Adjusted odds ratios and 95% confidence intervals describing the risk of death from each procedure at high-volume hospitals compared with at low-volume hospitals.
Observed operative mortality by hospital volume.
Finlayson EVA, Goodney PP, Birkmeyer JD. Hospital Volume and Operative Mortality in Cancer SurgeryA National Study. Arch Surg. 2003;138(7):721-725. doi:10.1001/archsurg.138.7.721
Although initiatives to regionalize cancer surgery are already under way, the relative importance of volume in cancer surgery is disputed.
We examined surgical mortality with 8 cancer resections in the US population to better quantify the influence of hospital volume.
Using information from the all-payer Nationwide Inpatient Sample (1995-1997), we examined mortality with 8 cancer resections (N = 195 152). After dividing patients into 3 evenly sized volume groups based on hospital procedure volume (low, medium, and high), we used regression techniques to describe relationships between hospital volume and in-hospital mortality, adjusting for patient characteristics.
Trends toward lower operative risks at high-volume hospitals were observed for 7 of the 8 procedures. However, differences between low- and highhigh-volume hospitals were statistically significant for only 3 operations (esophagectomy, 15.0% vs 6.5%; pancreatic resection, 13.1% vs 2.5%; and pulmonary lobectomy, 10.1% vs 8.9%, respectively). Although they did not reach statistical significance, absolute differences in mortality between low- and high-volume hospitals were greater than 1% for the following 3 procedures: gastrectomy, 8.7% vs 6.9%; cystectomy, 3.6% vs 2.5%; and pneumonectomy, 10.6% vs 8.9%, respectively. Mortality reductions for nephrectomy and colectomy were small. In general, in terms of absolute differences in mortality, the effect of volume was greatest in elderly patients.
Operative mortality decreases with increasing hospital volume for several cancer resections. However, volume may be most important in patients who are older and at higher risk.
BECAUSE NUMEROUS studies have documented that high-volume hospitals achieve lower operative mortality with selected high-risk procedures,1- 5 policymakers are developing strategies for concentrating selected procedures in higher-volume hospitals. The Leapfrog Group, a large coalition of private and public purchasers, is encouraging patients undergoing 5 high-risk procedures to seek care at high-volume hospitals.6 Web sites sponsored by consumer groups and private organizations are providing patients with information about volume at hospitals near them. The Center for Medicare and Medicaid Services is also exploring ways to communicate volume information to patients.
The relative importance of volume in cancer surgery, however, is disputed. Although many studies have documented an association between volume and operative mortality, the populations on which these studies were based—and thus their results—have varied widely. Some studies have relied on the National Cancer Data Base.7 This database consists of data that are contributed voluntarily by participating hospitals and their outcomes may not reflect results achieved at nonparticipating hospitals. Other studies of surgery for pancreatic, colon, and thyroid cancer have relied on regional or state-level databases.3- 5,8- 14 Their findings may not be safely extrapolated to other settings. Although some studies used nationally representative data, they have most often used the Medicare database that includes primarily patients aged 65 years and older.1,2,15 It is unclear whether operative mortality rates observed in this population are generalizable to younger patients. To better delineate the importance of hospital volume for patients undergoing 8 different cancer procedures, we studied operative mortality in the all-payer Nationwide Inpatient Sample (NIS), a nationally representative database.
We used data from the 1995-1997 NIS. The NIS is a large national database containing hospital discharge data from approximately 7 million hospital stays that contain information about patients from all payers. It is a 20% sample of all US nonfederal hospitals and contains data from approximately 1000 hospitals in 22 states. Hospitals are selected to represent 5 strata of hospital characteristics: ownership-control, bed size, teaching status, rural-urban location, and geographic region. Weights based on the sampling probabilities for each stratum are used in analysis so that the sample hospitals are representative of all US hospitals.
We identified all discharges for the 8 procedures during the 3-year period (1995-1997) using appropriate procedure codes from the International Classification of Diseases, Ninth Revision.15 To identify only resections for cancer, we restricted our sample to patients who had an accompanying cancer diagnosis code (related to the procedure of interest). To best reflect overall institutional experience with each operation, we collapsed lobectomy and pneumonectomy (lung resection) in determining hospital volume. Volume-outcome relationships were assessed separately for all 8 procedures. The primary outcome measure, operative mortality, was defined as death before hospital discharge (30-day mortality cannot be assessed in NIS).
Using unique hospital identification numbers, we assessed hospital volume by calculating the average annual volume of each of the 7 procedure volume categories over the 3 study years. Using appropriate weights, we calculated the average annual volume for each hospital separately for each of the 7 procedure volume categories.
Volume was first examined as a continuous variable. For presentation, we also considered volume as a categorical measure. In defining hospital volume categories for each procedure, we used volume cutoffs that sorted patients into 3 evenly sized groups: low, medium, and high volume (Table 1).
We used multiple logistic regression to assess the relationships between hospital volume and operative mortality, adjusting for patient characteristics.16 We used the patient as the unit of analysis, with the exposure (volume) measured at the hospital level. We first fitted separate models for each procedure. Models were run on the entire cohort and separately by age subgroups (<65 vs ≥65 years). The logarithm of hospital volume was used to establish the general form and determine statistical significance of the volume-outcome relationship. Hospital volume was also divided in terciles for exploratory and display purposes.
We adjusted for patient age group (<35, 35-44, 45-54, 55-64, 65-69, 70-74, 75-79, 80-84, ≥85 years), sex and race (black, nonblack), year of procedure, acuity of the admission (elective, urgent-emergent), patient comorbidities, and median Social Security income.16 Because patient-level socioeconomic status information is unavailable, household income was assessed at the ZIP code level (from the 1990 US Census Bureau file).
Patient comorbidities were identified using information from each patient discharge abstract. Up to 15 International Classification of Diseases, Ninth Revision diagnosis and surgical codes were used to assess comorbidity. We modified the definition in each procedure cohort to exclude conditions likely to (1) reflect the primary diagnosis or (2) represent postoperative complications (eg, acute myocardial infarction or renal failure). Individual comorbidities were then combined into a summary Charlson comorbidity index score.17
We adjusted for the net effect of clustering of mortality within hospitals by using overdispersed binary logistic models, clustering by hospital.18 Accounting for clustering of patients within hospitals had little effect on the SE of the estimates and, therefore, the width of the confidence intervals. All analyses were performed using sampling weights; all P values are 2-tailed.
Between 1995 and 1997, more than 195 000 patients in the NIS database underwent 1 of the 8 procedures. Volume criteria defining the 3 hospital volume strata varied markedly by procedure, reflecting the relative frequency with which each is performed (Table 1). With pancreatectomy, cystectomy, and esophagectomy, for example, about one third of the patients (the low-volume group) had surgery at hospitals performing fewer than 4 such procedures annually. Conversely, with colectomy, the volume threshold distinguishing low- and medium-volume hospitals was 61 procedures per year.
Patient characteristics did not vary systematically by hospital volume strata (Table 2). While patients tended to be younger at high-volume hospitals with several procedures, Charlson comorbidity indices tended to be slightly higher at higher-volume hospitals. For most procedures, there were modest trends toward increasing mean Social Security income with hospital volume. Differences in admission acuity were modest and varied by procedure. Sex and race did not vary by volume.
Overall, there were statistically significant relationships between hospital volume and observed operative mortality for 5 of the 8 procedures—colectomy, gastrectomy, esophagectomy, pancreatic resection, and pulmonary lobectomy. However, differences were significant for only 3 procedures after risk adjustment (Figure 1). Risk adjustment tended to dampen volume effects for colectomy and gastrectomy. In general, however, adjusting for patient characteristics had little effect on adjusted odds ratios of mortality by volume for most procedures.
In terms of absolute mortality rates, the importance of hospital volume varied markedly by procedure (Figure 2). For example, with pancreatic resection, observed mortality rates at low-volume hospitals were more than 10% higher than at high-volume hospitals (13.1% vs 2.5%, respectively). Relatively large risk differences were also observed for esophagectomy (low-volume vs high-volume hospitals, 15% vs 6.5%). Conversely, trivial mortality differences between low- and high-volume hospitals were observed with nephrectomy (<0.1%). For most procedures, absolute differences in observed mortality rates between low- and high-volume hospitals ranged between 0.5% and 2.0%.
We repeated our analyses stratifying by age, that is, patients younger than 65 years and patients 65 years and older (Table 3). In relative terms (odds ratios of mortality), volume-outcome associations did not differ substantially by age. As illustrated by the large confidence intervals with some procedures, this subgroup analysis was hindered by small sample sizes, which limit statistical inferences for uncommon procedures (eg, cystectomy). However, adjusted odds ratios of mortality (high- vs low-volume hospitals) in patients younger than 65 years and patients 65 years and older were strongly correlated (r = 0.92 after excluding cystectomy).
For all procedures, mortality rates were substantially higher in older patients (Table 3). For this reason, the absolute differences in operative mortality by volume tended to be larger in older patients. For pancreatic resection and esophagectomy, absolute differences in operative mortality between high- and low-volume hospitals were greater than 10% for older patients. In younger patients, operative mortality reductions at high-volume hospitals were less than 5% for the same 2 procedures. For colectomy and pulmonary resection, absolute mortality reductions were only slightly higher in older patients.
In this nationwide study, higher-volume hospitals achieved lower operative mortality rates for several major cancer resections. However, the magnitude of the difference between low- and high-volume hospitals varied across procedures. Dramatic mortality differences between low- and high-volume hospitals were observed for esophagectomy (8.5%) and pancreatic resection (10.5%). Although all were not statistically significant, absolute differences in mortality between low- and high-volume hospitals were greater than 1% for gastrectomy, cystectomy, and pneumonectomy. In contrast, volume-related differences for colectomy and nephrectomy were small. Volume-related absolute differences in mortality were greater in older patients.
Our findings that higher-volume hospitals have lower mortality rates in cancer surgery are consistent with those of numerous state-level studies. These studies have documented strong volume-outcome associations for pancreatic resection and esophagectomy.9,10,12- 14 Based on data from New York State, Hannan et al5 noted that mortality rates for patients undergoing colectomy, gastrectomy, and pulmonary lobectomy for cancer at high-volume hospitals were significantly lower than at low-volume hospitals. These analyses may not accurately represent nationwide performance since surgical performance is known to vary across states.19
Although our findings are also consistent with national studies, mortality reductions were not as large.1,2,15 There are several potential reasons why our findings differ. First, because baseline operative risks are higher in Medicare recipients than in the general population, absolute mortality reductions are likely to be greater. We confirmed this in our age-stratified analysis, where differences in mortality between low- and high-volume hospitals were greater in older patients. Second, the NIS is not a 100% national sample. Although the data are weighted to represent the general population, the states sampled by the NIS may not perfectly reflect national surgical performance. Third, because the NIS provides only in-hospital mortality, our mortality rates may be lower than in analyses that consider 30-day mortality as well.20 Fourth, because our sample was substantially smaller than the national Medicare database, our analysis may not have had the power to detect important differences in operative mortality. For example, although mortality reductions were greater for cystectomy than pulmonary lobectomy, these reductions did not achieve statistical significance because fewer patients underwent cystectomy—about 1700 patients in each volume stratum. For our analysis to detect a significant difference, there would need to be more than 3700 patients in each volume stratum. Fifth, because of sample size constraints, we were only able to define 3 volume categories for each procedure. Larger studies have had more volume strata (eg, quintiles) and, thus, have compared performance at greater extremes of volume.
Because we relied on administrative data, we may not have fully accounted for case-mix differences across hospital volume strata. Administrative data are limited in their ability to discern severity of illness, and the problems of using administrative data for risk adjustment are well known.21- 24 However, although patients tended to be slightly older at low-volume hospitals for several procedures, comorbidity did not vary systematically by volume. Although we cannot rule out unmeasured differences in case-mix across volume strata, there is little evidence that patients at low-volume hospitals are sicker. However, studies based on clinical data have not reported weaker volume-outcome relationships than those based on administrative data. In fact, of the 16 studies with the highest-quality scores in a recent structured literature review conducted by the Institute of Medicine (all based on clinical data), all reported statistically significant relationships between volume and mortality.25
While our study provides further evidence that volume-based referral initiatives would save lives, such policies have several indirect harms for patients and hospitals. For patients in isolated rural areas, regionalizing surgery could imply unreasonable travel burdens, delays in initial evaluation, and problems with continuity of care after surgery. Loss of surgical caseload at small rural hospitals could threaten their financial viability or their ability to recruit and retain surgeons. In addition, loss of procedure volume would also adversely affect a hospital's or surgeon's ability to manage emergent cases (eg, emergent colon surgery) and proficiency in related elective procedures.
To avoid these potential harms, volume-based referral initiatives need to target both procedure and patient groups carefully. First, policymakers should focus on procedures with the largest volume outcome associations (eg, esophagectomy and pancreatic resection). Because these procedures also tend to be relatively infrequent, volume-based referral would be less disruptive for both patients and hospitals. Second, policymakers should focus on patient groups likely to benefit most. Prior work has suggested that in coronary bypass surgery, hospital volume is most important for high-risk patients.26 Additional work is needed to further characterize the types of patients who would benefit most from referral to high-volume hospitals for other procedures. Meanwhile, as long as differences in mortality across hospitals persist, policies concentrating selected procedures at high-volume hospitals will continue to gain support. The challenge for both surgeons and policy makers will be to identify strategies that maximize patient benefit while minimizing harms.
Corresponding author: Emily V. A. Finlayson, MD, Department of Surgery, Room S-343, University of California, San Francisco, San Francisco, CA 94143 (e-mail: firstname.lastname@example.org).
Accepted for publication December 21, 2002.
This study was supported by the Center for Medicare and Medicaid Services, grant R01 HS10141-01 from the Agency for Healthcare Research and Quality, Bethesda, Md, and a Career Development Award from the VA Health Services Research and Development Program, Washington, DC (Dr Birkmeyer).
The views expressed herein do not necessarily represent the views of the Center for Medicare and Medicaid Services, the Department of Veterans Affairs, or the US government.