All 1302 patients with bladder carcinoma (International Classification of Diseases, 9th revision [ICD-9] codes 188.0–188.9 and 236.7) who underwent total cystectomy (ICD-9 codes 57.7, 57.71, 57.79, and 68.8) between January 1, 1999 and December 31, 2001 were identified from a statewide claims database. Patients with carcinoma in situ and those who underwent partial cystectomy were excluded.
Patients were identified from the Texas Hospital Discharge Public Use Data File, which includes claims from all hospitals in Texas except Veterans Administration and military hospitals. Clinical, demographic, and economic data obtained from these claims were linked to socioeconomic data from the 2000 U.S. Census (by patients' postal code) and to hospital ownership, facility, and staffing information from the Center for Medicare and Medicaid Services' (CMS) Hospital Cost Report Information System, Provider of Services files, and the American Hospital Association Survey (AHA) (by hospital provider number).23–25
Two outcomes, inpatient mortality and postoperative complications, were evaluated. Inpatient mortality was identified by discharge status coding from each claim. Complications were identified by searching the 10 diagnosis fields in each claim for ICD-9 codes corresponding with algorithms developed by Iezzoni et al. for use with administrative data sets.26 The complications evaluated (and their respective ICD-9 codes) included bacteremia (038.xx and 790.7), wound infection (998.59 and 998.51), pulmonary compromise (514, 518.4, 518.5, 518.81, 518.82, 799.0, and 799.1), pneumonia (481, 482.00–482.99, 483, 485, and 486), deep venous thrombosis (451.11, 451.19, 451.2, 451,81, and 453.8), pulmonary embolus (415.1), reoperation (procedures: 54.12 and 54.61), postoperative coma or shock (780, 780.03, 780.03, and 998), acute myocardial infarction (410.00–410.91), arrhythmia (426, 427.41, 427.42, and 429.4), and cardiac arrest or shock (37.91, 427.5, and 799.1). Preexisting conditions were excluded by excluding conditions coded in either the primary or admitting diagnosis fields. Preexisting conditions also were excluded logically because of the elective nature of the timing of radical cystectomy. Because cystectomy is very rarely an emergency, it typically would not be attempted if conditions such as bacteremia, pneumonia, or myocardial infarction were present. The specific complications were chosen because of their correlation with the quality of perioperative care and because of their high sensitivity and specificity in comparisons of claims for surgical procedures with medical records.27, 28 Although claims data were reported to have a low sensitivity for medical complications in previous studies (35%), the sensitivity exceeded 60% for surgical complications. We included only those with a sensitivity ≥ 60%. Failure to rescue was defined as inpatient mortality after 1 of the 10 complications occurred.29
Patient-level factors included age, gender, race (white, African American, Asian, and others), Hispanic ethnicity, and distance from the closest high-volume hospital, which was computed between the center point of each patient's ZIP code of residence and the center point of the ZIP code of the closest high-volume hospital. Comorbid conditions were coded using the Dartmouth Manitoba adaptation of the Charlson comorbidity score.30, 31 The six procedure fields from each claim were searched for lymph node dissections and urinary diversion procedures. Patient-level socioeconomic information included payor (commercial, Medicare, Medicaid, self, and others) and type of health plan (health maintenance organization [HMO], preferred provider organization [PPO], or fee for service). Patients' educational level, income, and primary language were not available. However, the percentages of high school graduates and English speakers and the median family income corresponding to each patient's ZIP code were obtained from the 2000 U.S. census. These factors have been shown to be valid indicators of socioeconomic status.32, 33 For analysis, these factors were dichotomized based on the U.S. median value from the 2000 U.S. census (80.4% of U.S. residents were high school graduates, 82.1% spoke English, and had a median income of $41,994).
Characteristics of the hospitals were obtained from the CMS Hospital Cost Report Information System and Provider of Services files, and from the AHA Database, by matching based on hospital provider numbers and verifying the name and address of the provider. All 133 hospitals were matched to the CMS and AHA sources. Hospital ownership, for-profit and teaching status, number of beds and critical care beds, occupancy, and annual number of surgical procedures performed were studied. Staffing-to-occupied bed ratios were constructed based on the number of professional staff (registered nurses, licensed practical nurses, and respiratory therapists) and the mean annual number of occupied bed days.34, 35 Surgeon identifiers were not reported by the majority of Texas hospitals; therefore, analysis of surgeon volume was not possible.
Many previous authors have categorized volume of procedures2–4, 6 whereas others have advocated analysis of volume as a continuous variable.1, 5, 17, 18 We categorized volume to facilitate decision-making by identifying a threshold above or below which action is appropriate. The cohort was divided into 3, approximately equal groups, such that each group contained at least 25% of patients and at least 5 hospitals, and hospitals with the same volume were in the same group. After this procedure, there were 105 low-volume hospitals (≤ 3 cystectomies performed annually), 23 moderate-volume hospitals (4–10 cystectomies performed annually), and 5 high-volume hospitals (> 10 cystectomies performed annually). This method resulted in the same thresholds used for the presentation of data in prior studies of cystectomy.22
Differences with regard to patient and hospital factors among the volume groups were tested univariately using the Pearson chi-square test for categoric variables and the Student t test for continuous variables. The impact of these factors on mortality and complication rates was examined using generalized estimating equations (GEE), with a compound symmetric working correlation matrix, to adjust for the clustering of multiple patients within hospitals. Because many patient and hospital factors are correlated, multiple-variable GEE models (with logit links) were developed to estimate the unique contribution of each factor to inpatient mortality and to the development of complications. Two sets of models were developed. The first model replicated previous analyses of volume-outcome correlations in cancer surgeries. Patient factors with a P value of less than 0.10 on univariate analysis and one hospital factor (annual volume of cystectomies) were examined using GEE to adjust for within-hospital correlation. In the second model, hospital factors with a P value less than 0.10 on univariate analysis were added. Finally, we developed a multiple-variable GEE model of failure to rescue using all patient and hospital factors. All multiple-variable models were adjusted for advanced age and comorbidities; all P values were two-sided.