded. This is
a method for accounting for competing causes of death
when exact cause-of-death information is incomplete or
unknown. We computed 5-year CS by ethnicity and
stage, and further grouped stage results by age (>65 vs
<65 years), and sex.
Results
After eliminating 3195 cases that did not meet the case
selection criteria, a total of 20 018 gastric cancer patients
were included in this analysis. When grouped by
summary stage, 21% had localized disease, 10% had
regional disease by direct extension only, 10% had
regional disease by lymph node involvement only, 18%
had regional disease by both direct extension and lymph
node involvement, and 41% had distant site/nodal
disease. Patients were further subsetted by sex and age,
as shown in Table 1, and the number of patients in each
ethnic group is shown in Table 2.

S.J. Wang et al.: Conditional survival in gastric cancer 155
and 46% for age >65). When examined by sex and stage,
CS for females was generally higher than for males, with
the largest gains over time seen in patients with distant
disease — for females, CS rose from 3% at diagnosis to
67% after 5 years, and for males, CS rose from 2% at
diagnosis to 61% after 5 years.
Discussion
In this study, we found that various prognostic factors
affect the conditional survival (CS) of patients with
gastric cancer. From other studies of CS, it is known
that patients with poorer initial prognoses often see the
largest increases in CS over time [6,8,10,11]. Our study
confi rmed this fi nding, and also found that CS appeared
to increase the most for American Indians/Alaskans,
women, and those less than 65 years of age with
advanced stage disease. Our data appear to correspond
with previously published data regarding the infl uence
of ethnicity, sex, and age on gastric cancer survival at
diagnosis. Asian ethnicity is associated with distinct
characteristics at presentation and has previously been
shown to exhibit more favorable survival as compared
with that in non-Asians, even when adjusted for age,
Table 1. Number (%) of patients, grouped by age, sex, and summary stage (n = 20 018)
Regional- Regional- Regional-
Localized direct extension lymph node(s) both Distant mets
Age < 65 1126 (27) 568 (29) 807 (38) 1408 (38) 3277 (40)
Age > 65 3013 (73) 1386 (71) 1304 (62) 2270 (62) 4859 (60)
Men 2527 (61) 1187 (61) 1357 (64) 2410 (66) 5247 (64)
Women 1612 (39) 767 (39) 754 (36) 1268 (34) 2889 (36)
Totals 4139 (100) 1954 (100) 2111 (100) 3678 (100) 8136 (100)
Table 2. Number (%) of patients in each ethnic group
White 13 950 (70)
Black 2 368 (12)
American Indian/Alaskan Native 196 (1)
Asian/Pacifi c Islander 3 476 (17)
Other/Unknown 28 (<1)
Total 20 018 (100)
Overall Survival
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0 1 2 3 4 5 6 7 8 9 10
Years after Diagnosis
O
ve
ra
ll
S
ur
vi
va
l (
R
el
at
iv
e)
Localized
Regional-
Direct Ext
Regional-
Nodes
Regional-
Both
Distant
Mets
Fig. 1. Ten-year (relative) overall sur-
vival curves by Surveillance, Epidemiol-
ogy, and End Results (SEER) summary
stage. These data were used to calculate
the 5-year relative conditional survival
probabilities. Ext, extension; mets,
metastases
The 10-year relative overall survival data (Fig. 1)
were used to calculate 5-year relative CS, shown in Figs.
2 and 3. Error bars depict the 95% confi dence intervals
for each CS statistic. Overall, while higher stage (distant
disease) patients had lower CS at diagnosis compared
to other stages, these patients also saw the greatest
increases in CS as more time elapsed from diagnosis
(Fig. 2). When grouped by ethnicity (Fig. 3), Asians had
the quickest sustained increase in CS, but the largest
differences in CS from year 0 to year 5 were in Ameri-
can Indians/Alaskan Natives. When grouped by age and
stage, patients over 65 years of age with localized and
distant disease did worse than those under 65 for all
times following diagnosis. For patients over age 65
with distant disease, this discrepancy in CS appeared to
widen over time (at 5 years, CS was 77% for age <65,

156 S.J. Wang et al.: Conditional survival in gastric cancer
and location [19,20]. Our fi ndings show that Asians con-
tinue to have more favorable CS compared to non-
Asians, even out to 5 years from the time of diagnosis.
As time progresses from diagnosis, CS provides
more relevant prognostic information compared to more
commonly reported static survival statistics, such as 5-
year overall survival. CS information is potentially of
great interest to patients, their clinicians, and research-
ers. When patients who are seen in follow-up inquire
about their current prognosis, they should be given an
accurate risk assessment that accounts for time already
survived since diagnosis. Psychologically, the ability of
patients to more accurately quantify their improvement
in prognosis over time may be of great benefi t. Accord-
ingly, every effort should be made to communicate this
changing risk profi le in terms that are accessible to the
lay patient. Five-year CS probability is an easily under-
standable measure that can be used to accurately portray
to a patient their current risk profi le.
Clinicians can also make use of CS data to implement
more evidence-driven approaches to plan an appro-
priate post-therapy surveillance schedule based on a
patient’s changing risk. Many physicians arbitrarily
taper follow-up visit frequency after 2–3 years, but often
All Patients By SEER Summary Stage
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Years after Diagnosis
5-
Y
ea
r
R
el
at
iv
e
C
on
di
tio
na
l S
ur
vi
va
l
Localized 61% 71% 78% 82% 85% 85%
Regional-Direct Extension 26% 45% 57% 72% 77% 79%
Regional-Lymph Nodes 34% 41% 53% 63% 72% 79%
Regional-Both 15% 22% 37% 51% 63% 67%
Distant Mets 2% 12% 29% 46% 58% 64%
0 1 2 3 4 5
Fig. 2. Five-year relative conditional sur-
vival by SEER summary stage, as a function
of elapsed time since diagnosis. Each bar
represents the probability of surviving an
additional 5 years, after having already sur-
vived for 0 to 5 years since diagnosis
All Patients By Ethnicity
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Years after Diagnosis
5-
Y
ea
r
R
el
at
iv
e
C
on
di
tio
na
l S
ur
vi
va
l
White 20% 40% 56% 68% 75% 78%
Black 22% 45% 60% 69% 75% 79%
Am Indian/Alaskan 16% 32% 41% 53% 61% 77%
Asian/Pacific Islander 31% 51% 68% 78% 83% 86%
0 1 2 3 4 5
Fig. 3. Five-year relative conditional sur-
vival by ethnicity, as a function of elapsed
time since diagnosis

S.J. Wang et al.: Conditional survival in gastric cancer 157
without evidence to justify whether this is an appropri-
ate timeframe. The determination of optimum follow-
up testing frequency and duration should ideally be
based on the patient’s disease risk, rather than simply
on custom or tradition. For instance, the CS data in
the present study show that, for patients with regional
lymph node involvement who survive for more than 3
years from diagnosis, subsequent mortality outcomes
are comparable to those in patients with localized
disease at diagnosis. Consequently, if one uses a particu-
lar follow-up surveillance schedule for patients with
localized disease for the fi rst year after diagnosis and
treatment, it stands to reason that patients with regional
lymph node involvement should be similarly followed
at 3 years out from diagnosis.
When researchers design clinical trials, the study
duration is predicted based on the estimated length of
follow-up needed to see a signifi cant result in a given
endpoint. These study duration estimates have impor-
tant implications in terms of the economic costs of a
trial and the timeliness in which trial results can be
reported. Gill and Sargent [21] have proposed that
alternative surrogate endpoints, such as disease-free
survival, may be appropriate in certain situations, as a
means to accelerate the completion of adjuvant clinical
trials. Examination of how CS data change over time
may be another method to ascertain an appropriate trial
follow-up time, because these data effectively quantify
the remaining risk to a patient after a given survival
time.
It is imperative for physicians to more specifi cally
characterize, utilize, and communicate patient risk pro-
fi les as accurately as possible. Models such as Adjuvant!
Online [22] or the Memorial Sloan-Kettering Cancer
Center (MSKCC) risk nomograms [23–27] are espe-
cially inviting as a mechanism for initial risk stratifi ca-
tion. We are currently investigating the incorporation
of a CS component into a similar predictive model that
could be used to allow patients and physicians to re-
evaluate a patient’s prognosis as it changes over time.
Such tools could potentially afford more appropriate
management of disease as a part of the physician’s sur-
veillance algorithm.
Utilization of the SEER dataset represents an effort
to determine CS parameters based on large population
cohorts by using the single largest domestic cancer case
data registry. The SEER dataset, while geographically
limited, represents the best large-scale pool of patient
data, and allows us to make reasonable estimates of CS
that are generally applicable for the United States pop-
ulation. Nonetheless, several limitations must be noted.
Increased cohort size comes at the cost of treatment
homogeneity. Treatment data reported in SEER are
limited (i.e., no chemotherapy regimen or radiotherapy
dose is recorded), so we did not attempt a subanalysis
by treatment modality. Also, the sample size for some
subgroups is small, and is refl ected in the larger confi -
dence interval bars seen in some groups, such as for
patients with distant disease and those in the American
Indian/Alaskan subgroup. This precludes our ability to
make more defi nite generalizations regarding the sig-
nifi cance of differences between these groups. Also, the
small number of patients in the minority ethnic catego-
ries precluded our ability to subanalyze ethnicity by
stage. Because SEER does not include information on
disease recurrence, and because cause-of-death infor-
mation is not always reliable, we are unable to analyze
other endpoints, such as time to recurrence, or disease-
free survival. Finally, historical survival data collected
over an extended period of time may not refl ect current
practices in oncology. CS should be considered within
the individual context of patient care, where specifi c
risk factors must be carefully considered.
Throughout this analysis, relative survival rates are
reported, defi ned as observed survival divided by the
expected survival for that patient. Because cause-of-
death information is often unreliable in the SEER data-
base [28], relative survival is a useful alternative measure
that assesses the proportion of excess deaths that occur
in gastric cancer patients compared to the general popu-
lation. It serves to adjust for differences in observed
survival that may be due to competing causes of death
[5]. Expected survival rates were obtained from the US
SEER 1970, 1980, and 1990 expected rate tables by
matching patients for age, sex, ethnicity, and the date at
which age was coded.
In summary, our data show that CS for gastric cancer
patients changes over time, and CS can be used as an
important adjunct to traditional survival statistics. We
have found that CS in gastric cancer shows the largest
increases over time in patients under age 65, and in
those with advanced stage disease. We hope that CS
calculations will become more common in cancer report-
ing, and that extrapolation from other epidemiological
and clinical trial datasets may allow future optimization
of risk stratifi cation for gastric cancer patients. Condi-
tional survival data provide patients, clinicians, and
researchers with more accurate prognostic information
about how risk changes over time for cancer survivors.
References
1. American Cancer Society. Cancer facts and fi gures 2006 [online]
2006 [cited 12-12-2006]. http://www.cancer.org.
2. Blanke CD, Coia LR, Schwarz RE, Bonin SR. Gastric cancer. In:
Pazdur R, Coia LR, Hoskins WJ, Wagman LD, editors. Cancer
management: a multidisciplinary approach. 9th ed. Lawrence, KS:
CMP Media, LLC; 2005. p. 279–92.
3. Pisters PWT, Kelsen DP, Powell SM, Tepper JE. Cancer of the
stomach. In: Devita VT, Hellman S, Rosenberg SA, editors.