Mary Vaughan‐Sarrazin, PhD

Study Sites

From July 1, 2001 to June 30, 2003, the MCH trial1113 was a prospective, multicenter, observational trial of the care provided by hospitalists to patients admitted to general medical services at 6 academic medical institutions. There were 31,000 consecutive admissions to the general medical services of these participating sites: University of Chicago (Chicago, IL), University of Wisconsin Hospital (Madison, WI), University of Iowa (Iowa City, IA), University of California at San Francisco (San Francisco, CA), University of New Mexico (Albuquerque, NM), and Brigham and Women's Hospital (Boston, MA). The study was approved by the institutional review boards (IRBs) at each of the 6 participating institutions.

MCH Study Patients

Patients were eligible if they were admitted to the general medical services under the care of a hospitalist or nonhospitalist physician. Regardless of the admitting provider, each medical service was composed of rotating senior and junior resident physicians in all 6 sites. Furthermore, patients were 18 years of age or older, and were able to give consent themselves or had an appropriate proxy. Patients with mini‐mental status score of 17 (out of 22), admitted under their primary care physician or to an inpatient gastroenterology service, or transferred from another hospital, were excluded. The MCH study was designed to study the outcomes and efficiency in patients admitted for CHF, pneumonia, UGIH, and end‐of‐life care.

Acute UGIH Patients

Within the MCH‐eligible patients, we identified those with acute UGIH using the following International Classification of Diseases, 9th edition (ICD‐9) codes assigned at discharge: esophageal varices with hemorrhage (456.0, 456.20); Mallory‐Weiss syndrome (530.7); gastric ulcer with hemorrhage (531.00531.61); duodenal ulcer with hemorrhage (532.00532.61); peptic ulcer, site unspecified, with hemorrhage (533.00533.61); gastrojejunal ulcer with hemorrhage (534.00534.61); gastritis with hemorrhage (535.61); angiodysplasia of stomach/duodenum with hemorrhage (537.83); and hematemesis (578.0, 578.9). We also confirmed the diagnosis of UGIH by reviewing patient medical records for observed hematemesis, nasogastric tube aspirate with gross or hemoccult blood, or clinical history of hematemesis, melena, or hematochezia.14, 15

Data

All data were obtained from the 6 hospitals' administrative records, patient interviews, and medical chart abstractions. Dates of admission and discharge, ICD‐9 diagnosis codes, insurance type, age, race, and gender were obtained from administrative data. One‐month follow‐up telephone interviews assessed whether or not patient had any follow‐up appointment or hospital readmissions. Trained abstractors from each site performed manual chart reviews using a standard data collection sheet. The ICD‐9 code designation and chart abstraction methodology were developed prior to the initiation of the study to ensure consistent data collection and reduce bias.

The following data elements were collected: comorbidities, endoscopic findings, inpatient mortality, clinical evidence of rebleeding, endoscopic treatment or gastrointestinal (GI) surgery to control bleeding, repeat EGD, ICU transfer, decompensated comorbid illness requiring continued hospitalization, and blood transfusion (packed red cells, plasma, platelets). Clinical evidence of rebleeding was defined as either hematemesis or melena with decrease in hemoglobin of 2 g in 24 hours with or without hemodynamic compromise.14, 15 For the purpose of this study, recurrent bleeding was defined as clinical evidence of rebleeding, emergency GI surgery for control of UGIH, or repeat EGD before discharge. Furthermore, a composite endpoint termed total complications encompassed all adverse outcomes related to the UGIH hospitalization. The 30‐day readmission variable was defined using readmission identified in administrative records and a 30‐day follow‐up phone call. To guard against recall bias, self‐report data was only included for nonsite admissions.

We defined efficiency in terms of costs and LOS. Total hospital costs were measured using the TSI cost accounting system (Transition Systems, Inc., Boston, MA; now Eclipsys Corporation)16, 17 at 5 out of the 6 participating sites. TSI is a hospital cost accounting software system that integrates resource utilization and financial data already recorded in other hospital databases (such as the billing system, payroll system, and general ledger system).17 Hospital LOS was defined as the number of days from patient admission to the general medicine service until patient discharge.

Provider Specialization: Hospitalists vs. Nonhospitalists

The study was designed as a natural experiment based on a call cycle. The hospitalist‐led teams at each institution alternated in a 4‐day or 5‐day general medicine call cycle with teams led by traditional academic internal medicine attending physicians. All patients were assigned to teams according to their position in the call cycle without regard to whether the attending physician was a hospitalist or a nonhospitalist. Hospitalists are physicians whose primary professional focus is the general medical care of hospitalized patients.18, 19 As previously reported in a related MCH work,11 a hospitalist was also defined as a provider who spends at least 25% of his or her time on an academic inpatient general medicine service. Nonhospitalist physicians were most often outpatient general internal medicine faculty or subspecialists, who attended 1 month per year. Physicians were classified as hospitalists or nonhospitalists according to the designations provided by each site.

UGIH‐specific Confounders

From chart abstraction, we captured severity of illness, comorbidity, and performance of early EGD, variables that can confound analysis in UGIH. To capture severity of illness, a complete Rockall risk score was calculated for each patient. The complete Rockall uses 3 clinical variables (age, shock, and comorbidity) and 2 endoscopic variables (endoscopic diagnosis and stigmata of recent hemorrhage).5, 20 A complete Rockall score of 2 is considered low‐risk for rebleeding or death following admission.21, 22 The accepted definition of low‐risk is <5% recurrent bleeding and <1% mortality risk. A complete Rockall score of 3 to 5 is considered moderate‐risk while 6 is considered high‐risk. Comorbidity was measured using the Charlson comorbidity index.23 Performance of early endoscopy, usually defined as endoscopy performed within 24 hours from presentation, was previously shown to decrease LOS and need for surgical intervention in patients with acute UGIH.24, 25 Documented times of presentation to the emergency department and time of endoscopy performance were collected to calculate for the rate of early endoscopy in our study population.

Statistical Analysis

All statistical analyses were performed using SAS Version 9.1 for Windows (SAS Institute, Cary, NC).

Differences in baseline demographic characteristics of patients and their endoscopic findings were compared between the 2 types of providers. Univariate analyses were also performed to compare the differences in adverse outcomes, LOS, and costs between patients cared for by hospitalists and nonhospitalists. Chi‐square tests were used for categorical variables; while both Wilcoxon rank sum test and Student's t test were used in the analysis of continuous variables.

Next, we performed multivariable analyses to determine the independent association between hospitalist care and the odds of the patients having certain outcomes. However, to prevent overfitting, we only developed regression models for adverse outcomes that have at least 20% event rate.

Multivariable regression models were developed separately for LOS and costs. In contrast with the models on outcomes, analyses of LOS and costs were restricted to: (1) patients who were discharged alive; and (2) to cases with LOS and costs values within 3 standard deviations (SDs) of the mean because of the skewed nature of these data.

All models were adjusted for age, gender, race, insurance type, complete Rockall risk score, performance of early EGD, Charlson comorbidity index, and study site. Final candidate variables in the models were chosen based on stepwise selection, a method very similar to forward selection except that variables selected for the model do not necessarily remain in the model. Effects were entered into and then removed from the model in such a way that each forward selection step can be followed by 1 or more backward elimination steps. The stepwise selection was terminated if no further effect can be added to the model or if the current model was identical to the previous model. The stepwise selection model was generated using statistical criterion of alpha = 0.05 for entry and elimination from the model. Variables that can be a profound source of variation, such as study site and treating physician, were included in the model irrespective of their statistical significance.

To account for clustering of patients treated by the same physician, we used multilevel modeling with SAS PROC GLIMMIX (with random effects). For outcomes (categorical variables), we utilized models with logit‐link and binomial‐distributed errors. As for efficiency (continuous variables with skewed distribution), the multivariable analyses used a generalized linear model with log‐link and assuming gamma‐distributed errors.

Patient Characteristics and Endoscopic Diagnoses

Out of 31,000 patients, the study identified a total of 566 patients (1.8%) with acute UGIH (Table 1). However, 116 patients transferred from another hospital were excluded as their initial management was provided elsewhere, giving a final study sample of 450 patients. Overall, there are 163 admitting physicians from 6 sites, with 39 (24%) classified as hospitalists and 124 (76%) as nonhospitalists. Forty‐two percent (177/450) of patients were cared for by hospitalists. Compared to nonhospitalists, patients admitted to the hospitalist service were older (62.8 vs. 57.7 years, P < 0.01) and with third‐party payor mix differences (P < 0.01). However, there were no statistical differences between patients attended by hospitalists and nonhospitalists with regard to Complete Rockall risk score, Charlson comorbidity index, performance of early endoscopy, and mean hemoglobin values on admission. Upper endoscopy was performed in all patients with distribution of the 3 most common diagnoses being similar (P > 0.05) between hospitalists and nonhospitalists: erosive disease (49.7% vs. 54.6%), peptic ulcer disease (PUD) (48% vs. 46.9%), and varices (18.6% vs. 14.7%).

Clinical Outcomes

Between hospitalists and nonhospitalists, unadjusted outcomes were similar (P > 0.05) for mortality (2.3% vs. 0.4%), recurrent bleeding (11% vs. 11%), need for endoscopic therapy (24% vs. 22%), ICU‐transfer and decompensation (15% vs. 15%), as well as an overall composite measure of any complication (79% vs. 72%) (Table 2). However, the hospitalist‐led teams performed more blood transfusions (74% vs. 63%, P = 0.02) and readmission rates were higher (7.3% vs. 3.3%, P = 0.05).

Because of the low event rate of certain adverse outcomes (<20%), we were only able to perform adjusted analyses on 4 outcomes: need for endoscopic therapy (odds ratio [OR], 0.82; 95% confidence interval [CI], 0.491.37), ICU transfer and decompensation (OR, 0.82; 95% CI, 0.451.52), blood transfusion (OR, 1.30; 95% CI, 0.822.04), and any complication (OR, 1.18; 95% CI, 0.711.96). Since outcome differences disappeared after controlling for confounders, the data suggest that overall care provided by hospitalists and nonhospitalists might be equivalenteven in certain outcomes that we were unable to substantiate using multivariable methods.

Efficiency

Efficiency, as measured by LOS and costs, are presented both as means and medians in univariate analyses in Table 2. Median LOS was similar for hospitalist‐led and nonhospitalist‐led teams (4 days). Despite having similar LOS, the median costs of acute UGIH in patients cared for by hospitalists were higher ($7,359.00 vs. $6,181.00; P < 0.01).

After adjusting for demographic factors, Rockall risk score, comorbidity, early EGD, and hospital site, LOS remained similar between the 2 groups. On the other hand, the adjusted cost for UGIH patients cared for by hospitalists and nonhospitalists persisted, with hospitalist care costs $1,502.40 more than their nonhospitalist counterparts (Table 3).

Study Sites

From July 1, 2001 to June 30, 2003, the MCH trial1113 was a prospective, multicenter, observational trial of the care provided by hospitalists to patients admitted to general medical services at 6 academic medical institutions. There were 31,000 consecutive admissions to the general medical services of these participating sites: University of Chicago (Chicago, IL), University of Wisconsin Hospital (Madison, WI), University of Iowa (Iowa City, IA), University of California at San Francisco (San Francisco, CA), University of New Mexico (Albuquerque, NM), and Brigham and Women's Hospital (Boston, MA). The study was approved by the institutional review boards (IRBs) at each of the 6 participating institutions.

MCH Study Patients

Patients were eligible if they were admitted to the general medical services under the care of a hospitalist or nonhospitalist physician. Regardless of the admitting provider, each medical service was composed of rotating senior and junior resident physicians in all 6 sites. Furthermore, patients were 18 years of age or older, and were able to give consent themselves or had an appropriate proxy. Patients with mini‐mental status score of 17 (out of 22), admitted under their primary care physician or to an inpatient gastroenterology service, or transferred from another hospital, were excluded. The MCH study was designed to study the outcomes and efficiency in patients admitted for CHF, pneumonia, UGIH, and end‐of‐life care.

Acute UGIH Patients

Within the MCH‐eligible patients, we identified those with acute UGIH using the following International Classification of Diseases, 9th edition (ICD‐9) codes assigned at discharge: esophageal varices with hemorrhage (456.0, 456.20); Mallory‐Weiss syndrome (530.7); gastric ulcer with hemorrhage (531.00531.61); duodenal ulcer with hemorrhage (532.00532.61); peptic ulcer, site unspecified, with hemorrhage (533.00533.61); gastrojejunal ulcer with hemorrhage (534.00534.61); gastritis with hemorrhage (535.61); angiodysplasia of stomach/duodenum with hemorrhage (537.83); and hematemesis (578.0, 578.9). We also confirmed the diagnosis of UGIH by reviewing patient medical records for observed hematemesis, nasogastric tube aspirate with gross or hemoccult blood, or clinical history of hematemesis, melena, or hematochezia.14, 15

Data

All data were obtained from the 6 hospitals' administrative records, patient interviews, and medical chart abstractions. Dates of admission and discharge, ICD‐9 diagnosis codes, insurance type, age, race, and gender were obtained from administrative data. One‐month follow‐up telephone interviews assessed whether or not patient had any follow‐up appointment or hospital readmissions. Trained abstractors from each site performed manual chart reviews using a standard data collection sheet. The ICD‐9 code designation and chart abstraction methodology were developed prior to the initiation of the study to ensure consistent data collection and reduce bias.

The following data elements were collected: comorbidities, endoscopic findings, inpatient mortality, clinical evidence of rebleeding, endoscopic treatment or gastrointestinal (GI) surgery to control bleeding, repeat EGD, ICU transfer, decompensated comorbid illness requiring continued hospitalization, and blood transfusion (packed red cells, plasma, platelets). Clinical evidence of rebleeding was defined as either hematemesis or melena with decrease in hemoglobin of 2 g in 24 hours with or without hemodynamic compromise.14, 15 For the purpose of this study, recurrent bleeding was defined as clinical evidence of rebleeding, emergency GI surgery for control of UGIH, or repeat EGD before discharge. Furthermore, a composite endpoint termed total complications encompassed all adverse outcomes related to the UGIH hospitalization. The 30‐day readmission variable was defined using readmission identified in administrative records and a 30‐day follow‐up phone call. To guard against recall bias, self‐report data was only included for nonsite admissions.

We defined efficiency in terms of costs and LOS. Total hospital costs were measured using the TSI cost accounting system (Transition Systems, Inc., Boston, MA; now Eclipsys Corporation)16, 17 at 5 out of the 6 participating sites. TSI is a hospital cost accounting software system that integrates resource utilization and financial data already recorded in other hospital databases (such as the billing system, payroll system, and general ledger system).17 Hospital LOS was defined as the number of days from patient admission to the general medicine service until patient discharge.

Provider Specialization: Hospitalists vs. Nonhospitalists

The study was designed as a natural experiment based on a call cycle. The hospitalist‐led teams at each institution alternated in a 4‐day or 5‐day general medicine call cycle with teams led by traditional academic internal medicine attending physicians. All patients were assigned to teams according to their position in the call cycle without regard to whether the attending physician was a hospitalist or a nonhospitalist. Hospitalists are physicians whose primary professional focus is the general medical care of hospitalized patients.18, 19 As previously reported in a related MCH work,11 a hospitalist was also defined as a provider who spends at least 25% of his or her time on an academic inpatient general medicine service. Nonhospitalist physicians were most often outpatient general internal medicine faculty or subspecialists, who attended 1 month per year. Physicians were classified as hospitalists or nonhospitalists according to the designations provided by each site.

UGIH‐specific Confounders

From chart abstraction, we captured severity of illness, comorbidity, and performance of early EGD, variables that can confound analysis in UGIH. To capture severity of illness, a complete Rockall risk score was calculated for each patient. The complete Rockall uses 3 clinical variables (age, shock, and comorbidity) and 2 endoscopic variables (endoscopic diagnosis and stigmata of recent hemorrhage).5, 20 A complete Rockall score of 2 is considered low‐risk for rebleeding or death following admission.21, 22 The accepted definition of low‐risk is <5% recurrent bleeding and <1% mortality risk. A complete Rockall score of 3 to 5 is considered moderate‐risk while 6 is considered high‐risk. Comorbidity was measured using the Charlson comorbidity index.23 Performance of early endoscopy, usually defined as endoscopy performed within 24 hours from presentation, was previously shown to decrease LOS and need for surgical intervention in patients with acute UGIH.24, 25 Documented times of presentation to the emergency department and time of endoscopy performance were collected to calculate for the rate of early endoscopy in our study population.

Statistical Analysis

All statistical analyses were performed using SAS Version 9.1 for Windows (SAS Institute, Cary, NC).

Differences in baseline demographic characteristics of patients and their endoscopic findings were compared between the 2 types of providers. Univariate analyses were also performed to compare the differences in adverse outcomes, LOS, and costs between patients cared for by hospitalists and nonhospitalists. Chi‐square tests were used for categorical variables; while both Wilcoxon rank sum test and Student's t test were used in the analysis of continuous variables.

Next, we performed multivariable analyses to determine the independent association between hospitalist care and the odds of the patients having certain outcomes. However, to prevent overfitting, we only developed regression models for adverse outcomes that have at least 20% event rate.

Multivariable regression models were developed separately for LOS and costs. In contrast with the models on outcomes, analyses of LOS and costs were restricted to: (1) patients who were discharged alive; and (2) to cases with LOS and costs values within 3 standard deviations (SDs) of the mean because of the skewed nature of these data.

All models were adjusted for age, gender, race, insurance type, complete Rockall risk score, performance of early EGD, Charlson comorbidity index, and study site. Final candidate variables in the models were chosen based on stepwise selection, a method very similar to forward selection except that variables selected for the model do not necessarily remain in the model. Effects were entered into and then removed from the model in such a way that each forward selection step can be followed by 1 or more backward elimination steps. The stepwise selection was terminated if no further effect can be added to the model or if the current model was identical to the previous model. The stepwise selection model was generated using statistical criterion of alpha = 0.05 for entry and elimination from the model. Variables that can be a profound source of variation, such as study site and treating physician, were included in the model irrespective of their statistical significance.

To account for clustering of patients treated by the same physician, we used multilevel modeling with SAS PROC GLIMMIX (with random effects). For outcomes (categorical variables), we utilized models with logit‐link and binomial‐distributed errors. As for efficiency (continuous variables with skewed distribution), the multivariable analyses used a generalized linear model with log‐link and assuming gamma‐distributed errors.

Patient Characteristics and Endoscopic Diagnoses

Out of 31,000 patients, the study identified a total of 566 patients (1.8%) with acute UGIH (Table 1). However, 116 patients transferred from another hospital were excluded as their initial management was provided elsewhere, giving a final study sample of 450 patients. Overall, there are 163 admitting physicians from 6 sites, with 39 (24%) classified as hospitalists and 124 (76%) as nonhospitalists. Forty‐two percent (177/450) of patients were cared for by hospitalists. Compared to nonhospitalists, patients admitted to the hospitalist service were older (62.8 vs. 57.7 years, P < 0.01) and with third‐party payor mix differences (P < 0.01). However, there were no statistical differences between patients attended by hospitalists and nonhospitalists with regard to Complete Rockall risk score, Charlson comorbidity index, performance of early endoscopy, and mean hemoglobin values on admission. Upper endoscopy was performed in all patients with distribution of the 3 most common diagnoses being similar (P > 0.05) between hospitalists and nonhospitalists: erosive disease (49.7% vs. 54.6%), peptic ulcer disease (PUD) (48% vs. 46.9%), and varices (18.6% vs. 14.7%).

Clinical Outcomes

Between hospitalists and nonhospitalists, unadjusted outcomes were similar (P > 0.05) for mortality (2.3% vs. 0.4%), recurrent bleeding (11% vs. 11%), need for endoscopic therapy (24% vs. 22%), ICU‐transfer and decompensation (15% vs. 15%), as well as an overall composite measure of any complication (79% vs. 72%) (Table 2). However, the hospitalist‐led teams performed more blood transfusions (74% vs. 63%, P = 0.02) and readmission rates were higher (7.3% vs. 3.3%, P = 0.05).

Because of the low event rate of certain adverse outcomes (<20%), we were only able to perform adjusted analyses on 4 outcomes: need for endoscopic therapy (odds ratio [OR], 0.82; 95% confidence interval [CI], 0.491.37), ICU transfer and decompensation (OR, 0.82; 95% CI, 0.451.52), blood transfusion (OR, 1.30; 95% CI, 0.822.04), and any complication (OR, 1.18; 95% CI, 0.711.96). Since outcome differences disappeared after controlling for confounders, the data suggest that overall care provided by hospitalists and nonhospitalists might be equivalenteven in certain outcomes that we were unable to substantiate using multivariable methods.

Efficiency

Efficiency, as measured by LOS and costs, are presented both as means and medians in univariate analyses in Table 2. Median LOS was similar for hospitalist‐led and nonhospitalist‐led teams (4 days). Despite having similar LOS, the median costs of acute UGIH in patients cared for by hospitalists were higher ($7,359.00 vs. $6,181.00; P < 0.01).

After adjusting for demographic factors, Rockall risk score, comorbidity, early EGD, and hospital site, LOS remained similar between the 2 groups. On the other hand, the adjusted cost for UGIH patients cared for by hospitalists and nonhospitalists persisted, with hospitalist care costs $1,502.40 more than their nonhospitalist counterparts (Table 3).

Study Sites

From July 1, 2001 to June 30, 2003, the MCH trial1113 was a prospective, multicenter, observational trial of the care provided by hospitalists to patients admitted to general medical services at 6 academic medical institutions. There were 31,000 consecutive admissions to the general medical services of these participating sites: University of Chicago (Chicago, IL), University of Wisconsin Hospital (Madison, WI), University of Iowa (Iowa City, IA), University of California at San Francisco (San Francisco, CA), University of New Mexico (Albuquerque, NM), and Brigham and Women's Hospital (Boston, MA). The study was approved by the institutional review boards (IRBs) at each of the 6 participating institutions.

MCH Study Patients

Patients were eligible if they were admitted to the general medical services under the care of a hospitalist or nonhospitalist physician. Regardless of the admitting provider, each medical service was composed of rotating senior and junior resident physicians in all 6 sites. Furthermore, patients were 18 years of age or older, and were able to give consent themselves or had an appropriate proxy. Patients with mini‐mental status score of 17 (out of 22), admitted under their primary care physician or to an inpatient gastroenterology service, or transferred from another hospital, were excluded. The MCH study was designed to study the outcomes and efficiency in patients admitted for CHF, pneumonia, UGIH, and end‐of‐life care.

Acute UGIH Patients

Within the MCH‐eligible patients, we identified those with acute UGIH using the following International Classification of Diseases, 9th edition (ICD‐9) codes assigned at discharge: esophageal varices with hemorrhage (456.0, 456.20); Mallory‐Weiss syndrome (530.7); gastric ulcer with hemorrhage (531.00531.61); duodenal ulcer with hemorrhage (532.00532.61); peptic ulcer, site unspecified, with hemorrhage (533.00533.61); gastrojejunal ulcer with hemorrhage (534.00534.61); gastritis with hemorrhage (535.61); angiodysplasia of stomach/duodenum with hemorrhage (537.83); and hematemesis (578.0, 578.9). We also confirmed the diagnosis of UGIH by reviewing patient medical records for observed hematemesis, nasogastric tube aspirate with gross or hemoccult blood, or clinical history of hematemesis, melena, or hematochezia.14, 15

Data

All data were obtained from the 6 hospitals' administrative records, patient interviews, and medical chart abstractions. Dates of admission and discharge, ICD‐9 diagnosis codes, insurance type, age, race, and gender were obtained from administrative data. One‐month follow‐up telephone interviews assessed whether or not patient had any follow‐up appointment or hospital readmissions. Trained abstractors from each site performed manual chart reviews using a standard data collection sheet. The ICD‐9 code designation and chart abstraction methodology were developed prior to the initiation of the study to ensure consistent data collection and reduce bias.

The following data elements were collected: comorbidities, endoscopic findings, inpatient mortality, clinical evidence of rebleeding, endoscopic treatment or gastrointestinal (GI) surgery to control bleeding, repeat EGD, ICU transfer, decompensated comorbid illness requiring continued hospitalization, and blood transfusion (packed red cells, plasma, platelets). Clinical evidence of rebleeding was defined as either hematemesis or melena with decrease in hemoglobin of 2 g in 24 hours with or without hemodynamic compromise.14, 15 For the purpose of this study, recurrent bleeding was defined as clinical evidence of rebleeding, emergency GI surgery for control of UGIH, or repeat EGD before discharge. Furthermore, a composite endpoint termed total complications encompassed all adverse outcomes related to the UGIH hospitalization. The 30‐day readmission variable was defined using readmission identified in administrative records and a 30‐day follow‐up phone call. To guard against recall bias, self‐report data was only included for nonsite admissions.

We defined efficiency in terms of costs and LOS. Total hospital costs were measured using the TSI cost accounting system (Transition Systems, Inc., Boston, MA; now Eclipsys Corporation)16, 17 at 5 out of the 6 participating sites. TSI is a hospital cost accounting software system that integrates resource utilization and financial data already recorded in other hospital databases (such as the billing system, payroll system, and general ledger system).17 Hospital LOS was defined as the number of days from patient admission to the general medicine service until patient discharge.

Provider Specialization: Hospitalists vs. Nonhospitalists

The study was designed as a natural experiment based on a call cycle. The hospitalist‐led teams at each institution alternated in a 4‐day or 5‐day general medicine call cycle with teams led by traditional academic internal medicine attending physicians. All patients were assigned to teams according to their position in the call cycle without regard to whether the attending physician was a hospitalist or a nonhospitalist. Hospitalists are physicians whose primary professional focus is the general medical care of hospitalized patients.18, 19 As previously reported in a related MCH work,11 a hospitalist was also defined as a provider who spends at least 25% of his or her time on an academic inpatient general medicine service. Nonhospitalist physicians were most often outpatient general internal medicine faculty or subspecialists, who attended 1 month per year. Physicians were classified as hospitalists or nonhospitalists according to the designations provided by each site.

UGIH‐specific Confounders

From chart abstraction, we captured severity of illness, comorbidity, and performance of early EGD, variables that can confound analysis in UGIH. To capture severity of illness, a complete Rockall risk score was calculated for each patient. The complete Rockall uses 3 clinical variables (age, shock, and comorbidity) and 2 endoscopic variables (endoscopic diagnosis and stigmata of recent hemorrhage).5, 20 A complete Rockall score of 2 is considered low‐risk for rebleeding or death following admission.21, 22 The accepted definition of low‐risk is <5% recurrent bleeding and <1% mortality risk. A complete Rockall score of 3 to 5 is considered moderate‐risk while 6 is considered high‐risk. Comorbidity was measured using the Charlson comorbidity index.23 Performance of early endoscopy, usually defined as endoscopy performed within 24 hours from presentation, was previously shown to decrease LOS and need for surgical intervention in patients with acute UGIH.24, 25 Documented times of presentation to the emergency department and time of endoscopy performance were collected to calculate for the rate of early endoscopy in our study population.

Statistical Analysis

All statistical analyses were performed using SAS Version 9.1 for Windows (SAS Institute, Cary, NC).

Differences in baseline demographic characteristics of patients and their endoscopic findings were compared between the 2 types of providers. Univariate analyses were also performed to compare the differences in adverse outcomes, LOS, and costs between patients cared for by hospitalists and nonhospitalists. Chi‐square tests were used for categorical variables; while both Wilcoxon rank sum test and Student's t test were used in the analysis of continuous variables.

Next, we performed multivariable analyses to determine the independent association between hospitalist care and the odds of the patients having certain outcomes. However, to prevent overfitting, we only developed regression models for adverse outcomes that have at least 20% event rate.

Multivariable regression models were developed separately for LOS and costs. In contrast with the models on outcomes, analyses of LOS and costs were restricted to: (1) patients who were discharged alive; and (2) to cases with LOS and costs values within 3 standard deviations (SDs) of the mean because of the skewed nature of these data.

All models were adjusted for age, gender, race, insurance type, complete Rockall risk score, performance of early EGD, Charlson comorbidity index, and study site. Final candidate variables in the models were chosen based on stepwise selection, a method very similar to forward selection except that variables selected for the model do not necessarily remain in the model. Effects were entered into and then removed from the model in such a way that each forward selection step can be followed by 1 or more backward elimination steps. The stepwise selection was terminated if no further effect can be added to the model or if the current model was identical to the previous model. The stepwise selection model was generated using statistical criterion of alpha = 0.05 for entry and elimination from the model. Variables that can be a profound source of variation, such as study site and treating physician, were included in the model irrespective of their statistical significance.

To account for clustering of patients treated by the same physician, we used multilevel modeling with SAS PROC GLIMMIX (with random effects). For outcomes (categorical variables), we utilized models with logit‐link and binomial‐distributed errors. As for efficiency (continuous variables with skewed distribution), the multivariable analyses used a generalized linear model with log‐link and assuming gamma‐distributed errors.

Patient Characteristics and Endoscopic Diagnoses

Out of 31,000 patients, the study identified a total of 566 patients (1.8%) with acute UGIH (Table 1). However, 116 patients transferred from another hospital were excluded as their initial management was provided elsewhere, giving a final study sample of 450 patients. Overall, there are 163 admitting physicians from 6 sites, with 39 (24%) classified as hospitalists and 124 (76%) as nonhospitalists. Forty‐two percent (177/450) of patients were cared for by hospitalists. Compared to nonhospitalists, patients admitted to the hospitalist service were older (62.8 vs. 57.7 years, P < 0.01) and with third‐party payor mix differences (P < 0.01). However, there were no statistical differences between patients attended by hospitalists and nonhospitalists with regard to Complete Rockall risk score, Charlson comorbidity index, performance of early endoscopy, and mean hemoglobin values on admission. Upper endoscopy was performed in all patients with distribution of the 3 most common diagnoses being similar (P > 0.05) between hospitalists and nonhospitalists: erosive disease (49.7% vs. 54.6%), peptic ulcer disease (PUD) (48% vs. 46.9%), and varices (18.6% vs. 14.7%).

Clinical Outcomes

Between hospitalists and nonhospitalists, unadjusted outcomes were similar (P > 0.05) for mortality (2.3% vs. 0.4%), recurrent bleeding (11% vs. 11%), need for endoscopic therapy (24% vs. 22%), ICU‐transfer and decompensation (15% vs. 15%), as well as an overall composite measure of any complication (79% vs. 72%) (Table 2). However, the hospitalist‐led teams performed more blood transfusions (74% vs. 63%, P = 0.02) and readmission rates were higher (7.3% vs. 3.3%, P = 0.05).

Because of the low event rate of certain adverse outcomes (<20%), we were only able to perform adjusted analyses on 4 outcomes: need for endoscopic therapy (odds ratio [OR], 0.82; 95% confidence interval [CI], 0.491.37), ICU transfer and decompensation (OR, 0.82; 95% CI, 0.451.52), blood transfusion (OR, 1.30; 95% CI, 0.822.04), and any complication (OR, 1.18; 95% CI, 0.711.96). Since outcome differences disappeared after controlling for confounders, the data suggest that overall care provided by hospitalists and nonhospitalists might be equivalenteven in certain outcomes that we were unable to substantiate using multivariable methods.

Efficiency

Efficiency, as measured by LOS and costs, are presented both as means and medians in univariate analyses in Table 2. Median LOS was similar for hospitalist‐led and nonhospitalist‐led teams (4 days). Despite having similar LOS, the median costs of acute UGIH in patients cared for by hospitalists were higher ($7,359.00 vs. $6,181.00; P < 0.01).

After adjusting for demographic factors, Rockall risk score, comorbidity, early EGD, and hospital site, LOS remained similar between the 2 groups. On the other hand, the adjusted cost for UGIH patients cared for by hospitalists and nonhospitalists persisted, with hospitalist care costs $1,502.40 more than their nonhospitalist counterparts (Table 3).