Schrag D, Cramer LD, Bach PB, Cohen AM, Warren JL, Begg CB. Influence of Hospital Procedure Volume on Outcomes Following Surgery for Colon Cancer. JAMA. 2000;284(23):3028-3035. doi:10.1001/jama.284.23.3028
Author Affiliations: Departments of Epidemiology and Biostatistics (Drs Schrag, Bach, and Begg and Ms Cramer), Medicine (Drs Schrag and Bach), and Surgery (Dr Cohen), Health Outcomes Research Group, Memorial Sloan-Kettering Cancer Center, New York, NY; and the Applied Research Branch, National Cancer Institute (Dr Warren).
Context Survival following high-risk cancer surgery, such as pancreatectomy
and esophagectomy, is superior at hospitals where high volumes of these procedures
are performed. Conflicting evidence exists as to whether the association between
hospital experience and favorable health outcomes also applies to more frequently
performed operations, such as those for colon cancer.
Objective To determine whether hospital procedure volume predicts survival following
colon cancer surgery.
Design, Setting, and Participants Retrospective cohort study of data from the Surveillance, Epidemiology
and End Results–Medicare linked database on 27 986 colon cancer
patients aged 65 years and older who had surgical resection for primary adenocarcinoma
diagnosed between 1991 and 1996.
Main Outcome Measures Thirty-day postoperative mortality and overall and cancer-specific long-term
survival, by hospital procedure volume.
Results We found small differences in 30-day postoperative mortality for patients
treated at low- vs high-volume hospitals (3.5% at hospitals in the top-volume
quartile vs 5.5% at hospitals in the bottom-volume quartile). However, the
correlation was statistically significant and persisted after adjusting for
age at diagnosis, sex, race, cancer stage, comorbid illness, socioeconomic
status, and acuity of hospitalization (P<.001).
The association was evident for subgroups with stage I, II, and III disease.
Hospital volume directly correlated with survival beyond 30 days and also
was not attributable to differences in case mix (P<.001).
The association between hospital volume and long-term survival was concentrated
among patients with stage II and III disease (P<.001
for both). Among stage III patients, variation in use of adjuvant chemotherapy
did not explain this finding.
Conclusion Our data suggest that hospital procedure volume predicts clinical outcomes
following surgery for colon cancer, although the absolute magnitudes of these
differences are modest in comparison with the variation observed for higher-risk
The complexity of health care processes makes identification and measurement
of the critical components of high-quality care especially challenging.1 Because hospital procedure volume is relatively easy
to measure and is assumed to be a proxy for experience, it has long been examined
as a predictor of clinical outcomes, and a volume-outcome relationship has
been observed for a wide variety of surgeries.2- 13
Concentration of surgeries in high-volume centers has been considered a strategy
to improve the quality of care,2,14- 16
and in select instances, policies to achieve this goal have been implemented.17
For cancer patients, large population-based studies have demonstrated
that hospital procedure volume can have a profound effect on outcomes following
operations associated with high mortality, such as pancreatectomy.18- 20 Some prior studies
have suggested that a volume-outcome effect may also exist for colon cancer
surgery, which is performed more frequently but with less substantial morbidity
and mortality.3- 5,21- 26
However, these analyses have been limited by either sample size, lack of population-based
case ascertainment, geographic diversity, insufficient clinical detail for
risk adjustment, or incomplete mortality data. Recently, linkage of the Surveillance,
Epidemiology, and End Results (SEER) registries to Medicare claims created
a resource that combines the necessary ingredients for volume-outcome analyses
for the US population aged 65 years and older.27
We used SEER-Medicare data to identify a population-based cohort of
colon cancer patients to determine whether hospital procedure volume predicts
short- and long-term survival following primary surgery. We hypothesized that
hospital volume would predict both postoperative mortality and long-term survival,
but anticipated that the absolute magnitude of these associations would be
smaller than those observed for infrequently performed higher-risk operations,
such as pancreatectomy and esophagectomy.
The SEER registries ascertain all incident cancer cases diagnosed in
5 states and 6 US metropolitan areas, representing approximately 14% of the
US population.28 Information is collected on
each incident cancer, including the primary site and histology classified
according to the International Classification of Diseases
for Oncology, Second Revision (ICD-O-2)29
schema, the tumor stage at diagnosis, and patient demographics.
The Medicare program provides health coverage for 97% of the US population
aged 65 years and older. The Medicare Provider Analysis and Review files provide
details of all hospitalizations for persons eligible for Medicare Part A.
To receive payment, hospitals submit medical claims coding up to 10 diagnoses
and 10 procedures using the International Classification
of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) classification.
For the 96% of Medicare beneficiaries who opt for Part B coverage, claims
for care delivered in hospital outpatient departments and physicians' offices
are also recorded. Medicare records document date of death based on information
provided by the Social Security Administration. The SEER-Medicare data have
been linked to facilitate population-based studies of the outcomes of cancer
treatment. Ninety-four percent of patients in SEER aged 65 years and older
have been successfully linked to their Medicare records.27
All Medicare-enrolled patients aged 65 years and older diagnosed as
having primary colon cancer in a SEER area during the years 1991 to 1996 were
potentially eligible for inclusion in our study. Colon cancers were defined
using SEER codes for cancer sites 18.0 through 18.9, and 19.9; thus, tumors
arising in the rectosigmoid were included. We restricted our cohort to patients
with a histologic diagnosis of adenocarcinoma (SEER histology codes 8140,
8210, 8211, 8220, 8221, 8260, 8261, 8262, 8263, 8470, 8480, 8481, and 8490).
Diagnoses noted exclusively on death certificates or at autopsy were excluded,
as were those in which the month of diagnosis was unknown. Patients enrolled
in a health maintenance organization were excluded from our cohort (16.5%
of patients) because detailed claims are not submitted to Medicare by health
We searched Medicare claims for colon cancer surgeries performed within
6 months of primary diagnosis. Operations were defined according to the ICD-9-CM classification system (45.7x, 45.8, 48.4x, 48.5x,
and 48.6x); thus, patients operated on exclusively for local resection or
creation of an ostomy were excluded. Only patients undergoing surgery at hospitals
located in 1 of the 9 states containing 1 of 11 SEER registries were included
in the analysis because hospital volume could not be reliably measured at
institutions outside SEER areas.
The 3 outcomes assessed were 30-day postoperative mortality, overall
survival, and colon cancer-specific survival. Postoperative mortality was
defined as death within 30 days of hospitalization for surgical resection.
Date of hospitalization served as a proxy for date of surgery since it is
more reliably coded in the Medicare database. Survival was defined by the
interval from the date of hospitalization for resection until either death
as reported to Medicare, or December 31, 1998, when censoring occurred. SEER
reports cancer-specific mortality based on state death certificates but these
vital statistics were current only through December 31, 1996. Therefore, colon
cancer-specific survival was examined for the 98.5% of our cohort (27 561/27 986)
who had available state death certificates or had not died prior to this date.
We hypothesized that greater use of adjuvant chemotherapy at high-volume
hospitals might partially explain an observed volume-outcome relationship.
We examined use of adjuvant chemotherapy among patients with stage III cancer
who survived 3 months postoperatively, and were also enrolled in Medicare
Part B, which is required for coverage of outpatient services. Those patients
who had at least 1 claim for chemotherapy or its administration at any point
during the 3-month postoperative period were considered recipients.
Hospitals were ranked by volume according to the number of operations
performed between 1991 and 1996, an approach that has been previously validated.18 Examination of fluctuations of individual hospital
volumes on a year-by-year basis demonstrated great stability. To avoid the
possibility of selecting cut points with maximal P
values, primary statistical analyses were performed using the Mantel-Haenszel
test for trend without aggregation of the data into discrete volume categories.
To facilitate display of our results and to adjust survival in a Cox proportional
hazards model, we defined quartiles of hospital procedure volume (low, medium,
high, and very high) based on the volume of operations performed on members
of our cohort during the 6-year study period.
We used information on tumor size, nodal involvement, and spread coded
in the SEER database to stage patients according to the American Joint Committee
on Cancer schema. Patients with missing information about tumor extent, nodal
involvement, or metastases were classified as unstaged. To adjust for potential
confounding based on the severity of noncancer medical illness, we used Romano's30 modification of the comorbidity index originally
developed by Charlson.31 We examined all available
inpatient Medicare claims for the 12 months prior to the index surgical admission,
as well as claims during the index admission, and assigned patients the maximal
comorbidity observed. We used the median income in the census tract of residence
to adjust for differences in patients' economic status.
Colon cancer resections performed on an emergency basis may be associated
with high mortality. We used the Medicare claim code for emergent hospital
admission and ICD-9 codes for emergent indications
for surgery bowel obstruction (560.89, 560.90) and perforation (569.83) to
permit an adjustment in multivariable analysis.
The relationship between hospital procedure volume and postoperative
mortality was examined using the Mantel-Haenszel test for trend. While we
used the patient as the unit of analysis, we also performed a modified version
of the Mantel-Haenszel test that adjusts for within-hospital correlations
in the data.32 Multiple logistic regression,
with vital status at 30 days as the outcome and hospital procedure volume
as a continuous predictor, was used to adjust for potential confounding by
sex, race, age at diagnosis, cancer stage, comorbidity, socioeconomic status,
and the presence of emergent indications for surgery according to the categories
shown in Table 1.
The impact of hospital volume on survival is displayed using the Kaplan-Meier
method. The Cox proportional-hazards method was used to examine the effects
of potential confounders. The likelihood ratio test was used to compare a
model that had all variables except procedure volume with a model that had
all variables including procedure volume. All P values
are 2-sided. When we developed our research protocol, we calculated the effect
size that would allow for detection with a 2-sided significance level of .05
and 90% power for each planned analysis. With a sample size of 27 000
colon cancer patients and a 4% postoperative mortality rate, we had power
to detect a 0.8% difference between low- and high-volume hospitals; for a
subgroup of 7000 patients, we could detect a 1.5% difference; for a subgroup
of 3500 patients, a 2.1% difference.
A total of 47 495 Medicare-eligible patients received an antemortem
diagnosis of primary colon cancer at age 65 years and older during the years
1991-1996 in SEER areas. We sequentially excluded 2280 patients with histologies
other than adenocarcinoma, 71 lacking a month of diagnosis, and 2934 with
in situ tumors. Among the remaining 42 210 patients, 7999 were excluded
because they were enrolled in a health maintenance organization or not enrolled
in Medicare Part A at the time of diagnosis. From this group of 34 211,
a total of 28 475 had surgical resection, and among these, 27 986
had surgery performed at a hospital located in a SEER area. The demographic
and clinical characteristics of the 27 986 patients are displayed in Table 1.
In our cohort, colon cancer resections were performed at 611 different
hospitals between 1991 and 1996. Hospital volume over the 6-year period ranged
from 1 to 57 for the 440 low-volume hospitals (72%); 58 to 112 for the 89
medium-volume hospitals (15%); 113 to 165 for the 51 high-volume hospitals
(8%); and 166 to 383 for the 31 very high-volume hospitals (5%). The numbers
and characteristics of patients in each volume quartile category are shown
in Table 1. The age, sex, and
comorbidity of patients were similar across strata of hospital volume. Patients
with race coded as other than white or black, those with unstaged tumors,
and those with low socioeconomic status were more likely to undergo surgery
at a low-volume hospital.
As shown in Table 2, the
absolute magnitude of variation in 30-day postoperative mortality at hospitals
with different procedure volumes was small. For example, the difference in
30-day mortality for patients operated on at the very high-volume compared
with the low-volume hospitals was 2% (3.5% vs 5.5%). However, a consistent
association between higher postoperative mortality and lower surgical procedure
volume was evident (P<.001) and persisted after
inclusion of potential confounders in multivariable logistic regression (P<.001). In subgroup analyses, we found that hospital
volume was a significant predictor of mortality for patients with stage I,
II, and III tumors, but not for patients with stage IV or unstaged disease,
in which smaller sample sizes precluded our ability to detect effect sizes
of less than 2%.
The survival curves for patients treated at institutions in each volume
quartile illustrate a clear association between procedure volume and overall
survival (P<.001; Figure 1A). The difference in 5-year
mortality for patients operated on at the very high- vs the low-volume hospitals
was 4.4% (54.8%-50.4%; Table 2).
Since a 2% absolute mortality difference is evident at 30 days, 45% (2/4.4)
of the survival difference appears to be attributable to the immediate postoperative
period and 55% to more distal events. Figure
1B demonstrates that similar results are obtained for colon cancer-specific
Adjusted risk ratios and confidence intervals for overall mortality
for patients in each hospital volume category compared with patients operated
on at the very high-volume institutions are shown in Table 2. The volume-outcome association remained highly significant
for both overall mortality (P<.001) and colon
cancer-specific mortality (P<.001) after adjusting
for other variables.
We examined subgroups of patients with identical American Joint Committee
on Cancer tumor stage (Figure 2).
Whereas hospital volume was predictive of survival for patients with stage
II (P<.001 adjusted) and stage III disease (P<.001 adjusted), it was not significant for patients
with either stage I or stage IV disease.
Our adjusted models included variables associated with either shorter
colon cancer survival (advanced clinical stage, comorbidity, obstruction,
perforation) or shorter life expectancy (male sex, older age, black race,
and low socioeconomic status). Male sex, older age, black race, advanced clinical
stage, high comorbidity, low income, obstruction or perforation, and emergent
hospitalization were independent predictors of poor prognosis but adjusting
for these variables did not change our results.
Variation in synchronous hepatic resection (ICD-9-CM codes 50.22, 50.3, 50.4) did not confound the volume-outcome association
because it was performed at a similar low frequency (0.4%-0.6%) at hospitals
in each volume quartile. We considered the possibility that regional differences
in care might account for our results.33 Although
high-volume hospitals were more highly concentrated in some SEER regions (Connecticut
and Detroit, Mich) than in others (Iowa and Los Angeles, Calif), stratification
by registry did not change our results because there was minimal geographic
variation in mortality rates.
The pronounced association between volume and long-term survival for
patients with stage III disease (Figure 2) prompted us to examine patterns of postoperative chemotherapy
use as a possible process measure that might account for this observation.
Among the 6423 patients with stage III tumors who survived 3 months postoperatively
and were enrolled in Medicare Part B, 3519 (54.8%) had at least 1 Medicare
claim for chemotherapy within 3 months of surgery. Specifically, 51.2% of
patients at low-volume; 56.6% at medium-volume; 55.6% at high-volume; and
55.5% at very high-volume hospitals received chemotherapy within 3 months
of primary surgery. Although this trend was marginally significant (P = .02; and P = .07 when adjusted
for within-hospital correlation), when we added chemotherapy use to our Cox
model for stage III, hospital procedure volume remained a significant predictor
of survival in adjusted analyses.
Among a population-based cohort of Medicare beneficiaries with primary
colon cancer, we found that hospital procedure volume predicts both short-
and long-term survival following surgical resection. We failed to find any
evidence that underlying differences in the characteristics of patients accounted
for our results. Although the association between postoperative mortality
and hospital procedure volume is statistically significant, the absolute magnitude
of the difference (2%) was more modest than the 7% to 15% differences observed
for pancreatectomy and esophagectomy.18
Our results are consistent with those recently reported by Harmon et
al,3 who analyzed the association between hospital
volume and inhospital mortality for 9739 colorectal cancer patients treated
at 50 hospitals in Maryland. In our cohort, the top 5% (31/611) of hospitals
cared for 25% of patients; in their cohort, the top 12% (7/50) of hospitals
cared for 32% of patients. Both studies show that in the United States, colon
cancer surgery is currently performed at many hospitals with very low-case
volumes, that there is a statistically significant correlation between high
volume and favorable outcomes not attributable to differences in case mix,
and that the order of magnitude of the absolute postoperative mortality difference
is small (1.7%-2%).
Previous analyses demonstrating a relationship between hospital case
volume and clinical outcomes have suggested that selected procedures should
be regionalized and services restricted to centers performing a minimum number
of cases.2,15 To illustrate the
potential consequences of colon cancer surgery regionalization, we have calculated
the number of hospitals that would have to discontinue colon cancer surgery
and the number of patients who would need to be referred elsewhere if mandatory
minimum volume thresholds were implemented (Table 3). Under the optimistic assumptions that rerouting patients
to very high-volume hospitals would be feasible, cause no adverse consequences,
and achieve surgical outcomes similar to those for patients in the top-volume
quartile, our results show that regionalization would affect many institutions
and require relocation of many patients to obtain modest, although appreciable,
increments in survival. If our results are extrapolated to the approximately
70 000 colon cancer resections performed annually in the United States,
in a best-case scenario, rerouting patients treated at hospitals in the low-volume
quartile to the very high-volume quartile could potentially avert 350 postoperative
deaths and a total of 770 deaths 5 years after colon cancer surgery. Whether
regionalization of colon cancer surgery is warranted to achieve a benefit
of this magnitude should be a matter of public policy debate.
In urban areas such as Detroit and Atlanta, Ga, hospitals with large
caseloads are located in proximity to low-volume centers, but in areas such
as Iowa and Utah, regionalization policies mandating care in high-volume institutions
could require patients to travel long distances. As a result, the efficacy
and expense of alternatives to regionalization, such as continuing education
for surgical care teams at hospitals with low-case volumes, merit further
study. In addition, we concur with Hillner et al2
who recently emphasized that identification of the mechanisms underlying variation
in outcomes should facilitate initiatives tailored to address specific shortcomings
and is therefore a research priority.
Several concerns regarding our analysis must be noted. First, the potential
for inaccurate coding exists for any claims-based analysis.34- 36
However, the lack of ambiguity regarding the colon cancer diagnoses coupled
with the fact that complete coding for major surgical procedures favorably
affects hospital and physician reimbursement suggests that the claims-based
approach we used should be more accurate than it may be for other conditions.
Second, we determined surgical volume based only on the number of operations
performed in the Medicare population. However, this method has been validated
and Medicare case volume appears highly correlated with total volume.18
Although the SEER cohort is population-based, generalizability of our
analysis may be limited by the restriction of our study cohort to the subset
of Medicare-eligible patients older than 65 years who were not enrolled in
a health maintenance organization at diagnosis.37
Nevertheless, because the median age of colon cancer diagnosis is 71 years,28 and less than 20% of patients were health maintenance
organization enrollees, our cohort is representative of a substantial proportion
of patients in the United States.
Why do high-volume hospitals achieve superior outcomes? Our analysis
suggests that the answer is not attributable to differences in patient characteristics.
The 30-day mortality differences suggest that either intraoperative and/or
immediate postoperative care vary with institutional caseload. In part, the
effect may result from the skill of the individual surgeon.38
However, other analyses have shown that hospital volume is a stronger determinant
of outcomes than individual surgeon volume suggesting that access to an entire
team of health care professionals (surgeons, anesthesiologists, nurses, radiologists)
is important.3 It is unlikely that the association
between hospital volume and both overall and colon cancer-specific survival
is attributable to a single mechanism. Conceivably, surgeons at high-volume
hospitals perform more meticulous dissections. Evidence that local recurrences
were more common at low-volume hospitals would lend support to the interpretation
that surgical expertise is the primary determinant of outcome, but this detail
was not available from medical claims. Patients at high-volume hospitals may
undergo more careful postoperative surveillance or receive more intensive
subsequent treatment. However, no strong association between adjuvant chemotherapy
treatment and hospital procedure volume was evident. Our claims-based approach
did not permit us to examine cumulative dose or dose intensity of chemotherapy,
and such differences might account for at least a small proportion of the
Caution is warranted in interpreting our results and indeed those of
all volume-outcome studies.16,39,40
Our analysis cannot demonstrate the direction of any causal relationship between
volume and outcome. While we presume that high volumes contribute to good
outcomes it is also plausible that good outcomes lead to high volumes. We
emphasize that the implication of our analysis is not that colon cancer surgery
should be limited to high-volume institutions. Rather, we intend that it provoke
in-depth scrutiny of the processes of care at high-volume institutions that
determine their success and those at low-volume institutions that may account
for their relative shortcomings. Our results underscore the need for further
research to identify those specific features and processes of care that underlie
the volume-outcome relationship. This insight should help policymakers and
should enable hospitals to design and implement strategies to improve the
quality of care.